1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23 #ifndef KFD_PRIV_H_INCLUDED
24 #define KFD_PRIV_H_INCLUDED
25
26 #include <linux/hashtable.h>
27 #include <linux/mmu_notifier.h>
28 #include <linux/mutex.h>
29 #include <linux/types.h>
30 #include <linux/atomic.h>
31 #include <linux/workqueue.h>
32 #include <linux/spinlock.h>
33 #include <linux/kfd_ioctl.h>
34 #include <linux/idr.h>
35 #include <linux/kfifo.h>
36 #include <linux/seq_file.h>
37 #include <linux/kref.h>
38 #include <linux/sysfs.h>
39 #include <linux/device_cgroup.h>
40 #include <drm/drm_file.h>
41 #include <drm/drm_drv.h>
42 #include <drm/drm_device.h>
43 #include <drm/drm_ioctl.h>
44 #include <kgd_kfd_interface.h>
45 #include <linux/swap.h>
46
47 #include "amd_shared.h"
48 #include "amdgpu.h"
49
50 #define KFD_MAX_RING_ENTRY_SIZE 8
51
52 #define KFD_SYSFS_FILE_MODE 0444
53
54 /* GPU ID hash width in bits */
55 #define KFD_GPU_ID_HASH_WIDTH 16
56
57 /* Use upper bits of mmap offset to store KFD driver specific information.
58 * BITS[63:62] - Encode MMAP type
59 * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to
60 * BITS[45:0] - MMAP offset value
61 *
62 * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
63 * defines are w.r.t to PAGE_SIZE
64 */
65 #define KFD_MMAP_TYPE_SHIFT 62
66 #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT)
67 #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT)
68 #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT)
69 #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT)
70 #define KFD_MMAP_TYPE_MMIO (0x0ULL << KFD_MMAP_TYPE_SHIFT)
71
72 #define KFD_MMAP_GPU_ID_SHIFT 46
73 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \
74 << KFD_MMAP_GPU_ID_SHIFT)
75 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\
76 & KFD_MMAP_GPU_ID_MASK)
77 #define KFD_MMAP_GET_GPU_ID(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \
78 >> KFD_MMAP_GPU_ID_SHIFT)
79
80 /*
81 * When working with cp scheduler we should assign the HIQ manually or via
82 * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot
83 * definitions for Kaveri. In Kaveri only the first ME queues participates
84 * in the cp scheduling taking that in mind we set the HIQ slot in the
85 * second ME.
86 */
87 #define KFD_CIK_HIQ_PIPE 4
88 #define KFD_CIK_HIQ_QUEUE 0
89
90 /* Macro for allocating structures */
91 #define kfd_alloc_struct(ptr_to_struct) \
92 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
93
94 #define KFD_MAX_NUM_OF_PROCESSES 512
95 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
96
97 /*
98 * Size of the per-process TBA+TMA buffer: 2 pages
99 *
100 * The first page is the TBA used for the CWSR ISA code. The second
101 * page is used as TMA for user-mode trap handler setup in daisy-chain mode.
102 */
103 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
104 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE
105
106 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \
107 (KFD_MAX_NUM_OF_PROCESSES * \
108 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
109
110 #define KFD_KERNEL_QUEUE_SIZE 2048
111
112 #define KFD_UNMAP_LATENCY_MS (4000)
113
114 /*
115 * 512 = 0x200
116 * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
117 * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
118 * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
119 * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
120 * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
121 */
122 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
123
124
125 /*
126 * Kernel module parameter to specify maximum number of supported queues per
127 * device
128 */
129 extern int max_num_of_queues_per_device;
130
131
132 /* Kernel module parameter to specify the scheduling policy */
133 extern int sched_policy;
134
135 /*
136 * Kernel module parameter to specify the maximum process
137 * number per HW scheduler
138 */
139 extern int hws_max_conc_proc;
140
141 extern int cwsr_enable;
142
143 /*
144 * Kernel module parameter to specify whether to send sigterm to HSA process on
145 * unhandled exception
146 */
147 extern int send_sigterm;
148
149 /*
150 * This kernel module is used to simulate large bar machine on non-large bar
151 * enabled machines.
