2 * Copyright 2014 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
23 #ifndef KFD_PRIV_H_INCLUDED
24 #define KFD_PRIV_H_INCLUDED
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>
47 #include "amd_shared.h"
50 #define KFD_MAX_RING_ENTRY_SIZE 8
52 #define KFD_SYSFS_FILE_MODE 0444
54 /* GPU ID hash width in bits */
55 #define KFD_GPU_ID_HASH_WIDTH 16
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
62 * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
63 * defines are w.r.t to PAGE_SIZE
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)
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)
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
87 #define KFD_CIK_HIQ_PIPE 4
88 #define KFD_CIK_HIQ_QUEUE 0
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))
94 #define KFD_MAX_NUM_OF_PROCESSES 512
95 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
98 * Size of the per-process TBA+TMA buffer: 2 pages
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.
103 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
104 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE
106 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \
107 (KFD_MAX_NUM_OF_PROCESSES * \
108 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
110 #define KFD_KERNEL_QUEUE_SIZE 2048
112 #define KFD_UNMAP_LATENCY_MS (4000)
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.
122 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
126 * Kernel module parameter to specify maximum number of supported queues per
129 extern int max_num_of_queues_per_device;
132 /* Kernel module parameter to specify the scheduling policy */
133 extern int sched_policy;
136 * Kernel module parameter to specify the maximum process
137 * number per HW scheduler
139 extern int hws_max_conc_proc;
141 extern int cwsr_enable;
144 * Kernel module parameter to specify whether to send sigterm to HSA process on
145 * unhandled exception
147 extern int send_sigterm;
150 * This kernel module is used to simulate large bar machine on non-large bar
153 extern int debug_largebar;
156 * Ignore CRAT table during KFD initialization, can be used to work around
157 * broken CRAT tables on some AMD systems
159 extern int ignore_crat;
161 /* Set sh_mem_config.retry_disable on GFX v9 */
162 extern int amdgpu_noretry;
164 /* Halt if HWS hang is detected */
165 extern int halt_if_hws_hang;
167 /* Whether MEC FW support GWS barriers */
168 extern bool hws_gws_support;
170 /* Queue preemption timeout in ms */
171 extern int queue_preemption_timeout_ms;
174 * Don't evict process queues on vm fault
176 extern int amdgpu_no_queue_eviction_on_vm_fault;
178 /* Enable eviction debug messages */
179 extern bool debug_evictions;
182 cache_policy_coherent,
183 cache_policy_noncoherent
186 #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10)
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,
192 void (*interrupt_wq)(struct kfd_dev *dev,
193 const uint32_t *ih_ring_entry);
196 struct kfd_device_info {
197 enum amd_asic_type asic_family;
198 const char *asic_name;
199 const struct kfd_event_interrupt_class *event_interrupt_class;
200 unsigned int max_pasid_bits;
201 unsigned int max_no_of_hqd;
202 unsigned int doorbell_size;
203 size_t ih_ring_entry_size;
204 uint8_t num_of_watch_points;
205 uint16_t mqd_size_aligned;
207 bool needs_iommu_device;
208 bool needs_pci_atomics;
209 unsigned int num_sdma_engines;
210 unsigned int num_xgmi_sdma_engines;
211 unsigned int num_sdma_queues_per_engine;
215 uint32_t range_start;
222 struct kfd_vmid_info {
223 uint32_t first_vmid_kfd;
224 uint32_t last_vmid_kfd;
225 uint32_t vmid_num_kfd;
231 const struct kfd_device_info *device_info;
232 struct pci_dev *pdev;
233 struct drm_device *ddev;
235 unsigned int id; /* topology stub index */
237 phys_addr_t doorbell_base; /* Start of actual doorbells used by
238 * KFD. It is aligned for mapping
241 size_t doorbell_base_dw_offset; /* Offset from the start of the PCI
242 * doorbell BAR to the first KFD
243 * doorbell in dwords. GFX reserves
244 * the segment before this offset.
