xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_priv.h (revision 70cadefcc6160c575b04f763ada34c20e868d577)
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 	/* Lock for profiler process */
388 	struct mutex profiler_lock;
389 	/* Process currently holding the lock */
390 	struct kfd_process *profiler_process;
391 };
392 
393 enum kfd_mempool {
394 	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
395 	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
396 	KFD_MEMPOOL_FRAMEBUFFER = 3,
397 };
398 
399 /* Character device interface */
400 int kfd_chardev_init(void);
401 void kfd_chardev_exit(void);
402 void kfd_dev_unmap_mapping_range(loff_t const holebegin, loff_t const holelen);
403 
404 /**
405  * enum kfd_unmap_queues_filter - Enum for queue filters.
406  *
407  * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
408  *						running queues list.
409  *
410  * @KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: Preempts all non-static queues
411  *						in the run list.
412  *
413  * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
414  *						specific process.
415  *
416  */
417 enum kfd_unmap_queues_filter {
418 	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES = 1,
419 	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES = 2,
420 	KFD_UNMAP_QUEUES_FILTER_BY_PASID = 3
421 };
422 
423 /**
424  * enum kfd_queue_type - Enum for various queue types.
425  *
426  * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
427  *
428  * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
429  *
430  * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
431  *
432  * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
433  *
434  * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
435  *
436  * @KFD_QUEUE_TYPE_SDMA_BY_ENG_ID:  SDMA user mode queue with target SDMA engine ID.
437  */
438 enum kfd_queue_type  {
439 	KFD_QUEUE_TYPE_COMPUTE,
440 	KFD_QUEUE_TYPE_SDMA,
441 	KFD_QUEUE_TYPE_HIQ,
442 	KFD_QUEUE_TYPE_SDMA_XGMI,
443 	KFD_QUEUE_TYPE_SDMA_BY_ENG_ID
444 };
445 
446 enum kfd_queue_format {
447 	KFD_QUEUE_FORMAT_PM4,
448 	KFD_QUEUE_FORMAT_AQL
449 };
450 
451 enum KFD_QUEUE_PRIORITY {
452 	KFD_QUEUE_PRIORITY_MINIMUM = 0,
453 	KFD_QUEUE_PRIORITY_MAXIMUM = 15
454 };
455 
456 /**
457  * struct queue_properties
458  *
459  * @type: The queue type.
460  *
461  * @queue_id: Queue identifier.
462  *
463  * @queue_address: Queue ring buffer address.
464  *
465  * @queue_size: Queue ring buffer size.
466  *
467  * @priority: Defines the queue priority relative to other queues in the
468  * process.
469  * This is just an indication and HW scheduling may override the priority as
470  * necessary while keeping the relative prioritization.
471  * the priority granularity is from 0 to f which f is the highest priority.
472  * currently all queues are initialized with the highest priority.
473  *
474  * @queue_percent: This field is partially implemented and currently a zero in
475  * this field defines that the queue is non active.
476  *
477  * @read_ptr: User space address which points to the number of dwords the
478  * cp read from the ring buffer. This field updates automatically by the H/W.
479  *
480  * @write_ptr: Defines the number of dwords written to the ring buffer.
481  *
482  * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
483  * buffer. This field should be similar to write_ptr and the user should
484  * update this field after updating the write_ptr.
485  *
486  * @doorbell_off: The doorbell offset in the doorbell pci-bar.
487  *
488  * @is_interop: Defines if this is a interop queue. Interop queue means that
489  * the queue can access both graphics and compute resources.
490  *
491  * @is_evicted: Defines if the queue is evicted. Only active queues
492  * are evicted, rendering them inactive.
493  *
494  * @is_active: Defines if the queue is active or not. @is_active and
495  * @is_evicted are protected by the DQM lock.
496  *
497  * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
498  * @is_gws should be protected by the DQM lock, since changing it can yield the
499  * possibility of updating DQM state on number of GWS queues.
500  *
501  * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
502  * of the queue.
503  *
504  * This structure represents the queue properties for each queue no matter if
505  * it's user mode or kernel mode queue.
506  *
507  */
508 
509 struct queue_properties {
510 	enum kfd_queue_type type;
511 	enum kfd_queue_format format;
512 	unsigned int queue_id;
513 	uint64_t queue_address;
514 	uint64_t queue_size;
515 	uint64_t metadata_queue_size;
516 	uint32_t priority;
517 	uint32_t queue_percent;
518 	void __user *read_ptr;
519 	void __user *write_ptr;
520 	void __iomem *doorbell_ptr;
521 	uint32_t doorbell_off;
522 	bool is_interop;
523 	bool is_evicted;
524 	bool is_suspended;
525 	bool is_being_destroyed;
526 	bool is_active;
527 	bool is_gws;
528 	uint32_t pm4_target_xcc;
529 	bool is_dbg_wa;
530 	bool is_user_cu_masked;
531 	bool is_reset;
532 	/* Not relevant for user mode queues in cp scheduling */
533 	unsigned int vmid;
534 	/* Relevant only for sdma queues*/
535 	uint32_t sdma_engine_id;
536 	uint32_t sdma_queue_id;
537 	uint32_t sdma_vm_addr;
538 	/* Relevant only for VI */
539 	uint64_t eop_ring_buffer_address;
540 	uint32_t eop_ring_buffer_size;
541 	uint64_t ctx_save_restore_area_address;
542 	uint32_t ctx_save_restore_area_size;
543 	uint32_t ctl_stack_size;
544 	uint64_t tba_addr;
545 	uint64_t tma_addr;
546 	uint64_t exception_status;
547 
548 	struct amdgpu_bo *wptr_bo;
549 	struct amdgpu_bo *rptr_bo;
550 	struct amdgpu_bo *ring_bo;
551 	struct amdgpu_bo *eop_buf_bo;
552 	struct amdgpu_bo *cwsr_bo;
553 };
554 
555 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\
556 			    (q).queue_address != 0 &&	\
557 			    (q).queue_percent > 0 &&	\
558 			    !(q).is_evicted &&		\
559 			    !(q).is_suspended)
560 
561 enum mqd_update_flag {
562 	UPDATE_FLAG_DBG_WA_ENABLE = 1,
563 	UPDATE_FLAG_DBG_WA_DISABLE = 2,
564 	UPDATE_FLAG_IS_GWS = 4, /* quirk for gfx9 IP */
565 	UPDATE_FLAG_PERFCOUNT_ENABLE = 5,
566 	UPDATE_FLAG_PERFCOUNT_DISABLE = 6,
567 };
568 
569 struct mqd_update_info {
570 	union {
571 		struct {
572 			uint32_t count; /* Must be a multiple of 32 */
573 			uint32_t *ptr;
574 		} cu_mask;
575 	};
576 	enum mqd_update_flag update_flag;
577 };
578 
579 /**
580  * struct queue
581  *
582  * @list: Queue linked list.