152 */
153 extern int debug_largebar;
154
155 /*
156 * Ignore CRAT table during KFD initialization, can be used to work around
157 * broken CRAT tables on some AMD systems
158 */
159 extern int ignore_crat;
160
161 /* Set sh_mem_config.retry_disable on GFX v9 */
162 extern int amdgpu_noretry;
163
164 /* Halt if HWS hang is detected */
165 extern int halt_if_hws_hang;
166
167 /* Whether MEC FW support GWS barriers */
168 extern bool hws_gws_support;
169
170 /* Queue preemption timeout in ms */
171 extern int queue_preemption_timeout_ms;
172
173 /*
174 * Don't evict process queues on vm fault
175 */
176 extern int amdgpu_no_queue_eviction_on_vm_fault;
177
178 /* Enable eviction debug messages */
179 extern bool debug_evictions;
180
181 enum cache_policy {
182 cache_policy_coherent,
183 cache_policy_noncoherent
184 };
185
186 #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10)
187
188 struct kfd_event_interrupt_class {
189 bool (*interrupt_isr)(struct kfd_dev *dev,
190 const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
191 bool *patched_flag);
192 void (*interrupt_wq)(struct kfd_dev *dev,
193 const uint32_t *ih_ring_entry);
194 };
195
196 struct kfd_device_info {
197 enum amd_asic_type asic_family;
198 const char *asic_name;
199 uint32_t gfx_target_version;
200 const struct kfd_event_interrupt_class *event_interrupt_class;
201 unsigned int max_pasid_bits;
202 unsigned int max_no_of_hqd;
203 unsigned int doorbell_size;
204 size_t ih_ring_entry_size;
205 uint8_t num_of_watch_points;
206 uint16_t mqd_size_aligned;
207 bool supports_cwsr;
208 bool needs_iommu_device;
209 bool needs_pci_atomics;
210 uint32_t no_atomic_fw_version;
211 unsigned int num_sdma_engines;
212 unsigned int num_xgmi_sdma_engines;
213 unsigned int num_sdma_queues_per_engine;
214 };
215
216 struct kfd_mem_obj {
217 uint32_t range_start;
218 uint32_t range_end;
219 uint64_t gpu_addr;
220 uint32_t *cpu_ptr;
221 void *gtt_mem;
222 };
223
224 struct kfd_vmid_info {
225 uint32_t first_vmid_kfd;
226 uint32_t last_vmid_kfd;
227 uint32_t vmid_num_kfd;
228 };
229
230 struct kfd_dev {
231 struct kgd_dev *kgd;
232
233 const struct kfd_device_info *device_info;
234 struct pci_dev *pdev;
235 struct drm_device *ddev;
236
237 unsigned int id; /* topology stub index */
238
239 phys_addr_t doorbell_base; /* Start of actual doorbells used by
240 * KFD. It is aligned for mapping
241 * into user mode
242 */
243 size_t doorbell_base_dw_offset; /* Offset from the start of the PCI
244 * doorbell BAR to the first KFD
245 * doorbell in dwords. GFX reserves
246 * the segment before this offset.
247 */
248 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
249 * page used by kernel queue
250 */
251
252 struct kgd2kfd_shared_resources shared_resources;
253 struct kfd_vmid_info vm_info;
254
255 const struct kfd2kgd_calls *kfd2kgd;
256 struct mutex doorbell_mutex;
257 DECLARE_BITMAP(doorbell_available_index,
258 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
259
260 void *gtt_mem;
261 uint64_t gtt_start_gpu_addr;
262 void *gtt_start_cpu_ptr;
263 void *gtt_sa_bitmap;
264 struct mutex gtt_sa_lock;
265 unsigned int gtt_sa_chunk_size;
266 unsigned int gtt_sa_num_of_chunks;
267
268 /* Interrupts */
269 struct kfifo ih_fifo;
270 struct workqueue_struct *ih_wq;
271 struct work_struct interrupt_work;
272 spinlock_t interrupt_lock;
273
274 /* QCM Device instance */
275 struct device_queue_manager *dqm;
276
277 bool init_complete;
278 /*
279 * Interrupts of interest to KFD are copied
280 * from the HW ring into a SW ring.
281 */
282 bool interrupts_active;
283
284 /* Debug manager */
285 struct kfd_dbgmgr *dbgmgr;
286
287 /* Firmware versions */
288 uint16_t mec_fw_version;
289 uint16_t mec2_fw_version;
290 uint16_t sdma_fw_version;
291
292 /* Maximum process number mapped to HW scheduler */
293 unsigned int max_proc_per_quantum;
294
295 /* CWSR */
296 bool cwsr_enabled;
297 const void *cwsr_isa;
298 unsigned int cwsr_isa_size;
299
300 /* xGMI */
301 uint64_t hive_id;
302
303 bool pci_atomic_requested;
304
305 /* Use IOMMU v2 flag */
306 bool use_iommu_v2;
307
308 /* SRAM ECC flag */
309 atomic_t sram_ecc_flag;
310
311 /* Compute Profile ref. count */
312 atomic_t compute_profile;
313
314 /* Global GWS resource shared between processes */
315 void *gws;
316
317 /* Clients watching SMI events */
318 struct list_head smi_clients;
319 spinlock_t smi_lock;
320
321 uint32_t reset_seq_num;
322
323 struct ida doorbell_ida;
324 unsigned int max_doorbell_slices;
325
326 int noretry;
327
328 /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */
329 struct dev_pagemap pgmap;
330 };
331
332 enum kfd_mempool {
333 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
334 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
335 KFD_MEMPOOL_FRAMEBUFFER = 3,
336 };
337
338 /* Character device interface */
339 int kfd_chardev_init(void);
340 void kfd_chardev_exit(void);
341 struct device *kfd_chardev(void);
342
343 /**
344 * enum kfd_unmap_queues_filter - Enum for queue filters.
