xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_priv.h (revision 64b4c4cdfddd31d3a0ec468afccb7ff7cd926682)
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 <uapi/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_obj(*ptr_to_struct))
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 (AMDGPU_GPU_PAGE_SIZE * 2)
106 #define KFD_CWSR_TMA_OFFSET (AMDGPU_GPU_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 /*  KFD_UNMAP_LATENCY_MS is the timeout CP waiting for SDMA preemption. One XCC
115  *  can be associated to 2 SDMA engines. queue_preemption_timeout_ms is the time
116  *  driver waiting for CP returning the UNMAP_QUEUE fence. Thus the math is
117  *  queue_preemption_timeout_ms = sdma_preemption_time * 2 + cp workload
118  *  The format here makes CP workload 10% of total timeout
119  */
120 #define KFD_UNMAP_LATENCY_MS	\
121 	((queue_preemption_timeout_ms - queue_preemption_timeout_ms / 10) >> 1)
122 
123 #define KFD_MAX_SDMA_QUEUES	128
124 
125 /*
126  * 512 = 0x200
127  * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
128  * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
129  * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
130  * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
131  * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
132  */
133 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
134 
135 /**
136  * enum kfd_ioctl_flags - KFD ioctl flags
137  * Various flags that can be set in &amdkfd_ioctl_desc.flags to control how
138  * userspace can use a given ioctl.
139  */
140 enum kfd_ioctl_flags {
141 	/*
142 	 * @KFD_IOC_FLAG_CHECKPOINT_RESTORE:
143 	 * Certain KFD ioctls such as AMDKFD_IOC_CRIU_OP can potentially
144 	 * perform privileged operations and load arbitrary data into MQDs and
145 	 * eventually HQD registers when the queue is mapped by HWS. In order to
146 	 * prevent this we should perform additional security checks.
147 	 *
148 	 * This is equivalent to callers with the CHECKPOINT_RESTORE capability.
149 	 *
150 	 * Note: Since earlier versions of docker do not support CHECKPOINT_RESTORE,
151 	 * we also allow ioctls with SYS_ADMIN capability.
152 	 */
153 	KFD_IOC_FLAG_CHECKPOINT_RESTORE = BIT(0),
154 };
155 /*
156  * Kernel module parameter to specify maximum number of supported queues per
157  * device
158  */
159 extern int max_num_of_queues_per_device;
160 
161 
162 /* Kernel module parameter to specify the scheduling policy */
163 extern int sched_policy;
164 
165 /*
166  * Kernel module parameter to specify the maximum process
167  * number per HW scheduler
168  */
169 extern int hws_max_conc_proc;
170 
171 extern int cwsr_enable;
172 
173 /*
174  * Kernel module parameter to specify whether to send sigterm to HSA process on
175  * unhandled exception
176  */
177 extern int send_sigterm;
178 
179 /*
180  * This kernel module is used to simulate large bar machine on non-large bar
181  * enabled machines.
182  */
183 extern int debug_largebar;
184 
185 /* Set sh_mem_config.retry_disable on GFX v9 */
186 extern int amdgpu_noretry;
187 
188 /* Halt if HWS hang is detected */
189 extern int halt_if_hws_hang;
190 
191 /* Whether MEC FW support GWS barriers */
192 extern bool hws_gws_support;
193 
194 /* Queue preemption timeout in ms */
195 extern int queue_preemption_timeout_ms;
196 
197 /*
198  * Don't evict process queues on vm fault
199  */
200 extern int amdgpu_no_queue_eviction_on_vm_fault;
201 
202 /* Enable eviction debug messages */
203 extern bool debug_evictions;
204 
205 extern struct mutex kfd_processes_mutex;
206 
207 enum cache_policy {
208 	cache_policy_coherent,
209 	cache_policy_noncoherent
210 };
211 
212 #define KFD_GC_VERSION(dev) (amdgpu_ip_version((dev)->adev, GC_HWIP, 0))
213 #define KFD_IS_SOC15(dev)   ((KFD_GC_VERSION(dev)) >= (IP_VERSION(9, 0, 1)))
214 #define KFD_SUPPORT_XNACK_PER_PROCESS(dev)\
215 	((KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2)) ||	\
216 	 (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) ||	\
217 	 (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 4)) ||	\
218 	 (KFD_GC_VERSION(dev) == IP_VERSION(9, 5, 0)))
219 
220 struct kfd_node;
221 
222 struct kfd_event_interrupt_class {
223 	bool (*interrupt_isr)(struct kfd_node *dev,
224 			const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
225 			bool *patched_flag);
226 	void (*interrupt_wq)(struct kfd_node *dev,
227 			const uint32_t *ih_ring_entry);
228 };
229 
230 struct kfd_device_info {
231 	uint32_t gfx_target_version;
232 	const struct kfd_event_interrupt_class *event_interrupt_class;
233 	unsigned int max_pasid_bits;
234 	unsigned int max_no_of_hqd;
235 	unsigned int doorbell_size;
236 	size_t ih_ring_entry_size;
237 	uint8_t num_of_watch_points;
238 	uint16_t mqd_size_aligned;
239 	bool supports_cwsr;
240 	bool needs_pci_atomics;
241 	uint32_t no_atomic_fw_version;
242 	unsigned int num_sdma_queues_per_engine;
243 	unsigned int num_reserved_sdma_queues_per_engine;
244 };
245 
246 unsigned int kfd_get_num_sdma_engines(struct kfd_node *kdev);
247 unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *kdev);
248 
249 struct kfd_mem_obj {
250 	uint32_t range_start;
251 	uint32_t range_end;
252 	uint64_t gpu_addr;
253 	uint32_t *cpu_ptr;
254 	void *mem;
255 };
256 
257 struct kfd_vmid_info {
258 	uint32_t first_vmid_kfd;
259 	uint32_t last_vmid_kfd;
260 	uint32_t vmid_num_kfd;
261 };
262 
263 #define MAX_KFD_NODES	8
264 
265 struct kfd_dev;
266 
267 struct kfd_node {
268 	unsigned int node_id;
269 	struct amdgpu_device *adev;     /* Duplicated here along with keeping
270 					 * a copy in kfd_dev to save a hop
271 					 */
272 	const struct kfd2kgd_calls *kfd2kgd; /* Duplicated here along with
273 					      * keeping a copy in kfd_dev to
274 					      * save a hop
275 					      */
276 	struct kfd_vmid_info vm_info;
277 	unsigned int id;                /* topology stub index */
278 	uint32_t xcc_mask; /* Instance mask of XCCs present */
279 	struct amdgpu_xcp *xcp;
280 
281 	/* Interrupts */
282 	struct kfifo ih_fifo;
283 	struct work_struct interrupt_work;
284 	spinlock_t interrupt_lock;
285 
286 	/*
287 	 * Interrupts of interest to KFD are copied
288 	 * from the HW ring into a SW ring.
