xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_priv.h (revision 9fd2da71c301184d98fe37674ca8d017d1ce6600)
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(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 /*  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 	DECLARE_BITMAP(reserved_sdma_queues_bitmap, KFD_MAX_SDMA_QUEUES);
245 };
246 
247 unsigned int kfd_get_num_sdma_engines(struct kfd_node *kdev);
248 unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *kdev);
249 
250 struct kfd_mem_obj {
251 	uint32_t range_start;
252 	uint32_t range_end;
253 	uint64_t gpu_addr;
254 	uint32_t *cpu_ptr;
255 	void *gtt_mem;
256 };
257 
258 struct kfd_vmid_info {
259 	uint32_t first_vmid_kfd;
260 	uint32_t last_vmid_kfd;
261 	uint32_t vmid_num_kfd;
262 };
263 
264 #define MAX_KFD_NODES	8
265 
266 struct kfd_dev;
267 
268 struct kfd_node {
269 	unsigned int node_id;
270 	struct amdgpu_device *adev;     /* Duplicated here along with keeping
271 					 * a copy in kfd_dev to save a hop
272 					 */
273 	const struct kfd2kgd_calls *kfd2kgd; /* Duplicated here along with
274 					      * keeping a copy in kfd_dev to
275 					      * save a hop
276 					      */
277 	struct kfd_vmid_info vm_info;
278 	unsigned int id;                /* topology stub index */
279 	uint32_t xcc_mask; /* Instance mask of XCCs present */
280 	struct amdgpu_xcp *xcp;
281 
282 	/* Interrupts */
283 	struct kfifo ih_fifo;
284 	struct work_struct interrupt_work;
285 	spinlock_t interrupt_lock;
286 
287 	/*
288 	 * Interrupts of interest to KFD are copied
289 	 * from the HW ring into a SW ring.
290 	 */
291 	bool interrupts_active;
292 	uint32_t interrupt_bitmap; /* Only used for GFX 9.4.3 */
293 
294 	/* QCM Device instance */
295 	struct device_queue_manager *dqm;
296 
297 	/* Global GWS resource shared between processes */
298 	void *gws;
299 
300 	/* Clients watching SMI events */
301 	struct list_head smi_clients;
302 	spinlock_t smi_lock;
303 	uint32_t reset_seq_num;
304 
305 	/* SRAM ECC flag */
306 	atomic_t sram_ecc_flag;
307 
308 	/*spm process id */
309 	unsigned int spm_pasid;
310 
311 	/* Maximum process number mapped to HW scheduler */
312 	unsigned int max_proc_per_quantum;
313 
314 	unsigned int compute_vmid_bitmap;
315 
316 	struct kfd_local_mem_info local_mem_info;
317 
318 	struct kfd_dev *kfd;
319 
320 	/* Track per device allocated watch points */
321 	uint32_t alloc_watch_ids;
322 	spinlock_t watch_points_lock;
323 };
324 
325 struct kfd_dev {
326 	struct amdgpu_device *adev;
327 
328 	struct kfd_device_info device_info;
329 
330 	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
331 					   * page used by kernel queue
332 					   */
333 
334 	struct kgd2kfd_shared_resources shared_resources;
335 
336 	const struct kfd2kgd_calls *kfd2kgd;
337 	struct mutex doorbell_mutex;
338 
339 	void *gtt_mem;
340 	uint64_t gtt_start_gpu_addr;
341 	void *gtt_start_cpu_ptr;
342 	void *gtt_sa_bitmap;
343 	struct mutex gtt_sa_lock;
344 	unsigned int gtt_sa_chunk_size;
345 	unsigned int gtt_sa_num_of_chunks;
346 
347 	bool init_complete;
348 
349 	/* Firmware versions */
350 	uint16_t mec_fw_version;
351 	uint16_t mec2_fw_version;
352 	uint16_t sdma_fw_version;
353 
354 	/* CWSR */
355 	bool cwsr_enabled;
356 	const void *cwsr_isa;
357 	unsigned int cwsr_isa_size;
358 
359 	/* xGMI */
360 	uint64_t hive_id;
361 
362 	bool pci_atomic_requested;
363 
364 	/* Compute Profile ref. count */
365 	atomic_t compute_profile;
366 
367 	struct ida doorbell_ida;
368 	unsigned int max_doorbell_slices;
369 
370 	int noretry;
371 
372 	struct kfd_node *nodes[MAX_KFD_NODES];
373 	unsigned int num_nodes;
374 
375 	struct workqueue_struct *ih_wq;
376 
377 	/* Kernel doorbells for KFD device */
378 	struct amdgpu_bo *doorbells;
379 
380 	/* bitmap for dynamic doorbell allocation from doorbell object */
381 	unsigned long *doorbell_bitmap;
382 
383 	/* for dynamic partitioning */
384 	int kfd_dev_lock;
385 };
386 
387 enum kfd_mempool {
388 	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
389 	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
390 	KFD_MEMPOOL_FRAMEBUFFER = 3,
391 };
392 
393 /* Character device interface */
394 int kfd_chardev_init(void);
395 void kfd_chardev_exit(void);
396 
397 /**
398  * enum kfd_unmap_queues_filter - Enum for queue filters.
399  *
400  * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
401  *						running queues list.
402  *
403  * @KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: Preempts all non-static queues
404  *						in the run list.
405  *
406  * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
407  *						specific process.
408  *
409  */
410 enum kfd_unmap_queues_filter {
411 	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES = 1,
412 	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES = 2,
413 	KFD_UNMAP_QUEUES_FILTER_BY_PASID = 3
414 };
415 
416 /**
417  * enum kfd_queue_type - Enum for various queue types.
418  *
419  * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
420  *
421  * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
422  *
423  * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
424  *
425  * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
426  *
427  * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
428  *
429  * @KFD_QUEUE_TYPE_SDMA_BY_ENG_ID:  SDMA user mode queue with target SDMA engine ID.
