14a488a7aSOded Gabbay /* 24a488a7aSOded Gabbay * Copyright 2014 Advanced Micro Devices, Inc. 34a488a7aSOded Gabbay * 44a488a7aSOded Gabbay * Permission is hereby granted, free of charge, to any person obtaining a 54a488a7aSOded Gabbay * copy of this software and associated documentation files (the "Software"), 64a488a7aSOded Gabbay * to deal in the Software without restriction, including without limitation 74a488a7aSOded Gabbay * the rights to use, copy, modify, merge, publish, distribute, sublicense, 84a488a7aSOded Gabbay * and/or sell copies of the Software, and to permit persons to whom the 94a488a7aSOded Gabbay * Software is furnished to do so, subject to the following conditions: 104a488a7aSOded Gabbay * 114a488a7aSOded Gabbay * The above copyright notice and this permission notice shall be included in 124a488a7aSOded Gabbay * all copies or substantial portions of the Software. 134a488a7aSOded Gabbay * 144a488a7aSOded Gabbay * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 154a488a7aSOded Gabbay * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 164a488a7aSOded Gabbay * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 174a488a7aSOded Gabbay * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 184a488a7aSOded Gabbay * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 194a488a7aSOded Gabbay * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 204a488a7aSOded Gabbay * OTHER DEALINGS IN THE SOFTWARE. 214a488a7aSOded Gabbay */ 224a488a7aSOded Gabbay 234a488a7aSOded Gabbay #ifndef KFD_PRIV_H_INCLUDED 244a488a7aSOded Gabbay #define KFD_PRIV_H_INCLUDED 254a488a7aSOded Gabbay 264a488a7aSOded Gabbay #include <linux/hashtable.h> 274a488a7aSOded Gabbay #include <linux/mmu_notifier.h> 284a488a7aSOded Gabbay #include <linux/mutex.h> 294a488a7aSOded Gabbay #include <linux/types.h> 304a488a7aSOded Gabbay #include <linux/atomic.h> 314a488a7aSOded Gabbay #include <linux/workqueue.h> 324a488a7aSOded Gabbay #include <linux/spinlock.h> 3319f6d2a6SOded Gabbay #include <linux/kfd_ioctl.h> 34482f0777SFelix Kuehling #include <linux/idr.h> 3504ad47bdSAndres Rodriguez #include <linux/kfifo.h> 36851a645eSFelix Kuehling #include <linux/seq_file.h> 375ce10687SFelix Kuehling #include <linux/kref.h> 38de9f26bbSKent Russell #include <linux/sysfs.h> 396b855f7bSHarish Kasiviswanathan #include <linux/device_cgroup.h> 401cd4d9eeSStephen Rothwell #include <drm/drm_file.h> 411cd4d9eeSStephen Rothwell #include <drm/drm_drv.h> 421cd4d9eeSStephen Rothwell #include <drm/drm_device.h> 4399c7b309SLorenz Brun #include <drm/drm_ioctl.h> 444a488a7aSOded Gabbay #include <kgd_kfd_interface.h> 456d220a7eSAmber Lin #include <linux/swap.h> 464a488a7aSOded Gabbay 47e596b903SYong Zhao #include "amd_shared.h" 48e596b903SYong Zhao 49af47b390SLaura Abbott #define KFD_MAX_RING_ENTRY_SIZE 8 50af47b390SLaura Abbott 515b5c4e40SEvgeny Pinchuk #define KFD_SYSFS_FILE_MODE 0444 525b5c4e40SEvgeny Pinchuk 53df03ef93SHarish Kasiviswanathan /* GPU ID hash width in bits */ 54df03ef93SHarish Kasiviswanathan #define KFD_GPU_ID_HASH_WIDTH 16 55df03ef93SHarish Kasiviswanathan 56df03ef93SHarish Kasiviswanathan /* Use upper bits of mmap offset to store KFD driver specific information. 57df03ef93SHarish Kasiviswanathan * BITS[63:62] - Encode MMAP type 58df03ef93SHarish Kasiviswanathan * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to 59df03ef93SHarish Kasiviswanathan * BITS[45:0] - MMAP offset value 60df03ef93SHarish Kasiviswanathan * 61df03ef93SHarish Kasiviswanathan * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these 62df03ef93SHarish Kasiviswanathan * defines are w.r.t to PAGE_SIZE 63df03ef93SHarish Kasiviswanathan */ 6429453755SYong Zhao #define KFD_MMAP_TYPE_SHIFT 62 65df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT) 66df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT) 67df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT) 68df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT) 69d33ea570SOak Zeng #define KFD_MMAP_TYPE_MMIO (0x0ULL << KFD_MMAP_TYPE_SHIFT) 70df03ef93SHarish Kasiviswanathan 7129453755SYong Zhao #define KFD_MMAP_GPU_ID_SHIFT 46 72df03ef93SHarish Kasiviswanathan #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \ 73df03ef93SHarish Kasiviswanathan << KFD_MMAP_GPU_ID_SHIFT) 74df03ef93SHarish Kasiviswanathan #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\ 75df03ef93SHarish Kasiviswanathan & KFD_MMAP_GPU_ID_MASK) 7629453755SYong Zhao #define KFD_MMAP_GET_GPU_ID(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \ 77df03ef93SHarish Kasiviswanathan >> KFD_MMAP_GPU_ID_SHIFT) 78df03ef93SHarish Kasiviswanathan 79ed6e6a34SBen Goz /* 80ed6e6a34SBen Goz * When working with cp scheduler we should assign the HIQ manually or via 81e7016d8eSYong Zhao * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot 82ed6e6a34SBen Goz * definitions for Kaveri. In Kaveri only the first ME queues participates 83ed6e6a34SBen Goz * in the cp scheduling taking that in mind we set the HIQ slot in the 84ed6e6a34SBen Goz * second ME. 85ed6e6a34SBen Goz */ 86ed6e6a34SBen Goz #define KFD_CIK_HIQ_PIPE 4 87ed6e6a34SBen Goz #define KFD_CIK_HIQ_QUEUE 0 88ed6e6a34SBen Goz 895b5c4e40SEvgeny Pinchuk /* Macro for allocating structures */ 905b5c4e40SEvgeny Pinchuk #define kfd_alloc_struct(ptr_to_struct) \ 915b5c4e40SEvgeny Pinchuk ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 925b5c4e40SEvgeny Pinchuk 9319f6d2a6SOded Gabbay #define KFD_MAX_NUM_OF_PROCESSES 512 94b8cbab04SOded Gabbay #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 9519f6d2a6SOded Gabbay 9619f6d2a6SOded Gabbay /* 97373d7080SFelix Kuehling * Size of the per-process TBA+TMA buffer: 2 pages 98373d7080SFelix Kuehling * 99373d7080SFelix Kuehling * The first page is the TBA used for the CWSR ISA code. The second 100a4497974SRajneesh Bhardwaj * page is used as TMA for user-mode trap handler setup in daisy-chain mode. 101373d7080SFelix Kuehling */ 102373d7080SFelix Kuehling #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) 103373d7080SFelix Kuehling #define KFD_CWSR_TMA_OFFSET PAGE_SIZE 104373d7080SFelix Kuehling 10574523943SYong Zhao #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 10674523943SYong Zhao (KFD_MAX_NUM_OF_PROCESSES * \ 10774523943SYong Zhao KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 10874523943SYong Zhao 10974523943SYong Zhao #define KFD_KERNEL_QUEUE_SIZE 2048 11074523943SYong Zhao 11114328aa5SPhilip Cox #define KFD_UNMAP_LATENCY_MS (4000) 11214328aa5SPhilip Cox 113373d7080SFelix Kuehling /* 1141f86805aSYong Zhao * 512 = 0x200 1151f86805aSYong Zhao * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the 1161f86805aSYong Zhao * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA. 1171f86805aSYong Zhao * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC 1181f86805aSYong Zhao * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in 1191f86805aSYong Zhao * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE. 1201f86805aSYong Zhao */ 1211f86805aSYong Zhao #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512 1221f86805aSYong Zhao 1231f86805aSYong Zhao 1241f86805aSYong Zhao /* 125b8cbab04SOded Gabbay * Kernel module parameter to specify maximum number of supported queues per 126b8cbab04SOded Gabbay * device 12719f6d2a6SOded Gabbay */ 128b8cbab04SOded Gabbay extern int max_num_of_queues_per_device; 12919f6d2a6SOded Gabbay 130ed6e6a34SBen Goz 13131c21fecSBen Goz /* Kernel module parameter to specify the scheduling policy */ 13231c21fecSBen Goz extern int sched_policy; 13331c21fecSBen Goz 134a99c6d4fSFelix Kuehling /* 135a99c6d4fSFelix Kuehling * Kernel module parameter to specify the maximum process 136a99c6d4fSFelix Kuehling * number per HW scheduler 137a99c6d4fSFelix Kuehling */ 138a99c6d4fSFelix Kuehling extern int hws_max_conc_proc; 139a99c6d4fSFelix Kuehling 140373d7080SFelix Kuehling extern int cwsr_enable; 141373d7080SFelix Kuehling 14281663016SOded Gabbay /* 14381663016SOded Gabbay * Kernel module parameter to specify whether to send sigterm to HSA process on 14481663016SOded Gabbay * unhandled exception 14581663016SOded Gabbay */ 14681663016SOded Gabbay extern int send_sigterm; 14781663016SOded Gabbay 148ebcfd1e2SFelix Kuehling /* 149374200b1SFelix Kuehling * This kernel module is used to simulate large bar machine on non-large bar 150374200b1SFelix Kuehling * enabled machines. 