246 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
247 * page used by kernel queue
250 struct kgd2kfd_shared_resources shared_resources;
251 struct kfd_vmid_info vm_info;
253 const struct kfd2kgd_calls *kfd2kgd;
254 struct mutex doorbell_mutex;
255 DECLARE_BITMAP(doorbell_available_index,
256 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
259 uint64_t gtt_start_gpu_addr;
260 void *gtt_start_cpu_ptr;
262 struct mutex gtt_sa_lock;
263 unsigned int gtt_sa_chunk_size;
264 unsigned int gtt_sa_num_of_chunks;
267 struct kfifo ih_fifo;
268 struct workqueue_struct *ih_wq;
269 struct work_struct interrupt_work;
270 spinlock_t interrupt_lock;
272 /* QCM Device instance */
273 struct device_queue_manager *dqm;
277 * Interrupts of interest to KFD are copied
278 * from the HW ring into a SW ring.
280 bool interrupts_active;
283 struct kfd_dbgmgr *dbgmgr;
285 /* Firmware versions */
286 uint16_t mec_fw_version;
287 uint16_t mec2_fw_version;
288 uint16_t sdma_fw_version;
290 /* Maximum process number mapped to HW scheduler */
291 unsigned int max_proc_per_quantum;
295 const void *cwsr_isa;
296 unsigned int cwsr_isa_size;
301 bool pci_atomic_requested;
303 /* Use IOMMU v2 flag */
307 atomic_t sram_ecc_flag;
309 /* Compute Profile ref. count */
310 atomic_t compute_profile;
312 /* Global GWS resource shared between processes */
315 /* Clients watching SMI events */
316 struct list_head smi_clients;
319 uint32_t reset_seq_num;
321 struct ida doorbell_ida;
322 unsigned int max_doorbell_slices;
326 /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */
327 struct dev_pagemap pgmap;
331 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
332 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
333 KFD_MEMPOOL_FRAMEBUFFER = 3,
336 /* Character device interface */
337 int kfd_chardev_init(void);
338 void kfd_chardev_exit(void);
339 struct device *kfd_chardev(void);
342 * enum kfd_unmap_queues_filter - Enum for queue filters.
344 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
346 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
347 * running queues list.
349 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
353 enum kfd_unmap_queues_filter {
354 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
355 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
356 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
357 KFD_UNMAP_QUEUES_FILTER_BY_PASID
361 * enum kfd_queue_type - Enum for various queue types.
363 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
365 * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
367 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
369 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
371 * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
373 enum kfd_queue_type {
374 KFD_QUEUE_TYPE_COMPUTE,
378 KFD_QUEUE_TYPE_SDMA_XGMI
381 enum kfd_queue_format {
382 KFD_QUEUE_FORMAT_PM4,
386 enum KFD_QUEUE_PRIORITY {
387 KFD_QUEUE_PRIORITY_MINIMUM = 0,
388 KFD_QUEUE_PRIORITY_MAXIMUM = 15
392 * struct queue_properties
394 * @type: The queue type.
396 * @queue_id: Queue identifier.
398 * @queue_address: Queue ring buffer address.
400 * @queue_size: Queue ring buffer size.
402 * @priority: Defines the queue priority relative to other queues in the
404 * This is just an indication and HW scheduling may override the priority as
405 * necessary while keeping the relative prioritization.
406 * the priority granularity is from 0 to f which f is the highest priority.
407 * currently all queues are initialized with the highest priority.
409 * @queue_percent: This field is partially implemented and currently a zero in
410 * this field defines that the queue is non active.
412 * @read_ptr: User space address which points to the number of dwords the
413 * cp read from the ring buffer. This field updates automatically by the H/W.
415 * @write_ptr: Defines the number of dwords written to the ring buffer.
417 * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
418 * buffer. This field should be similar to write_ptr and the user should
419 * update this field after updating the write_ptr.
421 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
423 * @is_interop: Defines if this is a interop queue. Interop queue means that
424 * the queue can access both graphics and compute resources.
426 * @is_evicted: Defines if the queue is evicted. Only active queues
427 * are evicted, rendering them inactive.
