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 page is the TBA used for the CWSR ISA code. The second
103 * page 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
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 /*
117 * 512 = 0x200
118 * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
119 * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
120 * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
121 * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
122 * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
123 */
124 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
125
126 /**
127 * enum kfd_ioctl_flags - KFD ioctl flags
128 * Various flags that can be set in &amdkfd_ioctl_desc.flags to control how
129 * userspace can use a given ioctl.
130 */
131 enum kfd_ioctl_flags {
132 /*
133 * @KFD_IOC_FLAG_CHECKPOINT_RESTORE:
134 * Certain KFD ioctls such as AMDKFD_IOC_CRIU_OP can potentially
135 * perform privileged operations and load arbitrary data into MQDs and
136 * eventually HQD registers when the queue is mapped by HWS. In order to
137 * prevent this we should perform additional security checks.
138 *
139 * This is equivalent to callers with the CHECKPOINT_RESTORE capability.
140 *
141 * Note: Since earlier versions of docker do not support CHECKPOINT_RESTORE,
142 * we also allow ioctls with SYS_ADMIN capability.
143 */
144 KFD_IOC_FLAG_CHECKPOINT_RESTORE = BIT(0),
145 };
146 /*
147 * Kernel module parameter to specify maximum number of supported queues per
148 * device
149 */
150 extern int max_num_of_queues_per_device;
151
152
153 /* Kernel module parameter to specify the scheduling policy */
154 extern int sched_policy;
155
156 /*
157 * Kernel module parameter to specify the maximum process
158 * number per HW scheduler
159 */
160 extern int hws_max_conc_proc;
161
162 extern int cwsr_enable;
163
164 /*
165 * Kernel module parameter to specify whether to send sigterm to HSA process on
166 * unhandled exception
167 */
168 extern int send_sigterm;
169
170 /*
171 * This kernel module is used to simulate large bar machine on non-large bar
172 * enabled machines.
173 */
174 extern int debug_largebar;
175
176 /*
177 * Ignore CRAT table during KFD initialization, can be used to work around
178 * broken CRAT tables on some AMD systems
179 */
180 extern int ignore_crat;
181
182 /* Set sh_mem_config.retry_disable on GFX v9 */
183 extern int amdgpu_noretry;
184
185 /* Halt if HWS hang is detected */
186 extern int halt_if_hws_hang;
187
188 /* Whether MEC FW support GWS barriers */
189 extern bool hws_gws_support;
190
191 /* Queue preemption timeout in ms */
192 extern int queue_preemption_timeout_ms;
193
194 /*
195 * Don't evict process queues on vm fault
196 */
197 extern int amdgpu_no_queue_eviction_on_vm_fault;
198
199 /* Enable eviction debug messages */
200 extern bool debug_evictions;
201
202 enum cache_policy {
203 cache_policy_coherent,
204 cache_policy_noncoherent
205 };
206
207 #define KFD_GC_VERSION(dev) ((dev)->adev->ip_versions[GC_HWIP][0])
208 #define KFD_IS_SOC15(dev) ((KFD_GC_VERSION(dev)) >= (IP_VERSION(9, 0, 1)))
209
210 struct kfd_event_interrupt_class {
211 bool (*interrupt_isr)(struct kfd_dev *dev,
212 const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
213 bool *patched_flag);
214 void (*interrupt_wq)(struct kfd_dev *dev,
215 const uint32_t *ih_ring_entry);
216 };
217
218 struct kfd_device_info {
219 uint32_t gfx_target_version;
220 const struct kfd_event_interrupt_class *event_interrupt_class;
221 unsigned int max_pasid_bits;
222 unsigned int max_no_of_hqd;
223 unsigned int doorbell_size;
224 size_t ih_ring_entry_size;
225 uint8_t num_of_watch_points;
226 uint16_t mqd_size_aligned;
227 bool supports_cwsr;
228 bool needs_iommu_device;
229 bool needs_pci_atomics;
230 uint32_t no_atomic_fw_version;
231 unsigned int num_sdma_queues_per_engine;
232 unsigned int num_reserved_sdma_queues_per_engine;
233 uint64_t reserved_sdma_queues_bitmap;
234 };
235
236 unsigned int kfd_get_num_sdma_engines(struct kfd_dev *kdev);
237 unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_dev *kdev);
238
239 struct kfd_mem_obj {
240 uint32_t range_start;
241 uint32_t range_end;
242 uint64_t gpu_addr;
243 uint32_t *cpu_ptr;
244 void *gtt_mem;
245 };
246
247 struct kfd_vmid_info {
248 uint32_t first_vmid_kfd;
249 uint32_t last_vmid_kfd;
250 uint32_t vmid_num_kfd;
251 };
252
253 struct kfd_dev {
254 struct amdgpu_device *adev;
255
256 struct kfd_device_info device_info;
257 struct pci_dev *pdev;
258 struct drm_device *ddev;
259
260 unsigned int id; /* topology stub index */
261
262 phys_addr_t doorbell_base; /* Start of actual doorbells used by
263 * KFD. It is aligned for mapping
264 * into user mode
265 */
266 size_t doorbell_base_dw_offset; /* Offset from the start of the PCI
267 * doorbell BAR to the first KFD
268 * doorbell in dwords. GFX reserves
269 * the segment before this offset.