583  *
584  * @mqd: The queue MQD (memory queue descriptor).
585  *
586  * @mqd_mem_obj: The MQD local gpu memory object.
587  *
588  * @gart_mqd_addr: The MQD gart mc address.
589  *
590  * @properties: The queue properties.
591  *
592  * @mec: Used only in no cp scheduling mode and identifies to micro engine id
593  *	 that the queue should be executed on.
594  *
595  * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
596  *	  id.
597  *
598  * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
599  *
600  * @process: The kfd process that created this queue.
601  *
602  * @device: The kfd device that created this queue.
603  *
604  * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
605  * otherwise.
606  *
607  * This structure represents user mode compute queues.
608  * It contains all the necessary data to handle such queues.
609  *
610  */
611 
612 struct queue {
613 	struct list_head list;
614 	void *mqd;
615 	struct kfd_mem_obj *mqd_mem_obj;
616 	uint64_t gart_mqd_addr;
617 	struct queue_properties properties;
618 
619 	uint32_t mec;
620 	uint32_t pipe;
621 	uint32_t queue;
622 
623 	unsigned int sdma_id;
624 	unsigned int doorbell_id;
625 
626 	struct kfd_process	*process;
627 	struct kfd_node		*device;
628 	void *gws;
629 
630 	/* procfs */
631 	struct kobject kobj;
632 
633 	void *gang_ctx_bo;
634 	uint64_t gang_ctx_gpu_addr;
635 	void *gang_ctx_cpu_ptr;
636 
637 	struct amdgpu_bo *wptr_bo_gart;
638 };
639 
640 enum KFD_MQD_TYPE {
641 	KFD_MQD_TYPE_HIQ = 0,		/* for hiq */
642 	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
643 	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
644 	KFD_MQD_TYPE_DIQ,		/* for diq */
645 	KFD_MQD_TYPE_MAX
646 };
647 
648 enum KFD_PIPE_PRIORITY {
649 	KFD_PIPE_PRIORITY_CS_LOW = 0,
650 	KFD_PIPE_PRIORITY_CS_MEDIUM,
651 	KFD_PIPE_PRIORITY_CS_HIGH
652 };
653 
654 struct scheduling_resources {
655 	unsigned int vmid_mask;
656 	enum kfd_queue_type type;
657 	uint64_t queue_mask;
658 	uint64_t gws_mask;
659 	uint32_t oac_mask;
660 	uint32_t gds_heap_base;
661 	uint32_t gds_heap_size;
662 };
663 
664 struct process_queue_manager {
665 	/* data */
666 	struct kfd_process	*process;
667 	struct list_head	queues;
668 	unsigned long		*queue_slot_bitmap;
669 };
670 
671 struct qcm_process_device {
672 	/* The Device Queue Manager that owns this data */
673 	struct device_queue_manager *dqm;
674 	struct process_queue_manager *pqm;
675 	/* Queues list */
676 	struct list_head queues_list;
677 	struct list_head priv_queue_list;
678 
679 	unsigned int queue_count;
680 	unsigned int vmid;
681 	bool is_debug;
682 	unsigned int evicted; /* eviction counter, 0=active */
683 
684 	/* This flag tells if we should reset all wavefronts on
685 	 * process termination
686 	 */
687 	bool reset_wavefronts;
688 
689 	/* This flag tells us if this process has a GWS-capable
690 	 * queue that will be mapped into the runlist. It's
691 	 * possible to request a GWS BO, but not have the queue
692 	 * currently mapped, and this changes how the MAP_PROCESS
693 	 * PM4 packet is configured.
