xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_priv.h (revision 6fa4bf3dce0668a96faca0024e382f4489a9cc9b)
1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #ifndef KFD_PRIV_H_INCLUDED
25 #define KFD_PRIV_H_INCLUDED
26 
27 #include <linux/hashtable.h>
28 #include <linux/mmu_notifier.h>
29 #include <linux/memremap.h>
30 #include <linux/mutex.h>
31 #include <linux/types.h>
32 #include <linux/atomic.h>
33 #include <linux/workqueue.h>
34 #include <linux/spinlock.h>
35 #include <linux/kfd_ioctl.h>
36 #include <linux/idr.h>
37 #include <linux/kfifo.h>
38 #include <linux/seq_file.h>
39 #include <linux/kref.h>
40 #include <linux/sysfs.h>
41 #include <linux/device_cgroup.h>
42 #include <drm/drm_file.h>
43 #include <drm/drm_drv.h>
44 #include <drm/drm_device.h>
45 #include <drm/drm_ioctl.h>
46 #include <kgd_kfd_interface.h>
47 #include <linux/swap.h>
48 
49 #include "amd_shared.h"
50 #include "amdgpu.h"
51 
52 #define KFD_MAX_RING_ENTRY_SIZE	8
53 
54 #define KFD_SYSFS_FILE_MODE 0444
55 
56 /* GPU ID hash width in bits */
57 #define KFD_GPU_ID_HASH_WIDTH 16
58 
59 /* Use upper bits of mmap offset to store KFD driver specific information.
60  * BITS[63:62] - Encode MMAP type
61  * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to
62  * BITS[45:0]  - MMAP offset value
63  *
64  * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
65  *  defines are w.r.t to PAGE_SIZE
66  */
67 #define KFD_MMAP_TYPE_SHIFT	62
68 #define KFD_MMAP_TYPE_MASK	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
69 #define KFD_MMAP_TYPE_DOORBELL	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
70 #define KFD_MMAP_TYPE_EVENTS	(0x2ULL << KFD_MMAP_TYPE_SHIFT)
71 #define KFD_MMAP_TYPE_RESERVED_MEM	(0x1ULL << KFD_MMAP_TYPE_SHIFT)
72 #define KFD_MMAP_TYPE_MMIO	(0x0ULL << KFD_MMAP_TYPE_SHIFT)
73 
74 #define KFD_MMAP_GPU_ID_SHIFT 46
75 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \
76 				<< KFD_MMAP_GPU_ID_SHIFT)
77 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\
78 				& KFD_MMAP_GPU_ID_MASK)
79 #define KFD_MMAP_GET_GPU_ID(offset)    ((offset & KFD_MMAP_GPU_ID_MASK) \
80 				>> KFD_MMAP_GPU_ID_SHIFT)
81 
82 /*
83  * When working with cp scheduler we should assign the HIQ manually or via
84  * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot
85  * definitions for Kaveri. In Kaveri only the first ME queues participates
86  * in the cp scheduling taking that in mind we set the HIQ slot in the
87  * second ME.
88  */
89 #define KFD_CIK_HIQ_PIPE 4
90 #define KFD_CIK_HIQ_QUEUE 0
91 
92 /* Macro for allocating structures */
93 #define kfd_alloc_struct(ptr_to_struct)	\
94 	((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
95 
96 #define KFD_MAX_NUM_OF_PROCESSES 512
97 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
98 
99 /*
100  * Size of the per-process TBA+TMA buffer: 2 pages
101  *
102  * The first chunk is the TBA used for the CWSR ISA code. The second
103  * chunk is used as TMA for user-mode trap handler setup in daisy-chain mode.
104  */
105 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
106 #define KFD_CWSR_TMA_OFFSET (PAGE_SIZE + 2048)
107 
108 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE		\
109 	(KFD_MAX_NUM_OF_PROCESSES *			\
110 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
111 
112 #define KFD_KERNEL_QUEUE_SIZE 2048
113 
114 #define KFD_UNMAP_LATENCY_MS	(4000)
115 
116 #define KFD_MAX_SDMA_QUEUES	128
117 
118 /*
119  * 512 = 0x200
120  * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
121  * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
122  * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
123  * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
124  * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
125  */
126 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
127 
128 /**
129  * enum kfd_ioctl_flags - KFD ioctl flags
130  * Various flags that can be set in &amdkfd_ioctl_desc.flags to control how
131  * userspace can use a given ioctl.
132  */
133 enum kfd_ioctl_flags {
134 	/*
135 	 * @KFD_IOC_FLAG_CHECKPOINT_RESTORE:
136 	 * Certain KFD ioctls such as AMDKFD_IOC_CRIU_OP can potentially
137 	 * perform privileged operations and load arbitrary data into MQDs and
138 	 * eventually HQD registers when the queue is mapped by HWS. In order to
139 	 * prevent this we should perform additional security checks.
140 	 *
141 	 * This is equivalent to callers with the CHECKPOINT_RESTORE capability.
142 	 *
143 	 * Note: Since earlier versions of docker do not support CHECKPOINT_RESTORE,
144 	 * we also allow ioctls with SYS_ADMIN capability.
145 	 */
146 	KFD_IOC_FLAG_CHECKPOINT_RESTORE = BIT(0),
147 };
148 /*
149  * Kernel module parameter to specify maximum number of supported queues per
150  * device
151  */
152 extern int max_num_of_queues_per_device;
153 
154 
155 /* Kernel module parameter to specify the scheduling policy */
156 extern int sched_policy;
157 
158 /*
159  * Kernel module parameter to specify the maximum process
160  * number per HW scheduler
161  */
162 extern int hws_max_conc_proc;
163 
164 extern int cwsr_enable;
165 
166 /*
167  * Kernel module parameter to specify whether to send sigterm to HSA process on
168  * unhandled exception
169  */
170 extern int send_sigterm;
171 
172 /*
173  * This kernel module is used to simulate large bar machine on non-large bar
174  * enabled machines.