151374200b1SFelix Kuehling */ 152374200b1SFelix Kuehling extern int debug_largebar; 153374200b1SFelix Kuehling 154374200b1SFelix Kuehling /* 155ebcfd1e2SFelix Kuehling * Ignore CRAT table during KFD initialization, can be used to work around 156ebcfd1e2SFelix Kuehling * broken CRAT tables on some AMD systems 157ebcfd1e2SFelix Kuehling */ 158ebcfd1e2SFelix Kuehling extern int ignore_crat; 159ebcfd1e2SFelix Kuehling 160a4497974SRajneesh Bhardwaj /* Set sh_mem_config.retry_disable on GFX v9 */ 16175ee6487SFelix Kuehling extern int amdgpu_noretry; 162bed4f110SFelix Kuehling 163a4497974SRajneesh Bhardwaj /* Halt if HWS hang is detected */ 1640e9a860cSYong Zhao extern int halt_if_hws_hang; 1650e9a860cSYong Zhao 166a4497974SRajneesh Bhardwaj /* Whether MEC FW support GWS barriers */ 16729e76462SOak Zeng extern bool hws_gws_support; 16829e76462SOak Zeng 169a4497974SRajneesh Bhardwaj /* Queue preemption timeout in ms */ 17014328aa5SPhilip Cox extern int queue_preemption_timeout_ms; 17114328aa5SPhilip Cox 172a4497974SRajneesh Bhardwaj /* Enable eviction debug messages */ 173b2057956SFelix Kuehling extern bool debug_evictions; 174b2057956SFelix Kuehling 175ed6e6a34SBen Goz enum cache_policy { 176ed6e6a34SBen Goz cache_policy_coherent, 177ed6e6a34SBen Goz cache_policy_noncoherent 178ed6e6a34SBen Goz }; 179ed6e6a34SBen Goz 180ef568db7SFelix Kuehling #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10) 181ef568db7SFelix Kuehling 182f3a39818SAndrew Lewycky struct kfd_event_interrupt_class { 183f3a39818SAndrew Lewycky bool (*interrupt_isr)(struct kfd_dev *dev, 18458e69886SLan Xiao const uint32_t *ih_ring_entry, uint32_t *patched_ihre, 18558e69886SLan Xiao bool *patched_flag); 186f3a39818SAndrew Lewycky void (*interrupt_wq)(struct kfd_dev *dev, 187f3a39818SAndrew Lewycky const uint32_t *ih_ring_entry); 188f3a39818SAndrew Lewycky }; 189f3a39818SAndrew Lewycky 1904a488a7aSOded Gabbay struct kfd_device_info { 191e596b903SYong Zhao enum amd_asic_type asic_family; 192c181159aSYong Zhao const char *asic_name; 193f3a39818SAndrew Lewycky const struct kfd_event_interrupt_class *event_interrupt_class; 1944a488a7aSOded Gabbay unsigned int max_pasid_bits; 195992839adSYair Shachar unsigned int max_no_of_hqd; 196ada2b29cSFelix Kuehling unsigned int doorbell_size; 1974a488a7aSOded Gabbay size_t ih_ring_entry_size; 198f7c826adSAlexey Skidanov uint8_t num_of_watch_points; 19919f6d2a6SOded Gabbay uint16_t mqd_size_aligned; 200373d7080SFelix Kuehling bool supports_cwsr; 20164d1c3a4SFelix Kuehling bool needs_iommu_device; 2023ee2d00cSFelix Kuehling bool needs_pci_atomics; 20398bb9222SYong Zhao unsigned int num_sdma_engines; 2041b4670f6SOak Zeng unsigned int num_xgmi_sdma_engines; 205d5094189SShaoyun Liu unsigned int num_sdma_queues_per_engine; 2064a488a7aSOded Gabbay }; 2074a488a7aSOded Gabbay 20836b5c08fSOded Gabbay struct kfd_mem_obj { 20936b5c08fSOded Gabbay uint32_t range_start; 21036b5c08fSOded Gabbay uint32_t range_end; 21136b5c08fSOded Gabbay uint64_t gpu_addr; 21236b5c08fSOded Gabbay uint32_t *cpu_ptr; 213b91d43ddSFelix Kuehling void *gtt_mem; 21436b5c08fSOded Gabbay }; 21536b5c08fSOded Gabbay 21644008d7aSYong Zhao struct kfd_vmid_info { 21744008d7aSYong Zhao uint32_t first_vmid_kfd; 21844008d7aSYong Zhao uint32_t last_vmid_kfd; 21944008d7aSYong Zhao uint32_t vmid_num_kfd; 22044008d7aSYong Zhao }; 22144008d7aSYong Zhao 2224a488a7aSOded Gabbay struct kfd_dev { 2234a488a7aSOded Gabbay struct kgd_dev *kgd; 2244a488a7aSOded Gabbay 2254a488a7aSOded Gabbay const struct kfd_device_info *device_info; 2264a488a7aSOded Gabbay struct pci_dev *pdev; 2273a0c3423SHarish Kasiviswanathan struct drm_device *ddev; 2284a488a7aSOded Gabbay 2294a488a7aSOded Gabbay unsigned int id; /* topology stub index */ 2304a488a7aSOded Gabbay 23119f6d2a6SOded Gabbay phys_addr_t doorbell_base; /* Start of actual doorbells used by 23219f6d2a6SOded Gabbay * KFD. It is aligned for mapping 23319f6d2a6SOded Gabbay * into user mode 23419f6d2a6SOded Gabbay */ 235339903faSYong Zhao size_t doorbell_base_dw_offset; /* Offset from the start of the PCI 236339903faSYong Zhao * doorbell BAR to the first KFD 237339903faSYong Zhao * doorbell in dwords. GFX reserves 238339903faSYong Zhao * the segment before this offset. 23919f6d2a6SOded Gabbay */ 24019f6d2a6SOded Gabbay u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 24119f6d2a6SOded Gabbay * page used by kernel queue 24219f6d2a6SOded Gabbay */ 24319f6d2a6SOded Gabbay 2444a488a7aSOded Gabbay struct kgd2kfd_shared_resources shared_resources; 24544008d7aSYong Zhao struct kfd_vmid_info vm_info; 2464a488a7aSOded Gabbay 247cea405b1SXihan Zhang const struct kfd2kgd_calls *kfd2kgd; 248cea405b1SXihan Zhang struct mutex doorbell_mutex; 249f761d8bdSJoe Perches DECLARE_BITMAP(doorbell_available_index, 250f761d8bdSJoe Perches KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 251cea405b1SXihan Zhang 25236b5c08fSOded Gabbay void *gtt_mem; 25336b5c08fSOded Gabbay uint64_t gtt_start_gpu_addr; 25436b5c08fSOded Gabbay void *gtt_start_cpu_ptr; 25536b5c08fSOded Gabbay void *gtt_sa_bitmap; 25636b5c08fSOded Gabbay struct mutex gtt_sa_lock; 25736b5c08fSOded Gabbay unsigned int gtt_sa_chunk_size; 25836b5c08fSOded Gabbay unsigned int gtt_sa_num_of_chunks; 25936b5c08fSOded Gabbay 2602249d558SAndrew Lewycky /* Interrupts */ 26104ad47bdSAndres Rodriguez struct kfifo ih_fifo; 26248e876a2SAndres Rodriguez struct workqueue_struct *ih_wq; 2632249d558SAndrew Lewycky struct work_struct interrupt_work; 2642249d558SAndrew Lewycky spinlock_t interrupt_lock; 2652249d558SAndrew Lewycky 266ed6e6a34SBen Goz /* QCM Device instance */ 267ed6e6a34SBen Goz struct device_queue_manager *dqm; 2684a488a7aSOded Gabbay 269ed6e6a34SBen Goz bool init_complete; 2702249d558SAndrew Lewycky /* 2712249d558SAndrew Lewycky * Interrupts of interest to KFD are copied 2722249d558SAndrew Lewycky * from the HW ring into a SW ring. 2732249d558SAndrew Lewycky */ 2742249d558SAndrew Lewycky bool interrupts_active; 275fbeb661bSYair Shachar 276fbeb661bSYair Shachar /* Debug manager */ 277fbeb661bSYair Shachar struct kfd_dbgmgr *dbgmgr; 278373d7080SFelix Kuehling 2795ade6c9cSFelix Kuehling /* Firmware versions */ 2805ade6c9cSFelix Kuehling uint16_t mec_fw_version; 28129633d0eSJoseph Greathouse uint16_t mec2_fw_version; 2825ade6c9cSFelix Kuehling uint16_t sdma_fw_version; 2835ade6c9cSFelix Kuehling 284a99c6d4fSFelix Kuehling /* Maximum process number mapped to HW scheduler */ 285a99c6d4fSFelix Kuehling unsigned int max_proc_per_quantum; 286a99c6d4fSFelix Kuehling 287373d7080SFelix Kuehling /* CWSR */ 288373d7080SFelix Kuehling bool cwsr_enabled; 289373d7080SFelix Kuehling const void *cwsr_isa; 290373d7080SFelix Kuehling unsigned int cwsr_isa_size; 2910c1690e3SShaoyun Liu 2920c1690e3SShaoyun Liu /* xGMI */ 2930c1690e3SShaoyun Liu uint64_t hive_id; 2940c663695SDivya Shikre 295d35f00d8SEric Huang bool pci_atomic_requested; 2969b54d201SEric Huang 2976127896fSHuang Rui /* Use IOMMU v2 flag */ 2986127896fSHuang Rui bool use_iommu_v2; 2996127896fSHuang Rui 3009b54d201SEric Huang /* SRAM ECC flag */ 3019b54d201SEric Huang atomic_t sram_ecc_flag; 302f756e631SHarish Kasiviswanathan 