345 *
346 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
347 *
348 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
349 * running queues list.
350 *
351 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
352 * specific process.
353 *
354 */
355 enum kfd_unmap_queues_filter {
356 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
357 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
358 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
359 KFD_UNMAP_QUEUES_FILTER_BY_PASID
360 };
361
362 /**
363 * enum kfd_queue_type - Enum for various queue types.
364 *
365 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
366 *
367 * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
368 *
369 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
370 *
371 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
372 *
373 * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
374 */
375 enum kfd_queue_type {
376 KFD_QUEUE_TYPE_COMPUTE,
377 KFD_QUEUE_TYPE_SDMA,
378 KFD_QUEUE_TYPE_HIQ,
379 KFD_QUEUE_TYPE_DIQ,
380 KFD_QUEUE_TYPE_SDMA_XGMI
381 };
382
383 enum kfd_queue_format {
384 KFD_QUEUE_FORMAT_PM4,
385 KFD_QUEUE_FORMAT_AQL
386 };
387
388 enum KFD_QUEUE_PRIORITY {
389 KFD_QUEUE_PRIORITY_MINIMUM = 0,
390 KFD_QUEUE_PRIORITY_MAXIMUM = 15
391 };
392
393 /**
394 * struct queue_properties
395 *
396 * @type: The queue type.
397 *
398 * @queue_id: Queue identifier.
399 *
400 * @queue_address: Queue ring buffer address.
401 *
402 * @queue_size: Queue ring buffer size.
403 *
404 * @priority: Defines the queue priority relative to other queues in the
405 * process.
406 * This is just an indication and HW scheduling may override the priority as
407 * necessary while keeping the relative prioritization.
408 * the priority granularity is from 0 to f which f is the highest priority.
409 * currently all queues are initialized with the highest priority.
410 *
411 * @queue_percent: This field is partially implemented and currently a zero in
412 * this field defines that the queue is non active.
413 *
414 * @read_ptr: User space address which points to the number of dwords the
415 * cp read from the ring buffer. This field updates automatically by the H/W.
416 *
417 * @write_ptr: Defines the number of dwords written to the ring buffer.
418 *
419 * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
420 * buffer. This field should be similar to write_ptr and the user should
421 * update this field after updating the write_ptr.
422 *
423 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
424 *
425 * @is_interop: Defines if this is a interop queue. Interop queue means that
426 * the queue can access both graphics and compute resources.
427 *
428 * @is_evicted: Defines if the queue is evicted. Only active queues
429 * are evicted, rendering them inactive.
430 *
431 * @is_active: Defines if the queue is active or not. @is_active and
432 * @is_evicted are protected by the DQM lock.
433 *
434 * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
435 * @is_gws should be protected by the DQM lock, since changing it can yield the
436 * possibility of updating DQM state on number of GWS queues.
437 *
438 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
439 * of the queue.
440 *
441 * This structure represents the queue properties for each queue no matter if
442 * it's user mode or kernel mode queue.
443 *
444 */
445 struct queue_properties {
446 enum kfd_queue_type type;
447 enum kfd_queue_format format;
448 unsigned int queue_id;
449 uint64_t queue_address;
450 uint64_t queue_size;
451 uint32_t priority;
452 uint32_t queue_percent;
453 uint32_t *read_ptr;
454 uint32_t *write_ptr;
455 void __iomem *doorbell_ptr;
456 uint32_t doorbell_off;
457 bool is_interop;
458 bool is_evicted;
459 bool is_active;
460 bool is_gws;
461 /* Not relevant for user mode queues in cp scheduling */
462 unsigned int vmid;
463 /* Relevant only for sdma queues*/
464 uint32_t sdma_engine_id;
465 uint32_t sdma_queue_id;
466 uint32_t sdma_vm_addr;
467 /* Relevant only for VI */
468 uint64_t eop_ring_buffer_address;
469 uint32_t eop_ring_buffer_size;
470 uint64_t ctx_save_restore_area_address;
471 uint32_t ctx_save_restore_area_size;
472 uint32_t ctl_stack_size;
473 uint64_t tba_addr;
474 uint64_t tma_addr;
475 /* Relevant for CU */
476 uint32_t cu_mask_count; /* Must be a multiple of 32 */
477 uint32_t *cu_mask;
478 };
479
480 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \
481 (q).queue_address != 0 && \
482 (q).queue_percent > 0 && \
483 !(q).is_evicted)
484
485 /**
486 * struct queue
487 *
488 * @list: Queue linked list.