175  */
176 extern int debug_largebar;
177 
178 /* Set sh_mem_config.retry_disable on GFX v9 */
179 extern int amdgpu_noretry;
180 
181 /* Halt if HWS hang is detected */
182 extern int halt_if_hws_hang;
183 
184 /* Whether MEC FW support GWS barriers */
185 extern bool hws_gws_support;
186 
187 /* Queue preemption timeout in ms */
188 extern int queue_preemption_timeout_ms;
189 
190 /*
191  * Don't evict process queues on vm fault
192  */
193 extern int amdgpu_no_queue_eviction_on_vm_fault;
194 
195 /* Enable eviction debug messages */
196 extern bool debug_evictions;
197 
198 extern struct mutex kfd_processes_mutex;
199 
200 enum cache_policy {
201 	cache_policy_coherent,
202 	cache_policy_noncoherent
203 };
204 
205 #define KFD_GC_VERSION(dev) (amdgpu_ip_version((dev)->adev, GC_HWIP, 0))
206 #define KFD_IS_SOC15(dev)   ((KFD_GC_VERSION(dev)) >= (IP_VERSION(9, 0, 1)))
207 #define KFD_SUPPORT_XNACK_PER_PROCESS(dev)\
208 	((KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2)) ||	\
209 	 (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) ||	\
210 	 (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 4)))
211 
212 struct kfd_node;
213 
214 struct kfd_event_interrupt_class {
215 	bool (*interrupt_isr)(struct kfd_node *dev,
216 			const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
217 			bool *patched_flag);
218 	void (*interrupt_wq)(struct kfd_node *dev,
219 			const uint32_t *ih_ring_entry);
220 };
221 
222 struct kfd_device_info {
223 	uint32_t gfx_target_version;
224 	const struct kfd_event_interrupt_class *event_interrupt_class;
225 	unsigned int max_pasid_bits;
226 	unsigned int max_no_of_hqd;
227 	unsigned int doorbell_size;
228 	size_t ih_ring_entry_size;
229 	uint8_t num_of_watch_points;
230 	uint16_t mqd_size_aligned;
231 	bool supports_cwsr;
232 	bool needs_pci_atomics;
233 	uint32_t no_atomic_fw_version;
234 	unsigned int num_sdma_queues_per_engine;
235 	unsigned int num_reserved_sdma_queues_per_engine;
236 	DECLARE_BITMAP(reserved_sdma_queues_bitmap, KFD_MAX_SDMA_QUEUES);
237 };
238 
239 unsigned int kfd_get_num_sdma_engines(struct kfd_node *kdev);
240 unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *kdev);
241 
242 struct kfd_mem_obj {
243 	uint32_t range_start;
244 	uint32_t range_end;
245 	uint64_t gpu_addr;
246 	uint32_t *cpu_ptr;
247 	void *gtt_mem;
248 };
249 
250 struct kfd_vmid_info {
251 	uint32_t first_vmid_kfd;
252 	uint32_t last_vmid_kfd;
253 	uint32_t vmid_num_kfd;
254 };
255 
256 #define MAX_KFD_NODES	8
257 
258 struct kfd_dev;
259 
260 struct kfd_node {
261 	unsigned int node_id;
262 	struct amdgpu_device *adev;     /* Duplicated here along with keeping
263 					 * a copy in kfd_dev to save a hop
264 					 */
265 	const struct kfd2kgd_calls *kfd2kgd; /* Duplicated here along with
266 					      * keeping a copy in kfd_dev to
267 					      * save a hop
268 					      */
269 	struct kfd_vmid_info vm_info;
270 	unsigned int id;                /* topology stub index */
271 	uint32_t xcc_mask; /* Instance mask of XCCs present */
272 	struct amdgpu_xcp *xcp;
273 
274 	/* Interrupts */
275 	struct kfifo ih_fifo;
276 	struct workqueue_struct *ih_wq;
277 	struct work_struct interrupt_work;
278 	spinlock_t interrupt_lock;
279 
280 	/*
281 	 * Interrupts of interest to KFD are copied
282 	 * from the HW ring into a SW ring.
283 	 */
284 	bool interrupts_active;
285 	uint32_t interrupt_bitmap; /* Only used for GFX 9.4.3 */
286 
287 	/* QCM Device instance */
288 	struct device_queue_manager *dqm;
289 
290 	/* Global GWS resource shared between processes */
291 	void *gws;
292 	bool gws_debug_workaround;
293 
294 	/* Clients watching SMI events */
295 	struct list_head smi_clients;
296 	spinlock_t smi_lock;
297 	uint32_t reset_seq_num;
298 
299 	/* SRAM ECC flag */
300 	atomic_t sram_ecc_flag;
301 
302 	/*spm process id */
303 	unsigned int spm_pasid;
304 
305 	/* Maximum process number mapped to HW scheduler */
306 	unsigned int max_proc_per_quantum;
307 
308 	unsigned int compute_vmid_bitmap;
309 
310 	struct kfd_local_mem_info local_mem_info;
311 
312 	struct kfd_dev *kfd;
313 };
314 
315 struct kfd_dev {
316 	struct amdgpu_device *adev;
317 
318 	struct kfd_device_info device_info;
319 
320 	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
321 					   * page used by kernel queue
322 					   */
323 
324 	struct kgd2kfd_shared_resources shared_resources;
325 
326 	const struct kfd2kgd_calls *kfd2kgd;
327 	struct mutex doorbell_mutex;
328 
329 	void *gtt_mem;
330 	uint64_t gtt_start_gpu_addr;
331 	void *gtt_start_cpu_ptr;
332 	void *gtt_sa_bitmap;
333 	struct mutex gtt_sa_lock;
334 	unsigned int gtt_sa_chunk_size;
335 	unsigned int gtt_sa_num_of_chunks;
336 
337 	bool init_complete;
338 
339 	/* Firmware versions */
340 	uint16_t mec_fw_version;
341 	uint16_t mec2_fw_version;
342 	uint16_t sdma_fw_version;
343 
344 	/* CWSR */
345 	bool cwsr_enabled;
346 	const void *cwsr_isa;
347 	unsigned int cwsr_isa_size;
348 
349 	/* xGMI */
350 	uint64_t hive_id;
351 
352 	bool pci_atomic_requested;
353 
354 	/* Compute Profile ref. count */
355 	atomic_t compute_profile;
356 
357 	struct ida doorbell_ida;
358 	unsigned int max_doorbell_slices;
359 
360 	int noretry;
361 
362 	struct kfd_node *nodes[MAX_KFD_NODES];
363 	unsigned int num_nodes;
364 
365 	/* Track per device allocated watch points */
366 	uint32_t alloc_watch_ids;
367 	spinlock_t watch_points_lock;
368 
369 	/* Kernel doorbells for KFD device */
370 	struct amdgpu_bo *doorbells;
371 
372 	/* bitmap for dynamic doorbell allocation from doorbell object */
373 	unsigned long *doorbell_bitmap;
374 };
375 
376 enum kfd_mempool {
377 	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
378 	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
379 	KFD_MEMPOOL_FRAMEBUFFER = 3,
380 };
381 
382 /* Character device interface */
383 int kfd_chardev_init(void);
384 void kfd_chardev_exit(void);
385 
386 /**
387  * enum kfd_unmap_queues_filter - Enum for queue filters.
388  *
389  * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
390  *						running queues list.
391  *
392  * @KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: Preempts all non-static queues
393  *						in the run list.
394  *
395  * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
396  *						specific process.
