brintos

brintos / linux-shallow public Read only

0
0
Text · 46.9 KiB · 26e48fd Raw
1572 lines · c
1/* SPDX-License-Identifier: GPL-2.0 OR MIT */2/*3 * Copyright 2014-2022 Advanced Micro Devices, Inc.4 *5 * Permission is hereby granted, free of charge, to any person obtaining a6 * copy of this software and associated documentation files (the "Software"),7 * to deal in the Software without restriction, including without limitation8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,9 * and/or sell copies of the Software, and to permit persons to whom the10 * Software is furnished to do so, subject to the following conditions:11 *12 * The above copyright notice and this permission notice shall be included in13 * all copies or substantial portions of the Software.14 *15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR21 * OTHER DEALINGS IN THE SOFTWARE.22 */23 24#ifndef KFD_PRIV_H_INCLUDED25#define KFD_PRIV_H_INCLUDED26 27#include <linux/hashtable.h>28#include <linux/mmu_notifier.h>29#include <linux/memremap.h>30#include <linux/mutex.h>31#include <linux/types.h>32#include <linux/atomic.h>33#include <linux/workqueue.h>34#include <linux/spinlock.h>35#include <linux/kfd_ioctl.h>36#include <linux/idr.h>37#include <linux/kfifo.h>38#include <linux/seq_file.h>39#include <linux/kref.h>40#include <linux/sysfs.h>41#include <linux/device_cgroup.h>42#include <drm/drm_file.h>43#include <drm/drm_drv.h>44#include <drm/drm_device.h>45#include <drm/drm_ioctl.h>46#include <kgd_kfd_interface.h>47#include <linux/swap.h>48 49#include "amd_shared.h"50#include "amdgpu.h"51 52#define KFD_MAX_RING_ENTRY_SIZE	853 54#define KFD_SYSFS_FILE_MODE 044455 56/* GPU ID hash width in bits */57#define KFD_GPU_ID_HASH_WIDTH 1658 59/* Use upper bits of mmap offset to store KFD driver specific information.60 * BITS[63:62] - Encode MMAP type61 * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to62 * BITS[45:0]  - MMAP offset value63 *64 * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these65 *  defines are w.r.t to PAGE_SIZE66 */67#define KFD_MMAP_TYPE_SHIFT	6268#define KFD_MMAP_TYPE_MASK	(0x3ULL << KFD_MMAP_TYPE_SHIFT)69#define KFD_MMAP_TYPE_DOORBELL	(0x3ULL << KFD_MMAP_TYPE_SHIFT)70#define KFD_MMAP_TYPE_EVENTS	(0x2ULL << KFD_MMAP_TYPE_SHIFT)71#define KFD_MMAP_TYPE_RESERVED_MEM	(0x1ULL << KFD_MMAP_TYPE_SHIFT)72#define KFD_MMAP_TYPE_MMIO	(0x0ULL << KFD_MMAP_TYPE_SHIFT)73 74#define KFD_MMAP_GPU_ID_SHIFT 4675#define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \76				<< KFD_MMAP_GPU_ID_SHIFT)77#define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\78				& KFD_MMAP_GPU_ID_MASK)79#define KFD_MMAP_GET_GPU_ID(offset)    ((offset & KFD_MMAP_GPU_ID_MASK) \80				>> KFD_MMAP_GPU_ID_SHIFT)81 82/*83 * When working with cp scheduler we should assign the HIQ manually or via84 * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot85 * definitions for Kaveri. In Kaveri only the first ME queues participates86 * in the cp scheduling taking that in mind we set the HIQ slot in the87 * second ME.88 */89#define KFD_CIK_HIQ_PIPE 490#define KFD_CIK_HIQ_QUEUE 091 92/* Macro for allocating structures */93#define kfd_alloc_struct(ptr_to_struct)	\94	((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))95 96#define KFD_MAX_NUM_OF_PROCESSES 51297#define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 102498 99/*100 * Size of the per-process TBA+TMA buffer: 2 pages101 *102 * The first chunk is the TBA used for the CWSR ISA code. The second103 * chunk is used as TMA for user-mode trap handler setup in daisy-chain mode.104 */105#define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)106#define KFD_CWSR_TMA_OFFSET (PAGE_SIZE + 2048)107 108#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE		\109	(KFD_MAX_NUM_OF_PROCESSES *			\110			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)111 112#define KFD_KERNEL_QUEUE_SIZE 2048113 114#define KFD_UNMAP_LATENCY_MS	(4000)115 116#define KFD_MAX_SDMA_QUEUES	128117 118/*119 * 512 = 0x200120 * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the121 * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.122 * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC123 * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in124 * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.125 */126#define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512127 128/**129 * enum kfd_ioctl_flags - KFD ioctl flags130 * Various flags that can be set in &amdkfd_ioctl_desc.flags to control how131 * userspace can use a given ioctl.132 */133enum kfd_ioctl_flags {134	/*135	 * @KFD_IOC_FLAG_CHECKPOINT_RESTORE:136	 * Certain KFD ioctls such as AMDKFD_IOC_CRIU_OP can potentially137	 * perform privileged operations and load arbitrary data into MQDs and138	 * eventually HQD registers when the queue is mapped by HWS. In order to139	 * prevent this we should perform additional security checks.140	 *141	 * This is equivalent to callers with the CHECKPOINT_RESTORE capability.142	 *143	 * Note: Since earlier versions of docker do not support CHECKPOINT_RESTORE,144	 * we also allow ioctls with SYS_ADMIN capability.145	 */146	KFD_IOC_FLAG_CHECKPOINT_RESTORE = BIT(0),147};148/*149 * Kernel module parameter to specify maximum number of supported queues per150 * device151 */152extern int max_num_of_queues_per_device;153 154 155/* Kernel module parameter to specify the scheduling policy */156extern int sched_policy;157 158/*159 * Kernel module parameter to specify the maximum process160 * number per HW scheduler161 */162extern int hws_max_conc_proc;163 164extern int cwsr_enable;165 166/*167 * Kernel module parameter to specify whether to send sigterm to HSA process on168 * unhandled exception169 */170extern int send_sigterm;171 172/*173 * This kernel module is used to simulate large bar machine on non-large bar174 * enabled machines.175 */176extern int debug_largebar;177 178/* Set sh_mem_config.retry_disable on GFX v9 */179extern int amdgpu_noretry;180 181/* Halt if HWS hang is detected */182extern int halt_if_hws_hang;183 184/* Whether MEC FW support GWS barriers */185extern bool hws_gws_support;186 187/* Queue preemption timeout in ms */188extern int queue_preemption_timeout_ms;189 190/*191 * Don't evict process queues on vm fault192 */193extern int amdgpu_no_queue_eviction_on_vm_fault;194 195/* Enable eviction