430  */
431 enum kfd_queue_type  {
432 	KFD_QUEUE_TYPE_COMPUTE,
433 	KFD_QUEUE_TYPE_SDMA,
434 	KFD_QUEUE_TYPE_HIQ,
435 	KFD_QUEUE_TYPE_DIQ,
436 	KFD_QUEUE_TYPE_SDMA_XGMI,
437 	KFD_QUEUE_TYPE_SDMA_BY_ENG_ID
438 };
439 
440 enum kfd_queue_format {
441 	KFD_QUEUE_FORMAT_PM4,
442 	KFD_QUEUE_FORMAT_AQL
443 };
444 
445 enum KFD_QUEUE_PRIORITY {
446 	KFD_QUEUE_PRIORITY_MINIMUM = 0,
447 	KFD_QUEUE_PRIORITY_MAXIMUM = 15
448 };
449 
450 /**
451  * struct queue_properties
452  *
453  * @type: The queue type.
454  *
455  * @queue_id: Queue identifier.
456  *
457  * @queue_address: Queue ring buffer address.
458  *
459  * @queue_size: Queue ring buffer size.
460  *
461  * @priority: Defines the queue priority relative to other queues in the
462  * process.
463  * This is just an indication and HW scheduling may override the priority as
464  * necessary while keeping the relative prioritization.
465  * the priority granularity is from 0 to f which f is the highest priority.
466  * currently all queues are initialized with the highest priority.
467  *
468  * @queue_percent: This field is partially implemented and currently a zero in
469  * this field defines that the queue is non active.
470  *
471  * @read_ptr: User space address which points to the number of dwords the
472  * cp read from the ring buffer. This field updates automatically by the H/W.
473  *
474  * @write_ptr: Defines the number of dwords written to the ring buffer.
475  *
476  * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
477  * buffer. This field should be similar to write_ptr and the user should
478  * update this field after updating the write_ptr.
479  *
480  * @doorbell_off: The doorbell offset in the doorbell pci-bar.
481  *
482  * @is_interop: Defines if this is a interop queue. Interop queue means that
483  * the queue can access both graphics and compute resources.
484  *
485  * @is_evicted: Defines if the queue is evicted. Only active queues
486  * are evicted, rendering them inactive.
487  *
488  * @is_active: Defines if the queue is active or not. @is_active and
489  * @is_evicted are protected by the DQM lock.
490  *
491  * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
492  * @is_gws should be protected by the DQM lock, since changing it can yield the
493  * possibility of updating DQM state on number of GWS queues.
494  *
495  * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
496  * of the queue.
497  *
498  * This structure represents the queue properties for each queue no matter if
499  * it's user mode or kernel mode queue.
500  *
501  */
502 
503 struct queue_properties {
504 	enum kfd_queue_type type;
505 	enum kfd_queue_format format;
506 	unsigned int queue_id;
507 	uint64_t queue_address;
508 	uint64_t  queue_size;
509 	uint32_t priority;
510 	uint32_t queue_percent;
511 	void __user *read_ptr;
512 	void __user *write_ptr;
513 	void __iomem *doorbell_ptr;
514 	uint32_t doorbell_off;
515 	bool is_interop;
516 	bool is_evicted;
517 	bool is_suspended;
518 	bool is_being_destroyed;
519 	bool is_active;
520 	bool is_gws;
521 	uint32_t pm4_target_xcc;
522 	bool is_dbg_wa;
523 	bool is_user_cu_masked;
524 	/* Not relevant for user mode queues in cp scheduling */
525 	unsigned int vmid;
526 	/* Relevant only for sdma queues*/
527 	uint32_t sdma_engine_id;
528 	uint32_t sdma_queue_id;
529 	uint32_t sdma_vm_addr;
530 	/* Relevant only for VI */
531 	uint64_t eop_ring_buffer_address;
532 	uint32_t eop_ring_buffer_size;
533 	uint64_t ctx_save_restore_area_address;
534 	uint32_t ctx_save_restore_area_size;
535 	uint32_t ctl_stack_size;
536 	uint64_t tba_addr;
537 	uint64_t tma_addr;
538 	uint64_t exception_status;
539 
540 	struct amdgpu_bo *wptr_bo;
541 	struct amdgpu_bo *rptr_bo;
542 	struct amdgpu_bo *ring_bo;
543 	struct amdgpu_bo *eop_buf_bo;
544 	struct amdgpu_bo *cwsr_bo;
545 };
546 
547 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\
548 			    (q).queue_address != 0 &&	\
549 			    (q).queue_percent > 0 &&	\
550 			    !(q).is_evicted &&		\
551 			    !(q).is_suspended)
552 
553 enum mqd_update_flag {
554 	UPDATE_FLAG_DBG_WA_ENABLE = 1,
555 	UPDATE_FLAG_DBG_WA_DISABLE = 2,
556 	UPDATE_FLAG_IS_GWS = 4, /* quirk for gfx9 IP */
557 };
558 
559 struct mqd_update_info {
560 	union {
561 		struct {
562 			uint32_t count; /* Must be a multiple of 32 */
563 			uint32_t *ptr;
564 		} cu_mask;
565 	};
566 	enum mqd_update_flag update_flag;
567 };
568 
569 /**
570  * struct queue
571  *
572  * @list: Queue linked list.
573  *
574  * @mqd: The queue MQD (memory queue descriptor).
575  *
576  * @mqd_mem_obj: The MQD local gpu memory object.
577  *
578  * @gart_mqd_addr: The MQD gart mc address.
579  *
580  * @properties: The queue properties.
581  *
582  * @mec: Used only in no cp scheduling mode and identifies to micro engine id
583  *	 that the queue should be executed on.
584  *
585  * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
586  *	  id.
587  *
588  * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
589  *
590  * @process: The kfd process that created this queue.
591  *
592  * @device: The kfd device that created this queue.
593  *
594  * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
595  * otherwise.
596  *
597  * This structure represents user mode compute queues.
598  * It contains all the necessary data to handle such queues.