397  *
398  */
399 enum kfd_unmap_queues_filter {
400 	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES = 1,
401 	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES = 2,
402 	KFD_UNMAP_QUEUES_FILTER_BY_PASID = 3
403 };
404 
405 /**
406  * enum kfd_queue_type - Enum for various queue types.
407  *
408  * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
409  *
410  * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
411  *
412  * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
413  *
414  * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
415  *
416  * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
417  */
418 enum kfd_queue_type  {
419 	KFD_QUEUE_TYPE_COMPUTE,
420 	KFD_QUEUE_TYPE_SDMA,
421 	KFD_QUEUE_TYPE_HIQ,
422 	KFD_QUEUE_TYPE_DIQ,
423 	KFD_QUEUE_TYPE_SDMA_XGMI
424 };
425 
426 enum kfd_queue_format {
427 	KFD_QUEUE_FORMAT_PM4,
428 	KFD_QUEUE_FORMAT_AQL
429 };
430 
431 enum KFD_QUEUE_PRIORITY {
432 	KFD_QUEUE_PRIORITY_MINIMUM = 0,
433 	KFD_QUEUE_PRIORITY_MAXIMUM = 15
434 };
435 
436 /**
437  * struct queue_properties
438  *
439  * @type: The queue type.
440  *
441  * @queue_id: Queue identifier.
442  *
443  * @queue_address: Queue ring buffer address.
444  *
445  * @queue_size: Queue ring buffer size.
446  *
447  * @priority: Defines the queue priority relative to other queues in the
448  * process.
449  * This is just an indication and HW scheduling may override the priority as
450  * necessary while keeping the relative prioritization.
451  * the priority granularity is from 0 to f which f is the highest priority.
452  * currently all queues are initialized with the highest priority.
453  *
454  * @queue_percent: This field is partially implemented and currently a zero in
455  * this field defines that the queue is non active.
456  *
457  * @read_ptr: User space address which points to the number of dwords the
458  * cp read from the ring buffer. This field updates automatically by the H/W.
459  *
460  * @write_ptr: Defines the number of dwords written to the ring buffer.
461  *
462  * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
463  * buffer. This field should be similar to write_ptr and the user should
464  * update this field after updating the write_ptr.
465  *
466  * @doorbell_off: The doorbell offset in the doorbell pci-bar.
467  *
468  * @is_interop: Defines if this is a interop queue. Interop queue means that
469  * the queue can access both graphics and compute resources.
470  *
471  * @is_evicted: Defines if the queue is evicted. Only active queues
472  * are evicted, rendering them inactive.
473  *
474  * @is_active: Defines if the queue is active or not. @is_active and
475  * @is_evicted are protected by the DQM lock.
476  *
477  * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
478  * @is_gws should be protected by the DQM lock, since changing it can yield the
479  * possibility of updating DQM state on number of GWS queues.
480  *
481  * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
482  * of the queue.
483  *
484  * This structure represents the queue properties for each queue no matter if
485  * it's user mode or kernel mode queue.
486  *
487  */
488 
489 struct queue_properties {
490 	enum kfd_queue_type type;
491 	enum kfd_queue_format format;
492 	unsigned int queue_id;
493 	uint64_t queue_address;
494 	uint64_t  queue_size;
495 	uint32_t priority;
496 	uint32_t queue_percent;
497 	void __user *read_ptr;
498 	void __user *write_ptr;
499 	void __iomem *doorbell_ptr;
500 	uint32_t doorbell_off;
501 	bool is_interop;
502 	bool is_evicted;
503 	bool is_suspended;
504 	bool is_being_destroyed;
505 	bool is_active;
506 	bool is_gws;
507 	uint32_t pm4_target_xcc;
508 	bool is_dbg_wa;
509 	bool is_user_cu_masked;
510 	/* Not relevant for user mode queues in cp scheduling */
511 	unsigned int vmid;
512 	/* Relevant only for sdma queues*/
513 	uint32_t sdma_engine_id;
514 	uint32_t sdma_queue_id;
515 	uint32_t sdma_vm_addr;
516 	/* Relevant only for VI */
517 	uint64_t eop_ring_buffer_address;
518 	uint32_t eop_ring_buffer_size;
519 	uint64_t ctx_save_restore_area_address;
520 	uint32_t ctx_save_restore_area_size;
521 	uint32_t ctl_stack_size;
522 	uint64_t tba_addr;
523 	uint64_t tma_addr;
524 	uint64_t exception_status;
525 
526 	struct amdgpu_bo *wptr_bo;
527 };
528 
529 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\
530 			    (q).queue_address != 0 &&	\
531 			    (q).queue_percent > 0 &&	\
532 			    !(q).is_evicted &&		\
533 			    !(q).is_suspended)
534 
535 enum mqd_update_flag {
536 	UPDATE_FLAG_DBG_WA_ENABLE = 1,
537 	UPDATE_FLAG_DBG_WA_DISABLE = 2,
538 	UPDATE_FLAG_IS_GWS = 4, /* quirk for gfx9 IP */
539 };
540 
541 struct mqd_update_info {
542 	union {
543 		struct {
544 			uint32_t count; /* Must be a multiple of 32 */
545 			uint32_t *ptr;
546 		} cu_mask;
547 	};
548 	enum mqd_update_flag update_flag;
549 };
550 
551 /**
552  * struct queue
553  *
554  * @list: Queue linked list.
555  *
556  * @mqd: The queue MQD (memory queue descriptor).
557  *
558  * @mqd_mem_obj: The MQD local gpu memory object.
559  *
560  * @gart_mqd_addr: The MQD gart mc address.
561  *
562  * @properties: The queue properties.
563  *
564  * @mec: Used only in no cp scheduling mode and identifies to micro engine id
565  *	 that the queue should be executed on.
566  *
567  * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
568  *	  id.
569  *
570  * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
571  *
572  * @process: The kfd process that created this queue.
573  *
574  * @device: The kfd device that created this queue.
575  *
576  * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
577  * otherwise.
578  *
579  * This structure represents user mode compute queues.
580  * It contains all the necessary data to handle such queues.
