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 2950c663695SDivya Shikre /* UUID */ 2960c663695SDivya Shikre uint64_t unique_id; 297d35f00d8SEric Huang 298d35f00d8SEric Huang bool pci_atomic_requested; 2999b54d201SEric Huang 3006127896fSHuang Rui /* Use IOMMU v2 flag */ 3016127896fSHuang Rui bool use_iommu_v2; 3026127896fSHuang Rui 3039b54d201SEric Huang /* SRAM ECC flag */ 3049b54d201SEric Huang atomic_t sram_ecc_flag; 305f756e631SHarish Kasiviswanathan 306f756e631SHarish Kasiviswanathan /* Compute Profile ref. count */ 307f756e631SHarish Kasiviswanathan atomic_t compute_profile; 308e09d4fc8SOak Zeng 309a4497974SRajneesh Bhardwaj /* Global GWS resource shared between processes */ 310e09d4fc8SOak Zeng void *gws; 311938a0650SAmber Lin 312938a0650SAmber Lin /* Clients watching SMI events */ 313938a0650SAmber Lin struct list_head smi_clients; 314938a0650SAmber Lin spinlock_t smi_lock; 31555977744SMukul Joshi 31655977744SMukul Joshi uint32_t reset_seq_num; 31759d7115dSMukul Joshi 31859d7115dSMukul Joshi struct ida doorbell_ida; 31959d7115dSMukul Joshi unsigned int max_doorbell_slices; 3204a488a7aSOded Gabbay }; 3214a488a7aSOded Gabbay 32219f6d2a6SOded Gabbay enum kfd_mempool { 32319f6d2a6SOded Gabbay KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 32419f6d2a6SOded Gabbay KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 32519f6d2a6SOded Gabbay KFD_MEMPOOL_FRAMEBUFFER = 3, 32619f6d2a6SOded Gabbay }; 32719f6d2a6SOded Gabbay 3284a488a7aSOded Gabbay /* Character device interface */ 3294a488a7aSOded Gabbay int kfd_chardev_init(void); 3304a488a7aSOded Gabbay void kfd_chardev_exit(void); 3314a488a7aSOded Gabbay struct device *kfd_chardev(void); 3324a488a7aSOded Gabbay 333241f24f8SBen Goz /** 334a4497974SRajneesh Bhardwaj * enum kfd_unmap_queues_filter - Enum for queue filters. 335241f24f8SBen Goz * 3367da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 337241f24f8SBen Goz * 3387da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 339241f24f8SBen Goz * running queues list. 340241f24f8SBen Goz * 3417da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 342241f24f8SBen Goz * specific process. 343241f24f8SBen Goz * 344241f24f8SBen Goz */ 3457da2bcf8SYong Zhao enum kfd_unmap_queues_filter { 3467da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 3477da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 3487da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 3497da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_BY_PASID 350241f24f8SBen Goz }; 35119f6d2a6SOded Gabbay 352ed8aab45SBen Goz /** 353a4497974SRajneesh Bhardwaj * enum kfd_queue_type - Enum for various queue types. 354ed8aab45SBen Goz * 355ed8aab45SBen Goz * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 356ed8aab45SBen Goz * 357a4497974SRajneesh Bhardwaj * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type. 358ed8aab45SBen Goz * 359ed8aab45SBen Goz * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 360ed8aab45SBen Goz * 361ed8aab45SBen Goz * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 362a4497974SRajneesh Bhardwaj * 363a4497974SRajneesh Bhardwaj * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface. 364ed8aab45SBen Goz */ 365ed8aab45SBen Goz enum kfd_queue_type { 366ed8aab45SBen Goz KFD_QUEUE_TYPE_COMPUTE, 367ed8aab45SBen Goz KFD_QUEUE_TYPE_SDMA, 368ed8aab45SBen Goz KFD_QUEUE_TYPE_HIQ, 3691b4670f6SOak Zeng KFD_QUEUE_TYPE_DIQ, 3701b4670f6SOak Zeng KFD_QUEUE_TYPE_SDMA_XGMI 371ed8aab45SBen Goz }; 372ed8aab45SBen Goz 3736e99df57SBen Goz enum kfd_queue_format { 3746e99df57SBen Goz KFD_QUEUE_FORMAT_PM4, 3756e99df57SBen Goz KFD_QUEUE_FORMAT_AQL 3766e99df57SBen Goz }; 3776e99df57SBen Goz 3780ccbc7cdSOak Zeng enum KFD_QUEUE_PRIORITY { 3790ccbc7cdSOak Zeng KFD_QUEUE_PRIORITY_MINIMUM = 0, 3800ccbc7cdSOak Zeng KFD_QUEUE_PRIORITY_MAXIMUM = 15 3810ccbc7cdSOak Zeng }; 3820ccbc7cdSOak Zeng 383ed8aab45SBen Goz /** 384ed8aab45SBen Goz * struct queue_properties 385ed8aab45SBen Goz * 386ed8aab45SBen Goz * @type: The queue type. 387ed8aab45SBen Goz * 388ed8aab45SBen Goz * @queue_id: Queue identifier. 389ed8aab45SBen Goz * 390ed8aab45SBen Goz * @queue_address: Queue ring buffer address. 