429 * @is_active: Defines if the queue is active or not. @is_active and
430 * @is_evicted are protected by the DQM lock.
432 * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
433 * @is_gws should be protected by the DQM lock, since changing it can yield the
434 * possibility of updating DQM state on number of GWS queues.
436 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
439 * This structure represents the queue properties for each queue no matter if
440 * it's user mode or kernel mode queue.
443 struct queue_properties {
444 enum kfd_queue_type type;
445 enum kfd_queue_format format;
446 unsigned int queue_id;
447 uint64_t queue_address;
450 uint32_t queue_percent;
453 void __iomem *doorbell_ptr;
454 uint32_t doorbell_off;
459 /* Not relevant for user mode queues in cp scheduling */
461 /* Relevant only for sdma queues*/
462 uint32_t sdma_engine_id;
463 uint32_t sdma_queue_id;
464 uint32_t sdma_vm_addr;
465 /* Relevant only for VI */
466 uint64_t eop_ring_buffer_address;
467 uint32_t eop_ring_buffer_size;
468 uint64_t ctx_save_restore_area_address;
469 uint32_t ctx_save_restore_area_size;
470 uint32_t ctl_stack_size;
473 /* Relevant for CU */
474 uint32_t cu_mask_count; /* Must be a multiple of 32 */
478 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \
479 (q).queue_address != 0 && \
480 (q).queue_percent > 0 && \
486 * @list: Queue linked list.
488 * @mqd: The queue MQD (memory queue descriptor).
490 * @mqd_mem_obj: The MQD local gpu memory object.
492 * @gart_mqd_addr: The MQD gart mc address.
494 * @properties: The queue properties.
496 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
497 * that the queue should be executed on.
499 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
502 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
504 * @process: The kfd process that created this queue.
506 * @device: The kfd device that created this queue.
508 * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
511 * This structure represents user mode compute queues.
512 * It contains all the necessary data to handle such queues.
517 struct list_head list;
519 struct kfd_mem_obj *mqd_mem_obj;
520 uint64_t gart_mqd_addr;
521 struct queue_properties properties;
527 unsigned int sdma_id;
528 unsigned int doorbell_id;
530 struct kfd_process *process;
531 struct kfd_dev *device;
539 KFD_MQD_TYPE_HIQ = 0, /* for hiq */
540 KFD_MQD_TYPE_CP, /* for cp queues and diq */
541 KFD_MQD_TYPE_SDMA, /* for sdma queues */
542 KFD_MQD_TYPE_DIQ, /* for diq */
546 enum KFD_PIPE_PRIORITY {
547 KFD_PIPE_PRIORITY_CS_LOW = 0,
548 KFD_PIPE_PRIORITY_CS_MEDIUM,
549 KFD_PIPE_PRIORITY_CS_HIGH
552 struct scheduling_resources {
553 unsigned int vmid_mask;
554 enum kfd_queue_type type;
558 uint32_t gds_heap_base;
559 uint32_t gds_heap_size;
562 struct process_queue_manager {
564 struct kfd_process *process;
565 struct list_head queues;
566 unsigned long *queue_slot_bitmap;
569 struct qcm_process_device {
570 /* The Device Queue Manager that owns this data */
571 struct device_queue_manager *dqm;
572 struct process_queue_manager *pqm;
574 struct list_head queues_list;
575 struct list_head priv_queue_list;
577 unsigned int queue_count;
580 unsigned int evicted; /* eviction counter, 0=active */
582 /* This flag tells if we should reset all wavefronts on
583 * process termination
585 bool reset_wavefronts;
587 /* This flag tells us if this process has a GWS-capable
588 * queue that will be mapped into the runlist. It's
589 * possible to request a GWS BO, but not have the queue
590 * currently mapped, and this changes how the MAP_PROCESS
591 * PM4 packet is configured.