581  *
582  */
583 
584 struct queue {
585 	struct list_head list;
586 	void *mqd;
587 	struct kfd_mem_obj *mqd_mem_obj;
588 	uint64_t gart_mqd_addr;
589 	struct queue_properties properties;
590 
591 	uint32_t mec;
592 	uint32_t pipe;
593 	uint32_t queue;
594 
595 	unsigned int sdma_id;
596 	unsigned int doorbell_id;
597 
598 	struct kfd_process	*process;
599 	struct kfd_node		*device;
600 	void *gws;
601 
602 	/* procfs */
603 	struct kobject kobj;
604 
605 	void *gang_ctx_bo;
606 	uint64_t gang_ctx_gpu_addr;
607 	void *gang_ctx_cpu_ptr;
608 
609 	struct amdgpu_bo *wptr_bo_gart;
610 };
611 
612 enum KFD_MQD_TYPE {
613 	KFD_MQD_TYPE_HIQ = 0,		/* for hiq */
614 	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
615 	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
616 	KFD_MQD_TYPE_DIQ,		/* for diq */
617 	KFD_MQD_TYPE_MAX
618 };
619 
620 enum KFD_PIPE_PRIORITY {
621 	KFD_PIPE_PRIORITY_CS_LOW = 0,
622 	KFD_PIPE_PRIORITY_CS_MEDIUM,
623 	KFD_PIPE_PRIORITY_CS_HIGH
624 };
625 
626 struct scheduling_resources {
627 	unsigned int vmid_mask;
628 	enum kfd_queue_type type;
629 	uint64_t queue_mask;
630 	uint64_t gws_mask;
631 	uint32_t oac_mask;
632 	uint32_t gds_heap_base;
633 	uint32_t gds_heap_size;
634 };
635 
636 struct process_queue_manager {
637 	/* data */
638 	struct kfd_process	*process;
639 	struct list_head	queues;
640 	unsigned long		*queue_slot_bitmap;
641 };
642 
643 struct qcm_process_device {
644 	/* The Device Queue Manager that owns this data */
645 	struct device_queue_manager *dqm;
646 	struct process_queue_manager *pqm;
647 	/* Queues list */
648 	struct list_head queues_list;
649 	struct list_head priv_queue_list;
650 
651 	unsigned int queue_count;
652 	unsigned int vmid;
653 	bool is_debug;
654 	unsigned int evicted; /* eviction counter, 0=active */
655 
656 	/* This flag tells if we should reset all wavefronts on
657 	 * process termination
658 	 */
659 	bool reset_wavefronts;
660 
661 	/* This flag tells us if this process has a GWS-capable
662 	 * queue that will be mapped into the runlist. It's
663 	 * possible to request a GWS BO, but not have the queue
664 	 * currently mapped, and this changes how the MAP_PROCESS
665 	 * PM4 packet is configured.
666 	 */
667 	bool mapped_gws_queue;
668 
669 	/* All the memory management data should be here too */
670 	uint64_t gds_context_area;
671 	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
672 	uint64_t page_table_base;
673 	uint32_t sh_mem_config;
674 	uint32_t sh_mem_bases;
675 	uint32_t sh_mem_ape1_base;
676 	uint32_t sh_mem_ape1_limit;
677 	uint32_t gds_size;
678 	uint32_t num_gws;
679 	uint32_t num_oac;
680 	uint32_t sh_hidden_private_base;
681 
682 	/* CWSR memory */
683 	struct kgd_mem *cwsr_mem;
684 	void *cwsr_kaddr;
685 	uint64_t cwsr_base;
686 	uint64_t tba_addr;
687 	uint64_t tma_addr;
688 
689 	/* IB memory */
690 	struct kgd_mem *ib_mem;
691 	uint64_t ib_base;
692 	void *ib_kaddr;
693 
694 	/* doorbells for kfd process */
695 	struct amdgpu_bo *proc_doorbells;
696 
697 	/* bitmap for dynamic doorbell allocation from the bo */
698 	unsigned long *doorbell_bitmap;
699 };
700 
701 /* KFD Memory Eviction */
702 
703 /* Approx. wait time before attempting to restore evicted BOs */
704 #define PROCESS_RESTORE_TIME_MS 100
705 /* Approx. back off time if restore fails due to lack of memory */
706 #define PROCESS_BACK_OFF_TIME_MS 100
707 /* Approx. time before evicting the process again */
708 #define PROCESS_ACTIVE_TIME_MS 10
709 
710 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
711  * idr_handle in the least significant 4 bytes
712  */
713 #define MAKE_HANDLE(gpu_id, idr_handle) \
714 	(((uint64_t)(gpu_id) << 32) + idr_handle)
715 #define GET_GPU_ID(handle) (handle >> 32)
716 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
717 
718 enum kfd_pdd_bound {
719 	PDD_UNBOUND = 0,
720 	PDD_BOUND,
721 	PDD_BOUND_SUSPENDED,
722 };
723 
724 #define MAX_SYSFS_FILENAME_LEN 15
725 
726 /*
727  * SDMA counter runs at 100MHz frequency.
728  * We display SDMA activity in microsecond granularity in sysfs.
729  * As a result, the divisor is 100.
730  */
731 #define SDMA_ACTIVITY_DIVISOR  100
732 
733 /* Data that is per-process-per device. */
734 struct kfd_process_device {
735 	/* The device that owns this data. */
736 	struct kfd_node *dev;
737 
738 	/* The process that owns this kfd_process_device. */
739 	struct kfd_process *process;
740 
741 	/* per-process-per device QCM data structure */
742 	struct qcm_process_device qpd;
743 
744 	/*Apertures*/
745 	uint64_t lds_base;
746 	uint64_t lds_limit;
747 	uint64_t gpuvm_base;
748 	uint64_t gpuvm_limit;
749 	uint64_t scratch_base;
750 	uint64_t scratch_limit;
751 
752 	/* VM context for GPUVM allocations */
753 	struct file *drm_file;
754 	void *drm_priv;
755 
756 	/* GPUVM allocations storage */
757 	struct idr alloc_idr;
758 
759 	/* Flag used to tell the pdd has dequeued from the dqm.
760 	 * This is used to prevent dev->dqm->ops.process_termination() from
761 	 * being called twice when it is already called in IOMMU callback
762 	 * function.
763 	 */
764 	bool already_dequeued;
765 	bool runtime_inuse;
766 
767 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
768 	enum kfd_pdd_bound bound;
769 
770 	/* VRAM usage */
771 	uint64_t vram_usage;
772 	struct attribute attr_vram;
773 	char vram_filename[MAX_SYSFS_FILENAME_LEN];
774 
775 	/* SDMA activity tracking */
776 	uint64_t sdma_past_activity_counter;
777 	struct attribute attr_sdma;
778 	char sdma_filename[MAX_SYSFS_FILENAME_LEN];
779 
780 	/* Eviction activity tracking */
781 	uint64_t last_evict_timestamp;
782 	atomic64_t evict_duration_counter;
783 	struct attribute attr_evict;
784 
785 	struct kobject *kobj_stats;
786 
787 	/*
788 	 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
789 	 * that is associated with device encoded by "this" struct instance. The
790 	 * value reflects CU usage by all of the waves launched by this process
791 	 * on this device. A very important property of occupancy parameter is
792 	 * that its value is a snapshot of current use.