694 	 */
695 	bool mapped_gws_queue;
696 
697 	/* All the memory management data should be here too */
698 	uint64_t gds_context_area;
699 	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
700 	uint64_t page_table_base;
701 	uint32_t sh_mem_config;
702 	uint32_t sh_mem_bases;
703 	uint32_t sh_mem_ape1_base;
704 	uint32_t sh_mem_ape1_limit;
705 	uint32_t gds_size;
706 	uint32_t num_gws;
707 	uint32_t num_oac;
708 	uint32_t sh_hidden_private_base;
709 	uint32_t vm_cntx_cntl;
710 
711 	/* CWSR memory */
712 	struct kgd_mem *cwsr_mem;
713 	void *cwsr_kaddr;
714 	uint64_t cwsr_base;
715 	uint64_t tba_addr;
716 	uint64_t tma_addr;
717 
718 	/* IB memory */
719 	struct kgd_mem *ib_mem;
720 	uint64_t ib_base;
721 	void *ib_kaddr;
722 
723 	/* doorbells for kfd process */
724 	struct amdgpu_bo *proc_doorbells;
725 
726 	/* bitmap for dynamic doorbell allocation from the bo */
727 	unsigned long *doorbell_bitmap;
728 };
729 
730 /* KFD Memory Eviction */
731 
732 /* Approx. wait time before attempting to restore evicted BOs */
733 #define PROCESS_RESTORE_TIME_MS 100
734 /* Approx. back off time if restore fails due to lack of memory */
735 #define PROCESS_BACK_OFF_TIME_MS 100
736 /* Approx. time before evicting the process again */
737 #define PROCESS_ACTIVE_TIME_MS 10
738 
739 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
740  * idr_handle in the least significant 4 bytes
741  */
742 #define MAKE_HANDLE(gpu_id, idr_handle) \
743 	(((uint64_t)(gpu_id) << 32) + idr_handle)
744 #define GET_GPU_ID(handle) (handle >> 32)
745 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
746 
747 enum kfd_pdd_bound {
748 	PDD_UNBOUND = 0,
749 	PDD_BOUND,
750 	PDD_BOUND_SUSPENDED,
751 };
752 
753 #define MAX_SYSFS_FILENAME_LEN 15
754 
755 /*
756  * SDMA counter runs at 100MHz frequency.
757  * We display SDMA activity in microsecond granularity in sysfs.
758  * As a result, the divisor is 100.
759  */
760 #define SDMA_ACTIVITY_DIVISOR  100
761 
762 /* Data that is per-process-per device. */
763 struct kfd_process_device {
764 	/* The device that owns this data. */
765 	struct kfd_node *dev;
766 
767 	/* The process that owns this kfd_process_device. */
768 	struct kfd_process *process;
769 
770 	/* per-process-per device QCM data structure */
771 	struct qcm_process_device qpd;
772 
773 	/*Apertures*/
774 	uint64_t lds_base;
775 	uint64_t lds_limit;
776 	uint64_t gpuvm_base;
777 	uint64_t gpuvm_limit;
778 	uint64_t scratch_base;
779 	uint64_t scratch_limit;
780 
781 	/* VM context for GPUVM allocations */
782 	struct file *drm_file;
783 	void *drm_priv;
784 
785 	/* GPUVM allocations storage */
786 	struct idr alloc_idr;
787 
788 	/* Flag used to tell the pdd has dequeued from the dqm.
789 	 * This is used to prevent dev->dqm->ops.process_termination() from
790 	 * being called twice when it is already called in IOMMU callback
791 	 * function.
792 	 */
793 	bool already_dequeued;
794 	bool runtime_inuse;
795 
796 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
797 	enum kfd_pdd_bound bound;
798 
799 	/* VRAM usage */
800 	atomic64_t vram_usage;
801 	struct attribute attr_vram;
802 	char vram_filename[MAX_SYSFS_FILENAME_LEN];
803 
804 	/* SDMA activity tracking */
805 	uint64_t sdma_past_activity_counter;
806 	struct attribute attr_sdma;
807 	char sdma_filename[MAX_SYSFS_FILENAME_LEN];
808 
809 	/* Eviction activity tracking */
810 	uint64_t last_evict_timestamp;
811 	atomic64_t evict_duration_counter;
812 	struct attribute attr_evict;
813 
814 	struct kobject *kobj_stats;
815 
816 	/*
817 	 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
818 	 * that is associated with device encoded by "this" struct instance. The
819 	 * value reflects CU usage by all of the waves launched by this process
820 	 * on this device. A very important property of occupancy parameter is
821 	 * that its value is a snapshot of current use.
822 	 *
823 	 * Following is to be noted regarding how this parameter is reported:
824 	 *
825 	 *  The number of waves that a CU can launch is limited by couple of
826 	 *  parameters. These are encoded by struct amdgpu_cu_info instance
827 	 *  that is part of every device definition. For GFX9 devices this
828 	 *  translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
829 	 *  do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
830 	 *  when they do use scratch memory. This could change for future
831 	 *  devices and therefore this example should be considered as a guide.
832 	 *
833 	 *  All CU's of a device are available for the process. This may not be true
834 	 *  under certain conditions - e.g. CU masking.
835 	 *
836 	 *  Finally number of CU's that are occupied by a process is affected by both
837 	 *  number of CU's a device has along with number of other competing processes
838 	 */
839 	struct attribute attr_cu_occupancy;
840 
841 	/* sysfs counters for GPU retry fault and page migration tracking */
842 	struct kobject *kobj_counters;
843 	struct attribute attr_faults;
844 	struct attribute attr_page_in;
845 	struct attribute attr_page_out;
846 	uint64_t faults;
847 	uint64_t page_in;
848 	uint64_t page_out;
849 
850 	/* Exception code status*/
851 	uint64_t exception_status;
852 	void *vm_fault_exc_data;
853 	size_t vm_fault_exc_data_size;
854 
855 	/* Tracks debug per-vmid request settings */
856 	uint32_t spi_dbg_override;
857 	uint32_t spi_dbg_launch_mode;
858 	uint32_t watch_points[4];
859 	uint32_t alloc_watch_ids;
860 
861 	/*
862 	 * If this process has been checkpointed before, then the user
863 	 * application will use the original gpu_id on the
864 	 * checkpointed node to refer to this device.