599  *
600  */
601 
602 struct queue {
603 	struct list_head list;
604 	void *mqd;
605 	struct kfd_mem_obj *mqd_mem_obj;
606 	uint64_t gart_mqd_addr;
607 	struct queue_properties properties;
608 
609 	uint32_t mec;
610 	uint32_t pipe;
611 	uint32_t queue;
612 
613 	unsigned int sdma_id;
614 	unsigned int doorbell_id;
615 
616 	struct kfd_process	*process;
617 	struct kfd_node		*device;
618 	void *gws;
619 
620 	/* procfs */
621 	struct kobject kobj;
622 
623 	void *gang_ctx_bo;
624 	uint64_t gang_ctx_gpu_addr;
625 	void *gang_ctx_cpu_ptr;
626 
627 	struct amdgpu_bo *wptr_bo_gart;
628 };
629 
630 enum KFD_MQD_TYPE {
631 	KFD_MQD_TYPE_HIQ = 0,		/* for hiq */
632 	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
633 	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
634 	KFD_MQD_TYPE_DIQ,		/* for diq */
635 	KFD_MQD_TYPE_MAX
636 };
637 
638 enum KFD_PIPE_PRIORITY {
639 	KFD_PIPE_PRIORITY_CS_LOW = 0,
640 	KFD_PIPE_PRIORITY_CS_MEDIUM,
641 	KFD_PIPE_PRIORITY_CS_HIGH
642 };
643 
644 struct scheduling_resources {
645 	unsigned int vmid_mask;
646 	enum kfd_queue_type type;
647 	uint64_t queue_mask;
648 	uint64_t gws_mask;
649 	uint32_t oac_mask;
650 	uint32_t gds_heap_base;
651 	uint32_t gds_heap_size;
652 };
653 
654 struct process_queue_manager {
655 	/* data */
656 	struct kfd_process	*process;
657 	struct list_head	queues;
658 	unsigned long		*queue_slot_bitmap;
659 };
660 
661 struct qcm_process_device {
662 	/* The Device Queue Manager that owns this data */
663 	struct device_queue_manager *dqm;
664 	struct process_queue_manager *pqm;
665 	/* Queues list */
666 	struct list_head queues_list;
667 	struct list_head priv_queue_list;
668 
669 	unsigned int queue_count;
670 	unsigned int vmid;
671 	bool is_debug;
672 	unsigned int evicted; /* eviction counter, 0=active */
673 
674 	/* This flag tells if we should reset all wavefronts on
675 	 * process termination
676 	 */
677 	bool reset_wavefronts;
678 
679 	/* This flag tells us if this process has a GWS-capable
680 	 * queue that will be mapped into the runlist. It's
681 	 * possible to request a GWS BO, but not have the queue
682 	 * currently mapped, and this changes how the MAP_PROCESS
683 	 * PM4 packet is configured.
684 	 */
685 	bool mapped_gws_queue;
686 
687 	/* All the memory management data should be here too */
688 	uint64_t gds_context_area;
689 	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
690 	uint64_t page_table_base;
691 	uint32_t sh_mem_config;
692 	uint32_t sh_mem_bases;
693 	uint32_t sh_mem_ape1_base;
694 	uint32_t sh_mem_ape1_limit;
695 	uint32_t gds_size;
696 	uint32_t num_gws;
697 	uint32_t num_oac;
698 	uint32_t sh_hidden_private_base;
699 
700 	/* CWSR memory */
701 	struct kgd_mem *cwsr_mem;
702 	void *cwsr_kaddr;
703 	uint64_t cwsr_base;
704 	uint64_t tba_addr;
705 	uint64_t tma_addr;
706 
707 	/* IB memory */
708 	struct kgd_mem *ib_mem;
709 	uint64_t ib_base;
710 	void *ib_kaddr;
711 
712 	/* doorbells for kfd process */
713 	struct amdgpu_bo *proc_doorbells;
714 
715 	/* bitmap for dynamic doorbell allocation from the bo */
716 	unsigned long *doorbell_bitmap;
717 };
718 
719 /* KFD Memory Eviction */
720 
721 /* Approx. wait time before attempting to restore evicted BOs */
722 #define PROCESS_RESTORE_TIME_MS 100
723 /* Approx. back off time if restore fails due to lack of memory */
724 #define PROCESS_BACK_OFF_TIME_MS 100
725 /* Approx. time before evicting the process again */
726 #define PROCESS_ACTIVE_TIME_MS 10
727 
728 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
729  * idr_handle in the least significant 4 bytes
730  */
731 #define MAKE_HANDLE(gpu_id, idr_handle) \
732 	(((uint64_t)(gpu_id) << 32) + idr_handle)
733 #define GET_GPU_ID(handle) (handle >> 32)
734 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
735 
736 enum kfd_pdd_bound {
737 	PDD_UNBOUND = 0,
738 	PDD_BOUND,
739 	PDD_BOUND_SUSPENDED,
740 };
741 
742 #define MAX_SYSFS_FILENAME_LEN 15
743 
744 /*
745  * SDMA counter runs at 100MHz frequency.
746  * We display SDMA activity in microsecond granularity in sysfs.
747  * As a result, the divisor is 100.
748  */
749 #define SDMA_ACTIVITY_DIVISOR  100
750 
751 /* Data that is per-process-per device. */
752 struct kfd_process_device {
753 	/* The device that owns this data. */
754 	struct kfd_node *dev;
755 
756 	/* The process that owns this kfd_process_device. */
757 	struct kfd_process *process;
758 
759 	/* per-process-per device QCM data structure */
760 	struct qcm_process_device qpd;
761 
762 	/*Apertures*/
763 	uint64_t lds_base;
764 	uint64_t lds_limit;
765 	uint64_t gpuvm_base;
766 	uint64_t gpuvm_limit;
767 	uint64_t scratch_base;
768 	uint64_t scratch_limit;
769 
770 	/* VM context for GPUVM allocations */
771 	struct file *drm_file;
772 	void *drm_priv;
773 
774 	/* GPUVM allocations storage */
775 	struct idr alloc_idr;
776 
777 	/* Flag used to tell the pdd has dequeued from the dqm.
778 	 * This is used to prevent dev->dqm->ops.process_termination() from
779 	 * being called twice when it is already called in IOMMU callback
780 	 * function.