289 	 */
290 	bool interrupts_active;
291 	uint32_t interrupt_bitmap; /* Only used for GFX 9.4.3 */
292 
293 	/* QCM Device instance */
294 	struct device_queue_manager *dqm;
295 
296 	/* Global GWS resource shared between processes */
297 	void *gws;
298 
299 	/* Clients watching SMI events */
300 	struct list_head smi_clients;
301 	spinlock_t smi_lock;
302 	uint32_t reset_seq_num;
303 
304 	/* SRAM ECC flag */
305 	atomic_t sram_ecc_flag;
306 
307 	/*spm process id */
308 	unsigned int spm_pasid;
309 
310 	/* Maximum process number mapped to HW scheduler */
311 	unsigned int max_proc_per_quantum;
312 
313 	unsigned int compute_vmid_bitmap;
314 
315 	struct kfd_local_mem_info local_mem_info;
316 
317 	struct kfd_dev *kfd;
318 
319 	/* Track per device allocated watch points */
320 	uint32_t alloc_watch_ids;
321 	spinlock_t watch_points_lock;
322 };
323 
324 struct kfd_dev {
325 	struct amdgpu_device *adev;
326 
327 	struct kfd_device_info device_info;
328 
329 	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
330 					   * page used by kernel queue
331 					   */
332 
333 	struct kgd2kfd_shared_resources shared_resources;
334 
335 	const struct kfd2kgd_calls *kfd2kgd;
336 	struct mutex doorbell_mutex;
337 
338 	void *gtt_mem;
339 	uint64_t gtt_start_gpu_addr;
340 	void *gtt_start_cpu_ptr;
341 	void *gtt_sa_bitmap;
342 	struct mutex gtt_sa_lock;
343 	unsigned int gtt_sa_chunk_size;
344 	unsigned int gtt_sa_num_of_chunks;
345 
346 	bool init_complete;
347 
348 	/* Firmware versions */
349 	uint16_t mec_fw_version;
350 	uint16_t mec2_fw_version;
351 	uint16_t sdma_fw_version;
352 
353 	/* CWSR */
354 	bool cwsr_enabled;
355 	const void *cwsr_isa;
356 	unsigned int cwsr_isa_size;
357 
358 	/* xGMI */
359 	uint64_t hive_id;
360 
361 	bool pci_atomic_requested;
362 
363 	/* Compute Profile ref. count */
364 	atomic_t compute_profile;
365 
366 	struct ida doorbell_ida;
367 	unsigned int max_doorbell_slices;
368 
369 	int noretry;
370 
371 	struct kfd_node *nodes[MAX_KFD_NODES];
372 	unsigned int num_nodes;
373 
374 	struct workqueue_struct *ih_wq;
375 
376 	/* Kernel doorbells for KFD device */
377 	struct amdgpu_bo *doorbells;
378 
379 	/* bitmap for dynamic doorbell allocation from doorbell object */
380 	unsigned long *doorbell_bitmap;
381 
382 	/* for dynamic partitioning */
383 	int kfd_dev_lock;
384 
385 	atomic_t kfd_processes_count;
386 };
387 
388 enum kfd_mempool {
389 	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
390 	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
391 	KFD_MEMPOOL_FRAMEBUFFER = 3,
392 };
393 
394 /* Character device interface */
395 int kfd_chardev_init(void);
396 void kfd_chardev_exit(void);
397 
398 /**
399  * enum kfd_unmap_queues_filter - Enum for queue filters.
400  *
401  * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
402  *						running queues list.
403  *
404  * @KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: Preempts all non-static queues
405  *						in the run list.
406  *
407  * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
408  *						specific process.
409  *
410  */
411 enum kfd_unmap_queues_filter {
412 	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES = 1,
413 	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES = 2,
414 	KFD_UNMAP_QUEUES_FILTER_BY_PASID = 3
415 };
416 
417 /**
418  * enum kfd_queue_type - Enum for various queue types.
419  *
420  * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
421  *
422  * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
423  *
424  * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
425  *
426  * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
427  *
428  * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
429  *
430  * @KFD_QUEUE_TYPE_SDMA_BY_ENG_ID:  SDMA user mode queue with target SDMA engine ID.
431  */
432 enum kfd_queue_type  {
433 	KFD_QUEUE_TYPE_COMPUTE,
434 	KFD_QUEUE_TYPE_SDMA,
435 	KFD_QUEUE_TYPE_HIQ,
436 	KFD_QUEUE_TYPE_SDMA_XGMI,
437 	KFD_QUEUE_TYPE_SDMA_BY_ENG_ID
438 };
439 
440 enum kfd_queue_format {
441 	KFD_QUEUE_FORMAT_PM4,
442 	KFD_QUEUE_FORMAT_AQL
443 };
444 
445 enum KFD_QUEUE_PRIORITY {
446 	KFD_QUEUE_PRIORITY_MINIMUM = 0,
447 	KFD_QUEUE_PRIORITY_MAXIMUM = 15
448 };
449 
450 /**
451  * struct queue_properties
452  *
453  * @type: The queue type.
454  *
455  * @queue_id: Queue identifier.
456  *
457  * @queue_address: Queue ring buffer address.
458  *
459  * @queue_size: Queue ring buffer size.
460  *
461  * @priority: Defines the queue priority relative to other queues in the
462  * process.
463  * This is just an indication and HW scheduling may override the priority as
464  * necessary while keeping the relative prioritization.
465  * the priority granularity is from 0 to f which f is the highest priority.
466  * currently all queues are initialized with the highest priority.
467  *
468  * @queue_percent: This field is partially implemented and currently a zero in
469  * this field defines that the queue is non active.
470  *
471  * @read_ptr: User space address which points to the number of dwords the
472  * cp read from the ring buffer. This field updates automatically by the H/W.
473  *
474  * @write_ptr: Defines the number of dwords written to the ring buffer.
475  *
476  * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
477  * buffer. This field should be similar to write_ptr and the user should
478  * update this field after updating the write_ptr.
479  *
480  * @doorbell_off: The doorbell offset in the doorbell pci-bar.
481  *
482  * @is_interop: Defines if this is a interop queue. Interop queue means that
483  * the queue can access both graphics and compute resources.
484  *
485  * @is_evicted: Defines if the queue is evicted. Only active queues
486  * are evicted, rendering them inactive.
487  *
488  * @is_active: Defines if the queue is active or not. @is_active and
489  * @is_evicted are protected by the DQM lock.
490  *
491  * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
492  * @is_gws should be protected by the DQM lock, since changing it can yield the
493  * possibility of updating DQM state on number of GWS queues.
494  *
495  * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
496  * of the queue.
497  *
498  * This structure represents the queue properties for each queue no matter if
499  * it's user mode or kernel mode queue.