270 */
271 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
272 * page used by kernel queue
273 */
274
275 struct kgd2kfd_shared_resources shared_resources;
276 struct kfd_vmid_info vm_info;
277 struct kfd_local_mem_info local_mem_info;
278
279 const struct kfd2kgd_calls *kfd2kgd;
280 struct mutex doorbell_mutex;
281 DECLARE_BITMAP(doorbell_available_index,
282 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
283
284 void *gtt_mem;
285 uint64_t gtt_start_gpu_addr;
286 void *gtt_start_cpu_ptr;
287 void *gtt_sa_bitmap;
288 struct mutex gtt_sa_lock;
289 unsigned int gtt_sa_chunk_size;
290 unsigned int gtt_sa_num_of_chunks;
291
292 /* Interrupts */
293 struct kfifo ih_fifo;
294 struct workqueue_struct *ih_wq;
295 struct work_struct interrupt_work;
296 spinlock_t interrupt_lock;
297
298 /* QCM Device instance */
299 struct device_queue_manager *dqm;
300
301 bool init_complete;
302 /*
303 * Interrupts of interest to KFD are copied
304 * from the HW ring into a SW ring.
305 */
306 bool interrupts_active;
307
308 /* Firmware versions */
309 uint16_t mec_fw_version;
310 uint16_t mec2_fw_version;
311 uint16_t sdma_fw_version;
312
313 /* Maximum process number mapped to HW scheduler */
314 unsigned int max_proc_per_quantum;
315
316 /* CWSR */
317 bool cwsr_enabled;
318 const void *cwsr_isa;
319 unsigned int cwsr_isa_size;
320
321 /* xGMI */
322 uint64_t hive_id;
323
324 bool pci_atomic_requested;
325
326 /* Use IOMMU v2 flag */
327 bool use_iommu_v2;
328
329 /* SRAM ECC flag */
330 atomic_t sram_ecc_flag;
331
332 /* Compute Profile ref. count */
333 atomic_t compute_profile;
334
335 /* Global GWS resource shared between processes */
336 void *gws;
337
338 /* Clients watching SMI events */
339 struct list_head smi_clients;
340 spinlock_t smi_lock;
341
342 uint32_t reset_seq_num;
343
344 struct ida doorbell_ida;
345 unsigned int max_doorbell_slices;
346
347 int noretry;
348
349 /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */
350 struct dev_pagemap pgmap;
351 };
352
353 enum kfd_mempool {
354 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
355 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
356 KFD_MEMPOOL_FRAMEBUFFER = 3,
357 };
358
359 /* Character device interface */
360 int kfd_chardev_init(void);
361 void kfd_chardev_exit(void);
362
363 /**
364 * enum kfd_unmap_queues_filter - Enum for queue filters.
365 *
366 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
367 * running queues list.
368 *
369 * @KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: Preempts all non-static queues
370 * in the run list.
371 *
372 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
373 * specific process.
374 *
375 */
376 enum kfd_unmap_queues_filter {
377 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES = 1,
378 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES = 2,
379 KFD_UNMAP_QUEUES_FILTER_BY_PASID = 3
380 };
381
382 /**
383 * enum kfd_queue_type - Enum for various queue types.
384 *
385 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
386 *
387 * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
388 *
389 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
390 *
391 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
392 *
393 * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
394 */
395 enum kfd_queue_type {
396 KFD_QUEUE_TYPE_COMPUTE,
397 KFD_QUEUE_TYPE_SDMA,
398 KFD_QUEUE_TYPE_HIQ,
399 KFD_QUEUE_TYPE_DIQ,
400 KFD_QUEUE_TYPE_SDMA_XGMI
401 };
402
403 enum kfd_queue_format {
404 KFD_QUEUE_FORMAT_PM4,
405 KFD_QUEUE_FORMAT_AQL
406 };
407
408 enum KFD_QUEUE_PRIORITY {
409 KFD_QUEUE_PRIORITY_MINIMUM = 0,
410 KFD_QUEUE_PRIORITY_MAXIMUM = 15
411 };
412
413 /**
414 * struct queue_properties
415 *
416 * @type: The queue type.
417 *
418 * @queue_id: Queue identifier.
419 *
420 * @queue_address: Queue ring buffer address.
421 *
422 * @queue_size: Queue ring buffer size.
423 *
424 * @priority: Defines the queue priority relative to other queues in the
425 * process.
426 * This is just an indication and HW scheduling may override the priority as
427 * necessary while keeping the relative prioritization.
428 * the priority granularity is from 0 to f which f is the highest priority.
429 * currently all queues are initialized with the highest priority.
430 *
431 * @queue_percent: This field is partially implemented and currently a zero in
432 * this field defines that the queue is non active.
433 *
434 * @read_ptr: User space address which points to the number of dwords the
435 * cp read from the ring buffer. This field updates automatically by the H/W.
436 *
437 * @write_ptr: Defines the number of dwords written to the ring buffer.
438 *
439 * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
440 * buffer. This field should be similar to write_ptr and the user should
441 * update this field after updating the write_ptr.