865 	 */
866 	uint32_t user_gpu_id;
867 
868 	void *proc_ctx_bo;
869 	uint64_t proc_ctx_gpu_addr;
870 	void *proc_ctx_cpu_ptr;
871 
872 	/* Tracks queue reset status */
873 	bool has_reset_queue;
874 
875 	u32 pasid;
876 	/* Indicates this process has requested PTL stay disabled */
877 	bool ptl_disable_req;
878 };
879 
880 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
881 
882 struct svm_range_list {
883 	struct mutex			lock;
884 	struct rb_root_cached		objects;
885 	struct list_head		list;
886 	struct work_struct		deferred_list_work;
887 	struct list_head		deferred_range_list;
888 	struct list_head                criu_svm_metadata_list;
889 	spinlock_t			deferred_list_lock;
890 	atomic_t			evicted_ranges;
891 	atomic_t			drain_pagefaults;
892 	struct delayed_work		restore_work;
893 	DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);
894 	struct task_struct		*faulting_task;
895 	/* check point ts decides if page fault recovery need be dropped */
896 	uint64_t			checkpoint_ts[MAX_GPU_INSTANCE];
897 
898 	/* Default granularity to use in buffer migration
899 	 * and restoration of backing memory while handling
900 	 * recoverable page faults
901 	 */
902 	uint8_t default_granularity;
903 };
904 
905 /* Process data */
906 struct kfd_process {
907 	/*
908 	 * kfd_process are stored in an mm_struct*->kfd_process*
909 	 * hash table (kfd_processes in kfd_process.c)
910 	 */
911 	struct hlist_node kfd_processes;
912 
913 	/*
914 	 * Opaque pointer to mm_struct. We don't hold a reference to
915 	 * it so it should never be dereferenced from here. This is
916 	 * only used for looking up processes by their mm.
917 	 */
918 	void *mm;
919 
920 	struct kref ref;
921 	struct work_struct release_work;
922 
923 	struct mutex mutex;
924 
925 	/*
926 	 * In any process, the thread that started main() is the lead
927 	 * thread and outlives the rest.
928 	 * It is here because amd_iommu_bind_pasid wants a task_struct.
929 	 * It can also be used for safely getting a reference to the
930 	 * mm_struct of the process.
931 	 */
932 	struct task_struct *lead_thread;
933 
934 	/* We want to receive a notification when the mm_struct is destroyed */
935 	struct mmu_notifier mmu_notifier;
936 
937 	/*
938 	 * Array of kfd_process_device pointers,
939 	 * one for each device the process is using.
940 	 */
941 	struct kfd_process_device *pdds[MAX_GPU_INSTANCE];
942 	uint32_t n_pdds;
943 
944 	struct process_queue_manager pqm;
945 
946 	/*Is the user space process 32 bit?*/
947 	bool is_32bit_user_mode;
948 
949 	/* Event-related data */
950 	struct mutex event_mutex;
951 	/* Event ID allocator and lookup */
952 	struct idr event_idr;
953 	/* Event page */
954 	u64 signal_handle;
955 	struct kfd_signal_page *signal_page;
956 	size_t signal_mapped_size;
957 	size_t signal_event_count;
958 	bool signal_event_limit_reached;
959 
960 	/* Information used for memory eviction */
961 	void *kgd_process_info;
962 	/* Eviction fence that is attached to all the BOs of this process. The
963 	 * fence will be triggered during eviction and new one will be created
964 	 * during restore
965 	 */
966 	struct dma_fence __rcu *ef;
967 
968 	/* Work items for evicting and restoring BOs */
969 	struct delayed_work eviction_work;
970 	struct delayed_work restore_work;
971 	/* seqno of the last scheduled eviction */
972 	unsigned int last_eviction_seqno;
973 	/* Approx. the last timestamp (in jiffies) when the process was
974 	 * restored after an eviction
975 	 */
976 	unsigned long last_restore_timestamp;
977 
978 	/* Indicates device process is debug attached with reserved vmid. */
979 	bool debug_trap_enabled;
980 
981 	/* per-process-per device debug event fd file */
982 	struct file *dbg_ev_file;
983 
984 	/* If the process is a kfd debugger, we need to know so we can clean
985 	 * up at exit time.  If a process enables debugging on itself, it does
986 	 * its own clean-up, so we don't set the flag here.  We track this by
987 	 * counting the number of processes this process is debugging.
988 	 */
989 	atomic_t debugged_process_count;
990 
991 	/* If the process is a debugged, this is the debugger process */
992 	struct kfd_process *debugger_process;
993 
994 	/* Kobj for our procfs */
995 	struct kobject *kobj;
996 	struct kobject *kobj_queues;
997 	struct attribute attr_pasid;
998 
999 	/* Exception code enable mask and status */
1000 	uint64_t exception_enable_mask;
1001 	uint64_t exception_status;
1002 
1003 	/* Used to drain stale interrupts */
1004 	wait_queue_head_t wait_irq_drain;
1005 	bool irq_drain_is_open;
1006 
1007 	/* shared virtual memory registered by this process */
1008 	struct svm_range_list svms;
1009 
1010 	bool xnack_enabled;
1011 
1012 	/* Work area for debugger event writer worker. */
1013 	struct work_struct debug_event_workarea;
1014 
1015 	/* Tracks debug per-vmid request for debug flags */
1016 	u32 dbg_flags;
1017 
1018 	atomic_t poison;
1019 	/* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */
1020 	bool queues_paused;
1021 
1022 	/* Tracks runtime enable status */
1023 	struct semaphore runtime_enable_sema;
1024 	bool is_runtime_retry;
1025 	struct kfd_runtime_info runtime_info;
1026 
1027 	/* if gpu page fault sent to KFD */
1028 	bool gpu_page_fault;
1029 
1030 	/*kfd context id */
1031 	u16 context_id;
1032 
1033 	/* The primary kfd_process allocating IDs for its secondary kfd_process, 0 for primary kfd_process */
1034 	struct ida id_table;
1035 
1036 };
1037 
1038 #define KFD_PROCESS_TABLE_SIZE 8 /* bits: 256 entries */
1039 #define KFD_CONTEXT_ID_PRIMARY	0xFFFF
1040 #define KFD_CONTEXT_ID_MIN 0
1041 
1042 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
1043 extern struct srcu_struct kfd_processes_srcu;
1044 
1045 /**
1046  * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
1047  *
1048  * @filep: pointer to file structure.