303f756e631SHarish Kasiviswanathan /* Compute Profile ref. count */ 304f756e631SHarish Kasiviswanathan atomic_t compute_profile; 305e09d4fc8SOak Zeng 306a4497974SRajneesh Bhardwaj /* Global GWS resource shared between processes */ 307e09d4fc8SOak Zeng void *gws; 308938a0650SAmber Lin 309938a0650SAmber Lin /* Clients watching SMI events */ 310938a0650SAmber Lin struct list_head smi_clients; 311938a0650SAmber Lin spinlock_t smi_lock; 31255977744SMukul Joshi 31355977744SMukul Joshi uint32_t reset_seq_num; 31459d7115dSMukul Joshi 31559d7115dSMukul Joshi struct ida doorbell_ida; 31659d7115dSMukul Joshi unsigned int max_doorbell_slices; 3179b498efaSAlex Deucher 3189b498efaSAlex Deucher int noretry; 3194a488a7aSOded Gabbay }; 3204a488a7aSOded Gabbay 32119f6d2a6SOded Gabbay enum kfd_mempool { 32219f6d2a6SOded Gabbay KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 32319f6d2a6SOded Gabbay KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 32419f6d2a6SOded Gabbay KFD_MEMPOOL_FRAMEBUFFER = 3, 32519f6d2a6SOded Gabbay }; 32619f6d2a6SOded Gabbay 3274a488a7aSOded Gabbay /* Character device interface */ 3284a488a7aSOded Gabbay int kfd_chardev_init(void); 3294a488a7aSOded Gabbay void kfd_chardev_exit(void); 3304a488a7aSOded Gabbay struct device *kfd_chardev(void); 3314a488a7aSOded Gabbay 332241f24f8SBen Goz /** 333a4497974SRajneesh Bhardwaj * enum kfd_unmap_queues_filter - Enum for queue filters. 334241f24f8SBen Goz * 3357da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 336241f24f8SBen Goz * 3377da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 338241f24f8SBen Goz * running queues list. 339241f24f8SBen Goz * 3407da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 341241f24f8SBen Goz * specific process. 342241f24f8SBen Goz * 343241f24f8SBen Goz */ 3447da2bcf8SYong Zhao enum kfd_unmap_queues_filter { 3457da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 3467da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 3477da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 3487da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_BY_PASID 349241f24f8SBen Goz }; 35019f6d2a6SOded Gabbay 351ed8aab45SBen Goz /** 352a4497974SRajneesh Bhardwaj * enum kfd_queue_type - Enum for various queue types. 353ed8aab45SBen Goz * 354ed8aab45SBen Goz * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 355ed8aab45SBen Goz * 356a4497974SRajneesh Bhardwaj * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type. 357ed8aab45SBen Goz * 358ed8aab45SBen Goz * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 359ed8aab45SBen Goz * 360ed8aab45SBen Goz * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 361a4497974SRajneesh Bhardwaj * 362a4497974SRajneesh Bhardwaj * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface. 363ed8aab45SBen Goz */ 364ed8aab45SBen Goz enum kfd_queue_type { 365ed8aab45SBen Goz KFD_QUEUE_TYPE_COMPUTE, 366ed8aab45SBen Goz KFD_QUEUE_TYPE_SDMA, 367ed8aab45SBen Goz KFD_QUEUE_TYPE_HIQ, 3681b4670f6SOak Zeng KFD_QUEUE_TYPE_DIQ, 3691b4670f6SOak Zeng KFD_QUEUE_TYPE_SDMA_XGMI 370ed8aab45SBen Goz }; 371ed8aab45SBen Goz 3726e99df57SBen Goz enum kfd_queue_format { 3736e99df57SBen Goz KFD_QUEUE_FORMAT_PM4, 3746e99df57SBen Goz KFD_QUEUE_FORMAT_AQL 3756e99df57SBen Goz }; 3766e99df57SBen Goz 3770ccbc7cdSOak Zeng enum KFD_QUEUE_PRIORITY { 3780ccbc7cdSOak Zeng KFD_QUEUE_PRIORITY_MINIMUM = 0, 3790ccbc7cdSOak Zeng KFD_QUEUE_PRIORITY_MAXIMUM = 15 3800ccbc7cdSOak Zeng }; 3810ccbc7cdSOak Zeng 382ed8aab45SBen Goz /** 383ed8aab45SBen Goz * struct queue_properties 384ed8aab45SBen Goz * 385ed8aab45SBen Goz * @type: The queue type. 386ed8aab45SBen Goz * 387ed8aab45SBen Goz * @queue_id: Queue identifier. 388ed8aab45SBen Goz * 389ed8aab45SBen Goz * @queue_address: Queue ring buffer address. 390ed8aab45SBen Goz * 391ed8aab45SBen Goz * @queue_size: Queue ring buffer size. 392ed8aab45SBen Goz * 393ed8aab45SBen Goz * @priority: Defines the queue priority relative to other queues in the 394ed8aab45SBen Goz * process. 395ed8aab45SBen Goz * This is just an indication and HW scheduling may override the priority as 396ed8aab45SBen Goz * necessary while keeping the relative prioritization. 397ed8aab45SBen Goz * the priority granularity is from 0 to f which f is the highest priority. 398ed8aab45SBen Goz * currently all queues are initialized with the highest priority. 399ed8aab45SBen Goz * 400ed8aab45SBen Goz * @queue_percent: This field is partially implemented and currently a zero in 401ed8aab45SBen Goz * this field defines that the queue is non active. 402ed8aab45SBen Goz * 403ed8aab45SBen Goz * @read_ptr: User space address which points to the number of dwords the 404ed8aab45SBen Goz * cp read from the ring buffer. This field updates automatically by the H/W. 405ed8aab45SBen Goz * 406ed8aab45SBen Goz * @write_ptr: Defines the number of dwords written to the ring buffer. 407ed8aab45SBen Goz * 408a4497974SRajneesh Bhardwaj * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring 409a4497974SRajneesh Bhardwaj * buffer. This field should be similar to write_ptr and the user should 410a4497974SRajneesh Bhardwaj * update this field after updating the write_ptr. 411ed8aab45SBen Goz * 412ed8aab45SBen Goz * @doorbell_off: The doorbell offset in the doorbell pci-bar. 413ed8aab45SBen Goz * 4148eabaf54SKent Russell * @is_interop: Defines if this is a interop queue. Interop queue means that 4158eabaf54SKent Russell * the queue can access both graphics and compute resources. 416ed8aab45SBen Goz * 41726103436SFelix Kuehling * @is_evicted: Defines if the queue is evicted. Only active queues 41826103436SFelix Kuehling * are evicted, rendering them inactive. 41926103436SFelix Kuehling * 42026103436SFelix Kuehling * @is_active: Defines if the queue is active or not. @is_active and 42126103436SFelix Kuehling * @is_evicted are protected by the DQM lock. 422ed8aab45SBen Goz * 423b8020b03SJoseph Greathouse * @is_gws: Defines if the queue has been updated to be GWS-capable or not. 424b8020b03SJoseph Greathouse * @is_gws should be protected by the DQM lock, since changing it can yield the 425b8020b03SJoseph Greathouse * possibility of updating DQM state on number of GWS queues. 426b8020b03SJoseph Greathouse * 427ed8aab45SBen Goz * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 428ed8aab45SBen Goz * of the queue. 429ed8aab45SBen Goz * 430ed8aab45SBen Goz * This structure represents the queue properties for each queue no matter if 431ed8aab45SBen Goz * it's user mode or kernel mode queue. 432ed8aab45SBen Goz * 433ed8aab45SBen Goz */ 434ed8aab45SBen Goz struct queue_properties { 435ed8aab45SBen Goz enum kfd_queue_type type; 4366e99df57SBen Goz enum kfd_queue_format format; 437ed8aab45SBen Goz unsigned int queue_id; 438ed8aab45SBen Goz uint64_t queue_address; 439ed8aab45SBen Goz uint64_t queue_size; 440ed8aab45SBen Goz uint32_t priority; 441ed8aab45SBen Goz uint32_t queue_percent; 442ed8aab45SBen Goz uint32_t *read_ptr; 443ed8aab45SBen Goz uint32_t *write_ptr; 444ada2b29cSFelix Kuehling void __iomem *doorbell_ptr; 445ed8aab45SBen Goz uint32_t doorbell_off; 446ed8aab45SBen Goz bool is_interop; 44726103436SFelix Kuehling bool is_evicted; 448ed8aab45SBen Goz bool is_active; 449b8020b03SJoseph Greathouse bool is_gws; 450ed8aab45SBen Goz /* Not relevant for user mode queues in cp scheduling */ 451ed8aab45SBen Goz unsigned int vmid; 45277669eb8SBen Goz /* Relevant only for sdma queues*/ 45377669eb8SBen Goz uint32_t sdma_engine_id; 45477669eb8SBen Goz uint32_t sdma_queue_id; 45577669eb8SBen Goz uint32_t sdma_vm_addr; 456ff3d04a1SBen Goz /* Relevant only for VI */ 457ff3d04a1SBen Goz uint64_t eop_ring_buffer_address; 458ff3d04a1SBen Goz uint32_t eop_ring_buffer_size; 459ff3d04a1SBen Goz uint64_t ctx_save_restore_area_address; 460ff3d04a1SBen Goz uint32_t ctx_save_restore_area_size; 461373d7080SFelix Kuehling uint32_t ctl_stack_size; 462373d7080SFelix Kuehling uint64_t tba_addr; 463373d7080SFelix Kuehling uint64_t tma_addr; 46439e7f331SFelix Kuehling /* Relevant for CU */ 46539e7f331SFelix Kuehling uint32_t cu_mask_count; /* Must be a multiple of 32 */ 46639e7f331SFelix Kuehling uint32_t *cu_mask; 467ed8aab45SBen Goz }; 468ed8aab45SBen Goz 469bb2d2128SFelix Kuehling #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ 470bb2d2128SFelix Kuehling (q).queue_address != 0 && \ 471bb2d2128SFelix Kuehling (q).queue_percent > 0 && \ 472bb2d2128SFelix Kuehling !