442 *
443 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
444 *
445 * @is_interop: Defines if this is a interop queue. Interop queue means that
446 * the queue can access both graphics and compute resources.
447 *
448 * @is_evicted: Defines if the queue is evicted. Only active queues
449 * are evicted, rendering them inactive.
450 *
451 * @is_active: Defines if the queue is active or not. @is_active and
452 * @is_evicted are protected by the DQM lock.
453 *
454 * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
455 * @is_gws should be protected by the DQM lock, since changing it can yield the
456 * possibility of updating DQM state on number of GWS queues.
457 *
458 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
459 * of the queue.
460 *
461 * This structure represents the queue properties for each queue no matter if
462 * it's user mode or kernel mode queue.
463 *
464 */
465
466 struct queue_properties {
467 enum kfd_queue_type type;
468 enum kfd_queue_format format;
469 unsigned int queue_id;
470 uint64_t queue_address;
471 uint64_t queue_size;
472 uint32_t priority;
473 uint32_t queue_percent;
474 uint32_t *read_ptr;
475 uint32_t *write_ptr;
476 void __iomem *doorbell_ptr;
477 uint32_t doorbell_off;
478 bool is_interop;
479 bool is_evicted;
480 bool is_active;
481 bool is_gws;
482 /* Not relevant for user mode queues in cp scheduling */
483 unsigned int vmid;
484 /* Relevant only for sdma queues*/
485 uint32_t sdma_engine_id;
486 uint32_t sdma_queue_id;
487 uint32_t sdma_vm_addr;
488 /* Relevant only for VI */
489 uint64_t eop_ring_buffer_address;
490 uint32_t eop_ring_buffer_size;
491 uint64_t ctx_save_restore_area_address;
492 uint32_t ctx_save_restore_area_size;
493 uint32_t ctl_stack_size;
494 uint64_t tba_addr;
495 uint64_t tma_addr;
496 };
497
498 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \
499 (q).queue_address != 0 && \
500 (q).queue_percent > 0 && \
501 !(q).is_evicted)
502
503 enum mqd_update_flag {
504 UPDATE_FLAG_CU_MASK = 0,
505 };
506
507 struct mqd_update_info {
508 union {
509 struct {
510 uint32_t count; /* Must be a multiple of 32 */
511 uint32_t *ptr;
512 } cu_mask;
513 };
514 enum mqd_update_flag update_flag;
515 };
516
517 /**
518 * struct queue
519 *
520 * @list: Queue linked list.
521 *
522 * @mqd: The queue MQD (memory queue descriptor).
523 *
524 * @mqd_mem_obj: The MQD local gpu memory object.
525 *
526 * @gart_mqd_addr: The MQD gart mc address.
527 *
528 * @properties: The queue properties.
529 *
530 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
531 * that the queue should be executed on.
532 *
533 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
534 * id.
535 *
536 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
537 *
538 * @process: The kfd process that created this queue.
539 *
540 * @device: The kfd device that created this queue.
541 *
542 * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
543 * otherwise.
544 *
545 * This structure represents user mode compute queues.
546 * It contains all the necessary data to handle such queues.
547 *
548 */
549
550 struct queue {
551 struct list_head list;
552 void *mqd;
553 struct kfd_mem_obj *mqd_mem_obj;
554 uint64_t gart_mqd_addr;
555 struct queue_properties properties;
556
557 uint32_t mec;
558 uint32_t pipe;
559 uint32_t queue;
560
561 unsigned int sdma_id;
562 unsigned int doorbell_id;
563
564 struct kfd_process *process;
565 struct kfd_dev *device;
566 void *gws;
567
568 /* procfs */
569 struct kobject kobj;
570
571 void *gang_ctx_bo;
572 uint64_t gang_ctx_gpu_addr;
573 void *gang_ctx_cpu_ptr;
574
575 struct amdgpu_bo *wptr_bo;
576 };
577
578 enum KFD_MQD_TYPE {
579 KFD_MQD_TYPE_HIQ = 0, /* for hiq */
580 KFD_MQD_TYPE_CP, /* for cp queues and diq */
581 KFD_MQD_TYPE_SDMA, /* for sdma queues */
582 KFD_MQD_TYPE_DIQ, /* for diq */
583 KFD_MQD_TYPE_MAX
584 };
585
586 enum KFD_PIPE_PRIORITY {
587 KFD_PIPE_PRIORITY_CS_LOW = 0,
588 KFD_PIPE_PRIORITY_CS_MEDIUM,
589 KFD_PIPE_PRIORITY_CS_HIGH
590 };
591
592 struct scheduling_resources {
593 unsigned int vmid_mask;
594 enum kfd_queue_type type;
595 uint64_t queue_mask;
596 uint64_t gws_mask;
597 uint32_t oac_mask;
598 uint32_t gds_heap_base;
599 uint32_t gds_heap_size;
600 };
601
602 struct process_queue_manager {
603 /* data */
604 struct kfd_process *process;
605 struct list_head queues;
606 unsigned long *queue_slot_bitmap;
607 };
608
609 struct qcm_process_device {
610 /* The Device Queue Manager that owns this data */
611 struct device_queue_manager *dqm;
612 struct process_queue_manager *pqm;
613 /* Queues list */
614 struct list_head queues_list;
615 struct list_head priv_queue_list;
616
617 unsigned int queue_count;
618 unsigned int vmid;
619 bool is_debug;
620 unsigned int evicted; /* eviction counter, 0=active */
621
622 /* This flag tells if we should reset all wavefronts on
623 * process termination
624 */
625 bool reset_wavefronts;
626
627 /* This flag tells us if this process has a GWS-capable
628 * queue that will be mapped into the runlist. It's
629 * possible to request a GWS BO, but not have the queue
630 * currently mapped, and this changes how the MAP_PROCESS
631 * PM4 packet is configured.