500  *
501  */
502 
503 struct queue_properties {
504 	enum kfd_queue_type type;
505 	enum kfd_queue_format format;
506 	unsigned int queue_id;
507 	uint64_t queue_address;
508 	uint64_t queue_size;
509 	uint64_t metadata_queue_size;
510 	uint32_t priority;
511 	uint32_t queue_percent;
512 	void __user *read_ptr;
513 	void __user *write_ptr;
514 	void __iomem *doorbell_ptr;
515 	uint32_t doorbell_off;
516 	bool is_interop;
517 	bool is_evicted;
518 	bool is_suspended;
519 	bool is_being_destroyed;
520 	bool is_active;
521 	bool is_gws;
522 	uint32_t pm4_target_xcc;
523 	bool is_dbg_wa;
524 	bool is_user_cu_masked;
525 	bool is_reset;
526 	/* Not relevant for user mode queues in cp scheduling */
527 	unsigned int vmid;
528 	/* Relevant only for sdma queues*/
529 	uint32_t sdma_engine_id;
530 	uint32_t sdma_queue_id;
531 	uint32_t sdma_vm_addr;
532 	/* Relevant only for VI */
533 	uint64_t eop_ring_buffer_address;
534 	uint32_t eop_ring_buffer_size;
535 	uint64_t ctx_save_restore_area_address;
536 	uint32_t ctx_save_restore_area_size;
537 	uint32_t ctl_stack_size;
538 	uint64_t tba_addr;
539 	uint64_t tma_addr;
540 	uint64_t exception_status;
541 
542 	struct amdgpu_bo *wptr_bo;
543 	struct amdgpu_bo *rptr_bo;
544 	struct amdgpu_bo *ring_bo;
545 	struct amdgpu_bo *eop_buf_bo;
546 	struct amdgpu_bo *cwsr_bo;
547 };
548 
549 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\
550 			    (q).queue_address != 0 &&	\
551 			    (q).queue_percent > 0 &&	\
552 			    !(q).is_evicted &&		\
553 			    !(q).is_suspended)
554 
555 enum mqd_update_flag {
556 	UPDATE_FLAG_DBG_WA_ENABLE = 1,
557 	UPDATE_FLAG_DBG_WA_DISABLE = 2,
558 	UPDATE_FLAG_IS_GWS = 4, /* quirk for gfx9 IP */
559 };
560 
561 struct mqd_update_info {
562 	union {
563 		struct {
564 			uint32_t count; /* Must be a multiple of 32 */
565 			uint32_t *ptr;
566 		} cu_mask;
567 	};
568 	enum mqd_update_flag update_flag;
569 };
570 
571 /**
572  * struct queue
573  *
574  * @list: Queue linked list.
575  *
576  * @mqd: The queue MQD (memory queue descriptor).
577  *
578  * @mqd_mem_obj: The MQD local gpu memory object.
579  *
580  * @gart_mqd_addr: The MQD gart mc address.
581  *
582  * @properties: The queue properties.
583  *
584  * @mec: Used only in no cp scheduling mode and identifies to micro engine id
585  *	 that the queue should be executed on.
586  *
587  * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
588  *	  id.
589  *
590  * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
591  *
592  * @process: The kfd process that created this queue.
593  *
594  * @device: The kfd device that created this queue.
595  *
596  * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
597  * otherwise.
598  *
599  * This structure represents user mode compute queues.
600  * It contains all the necessary data to handle such queues.
601  *
602  */
603 
604 struct queue {
605 	struct list_head list;
606 	void *mqd;
607 	struct kfd_mem_obj *mqd_mem_obj;
608 	uint64_t gart_mqd_addr;
609 	struct queue_properties properties;
610 
611 	uint32_t mec;
612 	uint32_t pipe;
613 	uint32_t queue;
614 
615 	unsigned int sdma_id;
616 	unsigned int doorbell_id;
617 
618 	struct kfd_process	*process;
619 	struct kfd_node		*device;
620 	void *gws;
621 
622 	/* procfs */
623 	struct kobject kobj;
624 
625 	void *gang_ctx_bo;
626 	uint64_t gang_ctx_gpu_addr;
627 	void *gang_ctx_cpu_ptr;
628 
629 	struct amdgpu_bo *wptr_bo_gart;
630 };
631 
632 enum KFD_MQD_TYPE {
633 	KFD_MQD_TYPE_HIQ = 0,		/* for hiq */
634 	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
635 	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
636 	KFD_MQD_TYPE_DIQ,		/* for diq */
637 	KFD_MQD_TYPE_MAX
638 };
639 
640 enum KFD_PIPE_PRIORITY {
641 	KFD_PIPE_PRIORITY_CS_LOW = 0,
642 	KFD_PIPE_PRIORITY_CS_MEDIUM,
643 	KFD_PIPE_PRIORITY_CS_HIGH
644 };
645 
646 struct scheduling_resources {
647 	unsigned int vmid_mask;
648 	enum kfd_queue_type type;
649 	uint64_t queue_mask;
650 	uint64_t gws_mask;
651 	uint32_t oac_mask;
652 	uint32_t gds_heap_base;
653 	uint32_t gds_heap_size;
654 };
655 
656 struct process_queue_manager {
657 	/* data */
658 	struct kfd_process	*process;
659 	struct list_head	queues;
660 	unsigned long		*queue_slot_bitmap;
661 };
662 
663 struct qcm_process_device {
664 	/* The Device Queue Manager that owns this data */
665 	struct device_queue_manager *dqm;
666 	struct process_queue_manager *pqm;
667 	/* Queues list */
668 	struct list_head queues_list;
669 	struct list_head priv_queue_list;
670 
671 	unsigned int queue_count;
672 	unsigned int vmid;
673 	bool is_debug;
674 	unsigned int evicted; /* eviction counter, 0=active */
675 
676 	/* This flag tells if we should reset all wavefronts on
677 	 * process termination
678 	 */
679 	bool reset_wavefronts;
680 
681 	/* This flag tells us if this process has a GWS-capable
682 	 * queue that will be mapped into the runlist. It's
683 	 * possible to request a GWS BO, but not have the queue
684 	 * currently mapped, and this changes how the MAP_PROCESS
685 	 * PM4 packet is configured.