debug messages */196extern bool debug_evictions;197 198extern struct mutex kfd_processes_mutex;199 200enum cache_policy {201	cache_policy_coherent,202	cache_policy_noncoherent203};204 205#define KFD_GC_VERSION(dev) (amdgpu_ip_version((dev)->adev, GC_HWIP, 0))206#define KFD_IS_SOC15(dev)   ((KFD_GC_VERSION(dev)) >= (IP_VERSION(9, 0, 1)))207#define KFD_SUPPORT_XNACK_PER_PROCESS(dev)\208	((KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2)) ||	\209	 (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) ||	\210	 (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 4)))211 212struct kfd_node;213 214struct kfd_event_interrupt_class {215	bool (*interrupt_isr)(struct kfd_node *dev,216			const uint32_t *ih_ring_entry, uint32_t *patched_ihre,217			bool *patched_flag);218	void (*interrupt_wq)(struct kfd_node *dev,219			const uint32_t *ih_ring_entry);220};221 222struct kfd_device_info {223	uint32_t gfx_target_version;224	const struct kfd_event_interrupt_class *event_interrupt_class;225	unsigned int max_pasid_bits;226	unsigned int max_no_of_hqd;227	unsigned int doorbell_size;228	size_t ih_ring_entry_size;229	uint8_t num_of_watch_points;230	uint16_t mqd_size_aligned;231	bool supports_cwsr;232	bool needs_pci_atomics;233	uint32_t no_atomic_fw_version;234	unsigned int num_sdma_queues_per_engine;235	unsigned int num_reserved_sdma_queues_per_engine;236	DECLARE_BITMAP(reserved_sdma_queues_bitmap, KFD_MAX_SDMA_QUEUES);237};238 239unsigned int kfd_get_num_sdma_engines(struct kfd_node *kdev);240unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *kdev);241 242struct kfd_mem_obj {243	uint32_t range_start;244	uint32_t range_end;245	uint64_t gpu_addr;246	uint32_t *cpu_ptr;247	void *gtt_mem;248};249 250struct kfd_vmid_info {251	uint32_t first_vmid_kfd;252	uint32_t last_vmid_kfd;253	uint32_t vmid_num_kfd;254};255 256#define MAX_KFD_NODES	8257 258struct kfd_dev;259 260struct kfd_node {261	unsigned int node_id;262	struct amdgpu_device *adev;     /* Duplicated here along with keeping263					 * a copy in kfd_dev to save a hop264					 */265	const struct kfd2kgd_calls *kfd2kgd; /* Duplicated here along with266					      * keeping a copy in kfd_dev to267					      * save a hop268					      */269	struct kfd_vmid_info vm_info;270	unsigned int id;                /* topology stub index */271	uint32_t xcc_mask; /* Instance mask of XCCs present */272	struct amdgpu_xcp *xcp;273 274	/* Interrupts */275	struct kfifo ih_fifo;276	struct workqueue_struct *ih_wq;277	struct work_struct interrupt_work;278	spinlock_t interrupt_lock;279 280	/*281	 * Interrupts of interest to KFD are copied282	 * from the HW ring into a SW ring.283	 */284	bool interrupts_active;285	uint32_t interrupt_bitmap; /* Only used for GFX 9.4.3 */286 287	/* QCM Device instance */288	struct device_queue_manager *dqm;289 290	/* Global GWS resource shared between processes */291	void *gws;292	bool gws_debug_workaround;293 294	/* Clients watching SMI events */295	struct list_head smi_clients;296	spinlock_t smi_lock;297	uint32_t reset_seq_num;298 299	/* SRAM ECC flag */300	atomic_t sram_ecc_flag;301 302	/*spm process id */303	unsigned int spm_pasid;304 305	/* Maximum process number mapped to HW scheduler */306	unsigned int max_proc_per_quantum;307 308	unsigned int compute_vmid_bitmap;309 310	struct kfd_local_mem_info local_mem_info;311 312	struct kfd_dev *kfd;313 314	/* Track per device allocated watch points */315	uint32_t alloc_watch_ids;316	spinlock_t watch_points_lock;317};318 319struct kfd_dev {320	struct amdgpu_device *adev;321 322	struct kfd_device_info device_info;323 324	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells325					   * page used by kernel queue326					   */327 328	struct kgd2kfd_shared_resources shared_resources;329 330	const struct kfd2kgd_calls *kfd2kgd;331	struct mutex doorbell_mutex;332 333	void *gtt_mem;334	uint64_t gtt_start_gpu_addr;335	void *gtt_start_cpu_ptr;336	void *gtt_sa_bitmap;337	struct mutex gtt_sa_lock;338	unsigned int gtt_sa_chunk_size;339	unsigned int gtt_sa_num_of_chunks;340 341	bool init_complete;342 343	/* Firmware versions */344	uint16_t mec_fw_version;345	uint16_t mec2_fw_version;346	uint16_t sdma_fw_version;347 348	/* CWSR */349	bool cwsr_enabled;350	const void *cwsr_isa;351	unsigned int cwsr_isa_size;352 353	/* xGMI */354	uint64_t hive_id;355 356	bool pci_atomic_requested;357 358	/* Compute Profile ref. count */359	atomic_t compute_profile;360 361	struct ida doorbell_ida;362	unsigned int max_doorbell_slices;363 364	int noretry;365 366	struct kfd_node *nodes[MAX_KFD_NODES];367	unsigned int num_nodes;368 369	/* Kernel doorbells for KFD device */370	struct amdgpu_bo *doorbells;371 372	/* bitmap for dynamic doorbell allocation from doorbell object */373	unsigned long *doorbell_bitmap;374};375 376enum kfd_mempool {377	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,378	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,379	KFD_MEMPOOL_FRAMEBUFFER = 3,380};381 382/* Character device interface */383int kfd_chardev_init(void);384void kfd_chardev_exit(void);385 386/**387 * enum kfd_unmap_queues_filter - Enum for queue filters.388 *389 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the390 *						running queues list.391 *392 * @KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: Preempts all non-static queues393 *						in the run list.394 *395 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to396 *						specific process.397 *398 */399enum kfd_unmap_queues_filter {400	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES = 1,401	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES = 2,402	KFD_UNMAP_QUEUES_FILTER_BY_PASID = 3403};404 405/**406 * enum kfd_queue_type - Enum for various queue types.407 *408 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.409 *410 * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.411 *412 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.413 *414 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.415 *416 * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.417 *418 * @KFD_QUEUE_TYPE_SDMA_BY_ENG_ID:  SDMA user mode queue with target SDMA engine ID.419 */420enum kfd_queue_type  {421	KFD_QUEUE_TYPE_COMPUTE,422	KFD_QUEUE_TYPE_SDMA,423	KFD_QUEUE_TYPE_HIQ,424	KFD_QUEUE_TYPE_DIQ,425	KFD_QUEUE_TYPE_SDMA_XGMI,426	KFD_QUEUE_TYPE_SDMA_BY_ENG_ID427};428 429enum