632 */
633 bool mapped_gws_queue;
634
635 /* All the memory management data should be here too */
636 uint64_t gds_context_area;
637 /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
638 uint64_t page_table_base;
639 uint32_t sh_mem_config;
640 uint32_t sh_mem_bases;
641 uint32_t sh_mem_ape1_base;
642 uint32_t sh_mem_ape1_limit;
643 uint32_t gds_size;
644 uint32_t num_gws;
645 uint32_t num_oac;
646 uint32_t sh_hidden_private_base;
647
648 /* CWSR memory */
649 struct kgd_mem *cwsr_mem;
650 void *cwsr_kaddr;
651 uint64_t cwsr_base;
652 uint64_t tba_addr;
653 uint64_t tma_addr;
654
655 /* IB memory */
656 struct kgd_mem *ib_mem;
657 uint64_t ib_base;
658 void *ib_kaddr;
659
660 /* doorbell resources per process per device */
661 unsigned long *doorbell_bitmap;
662 };
663
664 /* KFD Memory Eviction */
665
666 /* Approx. wait time before attempting to restore evicted BOs */
667 #define PROCESS_RESTORE_TIME_MS 100
668 /* Approx. back off time if restore fails due to lack of memory */
669 #define PROCESS_BACK_OFF_TIME_MS 100
670 /* Approx. time before evicting the process again */
671 #define PROCESS_ACTIVE_TIME_MS 10
672
673 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
674 * idr_handle in the least significant 4 bytes
675 */
676 #define MAKE_HANDLE(gpu_id, idr_handle) \
677 (((uint64_t)(gpu_id) << 32) + idr_handle)
678 #define GET_GPU_ID(handle) (handle >> 32)
679 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
680
681 enum kfd_pdd_bound {
682 PDD_UNBOUND = 0,
683 PDD_BOUND,
684 PDD_BOUND_SUSPENDED,
685 };
686
687 #define MAX_SYSFS_FILENAME_LEN 15
688
689 /*
690 * SDMA counter runs at 100MHz frequency.
691 * We display SDMA activity in microsecond granularity in sysfs.
692 * As a result, the divisor is 100.
693 */
694 #define SDMA_ACTIVITY_DIVISOR 100
695
696 /* Data that is per-process-per device. */
697 struct kfd_process_device {
698 /* The device that owns this data. */
699 struct kfd_dev *dev;
700
701 /* The process that owns this kfd_process_device. */
702 struct kfd_process *process;
703
704 /* per-process-per device QCM data structure */
705 struct qcm_process_device qpd;
706
707 /*Apertures*/
708 uint64_t lds_base;
709 uint64_t lds_limit;
710 uint64_t gpuvm_base;
711 uint64_t gpuvm_limit;
712 uint64_t scratch_base;
713 uint64_t scratch_limit;
714
715 /* VM context for GPUVM allocations */
716 struct file *drm_file;
717 void *drm_priv;
718 atomic64_t tlb_seq;
719
720 /* GPUVM allocations storage */
721 struct idr alloc_idr;
722
723 /* Flag used to tell the pdd has dequeued from the dqm.
724 * This is used to prevent dev->dqm->ops.process_termination() from
725 * being called twice when it is already called in IOMMU callback
726 * function.
727 */
728 bool already_dequeued;
729 bool runtime_inuse;
730
731 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
732 enum kfd_pdd_bound bound;
733
734 /* VRAM usage */
735 uint64_t vram_usage;
736 struct attribute attr_vram;
737 char vram_filename[MAX_SYSFS_FILENAME_LEN];
738
739 /* SDMA activity tracking */
740 uint64_t sdma_past_activity_counter;
741 struct attribute attr_sdma;
742 char sdma_filename[MAX_SYSFS_FILENAME_LEN];
743
744 /* Eviction activity tracking */
745 uint64_t last_evict_timestamp;
746 atomic64_t evict_duration_counter;
747 struct attribute attr_evict;
748
749 struct kobject *kobj_stats;
750 unsigned int doorbell_index;
751
752 /*
753 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
754 * that is associated with device encoded by "this" struct instance. The
755 * value reflects CU usage by all of the waves launched by this process
756 * on this device. A very important property of occupancy parameter is
757 * that its value is a snapshot of current use.
758 *
759 * Following is to be noted regarding how this parameter is reported:
760 *
761 * The number of waves that a CU can launch is limited by couple of
762 * parameters. These are encoded by struct amdgpu_cu_info instance
763 * that is part of every device definition. For GFX9 devices this
764 * translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
765 * do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
766 * when they do use scratch memory. This could change for future
767 * devices and therefore this example should be considered as a guide.