686 	 */
687 	bool mapped_gws_queue;
688 
689 	/* All the memory management data should be here too */
690 	uint64_t gds_context_area;
691 	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
692 	uint64_t page_table_base;
693 	uint32_t sh_mem_config;
694 	uint32_t sh_mem_bases;
695 	uint32_t sh_mem_ape1_base;
696 	uint32_t sh_mem_ape1_limit;
697 	uint32_t gds_size;
698 	uint32_t num_gws;
699 	uint32_t num_oac;
700 	uint32_t sh_hidden_private_base;
701 	uint32_t vm_cntx_cntl;
702 
703 	/* CWSR memory */
704 	struct kgd_mem *cwsr_mem;
705 	void *cwsr_kaddr;
706 	uint64_t cwsr_base;
707 	uint64_t tba_addr;
708 	uint64_t tma_addr;
709 
710 	/* IB memory */
711 	struct kgd_mem *ib_mem;
712 	uint64_t ib_base;
713 	void *ib_kaddr;
714 
715 	/* doorbells for kfd process */
716 	struct amdgpu_bo *proc_doorbells;
717 
718 	/* bitmap for dynamic doorbell allocation from the bo */
719 	unsigned long *doorbell_bitmap;
720 };
721 
722 /* KFD Memory Eviction */
723 
724 /* Approx. wait time before attempting to restore evicted BOs */
725 #define PROCESS_RESTORE_TIME_MS 100
726 /* Approx. back off time if restore fails due to lack of memory */
727 #define PROCESS_BACK_OFF_TIME_MS 100
728 /* Approx. time before evicting the process again */
729 #define PROCESS_ACTIVE_TIME_MS 10
730 
731 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
732  * idr_handle in the least significant 4 bytes
733  */
734 #define MAKE_HANDLE(gpu_id, idr_handle) \
735 	(((uint64_t)(gpu_id) << 32) + idr_handle)
736 #define GET_GPU_ID(handle) (handle >> 32)
737 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
738 
739 enum kfd_pdd_bound {
740 	PDD_UNBOUND = 0,
741 	PDD_BOUND,
742 	PDD_BOUND_SUSPENDED,
743 };
744 
745 #define MAX_SYSFS_FILENAME_LEN 15
746 
747 /*
748  * SDMA counter runs at 100MHz frequency.
749  * We display SDMA activity in microsecond granularity in sysfs.
750  * As a result, the divisor is 100.
751  */
752 #define SDMA_ACTIVITY_DIVISOR  100
753 
754 /* Data that is per-process-per device. */
755 struct kfd_process_device {
756 	/* The device that owns this data. */
757 	struct kfd_node *dev;
758 
759 	/* The process that owns this kfd_process_device. */
760 	struct kfd_process *process;
761 
762 	/* per-process-per device QCM data structure */
763 	struct qcm_process_device qpd;
764 
765 	/*Apertures*/
766 	uint64_t lds_base;
767 	uint64_t lds_limit;
768 	uint64_t gpuvm_base;
769 	uint64_t gpuvm_limit;
770 	uint64_t scratch_base;
771 	uint64_t scratch_limit;
772 
773 	/* VM context for GPUVM allocations */
774 	struct file *drm_file;
775 	void *drm_priv;
776 
777 	/* GPUVM allocations storage */
778 	struct idr alloc_idr;
779 
780 	/* Flag used to tell the pdd has dequeued from the dqm.
781 	 * This is used to prevent dev->dqm->ops.process_termination() from
782 	 * being called twice when it is already called in IOMMU callback
783 	 * function.
784 	 */
785 	bool already_dequeued;
786 	bool runtime_inuse;
787 
788 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
789 	enum kfd_pdd_bound bound;
790 
791 	/* VRAM usage */
792 	atomic64_t vram_usage;
793 	struct attribute attr_vram;
794 	char vram_filename[MAX_SYSFS_FILENAME_LEN];
795 
796 	/* SDMA activity tracking */
797 	uint64_t sdma_past_activity_counter;
798 	struct attribute attr_sdma;
799 	char sdma_filename[MAX_SYSFS_FILENAME_LEN];
800 
801 	/* Eviction activity tracking */
802 	uint64_t last_evict_timestamp;
803 	atomic64_t evict_duration_counter;
804 	struct attribute attr_evict;
805 
806 	struct kobject *kobj_stats;
807 
808 	/*
809 	 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
810 	 * that is associated with device encoded by "this" struct instance. The
811 	 * value reflects CU usage by all of the waves launched by this process
812 	 * on this device. A very important property of occupancy parameter is
813 	 * that its value is a snapshot of current use.
814 	 *
815 	 * Following is to be noted regarding how this parameter is reported:
816 	 *
817 	 *  The number of waves that a CU can launch is limited by couple of
818 	 *  parameters. These are encoded by struct amdgpu_cu_info instance
819 	 *  that is part of every device definition. For GFX9 devices this
820 	 *  translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
821 	 *  do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
822 	 *  when they do use scratch memory. This could change for future
823 	 *  devices and therefore this example should be considered as a guide.
824 	 *
825 	 *  All CU's of a device are available for the process. This may not be true
826 	 *  under certain conditions - e.g. CU masking.
827 	 *
828 	 *  Finally number of CU's that are occupied by a process is affected by both
829 	 *  number of CU's a device has along with number of other competing processes
830 	 */
831 	struct attribute attr_cu_occupancy;
832 
833 	/* sysfs counters for GPU retry fault and page migration tracking */
834 	struct kobject *kobj_counters;
835 	struct attribute attr_faults;
836 	struct attribute attr_page_in;
837 	struct attribute attr_page_out;
838 	uint64_t faults;
839 	uint64_t page_in;
840 	uint64_t page_out;
841 
842 	/* Exception code status*/
843 	uint64_t exception_status;
844 	void *vm_fault_exc_data;
845 	size_t vm_fault_exc_data_size;
846 
847 	/* Tracks debug per-vmid request settings */
848 	uint32_t spi_dbg_override;
849 	uint32_t spi_dbg_launch_mode;
850 	uint32_t watch_points[4];
851 	uint32_t alloc_watch_ids;
852 
853 	/*
854 	 * If this process has been checkpointed before, then the user
855 	 * application will use the original gpu_id on the
856 	 * checkpointed node to refer to this device.
857 	 */
858 	uint32_t user_gpu_id;
859 
860 	void *proc_ctx_bo;
861 	uint64_t proc_ctx_gpu_addr;
862 	void *proc_ctx_cpu_ptr;
863 
864 	/* Tracks queue reset status */
865 	bool has_reset_queue;
866 
867 	u32 pasid;
868 };
869 
870 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
871 
872 struct svm_range_list {
873 	struct mutex			lock;
874 	struct rb_root_cached		objects;
875 	struct list_head		list;
876 	struct work_struct		deferred_list_work;
877 	struct list_head		deferred_range_list;
878 	struct list_head                criu_svm_metadata_list;
879 	spinlock_t			deferred_list_lock;
880 	atomic_t			evicted_ranges;
881 	atomic_t			drain_pagefaults;
882 	struct delayed_work		restore_work;
883 	DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);
884 	struct task_struct		*faulting_task;
885 	/* check point ts decides if page fault recovery need be dropped */
886 	uint64_t			checkpoint_ts[MAX_GPU_INSTANCE];
887 
888 	/* Default granularity to use in buffer migration
889 	 * and restoration of backing memory while handling
890 	 * recoverable page faults
891 	 */
892 	uint8_t default_granularity;
893 };
894 
895 /* Process data */
896 struct kfd_process {
897 	/*
898 	 * kfd_process are stored in an mm_struct*->kfd_process*
899 	 * hash table (kfd_processes in kfd_process.c)
900 	 */
901 	struct hlist_node kfd_processes;
902 
903 	/*
904 	 * Opaque pointer to mm_struct. We don't hold a reference to
905 	 * it so it should never be dereferenced from here. This is
906 	 * only used for looking up processes by their mm.