781 	 */
782 	bool already_dequeued;
783 	bool runtime_inuse;
784 
785 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
786 	enum kfd_pdd_bound bound;
787 
788 	/* VRAM usage */
789 	atomic64_t vram_usage;
790 	struct attribute attr_vram;
791 	char vram_filename[MAX_SYSFS_FILENAME_LEN];
792 
793 	/* SDMA activity tracking */
794 	uint64_t sdma_past_activity_counter;
795 	struct attribute attr_sdma;
796 	char sdma_filename[MAX_SYSFS_FILENAME_LEN];
797 
798 	/* Eviction activity tracking */
799 	uint64_t last_evict_timestamp;
800 	atomic64_t evict_duration_counter;
801 	struct attribute attr_evict;
802 
803 	struct kobject *kobj_stats;
804 
805 	/*
806 	 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
807 	 * that is associated with device encoded by "this" struct instance. The
808 	 * value reflects CU usage by all of the waves launched by this process
809 	 * on this device. A very important property of occupancy parameter is
810 	 * that its value is a snapshot of current use.
811 	 *
812 	 * Following is to be noted regarding how this parameter is reported:
813 	 *
814 	 *  The number of waves that a CU can launch is limited by couple of
815 	 *  parameters. These are encoded by struct amdgpu_cu_info instance
816 	 *  that is part of every device definition. For GFX9 devices this
817 	 *  translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
818 	 *  do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
819 	 *  when they do use scratch memory. This could change for future
820 	 *  devices and therefore this example should be considered as a guide.
821 	 *
822 	 *  All CU's of a device are available for the process. This may not be true
823 	 *  under certain conditions - e.g. CU masking.
824 	 *
825 	 *  Finally number of CU's that are occupied by a process is affected by both
826 	 *  number of CU's a device has along with number of other competing processes
827 	 */
828 	struct attribute attr_cu_occupancy;
829 
830 	/* sysfs counters for GPU retry fault and page migration tracking */
831 	struct kobject *kobj_counters;
832 	struct attribute attr_faults;
833 	struct attribute attr_page_in;
834 	struct attribute attr_page_out;
835 	uint64_t faults;
836 	uint64_t page_in;
837 	uint64_t page_out;
838 
839 	/* Exception code status*/
840 	uint64_t exception_status;
841 	void *vm_fault_exc_data;
842 	size_t vm_fault_exc_data_size;
843 
844 	/* Tracks debug per-vmid request settings */
845 	uint32_t spi_dbg_override;
846 	uint32_t spi_dbg_launch_mode;
847 	uint32_t watch_points[4];
848 	uint32_t alloc_watch_ids;
849 
850 	/*
851 	 * If this process has been checkpointed before, then the user
852 	 * application will use the original gpu_id on the
853 	 * checkpointed node to refer to this device.
854 	 */
855 	uint32_t user_gpu_id;
856 
857 	void *proc_ctx_bo;
858 	uint64_t proc_ctx_gpu_addr;
859 	void *proc_ctx_cpu_ptr;
860 
861 	/* Tracks queue reset status */
862 	bool has_reset_queue;
863 
864 	u32 pasid;
865 };
866 
867 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
868 
869 struct svm_range_list {
870 	struct mutex			lock;
871 	struct rb_root_cached		objects;
872 	struct list_head		list;
873 	struct work_struct		deferred_list_work;
874 	struct list_head		deferred_range_list;
875 	struct list_head                criu_svm_metadata_list;
876 	spinlock_t			deferred_list_lock;
877 	atomic_t			evicted_ranges;
878 	atomic_t			drain_pagefaults;
879 	struct delayed_work		restore_work;
880 	DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);
881 	struct task_struct		*faulting_task;
882 	/* check point ts decides if page fault recovery need be dropped */
883 	uint64_t			checkpoint_ts[MAX_GPU_INSTANCE];
884 
885 	/* Default granularity to use in buffer migration
886 	 * and restoration of backing memory while handling
887 	 * recoverable page faults
888 	 */
889 	uint8_t default_granularity;
890 };
891 
892 /* Process data */
893 struct kfd_process {
894 	/*
895 	 * kfd_process are stored in an mm_struct*->kfd_process*
896 	 * hash table (kfd_processes in kfd_process.c)
897 	 */
898 	struct hlist_node kfd_processes;
899 
900 	/*
901 	 * Opaque pointer to mm_struct. We don't hold a reference to
902 	 * it so it should never be dereferenced from here. This is
903 	 * only used for looking up processes by their mm.
904 	 */
905 	void *mm;
906 
907 	struct kref ref;
908 	struct work_struct release_work;
909 
910 	struct mutex mutex;
911 
912 	/*
913 	 * In any process, the thread that started main() is the lead
914 	 * thread and outlives the rest.
915 	 * It is here because amd_iommu_bind_pasid wants a task_struct.
916 	 * It can also be used for safely getting a reference to the
917 	 * mm_struct of the process.
918 	 */
919 	struct task_struct *lead_thread;
920 
921 	/* We want to receive a notification when the mm_struct is destroyed */
922 	struct mmu_notifier mmu_notifier;
923 
924 	/*
925 	 * Array of kfd_process_device pointers,
926 	 * one for each device the process is using.
927 	 */
928 	struct kfd_process_device *pdds[MAX_GPU_INSTANCE];
929 	uint32_t n_pdds;
930 
931 	struct process_queue_manager pqm;
932 
933 	/*Is the user space process 32 bit?*/
934 	bool is_32bit_user_mode;
935 
936 	/* Event-related data */
937 	struct mutex event_mutex;
938 	/* Event ID allocator and lookup */
939 	struct idr event_idr;
940 	/* Event page */
941 	u64 signal_handle;
942 	struct kfd_signal_page *signal_page;
943 	size_t signal_mapped_size;
944 	size_t signal_event_count;
945 	bool signal_event_limit_reached;
946 
947 	/* Information used for memory eviction */
948 	void *kgd_process_info;
949 	/* Eviction fence that is attached to all the BOs of this process. The
950 	 * fence will be triggered during eviction and new one will be created
951 	 * during restore
952 	 */
953 	struct dma_fence __rcu *ef;
954 
955 	/* Work items for evicting and restoring BOs */
956 	struct delayed_work eviction_work;
957 	struct delayed_work restore_work;
958 	/* seqno of the last scheduled eviction */
959 	unsigned int last_eviction_seqno;
960 	/* Approx. the last timestamp (in jiffies) when the process was
961 	 * restored after an eviction
962 	 */
963 	unsigned long last_restore_timestamp;
964 
965 	/* Indicates device process is debug attached with reserved vmid. */
966 	bool debug_trap_enabled;
967 
968 	/* per-process-per device debug event fd file */
969 	struct file *dbg_ev_file;
970 
971 	/* If the process is a kfd debugger, we need to know so we can clean
972 	 * up at exit time.  If a process enables debugging on itself, it does
973 	 * its own clean-up, so we don't set the flag here.  We track this by
974 	 * counting the number of processes this process is debugging.