768 *
769 * All CU's of a device are available for the process. This may not be true
770 * under certain conditions - e.g. CU masking.
771 *
772 * Finally number of CU's that are occupied by a process is affected by both
773 * number of CU's a device has along with number of other competing processes
774 */
775 struct attribute attr_cu_occupancy;
776
777 /* sysfs counters for GPU retry fault and page migration tracking */
778 struct kobject *kobj_counters;
779 struct attribute attr_faults;
780 struct attribute attr_page_in;
781 struct attribute attr_page_out;
782 uint64_t faults;
783 uint64_t page_in;
784 uint64_t page_out;
785 /*
786 * If this process has been checkpointed before, then the user
787 * application will use the original gpu_id on the
788 * checkpointed node to refer to this device.
789 */
790 uint32_t user_gpu_id;
791
792 void *proc_ctx_bo;
793 uint64_t proc_ctx_gpu_addr;
794 void *proc_ctx_cpu_ptr;
795 };
796
797 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
798
799 struct svm_range_list {
800 struct mutex lock;
801 struct rb_root_cached objects;
802 struct list_head list;
803 struct work_struct deferred_list_work;
804 struct list_head deferred_range_list;
805 struct list_head criu_svm_metadata_list;
806 spinlock_t deferred_list_lock;
807 atomic_t evicted_ranges;
808 atomic_t drain_pagefaults;
809 struct delayed_work restore_work;
810 DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);
811 struct task_struct *faulting_task;
812 };
813
814 /* Process data */
815 struct kfd_process {
816 /*
817 * kfd_process are stored in an mm_struct*->kfd_process*
818 * hash table (kfd_processes in kfd_process.c)
819 */
820 struct hlist_node kfd_processes;
821
822 /*
823 * Opaque pointer to mm_struct. We don't hold a reference to
824 * it so it should never be dereferenced from here. This is
825 * only used for looking up processes by their mm.
826 */
827 void *mm;
828
829 struct kref ref;
830 struct work_struct release_work;
831
832 struct mutex mutex;
833
834 /*
835 * In any process, the thread that started main() is the lead
836 * thread and outlives the rest.
837 * It is here because amd_iommu_bind_pasid wants a task_struct.
838 * It can also be used for safely getting a reference to the
839 * mm_struct of the process.
840 */
841 struct task_struct *lead_thread;
842
843 /* We want to receive a notification when the mm_struct is destroyed */
844 struct mmu_notifier mmu_notifier;
845
846 u32 pasid;
847
848 /*
849 * Array of kfd_process_device pointers,
850 * one for each device the process is using.
851 */
852 struct kfd_process_device *pdds[MAX_GPU_INSTANCE];
853 uint32_t n_pdds;
854
855 struct process_queue_manager pqm;
856
857 /*Is the user space process 32 bit?*/
858 bool is_32bit_user_mode;
859
860 /* Event-related data */
861 struct mutex event_mutex;
862 /* Event ID allocator and lookup */
863 struct idr event_idr;
864 /* Event page */
865 u64 signal_handle;
866 struct kfd_signal_page *signal_page;
867 size_t signal_mapped_size;
868 size_t signal_event_count;
869 bool signal_event_limit_reached;
870
871 /* Information used for memory eviction */
872 void *kgd_process_info;
873 /* Eviction fence that is attached to all the BOs of this process. The
874 * fence will be triggered during eviction and new one will be created
875 * during restore
876 */
877 struct dma_fence *ef;
878
879 /* Work items for evicting and restoring BOs */
880 struct delayed_work eviction_work;
881 struct delayed_work restore_work;
882 /* seqno of the last scheduled eviction */
883 unsigned int last_eviction_seqno;
884 /* Approx. the last timestamp (in jiffies) when the process was
885 * restored after an eviction
886 */
887 unsigned long last_restore_timestamp;
888
889 /* Kobj for our procfs */
890 struct kobject *kobj;
891 struct kobject *kobj_queues;
892 struct attribute attr_pasid;
893
894 /* shared virtual memory registered by this process */
895 struct svm_range_list svms;
896
897 bool xnack_enabled;
898
899 atomic_t poison;
900 /* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */
901 bool queues_paused;
902 };
903
904 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
905 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
906 extern struct srcu_struct kfd_processes_srcu;
907
908 /**
909 * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
910 *
911 * @filep: pointer to file structure.
912 * @p: amdkfd process pointer.
913 * @data: pointer to arg that was copied from user.
914 *
915 * Return: returns ioctl completion code.