(q).is_evicted) 473bb2d2128SFelix Kuehling 474ed8aab45SBen Goz /** 475ed8aab45SBen Goz * struct queue 476ed8aab45SBen Goz * 477ed8aab45SBen Goz * @list: Queue linked list. 478ed8aab45SBen Goz * 479a4497974SRajneesh Bhardwaj * @mqd: The queue MQD (memory queue descriptor). 480ed8aab45SBen Goz * 481ed8aab45SBen Goz * @mqd_mem_obj: The MQD local gpu memory object. 482ed8aab45SBen Goz * 483ed8aab45SBen Goz * @gart_mqd_addr: The MQD gart mc address. 484ed8aab45SBen Goz * 485ed8aab45SBen Goz * @properties: The queue properties. 486ed8aab45SBen Goz * 487ed8aab45SBen Goz * @mec: Used only in no cp scheduling mode and identifies to micro engine id 488a4497974SRajneesh Bhardwaj * that the queue should be executed on. 489ed8aab45SBen Goz * 4908eabaf54SKent Russell * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 4918eabaf54SKent Russell * id. 492ed8aab45SBen Goz * 493ed8aab45SBen Goz * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 494ed8aab45SBen Goz * 495ed8aab45SBen Goz * @process: The kfd process that created this queue. 496ed8aab45SBen Goz * 497ed8aab45SBen Goz * @device: The kfd device that created this queue. 498ed8aab45SBen Goz * 499eb82da1dSOak Zeng * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL 500eb82da1dSOak Zeng * otherwise. 501eb82da1dSOak Zeng * 502ed8aab45SBen Goz * This structure represents user mode compute queues. 503ed8aab45SBen Goz * It contains all the necessary data to handle such queues. 504ed8aab45SBen Goz * 505ed8aab45SBen Goz */ 506ed8aab45SBen Goz 507ed8aab45SBen Goz struct queue { 508ed8aab45SBen Goz struct list_head list; 509ed8aab45SBen Goz void *mqd; 510ed8aab45SBen Goz struct kfd_mem_obj *mqd_mem_obj; 511ed8aab45SBen Goz uint64_t gart_mqd_addr; 512ed8aab45SBen Goz struct queue_properties properties; 513ed8aab45SBen Goz 514ed8aab45SBen Goz uint32_t mec; 515ed8aab45SBen Goz uint32_t pipe; 516ed8aab45SBen Goz uint32_t queue; 517ed8aab45SBen Goz 51877669eb8SBen Goz unsigned int sdma_id; 519ef568db7SFelix Kuehling unsigned int doorbell_id; 52077669eb8SBen Goz 521ed8aab45SBen Goz struct kfd_process *process; 522ed8aab45SBen Goz struct kfd_dev *device; 523eb82da1dSOak Zeng void *gws; 5246d220a7eSAmber Lin 5256d220a7eSAmber Lin /* procfs */ 5266d220a7eSAmber Lin struct kobject kobj; 527ed8aab45SBen Goz }; 528ed8aab45SBen Goz 5296e99df57SBen Goz enum KFD_MQD_TYPE { 530d7c0b047SYong Zhao KFD_MQD_TYPE_HIQ = 0, /* for hiq */ 53185d258f9SBen Goz KFD_MQD_TYPE_CP, /* for cp queues and diq */ 53285d258f9SBen Goz KFD_MQD_TYPE_SDMA, /* for sdma queues */ 53359f650a0SOak Zeng KFD_MQD_TYPE_DIQ, /* for diq */ 5346e99df57SBen Goz KFD_MQD_TYPE_MAX 5356e99df57SBen Goz }; 5366e99df57SBen Goz 5370ccbc7cdSOak Zeng enum KFD_PIPE_PRIORITY { 5380ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_LOW = 0, 5390ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_MEDIUM, 5400ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_HIGH 5410ccbc7cdSOak Zeng }; 5420ccbc7cdSOak Zeng 543241f24f8SBen Goz struct scheduling_resources { 544241f24f8SBen Goz unsigned int vmid_mask; 545241f24f8SBen Goz enum kfd_queue_type type; 546241f24f8SBen Goz uint64_t queue_mask; 547241f24f8SBen Goz uint64_t gws_mask; 548241f24f8SBen Goz uint32_t oac_mask; 549241f24f8SBen Goz uint32_t gds_heap_base; 550241f24f8SBen Goz uint32_t gds_heap_size; 551241f24f8SBen Goz }; 552241f24f8SBen Goz 553241f24f8SBen Goz struct process_queue_manager { 554241f24f8SBen Goz /* data */ 555241f24f8SBen Goz struct kfd_process *process; 556241f24f8SBen Goz struct list_head queues; 557241f24f8SBen Goz unsigned long *queue_slot_bitmap; 558241f24f8SBen Goz }; 559241f24f8SBen Goz 560241f24f8SBen Goz struct qcm_process_device { 561241f24f8SBen Goz /* The Device Queue Manager that owns this data */ 562241f24f8SBen Goz struct device_queue_manager *dqm; 563241f24f8SBen Goz struct process_queue_manager *pqm; 564241f24f8SBen Goz /* Queues list */ 565241f24f8SBen Goz struct list_head queues_list; 566241f24f8SBen Goz struct list_head priv_queue_list; 567241f24f8SBen Goz 568241f24f8SBen Goz unsigned int queue_count; 569241f24f8SBen Goz unsigned int vmid; 570241f24f8SBen Goz bool is_debug; 57126103436SFelix Kuehling unsigned int evicted; /* eviction counter, 0=active */ 5729fd3f1bfSFelix Kuehling 5739fd3f1bfSFelix Kuehling /* This flag tells if we should reset all wavefronts on 5749fd3f1bfSFelix Kuehling * process termination 5759fd3f1bfSFelix Kuehling */ 5769fd3f1bfSFelix Kuehling bool reset_wavefronts; 5779fd3f1bfSFelix Kuehling 578b8020b03SJoseph Greathouse /* This flag tells us if this process has a GWS-capable 579b8020b03SJoseph Greathouse * queue that will be mapped into the runlist. It's 580b8020b03SJoseph Greathouse * possible to request a GWS BO, but not have the queue 581b8020b03SJoseph Greathouse * currently mapped, and this changes how the MAP_PROCESS 582b8020b03SJoseph Greathouse * PM4 packet is configured. 583b8020b03SJoseph Greathouse */ 584b8020b03SJoseph Greathouse bool mapped_gws_queue; 585b8020b03SJoseph Greathouse 586a4497974SRajneesh Bhardwaj /* All the memory management data should be here too */ 587241f24f8SBen Goz uint64_t gds_context_area; 588435e2f97SYong Zhao /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */ 589e715c6d0SShaoyun Liu uint64_t page_table_base; 590241f24f8SBen Goz uint32_t sh_mem_config; 591241f24f8SBen Goz uint32_t sh_mem_bases; 592241f24f8SBen Goz uint32_t sh_mem_ape1_base; 593241f24f8SBen Goz uint32_t sh_mem_ape1_limit; 594241f24f8SBen Goz uint32_t gds_size; 595241f24f8SBen Goz uint32_t num_gws; 596241f24f8SBen Goz uint32_t num_oac; 5976a1c9510SMoses Reuben uint32_t sh_hidden_private_base; 598373d7080SFelix Kuehling 599373d7080SFelix Kuehling /* CWSR memory */ 600373d7080SFelix Kuehling void *cwsr_kaddr; 601d01994c2SFelix Kuehling uint64_t cwsr_base; 602373d7080SFelix Kuehling uint64_t tba_addr; 603373d7080SFelix Kuehling uint64_t tma_addr; 604d01994c2SFelix Kuehling 605d01994c2SFelix Kuehling /* IB memory */ 606d01994c2SFelix Kuehling uint64_t ib_base; 607552764b6SFelix Kuehling void *ib_kaddr; 608ef568db7SFelix Kuehling 609ef568db7SFelix Kuehling /* doorbell resources per process per device */ 610ef568db7SFelix Kuehling unsigned long *doorbell_bitmap; 611241f24f8SBen Goz }; 612241f24f8SBen Goz 61326103436SFelix Kuehling /* KFD Memory Eviction */ 61426103436SFelix Kuehling 61526103436SFelix Kuehling /* Approx. wait time before attempting to restore evicted BOs */ 61626103436SFelix Kuehling #define PROCESS_RESTORE_TIME_MS 100 61726103436SFelix Kuehling /* Approx. back off time if restore fails due to lack of memory */ 61826103436SFelix Kuehling #define PROCESS_BACK_OFF_TIME_MS 100 61926103436SFelix Kuehling /* Approx. time before evicting the process again */ 62026103436SFelix Kuehling #define PROCESS_ACTIVE_TIME_MS 10 62126103436SFelix Kuehling 6225ec7e028SFelix Kuehling /* 8 byte handle containing GPU ID in the most significant 4 bytes and 6235ec7e028SFelix Kuehling * idr_handle in the least significant 4 bytes 6245ec7e028SFelix Kuehling */ 6255ec7e028SFelix Kuehling #define MAKE_HANDLE(gpu_id, idr_handle) \ 6265ec7e028SFelix Kuehling (((uint64_t)(gpu_id) << 32) + idr_handle) 6275ec7e028SFelix Kuehling #define GET_GPU_ID(handle) (handle >> 32) 6285ec7e028SFelix Kuehling #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 6295ec7e028SFelix Kuehling 630733fa1f7SYong Zhao enum kfd_pdd_bound { 631733fa1f7SYong Zhao PDD_UNBOUND = 0, 632733fa1f7SYong Zhao PDD_BOUND, 633733fa1f7SYong Zhao PDD_BOUND_SUSPENDED, 634733fa1f7SYong Zhao }; 635733fa1f7SYong Zhao 6364327bed2SPhilip Cox #define MAX_SYSFS_FILENAME_LEN 15 63732cb59f3SMukul Joshi 63832cb59f3SMukul Joshi /* 63932cb59f3SMukul Joshi * SDMA counter runs at 100MHz frequency. 