907 	 */
908 	void *mm;
909 
910 	struct kref ref;
911 	struct work_struct release_work;
912 
913 	struct mutex mutex;
914 
915 	/*
916 	 * In any process, the thread that started main() is the lead
917 	 * thread and outlives the rest.
918 	 * It is here because amd_iommu_bind_pasid wants a task_struct.
919 	 * It can also be used for safely getting a reference to the
920 	 * mm_struct of the process.
921 	 */
922 	struct task_struct *lead_thread;
923 
924 	/* We want to receive a notification when the mm_struct is destroyed */
925 	struct mmu_notifier mmu_notifier;
926 
927 	/*
928 	 * Array of kfd_process_device pointers,
929 	 * one for each device the process is using.
930 	 */
931 	struct kfd_process_device *pdds[MAX_GPU_INSTANCE];
932 	uint32_t n_pdds;
933 
934 	struct process_queue_manager pqm;
935 
936 	/*Is the user space process 32 bit?*/
937 	bool is_32bit_user_mode;
938 
939 	/* Event-related data */
940 	struct mutex event_mutex;
941 	/* Event ID allocator and lookup */
942 	struct idr event_idr;
943 	/* Event page */
944 	u64 signal_handle;
945 	struct kfd_signal_page *signal_page;
946 	size_t signal_mapped_size;
947 	size_t signal_event_count;
948 	bool signal_event_limit_reached;
949 
950 	/* Information used for memory eviction */
951 	void *kgd_process_info;
952 	/* Eviction fence that is attached to all the BOs of this process. The
953 	 * fence will be triggered during eviction and new one will be created
954 	 * during restore
955 	 */
956 	struct dma_fence __rcu *ef;
957 
958 	/* Work items for evicting and restoring BOs */
959 	struct delayed_work eviction_work;
960 	struct delayed_work restore_work;
961 	/* seqno of the last scheduled eviction */
962 	unsigned int last_eviction_seqno;
963 	/* Approx. the last timestamp (in jiffies) when the process was
964 	 * restored after an eviction
965 	 */
966 	unsigned long last_restore_timestamp;
967 
968 	/* Indicates device process is debug attached with reserved vmid. */
969 	bool debug_trap_enabled;
970 
971 	/* per-process-per device debug event fd file */
972 	struct file *dbg_ev_file;
973 
974 	/* If the process is a kfd debugger, we need to know so we can clean
975 	 * up at exit time.  If a process enables debugging on itself, it does
976 	 * its own clean-up, so we don't set the flag here.  We track this by
977 	 * counting the number of processes this process is debugging.
978 	 */
979 	atomic_t debugged_process_count;
980 
981 	/* If the process is a debugged, this is the debugger process */
982 	struct kfd_process *debugger_process;
983 
984 	/* Kobj for our procfs */
985 	struct kobject *kobj;
986 	struct kobject *kobj_queues;
987 	struct attribute attr_pasid;
988 
989 	/* Exception code enable mask and status */
990 	uint64_t exception_enable_mask;
991 	uint64_t exception_status;
992 
993 	/* Used to drain stale interrupts */
994 	wait_queue_head_t wait_irq_drain;
995 	bool irq_drain_is_open;
996 
997 	/* shared virtual memory registered by this process */
998 	struct svm_range_list svms;
999 
1000 	bool xnack_enabled;
1001 
1002 	/* Work area for debugger event writer worker. */
1003 	struct work_struct debug_event_workarea;
1004 
1005 	/* Tracks debug per-vmid request for debug flags */
1006 	u32 dbg_flags;
1007 
1008 	atomic_t poison;
1009 	/* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */
1010 	bool queues_paused;
1011 
1012 	/* Tracks runtime enable status */
1013 	struct semaphore runtime_enable_sema;
1014 	bool is_runtime_retry;
1015 	struct kfd_runtime_info runtime_info;
1016 
1017 	/* if gpu page fault sent to KFD */
1018 	bool gpu_page_fault;
1019 
1020 	/*kfd context id */
1021 	u16 context_id;
1022 
1023 	/* The primary kfd_process allocating IDs for its secondary kfd_process, 0 for primary kfd_process */
1024 	struct ida id_table;
1025 
1026 };
1027 
1028 #define KFD_PROCESS_TABLE_SIZE 8 /* bits: 256 entries */
1029 #define KFD_CONTEXT_ID_PRIMARY	0xFFFF
1030 #define KFD_CONTEXT_ID_MIN 0
1031 
1032 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
1033 extern struct srcu_struct kfd_processes_srcu;
1034 
1035 /**
1036  * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
1037  *
1038  * @filep: pointer to file structure.
1039  * @p: amdkfd process pointer.
1040  * @data: pointer to arg that was copied from user.
1041  *
1042  * Return: returns ioctl completion code.