391ed8aab45SBen Goz * 392ed8aab45SBen Goz * @queue_size: Queue ring buffer size. 393ed8aab45SBen Goz * 394ed8aab45SBen Goz * @priority: Defines the queue priority relative to other queues in the 395ed8aab45SBen Goz * process. 396ed8aab45SBen Goz * This is just an indication and HW scheduling may override the priority as 397ed8aab45SBen Goz * necessary while keeping the relative prioritization. 398ed8aab45SBen Goz * the priority granularity is from 0 to f which f is the highest priority. 399ed8aab45SBen Goz * currently all queues are initialized with the highest priority. 400ed8aab45SBen Goz * 401ed8aab45SBen Goz * @queue_percent: This field is partially implemented and currently a zero in 402ed8aab45SBen Goz * this field defines that the queue is non active. 403ed8aab45SBen Goz * 404ed8aab45SBen Goz * @read_ptr: User space address which points to the number of dwords the 405ed8aab45SBen Goz * cp read from the ring buffer. This field updates automatically by the H/W. 406ed8aab45SBen Goz * 407ed8aab45SBen Goz * @write_ptr: Defines the number of dwords written to the ring buffer. 408ed8aab45SBen Goz * 409a4497974SRajneesh Bhardwaj * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring 410a4497974SRajneesh Bhardwaj * buffer. This field should be similar to write_ptr and the user should 411a4497974SRajneesh Bhardwaj * update this field after updating the write_ptr. 412ed8aab45SBen Goz * 413ed8aab45SBen Goz * @doorbell_off: The doorbell offset in the doorbell pci-bar. 414ed8aab45SBen Goz * 4158eabaf54SKent Russell * @is_interop: Defines if this is a interop queue. Interop queue means that 4168eabaf54SKent Russell * the queue can access both graphics and compute resources. 417ed8aab45SBen Goz * 41826103436SFelix Kuehling * @is_evicted: Defines if the queue is evicted. Only active queues 41926103436SFelix Kuehling * are evicted, rendering them inactive. 42026103436SFelix Kuehling * 42126103436SFelix Kuehling * @is_active: Defines if the queue is active or not. @is_active and 42226103436SFelix Kuehling * @is_evicted are protected by the DQM lock. 423ed8aab45SBen Goz * 424b8020b03SJoseph Greathouse * @is_gws: Defines if the queue has been updated to be GWS-capable or not. 425b8020b03SJoseph Greathouse * @is_gws should be protected by the DQM lock, since changing it can yield the 426b8020b03SJoseph Greathouse * possibility of updating DQM state on number of GWS queues. 427b8020b03SJoseph Greathouse * 428ed8aab45SBen Goz * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 429ed8aab45SBen Goz * of the queue. 430ed8aab45SBen Goz * 431ed8aab45SBen Goz * This structure represents the queue properties for each queue no matter if 432ed8aab45SBen Goz * it's user mode or kernel mode queue. 433ed8aab45SBen Goz * 434ed8aab45SBen Goz */ 435ed8aab45SBen Goz struct queue_properties { 436ed8aab45SBen Goz enum kfd_queue_type type; 4376e99df57SBen Goz enum kfd_queue_format format; 438ed8aab45SBen Goz unsigned int queue_id; 439ed8aab45SBen Goz uint64_t queue_address; 440ed8aab45SBen Goz uint64_t queue_size; 441ed8aab45SBen Goz uint32_t priority; 442ed8aab45SBen Goz uint32_t queue_percent; 443ed8aab45SBen Goz uint32_t *read_ptr; 444ed8aab45SBen Goz uint32_t *write_ptr; 445ada2b29cSFelix Kuehling void __iomem *doorbell_ptr; 446ed8aab45SBen Goz uint32_t doorbell_off; 447ed8aab45SBen Goz bool is_interop; 44826103436SFelix Kuehling bool is_evicted; 449ed8aab45SBen Goz bool is_active; 450b8020b03SJoseph Greathouse bool is_gws; 451ed8aab45SBen Goz /* Not relevant for user mode queues in cp scheduling */ 452ed8aab45SBen Goz unsigned int vmid; 45377669eb8SBen Goz /* Relevant only for sdma queues*/ 45477669eb8SBen Goz uint32_t sdma_engine_id; 45577669eb8SBen Goz uint32_t sdma_queue_id; 45677669eb8SBen Goz uint32_t sdma_vm_addr; 457ff3d04a1SBen Goz /* Relevant only for VI */ 458ff3d04a1SBen Goz uint64_t eop_ring_buffer_address; 459ff3d04a1SBen Goz uint32_t eop_ring_buffer_size; 460ff3d04a1SBen Goz uint64_t ctx_save_restore_area_address; 461ff3d04a1SBen Goz uint32_t ctx_save_restore_area_size; 462373d7080SFelix Kuehling uint32_t ctl_stack_size; 463373d7080SFelix Kuehling uint64_t tba_addr; 464373d7080SFelix Kuehling uint64_t tma_addr; 46539e7f331SFelix Kuehling /* Relevant for CU */ 46639e7f331SFelix Kuehling uint32_t cu_mask_count; /* Must be a multiple of 32 */ 46739e7f331SFelix Kuehling uint32_t *cu_mask; 468ed8aab45SBen Goz }; 469ed8aab45SBen Goz 470bb2d2128SFelix Kuehling #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ 471bb2d2128SFelix Kuehling (q).queue_address != 0 && \ 472bb2d2128SFelix Kuehling (q).queue_percent > 0 && \ 473bb2d2128SFelix Kuehling !