64032cb59f3SMukul Joshi * We display SDMA activity in microsecond granularity in sysfs. 64132cb59f3SMukul Joshi * As a result, the divisor is 100. 64232cb59f3SMukul Joshi */ 64332cb59f3SMukul Joshi #define SDMA_ACTIVITY_DIVISOR 100 644d4566deeSMukul Joshi 64519f6d2a6SOded Gabbay /* Data that is per-process-per device. */ 64619f6d2a6SOded Gabbay struct kfd_process_device { 64719f6d2a6SOded Gabbay /* 64819f6d2a6SOded Gabbay * List of all per-device data for a process. 64919f6d2a6SOded Gabbay * Starts from kfd_process.per_device_data. 65019f6d2a6SOded Gabbay */ 65119f6d2a6SOded Gabbay struct list_head per_device_list; 65219f6d2a6SOded Gabbay 65319f6d2a6SOded Gabbay /* The device that owns this data. */ 65419f6d2a6SOded Gabbay struct kfd_dev *dev; 65519f6d2a6SOded Gabbay 6569fd3f1bfSFelix Kuehling /* The process that owns this kfd_process_device. */ 6579fd3f1bfSFelix Kuehling struct kfd_process *process; 65819f6d2a6SOded Gabbay 65945102048SBen Goz /* per-process-per device QCM data structure */ 66045102048SBen Goz struct qcm_process_device qpd; 66145102048SBen Goz 66219f6d2a6SOded Gabbay /*Apertures*/ 66319f6d2a6SOded Gabbay uint64_t lds_base; 66419f6d2a6SOded Gabbay uint64_t lds_limit; 66519f6d2a6SOded Gabbay uint64_t gpuvm_base; 66619f6d2a6SOded Gabbay uint64_t gpuvm_limit; 66719f6d2a6SOded Gabbay uint64_t scratch_base; 66819f6d2a6SOded Gabbay uint64_t scratch_limit; 66919f6d2a6SOded Gabbay 670403575c4SFelix Kuehling /* VM context for GPUVM allocations */ 671b84394e2SFelix Kuehling struct file *drm_file; 672403575c4SFelix Kuehling void *vm; 673403575c4SFelix Kuehling 67452b29d73SFelix Kuehling /* GPUVM allocations storage */ 67552b29d73SFelix Kuehling struct idr alloc_idr; 67652b29d73SFelix Kuehling 6779fd3f1bfSFelix Kuehling /* Flag used to tell the pdd has dequeued from the dqm. 6789fd3f1bfSFelix Kuehling * This is used to prevent dev->dqm->ops.process_termination() from 6799fd3f1bfSFelix Kuehling * being called twice when it is already called in IOMMU callback 6809fd3f1bfSFelix Kuehling * function. 681a82918f1SBen Goz */ 6829fd3f1bfSFelix Kuehling bool already_dequeued; 6839593f4d6SRajneesh Bhardwaj bool runtime_inuse; 68464d1c3a4SFelix Kuehling 68564d1c3a4SFelix Kuehling /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 68664d1c3a4SFelix Kuehling enum kfd_pdd_bound bound; 687d4566deeSMukul Joshi 688d4566deeSMukul Joshi /* VRAM usage */ 689d4566deeSMukul Joshi uint64_t vram_usage; 690d4566deeSMukul Joshi struct attribute attr_vram; 69132cb59f3SMukul Joshi char vram_filename[MAX_SYSFS_FILENAME_LEN]; 69232cb59f3SMukul Joshi 69332cb59f3SMukul Joshi /* SDMA activity tracking */ 69432cb59f3SMukul Joshi uint64_t sdma_past_activity_counter; 69532cb59f3SMukul Joshi struct attribute attr_sdma; 69632cb59f3SMukul Joshi char sdma_filename[MAX_SYSFS_FILENAME_LEN]; 6974327bed2SPhilip Cox 6984327bed2SPhilip Cox /* Eviction activity tracking */ 6994327bed2SPhilip Cox uint64_t last_evict_timestamp; 7004327bed2SPhilip Cox atomic64_t evict_duration_counter; 7014327bed2SPhilip Cox struct attribute attr_evict; 7024327bed2SPhilip Cox 7034327bed2SPhilip Cox struct kobject *kobj_stats; 70459d7115dSMukul Joshi unsigned int doorbell_index; 705f2fa07b3SRamesh Errabolu 706f2fa07b3SRamesh Errabolu /* 707f2fa07b3SRamesh Errabolu * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process 708f2fa07b3SRamesh Errabolu * that is associated with device encoded by "this" struct instance. The 709f2fa07b3SRamesh Errabolu * value reflects CU usage by all of the waves launched by this process 710f2fa07b3SRamesh Errabolu * on this device. A very important property of occupancy parameter is 711f2fa07b3SRamesh Errabolu * that its value is a snapshot of current use. 712f2fa07b3SRamesh Errabolu * 713f2fa07b3SRamesh Errabolu * Following is to be noted regarding how this parameter is reported: 714f2fa07b3SRamesh Errabolu * 715f2fa07b3SRamesh Errabolu * The number of waves that a CU can launch is limited by couple of 716f2fa07b3SRamesh Errabolu * parameters. These are encoded by struct amdgpu_cu_info instance 717f2fa07b3SRamesh Errabolu * that is part of every device definition. For GFX9 devices this 718f2fa07b3SRamesh Errabolu * translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves 719f2fa07b3SRamesh Errabolu * do not use scratch memory and 32 waves (max_scratch_slots_per_cu) 720f2fa07b3SRamesh Errabolu * when they do use scratch memory. This could change for future 721f2fa07b3SRamesh Errabolu * devices and therefore this example should be considered as a guide. 722f2fa07b3SRamesh Errabolu * 723f2fa07b3SRamesh Errabolu * All CU's of a device are available for the process. This may not be true 724f2fa07b3SRamesh Errabolu * under certain conditions - e.g. CU masking. 725f2fa07b3SRamesh Errabolu * 726f2fa07b3SRamesh Errabolu * Finally number of CU's that are occupied by a process is affected by both 727f2fa07b3SRamesh Errabolu * number of CU's a device has along with number of other competing processes 728f2fa07b3SRamesh Errabolu */ 729f2fa07b3SRamesh Errabolu struct attribute attr_cu_occupancy; 73019f6d2a6SOded Gabbay }; 73119f6d2a6SOded Gabbay 73252a5fdceSAlexey Skidanov #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 73352a5fdceSAlexey Skidanov 7344a488a7aSOded Gabbay /* Process data */ 7354a488a7aSOded Gabbay struct kfd_process { 73619f6d2a6SOded Gabbay /* 73719f6d2a6SOded Gabbay * kfd_process are stored in an mm_struct*->kfd_process* 73819f6d2a6SOded Gabbay * hash table (kfd_processes in kfd_process.c) 73919f6d2a6SOded Gabbay */ 74019f6d2a6SOded Gabbay struct hlist_node kfd_processes; 74119f6d2a6SOded Gabbay 7429b56bb11SFelix Kuehling /* 7439b56bb11SFelix Kuehling * Opaque pointer to mm_struct. We don't hold a reference to 7449b56bb11SFelix Kuehling * it so it should never be dereferenced from here. This is 7459b56bb11SFelix Kuehling * only used for looking up processes by their mm. 7469b56bb11SFelix Kuehling */ 7479b56bb11SFelix Kuehling void *mm; 74819f6d2a6SOded Gabbay 7495ce10687SFelix Kuehling struct kref ref; 7505ce10687SFelix Kuehling struct work_struct release_work; 7515ce10687SFelix Kuehling 75219f6d2a6SOded Gabbay struct mutex mutex; 75319f6d2a6SOded Gabbay 75419f6d2a6SOded Gabbay /* 75519f6d2a6SOded Gabbay * In any process, the thread that started main() is the lead 75619f6d2a6SOded Gabbay * thread and outlives the rest. 75719f6d2a6SOded Gabbay * It is here because amd_iommu_bind_pasid wants a task_struct. 758894a8293SFelix Kuehling * It can also be used for safely getting a reference to the 759894a8293SFelix Kuehling * mm_struct of the process. 76019f6d2a6SOded Gabbay */ 76119f6d2a6SOded Gabbay struct task_struct *lead_thread; 76219f6d2a6SOded Gabbay 76319f6d2a6SOded Gabbay /* We want to receive a notification when the mm_struct is destroyed */ 76419f6d2a6SOded Gabbay struct mmu_notifier mmu_notifier; 76519f6d2a6SOded Gabbay 766c7b6bac9SFenghua Yu u32 pasid; 76719f6d2a6SOded Gabbay 76819f6d2a6SOded Gabbay /* 76919f6d2a6SOded Gabbay * List of kfd_process_device structures, 77019f6d2a6SOded Gabbay * one for each device the process is using. 