916 */
917 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
918 void *data);
919
920 struct amdkfd_ioctl_desc {
921 unsigned int cmd;
922 int flags;
923 amdkfd_ioctl_t *func;
924 unsigned int cmd_drv;
925 const char *name;
926 };
927 bool kfd_dev_is_large_bar(struct kfd_dev *dev);
928
929 int kfd_process_create_wq(void);
930 void kfd_process_destroy_wq(void);
931 void kfd_cleanup_processes(void);
932 struct kfd_process *kfd_create_process(struct file *filep);
933 struct kfd_process *kfd_get_process(const struct task_struct *task);
934 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid);
935 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
936
937 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
938 int kfd_process_gpuid_from_adev(struct kfd_process *p,
939 struct amdgpu_device *adev, uint32_t *gpuid,
940 uint32_t *gpuidx);
kfd_process_gpuid_from_gpuidx(struct kfd_process * p,uint32_t gpuidx,uint32_t * gpuid)941 static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
942 uint32_t gpuidx, uint32_t *gpuid) {
943 return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
944 }
kfd_process_device_from_gpuidx(struct kfd_process * p,uint32_t gpuidx)945 static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
946 struct kfd_process *p, uint32_t gpuidx) {
947 return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;
948 }
949
950 void kfd_unref_process(struct kfd_process *p);
951 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger);
952 int kfd_process_restore_queues(struct kfd_process *p);
953 void kfd_suspend_all_processes(void);
954 int kfd_resume_all_processes(void);
955
956 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *process,
957 uint32_t gpu_id);
958
959 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id);
960
961 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
962 struct file *drm_file);
963 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
964 struct kfd_process *p);
965 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
966 struct kfd_process *p);
967 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
968 struct kfd_process *p);
969
970 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
971
972 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
973 struct vm_area_struct *vma);
974
975 /* KFD process API for creating and translating handles */
976 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
977 void *mem);
978 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
979 int handle);
980 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
981 int handle);
982 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid);
983
984 /* PASIDs */
985 int kfd_pasid_init(void);
986 void kfd_pasid_exit(void);
987 bool kfd_set_pasid_limit(unsigned int new_limit);
988 unsigned int kfd_get_pasid_limit(void);
989 u32 kfd_pasid_alloc(void);
990 void kfd_pasid_free(u32 pasid);
991
992 /* Doorbells */
993 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
994 int kfd_doorbell_init(struct kfd_dev *kfd);
995 void kfd_doorbell_fini(struct kfd_dev *kfd);
996 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
997 struct vm_area_struct *vma);
998 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
999 unsigned int *doorbell_off);
1000 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
1001 u32 read_kernel_doorbell(u32 __iomem *db);
1002 void write_kernel_doorbell(void __iomem *db, u32 value);
1003 void write_kernel_doorbell64(void __iomem *db, u64 value);
1004 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
1005 struct kfd_process_device *pdd,
1006 unsigned int doorbell_id);
1007 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
1008 int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
1009 unsigned int *doorbell_index);
1010 void kfd_free_process_doorbells(struct kfd_dev *kfd,
1011 unsigned int doorbell_index);
1012 /* GTT Sub-Allocator */
1013
1014 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
1015 struct kfd_mem_obj **mem_obj);
1016
1017 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
1018
1019 extern struct device *kfd_device;
1020
1021 /* KFD's procfs */
1022 void kfd_procfs_init(void);
1023 void kfd_procfs_shutdown(void);
1024 int kfd_procfs_add_queue(struct queue *q);
1025 void kfd_procfs_del_queue(struct queue *q);
1026
1027 /* Topology */
1028 int kfd_topology_init(void);
1029 void kfd_topology_shutdown(void);
1030 int kfd_topology_add_device(struct kfd_dev *gpu);
1031 int kfd_topology_remove_device(struct kfd_dev *gpu);
1032 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
1033 uint32_t proximity_domain);
1034 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
1035 uint32_t proximity_domain);
1036 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
1037 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
1038 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
1039 struct kfd_dev *kfd_device_by_adev(const struct amdgpu_device *adev);
1040 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
1041 int kfd_numa_node_to_apic_id(int numa_node_id);
1042 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu);
1043
1044 /* Interrupts */
1045 int kfd_interrupt_init(struct kfd_dev *dev);
1046 void kfd_interrupt_exit(struct kfd_dev *dev);
1047 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry);
1048 bool interrupt_is_wanted(struct kfd_dev *dev,
1049 const uint32_t *ih_ring_entry,
1050 uint32_t *patched_ihre, bool *flag);
1051
1052 /* amdkfd Apertures */
1053 int kfd_init_apertures(struct kfd_process *process);
1054
1055 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1056 uint64_t tba_addr,
1057 uint64_t tma_addr);
1058
1059 /* CRIU */
1060 /*
1061 * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private
1062 * structures:
1063 * kfd_criu_process_priv_data
1064 * kfd_criu_device_priv_data
1065 * kfd_criu_bo_priv_data
1066 * kfd_criu_queue_priv_data