593 bool mapped_gws_queue;
595 /* All the memory management data should be here too */
596 uint64_t gds_context_area;
597 /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
598 uint64_t page_table_base;
599 uint32_t sh_mem_config;
600 uint32_t sh_mem_bases;
601 uint32_t sh_mem_ape1_base;
602 uint32_t sh_mem_ape1_limit;
606 uint32_t sh_hidden_private_base;
618 /* doorbell resources per process per device */
619 unsigned long *doorbell_bitmap;
622 /* KFD Memory Eviction */
624 /* Approx. wait time before attempting to restore evicted BOs */
625 #define PROCESS_RESTORE_TIME_MS 100
626 /* Approx. back off time if restore fails due to lack of memory */
627 #define PROCESS_BACK_OFF_TIME_MS 100
628 /* Approx. time before evicting the process again */
629 #define PROCESS_ACTIVE_TIME_MS 10
631 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
632 * idr_handle in the least significant 4 bytes
634 #define MAKE_HANDLE(gpu_id, idr_handle) \
635 (((uint64_t)(gpu_id) << 32) + idr_handle)
636 #define GET_GPU_ID(handle) (handle >> 32)
637 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
645 #define MAX_SYSFS_FILENAME_LEN 15
648 * SDMA counter runs at 100MHz frequency.
649 * We display SDMA activity in microsecond granularity in sysfs.
650 * As a result, the divisor is 100.
652 #define SDMA_ACTIVITY_DIVISOR 100
654 /* Data that is per-process-per device. */
655 struct kfd_process_device {
656 /* The device that owns this data. */
659 /* The process that owns this kfd_process_device. */
660 struct kfd_process *process;
662 /* per-process-per device QCM data structure */
663 struct qcm_process_device qpd;
669 uint64_t gpuvm_limit;
670 uint64_t scratch_base;
671 uint64_t scratch_limit;
673 /* VM context for GPUVM allocations */
674 struct file *drm_file;
677 /* GPUVM allocations storage */
678 struct idr alloc_idr;
680 /* Flag used to tell the pdd has dequeued from the dqm.
681 * This is used to prevent dev->dqm->ops.process_termination() from
682 * being called twice when it is already called in IOMMU callback
685 bool already_dequeued;
688 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
689 enum kfd_pdd_bound bound;
693 struct attribute attr_vram;
694 char vram_filename[MAX_SYSFS_FILENAME_LEN];
696 /* SDMA activity tracking */
697 uint64_t sdma_past_activity_counter;
698 struct attribute attr_sdma;
699 char sdma_filename[MAX_SYSFS_FILENAME_LEN];
701 /* Eviction activity tracking */
702 uint64_t last_evict_timestamp;
703 atomic64_t evict_duration_counter;
704 struct attribute attr_evict;
706 struct kobject *kobj_stats;
707 unsigned int doorbell_index;
710 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
711 * that is associated with device encoded by "this" struct instance. The
712 * value reflects CU usage by all of the waves launched by this process
713 * on this device. A very important property of occupancy parameter is
714 * that its value is a snapshot of current use.
716 * Following is to be noted regarding how this parameter is reported:
718 * The number of waves that a CU can launch is limited by couple of
719 * parameters. These are encoded by struct amdgpu_cu_info instance
720 * that is part of every device definition. For GFX9 devices this
721 * translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
722 * do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
723 * when they do use scratch memory. This could change for future
724 * devices and therefore this example should be considered as a guide.
726 * All CU's of a device are available for the process. This may not be true
727 * under certain conditions - e.g. CU masking.
729 * Finally number of CU's that are occupied by a process is affected by both
730 * number of CU's a device has along with number of other competing processes
732 struct attribute attr_cu_occupancy;
734 /* sysfs counters for GPU retry fault and page migration tracking */
735 struct kobject *kobj_counters;
736 struct attribute attr_faults;
737 struct attribute attr_page_in;
738 struct attribute attr_page_out;
744 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
746 struct svm_range_list {
748 struct rb_root_cached objects;
749 struct list_head list;
750 struct work_struct deferred_list_work;
751 struct list_head deferred_range_list;
752 spinlock_t deferred_list_lock;
753 atomic_t evicted_ranges;
754 struct delayed_work restore_work;
755 DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);
761 * kfd_process are stored in an mm_struct*->kfd_process*
762 * hash table (kfd_processes in kfd_process.c)
764 struct hlist_node kfd_processes;
767 * Opaque pointer to mm_struct. We don't hold a reference to
768 * it so it should never be dereferenced from here. This is
769 * only used for looking up processes by their mm.