489 *
490 * @mqd: The queue MQD (memory queue descriptor).
491 *
492 * @mqd_mem_obj: The MQD local gpu memory object.
493 *
494 * @gart_mqd_addr: The MQD gart mc address.
495 *
496 * @properties: The queue properties.
497 *
498 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
499 * that the queue should be executed on.
500 *
501 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
502 * id.
503 *
504 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
505 *
506 * @process: The kfd process that created this queue.
507 *
508 * @device: The kfd device that created this queue.
509 *
510 * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
511 * otherwise.
512 *
513 * This structure represents user mode compute queues.
514 * It contains all the necessary data to handle such queues.
515 *
516 */
517
518 struct queue {
519 struct list_head list;
520 void *mqd;
521 struct kfd_mem_obj *mqd_mem_obj;
522 uint64_t gart_mqd_addr;
523 struct queue_properties properties;
524
525 uint32_t mec;
526 uint32_t pipe;
527 uint32_t queue;
528
529 unsigned int sdma_id;
530 unsigned int doorbell_id;
531
532 struct kfd_process *process;
533 struct kfd_dev *device;
534 void *gws;
535
536 /* procfs */
537 struct kobject kobj;
538 };
539
540 enum KFD_MQD_TYPE {
541 KFD_MQD_TYPE_HIQ = 0, /* for hiq */
542 KFD_MQD_TYPE_CP, /* for cp queues and diq */
543 KFD_MQD_TYPE_SDMA, /* for sdma queues */
544 KFD_MQD_TYPE_DIQ, /* for diq */
545 KFD_MQD_TYPE_MAX
546 };
547
548 enum KFD_PIPE_PRIORITY {
549 KFD_PIPE_PRIORITY_CS_LOW = 0,
550 KFD_PIPE_PRIORITY_CS_MEDIUM,
551 KFD_PIPE_PRIORITY_CS_HIGH
552 };
553
554 struct scheduling_resources {
555 unsigned int vmid_mask;
556 enum kfd_queue_type type;
557 uint64_t queue_mask;
558 uint64_t gws_mask;
559 uint32_t oac_mask;
560 uint32_t gds_heap_base;
561 uint32_t gds_heap_size;
562 };
563
564 struct process_queue_manager {
565 /* data */
566 struct kfd_process *process;
567 struct list_head queues;
568 unsigned long *queue_slot_bitmap;
569 };
570
571 struct qcm_process_device {
572 /* The Device Queue Manager that owns this data */
573 struct device_queue_manager *dqm;
574 struct process_queue_manager *pqm;
575 /* Queues list */
576 struct list_head queues_list;
577 struct list_head priv_queue_list;
578
579 unsigned int queue_count;
580 unsigned int vmid;
581 bool is_debug;
582 unsigned int evicted; /* eviction counter, 0=active */
583
584 /* This flag tells if we should reset all wavefronts on
585 * process termination
586 */
587 bool reset_wavefronts;
588
589 /* This flag tells us if this process has a GWS-capable
590 * queue that will be mapped into the runlist. It's
591 * possible to request a GWS BO, but not have the queue
592 * currently mapped, and this changes how the MAP_PROCESS
593 * PM4 packet is configured.
594 */
595 bool mapped_gws_queue;
596
597 /* All the memory management data should be here too */
598 uint64_t gds_context_area;
599 /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
600 uint64_t page_table_base;
601 uint32_t sh_mem_config;
602 uint32_t sh_mem_bases;
603 uint32_t sh_mem_ape1_base;
604 uint32_t sh_mem_ape1_limit;
605 uint32_t gds_size;
606 uint32_t num_gws;
607 uint32_t num_oac;
608 uint32_t sh_hidden_private_base;
609
610 /* CWSR memory */
611 void *cwsr_kaddr;
612 uint64_t cwsr_base;
613 uint64_t tba_addr;
614 uint64_t tma_addr;
615
616 /* IB memory */
617 uint64_t ib_base;
618 void *ib_kaddr;
619
620 /* doorbell resources per process per device */
621 unsigned long *doorbell_bitmap;
622 };
623
624 /* KFD Memory Eviction */
625
626 /* Approx. wait time before attempting to restore evicted BOs */
627 #define PROCESS_RESTORE_TIME_MS 100
628 /* Approx. back off time if restore fails due to lack of memory */
629 #define PROCESS_BACK_OFF_TIME_MS 100
630 /* Approx. time before evicting the process again */
631 #define PROCESS_ACTIVE_TIME_MS 10
632
633 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
634 * idr_handle in the least significant 4 bytes
635 */
636 #define MAKE_HANDLE(gpu_id, idr_handle) \
637 (((uint64_t)(gpu_id) << 32) + idr_handle)
638 #define GET_GPU_ID(handle) (handle >> 32)
639 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
640
641 enum kfd_pdd_bound {
642 PDD_UNBOUND = 0,
643 PDD_BOUND,
644 PDD_BOUND_SUSPENDED,
645 };
646
647 #define MAX_SYSFS_FILENAME_LEN 15
648
649 /*
650 * SDMA counter runs at 100MHz frequency.