kfd_queue_format {430	KFD_QUEUE_FORMAT_PM4,431	KFD_QUEUE_FORMAT_AQL432};433 434enum KFD_QUEUE_PRIORITY {435	KFD_QUEUE_PRIORITY_MINIMUM = 0,436	KFD_QUEUE_PRIORITY_MAXIMUM = 15437};438 439/**440 * struct queue_properties441 *442 * @type: The queue type.443 *444 * @queue_id: Queue identifier.445 *446 * @queue_address: Queue ring buffer address.447 *448 * @queue_size: Queue ring buffer size.449 *450 * @priority: Defines the queue priority relative to other queues in the451 * process.452 * This is just an indication and HW scheduling may override the priority as453 * necessary while keeping the relative prioritization.454 * the priority granularity is from 0 to f which f is the highest priority.455 * currently all queues are initialized with the highest priority.456 *457 * @queue_percent: This field is partially implemented and currently a zero in458 * this field defines that the queue is non active.459 *460 * @read_ptr: User space address which points to the number of dwords the461 * cp read from the ring buffer. This field updates automatically by the H/W.462 *463 * @write_ptr: Defines the number of dwords written to the ring buffer.464 *465 * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring466 * buffer. This field should be similar to write_ptr and the user should467 * update this field after updating the write_ptr.468 *469 * @doorbell_off: The doorbell offset in the doorbell pci-bar.470 *471 * @is_interop: Defines if this is a interop queue. Interop queue means that472 * the queue can access both graphics and compute resources.473 *474 * @is_evicted: Defines if the queue is evicted. Only active queues475 * are evicted, rendering them inactive.476 *477 * @is_active: Defines if the queue is active or not. @is_active and478 * @is_evicted are protected by the DQM lock.479 *480 * @is_gws: Defines if the queue has been updated to be GWS-capable or not.481 * @is_gws should be protected by the DQM lock, since changing it can yield the482 * possibility of updating DQM state on number of GWS queues.483 *484 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid485 * of the queue.486 *487 * This structure represents the queue properties for each queue no matter if488 * it's user mode or kernel mode queue.489 *490 */491 492struct queue_properties {493	enum kfd_queue_type type;494	enum kfd_queue_format format;495	unsigned int queue_id;496	uint64_t queue_address;497	uint64_t  queue_size;498	uint32_t priority;499	uint32_t queue_percent;500	void __user *read_ptr;501	void __user *write_ptr;502	void __iomem *doorbell_ptr;503	uint32_t doorbell_off;504	bool is_interop;505	bool is_evicted;506	bool is_suspended;507	bool is_being_destroyed;508	bool is_active;509	bool is_gws;510	uint32_t pm4_target_xcc;511	bool is_dbg_wa;512	bool is_user_cu_masked;513	/* Not relevant for user mode queues in cp scheduling */514	unsigned int vmid;515	/* Relevant only for sdma queues*/516	uint32_t sdma_engine_id;517	uint32_t sdma_queue_id;518	uint32_t sdma_vm_addr;519	/* Relevant only for VI */520	uint64_t eop_ring_buffer_address;521	uint32_t eop_ring_buffer_size;522	uint64_t ctx_save_restore_area_address;523	uint32_t ctx_save_restore_area_size;524	uint32_t ctl_stack_size;525	uint64_t tba_addr;526	uint64_t tma_addr;527	uint64_t exception_status;528 529	struct amdgpu_bo *wptr_bo;530	struct amdgpu_bo *rptr_bo;531	struct amdgpu_bo *ring_bo;532	struct amdgpu_bo *eop_buf_bo;533	struct amdgpu_bo *cwsr_bo;534};535 536#define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\537			    (q).queue_address != 0 &&	\538			    (q).queue_percent > 0 &&	\539			    !(q).is_evicted &&		\540			    !(q).is_suspended)541 542enum mqd_update_flag {543	UPDATE_FLAG_DBG_WA_ENABLE = 1,544	UPDATE_FLAG_DBG_WA_DISABLE = 2,545	UPDATE_FLAG_IS_GWS = 4, /* quirk for gfx9 IP */546};547 548struct mqd_update_info {549	union {550		struct {551			uint32_t count; /* Must be a multiple of 32 */552			uint32_t *ptr;553		} cu_mask;554	};555	enum mqd_update_flag update_flag;556};557 558/**559 * struct queue560 *561 * @list: Queue linked list.562 *563 * @mqd: The queue MQD (memory queue descriptor).564 *565 * @mqd_mem_obj: The MQD local gpu memory object.566 *567 * @gart_mqd_addr: The MQD gart mc address.568 *569 * @properties: The queue properties.570 *571 * @mec: Used only in no cp scheduling mode and identifies to micro engine id572 *	 that the queue should be executed on.573 *574 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe575 *	  id.576 *577 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.578 *579 * @process: The kfd process that created this queue.580 *581 * @device: The kfd device that created this queue.582 *583 * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL584 * otherwise.585 *586 * This structure represents user mode compute queues.587 * It contains all the necessary data to handle such queues.588 *589 */590 591struct queue {592	struct list_head list;593	void *mqd;594	struct kfd_mem_obj *mqd_mem_obj;595	uint64_t gart_mqd_addr;596	struct queue_properties properties;597 598	uint32_t mec;599	uint32_t pipe;600	uint32_t queue;601 602	unsigned int sdma_id;603	unsigned int doorbell_id;604 605	struct kfd_process	*process;606	struct kfd_node		*device;607	void *gws;608 609	/* procfs */610	struct kobject kobj;611 612	void *gang_ctx_bo;613	uint64_t gang_ctx_gpu_addr;614	void *gang_ctx_cpu_ptr;615 616	struct amdgpu_bo *wptr_bo_gart;617};618 619enum KFD_MQD_TYPE {620	KFD_MQD_TYPE_HIQ = 0,		/* for hiq */621	KFD_MQD_TYPE_CP,		/* for cp queues and diq */622	KFD_MQD_TYPE_SDMA,		/* for sdma queues */623	KFD_MQD_TYPE_DIQ,		/* for diq */624	KFD_MQD_TYPE_MAX625};626 627enum KFD_PIPE_PRIORITY {628	KFD_PIPE_PRIORITY_CS_LOW = 0,629	KFD_PIPE_PRIORITY_CS_MEDIUM,630	KFD_PIPE_PRIORITY_CS_HIGH631};632 633struct scheduling_resources {634	unsigned int vmid_mask;635	enum kfd_queue_type type;636	uint64_t queue_mask;637	uint64_t gws_mask;638	uint32_t oac_mask;639	uint32_t gds_heap_base;640	uint32_t gds_heap_size;641};642 643struct process_queue_manager {644	/* data */645	struct kfd_process	*process;646	struct list_head	queues;647	unsigned long		*queue_slot_bitmap;648};649 650struct qcm_process_device {651	/* The Device Queue Manager that owns this data */652	struct device_queue_manager *dqm;653	struct process_queue_manager *pqm;654	/* Queues list */655	struct list_head queues_list;656	