975 	 */
976 	atomic_t debugged_process_count;
977 
978 	/* If the process is a debugged, this is the debugger process */
979 	struct kfd_process *debugger_process;
980 
981 	/* Kobj for our procfs */
982 	struct kobject *kobj;
983 	struct kobject *kobj_queues;
984 	struct attribute attr_pasid;
985 
986 	/* Keep track cwsr init */
987 	bool has_cwsr;
988 
989 	/* Exception code enable mask and status */
990 	uint64_t exception_enable_mask;
991 	uint64_t exception_status;
992 
993 	/* Used to drain stale interrupts */
994 	wait_queue_head_t wait_irq_drain;
995 	bool irq_drain_is_open;
996 
997 	/* shared virtual memory registered by this process */
998 	struct svm_range_list svms;
999 
1000 	bool xnack_enabled;
1001 
1002 	/* Work area for debugger event writer worker. */
1003 	struct work_struct debug_event_workarea;
1004 
1005 	/* Tracks debug per-vmid request for debug flags */
1006 	u32 dbg_flags;
1007 
1008 	atomic_t poison;
1009 	/* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */
1010 	bool queues_paused;
1011 
1012 	/* Tracks runtime enable status */
1013 	struct semaphore runtime_enable_sema;
1014 	bool is_runtime_retry;
1015 	struct kfd_runtime_info runtime_info;
1016 
1017 	/* if gpu page fault sent to KFD */
1018 	bool gpu_page_fault;
1019 };
1020 
1021 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
1022 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
1023 extern struct srcu_struct kfd_processes_srcu;
1024 
1025 /**
1026  * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
1027  *
1028  * @filep: pointer to file structure.
1029  * @p: amdkfd process pointer.
1030  * @data: pointer to arg that was copied from user.
1031  *
1032  * Return: returns ioctl completion code.
1033  */
1034 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
1035 				void *data);
1036 
1037 struct amdkfd_ioctl_desc {
1038 	unsigned int cmd;
1039 	int flags;
1040 	amdkfd_ioctl_t *func;
1041 	unsigned int cmd_drv;
1042 	const char *name;
1043 };
1044 bool kfd_dev_is_large_bar(struct kfd_node *dev);
1045 
1046 int kfd_process_create_wq(void);
1047 void kfd_process_destroy_wq(void);
1048 void kfd_cleanup_processes(void);
1049 struct kfd_process *kfd_create_process(struct task_struct *thread);
1050 struct kfd_process *kfd_get_process(const struct task_struct *task);
1051 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid,
1052 						 struct kfd_process_device **pdd);
1053 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
1054 
1055 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
1056 int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
1057 				uint32_t *gpuid, uint32_t *gpuidx);
1058 static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
1059 				uint32_t gpuidx, uint32_t *gpuid) {
1060 	return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
1061 }
1062 static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
1063 				struct kfd_process *p, uint32_t gpuidx) {
1064 	return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;
1065 }
1066 
1067 void kfd_unref_process(struct kfd_process *p);
1068 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger);
1069 int kfd_process_restore_queues(struct kfd_process *p);
1070 void kfd_suspend_all_processes(void);
1071 int kfd_resume_all_processes(void);
1072 
1073 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *process,
1074 							 uint32_t gpu_id);
1075 
1076 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id);
1077 
1078 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1079 			       struct file *drm_file);
1080 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1081 						struct kfd_process *p);
1082 struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1083 							struct kfd_process *p);
1084 struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1085 							struct kfd_process *p);
1086 
1087 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
1088 
1089 int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
1090 			  struct vm_area_struct *vma);
1091 
1092 /* KFD process API for creating and translating handles */
1093 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1094 					void *mem);
1095 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
1096 					int handle);
1097 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1098 					int handle);
1099 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid);
1100 
1101 /* PASIDs */
1102 int kfd_pasid_init(void);
1103 void kfd_pasid_exit(void);
1104 u32 kfd_pasid_alloc(void);
1105 void kfd_pasid_free(u32 pasid);
1106 
1107 /* Doorbells */
1108 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
1109 int kfd_doorbell_init(struct kfd_dev *kfd);
1110 void kfd_doorbell_fini(struct kfd_dev *kfd);
1111 int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process,
1112 		      struct vm_area_struct *vma);
1113 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
1114 					unsigned int *doorbell_off);
1115 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
1116 u32 read_kernel_doorbell(u32 __iomem *db);
1117 void write_kernel_doorbell(void __iomem *db, u32 value);
1118 void write_kernel_doorbell64(void __iomem *db, u64 value);
1119 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