651 * We display SDMA activity in microsecond granularity in sysfs.
652 * As a result, the divisor is 100.
653 */
654 #define SDMA_ACTIVITY_DIVISOR 100
655
656 /* Data that is per-process-per device. */
657 struct kfd_process_device {
658 /* The device that owns this data. */
659 struct kfd_dev *dev;
660
661 /* The process that owns this kfd_process_device. */
662 struct kfd_process *process;
663
664 /* per-process-per device QCM data structure */
665 struct qcm_process_device qpd;
666
667 /*Apertures*/
668 uint64_t lds_base;
669 uint64_t lds_limit;
670 uint64_t gpuvm_base;
671 uint64_t gpuvm_limit;
672 uint64_t scratch_base;
673 uint64_t scratch_limit;
674
675 /* VM context for GPUVM allocations */
676 struct file *drm_file;
677 void *drm_priv;
678
679 /* GPUVM allocations storage */
680 struct idr alloc_idr;
681
682 /* Flag used to tell the pdd has dequeued from the dqm.
683 * This is used to prevent dev->dqm->ops.process_termination() from
684 * being called twice when it is already called in IOMMU callback
685 * function.
686 */
687 bool already_dequeued;
688 bool runtime_inuse;
689
690 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
691 enum kfd_pdd_bound bound;
692
693 /* VRAM usage */
694 uint64_t vram_usage;
695 struct attribute attr_vram;
696 char vram_filename[MAX_SYSFS_FILENAME_LEN];
697
698 /* SDMA activity tracking */
699 uint64_t sdma_past_activity_counter;
700 struct attribute attr_sdma;
701 char sdma_filename[MAX_SYSFS_FILENAME_LEN];
702
703 /* Eviction activity tracking */
704 uint64_t last_evict_timestamp;
705 atomic64_t evict_duration_counter;
706 struct attribute attr_evict;
707
708 struct kobject *kobj_stats;
709 unsigned int doorbell_index;
710
711 /*
712 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
713 * that is associated with device encoded by "this" struct instance. The
714 * value reflects CU usage by all of the waves launched by this process
715 * on this device. A very important property of occupancy parameter is
716 * that its value is a snapshot of current use.
717 *
718 * Following is to be noted regarding how this parameter is reported:
719 *
720 * The number of waves that a CU can launch is limited by couple of
721 * parameters. These are encoded by struct amdgpu_cu_info instance
722 * that is part of every device definition. For GFX9 devices this
723 * translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
724 * do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
725 * when they do use scratch memory. This could change for future
726 * devices and therefore this example should be considered as a guide.
727 *
728 * All CU's of a device are available for the process. This may not be true
729 * under certain conditions - e.g. CU masking.
730 *
731 * Finally number of CU's that are occupied by a process is affected by both
732 * number of CU's a device has along with number of other competing processes
733 */
734 struct attribute attr_cu_occupancy;
735
736 /* sysfs counters for GPU retry fault and page migration tracking */
737 struct kobject *kobj_counters;
738 struct attribute attr_faults;
739 struct attribute attr_page_in;
740 struct attribute attr_page_out;
741 uint64_t faults;
742 uint64_t page_in;
743 uint64_t page_out;
744 };
745
746 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
747
748 struct svm_range_list {
749 struct mutex lock;
750 struct rb_root_cached objects;
751 struct list_head list;
752 struct work_struct deferred_list_work;
753 struct list_head deferred_range_list;
754 spinlock_t deferred_list_lock;
755 atomic_t evicted_ranges;
756 struct delayed_work restore_work;
757 DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);
758 struct task_struct *faulting_task;
759 };
760
761 /* Process data */
762 struct kfd_process {
763 /*
764 * kfd_process are stored in an mm_struct*->kfd_process*
765 * hash table (kfd_processes in kfd_process.c)
766 */
767 struct hlist_node kfd_processes;
768
769 /*
770 * Opaque pointer to mm_struct. We don't hold a reference to
771 * it so it should never be dereferenced from here. This is
772 * only used for looking up processes by their mm.
