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 <kgd_kfd_interface.h>
36 #define KFD_SYSFS_FILE_MODE 0444
39 * When working with cp scheduler we should assign the HIQ manually or via
40 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot
41 * definitions for Kaveri. In Kaveri only the first ME queues participates
42 * in the cp scheduling taking that in mind we set the HIQ slot in the
45 #define KFD_CIK_HIQ_PIPE 4
46 #define KFD_CIK_HIQ_QUEUE 0
48 /* GPU ID hash width in bits */
49 #define KFD_GPU_ID_HASH_WIDTH 16
51 /* Macro for allocating structures */
52 #define kfd_alloc_struct(ptr_to_struct) \
53 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
55 /* Kernel module parameter to specify maximum number of supported processes */
56 extern int max_num_of_processes
;
58 #define KFD_MAX_NUM_OF_PROCESSES_DEFAULT 32
59 #define KFD_MAX_NUM_OF_PROCESSES 512
62 * Kernel module parameter to specify maximum number of supported queues
65 extern int max_num_of_queues_per_process
;
67 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS_DEFAULT 128
68 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
70 #define KFD_KERNEL_QUEUE_SIZE 2048
72 /* Kernel module parameter to specify the scheduling policy */
73 extern int sched_policy
;
76 * enum kfd_sched_policy
78 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
79 * scheduling. In this scheduling mode we're using the firmware code to
80 * schedule the user mode queues and kernel queues such as HIQ and DIQ.
81 * the HIQ queue is used as a special queue that dispatches the configuration
82 * to the cp and the user mode queues list that are currently running.
83 * the DIQ queue is a debugging queue that dispatches debugging commands to the
85 * in this scheduling mode user mode queues over subscription feature is
88 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
89 * subscription feature disabled.
91 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
92 * set the command processor registers and sets the queues "manually". This
93 * mode is used *ONLY* for debugging proposes.
96 enum kfd_sched_policy
{
97 KFD_SCHED_POLICY_HWS
= 0,
98 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION
,
99 KFD_SCHED_POLICY_NO_HWS
103 cache_policy_coherent
,
104 cache_policy_noncoherent
107 struct kfd_device_info
{
108 unsigned int max_pasid_bits
;
109 size_t ih_ring_entry_size
;
110 uint16_t mqd_size_aligned
;
116 const struct kfd_device_info
*device_info
;
117 struct pci_dev
*pdev
;
119 unsigned int id
; /* topology stub index */
121 phys_addr_t doorbell_base
; /* Start of actual doorbells used by
122 * KFD. It is aligned for mapping
125 size_t doorbell_id_offset
; /* Doorbell offset (from KFD doorbell
126 * to HW doorbell, GFX reserved some
129 size_t doorbell_process_limit
; /* Number of processes we have doorbell
132 u32 __iomem
*doorbell_kernel_ptr
; /* This is a pointer for a doorbells
133 * page used by kernel queue
136 struct kgd2kfd_shared_resources shared_resources
;
138 void *interrupt_ring
;
139 size_t interrupt_ring_size
;
140 atomic_t interrupt_ring_rptr
;
141 atomic_t interrupt_ring_wptr
;
142 struct work_struct interrupt_work
;
143 spinlock_t interrupt_lock
;
145 /* QCM Device instance */
146 struct device_queue_manager
*dqm
;
150 * Interrupts of interest to KFD are copied
151 * from the HW ring into a SW ring.
153 bool interrupts_active
;
156 /* KGD2KFD callbacks */
157 void kgd2kfd_exit(void);
158 struct kfd_dev
*kgd2kfd_probe(struct kgd_dev
*kgd
, struct pci_dev
*pdev
);
159 bool kgd2kfd_device_init(struct kfd_dev
*kfd
,
160 const struct kgd2kfd_shared_resources
*gpu_resources
);
161 void kgd2kfd_device_exit(struct kfd_dev
*kfd
);
163 extern const struct kfd2kgd_calls
*kfd2kgd
;
172 KFD_MEMPOOL_SYSTEM_CACHEABLE
= 1,
173 KFD_MEMPOOL_SYSTEM_WRITECOMBINE
= 2,
174 KFD_MEMPOOL_FRAMEBUFFER
= 3,
177 /* Character device interface */
178 int kfd_chardev_init(void);
179 void kfd_chardev_exit(void);
180 struct device
*kfd_chardev(void);
183 * enum kfd_preempt_type_filter
185 * @KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE: Preempts single queue.