77119f6d2a6SOded Gabbay */ 77219f6d2a6SOded Gabbay struct list_head per_device_data; 77319f6d2a6SOded Gabbay 77445102048SBen Goz struct process_queue_manager pqm; 77545102048SBen Goz 77619f6d2a6SOded Gabbay /*Is the user space process 32 bit?*/ 77719f6d2a6SOded Gabbay bool is_32bit_user_mode; 778f3a39818SAndrew Lewycky 779f3a39818SAndrew Lewycky /* Event-related data */ 780f3a39818SAndrew Lewycky struct mutex event_mutex; 781482f0777SFelix Kuehling /* Event ID allocator and lookup */ 782482f0777SFelix Kuehling struct idr event_idr; 78350cb7dd9SFelix Kuehling /* Event page */ 78450cb7dd9SFelix Kuehling struct kfd_signal_page *signal_page; 785b9a5d0a5SFelix Kuehling size_t signal_mapped_size; 786f3a39818SAndrew Lewycky size_t signal_event_count; 787c986169fSFelix Kuehling bool signal_event_limit_reached; 788403575c4SFelix Kuehling 789403575c4SFelix Kuehling /* Information used for memory eviction */ 790403575c4SFelix Kuehling void *kgd_process_info; 791403575c4SFelix Kuehling /* Eviction fence that is attached to all the BOs of this process. The 792403575c4SFelix Kuehling * fence will be triggered during eviction and new one will be created 793403575c4SFelix Kuehling * during restore 794403575c4SFelix Kuehling */ 795403575c4SFelix Kuehling struct dma_fence *ef; 79626103436SFelix Kuehling 79726103436SFelix Kuehling /* Work items for evicting and restoring BOs */ 79826103436SFelix Kuehling struct delayed_work eviction_work; 79926103436SFelix Kuehling struct delayed_work restore_work; 80026103436SFelix Kuehling /* seqno of the last scheduled eviction */ 80126103436SFelix Kuehling unsigned int last_eviction_seqno; 80226103436SFelix Kuehling /* Approx. the last timestamp (in jiffies) when the process was 80326103436SFelix Kuehling * restored after an eviction 80426103436SFelix Kuehling */ 80526103436SFelix Kuehling unsigned long last_restore_timestamp; 806de9f26bbSKent Russell 807de9f26bbSKent Russell /* Kobj for our procfs */ 808de9f26bbSKent Russell struct kobject *kobj; 8096d220a7eSAmber Lin struct kobject *kobj_queues; 810de9f26bbSKent Russell struct attribute attr_pasid; 8114a488a7aSOded Gabbay }; 8124a488a7aSOded Gabbay 81364d1c3a4SFelix Kuehling #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 81464d1c3a4SFelix Kuehling extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 81564d1c3a4SFelix Kuehling extern struct srcu_struct kfd_processes_srcu; 81664d1c3a4SFelix Kuehling 81776baee6cSOded Gabbay /** 818a4497974SRajneesh Bhardwaj * typedef amdkfd_ioctl_t - typedef for ioctl function pointer. 81976baee6cSOded Gabbay * 820a4497974SRajneesh Bhardwaj * @filep: pointer to file structure. 821a4497974SRajneesh Bhardwaj * @p: amdkfd process pointer. 822a4497974SRajneesh Bhardwaj * @data: pointer to arg that was copied from user. 823a4497974SRajneesh Bhardwaj * 824a4497974SRajneesh Bhardwaj * Return: returns ioctl completion code. 82576baee6cSOded Gabbay */ 82676baee6cSOded Gabbay typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 82776baee6cSOded Gabbay void *data); 82876baee6cSOded Gabbay 82976baee6cSOded Gabbay struct amdkfd_ioctl_desc { 83076baee6cSOded Gabbay unsigned int cmd; 83176baee6cSOded Gabbay int flags; 83276baee6cSOded Gabbay amdkfd_ioctl_t *func; 83376baee6cSOded Gabbay unsigned int cmd_drv; 83476baee6cSOded Gabbay const char *name; 83576baee6cSOded Gabbay }; 83667f7cf9fSshaoyunl bool kfd_dev_is_large_bar(struct kfd_dev *dev); 83776baee6cSOded Gabbay 8381679ae8fSFelix Kuehling int kfd_process_create_wq(void); 83919f6d2a6SOded Gabbay void kfd_process_destroy_wq(void); 840373d7080SFelix Kuehling struct kfd_process *kfd_create_process(struct file *filep); 84119f6d2a6SOded Gabbay struct kfd_process *kfd_get_process(const struct task_struct *); 842c7b6bac9SFenghua Yu struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid); 84326103436SFelix Kuehling struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 844abb208a8SFelix Kuehling void kfd_unref_process(struct kfd_process *p); 8456b95e797SFelix Kuehling int kfd_process_evict_queues(struct kfd_process *p); 8466b95e797SFelix Kuehling int kfd_process_restore_queues(struct kfd_process *p); 84726103436SFelix Kuehling void kfd_suspend_all_processes(void); 84826103436SFelix Kuehling int kfd_resume_all_processes(void); 84919f6d2a6SOded Gabbay 850b84394e2SFelix Kuehling int kfd_process_device_init_vm(struct kfd_process_device *pdd, 851b84394e2SFelix Kuehling struct file *drm_file); 85264c7f8cfSBen Goz struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 85364c7f8cfSBen Goz struct kfd_process *p); 85419f6d2a6SOded Gabbay struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 855093c7d8cSAlexey Skidanov struct kfd_process *p); 856093c7d8cSAlexey Skidanov struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 857093c7d8cSAlexey Skidanov struct kfd_process *p); 85819f6d2a6SOded Gabbay 859df03ef93SHarish Kasiviswanathan int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 860373d7080SFelix Kuehling struct vm_area_struct *vma); 861373d7080SFelix Kuehling 86252b29d73SFelix Kuehling /* KFD process API for creating and translating handles */ 86352b29d73SFelix Kuehling int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 86452b29d73SFelix Kuehling void *mem); 86552b29d73SFelix Kuehling void *kfd_process_device_translate_handle(struct kfd_process_device *p, 86652b29d73SFelix Kuehling int handle); 86752b29d73SFelix Kuehling void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 86852b29d73SFelix Kuehling int handle); 86952b29d73SFelix Kuehling 870775921edSAlexey Skidanov /* Process device data iterator */ 8718eabaf54SKent Russell struct kfd_process_device *kfd_get_first_process_device_data( 8728eabaf54SKent Russell struct kfd_process *p); 8738eabaf54SKent Russell struct kfd_process_device *kfd_get_next_process_device_data( 8748eabaf54SKent Russell struct kfd_process *p, 875775921edSAlexey Skidanov struct kfd_process_device *pdd); 876775921edSAlexey Skidanov bool kfd_has_process_device_data(struct kfd_process *p); 877775921edSAlexey Skidanov 87819f6d2a6SOded Gabbay /* PASIDs */ 87919f6d2a6SOded Gabbay int kfd_pasid_init(void); 88019f6d2a6SOded Gabbay void kfd_pasid_exit(void); 88119f6d2a6SOded Gabbay bool kfd_set_pasid_limit(unsigned int new_limit); 88219f6d2a6SOded Gabbay unsigned int kfd_get_pasid_limit(void); 883c7b6bac9SFenghua Yu u32 kfd_pasid_alloc(void); 884c7b6bac9SFenghua Yu void kfd_pasid_free(u32 pasid); 88519f6d2a6SOded Gabbay 88619f6d2a6SOded Gabbay /* Doorbells */ 887ef568db7SFelix Kuehling size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 888735df2baSFelix Kuehling int kfd_doorbell_init(struct kfd_dev *kfd); 889735df2baSFelix Kuehling void kfd_doorbell_fini(struct kfd_dev *kfd); 890df03ef93SHarish Kasiviswanathan int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 891df03ef93SHarish Kasiviswanathan struct vm_area_struct *vma); 892ada2b29cSFelix Kuehling void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 89319f6d2a6SOded Gabbay unsigned int *doorbell_off); 89419f6d2a6SOded Gabbay void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 89519f6d2a6SOded Gabbay u32 read_kernel_doorbell(u32 __iomem *db); 896ada2b29cSFelix Kuehling void write_kernel_doorbell(void __iomem *db, u32 value); 8979d7d0248SFelix Kuehling void write_kernel_doorbell64(void __iomem *db, u64 value); 898339903faSYong Zhao unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, 89959d7115dSMukul Joshi struct kfd_process_device *pdd, 900ef568db7SFelix Kuehling unsigned int doorbell_id); 90159d7115dSMukul Joshi phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd); 90259d7115dSMukul Joshi int kfd_alloc_process_doorbells(struct kfd_dev *kfd, 90359d7115dSMukul Joshi unsigned int *doorbell_index); 90459d7115dSMukul Joshi void kfd_free_process_doorbells(struct kfd_dev *kfd, 90559d7115dSMukul Joshi