1049  * @p: amdkfd process pointer.
1050  * @data: pointer to arg that was copied from user.
1051  *
1052  * Return: returns ioctl completion code.
1053  */
1054 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
1055 				void *data);
1056 
1057 typedef int amdkfd_ioctl_validate_t(void *kdata, unsigned int usize);
1058 
1059 struct amdkfd_ioctl_desc {
1060 	unsigned int cmd;
1061 	int flags;
1062 	amdkfd_ioctl_t *func;
1063 	amdkfd_ioctl_validate_t *validate;
1064 	unsigned int cmd_drv;
1065 	const char *name;
1066 };
1067 bool kfd_dev_is_large_bar(struct kfd_node *dev);
1068 
1069 struct kfd_process *create_process(const struct task_struct *thread, bool primary);
1070 int kfd_process_create_wq(void);
1071 void kfd_process_destroy_wq(void);
1072 void kfd_cleanup_processes(void);
1073 struct kfd_process *kfd_create_process(struct task_struct *thread);
1074 int kfd_create_process_sysfs(struct kfd_process *process);
1075 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid,
1076 						 struct kfd_process_device **pdd);
1077 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
1078 struct kfd_process *kfd_lookup_process_by_id(const struct mm_struct *mm, u16 id);
1079 
1080 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
1081 int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
1082 				uint32_t *gpuid, uint32_t *gpuidx);
1083 static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
1084 				uint32_t gpuidx, uint32_t *gpuid) {
1085 	return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
1086 }
1087 static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
1088 				struct kfd_process *p, uint32_t gpuidx) {
1089 	return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;
1090 }
1091 
1092 void kfd_unref_process(struct kfd_process *p);
1093 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger);
1094 int kfd_process_restore_queues(struct kfd_process *p);
1095 void kfd_suspend_all_processes(void);
1096 int kfd_resume_all_processes(void);
1097 
1098 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *process,
1099 							 uint32_t gpu_id);
1100 
1101 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id);
1102 
1103 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1104 			       struct file *drm_file);
1105 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1106 						struct kfd_process *p);
1107 struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1108 							struct kfd_process *p);
1109 struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1110 							struct kfd_process *p);
1111 
1112 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
1113 
1114 void kfd_process_notifier_release_internal(struct kfd_process *p);
1115 
1116 /* KFD process API for creating and translating handles */
1117 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1118 					void *mem);
1119 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
1120 					int handle);
1121 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1122 					int handle);
1123 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid);
1124 
1125 /* PASIDs */
1126 int kfd_pasid_init(void);
1127 void kfd_pasid_exit(void);
1128 u32 kfd_pasid_alloc(void);
1129 void kfd_pasid_free(u32 pasid);
1130 
1131 /* Doorbells */
1132 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
1133 int kfd_doorbell_init(struct kfd_dev *kfd);
1134 void kfd_doorbell_fini(struct kfd_dev *kfd);
1135 int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process,
1136 		      struct vm_area_struct *vma);
1137 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
1138 					unsigned int *doorbell_off);
1139 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
1140 u32 read_kernel_doorbell(u32 __iomem *db);
1141 void write_kernel_doorbell(void __iomem *db, u32 value);
1142 void write_kernel_doorbell64(void __iomem *db, u64 value);
1143 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
1144 					struct kfd_process_device *pdd,
1145 					unsigned int doorbell_id);
1146 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
1147 int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
1148 				struct kfd_process_device *pdd);
1149 void kfd_free_process_doorbells(struct kfd_dev *kfd,
1150 				struct kfd_process_device *pdd);
1151 /* GTT Sub-Allocator */
1152 
1153 int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size,
1154 			struct kfd_mem_obj **mem_obj);
1155 
1156 int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj);
1157 
1158 extern struct device *kfd_device;
1159 
1160 /* KFD's procfs */
1161 void kfd_procfs_init(void);
1162 void kfd_procfs_shutdown(void);
1163 int kfd_procfs_add_queue(struct queue *q);
1164 void kfd_procfs_del_queue(struct queue *q);
1165 
1166 /* Topology */
1167 int kfd_topology_init(void);
1168 void kfd_topology_shutdown(void);
1169 int kfd_topology_add_device(struct kfd_node *gpu);
1170 int kfd_topology_remove_device(struct kfd_node *gpu);
1171 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
1172 						uint32_t proximity_domain);
1173 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
1174 						uint32_t proximity_domain);
1175 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
1176 struct kfd_node *kfd_device_by_id(uint32_t gpu_id);
1177 static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t node_id,
1178 					uint32_t vmid)
1179 {
1180 	return (node->interrupt_bitmap & (1 << node_id)) != 0 &&
1181 	       (node->compute_vmid_bitmap & (1 << vmid)) != 0;
1182 }
1183 static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev,
1184 					uint32_t node_id, uint32_t vmid) {
1185 	struct kfd_dev *dev = adev->kfd.dev;
1186 	uint32_t i;
1187 
1188 	/*
1189 	 * On multi-aid system, attempt per-node matching. Otherwise,
1190 	 * fall back to the first node.