struct list_head priv_queue_list;657 658	unsigned int queue_count;659	unsigned int vmid;660	bool is_debug;661	unsigned int evicted; /* eviction counter, 0=active */662 663	/* This flag tells if we should reset all wavefronts on664	 * process termination665	 */666	bool reset_wavefronts;667 668	/* This flag tells us if this process has a GWS-capable669	 * queue that will be mapped into the runlist. It's670	 * possible to request a GWS BO, but not have the queue671	 * currently mapped, and this changes how the MAP_PROCESS672	 * PM4 packet is configured.673	 */674	bool mapped_gws_queue;675 676	/* All the memory management data should be here too */677	uint64_t gds_context_area;678	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */679	uint64_t page_table_base;680	uint32_t sh_mem_config;681	uint32_t sh_mem_bases;682	uint32_t sh_mem_ape1_base;683	uint32_t sh_mem_ape1_limit;684	uint32_t gds_size;685	uint32_t num_gws;686	uint32_t num_oac;687	uint32_t sh_hidden_private_base;688 689	/* CWSR memory */690	struct kgd_mem *cwsr_mem;691	void *cwsr_kaddr;692	uint64_t cwsr_base;693	uint64_t tba_addr;694	uint64_t tma_addr;695 696	/* IB memory */697	struct kgd_mem *ib_mem;698	uint64_t ib_base;699	void *ib_kaddr;700 701	/* doorbells for kfd process */702	struct amdgpu_bo *proc_doorbells;703 704	/* bitmap for dynamic doorbell allocation from the bo */705	unsigned long *doorbell_bitmap;706};707 708/* KFD Memory Eviction */709 710/* Approx. wait time before attempting to restore evicted BOs */711#define PROCESS_RESTORE_TIME_MS 100712/* Approx. back off time if restore fails due to lack of memory */713#define PROCESS_BACK_OFF_TIME_MS 100714/* Approx. time before evicting the process again */715#define PROCESS_ACTIVE_TIME_MS 10716 717/* 8 byte handle containing GPU ID in the most significant 4 bytes and718 * idr_handle in the least significant 4 bytes719 */720#define MAKE_HANDLE(gpu_id, idr_handle) \721	(((uint64_t)(gpu_id) << 32) + idr_handle)722#define GET_GPU_ID(handle) (handle >> 32)723#define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)724 725enum kfd_pdd_bound {726	PDD_UNBOUND = 0,727	PDD_BOUND,728	PDD_BOUND_SUSPENDED,729};730 731#define MAX_SYSFS_FILENAME_LEN 15732 733/*734 * SDMA counter runs at 100MHz frequency.735 * We display SDMA activity in microsecond granularity in sysfs.736 * As a result, the divisor is 100.737 */738#define SDMA_ACTIVITY_DIVISOR  100739 740/* Data that is per-process-per device. */741struct kfd_process_device {742	/* The device that owns this data. */743	struct kfd_node *dev;744 745	/* The process that owns this kfd_process_device. */746	struct kfd_process *process;747 748	/* per-process-per device QCM data structure */749	struct qcm_process_device qpd;750 751	/*Apertures*/752	uint64_t lds_base;753	uint64_t lds_limit;754	uint64_t gpuvm_base;755	uint64_t gpuvm_limit;756	uint64_t scratch_base;757	uint64_t scratch_limit;758 759	/* VM context for GPUVM allocations */760	struct file *drm_file;761	void *drm_priv;762 763	/* GPUVM allocations storage */764	struct idr alloc_idr;765 766	/* Flag used to tell the pdd has dequeued from the dqm.767	 * This is used to prevent dev->dqm->ops.process_termination() from768	 * being called twice when it is already called in IOMMU callback769	 * function.770	 */771	bool already_dequeued;772	bool runtime_inuse;773 774	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */775	enum kfd_pdd_bound bound;776 777	/* VRAM usage */778	atomic64_t vram_usage;779	struct attribute attr_vram;780	char vram_filename[MAX_SYSFS_FILENAME_LEN];781 782	/* SDMA activity tracking */783	uint64_t sdma_past_activity_counter;784	struct attribute attr_sdma;785	char sdma_filename[MAX_SYSFS_FILENAME_LEN];786 787	/* Eviction activity tracking */788	uint64_t last_evict_timestamp;789	atomic64_t evict_duration_counter;790	struct attribute attr_evict;791 792	struct kobject *kobj_stats;793 794	/*795	 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process796	 * that is associated with device encoded by "this" struct instance. The797	 * value reflects CU usage by all of the waves launched by this process798	 * on this device. A very important property of occupancy parameter is799	 * that its value is a snapshot of current use.800	 *801	 * Following is to be noted regarding how this parameter is reported:802	 *803	 *  The number of waves that a CU can launch is limited by couple of804	 *  parameters. These are encoded by struct amdgpu_cu_info instance805	 *  that is part of every device definition. For GFX9 devices this806	 *  translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves807	 *  do not use scratch memory and 32 waves (max_scratch_slots_per_cu)808	 *  when they do use scratch memory. This could change for future809	 *  devices and therefore this example should be considered as a guide.810	 *811	 *  All CU's of a device are available for the process. This may not be true812	 *  under certain conditions - e.g. CU masking.813	 *814	 *  Finally number of CU's that are occupied by a process is affected by both815	 *  number of CU's a device has along with number of other competing processes816	 */817	struct attribute attr_cu_occupancy;818 819	/* sysfs counters for GPU retry fault and page migration tracking */820	struct kobject *kobj_counters;821	struct attribute attr_faults;822	struct attribute attr_page_in;823	struct attribute attr_page_out;824	uint64_t faults;825	uint64_t page_in;826	uint64_t page_out;827 828	/* Exception code status*/829	uint64_t exception_status;830	void *vm_fault_exc_data;831	size_t vm_fault_exc_data_size;832 833	/* Tracks debug per-vmid request settings */834	uint32_t spi_dbg_override;835	uint32_t spi_dbg_launch_mode;836	uint32_t watch_points[4];837	uint32_t alloc_watch_ids;838 839	/*840	 * If this process has been checkpointed before, then the user841	 * application will use the original gpu_id on the842	 * checkpointed node to refer to this device.843	 */844	uint32_t user_gpu_id;845 846	void *proc_ctx_bo;847	uint64_t proc_ctx_gpu_addr;848	void *proc_ctx_cpu_ptr;849 850	/* Tracks queue reset status */851	bool has_reset_queue;852};853 854#define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)855 856struct svm_range_list {857	struct mutex			lock;858	struct rb_root_cached		objects;859	struct list_head		list;860	struct work_struct		