773 */
774 void *mm;
775
776 struct kref ref;
777 struct work_struct release_work;
778
779 struct mutex mutex;
780
781 /*
782 * In any process, the thread that started main() is the lead
783 * thread and outlives the rest.
784 * It is here because amd_iommu_bind_pasid wants a task_struct.
785 * It can also be used for safely getting a reference to the
786 * mm_struct of the process.
787 */
788 struct task_struct *lead_thread;
789
790 /* We want to receive a notification when the mm_struct is destroyed */
791 struct mmu_notifier mmu_notifier;
792
793 u32 pasid;
794
795 /*
796 * Array of kfd_process_device pointers,
797 * one for each device the process is using.
798 */
799 struct kfd_process_device *pdds[MAX_GPU_INSTANCE];
800 uint32_t n_pdds;
801
802 struct process_queue_manager pqm;
803
804 /*Is the user space process 32 bit?*/
805 bool is_32bit_user_mode;
806
807 /* Event-related data */
808 struct mutex event_mutex;
809 /* Event ID allocator and lookup */
810 struct idr event_idr;
811 /* Event page */
812 struct kfd_signal_page *signal_page;
813 size_t signal_mapped_size;
814 size_t signal_event_count;
815 bool signal_event_limit_reached;
816
817 /* Information used for memory eviction */
818 void *kgd_process_info;
819 /* Eviction fence that is attached to all the BOs of this process. The
820 * fence will be triggered during eviction and new one will be created
821 * during restore
822 */
823 struct dma_fence *ef;
824
825 /* Work items for evicting and restoring BOs */
826 struct delayed_work eviction_work;
827 struct delayed_work restore_work;
828 /* seqno of the last scheduled eviction */
829 unsigned int last_eviction_seqno;
830 /* Approx. the last timestamp (in jiffies) when the process was
831 * restored after an eviction
832 */
833 unsigned long last_restore_timestamp;
834
835 /* Kobj for our procfs */
836 struct kobject *kobj;
837 struct kobject *kobj_queues;
838 struct attribute attr_pasid;
839
840 /* shared virtual memory registered by this process */
841 struct svm_range_list svms;
842
843 bool xnack_enabled;
844 };
845
846 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
847 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
848 extern struct srcu_struct kfd_processes_srcu;
849
850 /**
851 * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
852 *
853 * @filep: pointer to file structure.
854 * @p: amdkfd process pointer.
855 * @data: pointer to arg that was copied from user.
856 *
857 * Return: returns ioctl completion code.
858 */
859 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
860 void *data);
861
862 struct amdkfd_ioctl_desc {
863 unsigned int cmd;
864 int flags;
865 amdkfd_ioctl_t *func;
866 unsigned int cmd_drv;
867 const char *name;
868 };
869 bool kfd_dev_is_large_bar(struct kfd_dev *dev);
870
871 int kfd_process_create_wq(void);
872 void kfd_process_destroy_wq(void);
873 struct kfd_process *kfd_create_process(struct file *filep);
874 struct kfd_process *kfd_get_process(const struct task_struct *);
875 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid);
876 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
877
878 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
879 int kfd_process_gpuid_from_kgd(struct kfd_process *p,
880 struct amdgpu_device *adev, uint32_t *gpuid,
881 uint32_t *gpuidx);
kfd_process_gpuid_from_gpuidx(struct kfd_process * p,uint32_t gpuidx,uint32_t * gpuid)882 static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
883 uint32_t gpuidx, uint32_t *gpuid) {
884 return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
885 }
kfd_process_device_from_gpuidx(struct kfd_process * p,uint32_t gpuidx)886 static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
887 struct kfd_process *p, uint32_t gpuidx) {
888 return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;
889 }
890
891 void kfd_unref_process(struct kfd_process *p);
892 int kfd_process_evict_queues(struct kfd_process *p);
893 int kfd_process_restore_queues(struct kfd_process *p);
894 void kfd_suspend_all_processes(void);
895 int kfd_resume_all_processes(void);
896
897 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
898 struct file *drm_file);
899 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
900 struct kfd_process *p);
901 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
902 struct kfd_process *p);
903 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
904 struct kfd_process *p);
905
906 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
907
908 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
909 struct vm_area_struct *vma);
910
911 /* KFD process API for creating and translating handles */
912 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
913 void *mem);
914 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
915 int handle);
916 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
917 int handle);
918
919 /* PASIDs */
920 int kfd_pasid_init(void);
921 void kfd_pasid_exit(void);
922 bool kfd_set_pasid_limit(unsigned int new_limit);
923 unsigned int kfd_get_pasid_limit(void);
924 u32 kfd_pasid_alloc(void);
925 void kfd_pasid_free(u32 pasid);
926
927 /* Doorbells */
928 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
929 int kfd_doorbell_init(struct kfd_dev *kfd);
930 void kfd_doorbell_fini(struct kfd_dev *kfd);
931 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