(q).is_evicted) 474bb2d2128SFelix Kuehling 475ed8aab45SBen Goz /** 476ed8aab45SBen Goz * struct queue 477ed8aab45SBen Goz * 478ed8aab45SBen Goz * @list: Queue linked list. 479ed8aab45SBen Goz * 480a4497974SRajneesh Bhardwaj * @mqd: The queue MQD (memory queue descriptor). 481ed8aab45SBen Goz * 482ed8aab45SBen Goz * @mqd_mem_obj: The MQD local gpu memory object. 483ed8aab45SBen Goz * 484ed8aab45SBen Goz * @gart_mqd_addr: The MQD gart mc address. 485ed8aab45SBen Goz * 486ed8aab45SBen Goz * @properties: The queue properties. 487ed8aab45SBen Goz * 488ed8aab45SBen Goz * @mec: Used only in no cp scheduling mode and identifies to micro engine id 489a4497974SRajneesh Bhardwaj * that the queue should be executed on. 490ed8aab45SBen Goz * 4918eabaf54SKent Russell * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 4928eabaf54SKent Russell * id. 493ed8aab45SBen Goz * 494ed8aab45SBen Goz * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 495ed8aab45SBen Goz * 496ed8aab45SBen Goz * @process: The kfd process that created this queue. 497ed8aab45SBen Goz * 498ed8aab45SBen Goz * @device: The kfd device that created this queue. 499ed8aab45SBen Goz * 500eb82da1dSOak Zeng * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL 501eb82da1dSOak Zeng * otherwise. 502eb82da1dSOak Zeng * 503ed8aab45SBen Goz * This structure represents user mode compute queues. 504ed8aab45SBen Goz * It contains all the necessary data to handle such queues. 505ed8aab45SBen Goz * 506ed8aab45SBen Goz */ 507ed8aab45SBen Goz 508ed8aab45SBen Goz struct queue { 509ed8aab45SBen Goz struct list_head list; 510ed8aab45SBen Goz void *mqd; 511ed8aab45SBen Goz struct kfd_mem_obj *mqd_mem_obj; 512ed8aab45SBen Goz uint64_t gart_mqd_addr; 513ed8aab45SBen Goz struct queue_properties properties; 514ed8aab45SBen Goz 515ed8aab45SBen Goz uint32_t mec; 516ed8aab45SBen Goz uint32_t pipe; 517ed8aab45SBen Goz uint32_t queue; 518ed8aab45SBen Goz 51977669eb8SBen Goz unsigned int sdma_id; 520ef568db7SFelix Kuehling unsigned int doorbell_id; 52177669eb8SBen Goz 522ed8aab45SBen Goz struct kfd_process *process; 523ed8aab45SBen Goz struct kfd_dev *device; 524eb82da1dSOak Zeng void *gws; 5256d220a7eSAmber Lin 5266d220a7eSAmber Lin /* procfs */ 5276d220a7eSAmber Lin struct kobject kobj; 528ed8aab45SBen Goz }; 529ed8aab45SBen Goz 5306e99df57SBen Goz enum KFD_MQD_TYPE { 531d7c0b047SYong Zhao KFD_MQD_TYPE_HIQ = 0, /* for hiq */ 53285d258f9SBen Goz KFD_MQD_TYPE_CP, /* for cp queues and diq */ 53385d258f9SBen Goz KFD_MQD_TYPE_SDMA, /* for sdma queues */ 53459f650a0SOak Zeng KFD_MQD_TYPE_DIQ, /* for diq */ 5356e99df57SBen Goz KFD_MQD_TYPE_MAX 5366e99df57SBen Goz }; 5376e99df57SBen Goz 5380ccbc7cdSOak Zeng enum KFD_PIPE_PRIORITY { 5390ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_LOW = 0, 5400ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_MEDIUM, 5410ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_HIGH 5420ccbc7cdSOak Zeng }; 5430ccbc7cdSOak Zeng 544241f24f8SBen Goz struct scheduling_resources { 545241f24f8SBen Goz unsigned int vmid_mask; 546241f24f8SBen Goz enum kfd_queue_type type; 547241f24f8SBen Goz uint64_t queue_mask; 548241f24f8SBen Goz uint64_t gws_mask; 549241f24f8SBen Goz uint32_t oac_mask; 550241f24f8SBen Goz uint32_t gds_heap_base; 551241f24f8SBen Goz uint32_t gds_heap_size; 552241f24f8SBen Goz }; 553241f24f8SBen Goz 554241f24f8SBen Goz struct process_queue_manager { 555241f24f8SBen Goz /* data */ 556241f24f8SBen Goz struct kfd_process *process; 557241f24f8SBen Goz struct list_head queues; 558241f24f8SBen Goz unsigned long *queue_slot_bitmap; 559241f24f8SBen Goz }; 560241f24f8SBen Goz 561241f24f8SBen Goz struct qcm_process_device { 562241f24f8SBen Goz /* The Device Queue Manager that owns this data */ 563241f24f8SBen Goz struct device_queue_manager *dqm; 564241f24f8SBen Goz struct process_queue_manager *pqm; 565241f24f8SBen Goz /* Queues list */ 566241f24f8SBen Goz struct