793 	 *
794 	 * Following is to be noted regarding how this parameter is reported:
795 	 *
796 	 *  The number of waves that a CU can launch is limited by couple of
797 	 *  parameters. These are encoded by struct amdgpu_cu_info instance
798 	 *  that is part of every device definition. For GFX9 devices this
799 	 *  translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
800 	 *  do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
801 	 *  when they do use scratch memory. This could change for future
802 	 *  devices and therefore this example should be considered as a guide.
803 	 *
804 	 *  All CU's of a device are available for the process. This may not be true
805 	 *  under certain conditions - e.g. CU masking.
806 	 *
807 	 *  Finally number of CU's that are occupied by a process is affected by both
808 	 *  number of CU's a device has along with number of other competing processes
809 	 */
810 	struct attribute attr_cu_occupancy;
811 
812 	/* sysfs counters for GPU retry fault and page migration tracking */
813 	struct kobject *kobj_counters;
814 	struct attribute attr_faults;
815 	struct attribute attr_page_in;
816 	struct attribute attr_page_out;
817 	uint64_t faults;
818 	uint64_t page_in;
819 	uint64_t page_out;
820 
821 	/* Exception code status*/
822 	uint64_t exception_status;
823 	void *vm_fault_exc_data;
824 	size_t vm_fault_exc_data_size;
825 
826 	/* Tracks debug per-vmid request settings */
827 	uint32_t spi_dbg_override;
828 	uint32_t spi_dbg_launch_mode;
829 	uint32_t watch_points[4];
830 	uint32_t alloc_watch_ids;
831 
832 	/*
833 	 * If this process has been checkpointed before, then the user
834 	 * application will use the original gpu_id on the
835 	 * checkpointed node to refer to this device.
836 	 */
837 	uint32_t user_gpu_id;
838 
839 	void *proc_ctx_bo;
840 	uint64_t proc_ctx_gpu_addr;
841 	void *proc_ctx_cpu_ptr;
842 };
843 
844 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
845 
846 struct svm_range_list {
847 	struct mutex			lock;
848 	struct rb_root_cached		objects;
849 	struct list_head		list;
850 	struct work_struct		deferred_list_work;
851 	struct list_head		deferred_range_list;
852 	struct list_head                criu_svm_metadata_list;
853 	spinlock_t			deferred_list_lock;
854 	atomic_t			evicted_ranges;
855 	atomic_t			drain_pagefaults;
856 	struct delayed_work		restore_work;
857 	DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);
858 	struct task_struct		*faulting_task;
859 };
860 
861 /* Process data */
862 struct kfd_process {
863 	/*
864 	 * kfd_process are stored in an mm_struct*->kfd_process*
865 	 * hash table (kfd_processes in kfd_process.c)
866 	 */
867 	struct hlist_node kfd_processes;
868 
869 	/*
870 	 * Opaque pointer to mm_struct. We don't hold a reference to
871 	 * it so it should never be dereferenced from here. This is
872 	 * only used for looking up processes by their mm.
873 	 */
874 	void *mm;
875 
876 	struct kref ref;
877 	struct work_struct release_work;
878 
879 	struct mutex mutex;
880 
881 	/*
882 	 * In any process, the thread that started main() is the lead
883 	 * thread and outlives the rest.
884 	 * It is here because amd_iommu_bind_pasid wants a task_struct.
885 	 * It can also be used for safely getting a reference to the
886 	 * mm_struct of the process.
887 	 */
888 	struct task_struct *lead_thread;
889 
890 	/* We want to receive a notification when the mm_struct is destroyed */
891 	struct mmu_notifier mmu_notifier;
892 
893 	u32 pasid;
894 
895 	/*
896 	 * Array of kfd_process_device pointers,
897 	 * one for each device the process is using.
898 	 */
899 	struct kfd_process_device *pdds[MAX_GPU_INSTANCE];
900 	uint32_t n_pdds;
901 
902 	struct process_queue_manager pqm;
903 
904 	/*Is the user space process 32 bit?*/
905 	bool is_32bit_user_mode;
906 
907 	/* Event-related data */
908 	struct mutex event_mutex;
909 	/* Event ID allocator and lookup */
910 	struct idr event_idr;
911 	/* Event page */
912 	u64 signal_handle;
913 	struct kfd_signal_page *signal_page;
914 	size_t signal_mapped_size;
915 	size_t signal_event_count;
916 	bool signal_event_limit_reached;
917 
918 	/* Information used for memory eviction */
919 	void *kgd_process_info;
920 	/* Eviction fence that is attached to all the BOs of this process. The
921 	 * fence will be triggered during eviction and new one will be created
922 	 * during restore
923 	 */
924 	struct dma_fence __rcu *ef;
925 
926 	/* Work items for evicting and restoring BOs */
927 	struct delayed_work eviction_work;
928 	struct delayed_work restore_work;
929 	/* seqno of the last scheduled eviction */
930 	unsigned int last_eviction_seqno;
931 	/* Approx. the last timestamp (in jiffies) when the process was
932 	 * restored after an eviction
933 	 */
934 	unsigned long last_restore_timestamp;
935 
936 	/* Indicates device process is debug attached with reserved vmid. */
937 	bool debug_trap_enabled;
938 
939 	/* per-process-per device debug event fd file */
940 	struct file *dbg_ev_file;
941 
942 	/* If the process is a kfd debugger, we need to know so we can clean
943 	 * up at exit time.  If a process enables debugging on itself, it does
944 	 * its own clean-up, so we don't set the flag here.  We track this by
945 	 * counting the number of processes this process is debugging.
946 	 */
947 	atomic_t debugged_process_count;
948 
949 	/* If the process is a debugged, this is the debugger process */
950 	struct kfd_process *debugger_process;
951 
952 	/* Kobj for our procfs */
953 	struct kobject *kobj;
954 	struct kobject *kobj_queues;
955 	struct attribute attr_pasid;
956 
957 	/* Keep track cwsr init */
958 	bool has_cwsr;
959 
960 	/* Exception code enable mask and status */
961 	uint64_t exception_enable_mask;
962 	uint64_t exception_status;
963 
964 	/* Used to drain stale interrupts */
965 	wait_queue_head_t wait_irq_drain;
966 	bool irq_drain_is_open;
967 
968 	/* shared virtual memory registered by this process */
969 	struct svm_range_list svms;
970 
971 	bool xnack_enabled;
972 
973 	/* Work area for debugger event writer worker. */
974 	struct work_struct debug_event_workarea;
975 
976 	/* Tracks debug per-vmid request for debug flags */
977 	u32 dbg_flags;
978 
979 	atomic_t poison;
980 	/* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */
981 	bool queues_paused;
982 
983 	/* Tracks runtime enable status */
984 	struct semaphore runtime_enable_sema;
985 	bool is_runtime_retry;
986 	struct kfd_runtime_info runtime_info;
987 };
988 
989 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
990 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
991 extern struct srcu_struct kfd_processes_srcu;
992 
993 /**
994  * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
995  *
996  * @filep: pointer to file structure.