unsigned int doorbell_index); 9066e81090bSOded Gabbay /* GTT Sub-Allocator */ 9076e81090bSOded Gabbay 9086e81090bSOded Gabbay int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 9096e81090bSOded Gabbay struct kfd_mem_obj **mem_obj); 9106e81090bSOded Gabbay 9116e81090bSOded Gabbay int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 9126e81090bSOded Gabbay 9134a488a7aSOded Gabbay extern struct device *kfd_device; 9144a488a7aSOded Gabbay 915de9f26bbSKent Russell /* KFD's procfs */ 916de9f26bbSKent Russell void kfd_procfs_init(void); 917de9f26bbSKent Russell void kfd_procfs_shutdown(void); 9186d220a7eSAmber Lin int kfd_procfs_add_queue(struct queue *q); 9196d220a7eSAmber Lin void kfd_procfs_del_queue(struct queue *q); 920de9f26bbSKent Russell 9215b5c4e40SEvgeny Pinchuk /* Topology */ 9225b5c4e40SEvgeny Pinchuk int kfd_topology_init(void); 9235b5c4e40SEvgeny Pinchuk void kfd_topology_shutdown(void); 9245b5c4e40SEvgeny Pinchuk int kfd_topology_add_device(struct kfd_dev *gpu); 9255b5c4e40SEvgeny Pinchuk int kfd_topology_remove_device(struct kfd_dev *gpu); 9263a87177eSHarish Kasiviswanathan struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 9273a87177eSHarish Kasiviswanathan uint32_t proximity_domain); 92844d8cc6fSYong Zhao struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); 9295b5c4e40SEvgeny Pinchuk struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 9305b5c4e40SEvgeny Pinchuk struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 9311dde0ea9SFelix Kuehling struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd); 9326d82eb0eSHarish Kasiviswanathan int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 933520b8fb7SFelix Kuehling int kfd_numa_node_to_apic_id(int numa_node_id); 9346127896fSHuang Rui void kfd_double_confirm_iommu_support(struct kfd_dev *gpu); 9355b5c4e40SEvgeny Pinchuk 9364a488a7aSOded Gabbay /* Interrupts */ 9372249d558SAndrew Lewycky int kfd_interrupt_init(struct kfd_dev *dev); 9382249d558SAndrew Lewycky void kfd_interrupt_exit(struct kfd_dev *dev); 9392249d558SAndrew Lewycky bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 94058e69886SLan Xiao bool interrupt_is_wanted(struct kfd_dev *dev, 94158e69886SLan Xiao const uint32_t *ih_ring_entry, 94258e69886SLan Xiao uint32_t *patched_ihre, bool *flag); 9434a488a7aSOded Gabbay 94419f6d2a6SOded Gabbay /* amdkfd Apertures */ 94519f6d2a6SOded Gabbay int kfd_init_apertures(struct kfd_process *process); 94619f6d2a6SOded Gabbay 947*7c9631afSJay Cornwall void kfd_process_set_trap_handler(struct qcm_process_device *qpd, 948*7c9631afSJay Cornwall uint64_t tba_addr, 949*7c9631afSJay Cornwall uint64_t tma_addr); 950*7c9631afSJay Cornwall 951ed6e6a34SBen Goz /* Queue Context Management */ 952e88a614cSEdward O'Callaghan int init_queue(struct queue **q, const struct queue_properties *properties); 953ed6e6a34SBen Goz void uninit_queue(struct queue *q); 95445102048SBen Goz void print_queue_properties(struct queue_properties *q); 955ed6e6a34SBen Goz void print_queue(struct queue *q); 956ed6e6a34SBen Goz 9574b8f589bSBen Goz struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 9584b8f589bSBen Goz struct kfd_dev *dev); 959ee04955aSFelix Kuehling struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 960ee04955aSFelix Kuehling struct kfd_dev *dev); 9614b8f589bSBen Goz struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 9624b8f589bSBen Goz struct kfd_dev *dev); 963ee04955aSFelix Kuehling struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 964ee04955aSFelix Kuehling struct kfd_dev *dev); 965b91d43ddSFelix Kuehling struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 966b91d43ddSFelix Kuehling struct kfd_dev *dev); 96714328aa5SPhilip Cox struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, 96814328aa5SPhilip Cox struct kfd_dev *dev); 96964c7f8cfSBen Goz struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 97064c7f8cfSBen Goz void device_queue_manager_uninit(struct device_queue_manager *dqm); 971241f24f8SBen Goz struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 972241f24f8SBen Goz enum kfd_queue_type type); 973c2a77fdeSFelix Kuehling void kernel_queue_uninit(struct kernel_queue *kq, bool hanging); 974c7b6bac9SFenghua Yu int kfd_process_vm_fault(struct device_queue_manager *dqm, u32 pasid); 975241f24f8SBen Goz 97645102048SBen Goz /* Process Queue Manager */ 97745102048SBen Goz struct process_queue_node { 97845102048SBen Goz struct queue *q; 97945102048SBen Goz struct kernel_queue *kq; 98045102048SBen Goz struct list_head process_queue_list; 98145102048SBen Goz }; 98245102048SBen Goz 9839fd3f1bfSFelix Kuehling void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 9849fd3f1bfSFelix Kuehling void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 98545102048SBen Goz int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 98645102048SBen Goz void pqm_uninit(struct process_queue_manager *pqm); 98745102048SBen Goz int pqm_create_queue(struct process_queue_manager *pqm, 98845102048SBen Goz struct kfd_dev *dev, 98945102048SBen Goz struct file *f, 99045102048SBen Goz struct queue_properties *properties, 991e47a8b52SYong Zhao unsigned int *qid, 992e47a8b52SYong Zhao uint32_t *p_doorbell_offset_in_process); 99345102048SBen Goz int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 99445102048SBen Goz int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 99545102048SBen Goz struct queue_properties *p); 99639e7f331SFelix Kuehling int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid, 99739e7f331SFelix Kuehling struct queue_properties *p); 998eb82da1dSOak Zeng int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, 999eb82da1dSOak Zeng void *gws); 1000fbeb661bSYair Shachar struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 1001fbeb661bSYair Shachar unsigned int qid); 10025bb4b78bSOak Zeng struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, 10035bb4b78bSOak Zeng unsigned int qid); 10045df099e8SJay Cornwall int pqm_get_wave_state(struct process_queue_manager *pqm, 10055df099e8SJay Cornwall unsigned int qid, 10065df099e8SJay Cornwall void __user *ctl_stack, 10075df099e8SJay Cornwall u32 *ctl_stack_used_size, 10085df099e8SJay Cornwall u32 *save_area_used_size); 100945102048SBen Goz 1010788bf83dSYair Shachar int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 1011788bf83dSYair Shachar unsigned int fence_value, 10128c72c3d7SYong Zhao unsigned int timeout_ms); 1013788bf83dSYair Shachar 1014ed6e6a34SBen Goz /* Packet Manager */ 1015ed6e6a34SBen Goz 101664c7f8cfSBen Goz #define KFD_FENCE_COMPLETED (100) 101764c7f8cfSBen Goz #define KFD_FENCE_INIT (10) 1018241f24f8SBen Goz 1019ed6e6a34SBen Goz struct packet_manager { 1020ed6e6a34SBen Goz struct device_queue_manager *dqm; 1021ed6e6a34SBen Goz struct kernel_queue *priv_queue; 1022ed6e6a34SBen Goz struct mutex lock; 1023ed6e6a34SBen Goz bool allocated; 1024ed6e6a34SBen Goz struct kfd_mem_obj *ib_buffer_obj; 1025851a645eSFelix Kuehling unsigned int ib_size_bytes; 1026819ec5acSFelix Kuehling bool is_over_subscription; 1027f6e27ff1SFelix Kuehling 1028f6e27ff1SFelix Kuehling const struct packet_manager_funcs *pmf; 1029ed6e6a34SBen Goz }; 1030ed6e6a34SBen Goz 1031f6e27ff1SFelix Kuehling struct packet_manager_funcs { 1032f6e27ff1SFelix Kuehling /* Support ASIC-specific packet formats for PM4 packets */ 1033f6e27ff1SFelix Kuehling int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 1034f6e27ff1SFelix Kuehling struct qcm_process_device *qpd); 1035f6e27ff1SFelix Kuehling int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 1036f6e27ff1SFelix Kuehling uint64_t ib, size_t ib_size_in_dwords, bool chain); 1037f6e27ff1SFelix Kuehling int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 1038f6e27ff1SFelix Kuehling