deferred_list_work;861	struct list_head		deferred_range_list;862	struct list_head                criu_svm_metadata_list;863	spinlock_t			deferred_list_lock;864	atomic_t			evicted_ranges;865	atomic_t			drain_pagefaults;866	struct delayed_work		restore_work;867	DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);868	struct task_struct		*faulting_task;869	/* check point ts decides if page fault recovery need be dropped */870	uint64_t			checkpoint_ts[MAX_GPU_INSTANCE];871 872	/* Default granularity to use in buffer migration873	 * and restoration of backing memory while handling874	 * recoverable page faults875	 */876	uint8_t default_granularity;877};878 879/* Process data */880struct kfd_process {881	/*882	 * kfd_process are stored in an mm_struct*->kfd_process*883	 * hash table (kfd_processes in kfd_process.c)884	 */885	struct hlist_node kfd_processes;886 887	/*888	 * Opaque pointer to mm_struct. We don't hold a reference to889	 * it so it should never be dereferenced from here. This is890	 * only used for looking up processes by their mm.891	 */892	void *mm;893 894	struct kref ref;895	struct work_struct release_work;896 897	struct mutex mutex;898 899	/*900	 * In any process, the thread that started main() is the lead901	 * thread and outlives the rest.902	 * It is here because amd_iommu_bind_pasid wants a task_struct.903	 * It can also be used for safely getting a reference to the904	 * mm_struct of the process.905	 */906	struct task_struct *lead_thread;907 908	/* We want to receive a notification when the mm_struct is destroyed */909	struct mmu_notifier mmu_notifier;910 911	u32 pasid;912 913	/*914	 * Array of kfd_process_device pointers,915	 * one for each device the process is using.916	 */917	struct kfd_process_device *pdds[MAX_GPU_INSTANCE];918	uint32_t n_pdds;919 920	struct process_queue_manager pqm;921 922	/*Is the user space process 32 bit?*/923	bool is_32bit_user_mode;924 925	/* Event-related data */926	struct mutex event_mutex;927	/* Event ID allocator and lookup */928	struct idr event_idr;929	/* Event page */930	u64 signal_handle;931	struct kfd_signal_page *signal_page;932	size_t signal_mapped_size;933	size_t signal_event_count;934	bool signal_event_limit_reached;935 936	/* Information used for memory eviction */937	void *kgd_process_info;938	/* Eviction fence that is attached to all the BOs of this process. The939	 * fence will be triggered during eviction and new one will be created940	 * during restore941	 */942	struct dma_fence __rcu *ef;943 944	/* Work items for evicting and restoring BOs */945	struct delayed_work eviction_work;946	struct delayed_work restore_work;947	/* seqno of the last scheduled eviction */948	unsigned int last_eviction_seqno;949	/* Approx. the last timestamp (in jiffies) when the process was950	 * restored after an eviction951	 */952	unsigned long last_restore_timestamp;953 954	/* Indicates device process is debug attached with reserved vmid. */955	bool debug_trap_enabled;956 957	/* per-process-per device debug event fd file */958	struct file *dbg_ev_file;959 960	/* If the process is a kfd debugger, we need to know so we can clean961	 * up at exit time.  If a process enables debugging on itself, it does962	 * its own clean-up, so we don't set the flag here.  We track this by963	 * counting the number of processes this process is debugging.964	 */965	atomic_t debugged_process_count;966 967	/* If the process is a debugged, this is the debugger process */968	struct kfd_process *debugger_process;969 970	/* Kobj for our procfs */971	struct kobject *kobj;972	struct kobject *kobj_queues;973	struct attribute attr_pasid;974 975	/* Keep track cwsr init */976	bool has_cwsr;977 978	/* Exception code enable mask and status */979	uint64_t exception_enable_mask;980	uint64_t exception_status;981 982	/* Used to drain stale interrupts */983	wait_queue_head_t wait_irq_drain;984	bool irq_drain_is_open;985 986	/* shared virtual memory registered by this process */987	struct svm_range_list svms;988 989	bool xnack_enabled;990 991	/* Work area for debugger event writer worker. */992	struct work_struct debug_event_workarea;993 994	/* Tracks debug per-vmid request for debug flags */995	u32 dbg_flags;996 997	atomic_t poison;998	/* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */999	bool queues_paused;1000 1001	/* Tracks runtime enable status */1002	struct semaphore runtime_enable_sema;1003	bool is_runtime_retry;1004	struct kfd_runtime_info runtime_info;1005};1006 1007#define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */1008extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);1009extern struct srcu_struct kfd_processes_srcu;1010 1011/**1012 * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.1013 *1014 * @filep: pointer to file structure.1015 * @p: amdkfd process pointer.1016 * @data: pointer to arg that was copied from user.1017 *1018 * Return: returns ioctl completion code.1019 */1020typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,1021				void *data);1022 1023struct amdkfd_ioctl_desc {1024	unsigned int cmd;1025	int flags;1026	amdkfd_ioctl_t *func;1027	unsigned int cmd_drv;1028	const char *name;1029};1030bool kfd_dev_is_large_bar(struct kfd_node *dev);1031 1032int kfd_process_create_wq(void);1033void kfd_process_destroy_wq(void);1034void kfd_cleanup_processes(void);1035struct kfd_process *kfd_create_process(struct task_struct *thread);1036struct kfd_process *kfd_get_process(const struct task_struct *task);1037struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid);1038struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);1039 1040int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);1041int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,1042				uint32_t *gpuid, uint32_t *gpuidx);1043static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,1044				uint32_t gpuidx, uint32_t *gpuid) {1045	return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;1046}1047static inline struct kfd_process_device *kfd_process_device_from_gpuidx(1048				struct kfd_process *p, uint32_t gpuidx) {1049	return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;1050}1051 1052void kfd_unref_process(struct kfd_process *p);1053int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger);1054int kfd_process_restore_queues(struct kfd_process *p);1055void kfd_suspend_all_processes(void);1056int