997  * @p: amdkfd process pointer.
998  * @data: pointer to arg that was copied from user.
999  *
1000  * Return: returns ioctl completion code.
1001  */
1002 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
1003 				void *data);
1004 
1005 struct amdkfd_ioctl_desc {
1006 	unsigned int cmd;
1007 	int flags;
1008 	amdkfd_ioctl_t *func;
1009 	unsigned int cmd_drv;
1010 	const char *name;
1011 };
1012 bool kfd_dev_is_large_bar(struct kfd_node *dev);
1013 
1014 int kfd_process_create_wq(void);
1015 void kfd_process_destroy_wq(void);
1016 void kfd_cleanup_processes(void);
1017 struct kfd_process *kfd_create_process(struct task_struct *thread);
1018 struct kfd_process *kfd_get_process(const struct task_struct *task);
1019 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid);
1020 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
1021 
1022 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
1023 int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
1024 				uint32_t *gpuid, uint32_t *gpuidx);
1025 static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
1026 				uint32_t gpuidx, uint32_t *gpuid) {
1027 	return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
1028 }
1029 static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
1030 				struct kfd_process *p, uint32_t gpuidx) {
1031 	return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;
1032 }
1033 
1034 void kfd_unref_process(struct kfd_process *p);
1035 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger);
1036 int kfd_process_restore_queues(struct kfd_process *p);
1037 void kfd_suspend_all_processes(void);
1038 int kfd_resume_all_processes(void);
1039 
1040 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *process,
1041 							 uint32_t gpu_id);
1042 
1043 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id);
1044 
1045 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1046 			       struct file *drm_file);
1047 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1048 						struct kfd_process *p);
1049 struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1050 							struct kfd_process *p);
1051 struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1052 							struct kfd_process *p);
1053 
1054 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
1055 
1056 int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
1057 			  struct vm_area_struct *vma);
1058 
1059 /* KFD process API for creating and translating handles */
1060 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1061 					void *mem);
1062 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
1063 					int handle);
1064 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1065 					int handle);
1066 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid);
1067 
1068 /* PASIDs */
1069 int kfd_pasid_init(void);
1070 void kfd_pasid_exit(void);
1071 bool kfd_set_pasid_limit(unsigned int new_limit);
1072 unsigned int kfd_get_pasid_limit(void);
1073 u32 kfd_pasid_alloc(void);
1074 void kfd_pasid_free(u32 pasid);
1075 
1076 /* Doorbells */
1077 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
1078 int kfd_doorbell_init(struct kfd_dev *kfd);
1079 void kfd_doorbell_fini(struct kfd_dev *kfd);
1080 int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process,
1081 		      struct vm_area_struct *vma);
1082 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
1083 					unsigned int *doorbell_off);
1084 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
1085 u32 read_kernel_doorbell(u32 __iomem *db);
1086 void write_kernel_doorbell(void __iomem *db, u32 value);
1087 void write_kernel_doorbell64(void __iomem *db, u64 value);
1088 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
1089 					struct kfd_process_device *pdd,
1090 					unsigned int doorbell_id);
1091 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
1092 int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
1093 				struct kfd_process_device *pdd);
1094 void kfd_free_process_doorbells(struct kfd_dev *kfd,
1095 				struct kfd_process_device *pdd);
1096 /* GTT Sub-Allocator */
1097 
1098 int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size,
1099 			struct kfd_mem_obj **mem_obj);
1100 
1101 int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj);
1102 
1103 extern struct device *kfd_device;
1104 
1105 /* KFD's procfs */
1106 void kfd_procfs_init(void);
1107 void kfd_procfs_shutdown(void);
1108 int kfd_procfs_add_queue(struct queue *q);
1109 void kfd_procfs_del_queue(struct queue *q);
1110 
1111 /* Topology */
1112 int kfd_topology_init(void);
1113 void kfd_topology_shutdown(void);
1114 int kfd_topology_add_device(struct kfd_node *gpu);
1115 int kfd_topology_remove_device(struct kfd_node *gpu);
1116 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
1117 						uint32_t proximity_domain);
1118 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
1119 						uint32_t proximity_domain);
1120 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
1121 struct kfd_node *kfd_device_by_id(uint32_t gpu_id);
1122 struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev);
1123 static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t node_id,
1124 					uint32_t vmid)
1125 {
1126 	return (node->interrupt_bitmap & (1 << node_id)) != 0 &&
1127 	       (node->compute_vmid_bitmap & (1 << vmid)) != 0;
1128 }
1129 static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev,
1130 					uint32_t node_id, uint32_t vmid) {
1131 	struct kfd_dev *dev = adev->kfd.dev;
1132 	uint32_t i;
1133 
1134 	if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) &&
1135 	    KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 4))
1136 		return dev->nodes[0];
1137 
1138 	for (i = 0; i < dev->num_nodes; i++)
1139 		if (kfd_irq_is_from_node(dev->nodes[i], node_id, vmid))
1140 			return dev->nodes[i];
1141 
1142 	return NULL;
1143 }
1144 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev);
1145 int kfd_numa_node_to_apic_id(int numa_node_id);
1146 
1147 /* Interrupts */
1148 #define	KFD_IRQ_FENCE_CLIENTID	0xff
1149 #define	KFD_IRQ_FENCE_SOURCEID	0xff
1150 #define	KFD_IRQ_IS_FENCE(client, source)				\
1151 				((client) == KFD_IRQ_FENCE_CLIENTID &&	\
1152 				(source) == KFD_IRQ_FENCE_SOURCEID)
1153 int kfd_interrupt_init(struct kfd_node *dev);
1154 void kfd_interrupt_exit(struct kfd_node *dev);
1155 bool enqueue_ih_ring_entry(struct kfd_node *kfd, const void *ih_ring_entry);
1156 bool interrupt_is_wanted(struct kfd_node *dev,
1157 				const uint32_t *ih_ring_entry,
1158 				uint32_t *patched_ihre, bool *flag);
1159 int kfd_process_drain_interrupts(struct kfd_process_device *pdd);
1160 void kfd_process_close_interrupt_drain(unsigned int pasid);
1161 
1162 /* amdkfd Apertures */
1163 int kfd_init_apertures(struct kfd_process *process);
1164 
1165 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1166 				  uint64_t tba_addr,
1167 				  uint64_t tma_addr);