1067 * kfd_criu_event_priv_data
1068 * kfd_criu_svm_range_priv_data
1069 */
1070
1071 #define KFD_CRIU_PRIV_VERSION 1
1072
1073 struct kfd_criu_process_priv_data {
1074 uint32_t version;
1075 uint32_t xnack_mode;
1076 };
1077
1078 struct kfd_criu_device_priv_data {
1079 /* For future use */
1080 uint64_t reserved;
1081 };
1082
1083 struct kfd_criu_bo_priv_data {
1084 uint64_t user_addr;
1085 uint32_t idr_handle;
1086 uint32_t mapped_gpuids[MAX_GPU_INSTANCE];
1087 };
1088
1089 /*
1090 * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data,
1091 * kfd_criu_svm_range_priv_data is the object type
1092 */
1093 enum kfd_criu_object_type {
1094 KFD_CRIU_OBJECT_TYPE_QUEUE,
1095 KFD_CRIU_OBJECT_TYPE_EVENT,
1096 KFD_CRIU_OBJECT_TYPE_SVM_RANGE,
1097 };
1098
1099 struct kfd_criu_svm_range_priv_data {
1100 uint32_t object_type;
1101 uint64_t start_addr;
1102 uint64_t size;
1103 /* Variable length array of attributes */
1104 struct kfd_ioctl_svm_attribute attrs[];
1105 };
1106
1107 struct kfd_criu_queue_priv_data {
1108 uint32_t object_type;
1109 uint64_t q_address;
1110 uint64_t q_size;
1111 uint64_t read_ptr_addr;
1112 uint64_t write_ptr_addr;
1113 uint64_t doorbell_off;
1114 uint64_t eop_ring_buffer_address;
1115 uint64_t ctx_save_restore_area_address;
1116 uint32_t gpu_id;
1117 uint32_t type;
1118 uint32_t format;
1119 uint32_t q_id;
1120 uint32_t priority;
1121 uint32_t q_percent;
1122 uint32_t doorbell_id;
1123 uint32_t gws;
1124 uint32_t sdma_id;
1125 uint32_t eop_ring_buffer_size;
1126 uint32_t ctx_save_restore_area_size;
1127 uint32_t ctl_stack_size;
1128 uint32_t mqd_size;
1129 };
1130
1131 struct kfd_criu_event_priv_data {
1132 uint32_t object_type;
1133 uint64_t user_handle;
1134 uint32_t event_id;
1135 uint32_t auto_reset;
1136 uint32_t type;
1137 uint32_t signaled;
1138
1139 union {
1140 struct kfd_hsa_memory_exception_data memory_exception_data;
1141 struct kfd_hsa_hw_exception_data hw_exception_data;
1142 };
1143 };
1144
1145 int kfd_process_get_queue_info(struct kfd_process *p,
1146 uint32_t *num_queues,
1147 uint64_t *priv_data_sizes);
1148
1149 int kfd_criu_checkpoint_queues(struct kfd_process *p,
1150 uint8_t __user *user_priv_data,
1151 uint64_t *priv_data_offset);
1152
1153 int kfd_criu_restore_queue(struct kfd_process *p,
1154 uint8_t __user *user_priv_data,
1155 uint64_t *priv_data_offset,
1156 uint64_t max_priv_data_size);
1157
1158 int kfd_criu_checkpoint_events(struct kfd_process *p,
1159 uint8_t __user *user_priv_data,
1160 uint64_t *priv_data_offset);
1161
1162 int kfd_criu_restore_event(struct file *devkfd,
1163 struct kfd_process *p,
1164 uint8_t __user *user_priv_data,
1165 uint64_t *priv_data_offset,
1166 uint64_t max_priv_data_size);
1167 /* CRIU - End */
1168
1169 /* Queue Context Management */
1170 int init_queue(struct queue **q, const struct queue_properties *properties);
1171 void uninit_queue(struct queue *q);
1172 void print_queue_properties(struct queue_properties *q);
1173 void print_queue(struct queue *q);
1174
1175 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
1176 struct kfd_dev *dev);
1177 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
1178 struct kfd_dev *dev);
1179 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
1180 struct kfd_dev *dev);
1181 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
1182 struct kfd_dev *dev);
1183 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
1184 struct kfd_dev *dev);
1185 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
1186 struct kfd_dev *dev);
1187 struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type,
1188 struct kfd_dev *dev);
1189 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
1190 void device_queue_manager_uninit(struct device_queue_manager *dqm);
1191 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
1192 enum kfd_queue_type type);
1193 void kernel_queue_uninit(struct kernel_queue *kq, bool hanging);
1194 int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid);
1195
1196 /* Process Queue Manager */
1197 struct process_queue_node {
1198 struct queue *q;
1199 struct kernel_queue *kq;
1200 struct list_head process_queue_list;
1201 };
1202
1203 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
1204 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
1205 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
1206 void pqm_uninit(struct process_queue_manager *pqm);
1207 int pqm_create_queue(struct process_queue_manager *pqm,
1208 struct kfd_dev *dev,
1209 struct file *f,
1210 struct queue_properties *properties,
1211 unsigned int *qid,
1212 struct amdgpu_bo *wptr_bo,
1213 const struct kfd_criu_queue_priv_data *q_data,
1214 const void *restore_mqd,
1215 const void *restore_ctl_stack,
1216 uint32_t *p_doorbell_offset_in_process);
1217 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
1218 int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid,
1219 struct queue_properties *p);
1220 int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid,
1221 struct mqd_update_info *minfo);
1222 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
1223 void *gws);
1224 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
1225 unsigned int qid);
1226 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1227 unsigned int qid);
1228 int pqm_get_wave_state(struct process_queue_manager *pqm,
1229 unsigned int qid,
1230 void __user *ctl_stack,
1231 u32 *ctl_stack_used_size,
1232 u32 *save_area_used_size);
1233
1234 int amdkfd_fence_wait_timeout(uint64_t *fence_addr,
1235 uint64_t fence_value,
1236 unsigned int timeout_ms);
1237
1238 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
1239 unsigned int qid,
1240 u32 *mqd_size,
1241 u32 *ctl_stack_size);
1242 /* Packet Manager */
1243
1244 #define KFD_FENCE_COMPLETED (100)
1245 #define KFD_FENCE_INIT (10)
1246
1247 struct packet_manager {
1248 struct device_queue_manager *dqm;