list_head queues_list; 567241f24f8SBen Goz struct list_head priv_queue_list; 568241f24f8SBen Goz 569241f24f8SBen Goz unsigned int queue_count; 570241f24f8SBen Goz unsigned int vmid; 571241f24f8SBen Goz bool is_debug; 57226103436SFelix Kuehling unsigned int evicted; /* eviction counter, 0=active */ 5739fd3f1bfSFelix Kuehling 5749fd3f1bfSFelix Kuehling /* This flag tells if we should reset all wavefronts on 5759fd3f1bfSFelix Kuehling * process termination 5769fd3f1bfSFelix Kuehling */ 5779fd3f1bfSFelix Kuehling bool reset_wavefronts; 5789fd3f1bfSFelix Kuehling 579b8020b03SJoseph Greathouse /* This flag tells us if this process has a GWS-capable 580b8020b03SJoseph Greathouse * queue that will be mapped into the runlist. It's 581b8020b03SJoseph Greathouse * possible to request a GWS BO, but not have the queue 582b8020b03SJoseph Greathouse * currently mapped, and this changes how the MAP_PROCESS 583b8020b03SJoseph Greathouse * PM4 packet is configured. 584b8020b03SJoseph Greathouse */ 585b8020b03SJoseph Greathouse bool mapped_gws_queue; 586b8020b03SJoseph Greathouse 587a4497974SRajneesh Bhardwaj /* All the memory management data should be here too */ 588241f24f8SBen Goz uint64_t gds_context_area; 589435e2f97SYong Zhao /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */ 590e715c6d0SShaoyun Liu uint64_t page_table_base; 591241f24f8SBen Goz uint32_t sh_mem_config; 592241f24f8SBen Goz uint32_t sh_mem_bases; 593241f24f8SBen Goz uint32_t sh_mem_ape1_base; 594241f24f8SBen Goz uint32_t sh_mem_ape1_limit; 595241f24f8SBen Goz uint32_t gds_size; 596241f24f8SBen Goz uint32_t num_gws; 597241f24f8SBen Goz uint32_t num_oac; 5986a1c9510SMoses Reuben uint32_t sh_hidden_private_base; 599373d7080SFelix Kuehling 600373d7080SFelix Kuehling /* CWSR memory */ 601373d7080SFelix Kuehling void *cwsr_kaddr; 602d01994c2SFelix Kuehling uint64_t cwsr_base; 603373d7080SFelix Kuehling uint64_t tba_addr; 604373d7080SFelix Kuehling uint64_t tma_addr; 605d01994c2SFelix Kuehling 606d01994c2SFelix Kuehling /* IB memory */ 607d01994c2SFelix Kuehling uint64_t ib_base; 608552764b6SFelix Kuehling void *ib_kaddr; 609ef568db7SFelix Kuehling 610ef568db7SFelix Kuehling /* doorbell resources per process per device */ 611ef568db7SFelix Kuehling unsigned long *doorbell_bitmap; 612241f24f8SBen Goz }; 613241f24f8SBen Goz 61426103436SFelix Kuehling /* KFD Memory Eviction */ 61526103436SFelix Kuehling 61626103436SFelix Kuehling /* Approx. wait time before attempting to restore evicted BOs */ 61726103436SFelix Kuehling #define PROCESS_RESTORE_TIME_MS 100 61826103436SFelix Kuehling /* Approx. back off time if restore fails due to lack of memory */ 61926103436SFelix Kuehling #define PROCESS_BACK_OFF_TIME_MS 100 62026103436SFelix Kuehling /* Approx. time before evicting the process again */ 62126103436SFelix Kuehling #define PROCESS_ACTIVE_TIME_MS 10 62226103436SFelix Kuehling 6235ec7e028SFelix Kuehling /* 8 byte handle containing GPU ID in the most significant 4 bytes and 6245ec7e028SFelix Kuehling * idr_handle in the least significant 4 bytes 6255ec7e028SFelix Kuehling */ 6265ec7e028SFelix Kuehling #define MAKE_HANDLE(gpu_id, idr_handle) \ 6275ec7e028SFelix Kuehling (((uint64_t)(gpu_id) << 32) + idr_handle) 6285ec7e028SFelix Kuehling #define GET_GPU_ID(handle) (handle >> 32) 6295ec7e028SFelix Kuehling #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 6305ec7e028SFelix Kuehling 631733fa1f7SYong Zhao enum kfd_pdd_bound { 632733fa1f7SYong Zhao PDD_UNBOUND = 0, 633733fa1f7SYong Zhao PDD_BOUND, 634733fa1f7SYong Zhao PDD_BOUND_SUSPENDED, 635733fa1f7SYong Zhao }; 636733fa1f7SYong Zhao 6374327bed2SPhilip Cox #define MAX_SYSFS_FILENAME_LEN 15 63832cb59f3SMukul Joshi 63932cb59f3SMukul Joshi /* 64032cb59f3SMukul Joshi * SDMA counter runs at 100MHz frequency. 64132cb59f3SMukul Joshi * We display SDMA activity in microsecond granularity in sysfs. 64232cb59f3SMukul Joshi * As a result, the divisor is 100. 64332cb59f3SMukul Joshi */ 64432cb59f3SMukul Joshi #define SDMA_ACTIVITY_DIVISOR 100 645d4566deeSMukul Joshi 64619f6d2a6SOded Gabbay /* Data that is per-process-per device. */ 64719f6d2a6SOded Gabbay struct kfd_process_device { 64819f6d2a6SOded Gabbay /* 64919f6d2a6SOded Gabbay * List of all per-device data for a process. 