1120 					struct kfd_process_device *pdd,
1121 					unsigned int doorbell_id);
1122 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
1123 int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
1124 				struct kfd_process_device *pdd);
1125 void kfd_free_process_doorbells(struct kfd_dev *kfd,
1126 				struct kfd_process_device *pdd);
1127 /* GTT Sub-Allocator */
1128 
1129 int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size,
1130 			struct kfd_mem_obj **mem_obj);
1131 
1132 int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj);
1133 
1134 extern struct device *kfd_device;
1135 
1136 /* KFD's procfs */
1137 void kfd_procfs_init(void);
1138 void kfd_procfs_shutdown(void);
1139 int kfd_procfs_add_queue(struct queue *q);
1140 void kfd_procfs_del_queue(struct queue *q);
1141 
1142 /* Topology */
1143 int kfd_topology_init(void);
1144 void kfd_topology_shutdown(void);
1145 int kfd_topology_add_device(struct kfd_node *gpu);
1146 int kfd_topology_remove_device(struct kfd_node *gpu);
1147 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
1148 						uint32_t proximity_domain);
1149 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
1150 						uint32_t proximity_domain);
1151 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
1152 struct kfd_node *kfd_device_by_id(uint32_t gpu_id);
1153 static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t node_id,
1154 					uint32_t vmid)
1155 {
1156 	return (node->interrupt_bitmap & (1 << node_id)) != 0 &&
1157 	       (node->compute_vmid_bitmap & (1 << vmid)) != 0;
1158 }
1159 static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev,
1160 					uint32_t node_id, uint32_t vmid) {
1161 	struct kfd_dev *dev = adev->kfd.dev;
1162 	uint32_t i;
1163 
1164 	if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) &&
1165 	    KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 4) &&
1166 	    KFD_GC_VERSION(dev) != IP_VERSION(9, 5, 0))
1167 		return dev->nodes[0];
1168 
1169 	for (i = 0; i < dev->num_nodes; i++)
1170 		if (kfd_irq_is_from_node(dev->nodes[i], node_id, vmid))
1171 			return dev->nodes[i];
1172 
1173 	return NULL;
1174 }
1175 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev);
1176 int kfd_numa_node_to_apic_id(int numa_node_id);
1177 
1178 /* Interrupts */
1179 #define	KFD_IRQ_FENCE_CLIENTID	0xff
1180 #define	KFD_IRQ_FENCE_SOURCEID	0xff
1181 #define	KFD_IRQ_IS_FENCE(client, source)				\
1182 				((client) == KFD_IRQ_FENCE_CLIENTID &&	\
1183 				(source) == KFD_IRQ_FENCE_SOURCEID)
1184 int kfd_interrupt_init(struct kfd_node *dev);
1185 void kfd_interrupt_exit(struct kfd_node *dev);
1186 bool enqueue_ih_ring_entry(struct kfd_node *kfd, const void *ih_ring_entry);
1187 bool interrupt_is_wanted(struct kfd_node *dev,
1188 				const uint32_t *ih_ring_entry,
1189 				uint32_t *patched_ihre, bool *flag);
1190 int kfd_process_drain_interrupts(struct kfd_process_device *pdd);
1191 void kfd_process_close_interrupt_drain(unsigned int pasid);
1192 
1193 /* amdkfd Apertures */
1194 int kfd_init_apertures(struct kfd_process *process);
1195 
1196 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1197 				  uint64_t tba_addr,
1198 				  uint64_t tma_addr);
1199 void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1200 				     bool enabled);
1201 
1202 /* CWSR initialization */
1203 int kfd_process_init_cwsr_apu(struct kfd_process *process, struct file *filep);
1204 
1205 /* CRIU */
1206 /*
1207  * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private
1208  * structures:
1209  * kfd_criu_process_priv_data
1210  * kfd_criu_device_priv_data
1211  * kfd_criu_bo_priv_data
1212  * kfd_criu_queue_priv_data
1213  * kfd_criu_event_priv_data
1214  * kfd_criu_svm_range_priv_data
1215  */
1216 
1217 #define KFD_CRIU_PRIV_VERSION 1
1218 
1219 struct kfd_criu_process_priv_data {
1220 	uint32_t version;
1221 	uint32_t xnack_mode;
1222 };
1223 
1224 struct kfd_criu_device_priv_data {
1225 	/* For future use */
1226 	uint64_t reserved;
1227 };
1228 
1229 struct kfd_criu_bo_priv_data {
1230 	uint64_t user_addr;
1231 	uint32_t idr_handle;
1232 	uint32_t mapped_gpuids[MAX_GPU_INSTANCE];
1233 };
1234 
1235 /*
1236  * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data,
1237  * kfd_criu_svm_range_priv_data is the object type
1238  */
1239 enum kfd_criu_object_type {
1240 	KFD_CRIU_OBJECT_TYPE_QUEUE,
1241 	KFD_CRIU_OBJECT_TYPE_EVENT,
1242 	KFD_CRIU_OBJECT_TYPE_SVM_RANGE,
1243 };
1244 
1245 struct kfd_criu_svm_range_priv_data {
1246 	uint32_t object_type;
1247 	uint64_t start_addr;
1248 	uint64_t size;
1249 	/* Variable length array of attributes */
1250 	struct kfd_ioctl_svm_attribute attrs[];
1251 };
1252 
1253 struct kfd_criu_queue_priv_data {
1254 	uint32_t object_type;
1255 	uint64_t q_address;
1256 	uint64_t q_size;
1257 	uint64_t read_ptr_addr;
1258 	uint64_t write_ptr_addr;
1259 	uint64_t doorbell_off;
1260 	uint64_t eop_ring_buffer_address;
1261 	uint64_t ctx_save_restore_area_address;
1262 	uint32_t gpu_id;
1263 	uint32_t type;
1264 	uint32_t format;
1265 	uint32_t q_id;
1266 	uint32_t priority;
1267 	uint32_t q_percent;
1268 	uint32_t doorbell_id;
1269 	uint32_t gws;
1270 	uint32_t sdma_id;
1271 	uint32_t eop_ring_buffer_size;
1272 	uint32_t ctx_save_restore_area_size;
1273 	uint32_t ctl_stack_size;