1168 void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1169 				     bool enabled);
1170 
1171 /* CWSR initialization */
1172 int kfd_process_init_cwsr_apu(struct kfd_process *process, struct file *filep);
1173 
1174 /* CRIU */
1175 /*
1176  * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private
1177  * structures:
1178  * kfd_criu_process_priv_data
1179  * kfd_criu_device_priv_data
1180  * kfd_criu_bo_priv_data
1181  * kfd_criu_queue_priv_data
1182  * kfd_criu_event_priv_data
1183  * kfd_criu_svm_range_priv_data
1184  */
1185 
1186 #define KFD_CRIU_PRIV_VERSION 1
1187 
1188 struct kfd_criu_process_priv_data {
1189 	uint32_t version;
1190 	uint32_t xnack_mode;
1191 };
1192 
1193 struct kfd_criu_device_priv_data {
1194 	/* For future use */
1195 	uint64_t reserved;
1196 };
1197 
1198 struct kfd_criu_bo_priv_data {
1199 	uint64_t user_addr;
1200 	uint32_t idr_handle;
1201 	uint32_t mapped_gpuids[MAX_GPU_INSTANCE];
1202 };
1203 
1204 /*
1205  * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data,
1206  * kfd_criu_svm_range_priv_data is the object type
1207  */
1208 enum kfd_criu_object_type {
1209 	KFD_CRIU_OBJECT_TYPE_QUEUE,
1210 	KFD_CRIU_OBJECT_TYPE_EVENT,
1211 	KFD_CRIU_OBJECT_TYPE_SVM_RANGE,
1212 };
1213 
1214 struct kfd_criu_svm_range_priv_data {
1215 	uint32_t object_type;
1216 	uint64_t start_addr;
1217 	uint64_t size;
1218 	/* Variable length array of attributes */
1219 	struct kfd_ioctl_svm_attribute attrs[];
1220 };
1221 
1222 struct kfd_criu_queue_priv_data {
1223 	uint32_t object_type;
1224 	uint64_t q_address;
1225 	uint64_t q_size;
1226 	uint64_t read_ptr_addr;
1227 	uint64_t write_ptr_addr;
1228 	uint64_t doorbell_off;
1229 	uint64_t eop_ring_buffer_address;
1230 	uint64_t ctx_save_restore_area_address;
1231 	uint32_t gpu_id;
1232 	uint32_t type;
1233 	uint32_t format;
1234 	uint32_t q_id;
1235 	uint32_t priority;
1236 	uint32_t q_percent;
1237 	uint32_t doorbell_id;
1238 	uint32_t gws;
1239 	uint32_t sdma_id;
1240 	uint32_t eop_ring_buffer_size;
1241 	uint32_t ctx_save_restore_area_size;
1242 	uint32_t ctl_stack_size;
1243 	uint32_t mqd_size;
1244 };
1245 
1246 struct kfd_criu_event_priv_data {
1247 	uint32_t object_type;
1248 	uint64_t user_handle;
1249 	uint32_t event_id;
1250 	uint32_t auto_reset;
1251 	uint32_t type;
1252 	uint32_t signaled;
1253 
1254 	union {
1255 		struct kfd_hsa_memory_exception_data memory_exception_data;
1256 		struct kfd_hsa_hw_exception_data hw_exception_data;
1257 	};
1258 };
1259 
1260 int kfd_process_get_queue_info(struct kfd_process *p,
1261 			       uint32_t *num_queues,
1262 			       uint64_t *priv_data_sizes);
1263 
1264 int kfd_criu_checkpoint_queues(struct kfd_process *p,
1265 			 uint8_t __user *user_priv_data,
1266 			 uint64_t *priv_data_offset);
1267 
1268 int kfd_criu_restore_queue(struct kfd_process *p,
1269 			   uint8_t __user *user_priv_data,
1270 			   uint64_t *priv_data_offset,
1271 			   uint64_t max_priv_data_size);
1272 
1273 int kfd_criu_checkpoint_events(struct kfd_process *p,
1274 			 uint8_t __user *user_priv_data,
1275 			 uint64_t *priv_data_offset);
1276 
1277 int kfd_criu_restore_event(struct file *devkfd,
1278 			   struct kfd_process *p,
1279 			   uint8_t __user *user_priv_data,
1280 			   uint64_t *priv_data_offset,
1281 			   uint64_t max_priv_data_size);
1282 /* CRIU - End */
1283 
1284 /* Queue Context Management */
1285 int init_queue(struct queue **q, const struct queue_properties *properties);
1286 void uninit_queue(struct queue *q);
1287 void print_queue_properties(struct queue_properties *q);
1288 void print_queue(struct queue *q);
1289 int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_bo **pbo,
1290 			 u64 expected_size);
1291 int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);
1292 int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);
1293 
1294 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
1295 		struct kfd_node *dev);
1296 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
1297 		struct kfd_node *dev);
1298 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
1299 		struct kfd_node *dev);
1300 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
1301 		struct kfd_node *dev);
1302 struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type,
1303 		struct kfd_node *dev);
1304 struct mqd_manager *mqd_manager_init_v12(enum KFD_MQD_TYPE type,
1305 		struct kfd_node *dev);
1306 struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev);
1307 void device_queue_manager_uninit(struct device_queue_manager *dqm);
1308 struct kernel_queue *kernel_queue_init(struct kfd_node *dev,
1309 					enum kfd_queue_type type);
1310 void kernel_queue_uninit(struct kernel_queue *kq);
1311 int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid);
1312 
1313 /* Process Queue Manager */
1314 struct process_queue_node {
1315 	struct queue *q;
1316 	struct kernel_queue *kq;
1317 	struct list_head process_queue_list;
1318 };
1319 
1320 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
1321 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
1322 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
1323 void pqm_uninit(struct process_queue_manager *pqm);
1324 int pqm_create_queue(struct process_queue_manager *pqm,
1325 			    struct kfd_node *dev,
1326 			    struct file *f,
1327 			    struct queue_properties *properties,
1328 			    unsigned int *qid,
1329 			    const struct kfd_criu_queue_priv_data *q_data,
1330 			    const void *restore_mqd,
1331 			    const void *restore_ctl_stack,
1332 			    uint32_t *p_doorbell_offset_in_process);
1333 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
1334 int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid,
1335 			struct queue_properties *p);
1336 int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid,
1337 			struct mqd_update_info *minfo);
1338 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
1339 			void *gws);
1340 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
1341 						unsigned int qid);
1342 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1343 						unsigned int qid);
1344 int pqm_get_wave_state(struct process_queue_manager *pqm,
1345 		       unsigned int qid,
1346 		       void __user *ctl_stack,
1347 		       u32 *ctl_stack_used_size,
1348 		       u32 *save_area_used_size);
1349 int pqm_get_queue_snapshot(struct process_queue_manager *pqm,
1350 			   uint64_t exception_clear_mask,
1351 			   void __user *buf,
1352 			   int *num_qss_entries,
1353 			   uint32_t *entry_size);
1354 
1355 int amdkfd_fence_wait_timeout(struct device_queue_manager *dqm,
1356 			      uint64_t fence_value,
1357 			      unsigned int timeout_ms);
1358 
1359 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
1360 				  unsigned int qid,
1361 				  u32 *mqd_size,
1362 				  u32 *ctl_stack_size);
1363 /* Packet Manager */
1364 
1365 #define KFD_FENCE_COMPLETED (100)