65019f6d2a6SOded Gabbay * Starts from kfd_process.per_device_data. 65119f6d2a6SOded Gabbay */ 65219f6d2a6SOded Gabbay struct list_head per_device_list; 65319f6d2a6SOded Gabbay 65419f6d2a6SOded Gabbay /* The device that owns this data. */ 65519f6d2a6SOded Gabbay struct kfd_dev *dev; 65619f6d2a6SOded Gabbay 6579fd3f1bfSFelix Kuehling /* The process that owns this kfd_process_device. */ 6589fd3f1bfSFelix Kuehling struct kfd_process *process; 65919f6d2a6SOded Gabbay 66045102048SBen Goz /* per-process-per device QCM data structure */ 66145102048SBen Goz struct qcm_process_device qpd; 66245102048SBen Goz 66319f6d2a6SOded Gabbay /*Apertures*/ 66419f6d2a6SOded Gabbay uint64_t lds_base; 66519f6d2a6SOded Gabbay uint64_t lds_limit; 66619f6d2a6SOded Gabbay uint64_t gpuvm_base; 66719f6d2a6SOded Gabbay uint64_t gpuvm_limit; 66819f6d2a6SOded Gabbay uint64_t scratch_base; 66919f6d2a6SOded Gabbay uint64_t scratch_limit; 67019f6d2a6SOded Gabbay 671403575c4SFelix Kuehling /* VM context for GPUVM allocations */ 672b84394e2SFelix Kuehling struct file *drm_file; 673403575c4SFelix Kuehling void *vm; 674403575c4SFelix Kuehling 67552b29d73SFelix Kuehling /* GPUVM allocations storage */ 67652b29d73SFelix Kuehling struct idr alloc_idr; 67752b29d73SFelix Kuehling 6789fd3f1bfSFelix Kuehling /* Flag used to tell the pdd has dequeued from the dqm. 6799fd3f1bfSFelix Kuehling * This is used to prevent dev->dqm->ops.process_termination() from 6809fd3f1bfSFelix Kuehling * being called twice when it is already called in IOMMU callback 6819fd3f1bfSFelix Kuehling * function. 682a82918f1SBen Goz */ 6839fd3f1bfSFelix Kuehling bool already_dequeued; 6849593f4d6SRajneesh Bhardwaj bool runtime_inuse; 68564d1c3a4SFelix Kuehling 68664d1c3a4SFelix Kuehling /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 68764d1c3a4SFelix Kuehling enum kfd_pdd_bound bound; 688d4566deeSMukul Joshi 689d4566deeSMukul Joshi /* VRAM usage */ 690d4566deeSMukul Joshi uint64_t vram_usage; 691d4566deeSMukul Joshi struct attribute attr_vram; 69232cb59f3SMukul Joshi char vram_filename[MAX_SYSFS_FILENAME_LEN]; 69332cb59f3SMukul Joshi 69432cb59f3SMukul Joshi /* SDMA activity tracking */ 69532cb59f3SMukul Joshi uint64_t sdma_past_activity_counter; 69632cb59f3SMukul Joshi struct attribute attr_sdma; 69732cb59f3SMukul Joshi char sdma_filename[MAX_SYSFS_FILENAME_LEN]; 6984327bed2SPhilip Cox 6994327bed2SPhilip Cox /* Eviction activity tracking */ 7004327bed2SPhilip Cox uint64_t last_evict_timestamp; 7014327bed2SPhilip Cox atomic64_t evict_duration_counter; 7024327bed2SPhilip Cox struct attribute attr_evict; 7034327bed2SPhilip Cox 7044327bed2SPhilip Cox struct kobject *kobj_stats; 70559d7115dSMukul Joshi unsigned int doorbell_index; 70619f6d2a6SOded Gabbay }; 70719f6d2a6SOded Gabbay 70852a5fdceSAlexey Skidanov #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 70952a5fdceSAlexey Skidanov 7104a488a7aSOded Gabbay /* Process data */ 7114a488a7aSOded Gabbay struct kfd_process { 71219f6d2a6SOded Gabbay /* 71319f6d2a6SOded Gabbay * kfd_process are stored in an mm_struct*->kfd_process* 71419f6d2a6SOded Gabbay * hash table (kfd_processes in kfd_process.c) 71519f6d2a6SOded Gabbay */ 71619f6d2a6SOded Gabbay struct hlist_node kfd_processes; 71719f6d2a6SOded Gabbay 7189b56bb11SFelix Kuehling /* 7199b56bb11SFelix Kuehling * Opaque pointer to mm_struct. We don't hold a reference to 7209b56bb11SFelix Kuehling * it so it should never be dereferenced from here. This is 7219b56bb11SFelix Kuehling * only used for looking up processes by their mm. 7229b56bb11SFelix Kuehling */ 7239b56bb11SFelix Kuehling void *mm; 72419f6d2a6SOded Gabbay 7255ce10687SFelix Kuehling struct kref ref; 7265ce10687SFelix Kuehling struct work_struct release_work; 7275ce10687SFelix Kuehling 72819f6d2a6SOded Gabbay struct mutex mutex; 72919f6d2a6SOded Gabbay 73019f6d2a6SOded Gabbay /* 73119f6d2a6SOded Gabbay * In any process, the thread that started main() is the lead 73219f6d2a6SOded Gabbay * thread and outlives the rest. 73319f6d2a6SOded Gabbay * It is here because amd_iommu_bind_pasid wants a task_struct. 734894a8293SFelix Kuehling * It can also be used for safely getting a reference to the 735894a8293SFelix Kuehling * mm_struct of the process. 