1191 	 */
1192 	if (!amdgpu_is_multi_aid(adev))
1193 		return dev->nodes[0];
1194 
1195 	for (i = 0; i < dev->num_nodes; i++)
1196 		if (kfd_irq_is_from_node(dev->nodes[i], node_id, vmid))
1197 			return dev->nodes[i];
1198 
1199 	return NULL;
1200 }
1201 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev);
1202 uint32_t kfd_topology_get_num_devices(void);
1203 int kfd_numa_node_to_apic_id(int numa_node_id);
1204 uint32_t kfd_gpu_node_num(void);
1205 
1206 /* Interrupts */
1207 #define	KFD_IRQ_FENCE_CLIENTID	0xff
1208 #define	KFD_IRQ_FENCE_SOURCEID	0xff
1209 #define	KFD_IRQ_IS_FENCE(client, source)				\
1210 				((client) == KFD_IRQ_FENCE_CLIENTID &&	\
1211 				(source) == KFD_IRQ_FENCE_SOURCEID)
1212 int kfd_interrupt_init(struct kfd_node *dev);
1213 void kfd_interrupt_exit(struct kfd_node *dev);
1214 bool enqueue_ih_ring_entry(struct kfd_node *kfd, const void *ih_ring_entry);
1215 bool interrupt_is_wanted(struct kfd_node *dev,
1216 				const uint32_t *ih_ring_entry,
1217 				uint32_t *patched_ihre, bool *flag);
1218 int kfd_process_drain_interrupts(struct kfd_process_device *pdd);
1219 void kfd_process_close_interrupt_drain(unsigned int pasid);
1220 
1221 /* amdkfd Apertures */
1222 int kfd_init_apertures(struct kfd_process *process);
1223 
1224 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1225 				  uint64_t tba_addr,
1226 				  uint64_t tma_addr);
1227 void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1228 				     bool enabled);
1229 
1230 /* CRIU */
1231 /*
1232  * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private
1233  * structures:
1234  * kfd_criu_process_priv_data
1235  * kfd_criu_device_priv_data
1236  * kfd_criu_bo_priv_data
1237  * kfd_criu_queue_priv_data
1238  * kfd_criu_event_priv_data
1239  * kfd_criu_svm_range_priv_data
1240  */
1241 
1242 #define KFD_CRIU_PRIV_VERSION 1
1243 
1244 struct kfd_criu_process_priv_data {
1245 	uint32_t version;
1246 	uint32_t xnack_mode;
1247 };
1248 
1249 struct kfd_criu_device_priv_data {
1250 	/* For future use */
1251 	uint64_t reserved;
1252 };
1253 
1254 struct kfd_criu_bo_priv_data {
1255 	uint64_t user_addr;
1256 	uint32_t idr_handle;
1257 	uint32_t mapped_gpuids[MAX_GPU_INSTANCE];
1258 };
1259 
1260 /*
1261  * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data,
1262  * kfd_criu_svm_range_priv_data is the object type
1263  */
1264 enum kfd_criu_object_type {
1265 	KFD_CRIU_OBJECT_TYPE_QUEUE,
1266 	KFD_CRIU_OBJECT_TYPE_EVENT,
1267 	KFD_CRIU_OBJECT_TYPE_SVM_RANGE,
1268 };
1269 
1270 struct kfd_criu_svm_range_priv_data {
1271 	uint32_t object_type;
1272 	uint64_t start_addr;
1273 	uint64_t size;
1274 	/* Variable length array of attributes */
1275 	struct kfd_ioctl_svm_attribute attrs[];
1276 };
1277 
1278 struct kfd_criu_queue_priv_data {
1279 	uint32_t object_type;
1280 	uint64_t q_address;
1281 	uint64_t q_size;
1282 	uint64_t read_ptr_addr;
1283 	uint64_t write_ptr_addr;
1284 	uint64_t doorbell_off;
1285 	uint64_t eop_ring_buffer_address;
1286 	uint64_t ctx_save_restore_area_address;
1287 	uint32_t gpu_id;
1288 	uint32_t type;
1289 	uint32_t format;
1290 	uint32_t q_id;
1291 	uint32_t priority;
1292 	uint32_t q_percent;
1293 	uint32_t doorbell_id;
1294 	uint32_t gws;
1295 	uint32_t sdma_id;
1296 	uint32_t eop_ring_buffer_size;
1297 	uint32_t ctx_save_restore_area_size;
1298 	uint32_t ctl_stack_size;
1299 	uint32_t mqd_size;
1300 };
1301 
1302 struct kfd_criu_event_priv_data {
1303 	uint32_t object_type;
1304 	uint64_t user_handle;
1305 	uint32_t event_id;
1306 	uint32_t auto_reset;
1307 	uint32_t type;
1308 	uint32_t signaled;
1309 
1310 	union {
1311 		struct kfd_hsa_memory_exception_data memory_exception_data;
1312 		struct kfd_hsa_hw_exception_data hw_exception_data;
1313 	};
1314 };
1315 
1316 int kfd_process_get_queue_info(struct kfd_process *p,
1317 			       uint32_t *num_queues,
1318 			       uint64_t *priv_data_sizes);
1319 
1320 int kfd_criu_checkpoint_queues(struct kfd_process *p,
1321 			 uint8_t __user *user_priv_data,
1322 			 uint64_t *priv_data_offset);
1323 
1324 int kfd_criu_restore_queue(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 
1329 int kfd_criu_checkpoint_events(struct kfd_process *p,
1330 			 uint8_t __user *user_priv_data,
1331 			 uint64_t *priv_data_offset);
1332 
1333 int kfd_criu_restore_event(struct file *devkfd,
1334 			   struct kfd_process *p,
1335 			   uint8_t __user *user_priv_data,
1336 			   uint64_t *priv_data_offset,
1337 			   uint64_t max_priv_data_size);
1338 /* CRIU - End */
1339 
1340 /* Queue Context Management */
1341 int init_queue(struct queue **q, const struct queue_properties *properties);
1342 void uninit_queue(struct queue *q);
1343 void print_queue_properties(struct queue_properties *q);
1344 void print_queue(struct queue *q);
1345 int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_bo **pbo,
1346 			 u64 expected_size);
1347 void kfd_queue_buffer_put(struct amdgpu_bo **bo);
1348 int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);
1349 int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);
1350 void kfd_queue_unref_bo_va(struct amdgpu_vm *vm, struct amdgpu_bo **bo);
1351 int kfd_queue_unref_bo_vas(struct kfd_process_device *pdd,
1352 			   struct queue_properties *properties);
1353 void kfd_queue_ctx_save_restore_size(struct kfd_topology_device *dev);
1354 
1355 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
1356 		struct kfd_node *dev);
1357 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
1358 		struct kfd_node *dev);
1359 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
1360 		struct kfd_node *dev);
1361 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