774 struct work_struct release_work;
779 * In any process, the thread that started main() is the lead
780 * thread and outlives the rest.
781 * It is here because amd_iommu_bind_pasid wants a task_struct.
782 * It can also be used for safely getting a reference to the
783 * mm_struct of the process.
785 struct task_struct *lead_thread;
787 /* We want to receive a notification when the mm_struct is destroyed */
788 struct mmu_notifier mmu_notifier;
793 * Array of kfd_process_device pointers,
794 * one for each device the process is using.
796 struct kfd_process_device *pdds[MAX_GPU_INSTANCE];
799 struct process_queue_manager pqm;
801 /*Is the user space process 32 bit?*/
802 bool is_32bit_user_mode;
804 /* Event-related data */
805 struct mutex event_mutex;
806 /* Event ID allocator and lookup */
807 struct idr event_idr;
809 struct kfd_signal_page *signal_page;
810 size_t signal_mapped_size;
811 size_t signal_event_count;
812 bool signal_event_limit_reached;
814 /* Information used for memory eviction */
815 void *kgd_process_info;
816 /* Eviction fence that is attached to all the BOs of this process. The
817 * fence will be triggered during eviction and new one will be created
820 struct dma_fence *ef;
822 /* Work items for evicting and restoring BOs */
823 struct delayed_work eviction_work;
824 struct delayed_work restore_work;
825 /* seqno of the last scheduled eviction */
826 unsigned int last_eviction_seqno;
827 /* Approx. the last timestamp (in jiffies) when the process was
828 * restored after an eviction
830 unsigned long last_restore_timestamp;
832 /* Kobj for our procfs */
833 struct kobject *kobj;
834 struct kobject *kobj_queues;
835 struct attribute attr_pasid;
837 /* shared virtual memory registered by this process */
838 struct svm_range_list svms;
843 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
844 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
845 extern struct srcu_struct kfd_processes_srcu;
848 * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
850 * @filep: pointer to file structure.
851 * @p: amdkfd process pointer.
852 * @data: pointer to arg that was copied from user.
854 * Return: returns ioctl completion code.
856 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
859 struct amdkfd_ioctl_desc {
862 amdkfd_ioctl_t *func;
863 unsigned int cmd_drv;
866 bool kfd_dev_is_large_bar(struct kfd_dev *dev);
868 int kfd_process_create_wq(void);
869 void kfd_process_destroy_wq(void);
870 struct kfd_process *kfd_create_process(struct file *filep);
871 struct kfd_process *kfd_get_process(const struct task_struct *);
872 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid);
873 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
875 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
876 int kfd_process_gpuid_from_kgd(struct kfd_process *p,
877 struct amdgpu_device *adev, uint32_t *gpuid,
879 static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
880 uint32_t gpuidx, uint32_t *gpuid) {
881 return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
883 static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
884 struct kfd_process *p, uint32_t gpuidx) {
885 return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;
888 void kfd_unref_process(struct kfd_process *p);
889 int kfd_process_evict_queues(struct kfd_process *p);
890 int kfd_process_restore_queues(struct kfd_process *p);
891 void kfd_suspend_all_processes(void);
892 int kfd_resume_all_processes(void);
894 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
895 struct file *drm_file);
896 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
897 struct kfd_process *p);
898 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
899 struct kfd_process *p);
900 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
901 struct kfd_process *p);
903 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
905 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
906 struct vm_area_struct *vma);
908 /* KFD process API for creating and translating handles */
909 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
911 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
913 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
917 int kfd_pasid_init(void);
918 void kfd_pasid_exit(void);
919 bool kfd_set_pasid_limit(unsigned int new_limit);
920 unsigned int kfd_get_pasid_limit(void);
921 u32 kfd_pasid_alloc(void);
922 void kfd_pasid_free(u32 pasid);
925 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
926 int kfd_doorbell_init(struct kfd_dev *kfd);
927 void kfd_doorbell_fini(struct kfd_dev *kfd);
928 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
929 struct vm_area_struct *vma);
930 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
931 unsigned int *doorbell_off);
932 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