73619f6d2a6SOded Gabbay */ 73719f6d2a6SOded Gabbay struct task_struct *lead_thread; 73819f6d2a6SOded Gabbay 73919f6d2a6SOded Gabbay /* We want to receive a notification when the mm_struct is destroyed */ 74019f6d2a6SOded Gabbay struct mmu_notifier mmu_notifier; 74119f6d2a6SOded Gabbay 7426027b1bfSYong Zhao uint16_t pasid; 74319f6d2a6SOded Gabbay 74419f6d2a6SOded Gabbay /* 74519f6d2a6SOded Gabbay * List of kfd_process_device structures, 74619f6d2a6SOded Gabbay * one for each device the process is using. 74719f6d2a6SOded Gabbay */ 74819f6d2a6SOded Gabbay struct list_head per_device_data; 74919f6d2a6SOded Gabbay 75045102048SBen Goz struct process_queue_manager pqm; 75145102048SBen Goz 75219f6d2a6SOded Gabbay /*Is the user space process 32 bit?*/ 75319f6d2a6SOded Gabbay bool is_32bit_user_mode; 754f3a39818SAndrew Lewycky 755f3a39818SAndrew Lewycky /* Event-related data */ 756f3a39818SAndrew Lewycky struct mutex event_mutex; 757482f0777SFelix Kuehling /* Event ID allocator and lookup */ 758482f0777SFelix Kuehling struct idr event_idr; 75950cb7dd9SFelix Kuehling /* Event page */ 76050cb7dd9SFelix Kuehling struct kfd_signal_page *signal_page; 761b9a5d0a5SFelix Kuehling size_t signal_mapped_size; 762f3a39818SAndrew Lewycky size_t signal_event_count; 763c986169fSFelix Kuehling bool signal_event_limit_reached; 764403575c4SFelix Kuehling 765403575c4SFelix Kuehling /* Information used for memory eviction */ 766403575c4SFelix Kuehling void *kgd_process_info; 767403575c4SFelix Kuehling /* Eviction fence that is attached to all the BOs of this process. The 768403575c4SFelix Kuehling * fence will be triggered during eviction and new one will be created 769403575c4SFelix Kuehling * during restore 770403575c4SFelix Kuehling */ 771403575c4SFelix Kuehling struct dma_fence *ef; 77226103436SFelix Kuehling 77326103436SFelix Kuehling /* Work items for evicting and restoring BOs */ 77426103436SFelix Kuehling struct delayed_work eviction_work; 77526103436SFelix Kuehling struct delayed_work restore_work; 77626103436SFelix Kuehling /* seqno of the last scheduled eviction */ 77726103436SFelix Kuehling unsigned int last_eviction_seqno; 77826103436SFelix Kuehling /* Approx. the last timestamp (in jiffies) when the process was 77926103436SFelix Kuehling * restored after an eviction 78026103436SFelix Kuehling */ 78126103436SFelix Kuehling unsigned long last_restore_timestamp; 782de9f26bbSKent Russell 783de9f26bbSKent Russell /* Kobj for our procfs */ 784de9f26bbSKent Russell struct kobject *kobj; 7856d220a7eSAmber Lin struct kobject *kobj_queues; 786de9f26bbSKent Russell struct attribute attr_pasid; 7874a488a7aSOded Gabbay }; 7884a488a7aSOded Gabbay 78964d1c3a4SFelix Kuehling #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 79064d1c3a4SFelix Kuehling extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 79164d1c3a4SFelix Kuehling extern struct srcu_struct kfd_processes_srcu; 79264d1c3a4SFelix Kuehling 79376baee6cSOded Gabbay /** 794a4497974SRajneesh Bhardwaj * typedef amdkfd_ioctl_t - typedef for ioctl function pointer. 79576baee6cSOded Gabbay * 796a4497974SRajneesh Bhardwaj * @filep: pointer to file structure. 797a4497974SRajneesh Bhardwaj * @p: amdkfd process pointer. 798a4497974SRajneesh Bhardwaj * @data: pointer to arg that was copied from user. 799a4497974SRajneesh Bhardwaj * 800a4497974SRajneesh Bhardwaj * Return: returns ioctl completion code. 80176baee6cSOded Gabbay */ 80276baee6cSOded Gabbay typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 80376baee6cSOded Gabbay void *data); 80476baee6cSOded Gabbay 80576baee6cSOded Gabbay struct amdkfd_ioctl_desc { 80676baee6cSOded Gabbay unsigned int cmd; 80776baee6cSOded Gabbay int flags; 80876baee6cSOded Gabbay amdkfd_ioctl_t *func; 80976baee6cSOded Gabbay unsigned int cmd_drv; 81076baee6cSOded Gabbay const char *name; 81176baee6cSOded Gabbay }; 81267f7cf9fSshaoyunl bool kfd_dev_is_large_bar(struct kfd_dev *dev); 81376baee6cSOded Gabbay 8141679ae8fSFelix Kuehling int kfd_process_create_wq(void); 81519f6d2a6SOded Gabbay void kfd_process_destroy_wq(void); 816373d7080SFelix Kuehling struct kfd_process *kfd_create_process(struct file *filep); 81719f6d2a6SOded Gabbay struct kfd_process *kfd_get_process(const struct task_struct *); 818f3a39818SAndrew Lewycky struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid); 81926103436SFelix Kuehling struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 820abb208a8SFelix Kuehling void kfd_unref_process(struct kfd_process *p); 8216b95e797SFelix Kuehling int kfd_process_evict_queues(struct kfd_process *p); 8226b95e797SFelix Kuehling int kfd_process_restore_queues(struct kfd_process *p); 82326103436SFelix Kuehling void kfd_suspend_all_processes(void); 82426103436SFelix Kuehling int kfd_resume_all_processes(void); 82519f6d2a6SOded Gabbay 826b84394e2SFelix Kuehling int kfd_process_device_init_vm(struct kfd_process_device *pdd, 827b84394e2SFelix Kuehling struct file *drm_file); 82864c7f8cfSBen Goz struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 82964c7f8cfSBen Goz struct kfd_process *p); 83019f6d2a6SOded Gabbay struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 831093c7d8cSAlexey Skidanov struct kfd_process *p); 832093c7d8cSAlexey Skidanov struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 833093c7d8cSAlexey Skidanov struct kfd_process *p); 83419f6d2a6SOded Gabbay 835df03ef93SHarish Kasiviswanathan int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 836373d7080SFelix Kuehling struct vm_area_struct *vma); 837373d7080SFelix Kuehling 83852b29d73SFelix Kuehling /* KFD process API for creating and translating handles */ 83952b29d73SFelix Kuehling int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 84052b29d73SFelix Kuehling void *mem); 84152b29d73SFelix Kuehling void *kfd_process_device_translate_handle(struct kfd_process_device *p, 84252b29d73SFelix Kuehling int handle); 84352b29d73SFelix Kuehling void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 84452b29d73SFelix Kuehling int handle); 84552b29d73SFelix Kuehling 846775921edSAlexey Skidanov /* Process device data iterator */ 8478eabaf54SKent Russell struct kfd_process_device *kfd_get_first_process_device_data( 8488eabaf54SKent Russell struct kfd_process *p); 8498eabaf54SKent Russell struct kfd_process_device *kfd_get_next_process_device_data( 8508eabaf54SKent Russell struct kfd_process *p, 851775921edSAlexey Skidanov struct kfd_process_device *pdd); 852775921edSAlexey Skidanov bool kfd_has_process_device_data(struct kfd_process *p); 853775921edSAlexey Skidanov 85419f6d2a6SOded Gabbay /* PASIDs */ 85519f6d2a6SOded Gabbay int kfd_pasid_init(void); 85619f6d2a6SOded Gabbay void kfd_pasid_exit(void); 85719f6d2a6SOded Gabbay bool kfd_set_pasid_limit(unsigned int new_limit); 85819f6d2a6SOded Gabbay unsigned int kfd_get_pasid_limit(void); 85919f6d2a6SOded Gabbay unsigned int kfd_pasid_alloc(void); 86019f6d2a6SOded Gabbay void kfd_pasid_free(unsigned int pasid); 86119f6d2a6SOded Gabbay 86219f6d2a6SOded Gabbay /* Doorbells */ 863ef568db7SFelix Kuehling size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 864735df2baSFelix Kuehling int kfd_doorbell_init(struct kfd_dev *kfd); 865735df2baSFelix Kuehling void kfd_doorbell_fini(struct kfd_dev *kfd); 866df03ef93SHarish Kasiviswanathan int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 867df03ef93SHarish Kasiviswanathan struct vm_area_struct *vma); 868ada2b29cSFelix Kuehling void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 86919f6d2a6SOded Gabbay unsigned int *doorbell_off); 87019f6d2a6SOded Gabbay void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 87119f6d2a6SOded Gabbay u32 read_kernel_doorbell(u32 __iomem *db); 872ada2b29cSFelix Kuehling void write_kernel_doorbell(void __iomem *db, u32 value); 8739d7d0248SFelix Kuehling void write_kernel_doorbell64(void __iomem *db, u64 value); 874339903faSYong Zhao unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, 87559d7115dSMukul Joshi struct kfd_process_device *pdd, 876ef568db7SFelix Kuehling unsigned int doorbell_id); 87759d7115dSMukul Joshi phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd); 87859d7115dSMukul Joshi int kfd_alloc_process_doorbells(struct kfd_dev *kfd, 87959d7115dSMukul Joshi unsigned int *doorbell_index); 88059d7115dSMukul Joshi void kfd_free_process_doorbells(struct kfd_dev *kfd, 88159d7115dSMukul Joshi unsigned int doorbell_index); 8826e81090bSOded Gabbay /* GTT Sub-Allocator */ 8836e81090bSOded Gabbay 