1362 		struct kfd_node *dev);
1363 struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type,
1364 		struct kfd_node *dev);
1365 struct mqd_manager *mqd_manager_init_v12(enum KFD_MQD_TYPE type,
1366 		struct kfd_node *dev);
1367 struct mqd_manager *mqd_manager_init_v12_1(enum KFD_MQD_TYPE type,
1368 		struct kfd_node *dev);
1369 struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev);
1370 void device_queue_manager_uninit(struct device_queue_manager *dqm);
1371 struct kernel_queue *kernel_queue_init(struct kfd_node *dev,
1372 					enum kfd_queue_type type);
1373 void kernel_queue_uninit(struct kernel_queue *kq);
1374 int kfd_evict_process_device(struct kfd_process_device *pdd);
1375 int kfd_dqm_suspend_bad_queue_mes(struct kfd_node *knode, u32 pasid, u32 doorbell_id);
1376 
1377 /* Process Queue Manager */
1378 struct process_queue_node {
1379 	struct queue *q;
1380 	struct kernel_queue *kq;
1381 	struct list_head process_queue_list;
1382 };
1383 
1384 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
1385 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
1386 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
1387 void pqm_uninit(struct process_queue_manager *pqm);
1388 int pqm_create_queue(struct process_queue_manager *pqm,
1389 			    struct kfd_node *dev,
1390 			    struct queue_properties *properties,
1391 			    unsigned int *qid,
1392 			    const struct kfd_criu_queue_priv_data *q_data,
1393 			    const void *restore_mqd,
1394 			    const void *restore_ctl_stack,
1395 			    uint32_t *p_doorbell_offset_in_process);
1396 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
1397 int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid,
1398 			struct queue_properties *p);
1399 int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid,
1400 			struct mqd_update_info *minfo);
1401 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
1402 			void *gws);
1403 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1404 						unsigned int qid);
1405 int pqm_get_wave_state(struct process_queue_manager *pqm,
1406 		       unsigned int qid,
1407 		       void __user *ctl_stack,
1408 		       u32 *ctl_stack_used_size,
1409 		       u32 *save_area_used_size);
1410 int pqm_get_queue_snapshot(struct process_queue_manager *pqm,
1411 			   uint64_t exception_clear_mask,
1412 			   void __user *buf,
1413 			   int *num_qss_entries,
1414 			   uint32_t *entry_size);
1415 
1416 int amdkfd_fence_wait_timeout(struct device_queue_manager *dqm,
1417 			      uint64_t fence_value,
1418 			      unsigned int timeout_ms);
1419 
1420 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
1421 				  unsigned int qid,
1422 				  u32 *mqd_size,
1423 				  u32 *ctl_stack_size);
1424 /* Packet Manager */
1425 
1426 #define KFD_FENCE_COMPLETED (100)
1427 #define KFD_FENCE_INIT   (10)
1428 
1429 /**
1430  * enum kfd_config_dequeue_wait_counts_cmd - Command for configuring
1431  *  dequeue wait counts.
1432  *
1433  * @KFD_DEQUEUE_WAIT_INIT: Set optimized dequeue wait counts for a
1434  *	certain ASICs. For these ASICs, this is default value used by RESET
1435  * @KFD_DEQUEUE_WAIT_RESET: Reset dequeue wait counts to the optimized value
1436  *	for certain ASICs. For others set it to default hardware reset value
1437  * @KFD_DEQUEUE_WAIT_SET_SCH_WAVE: Set context switch latency wait
1438  *
1439  */
1440 enum kfd_config_dequeue_wait_counts_cmd {
1441 	KFD_DEQUEUE_WAIT_INIT = 1,
1442 	KFD_DEQUEUE_WAIT_RESET = 2,
1443 	KFD_DEQUEUE_WAIT_SET_SCH_WAVE = 3
1444 };
1445 
1446 
1447 struct packet_manager {
1448 	struct device_queue_manager *dqm;
1449 	struct kernel_queue *priv_queue;
1450 	struct mutex lock;
1451 	bool allocated;
1452 	struct kfd_mem_obj *ib_buffer_obj;
1453 	unsigned int ib_size_bytes;
1454 	bool is_over_subscription;
1455 
1456 	const struct packet_manager_funcs *pmf;
1457 };
1458 
1459 struct packet_manager_funcs {
1460 	/* Support ASIC-specific packet formats for PM4 packets */
1461 	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
1462 			struct qcm_process_device *qpd);
1463 	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
1464 			uint64_t ib, size_t ib_size_in_dwords, bool chain);
1465 	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
1466 			struct scheduling_resources *res);
1467 	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
1468 			struct queue *q, bool is_static);
1469 	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
1470 			enum kfd_unmap_queues_filter mode,
1471 			uint32_t filter_param, bool reset);
1472 	int (*config_dequeue_wait_counts)(struct packet_manager *pm, uint32_t *buffer,
1473 			enum kfd_config_dequeue_wait_counts_cmd cmd, uint32_t value);
1474 	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
1475 			uint64_t fence_address,	uint64_t fence_value);
1476 	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
1477 
1478 	/* Packet sizes */
1479 	int map_process_size;
1480 	int runlist_size;
1481 	int set_resources_size;
1482 	int map_queues_size;
1483 	int unmap_queues_size;
1484 	int config_dequeue_wait_counts_size;
1485 	int query_status_size;
1486 	int release_mem_size;
1487 };
1488 
1489 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
1490 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
1491 extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;
1492 
1493 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
1494 void pm_uninit(struct packet_manager *pm);
1495 int pm_send_set_resources(struct packet_manager *pm,
1496 				struct scheduling_resources *res);
1497 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
1498 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