1043  */
1044 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
1045 				void *data);
1046 
1047 typedef int amdkfd_ioctl_validate_t(void *kdata, unsigned int usize);
1048 
1049 struct amdkfd_ioctl_desc {
1050 	unsigned int cmd;
1051 	int flags;
1052 	amdkfd_ioctl_t *func;
1053 	amdkfd_ioctl_validate_t *validate;
1054 	unsigned int cmd_drv;
1055 	const char *name;
1056 };
1057 bool kfd_dev_is_large_bar(struct kfd_node *dev);
1058 
1059 struct kfd_process *create_process(const struct task_struct *thread, bool primary);
1060 int kfd_process_create_wq(void);
1061 void kfd_process_destroy_wq(void);
1062 void kfd_cleanup_processes(void);
1063 struct kfd_process *kfd_create_process(struct task_struct *thread);
1064 int kfd_create_process_sysfs(struct kfd_process *process);
1065 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid,
1066 						 struct kfd_process_device **pdd);
1067 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
1068 struct kfd_process *kfd_lookup_process_by_id(const struct mm_struct *mm, u16 id);
1069 
1070 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
1071 int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
1072 				uint32_t *gpuid, uint32_t *gpuidx);
1073 static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
1074 				uint32_t gpuidx, uint32_t *gpuid) {
1075 	return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
1076 }
1077 static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
1078 				struct kfd_process *p, uint32_t gpuidx) {
1079 	return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;
1080 }
1081 
1082 void kfd_unref_process(struct kfd_process *p);
1083 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger);
1084 int kfd_process_restore_queues(struct kfd_process *p);
1085 void kfd_suspend_all_processes(void);
1086 int kfd_resume_all_processes(void);
1087 
1088 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *process,
1089 							 uint32_t gpu_id);
1090 
1091 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id);
1092 
1093 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1094 			       struct file *drm_file);
1095 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1096 						struct kfd_process *p);
1097 struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1098 							struct kfd_process *p);
1099 struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1100 							struct kfd_process *p);
1101 
1102 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
1103 
1104 void kfd_process_notifier_release_internal(struct kfd_process *p);
1105 
1106 /* KFD process API for creating and translating handles */
1107 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1108 					void *mem);
1109 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
1110 					int handle);
1111 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1112 					int handle);
1113 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid);
1114 
1115 /* PASIDs */
1116 int kfd_pasid_init(void);
1117 void kfd_pasid_exit(void);
1118 u32 kfd_pasid_alloc(void);
1119 void kfd_pasid_free(u32 pasid);
1120 
1121 /* Doorbells */
1122 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
1123 int kfd_doorbell_init(struct kfd_dev *kfd);
1124 void kfd_doorbell_fini(struct kfd_dev *kfd);
1125 int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process,
1126 		      struct vm_area_struct *vma);
1127 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
1128 					unsigned int *doorbell_off);
1129 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
1130 u32 read_kernel_doorbell(u32 __iomem *db);
1131 void write_kernel_doorbell(void __iomem *db, u32 value);
1132 void write_kernel_doorbell64(void __iomem *db, u64 value);
1133 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
1134 					struct kfd_process_device *pdd,
1135 					unsigned int doorbell_id);
1136 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
1137 int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
1138 				struct kfd_process_device *pdd);
1139 void kfd_free_process_doorbells(struct kfd_dev *kfd,
1140 				struct kfd_process_device *pdd);
1141 /* GTT Sub-Allocator */
1142 
1143 int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size,
1144 			struct kfd_mem_obj **mem_obj);
1145 
1146 int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj);
1147 
1148 extern struct device *kfd_device;
1149 
1150 /* KFD's procfs */
1151 void kfd_procfs_init(void);
1152 void kfd_procfs_shutdown(void);
1153 int kfd_procfs_add_queue(struct queue *q);
1154 void kfd_procfs_del_queue(struct queue *q);
1155 
1156 /* Topology */
1157 int kfd_topology_init(void);
1158 void kfd_topology_shutdown(void);
1159 int kfd_topology_add_device(struct kfd_node *gpu);
1160 int kfd_topology_remove_device(struct kfd_node *gpu);
1161 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
1162 						uint32_t proximity_domain);
1163 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
1164 						uint32_t proximity_domain);
1165 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
1166 struct kfd_node *kfd_device_by_id(uint32_t gpu_id);
1167 static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t node_id,
1168 					uint32_t vmid)
1169 {
1170 	return (node->interrupt_bitmap & (1 << node_id)) != 0 &&
1171 	       (node->compute_vmid_bitmap & (1 << vmid)) != 0;
1172 }
1173 static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev,
1174 					uint32_t node_id, uint32_t vmid) {
1175 	struct kfd_dev *dev = adev->kfd.dev;
1176 	uint32_t i;
1177 
1178 	/*
1179 	 * On multi-aid system, attempt per-node matching. Otherwise,
1180 	 * fall back to the first node.
1181 	 */
1182 	if (!amdgpu_is_multi_aid(adev))
1183 		return dev->nodes[0];
1184 
1185 	for (i = 0; i < dev->num_nodes; i++)
1186 		if (kfd_irq_is_from_node(dev->nodes[i], node_id, vmid))
1187 			return dev->nodes[i];
1188 
1189 	return NULL;
1190 }
1191 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev);
1192 uint32_t kfd_topology_get_num_devices(void);
1193 int kfd_numa_node_to_apic_id(int numa_node_id);
1194 uint32_t kfd_gpu_node_num(void);
1195 
1196 /* Interrupts */
1197 #define	KFD_IRQ_FENCE_CLIENTID	0xff
1198 #define	KFD_IRQ_FENCE_SOURCEID	0xff
1199 #define	KFD_IRQ_IS_FENCE(client, source)				\
1200 				((client) == KFD_IRQ_FENCE_CLIENTID &&	\
1201 				(source) == KFD_IRQ_FENCE_SOURCEID)
1202 int kfd_interrupt_init(struct kfd_node *dev);
1203 void kfd_interrupt_exit(struct kfd_node *dev);
1204 bool enqueue_ih_ring_entry(struct kfd_node *kfd, const void *ih_ring_entry);
1205 bool interrupt_is_wanted(struct kfd_node *dev,
1206 				const uint32_t *ih_ring_entry,
1207 				uint32_t *patched_ihre, bool *flag);
1208 int kfd_process_drain_interrupts(struct kfd_process_device *pdd);
1209 void kfd_process_close_interrupt_drain(unsigned int pasid);
1210 
1211 /* amdkfd Apertures */
1212 int kfd_init_apertures(struct kfd_process *process);
1213 
1214 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1215 				  uint64_t tba_addr,
1216 				  uint64_t tma_addr);
1217 void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1218 				     bool enabled);
1219 
1220 /* CRIU */
1221 /*