933 u32 read_kernel_doorbell(u32 __iomem *db);
934 void write_kernel_doorbell(void __iomem *db, u32 value);
935 void write_kernel_doorbell64(void __iomem *db, u64 value);
936 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
937 struct kfd_process_device *pdd,
938 unsigned int doorbell_id);
939 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
940 int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
941 unsigned int *doorbell_index);
942 void kfd_free_process_doorbells(struct kfd_dev *kfd,
943 unsigned int doorbell_index);
944 /* GTT Sub-Allocator */
946 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
947 struct kfd_mem_obj **mem_obj);
949 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
951 extern struct device *kfd_device;
954 void kfd_procfs_init(void);
955 void kfd_procfs_shutdown(void);
956 int kfd_procfs_add_queue(struct queue *q);
957 void kfd_procfs_del_queue(struct queue *q);
960 int kfd_topology_init(void);
961 void kfd_topology_shutdown(void);
962 int kfd_topology_add_device(struct kfd_dev *gpu);
963 int kfd_topology_remove_device(struct kfd_dev *gpu);
964 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
965 uint32_t proximity_domain);
966 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
967 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
968 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
969 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd);
970 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
971 int kfd_numa_node_to_apic_id(int numa_node_id);
972 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu);
975 int kfd_interrupt_init(struct kfd_dev *dev);
976 void kfd_interrupt_exit(struct kfd_dev *dev);
977 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry);
978 bool interrupt_is_wanted(struct kfd_dev *dev,
979 const uint32_t *ih_ring_entry,
980 uint32_t *patched_ihre, bool *flag);
982 /* amdkfd Apertures */
983 int kfd_init_apertures(struct kfd_process *process);
985 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
989 /* Queue Context Management */
990 int init_queue(struct queue **q, const struct queue_properties *properties);
991 void uninit_queue(struct queue *q);
992 void print_queue_properties(struct queue_properties *q);
993 void print_queue(struct queue *q);
995 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
996 struct kfd_dev *dev);
997 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
998 struct kfd_dev *dev);
999 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
1000 struct kfd_dev *dev);
1001 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
1002 struct kfd_dev *dev);
1003 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
1004 struct kfd_dev *dev);
1005 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
1006 struct kfd_dev *dev);
1007 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
1008 void device_queue_manager_uninit(struct device_queue_manager *dqm);
1009 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
1010 enum kfd_queue_type type);
1011 void kernel_queue_uninit(struct kernel_queue *kq, bool hanging);
1012 int kfd_process_vm_fault(struct device_queue_manager *dqm, u32 pasid);
1014 /* Process Queue Manager */
1015 struct process_queue_node {
1017 struct kernel_queue *kq;
1018 struct list_head process_queue_list;
1021 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
1022 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
1023 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
1024 void pqm_uninit(struct process_queue_manager *pqm);
1025 int pqm_create_queue(struct process_queue_manager *pqm,
1026 struct kfd_dev *dev,
1028 struct queue_properties *properties,
1030 uint32_t *p_doorbell_offset_in_process);
1031 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
1032 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
1033 struct queue_properties *p);
1034 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
1035 struct queue_properties *p);
1036 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
1038 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
1040 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1042 int pqm_get_wave_state(struct process_queue_manager *pqm,
1044 void __user *ctl_stack,
1045 u32 *ctl_stack_used_size,
1046 u32 *save_area_used_size);
1048 int amdkfd_fence_wait_timeout(uint64_t *fence_addr,
1049 uint64_t fence_value,
1050 unsigned int timeout_ms);
1052 /* Packet Manager */
1054 #define KFD_FENCE_COMPLETED (100)
1055 #define KFD_FENCE_INIT (10)
1057 struct packet_manager {
1058 struct device_queue_manager *dqm;
1059 struct kernel_queue *priv_queue;
1062 struct kfd_mem_obj *ib_buffer_obj;
1063 unsigned int ib_size_bytes;
1064 bool is_over_subscription;
1066 const struct packet_manager_funcs *pmf;
1069 struct packet_manager_funcs {
1070 /* Support ASIC-specific packet formats for PM4 packets */