932 struct vm_area_struct *vma);
933 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
934 unsigned int *doorbell_off);
935 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
936 u32 read_kernel_doorbell(u32 __iomem *db);
937 void write_kernel_doorbell(void __iomem *db, u32 value);
938 void write_kernel_doorbell64(void __iomem *db, u64 value);
939 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
940 struct kfd_process_device *pdd,
941 unsigned int doorbell_id);
942 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
943 int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
944 unsigned int *doorbell_index);
945 void kfd_free_process_doorbells(struct kfd_dev *kfd,
946 unsigned int doorbell_index);
947 /* GTT Sub-Allocator */
948
949 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
950 struct kfd_mem_obj **mem_obj);
951
952 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
953
954 extern struct device *kfd_device;
955
956 /* KFD's procfs */
957 void kfd_procfs_init(void);
958 void kfd_procfs_shutdown(void);
959 int kfd_procfs_add_queue(struct queue *q);
960 void kfd_procfs_del_queue(struct queue *q);
961
962 /* Topology */
963 int kfd_topology_init(void);
964 void kfd_topology_shutdown(void);
965 int kfd_topology_add_device(struct kfd_dev *gpu);
966 int kfd_topology_remove_device(struct kfd_dev *gpu);
967 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
968 uint32_t proximity_domain);
969 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
970 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
971 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
972 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd);
973 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
974 int kfd_numa_node_to_apic_id(int numa_node_id);
975 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu);
976
977 /* Interrupts */
978 int kfd_interrupt_init(struct kfd_dev *dev);
979 void kfd_interrupt_exit(struct kfd_dev *dev);
980 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry);
981 bool interrupt_is_wanted(struct kfd_dev *dev,
982 const uint32_t *ih_ring_entry,
983 uint32_t *patched_ihre, bool *flag);
984
985 /* amdkfd Apertures */
986 int kfd_init_apertures(struct kfd_process *process);
987
988 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
989 uint64_t tba_addr,
990 uint64_t tma_addr);
991
992 /* Queue Context Management */
993 int init_queue(struct queue **q, const struct queue_properties *properties);
994 void uninit_queue(struct queue *q);
995 void print_queue_properties(struct queue_properties *q);
996 void print_queue(struct queue *q);
997
998 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
999 struct kfd_dev *dev);
1000 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
1001 struct kfd_dev *dev);
1002 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
1003 struct kfd_dev *dev);
1004 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
1005 struct kfd_dev *dev);
1006 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
1007 struct kfd_dev *dev);
1008 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
1009 struct kfd_dev *dev);
1010 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
1011 void device_queue_manager_uninit(struct device_queue_manager *dqm);
1012 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
1013 enum kfd_queue_type type);
1014 void kernel_queue_uninit(struct kernel_queue *kq, bool hanging);
1015 int kfd_process_vm_fault(struct device_queue_manager *dqm, u32 pasid);
1016
1017 /* Process Queue Manager */
1018 struct process_queue_node {
1019 struct queue *q;
1020 struct kernel_queue *kq;
1021 struct list_head process_queue_list;
1022 };
1023
1024 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
1025 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
1026 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
1027 void pqm_uninit(struct process_queue_manager *pqm);
1028 int pqm_create_queue(struct process_queue_manager *pqm,
1029 struct kfd_dev *dev,
1030 struct file *f,
1031 struct queue_properties *properties,
1032 unsigned int *qid,
1033 uint32_t *p_doorbell_offset_in_process);
1034 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
1035 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
1036 struct queue_properties *p);
1037 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
1038 struct queue_properties *p);
1039 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
1040 void *gws);
1041 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
1042 unsigned int qid);
1043 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1044 unsigned int qid);
1045 int pqm_get_wave_state(struct process_queue_manager *pqm,
1046 unsigned int qid,
1047 void __user *ctl_stack,
1048 u32 *ctl_stack_used_size,
1049 u32 *save_area_used_size);
1050
1051 int amdkfd_fence_wait_timeout(uint64_t *fence_addr,
1052 uint64_t fence_value,
1053 unsigned int timeout_ms);
1054
1055 /* Packet Manager */
1056
1057 #define KFD_FENCE_COMPLETED (100)
1058 #define KFD_FENCE_INIT (10)
1059
1060 struct packet_manager {
1061 struct device_queue_manager *dqm;
1062 struct kernel_queue *priv_queue;
1063 struct mutex lock;
1064 bool allocated;
1065 struct kfd_mem_obj *ib_buffer_obj;
1066 unsigned int ib_size_bytes;
1067 bool is_over_subscription;
1068
1069 const struct packet_manager_funcs *pmf;
1070 };
1071