1222  * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private
1223  * structures:
1224  * kfd_criu_process_priv_data
1225  * kfd_criu_device_priv_data
1226  * kfd_criu_bo_priv_data
1227  * kfd_criu_queue_priv_data
1228  * kfd_criu_event_priv_data
1229  * kfd_criu_svm_range_priv_data
1230  */
1231 
1232 #define KFD_CRIU_PRIV_VERSION 1
1233 
1234 struct kfd_criu_process_priv_data {
1235 	uint32_t version;
1236 	uint32_t xnack_mode;
1237 };
1238 
1239 struct kfd_criu_device_priv_data {
1240 	/* For future use */
1241 	uint64_t reserved;
1242 };
1243 
1244 struct kfd_criu_bo_priv_data {
1245 	uint64_t user_addr;
1246 	uint32_t idr_handle;
1247 	uint32_t mapped_gpuids[MAX_GPU_INSTANCE];
1248 };
1249 
1250 /*
1251  * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data,
1252  * kfd_criu_svm_range_priv_data is the object type
1253  */
1254 enum kfd_criu_object_type {
1255 	KFD_CRIU_OBJECT_TYPE_QUEUE,
1256 	KFD_CRIU_OBJECT_TYPE_EVENT,
1257 	KFD_CRIU_OBJECT_TYPE_SVM_RANGE,
1258 };
1259 
1260 struct kfd_criu_svm_range_priv_data {
1261 	uint32_t object_type;
1262 	uint64_t start_addr;
1263 	uint64_t size;
1264 	/* Variable length array of attributes */
1265 	struct kfd_ioctl_svm_attribute attrs[];
1266 };
1267 
1268 struct kfd_criu_queue_priv_data {
1269 	uint32_t object_type;
1270 	uint64_t q_address;
1271 	uint64_t q_size;
1272 	uint64_t read_ptr_addr;
1273 	uint64_t write_ptr_addr;
1274 	uint64_t doorbell_off;
1275 	uint64_t eop_ring_buffer_address;
1276 	uint64_t ctx_save_restore_area_address;
1277 	uint32_t gpu_id;
1278 	uint32_t type;
1279 	uint32_t format;
1280 	uint32_t q_id;
1281 	uint32_t priority;
1282 	uint32_t q_percent;
1283 	uint32_t doorbell_id;
1284 	uint32_t gws;
1285 	uint32_t sdma_id;
1286 	uint32_t eop_ring_buffer_size;
1287 	uint32_t ctx_save_restore_area_size;
1288 	uint32_t ctl_stack_size;
1289 	uint32_t mqd_size;
1290 };
1291 
1292 struct kfd_criu_event_priv_data {
1293 	uint32_t object_type;
1294 	uint64_t user_handle;
1295 	uint32_t event_id;
1296 	uint32_t auto_reset;
1297 	uint32_t type;
1298 	uint32_t signaled;
1299 
1300 	union {
1301 		struct kfd_hsa_memory_exception_data memory_exception_data;
1302 		struct kfd_hsa_hw_exception_data hw_exception_data;
1303 	};
1304 };
1305 
1306 int kfd_process_get_queue_info(struct kfd_process *p,
1307 			       uint32_t *num_queues,
1308 			       uint64_t *priv_data_sizes);
1309 
1310 int kfd_criu_checkpoint_queues(struct kfd_process *p,
1311 			 uint8_t __user *user_priv_data,
1312 			 uint64_t *priv_data_offset);
1313 
1314 int kfd_criu_restore_queue(struct kfd_process *p,
1315 			   uint8_t __user *user_priv_data,
1316 			   uint64_t *priv_data_offset,
1317 			   uint64_t max_priv_data_size);
1318 
1319 int kfd_criu_checkpoint_events(struct kfd_process *p,
1320 			 uint8_t __user *user_priv_data,
1321 			 uint64_t *priv_data_offset);
1322 
1323 int kfd_criu_restore_event(struct file *devkfd,
1324 			   struct kfd_process *p,
1325 			   uint8_t __user *user_priv_data,
1326 			   uint64_t *priv_data_offset,
1327 			   uint64_t max_priv_data_size);
1328 /* CRIU - End */
1329 
1330 /* Queue Context Management */
1331 int init_queue(struct queue **q, const struct queue_properties *properties);
1332 void uninit_queue(struct queue *q);
1333 void print_queue_properties(struct queue_properties *q);
1334 void print_queue(struct queue *q);
1335 int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_bo **pbo,
1336 			 u64 expected_size);
1337 void kfd_queue_buffer_put(struct amdgpu_bo **bo);
1338 int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);
1339 int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);
1340 void kfd_queue_unref_bo_va(struct amdgpu_vm *vm, struct amdgpu_bo **bo);
1341 int kfd_queue_unref_bo_vas(struct kfd_process_device *pdd,
1342 			   struct queue_properties *properties);
1343 void kfd_queue_ctx_save_restore_size(struct kfd_topology_device *dev);
1344 
1345 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
1346 		struct kfd_node *dev);
1347 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
1348 		struct kfd_node *dev);
1349 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
1350 		struct kfd_node *dev);
1351 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
1352 		struct kfd_node *dev);
1353 struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type,
1354 		struct kfd_node *dev);
1355 struct mqd_manager *mqd_manager_init_v12(enum KFD_MQD_TYPE type,
1356 		struct kfd_node *dev);
1357 struct mqd_manager *mqd_manager_init_v12_1(enum KFD_MQD_TYPE type,
1358 		struct kfd_node *dev);
1359 struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev);
1360 void device_queue_manager_uninit(struct device_queue_manager *dqm);
1361 struct kernel_queue *kernel_queue_init(struct kfd_node *dev,
1362 					enum kfd_queue_type type);
1363 void kernel_queue_uninit(struct kernel_queue *kq);
1364 int kfd_evict_process_device(struct kfd_process_device *pdd);
1365 int kfd_dqm_suspend_bad_queue_mes(struct kfd_node *knode, u32 pasid, u32 doorbell_id);
1366 
1367 /* Process Queue Manager */
1368 struct process_queue_node {
1369 	struct queue *q;
1370 	struct kernel_queue *kq;
1371 	struct list_head process_queue_list;
1372 };
1373 
1374 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
1375 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
1376 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
1377 void pqm_uninit(struct process_queue_manager *pqm);
1378 int pqm_create_queue(struct process_queue_manager *pqm,
1379 			    struct kfd_node *dev,
1380 			    struct queue_properties *properties,
1381 			    unsigned int *qid,
1382 			    const struct kfd_criu_queue_priv_data *q_data,
1383 			    const void *restore_mqd,
1384 			    const void *restore_ctl_stack,
1385 			    uint32_t *p_doorbell_offset_in_process);
1386 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
1387 int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid,
1388 			struct queue_properties *p);
1389 int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid,
1390 			struct mqd_update_info *minfo);
1391 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
1392 			void *gws);
1393 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1394 						unsigned int qid);
1395 int pqm_get_wave_state(struct process_queue_manager *pqm,
1396 		       unsigned int qid,
1397 		       void __user *ctl_stack,
1398 		       u32 *ctl_stack_used_size,
1399 		       u32 *save_area_used_size);
1400 int pqm_get_queue_snapshot(struct process_queue_manager *pqm,
1401 			   uint64_t exception_clear_mask,
1402 			   void __user *buf,
1403 			   int *num_qss_entries,
1404 			   uint32_t *entry_size);
1405 
1406 int amdkfd_fence_wait_timeout(struct device_queue_manager *dqm,
1407 			      uint64_t fence_value,
1408 			      unsigned int timeout_ms);
1409 
1410 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
1411 				  unsigned int qid,
1412 				  u32 *mqd_size,
1413 				  u32 *ctl_stack_size);
1414 /* Packet Manager */
1415 
1416 #define KFD_FENCE_COMPLETED (100)
1417 #define KFD_FENCE_INIT   (10)
1418 
1419 /**
1420  * enum kfd_config_dequeue_wait_counts_cmd - Command for configuring
1421  *  dequeue wait counts.