1274 	uint32_t mqd_size;
1275 };
1276 
1277 struct kfd_criu_event_priv_data {
1278 	uint32_t object_type;
1279 	uint64_t user_handle;
1280 	uint32_t event_id;
1281 	uint32_t auto_reset;
1282 	uint32_t type;
1283 	uint32_t signaled;
1284 
1285 	union {
1286 		struct kfd_hsa_memory_exception_data memory_exception_data;
1287 		struct kfd_hsa_hw_exception_data hw_exception_data;
1288 	};
1289 };
1290 
1291 int kfd_process_get_queue_info(struct kfd_process *p,
1292 			       uint32_t *num_queues,
1293 			       uint64_t *priv_data_sizes);
1294 
1295 int kfd_criu_checkpoint_queues(struct kfd_process *p,
1296 			 uint8_t __user *user_priv_data,
1297 			 uint64_t *priv_data_offset);
1298 
1299 int kfd_criu_restore_queue(struct kfd_process *p,
1300 			   uint8_t __user *user_priv_data,
1301 			   uint64_t *priv_data_offset,
1302 			   uint64_t max_priv_data_size);
1303 
1304 int kfd_criu_checkpoint_events(struct kfd_process *p,
1305 			 uint8_t __user *user_priv_data,
1306 			 uint64_t *priv_data_offset);
1307 
1308 int kfd_criu_restore_event(struct file *devkfd,
1309 			   struct kfd_process *p,
1310 			   uint8_t __user *user_priv_data,
1311 			   uint64_t *priv_data_offset,
1312 			   uint64_t max_priv_data_size);
1313 /* CRIU - End */
1314 
1315 /* Queue Context Management */
1316 int init_queue(struct queue **q, const struct queue_properties *properties);
1317 void uninit_queue(struct queue *q);
1318 void print_queue_properties(struct queue_properties *q);
1319 void print_queue(struct queue *q);
1320 int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_bo **pbo,
1321 			 u64 expected_size);
1322 void kfd_queue_buffer_put(struct amdgpu_bo **bo);
1323 int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);
1324 int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);
1325 void kfd_queue_unref_bo_va(struct amdgpu_vm *vm, struct amdgpu_bo **bo);
1326 int kfd_queue_unref_bo_vas(struct kfd_process_device *pdd,
1327 			   struct queue_properties *properties);
1328 void kfd_queue_ctx_save_restore_size(struct kfd_topology_device *dev);
1329 
1330 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
1331 		struct kfd_node *dev);
1332 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
1333 		struct kfd_node *dev);
1334 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
1335 		struct kfd_node *dev);
1336 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
1337 		struct kfd_node *dev);
1338 struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type,
1339 		struct kfd_node *dev);
1340 struct mqd_manager *mqd_manager_init_v12(enum KFD_MQD_TYPE type,
1341 		struct kfd_node *dev);
1342 struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev);
1343 void device_queue_manager_uninit(struct device_queue_manager *dqm);
1344 struct kernel_queue *kernel_queue_init(struct kfd_node *dev,
1345 					enum kfd_queue_type type);
1346 void kernel_queue_uninit(struct kernel_queue *kq);
1347 int kfd_evict_process_device(struct kfd_process_device *pdd);
1348 int kfd_dqm_suspend_bad_queue_mes(struct kfd_node *knode, u32 pasid, u32 doorbell_id);
1349 
1350 /* Process Queue Manager */
1351 struct process_queue_node {
1352 	struct queue *q;
1353 	struct kernel_queue *kq;
1354 	struct list_head process_queue_list;
1355 };
1356 
1357 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
1358 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
1359 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
1360 void pqm_uninit(struct process_queue_manager *pqm);
1361 int pqm_create_queue(struct process_queue_manager *pqm,
1362 			    struct kfd_node *dev,
1363 			    struct queue_properties *properties,
1364 			    unsigned int *qid,
1365 			    const struct kfd_criu_queue_priv_data *q_data,
1366 			    const void *restore_mqd,
1367 			    const void *restore_ctl_stack,
1368 			    uint32_t *p_doorbell_offset_in_process);
1369 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
1370 int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid,
1371 			struct queue_properties *p);
1372 int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid,
1373 			struct mqd_update_info *minfo);
1374 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
1375 			void *gws);
1376 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1377 						unsigned int qid);
1378 int pqm_get_wave_state(struct process_queue_manager *pqm,
1379 		       unsigned int qid,
1380 		       void __user *ctl_stack,
1381 		       u32 *ctl_stack_used_size,
1382 		       u32 *save_area_used_size);
1383 int pqm_get_queue_snapshot(struct process_queue_manager *pqm,
1384 			   uint64_t exception_clear_mask,
1385 			   void __user *buf,
1386 			   int *num_qss_entries,
1387 			   uint32_t *entry_size);
1388 
1389 int amdkfd_fence_wait_timeout(struct device_queue_manager *dqm,
1390 			      uint64_t fence_value,
1391 			      unsigned int timeout_ms);
1392 
1393 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
1394 				  unsigned int qid,
1395 				  u32 *mqd_size,
1396 				  u32 *ctl_stack_size);
1397 /* Packet Manager */
1398 
1399 #define KFD_FENCE_COMPLETED (100)
1400 #define KFD_FENCE_INIT   (10)
1401 
1402 /**
1403  * enum kfd_config_dequeue_wait_counts_cmd - Command for configuring
1404  *  dequeue wait counts.