187 * @KFD_PRERMPT_TYPE_FILTER_ALL_QUEUES: Preempts all queues in the
188 * running queues list.
190 * @KFD_PRERMPT_TYPE_FILTER_BY_PASID: Preempts queues that belongs to
194 enum kfd_preempt_type_filter
{
195 KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE
,
196 KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES
,
197 KFD_PREEMPT_TYPE_FILTER_BY_PASID
200 enum kfd_preempt_type
{
201 KFD_PREEMPT_TYPE_WAVEFRONT
,
202 KFD_PREEMPT_TYPE_WAVEFRONT_RESET
206 * enum kfd_queue_type
208 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
210 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
212 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
214 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
216 enum kfd_queue_type
{
217 KFD_QUEUE_TYPE_COMPUTE
,
223 enum kfd_queue_format
{
224 KFD_QUEUE_FORMAT_PM4
,
229 * struct queue_properties
231 * @type: The queue type.
233 * @queue_id: Queue identifier.
235 * @queue_address: Queue ring buffer address.
237 * @queue_size: Queue ring buffer size.
239 * @priority: Defines the queue priority relative to other queues in the
241 * This is just an indication and HW scheduling may override the priority as
242 * necessary while keeping the relative prioritization.
243 * the priority granularity is from 0 to f which f is the highest priority.
244 * currently all queues are initialized with the highest priority.
246 * @queue_percent: This field is partially implemented and currently a zero in
247 * this field defines that the queue is non active.
249 * @read_ptr: User space address which points to the number of dwords the
250 * cp read from the ring buffer. This field updates automatically by the H/W.
252 * @write_ptr: Defines the number of dwords written to the ring buffer.
254 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
255 * the queue ring buffer. This field should be similar to write_ptr and the user
256 * should update this field after he updated the write_ptr.
258 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
260 * @is_interop: Defines if this is a interop queue. Interop queue means that the
261 * queue can access both graphics and compute resources.
263 * @is_active: Defines if the queue is active or not.
265 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
268 * This structure represents the queue properties for each queue no matter if
269 * it's user mode or kernel mode queue.
272 struct queue_properties
{
273 enum kfd_queue_type type
;
274 enum kfd_queue_format format
;
275 unsigned int queue_id
;
276 uint64_t queue_address
;
279 uint32_t queue_percent
;
282 uint32_t *doorbell_ptr
;
283 uint32_t doorbell_off
;
286 /* Not relevant for user mode queues in cp scheduling */
293 * @list: Queue linked list.
295 * @mqd: The queue MQD.
297 * @mqd_mem_obj: The MQD local gpu memory object.
299 * @gart_mqd_addr: The MQD gart mc address.
301 * @properties: The queue properties.
303 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
304 * that the queue should be execute on.
306 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe id.
308 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
310 * @process: The kfd process that created this queue.
312 * @device: The kfd device that created this queue.
314 * This structure represents user mode compute queues.
315 * It contains all the necessary data to handle such queues.
320 struct list_head list
;
322 struct kfd_mem_obj
*mqd_mem_obj
;
323 uint64_t gart_mqd_addr
;
324 struct queue_properties properties
;
330 struct kfd_process
*process
;
331 struct kfd_dev
*device
;
335 * Please read the kfd_mqd_manager.h description.