kfd_resume_all_processes(void);1057 1058struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *process,1059							 uint32_t gpu_id);1060 1061int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id);1062 1063int kfd_process_device_init_vm(struct kfd_process_device *pdd,1064			       struct file *drm_file);1065struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,1066						struct kfd_process *p);1067struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,1068							struct kfd_process *p);1069struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,1070							struct kfd_process *p);1071 1072bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);1073 1074int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,1075			  struct vm_area_struct *vma);1076 1077/* KFD process API for creating and translating handles */1078int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,1079					void *mem);1080void *kfd_process_device_translate_handle(struct kfd_process_device *p,1081					int handle);1082void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,1083					int handle);1084struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid);1085 1086/* PASIDs */1087int kfd_pasid_init(void);1088void kfd_pasid_exit(void);1089bool kfd_set_pasid_limit(unsigned int new_limit);1090unsigned int kfd_get_pasid_limit(void);1091u32 kfd_pasid_alloc(void);1092void kfd_pasid_free(u32 pasid);1093 1094/* Doorbells */1095size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);1096int kfd_doorbell_init(struct kfd_dev *kfd);1097void kfd_doorbell_fini(struct kfd_dev *kfd);1098int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process,1099		      struct vm_area_struct *vma);1100void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,1101					unsigned int *doorbell_off);1102void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);1103u32 read_kernel_doorbell(u32 __iomem *db);1104void write_kernel_doorbell(void __iomem *db, u32 value);1105void write_kernel_doorbell64(void __iomem *db, u64 value);1106unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,1107					struct kfd_process_device *pdd,1108					unsigned int doorbell_id);1109phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);1110int kfd_alloc_process_doorbells(struct kfd_dev *kfd,1111				struct kfd_process_device *pdd);1112void kfd_free_process_doorbells(struct kfd_dev *kfd,1113				struct kfd_process_device *pdd);1114/* GTT Sub-Allocator */1115 1116int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size,1117			struct kfd_mem_obj **mem_obj);1118 1119int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj);1120 1121extern struct device *kfd_device;1122 1123/* KFD's procfs */1124void kfd_procfs_init(void);1125void kfd_procfs_shutdown(void);1126int kfd_procfs_add_queue(struct queue *q);1127void kfd_procfs_del_queue(struct queue *q);1128 1129/* Topology */1130int kfd_topology_init(void);1131void kfd_topology_shutdown(void);1132int kfd_topology_add_device(struct kfd_node *gpu);1133int kfd_topology_remove_device(struct kfd_node *gpu);1134struct kfd_topology_device *kfd_topology_device_by_proximity_domain(1135						uint32_t proximity_domain);1136struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(1137						uint32_t proximity_domain);1138struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);1139struct kfd_node *kfd_device_by_id(uint32_t gpu_id);1140struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev);1141static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t node_id,1142					uint32_t vmid)1143{1144	return (node->interrupt_bitmap & (1 << node_id)) != 0 &&1145	       (node->compute_vmid_bitmap & (1 << vmid)) != 0;1146}1147static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev,1148					uint32_t node_id, uint32_t vmid) {1149	struct kfd_dev *dev = adev->kfd.dev;1150	uint32_t i;1151 1152	if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) &&1153	    KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 4))1154		return dev->nodes[0];1155 1156	for (i = 0; i < dev->num_nodes; i++)1157		if (kfd_irq_is_from_node(dev->nodes[i], node_id, vmid))1158			return dev->nodes[i];1159 1160	return NULL;1161}1162int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev);1163int kfd_numa_node_to_apic_id(int numa_node_id);1164 1165/* Interrupts */1166#define	KFD_IRQ_FENCE_CLIENTID	0xff1167#define	KFD_IRQ_FENCE_SOURCEID	0xff1168#define	KFD_IRQ_IS_FENCE(client, source)				\1169				((client) == KFD_IRQ_FENCE_CLIENTID &&	\1170				(source) == KFD_IRQ_FENCE_SOURCEID)1171int kfd_interrupt_init(struct kfd_node *dev);1172void kfd_interrupt_exit(struct kfd_node *dev);1173bool enqueue_ih_ring_entry(struct kfd_node *kfd, const void *ih_ring_entry);1174bool interrupt_is_wanted(struct kfd_node *dev,1175				const uint32_t *ih_ring_entry,1176				uint32_t *patched_ihre, bool *flag);1177int kfd_process_drain_interrupts(struct kfd_process_device *pdd);1178void kfd_process_close_interrupt_drain(unsigned int pasid);1179 1180/* amdkfd Apertures */1181int kfd_init_apertures(struct kfd_process *process);1182 1183void kfd_process_set_trap_handler(struct qcm_process_device *qpd,1184				  uint64_t tba_addr,1185				  uint64_t tma_addr);1186void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,1187				     bool enabled);1188 1189/* CWSR initialization */1190int kfd_process_init_cwsr_apu(struct kfd_process *process, struct file *filep);1191 1192/* CRIU */1193/*1194 * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private1195 * structures:1196 * kfd_criu_process_priv_data1197 * kfd_criu_device_priv_data1198 * kfd_criu_bo_priv_data1199 * kfd_criu_queue_priv_data1200 * kfd_criu_event_priv_data1201 * kfd_criu_svm_range_priv_data1202 */1203 1204#define KFD_CRIU_PRIV_VERSION 11205 1206struct kfd_criu_process_priv_data {1207	uint32_t version;1208	uint32_t xnack_mode;1209};1210 1211struct kfd_criu_device_priv_data {1212	/* For future use */1213	uint64_t reserved;1214};1215 1216struct kfd_criu_bo_priv_data {1217	uint64_t user_addr;1218	uint32_t idr_handle;1219	uint32_t mapped_gpuids[MAX_GPU_INSTANCE];1220};1221 1222/*1223 * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data,1224 * kfd_criu_svm_range_priv_data is the object type1225 */1226enum kfd_criu_object_type {1227	KFD_CRIU_OBJECT_TYPE_QUEUE,1228	KFD_CRIU_OBJECT_TYPE_EVENT,1229	