struct scheduling_resources *res); 1039f6e27ff1SFelix Kuehling int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 1040f6e27ff1SFelix Kuehling struct queue *q, bool is_static); 1041f6e27ff1SFelix Kuehling int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 1042f6e27ff1SFelix Kuehling enum kfd_queue_type type, 1043f6e27ff1SFelix Kuehling enum kfd_unmap_queues_filter mode, 1044f6e27ff1SFelix Kuehling uint32_t filter_param, bool reset, 1045f6e27ff1SFelix Kuehling unsigned int sdma_engine); 1046f6e27ff1SFelix Kuehling int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 1047f6e27ff1SFelix Kuehling uint64_t fence_address, uint32_t fence_value); 1048f6e27ff1SFelix Kuehling int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 1049f6e27ff1SFelix Kuehling 1050f6e27ff1SFelix Kuehling /* Packet sizes */ 1051f6e27ff1SFelix Kuehling int map_process_size; 1052f6e27ff1SFelix Kuehling int runlist_size; 1053f6e27ff1SFelix Kuehling int set_resources_size; 1054f6e27ff1SFelix Kuehling int map_queues_size; 1055f6e27ff1SFelix Kuehling int unmap_queues_size; 1056f6e27ff1SFelix Kuehling int query_status_size; 1057f6e27ff1SFelix Kuehling int release_mem_size; 1058f6e27ff1SFelix Kuehling }; 1059f6e27ff1SFelix Kuehling 1060f6e27ff1SFelix Kuehling extern const struct packet_manager_funcs kfd_vi_pm_funcs; 1061454150b1SFelix Kuehling extern const struct packet_manager_funcs kfd_v9_pm_funcs; 1062f6e27ff1SFelix Kuehling 106364c7f8cfSBen Goz int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 1064c2a77fdeSFelix Kuehling void pm_uninit(struct packet_manager *pm, bool hanging); 106564c7f8cfSBen Goz int pm_send_set_resources(struct packet_manager *pm, 106664c7f8cfSBen Goz struct scheduling_resources *res); 106764c7f8cfSBen Goz int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 106864c7f8cfSBen Goz int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 106964c7f8cfSBen Goz uint32_t fence_value); 107064c7f8cfSBen Goz 107164c7f8cfSBen Goz int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 10727da2bcf8SYong Zhao enum kfd_unmap_queues_filter mode, 107364c7f8cfSBen Goz uint32_t filter_param, bool reset, 107464c7f8cfSBen Goz unsigned int sdma_engine); 107564c7f8cfSBen Goz 1076241f24f8SBen Goz void pm_release_ib(struct packet_manager *pm); 1077241f24f8SBen Goz 1078454150b1SFelix Kuehling /* Following PM funcs can be shared among VI and AI */ 1079454150b1SFelix Kuehling unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 108014328aa5SPhilip Cox 108119f6d2a6SOded Gabbay uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 108219f6d2a6SOded Gabbay 1083f3a39818SAndrew Lewycky /* Events */ 1084f3a39818SAndrew Lewycky extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 1085ca750681SFelix Kuehling extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 1086ca750681SFelix Kuehling 1087930c5ff4SAlexey Skidanov extern const struct kfd_device_global_init_class device_global_init_class_cik; 1088f3a39818SAndrew Lewycky 1089f3a39818SAndrew Lewycky void kfd_event_init_process(struct kfd_process *p); 1090f3a39818SAndrew Lewycky void kfd_event_free_process(struct kfd_process *p); 1091f3a39818SAndrew Lewycky int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 1092f3a39818SAndrew Lewycky int kfd_wait_on_events(struct kfd_process *p, 109359d3e8beSAlexey Skidanov uint32_t num_events, void __user *data, 1094f3a39818SAndrew Lewycky bool all, uint32_t user_timeout_ms, 1095fdf0c833SFelix Kuehling uint32_t *wait_result); 1096c7b6bac9SFenghua Yu void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, 1097f3a39818SAndrew Lewycky uint32_t valid_id_bits); 109859d3e8beSAlexey Skidanov void kfd_signal_iommu_event(struct kfd_dev *dev, 1099c7b6bac9SFenghua Yu u32 pasid, unsigned long address, 110059d3e8beSAlexey Skidanov bool is_write_requested, bool is_execute_requested); 1101c7b6bac9SFenghua Yu void kfd_signal_hw_exception_event(u32 pasid); 1102f3a39818SAndrew Lewycky int kfd_set_event(struct kfd_process *p, uint32_t event_id); 1103f3a39818SAndrew Lewycky int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 11040fc8011fSFelix Kuehling int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 11050fc8011fSFelix Kuehling uint64_t size); 1106f3a39818SAndrew Lewycky int kfd_event_create(struct file *devkfd, struct kfd_process *p, 1107f3a39818SAndrew Lewycky uint32_t event_type, bool auto_reset, uint32_t node_id, 1108f3a39818SAndrew Lewycky uint32_t *event_id, uint32_t *event_trigger_data, 1109f3a39818SAndrew Lewycky uint64_t *event_page_offset, uint32_t *event_slot_index); 1110f3a39818SAndrew Lewycky int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 1111f3a39818SAndrew Lewycky 1112c7b6bac9SFenghua Yu void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, 11132640c3faSshaoyunl struct kfd_vm_fault_info *info); 11142640c3faSshaoyunl 1115e42051d2SShaoyun Liu void kfd_signal_reset_event(struct kfd_dev *dev); 1116e42051d2SShaoyun Liu 1117403575c4SFelix Kuehling void kfd_flush_tlb(struct kfd_process_device *pdd); 1118403575c4SFelix Kuehling 1119c3447e81SBen Goz int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 1120c3447e81SBen Goz 1121e42051d2SShaoyun Liu bool kfd_is_locked(void); 1122e42051d2SShaoyun Liu 1123f756e631SHarish Kasiviswanathan /* Compute profile */ 1124f756e631SHarish Kasiviswanathan void kfd_inc_compute_active(struct kfd_dev *dev); 1125f756e631SHarish Kasiviswanathan void kfd_dec_compute_active(struct kfd_dev *dev); 1126f756e631SHarish Kasiviswanathan 11276b855f7bSHarish Kasiviswanathan /* Cgroup Support */ 11286b855f7bSHarish Kasiviswanathan /* Check with device cgroup if @kfd device is accessible */ 11296b855f7bSHarish Kasiviswanathan static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) 11306b855f7bSHarish Kasiviswanathan { 1131eec8fd02SOdin Ugedal #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF) 11326b855f7bSHarish Kasiviswanathan struct drm_device *ddev = kfd->ddev; 11336b855f7bSHarish Kasiviswanathan 113499c7b309SLorenz Brun return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR, 11356b855f7bSHarish Kasiviswanathan ddev->render->index, 11366b855f7bSHarish Kasiviswanathan DEVCG_ACC_WRITE | DEVCG_ACC_READ); 11376b855f7bSHarish Kasiviswanathan #else 11386b855f7bSHarish Kasiviswanathan return 0; 11396b855f7bSHarish Kasiviswanathan #endif 11406b855f7bSHarish Kasiviswanathan } 11416b855f7bSHarish Kasiviswanathan 1142851a645eSFelix Kuehling /* Debugfs */ 1143851a645eSFelix Kuehling #if defined(CONFIG_DEBUG_FS) 1144851a645eSFelix Kuehling 1145851a645eSFelix Kuehling void kfd_debugfs_init(void); 1146851a645eSFelix Kuehling void kfd_debugfs_fini(void); 1147851a645eSFelix Kuehling int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 1148851a645eSFelix Kuehling int pqm_debugfs_mqds(struct seq_file *m, void *data); 1149851a645eSFelix Kuehling int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 1150851a645eSFelix Kuehling int dqm_debugfs_hqds(struct seq_file *m, void *data); 1151851a645eSFelix Kuehling int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 1152851a645eSFelix Kuehling int pm_debugfs_runlist(struct seq_file *m, void *data); 1153851a645eSFelix Kuehling 1154a29ec470SShaoyun Liu int kfd_debugfs_hang_hws(struct kfd_dev *dev); 1155a29ec470SShaoyun Liu int pm_debugfs_hang_hws(struct packet_manager *pm); 1156a29ec470SShaoyun Liu int dqm_debugfs_execute_queues(struct device_queue_manager *dqm); 1157a29ec470SShaoyun Liu 1158851a645eSFelix Kuehling #else 1159851a645eSFelix Kuehling 1160851a645eSFelix Kuehling static inline void kfd_debugfs_init(void) {} 1161851a645eSFelix Kuehling static inline void kfd_debugfs_fini(void) {} 1162851a645eSFelix Kuehling 1163851a645eSFelix Kuehling #endif 1164851a645eSFelix Kuehling 11654a488a7aSOded Gabbay #endif 1166