1249 struct kernel_queue *priv_queue;
1250 struct mutex lock;
1251 bool allocated;
1252 struct kfd_mem_obj *ib_buffer_obj;
1253 unsigned int ib_size_bytes;
1254 bool is_over_subscription;
1255
1256 const struct packet_manager_funcs *pmf;
1257 };
1258
1259 struct packet_manager_funcs {
1260 /* Support ASIC-specific packet formats for PM4 packets */
1261 int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
1262 struct qcm_process_device *qpd);
1263 int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
1264 uint64_t ib, size_t ib_size_in_dwords, bool chain);
1265 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
1266 struct scheduling_resources *res);
1267 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
1268 struct queue *q, bool is_static);
1269 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
1270 enum kfd_unmap_queues_filter mode,
1271 uint32_t filter_param, bool reset);
1272 int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
1273 uint64_t fence_address, uint64_t fence_value);
1274 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
1275
1276 /* Packet sizes */
1277 int map_process_size;
1278 int runlist_size;
1279 int set_resources_size;
1280 int map_queues_size;
1281 int unmap_queues_size;
1282 int query_status_size;
1283 int release_mem_size;
1284 };
1285
1286 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
1287 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
1288 extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;
1289
1290 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
1291 void pm_uninit(struct packet_manager *pm, bool hanging);
1292 int pm_send_set_resources(struct packet_manager *pm,
1293 struct scheduling_resources *res);
1294 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
1295 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
1296 uint64_t fence_value);
1297
1298 int pm_send_unmap_queue(struct packet_manager *pm,
1299 enum kfd_unmap_queues_filter mode,
1300 uint32_t filter_param, bool reset);
1301
1302 void pm_release_ib(struct packet_manager *pm);
1303
1304 /* Following PM funcs can be shared among VI and AI */
1305 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
1306
1307 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
1308
1309 /* Events */
1310 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1311 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1312 extern const struct kfd_event_interrupt_class event_interrupt_class_v11;
1313
1314 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1315
1316 int kfd_event_init_process(struct kfd_process *p);
1317 void kfd_event_free_process(struct kfd_process *p);
1318 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1319 int kfd_wait_on_events(struct kfd_process *p,
1320 uint32_t num_events, void __user *data,
1321 bool all, uint32_t *user_timeout_ms,
1322 uint32_t *wait_result);
1323 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
1324 uint32_t valid_id_bits);
1325 void kfd_signal_iommu_event(struct kfd_dev *dev,
1326 u32 pasid, unsigned long address,
1327 bool is_write_requested, bool is_execute_requested);
1328 void kfd_signal_hw_exception_event(u32 pasid);
1329 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1330 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1331 int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset);
1332
1333 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1334 uint32_t event_type, bool auto_reset, uint32_t node_id,
1335 uint32_t *event_id, uint32_t *event_trigger_data,
1336 uint64_t *event_page_offset, uint32_t *event_slot_index);
1337
1338 int kfd_get_num_events(struct kfd_process *p);
1339 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1340
1341 void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
1342 struct kfd_vm_fault_info *info);
1343
1344 void kfd_signal_reset_event(struct kfd_dev *dev);
1345
1346 void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);
1347
1348 void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type);
1349
kfd_flush_tlb_after_unmap(struct kfd_dev * dev)1350 static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev)
1351 {
1352 return KFD_GC_VERSION(dev) > IP_VERSION(9, 4, 2) ||
1353 (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) ||
1354 KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0);
1355 }
1356
1357 bool kfd_is_locked(void);
1358
1359 /* Compute profile */
1360 void kfd_inc_compute_active(struct kfd_dev *dev);
1361 void kfd_dec_compute_active(struct kfd_dev *dev);
1362
1363 /* Cgroup Support */
1364 /* Check with device cgroup if @kfd device is accessible */
kfd_devcgroup_check_permission(struct kfd_dev * kfd)1365 static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)
1366 {
1367 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1368 struct drm_device *ddev = kfd->ddev;
1369
1370 return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1371 ddev->render->index,
1372 DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1373 #else
1374 return 0;
1375 #endif
1376 }
1377
1378 /* Debugfs */
1379 #if defined(CONFIG_DEBUG_FS)
1380
1381 void kfd_debugfs_init(void);
1382 void kfd_debugfs_fini(void);
1383 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1384 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1385 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1386 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1387 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1388 int pm_debugfs_runlist(struct seq_file *m, void *data);
1389
1390 int kfd_debugfs_hang_hws(struct kfd_dev *dev);
1391 int pm_debugfs_hang_hws(struct packet_manager *pm);
1392 int dqm_debugfs_hang_hws(struct device_queue_manager *dqm);
1393
1394 #else
1395
kfd_debugfs_init(void)1396 static inline void kfd_debugfs_init(void) {}
kfd_debugfs_fini(void)1397 static inline void kfd_debugfs_fini(void) {}
1398
1399 #endif
1400
1401 #endif
1402