KFD_CRIU_OBJECT_TYPE_SVM_RANGE,1230};1231 1232struct kfd_criu_svm_range_priv_data {1233	uint32_t object_type;1234	uint64_t start_addr;1235	uint64_t size;1236	/* Variable length array of attributes */1237	struct kfd_ioctl_svm_attribute attrs[];1238};1239 1240struct kfd_criu_queue_priv_data {1241	uint32_t object_type;1242	uint64_t q_address;1243	uint64_t q_size;1244	uint64_t read_ptr_addr;1245	uint64_t write_ptr_addr;1246	uint64_t doorbell_off;1247	uint64_t eop_ring_buffer_address;1248	uint64_t ctx_save_restore_area_address;1249	uint32_t gpu_id;1250	uint32_t type;1251	uint32_t format;1252	uint32_t q_id;1253	uint32_t priority;1254	uint32_t q_percent;1255	uint32_t doorbell_id;1256	uint32_t gws;1257	uint32_t sdma_id;1258	uint32_t eop_ring_buffer_size;1259	uint32_t ctx_save_restore_area_size;1260	uint32_t ctl_stack_size;1261	uint32_t mqd_size;1262};1263 1264struct kfd_criu_event_priv_data {1265	uint32_t object_type;1266	uint64_t user_handle;1267	uint32_t event_id;1268	uint32_t auto_reset;1269	uint32_t type;1270	uint32_t signaled;1271 1272	union {1273		struct kfd_hsa_memory_exception_data memory_exception_data;1274		struct kfd_hsa_hw_exception_data hw_exception_data;1275	};1276};1277 1278int kfd_process_get_queue_info(struct kfd_process *p,1279			       uint32_t *num_queues,1280			       uint64_t *priv_data_sizes);1281 1282int kfd_criu_checkpoint_queues(struct kfd_process *p,1283			 uint8_t __user *user_priv_data,1284			 uint64_t *priv_data_offset);1285 1286int kfd_criu_restore_queue(struct kfd_process *p,1287			   uint8_t __user *user_priv_data,1288			   uint64_t *priv_data_offset,1289			   uint64_t max_priv_data_size);1290 1291int kfd_criu_checkpoint_events(struct kfd_process *p,1292			 uint8_t __user *user_priv_data,1293			 uint64_t *priv_data_offset);1294 1295int kfd_criu_restore_event(struct file *devkfd,1296			   struct kfd_process *p,1297			   uint8_t __user *user_priv_data,1298			   uint64_t *priv_data_offset,1299			   uint64_t max_priv_data_size);1300/* CRIU - End */1301 1302/* Queue Context Management */1303int init_queue(struct queue **q, const struct queue_properties *properties);1304void uninit_queue(struct queue *q);1305void print_queue_properties(struct queue_properties *q);1306void print_queue(struct queue *q);1307int kfd_queue_buffer_get(struct amdgpu_vm *vm, void __user *addr, struct amdgpu_bo **pbo,1308			 u64 expected_size);1309void kfd_queue_buffer_put(struct amdgpu_bo **bo);1310int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);1311int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_properties *properties);1312void kfd_queue_unref_bo_va(struct amdgpu_vm *vm, struct amdgpu_bo **bo);1313int kfd_queue_unref_bo_vas(struct kfd_process_device *pdd,1314			   struct queue_properties *properties);1315void kfd_queue_ctx_save_restore_size(struct kfd_topology_device *dev);1316 1317struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,1318		struct kfd_node *dev);1319struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,1320		struct kfd_node *dev);1321struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,1322		struct kfd_node *dev);1323struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,1324		struct kfd_node *dev);1325struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type,1326		struct kfd_node *dev);1327struct mqd_manager *mqd_manager_init_v12(enum KFD_MQD_TYPE type,1328		struct kfd_node *dev);1329struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev);1330void device_queue_manager_uninit(struct device_queue_manager *dqm);1331struct kernel_queue *kernel_queue_init(struct kfd_node *dev,1332					enum kfd_queue_type type);1333void kernel_queue_uninit(struct kernel_queue *kq);1334int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid);1335int kfd_dqm_suspend_bad_queue_mes(struct kfd_node *knode, u32 pasid, u32 doorbell_id);1336 1337/* Process Queue Manager */1338struct process_queue_node {1339	struct queue *q;1340	struct kernel_queue *kq;1341	struct list_head process_queue_list;1342};1343 1344void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);1345void kfd_process_dequeue_from_all_devices(struct kfd_process *p);1346int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);1347void pqm_uninit(struct process_queue_manager *pqm);1348int pqm_create_queue(struct process_queue_manager *pqm,1349			    struct kfd_node *dev,1350			    struct file *f,1351			    struct queue_properties *properties,1352			    unsigned int *qid,1353			    const struct kfd_criu_queue_priv_data *q_data,1354			    const void *restore_mqd,1355			    const void *restore_ctl_stack,1356			    uint32_t *p_doorbell_offset_in_process);1357int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);1358int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid,1359			struct queue_properties *p);1360int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid,1361			struct mqd_update_info *minfo);1362int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,1363			void *gws);1364struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,1365						unsigned int qid);1366struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,1367						unsigned int qid);1368int pqm_get_wave_state(struct process_queue_manager *pqm,1369		       unsigned int qid,1370		       void __user *ctl_stack,1371		       u32 *ctl_stack_used_size,1372		       u32 *save_area_used_size);1373int pqm_get_queue_snapshot(struct process_queue_manager *pqm,1374			   uint64_t exception_clear_mask,1375			   void __user *buf,1376			   int *num_qss_entries,1377			   uint32_t *entry_size);1378 1379int amdkfd_fence_wait_timeout(struct device_queue_manager *dqm,1380			      uint64_t fence_value,1381			      unsigned int timeout_ms);1382 1383int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,1384				  unsigned int qid,1385				  u32 *mqd_size,1386				  u32 *ctl_stack_size);1387/* Packet Manager */1388 1389#define KFD_FENCE_COMPLETED (100)1390#define KFD_FENCE_INIT   (10)1391 1392struct packet_manager {1393	struct device_queue_manager *dqm;1394	struct kernel_queue *priv_queue;1395	struct mutex lock;1396	bool allocated;1397	struct kfd_mem_obj *ib_buffer_obj;1398	unsigned int ib_size_bytes;1399	bool is_over_subscription;1400 1401	const struct packet_manager_funcs *pmf;1402};1403 1404struct packet_manager_funcs {1405	/* Support ASIC-specific packet formats