1405  *
1406  * @KFD_DEQUEUE_WAIT_INIT: Set optimized dequeue wait counts for a
1407  *	certain ASICs. For these ASICs, this is default value used by RESET
1408  * @KFD_DEQUEUE_WAIT_RESET: Reset dequeue wait counts to the optimized value
1409  *	for certain ASICs. For others set it to default hardware reset value
1410  * @KFD_DEQUEUE_WAIT_SET_SCH_WAVE: Set context switch latency wait
1411  *
1412  */
1413 enum kfd_config_dequeue_wait_counts_cmd {
1414 	KFD_DEQUEUE_WAIT_INIT = 1,
1415 	KFD_DEQUEUE_WAIT_RESET = 2,
1416 	KFD_DEQUEUE_WAIT_SET_SCH_WAVE = 3
1417 };
1418 
1419 
1420 struct packet_manager {
1421 	struct device_queue_manager *dqm;
1422 	struct kernel_queue *priv_queue;
1423 	struct mutex lock;
1424 	bool allocated;
1425 	struct kfd_mem_obj *ib_buffer_obj;
1426 	unsigned int ib_size_bytes;
1427 	bool is_over_subscription;
1428 
1429 	const struct packet_manager_funcs *pmf;
1430 };
1431 
1432 struct packet_manager_funcs {
1433 	/* Support ASIC-specific packet formats for PM4 packets */
1434 	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
1435 			struct qcm_process_device *qpd);
1436 	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
1437 			uint64_t ib, size_t ib_size_in_dwords, bool chain);
1438 	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
1439 			struct scheduling_resources *res);
1440 	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
1441 			struct queue *q, bool is_static);
1442 	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
1443 			enum kfd_unmap_queues_filter mode,
1444 			uint32_t filter_param, bool reset);
1445 	int (*config_dequeue_wait_counts)(struct packet_manager *pm, uint32_t *buffer,
1446 			enum kfd_config_dequeue_wait_counts_cmd cmd, uint32_t value);
1447 	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
1448 			uint64_t fence_address,	uint64_t fence_value);
1449 	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
1450 
1451 	/* Packet sizes */
1452 	int map_process_size;
1453 	int runlist_size;
1454 	int set_resources_size;
1455 	int map_queues_size;
1456 	int unmap_queues_size;
1457 	int config_dequeue_wait_counts_size;
1458 	int query_status_size;
1459 	int release_mem_size;
1460 };
1461 
1462 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
1463 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
1464 extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;
1465 
1466 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
1467 void pm_uninit(struct packet_manager *pm);
1468 int pm_send_set_resources(struct packet_manager *pm,
1469 				struct scheduling_resources *res);
1470 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
1471 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
1472 				uint64_t fence_value);
1473 
1474 int pm_send_unmap_queue(struct packet_manager *pm,
1475 			enum kfd_unmap_queues_filter mode,
1476 			uint32_t filter_param, bool reset);
1477 
1478 void pm_release_ib(struct packet_manager *pm);
1479 
1480 int pm_config_dequeue_wait_counts(struct packet_manager *pm,
1481 			enum kfd_config_dequeue_wait_counts_cmd cmd,
1482 			uint32_t wait_counts_config);
1483 
1484 /* Following PM funcs can be shared among VI and AI */
1485 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
1486 
1487 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
1488 
1489 /* Events */
1490 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1491 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1492 extern const struct kfd_event_interrupt_class event_interrupt_class_v9_4_3;
1493 extern const struct kfd_event_interrupt_class event_interrupt_class_v10;
1494 extern const struct kfd_event_interrupt_class event_interrupt_class_v11;
1495 
1496 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1497 
1498 int kfd_event_init_process(struct kfd_process *p);
1499 void kfd_event_free_process(struct kfd_process *p);
1500 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1501 int kfd_wait_on_events(struct kfd_process *p,
1502 		       uint32_t num_events, void __user *data,
1503 		       bool all, uint32_t *user_timeout_ms,
1504 		       uint32_t *wait_result);
1505 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
1506 				uint32_t valid_id_bits);
1507 void kfd_signal_hw_exception_event(u32 pasid);
1508 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1509 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1510 int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset);
1511 
1512 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1513 		     uint32_t event_type, bool auto_reset, uint32_t node_id,
1514 		     uint32_t *event_id, uint32_t *event_trigger_data,
1515 		     uint64_t *event_page_offset, uint32_t *event_slot_index);
1516 
1517 int kfd_get_num_events(struct kfd_process *p);
1518 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1519 
1520 void kfd_signal_vm_fault_event_with_userptr(struct kfd_process *p, uint64_t gpu_va);
1521 
1522 void kfd_signal_vm_fault_event(struct kfd_process_device *pdd,
1523 				struct kfd_vm_fault_info *info,
1524 				struct kfd_hsa_memory_exception_data *data);
1525 
1526 void kfd_signal_reset_event(struct kfd_node *dev);
1527 
1528 void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid);
1529 
1530 static inline void kfd_flush_tlb(struct kfd_process_device *pdd,
1531 				 enum TLB_FLUSH_TYPE type)
1532 {
1533 	struct amdgpu_device *adev = pdd->dev->adev;
1534 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1535 
1536 	amdgpu_vm_flush_compute_tlb(adev, vm, type, pdd->dev->xcc_mask);
1537 }
1538 
1539 static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev)
1540 {
1541 	return KFD_GC_VERSION(dev) >= IP_VERSION(9, 4, 2) ||
1542 	       (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) ||
1543 	       KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0);
1544 }
1545 
1546 int kfd_send_exception_to_runtime(struct kfd_process *p,
1547 				unsigned int queue_id,
1548 				uint64_t error_reason);
1549 bool kfd_is_locked(struct kfd_dev *kfd);
1550 
1551 /* Compute profile */
1552 void kfd_inc_compute_active(struct kfd_node *dev);
1553 void kfd_dec_compute_active(struct kfd_node *dev);
1554 
1555 /* Cgroup Support */
1556 /* Check with device cgroup if @kfd device is accessible */
1557 static inline int kfd_devcgroup_check_permission(struct kfd_node *node)
1558 {
1559 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1560 	struct drm_device *ddev;
1561 
1562 	if (node->xcp)
1563 		ddev = node->xcp->ddev;
1564 	else
1565 		ddev = adev_to_drm(node->adev);
1566 
1567 	return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1568 					  ddev->render->index,
1569 					  DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1570 #else
1571 	return 0;
1572 #endif
1573 }
1574 
1575 static inline bool kfd_is_first_node(struct kfd_node *node)
1576 {
1577 	return (node == node->kfd->nodes[0]);
1578 }
1579 
1580 /* Debugfs */
1581 #if defined(CONFIG_DEBUG_FS)
1582 
1583 void kfd_debugfs_init(void);
1584 void kfd_debugfs_fini(void);
1585 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1586 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1587 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1588 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1589 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1590 int pm_debugfs_runlist(struct seq_file *m, void *data);
1591 
1592 int kfd_debugfs_hang_hws(struct kfd_node *dev);
1593 int pm_debugfs_hang_hws(struct packet_manager *pm);
1594 int dqm_debugfs_hang_hws(struct device_queue_manager *dqm);
1595 
1596 void kfd_debugfs_add_process(struct kfd_process *p);
1597 void kfd_debugfs_remove_process(struct kfd_process *p);
1598 
1599 #else
1600 
1601 static inline void kfd_debugfs_init(void) {}
1602 static inline void kfd_debugfs_fini(void) {}
1603 static inline void kfd_debugfs_add_process(struct kfd_process *p) {}
1604 static inline void kfd_debugfs_remove_process(struct kfd_process *p) {}
1605 
1606 #endif
1607 
1608 #endif
1609