1072 struct packet_manager_funcs {
1073 /* Support ASIC-specific packet formats for PM4 packets */
1074 int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
1075 struct qcm_process_device *qpd);
1076 int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
1077 uint64_t ib, size_t ib_size_in_dwords, bool chain);
1078 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
1079 struct scheduling_resources *res);
1080 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
1081 struct queue *q, bool is_static);
1082 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
1083 enum kfd_queue_type type,
1084 enum kfd_unmap_queues_filter mode,
1085 uint32_t filter_param, bool reset,
1086 unsigned int sdma_engine);
1087 int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
1088 uint64_t fence_address, uint64_t fence_value);
1089 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
1090
1091 /* Packet sizes */
1092 int map_process_size;
1093 int runlist_size;
1094 int set_resources_size;
1095 int map_queues_size;
1096 int unmap_queues_size;
1097 int query_status_size;
1098 int release_mem_size;
1099 };
1100
1101 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
1102 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
1103 extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;
1104
1105 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
1106 void pm_uninit(struct packet_manager *pm, bool hanging);
1107 int pm_send_set_resources(struct packet_manager *pm,
1108 struct scheduling_resources *res);
1109 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
1110 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
1111 uint64_t fence_value);
1112
1113 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
1114 enum kfd_unmap_queues_filter mode,
1115 uint32_t filter_param, bool reset,
1116 unsigned int sdma_engine);
1117
1118 void pm_release_ib(struct packet_manager *pm);
1119
1120 /* Following PM funcs can be shared among VI and AI */
1121 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
1122
1123 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
1124
1125 /* Events */
1126 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1127 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1128
1129 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1130
1131 void kfd_event_init_process(struct kfd_process *p);
1132 void kfd_event_free_process(struct kfd_process *p);
1133 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1134 int kfd_wait_on_events(struct kfd_process *p,
1135 uint32_t num_events, void __user *data,
1136 bool all, uint32_t user_timeout_ms,
1137 uint32_t *wait_result);
1138 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
1139 uint32_t valid_id_bits);
1140 void kfd_signal_iommu_event(struct kfd_dev *dev,
1141 u32 pasid, unsigned long address,
1142 bool is_write_requested, bool is_execute_requested);
1143 void kfd_signal_hw_exception_event(u32 pasid);
1144 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1145 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1146 int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
1147 uint64_t size);
1148 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1149 uint32_t event_type, bool auto_reset, uint32_t node_id,
1150 uint32_t *event_id, uint32_t *event_trigger_data,
1151 uint64_t *event_page_offset, uint32_t *event_slot_index);
1152 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1153
1154 void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
1155 struct kfd_vm_fault_info *info);
1156
1157 void kfd_signal_reset_event(struct kfd_dev *dev);
1158
1159 void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);
1160
1161 void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type);
1162
1163 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
1164
1165 bool kfd_is_locked(void);
1166
1167 /* Compute profile */
1168 void kfd_inc_compute_active(struct kfd_dev *dev);
1169 void kfd_dec_compute_active(struct kfd_dev *dev);
1170
1171 /* Cgroup Support */
1172 /* Check with device cgroup if @kfd device is accessible */
kfd_devcgroup_check_permission(struct kfd_dev * kfd)1173 static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)
1174 {
1175 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1176 struct drm_device *ddev = kfd->ddev;
1177
1178 return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1179 ddev->render->index,
1180 DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1181 #else
1182 return 0;
1183 #endif
1184 }
1185
1186 /* Debugfs */
1187 #if defined(CONFIG_DEBUG_FS)
1188
1189 void kfd_debugfs_init(void);
1190 void kfd_debugfs_fini(void);
1191 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1192 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1193 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1194 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1195 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1196 int pm_debugfs_runlist(struct seq_file *m, void *data);
1197
1198 int kfd_debugfs_hang_hws(struct kfd_dev *dev);
1199 int pm_debugfs_hang_hws(struct packet_manager *pm);
1200 int dqm_debugfs_hang_hws(struct device_queue_manager *dqm);
1201
1202 #else
1203
kfd_debugfs_init(void)1204 static inline void kfd_debugfs_init(void) {}
kfd_debugfs_fini(void)1205 static inline void kfd_debugfs_fini(void) {}
1206
1207 #endif
1208
1209 #endif
1210