for PM4 packets */1406	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,1407			struct qcm_process_device *qpd);1408	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,1409			uint64_t ib, size_t ib_size_in_dwords, bool chain);1410	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,1411			struct scheduling_resources *res);1412	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,1413			struct queue *q, bool is_static);1414	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,1415			enum kfd_unmap_queues_filter mode,1416			uint32_t filter_param, bool reset);1417	int (*set_grace_period)(struct packet_manager *pm, uint32_t *buffer,1418			uint32_t grace_period);1419	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,1420			uint64_t fence_address,	uint64_t fence_value);1421	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);1422 1423	/* Packet sizes */1424	int map_process_size;1425	int runlist_size;1426	int set_resources_size;1427	int map_queues_size;1428	int unmap_queues_size;1429	int set_grace_period_size;1430	int query_status_size;1431	int release_mem_size;1432};1433 1434extern const struct packet_manager_funcs kfd_vi_pm_funcs;1435extern const struct packet_manager_funcs kfd_v9_pm_funcs;1436extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;1437 1438int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);1439void pm_uninit(struct packet_manager *pm);1440int pm_send_set_resources(struct packet_manager *pm,1441				struct scheduling_resources *res);1442int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);1443int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,1444				uint64_t fence_value);1445 1446int pm_send_unmap_queue(struct packet_manager *pm,1447			enum kfd_unmap_queues_filter mode,1448			uint32_t filter_param, bool reset);1449 1450void pm_release_ib(struct packet_manager *pm);1451 1452int pm_update_grace_period(struct packet_manager *pm, uint32_t grace_period);1453 1454/* Following PM funcs can be shared among VI and AI */1455unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);1456 1457uint64_t kfd_get_number_elems(struct kfd_dev *kfd);1458 1459/* Events */1460extern const struct kfd_event_interrupt_class event_interrupt_class_cik;1461extern const struct kfd_event_interrupt_class event_interrupt_class_v9;1462extern const struct kfd_event_interrupt_class event_interrupt_class_v9_4_3;1463extern const struct kfd_event_interrupt_class event_interrupt_class_v10;1464extern const struct kfd_event_interrupt_class event_interrupt_class_v11;1465 1466extern const struct kfd_device_global_init_class device_global_init_class_cik;1467 1468int kfd_event_init_process(struct kfd_process *p);1469void kfd_event_free_process(struct kfd_process *p);1470int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);1471int kfd_wait_on_events(struct kfd_process *p,1472		       uint32_t num_events, void __user *data,1473		       bool all, uint32_t *user_timeout_ms,1474		       uint32_t *wait_result);1475void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,1476				uint32_t valid_id_bits);1477void kfd_signal_hw_exception_event(u32 pasid);1478int kfd_set_event(struct kfd_process *p, uint32_t event_id);1479int kfd_reset_event(struct kfd_process *p, uint32_t event_id);1480int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset);1481 1482int kfd_event_create(struct file *devkfd, struct kfd_process *p,1483		     uint32_t event_type, bool auto_reset, uint32_t node_id,1484		     uint32_t *event_id, uint32_t *event_trigger_data,1485		     uint64_t *event_page_offset, uint32_t *event_slot_index);1486 1487int kfd_get_num_events(struct kfd_process *p);1488int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);1489 1490void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid,1491				struct kfd_vm_fault_info *info,1492				struct kfd_hsa_memory_exception_data *data);1493 1494void kfd_signal_reset_event(struct kfd_node *dev);1495 1496void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid);1497 1498static inline void kfd_flush_tlb(struct kfd_process_device *pdd,1499				 enum TLB_FLUSH_TYPE type)1500{1501	struct amdgpu_device *adev = pdd->dev->adev;1502	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);1503 1504	amdgpu_vm_flush_compute_tlb(adev, vm, type, pdd->dev->xcc_mask);1505}1506 1507static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev)1508{1509	return KFD_GC_VERSION(dev) >= IP_VERSION(9, 4, 2) ||1510	       (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) ||1511	       KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0);1512}1513 1514int kfd_send_exception_to_runtime(struct kfd_process *p,1515				unsigned int queue_id,1516				uint64_t error_reason);1517bool kfd_is_locked(void);1518 1519/* Compute profile */1520void kfd_inc_compute_active(struct kfd_node *dev);1521void kfd_dec_compute_active(struct kfd_node *dev);1522 1523/* Cgroup Support */1524/* Check with device cgroup if @kfd device is accessible */1525static inline int kfd_devcgroup_check_permission(struct kfd_node *node)1526{1527#if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)1528	struct drm_device *ddev;1529 1530	if (node->xcp)1531		ddev = node->xcp->ddev;1532	else1533		ddev = adev_to_drm(node->adev);1534 1535	return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,1536					  ddev->render->index,1537					  DEVCG_ACC_WRITE | DEVCG_ACC_READ);1538#else1539	return 0;1540#endif1541}1542 1543static inline bool kfd_is_first_node(struct kfd_node *node)1544{1545	return (node == node->kfd->nodes[0]);1546}1547 1548/* Debugfs */1549#if defined(CONFIG_DEBUG_FS)1550 1551void kfd_debugfs_init(void);1552void kfd_debugfs_fini(void);1553int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);1554int pqm_debugfs_mqds(struct seq_file *m, void *data);1555int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);1556int dqm_debugfs_hqds(struct seq_file *m, void *data);1557int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);1558int pm_debugfs_runlist(struct seq_file *m, void *data);1559 1560int kfd_debugfs_hang_hws(struct kfd_node *dev);1561int pm_debugfs_hang_hws(struct packet_manager *pm);1562int dqm_debugfs_hang_hws(struct device_queue_manager *dqm);1563 1564#else1565 1566static inline void kfd_debugfs_init(void) {}1567static inline void kfd_debugfs_fini(void) {}1568 1569#endif1570 1571#endif1572