338 KFD_MQD_TYPE_CIK_COMPUTE
= 0, /* for no cp scheduling */
339 KFD_MQD_TYPE_CIK_HIQ
, /* for hiq */
340 KFD_MQD_TYPE_CIK_CP
, /* for cp queues and diq */
341 KFD_MQD_TYPE_CIK_SDMA
, /* for sdma queues */
345 struct scheduling_resources
{
346 unsigned int vmid_mask
;
347 enum kfd_queue_type type
;
351 uint32_t gds_heap_base
;
352 uint32_t gds_heap_size
;
355 struct process_queue_manager
{
357 struct kfd_process
*process
;
358 unsigned int num_concurrent_processes
;
359 struct list_head queues
;
360 unsigned long *queue_slot_bitmap
;
363 struct qcm_process_device
{
364 /* The Device Queue Manager that owns this data */
365 struct device_queue_manager
*dqm
;
366 struct process_queue_manager
*pqm
;
367 /* Device Queue Manager lock */
370 struct list_head queues_list
;
371 struct list_head priv_queue_list
;
373 unsigned int queue_count
;
377 * All the memory management data should be here too
379 uint64_t gds_context_area
;
380 uint32_t sh_mem_config
;
381 uint32_t sh_mem_bases
;
382 uint32_t sh_mem_ape1_base
;
383 uint32_t sh_mem_ape1_limit
;
384 uint32_t page_table_base
;
390 /* Data that is per-process-per device. */
391 struct kfd_process_device
{
393 * List of all per-device data for a process.
394 * Starts from kfd_process.per_device_data.
396 struct list_head per_device_list
;
398 /* The device that owns this data. */
402 /* per-process-per device QCM data structure */
403 struct qcm_process_device qpd
;
409 uint64_t gpuvm_limit
;
410 uint64_t scratch_base
;
411 uint64_t scratch_limit
;
413 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
420 * kfd_process are stored in an mm_struct*->kfd_process*
421 * hash table (kfd_processes in kfd_process.c)
423 struct hlist_node kfd_processes
;
425 struct mm_struct
*mm
;
430 * In any process, the thread that started main() is the lead
431 * thread and outlives the rest.
432 * It is here because amd_iommu_bind_pasid wants a task_struct.
434 struct task_struct
*lead_thread
;
436 /* We want to receive a notification when the mm_struct is destroyed */
437 struct mmu_notifier mmu_notifier
;
439 /* Use for delayed freeing of kfd_process structure */
445 * List of kfd_process_device structures,
446 * one for each device the process is using.
448 struct list_head per_device_data
;
450 struct process_queue_manager pqm
;
452 /* The process's queues. */
453 size_t queue_array_size
;
455 /* Size is queue_array_size, up to MAX_PROCESS_QUEUES. */
456 struct kfd_queue
**queues
;
458 unsigned long allocated_queue_bitmap
[DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS
, BITS_PER_LONG
)];
460 /*Is the user space process 32 bit?*/
461 bool is_32bit_user_mode
;
464 void kfd_process_create_wq(void);
465 void kfd_process_destroy_wq(void);
466 struct kfd_process
*kfd_create_process(const struct task_struct
*);
467 struct kfd_process
*kfd_get_process(const struct task_struct
*);
469 struct kfd_process_device
*kfd_bind_process_to_device(struct kfd_dev
*dev
,
470 struct kfd_process
*p
);
471 void kfd_unbind_process_from_device(struct kfd_dev
*dev
, unsigned int pasid
);
472 struct kfd_process_device
*kfd_get_process_device_data(struct kfd_dev
*dev
,
473 struct kfd_process
*p
,
477 int kfd_pasid_init(void);
478 void kfd_pasid_exit(void);
479 bool kfd_set_pasid_limit(unsigned int new_limit
);
480 unsigned int kfd_get_pasid_limit(void);
481 unsigned int kfd_pasid_alloc(void);
482 void kfd_pasid_free(unsigned int pasid
);
485 void kfd_doorbell_init(struct kfd_dev
*kfd
);
486 int kfd_doorbell_mmap(struct kfd_process
*process
, struct vm_area_struct
*vma
);
487 u32 __iomem
*kfd_get_kernel_doorbell(struct kfd_dev
*kfd
,
488 unsigned int *doorbell_off
);