1071 int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
1072 struct qcm_process_device *qpd);
1073 int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
1074 uint64_t ib, size_t ib_size_in_dwords, bool chain);
1075 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
1076 struct scheduling_resources *res);
1077 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
1078 struct queue *q, bool is_static);
1079 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
1080 enum kfd_queue_type type,
1081 enum kfd_unmap_queues_filter mode,
1082 uint32_t filter_param, bool reset,
1083 unsigned int sdma_engine);
1084 int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
1085 uint64_t fence_address, uint64_t fence_value);
1086 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
1089 int map_process_size;
1091 int set_resources_size;
1092 int map_queues_size;
1093 int unmap_queues_size;
1094 int query_status_size;
1095 int release_mem_size;
1098 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
1099 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
1100 extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;
1102 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
1103 void pm_uninit(struct packet_manager *pm, bool hanging);
1104 int pm_send_set_resources(struct packet_manager *pm,
1105 struct scheduling_resources *res);
1106 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
1107 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
1108 uint64_t fence_value);
1110 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
1111 enum kfd_unmap_queues_filter mode,
1112 uint32_t filter_param, bool reset,
1113 unsigned int sdma_engine);
1115 void pm_release_ib(struct packet_manager *pm);
1117 /* Following PM funcs can be shared among VI and AI */
1118 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
1120 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
1123 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1124 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1126 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1128 void kfd_event_init_process(struct kfd_process *p);
1129 void kfd_event_free_process(struct kfd_process *p);
1130 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1131 int kfd_wait_on_events(struct kfd_process *p,
1132 uint32_t num_events, void __user *data,
1133 bool all, uint32_t user_timeout_ms,
1134 uint32_t *wait_result);
1135 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
1136 uint32_t valid_id_bits);
1137 void kfd_signal_iommu_event(struct kfd_dev *dev,
1138 u32 pasid, unsigned long address,
1139 bool is_write_requested, bool is_execute_requested);
1140 void kfd_signal_hw_exception_event(u32 pasid);
1141 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1142 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1143 int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
1145 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1146 uint32_t event_type, bool auto_reset, uint32_t node_id,
1147 uint32_t *event_id, uint32_t *event_trigger_data,
1148 uint64_t *event_page_offset, uint32_t *event_slot_index);
1149 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1151 void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
1152 struct kfd_vm_fault_info *info);
1154 void kfd_signal_reset_event(struct kfd_dev *dev);
1156 void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);
1158 void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type);
1160 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
1162 bool kfd_is_locked(void);
1164 /* Compute profile */
1165 void kfd_inc_compute_active(struct kfd_dev *dev);
1166 void kfd_dec_compute_active(struct kfd_dev *dev);
1168 /* Cgroup Support */
1169 /* Check with device cgroup if @kfd device is accessible */
1170 static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)
1172 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1173 struct drm_device *ddev = kfd->ddev;
1175 return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1176 ddev->render->index,
1177 DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1184 #if defined(CONFIG_DEBUG_FS)
1186 void kfd_debugfs_init(void);
1187 void kfd_debugfs_fini(void);
1188 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1189 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1190 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1191 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1192 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1193 int pm_debugfs_runlist(struct seq_file *m, void *data);
1195 int kfd_debugfs_hang_hws(struct kfd_dev *dev);
1196 int pm_debugfs_hang_hws(struct packet_manager *pm);
1197 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm);
1201 static inline void kfd_debugfs_init(void) {}
1202 static inline void kfd_debugfs_fini(void) {}