1499 				uint64_t fence_value);
1500 
1501 int pm_send_unmap_queue(struct packet_manager *pm,
1502 			enum kfd_unmap_queues_filter mode,
1503 			uint32_t filter_param, bool reset);
1504 
1505 void pm_release_ib(struct packet_manager *pm);
1506 
1507 int pm_config_dequeue_wait_counts(struct packet_manager *pm,
1508 			enum kfd_config_dequeue_wait_counts_cmd cmd,
1509 			uint32_t wait_counts_config);
1510 
1511 /* Following PM funcs can be shared among VI and AI */
1512 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
1513 
1514 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
1515 
1516 /* Events */
1517 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1518 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1519 extern const struct kfd_event_interrupt_class event_interrupt_class_v9_4_3;
1520 extern const struct kfd_event_interrupt_class event_interrupt_class_v10;
1521 extern const struct kfd_event_interrupt_class event_interrupt_class_v11;
1522 extern const struct kfd_event_interrupt_class event_interrupt_class_v12_1;
1523 
1524 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1525 
1526 int kfd_event_init_process(struct kfd_process *p);
1527 void kfd_event_free_process(struct kfd_process *p);
1528 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1529 int kfd_wait_on_events(struct kfd_process *p,
1530 		       uint32_t num_events, void __user *data,
1531 		       bool all, uint32_t *user_timeout_ms,
1532 		       uint32_t *wait_result);
1533 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
1534 				uint32_t valid_id_bits, bool signal_mailbox_updated);
1535 void kfd_signal_hw_exception_event(u32 pasid);
1536 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1537 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1538 int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset);
1539 
1540 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1541 		     uint32_t event_type, bool auto_reset, uint32_t node_id,
1542 		     uint32_t *event_id, uint32_t *event_trigger_data,
1543 		     uint64_t *event_page_offset, uint32_t *event_slot_index);
1544 
1545 int kfd_get_num_events(struct kfd_process *p);
1546 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1547 
1548 void kfd_signal_vm_fault_event_with_userptr(struct kfd_process *p, uint64_t gpu_va);
1549 
1550 void kfd_signal_vm_fault_event(struct kfd_process_device *pdd,
1551 				struct kfd_vm_fault_info *info,
1552 				struct kfd_hsa_memory_exception_data *data);
1553 
1554 void kfd_signal_reset_event(struct kfd_node *dev);
1555 
1556 void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid);
1557 void kfd_signal_process_terminate_event(struct kfd_process *p);
1558 
1559 static inline void kfd_flush_tlb(struct kfd_process_device *pdd)
1560 {
1561 	struct amdgpu_device *adev = pdd->dev->adev;
1562 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1563 
1564 	amdgpu_vm_flush_compute_tlb(adev, vm, TLB_FLUSH_HEAVYWEIGHT,
1565 				    pdd->dev->xcc_mask);
1566 }
1567 
1568 static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev)
1569 {
1570 	return KFD_GC_VERSION(dev) >= IP_VERSION(9, 4, 2) ||
1571 	       (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) ||
1572 	       KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0);
1573 }
1574 
1575 int kfd_send_exception_to_runtime(struct kfd_process *p,
1576 				unsigned int queue_id,
1577 				uint64_t error_reason);
1578 bool kfd_is_locked(struct kfd_dev *kfd);
1579 
1580 /* Compute profile */
1581 void kfd_inc_compute_active(struct kfd_node *dev);
1582 void kfd_dec_compute_active(struct kfd_node *dev);
1583 
1584 /* Cgroup Support */
1585 /* Check with device cgroup if @kfd device is accessible */
1586 static inline int kfd_devcgroup_check_permission(struct kfd_node *node)
1587 {
1588 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1589 	struct drm_device *ddev;
1590 
1591 	if (node->xcp)
1592 		ddev = node->xcp->ddev;
1593 	else
1594 		ddev = adev_to_drm(node->adev);
1595 
1596 	return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1597 					  ddev->render->index,
1598 					  DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1599 #else
1600 	return 0;
1601 #endif
1602 }
1603 
1604 static inline bool kfd_is_first_node(struct kfd_node *node)
1605 {
1606 	return (node == node->kfd->nodes[0]);
1607 }
1608 
1609 /* PTL support */
1610 int kfd_ptl_disable_request(struct kfd_process_device *pdd,
1611 		struct kfd_process *p);
1612 int kfd_ptl_disable_release(struct kfd_process_device *pdd,
1613 		struct kfd_process *p);
1614 
1615 /* Debugfs */
1616 #if defined(CONFIG_DEBUG_FS)
1617 
1618 void kfd_debugfs_init(void);
1619 void kfd_debugfs_fini(void);
1620 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1621 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1622 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1623 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1624 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1625 int pm_debugfs_runlist(struct seq_file *m, void *data);
1626 
1627 int kfd_debugfs_hang_hws(struct kfd_node *dev);
1628 int pm_debugfs_hang_hws(struct packet_manager *pm);
1629 int dqm_debugfs_hang_hws(struct device_queue_manager *dqm);
1630 
1631 void kfd_debugfs_add_process(struct kfd_process *p);
1632 void kfd_debugfs_remove_process(struct kfd_process *p);
1633 
1634 #else
1635 
1636 static inline void kfd_debugfs_init(void) {}
1637 static inline void kfd_debugfs_fini(void) {}
1638 static inline void kfd_debugfs_add_process(struct kfd_process *p) {}
1639 static inline void kfd_debugfs_remove_process(struct kfd_process *p) {}
1640 
1641 #endif
1642 
1643 #endif
1644