489 void kfd_release_kernel_doorbell(struct kfd_dev
*kfd
, u32 __iomem
*db_addr
);
490 u32
read_kernel_doorbell(u32 __iomem
*db
);
491 void write_kernel_doorbell(u32 __iomem
*db
, u32 value
);
492 unsigned int kfd_queue_id_to_doorbell(struct kfd_dev
*kfd
,
493 struct kfd_process
*process
,
494 unsigned int queue_id
);
496 extern struct device
*kfd_device
;
499 int kfd_topology_init(void);
500 void kfd_topology_shutdown(void);
501 int kfd_topology_add_device(struct kfd_dev
*gpu
);
502 int kfd_topology_remove_device(struct kfd_dev
*gpu
);
503 struct kfd_dev
*kfd_device_by_id(uint32_t gpu_id
);
504 struct kfd_dev
*kfd_device_by_pci_dev(const struct pci_dev
*pdev
);
505 struct kfd_dev
*kfd_topology_enum_kfd_devices(uint8_t idx
);
508 int kfd_interrupt_init(struct kfd_dev
*dev
);
509 void kfd_interrupt_exit(struct kfd_dev
*dev
);
510 void kgd2kfd_interrupt(struct kfd_dev
*kfd
, const void *ih_ring_entry
);
511 bool enqueue_ih_ring_entry(struct kfd_dev
*kfd
, const void *ih_ring_entry
);
513 /* Power Management */
514 void kgd2kfd_suspend(struct kfd_dev
*kfd
);
515 int kgd2kfd_resume(struct kfd_dev
*kfd
);
517 /* amdkfd Apertures */
518 int kfd_init_apertures(struct kfd_process
*process
);
520 /* Queue Context Management */
521 inline uint32_t lower_32(uint64_t x
);
522 inline uint32_t upper_32(uint64_t x
);
524 int init_queue(struct queue
**q
, struct queue_properties properties
);
525 void uninit_queue(struct queue
*q
);
526 void print_queue_properties(struct queue_properties
*q
);
527 void print_queue(struct queue
*q
);
529 struct mqd_manager
*mqd_manager_init(enum KFD_MQD_TYPE type
,
530 struct kfd_dev
*dev
);
531 struct device_queue_manager
*device_queue_manager_init(struct kfd_dev
*dev
);
532 void device_queue_manager_uninit(struct device_queue_manager
*dqm
);
533 struct kernel_queue
*kernel_queue_init(struct kfd_dev
*dev
,
534 enum kfd_queue_type type
);
535 void kernel_queue_uninit(struct kernel_queue
*kq
);
537 /* Process Queue Manager */
538 struct process_queue_node
{
540 struct kernel_queue
*kq
;
541 struct list_head process_queue_list
;
544 int pqm_init(struct process_queue_manager
*pqm
, struct kfd_process
*p
);
545 void pqm_uninit(struct process_queue_manager
*pqm
);
546 int pqm_create_queue(struct process_queue_manager
*pqm
,
549 struct queue_properties
*properties
,
551 enum kfd_queue_type type
,
553 int pqm_destroy_queue(struct process_queue_manager
*pqm
, unsigned int qid
);
554 int pqm_update_queue(struct process_queue_manager
*pqm
, unsigned int qid
,
555 struct queue_properties
*p
);
559 #define KFD_HIQ_TIMEOUT (500)
561 #define KFD_FENCE_COMPLETED (100)
562 #define KFD_FENCE_INIT (10)
563 #define KFD_UNMAP_LATENCY (150)
565 struct packet_manager
{
566 struct device_queue_manager
*dqm
;
567 struct kernel_queue
*priv_queue
;
570 struct kfd_mem_obj
*ib_buffer_obj
;
573 int pm_init(struct packet_manager
*pm
, struct device_queue_manager
*dqm
);
574 void pm_uninit(struct packet_manager
*pm
);
575 int pm_send_set_resources(struct packet_manager
*pm
,
576 struct scheduling_resources
*res
);
577 int pm_send_runlist(struct packet_manager
*pm
, struct list_head
*dqm_queues
);
578 int pm_send_query_status(struct packet_manager
*pm
, uint64_t fence_address
,
579 uint32_t fence_value
);
581 int pm_send_unmap_queue(struct packet_manager
*pm
, enum kfd_queue_type type
,
582 enum kfd_preempt_type_filter mode
,
583 uint32_t filter_param
, bool reset
,
584 unsigned int sdma_engine
);
586 void pm_release_ib(struct packet_manager
*pm
);
588 uint64_t kfd_get_number_elems(struct kfd_dev
*kfd
);
589 phys_addr_t
kfd_get_process_doorbells(struct kfd_dev
*dev
,
590 struct kfd_process
*process
);