1422  *
1423  * @KFD_DEQUEUE_WAIT_INIT: Set optimized dequeue wait counts for a
1424  *	certain ASICs. For these ASICs, this is default value used by RESET
1425  * @KFD_DEQUEUE_WAIT_RESET: Reset dequeue wait counts to the optimized value
1426  *	for certain ASICs. For others set it to default hardware reset value
1427  * @KFD_DEQUEUE_WAIT_SET_SCH_WAVE: Set context switch latency wait
1428  *
1429  */
1430 enum kfd_config_dequeue_wait_counts_cmd {
1431 	KFD_DEQUEUE_WAIT_INIT = 1,
1432 	KFD_DEQUEUE_WAIT_RESET = 2,
1433 	KFD_DEQUEUE_WAIT_SET_SCH_WAVE = 3
1434 };
1435 
1436 
1437 struct packet_manager {
1438 	struct device_queue_manager *dqm;
1439 	struct kernel_queue *priv_queue;
1440 	struct mutex lock;
1441 	bool allocated;
1442 	struct kfd_mem_obj *ib_buffer_obj;
1443 	unsigned int ib_size_bytes;
1444 	bool is_over_subscription;
1445 
1446 	const struct packet_manager_funcs *pmf;
1447 };
1448 
1449 struct packet_manager_funcs {
1450 	/* Support ASIC-specific packet formats for PM4 packets */
1451 	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
1452 			struct qcm_process_device *qpd);
1453 	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
1454 			uint64_t ib, size_t ib_size_in_dwords, bool chain);
1455 	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
1456 			struct scheduling_resources *res);
1457 	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
1458 			struct queue *q, bool is_static);
1459 	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
1460 			enum kfd_unmap_queues_filter mode,
1461 			uint32_t filter_param, bool reset);
1462 	int (*config_dequeue_wait_counts)(struct packet_manager *pm, uint32_t *buffer,
1463 			enum kfd_config_dequeue_wait_counts_cmd cmd, uint32_t value);
1464 	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
1465 			uint64_t fence_address,	uint64_t fence_value);
1466 	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
1467 
1468 	/* Packet sizes */
1469 	int map_process_size;
1470 	int runlist_size;
1471 	int set_resources_size;
1472 	int map_queues_size;
1473 	int unmap_queues_size;
1474 	int config_dequeue_wait_counts_size;
1475 	int query_status_size;
1476 	int release_mem_size;
1477 };
1478 
1479 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
1480 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
1481 extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;
1482 
1483 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
1484 void pm_uninit(struct packet_manager *pm);
1485 int pm_send_set_resources(struct packet_manager *pm,
1486 				struct scheduling_resources *res);
1487 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
1488 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
1489 				uint64_t fence_value);
1490 
1491 int pm_send_unmap_queue(struct packet_manager *pm,
1492 			enum kfd_unmap_queues_filter mode,
1493 			uint32_t filter_param, bool reset);
1494 
1495 void pm_release_ib(struct packet_manager *pm);
1496 
1497 int pm_config_dequeue_wait_counts(struct packet_manager *pm,
1498 			enum kfd_config_dequeue_wait_counts_cmd cmd,
1499 			uint32_t wait_counts_config);
1500 
1501 /* Following PM funcs can be shared among VI and AI */
1502 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
1503 
1504 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
1505 
1506 /* Events */
1507 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1508 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1509 extern const struct kfd_event_interrupt_class event_interrupt_class_v9_4_3;
1510 extern const struct kfd_event_interrupt_class event_interrupt_class_v10;
1511 extern const struct kfd_event_interrupt_class event_interrupt_class_v11;
1512 extern const struct kfd_event_interrupt_class event_interrupt_class_v12_1;
1513 
1514 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1515 
1516 int kfd_event_init_process(struct kfd_process *p);
1517 void kfd_event_free_process(struct kfd_process *p);
1518 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1519 int kfd_wait_on_events(struct kfd_process *p,
1520 		       uint32_t num_events, void __user *data,
1521 		       bool all, uint32_t *user_timeout_ms,
1522 		       uint32_t *wait_result);
1523 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
1524 				uint32_t valid_id_bits, bool signal_mailbox_updated);
1525 void kfd_signal_hw_exception_event(u32 pasid);
1526 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1527 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1528 int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset);
1529 
1530 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1531 		     uint32_t event_type, bool auto_reset, uint32_t node_id,
1532 		     uint32_t *event_id, uint32_t *event_trigger_data,
1533 		     uint64_t *event_page_offset, uint32_t *event_slot_index);
1534 
1535 int kfd_get_num_events(struct kfd_process *p);
1536 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1537 
1538 void kfd_signal_vm_fault_event_with_userptr(struct kfd_process *p, uint64_t gpu_va);
1539 
1540 void kfd_signal_vm_fault_event(struct kfd_process_device *pdd,
1541 				struct kfd_vm_fault_info *info,
1542 				struct kfd_hsa_memory_exception_data *data);
1543 
1544 void kfd_signal_reset_event(struct kfd_node *dev);
1545 
1546 void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid);
1547 void kfd_signal_process_terminate_event(struct kfd_process *p);
1548 
1549 static inline void kfd_flush_tlb(struct kfd_process_device *pdd)
1550 {
1551 	struct amdgpu_device *adev = pdd->dev->adev;
1552 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1553 
1554 	amdgpu_vm_flush_compute_tlb(adev, vm, TLB_FLUSH_HEAVYWEIGHT,
1555 				    pdd->dev->xcc_mask);
1556 }
1557 
1558 static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev)
1559 {
1560 	return KFD_GC_VERSION(dev) >= IP_VERSION(9, 4, 2) ||
1561 	       (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) ||
1562 	       KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0);
1563 }
1564 
1565 int kfd_send_exception_to_runtime(struct kfd_process *p,
1566 				unsigned int queue_id,
1567 				uint64_t error_reason);
1568 bool kfd_is_locked(struct kfd_dev *kfd);
1569 
1570 /* Compute profile */
1571 void kfd_inc_compute_active(struct kfd_node *dev);
1572 void kfd_dec_compute_active(struct kfd_node *dev);
1573 
1574 /* Cgroup Support */
1575 /* Check with device cgroup if @kfd device is accessible */
1576 static inline int kfd_devcgroup_check_permission(struct kfd_node *node)
1577 {
1578 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1579 	struct drm_device *ddev;
1580 
1581 	if (node->xcp)
1582 		ddev = node->xcp->ddev;
1583 	else
1584 		ddev = adev_to_drm(node->adev);
1585 
1586 	return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1587 					  ddev->render->index,
1588 					  DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1589 #else
1590 	return 0;
1591 #endif
1592 }
1593 
1594 static inline bool kfd_is_first_node(struct kfd_node *node)
1595 {
1596 	return (node == node->kfd->nodes[0]);
1597 }
1598 
1599 /* Debugfs */
1600 #if defined(CONFIG_DEBUG_FS)
1601 
1602 void kfd_debugfs_init(void);
1603 void kfd_debugfs_fini(void);
1604 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1605 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1606 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1607 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1608 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1609 int pm_debugfs_runlist(struct seq_file *m, void *data);
1610 
1611 int kfd_debugfs_hang_hws(struct kfd_node *dev);
1612 int pm_debugfs_hang_hws(struct packet_manager *pm);
1613 int dqm_debugfs_hang_hws(struct device_queue_manager *dqm);
1614 
1615 void kfd_debugfs_add_process(struct kfd_process *p);
1616 void kfd_debugfs_remove_process(struct kfd_process *p);
1617 
1618 #else
1619 
1620 static inline void kfd_debugfs_init(void) {}
1621 static inline void kfd_debugfs_fini(void) {}
1622 static inline void kfd_debugfs_add_process(struct kfd_process *p) {}
1623 static inline void kfd_debugfs_remove_process(struct kfd_process *p) {}
1624 
1625 #endif
1626 
1627 #endif
1628