1366 #define KFD_FENCE_INIT   (10)
1367 
1368 struct packet_manager {
1369 	struct device_queue_manager *dqm;
1370 	struct kernel_queue *priv_queue;
1371 	struct mutex lock;
1372 	bool allocated;
1373 	struct kfd_mem_obj *ib_buffer_obj;
1374 	unsigned int ib_size_bytes;
1375 	bool is_over_subscription;
1376 
1377 	const struct packet_manager_funcs *pmf;
1378 };
1379 
1380 struct packet_manager_funcs {
1381 	/* Support ASIC-specific packet formats for PM4 packets */
1382 	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
1383 			struct qcm_process_device *qpd);
1384 	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
1385 			uint64_t ib, size_t ib_size_in_dwords, bool chain);
1386 	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
1387 			struct scheduling_resources *res);
1388 	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
1389 			struct queue *q, bool is_static);
1390 	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
1391 			enum kfd_unmap_queues_filter mode,
1392 			uint32_t filter_param, bool reset);
1393 	int (*set_grace_period)(struct packet_manager *pm, uint32_t *buffer,
1394 			uint32_t grace_period);
1395 	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
1396 			uint64_t fence_address,	uint64_t fence_value);
1397 	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
1398 
1399 	/* Packet sizes */
1400 	int map_process_size;
1401 	int runlist_size;
1402 	int set_resources_size;
1403 	int map_queues_size;
1404 	int unmap_queues_size;
1405 	int set_grace_period_size;
1406 	int query_status_size;
1407 	int release_mem_size;
1408 };
1409 
1410 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
1411 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
1412 extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;
1413 
1414 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
1415 void pm_uninit(struct packet_manager *pm);
1416 int pm_send_set_resources(struct packet_manager *pm,
1417 				struct scheduling_resources *res);
1418 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
1419 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
1420 				uint64_t fence_value);
1421 
1422 int pm_send_unmap_queue(struct packet_manager *pm,
1423 			enum kfd_unmap_queues_filter mode,
1424 			uint32_t filter_param, bool reset);
1425 
1426 void pm_release_ib(struct packet_manager *pm);
1427 
1428 int pm_update_grace_period(struct packet_manager *pm, uint32_t grace_period);
1429 
1430 /* Following PM funcs can be shared among VI and AI */
1431 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
1432 
1433 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
1434 
1435 /* Events */
1436 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1437 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1438 extern const struct kfd_event_interrupt_class event_interrupt_class_v9_4_3;
1439 extern const struct kfd_event_interrupt_class event_interrupt_class_v10;
1440 extern const struct kfd_event_interrupt_class event_interrupt_class_v11;
1441 
1442 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1443 
1444 int kfd_event_init_process(struct kfd_process *p);
1445 void kfd_event_free_process(struct kfd_process *p);
1446 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1447 int kfd_wait_on_events(struct kfd_process *p,
1448 		       uint32_t num_events, void __user *data,
1449 		       bool all, uint32_t *user_timeout_ms,
1450 		       uint32_t *wait_result);
1451 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
1452 				uint32_t valid_id_bits);
1453 void kfd_signal_hw_exception_event(u32 pasid);
1454 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1455 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1456 int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset);
1457 
1458 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1459 		     uint32_t event_type, bool auto_reset, uint32_t node_id,
1460 		     uint32_t *event_id, uint32_t *event_trigger_data,
1461 		     uint64_t *event_page_offset, uint32_t *event_slot_index);
1462 
1463 int kfd_get_num_events(struct kfd_process *p);
1464 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1465 
1466 void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid,
1467 				struct kfd_vm_fault_info *info,
1468 				struct kfd_hsa_memory_exception_data *data);
1469 
1470 void kfd_signal_reset_event(struct kfd_node *dev);
1471 
1472 void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid);
1473 
1474 static inline void kfd_flush_tlb(struct kfd_process_device *pdd,
1475 				 enum TLB_FLUSH_TYPE type)
1476 {
1477 	struct amdgpu_device *adev = pdd->dev->adev;
1478 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1479 
1480 	amdgpu_vm_flush_compute_tlb(adev, vm, type, pdd->dev->xcc_mask);
1481 }
1482 
1483 static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev)
1484 {
1485 	return KFD_GC_VERSION(dev) >= IP_VERSION(9, 4, 2) ||
1486 	       (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) ||
1487 	       KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0);
1488 }
1489 
1490 int kfd_send_exception_to_runtime(struct kfd_process *p,
1491 				unsigned int queue_id,
1492 				uint64_t error_reason);
1493 bool kfd_is_locked(void);
1494 
1495 /* Compute profile */
1496 void kfd_inc_compute_active(struct kfd_node *dev);
1497 void kfd_dec_compute_active(struct kfd_node *dev);
1498 
1499 /* Cgroup Support */
1500 /* Check with device cgroup if @kfd device is accessible */
1501 static inline int kfd_devcgroup_check_permission(struct kfd_node *node)
1502 {
1503 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1504 	struct drm_device *ddev;
1505 
1506 	if (node->xcp)
1507 		ddev = node->xcp->ddev;
1508 	else
1509 		ddev = adev_to_drm(node->adev);
1510 
1511 	return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1512 					  ddev->render->index,
1513 					  DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1514 #else
1515 	return 0;
1516 #endif
1517 }
1518 
1519 static inline bool kfd_is_first_node(struct kfd_node *node)
1520 {
1521 	return (node == node->kfd->nodes[0]);
1522 }
1523 
1524 /* Debugfs */
1525 #if defined(CONFIG_DEBUG_FS)
1526 
1527 void kfd_debugfs_init(void);
1528 void kfd_debugfs_fini(void);
1529 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1530 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1531 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1532 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1533 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1534 int pm_debugfs_runlist(struct seq_file *m, void *data);
1535 
1536 int kfd_debugfs_hang_hws(struct kfd_node *dev);
1537 int pm_debugfs_hang_hws(struct packet_manager *pm);
1538 int dqm_debugfs_hang_hws(struct device_queue_manager *dqm);
1539 
1540 #else
1541 
1542 static inline void kfd_debugfs_init(void) {}
1543 static inline void kfd_debugfs_fini(void) {}
1544 
1545 #endif
1546 
1547 #endif
1548