8846e81090bSOded Gabbay int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 8856e81090bSOded Gabbay struct kfd_mem_obj **mem_obj); 8866e81090bSOded Gabbay 8876e81090bSOded Gabbay int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 8886e81090bSOded Gabbay 8894a488a7aSOded Gabbay extern struct device *kfd_device; 8904a488a7aSOded Gabbay 891de9f26bbSKent Russell /* KFD's procfs */ 892de9f26bbSKent Russell void kfd_procfs_init(void); 893de9f26bbSKent Russell void kfd_procfs_shutdown(void); 8946d220a7eSAmber Lin int kfd_procfs_add_queue(struct queue *q); 8956d220a7eSAmber Lin void kfd_procfs_del_queue(struct queue *q); 896de9f26bbSKent Russell 8975b5c4e40SEvgeny Pinchuk /* Topology */ 8985b5c4e40SEvgeny Pinchuk int kfd_topology_init(void); 8995b5c4e40SEvgeny Pinchuk void kfd_topology_shutdown(void); 9005b5c4e40SEvgeny Pinchuk int kfd_topology_add_device(struct kfd_dev *gpu); 9015b5c4e40SEvgeny Pinchuk int kfd_topology_remove_device(struct kfd_dev *gpu); 9023a87177eSHarish Kasiviswanathan struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 9033a87177eSHarish Kasiviswanathan uint32_t proximity_domain); 90444d8cc6fSYong Zhao struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); 9055b5c4e40SEvgeny Pinchuk struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 9065b5c4e40SEvgeny Pinchuk struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 9071dde0ea9SFelix Kuehling struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd); 9086d82eb0eSHarish Kasiviswanathan int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 909520b8fb7SFelix Kuehling int kfd_numa_node_to_apic_id(int numa_node_id); 9106127896fSHuang Rui void kfd_double_confirm_iommu_support(struct kfd_dev *gpu); 9115b5c4e40SEvgeny Pinchuk 9124a488a7aSOded Gabbay /* Interrupts */ 9132249d558SAndrew Lewycky int kfd_interrupt_init(struct kfd_dev *dev); 9142249d558SAndrew Lewycky void kfd_interrupt_exit(struct kfd_dev *dev); 9152249d558SAndrew Lewycky bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 91658e69886SLan Xiao bool interrupt_is_wanted(struct kfd_dev *dev, 91758e69886SLan Xiao const uint32_t *ih_ring_entry, 91858e69886SLan Xiao uint32_t *patched_ihre, bool *flag); 9194a488a7aSOded Gabbay 92019f6d2a6SOded Gabbay /* amdkfd Apertures */ 92119f6d2a6SOded Gabbay int kfd_init_apertures(struct kfd_process *process); 92219f6d2a6SOded Gabbay 923ed6e6a34SBen Goz /* Queue Context Management */ 924e88a614cSEdward O'Callaghan int init_queue(struct queue **q, const struct queue_properties *properties); 925ed6e6a34SBen Goz void uninit_queue(struct queue *q); 92645102048SBen Goz void print_queue_properties(struct queue_properties *q); 927ed6e6a34SBen Goz void print_queue(struct queue *q); 928ed6e6a34SBen Goz 9294b8f589bSBen Goz struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 9304b8f589bSBen Goz struct kfd_dev *dev); 931ee04955aSFelix Kuehling struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 932ee04955aSFelix Kuehling struct kfd_dev *dev); 9334b8f589bSBen Goz struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 9344b8f589bSBen Goz struct kfd_dev *dev); 935ee04955aSFelix Kuehling struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 936ee04955aSFelix Kuehling struct kfd_dev *dev); 937b91d43ddSFelix Kuehling struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 938b91d43ddSFelix Kuehling struct kfd_dev *dev); 93914328aa5SPhilip Cox struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, 94014328aa5SPhilip Cox struct kfd_dev *dev); 94164c7f8cfSBen Goz struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 94264c7f8cfSBen Goz void device_queue_manager_uninit(struct device_queue_manager *dqm); 943241f24f8SBen Goz struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 944241f24f8SBen Goz enum kfd_queue_type type); 945c2a77fdeSFelix Kuehling void kernel_queue_uninit(struct kernel_queue *kq, bool hanging); 9462640c3faSshaoyunl int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid); 947241f24f8SBen Goz 94845102048SBen Goz /* Process Queue Manager */ 94945102048SBen Goz struct process_queue_node { 95045102048SBen Goz struct queue *q; 95145102048SBen Goz struct kernel_queue *kq; 95245102048SBen Goz struct list_head process_queue_list; 95345102048SBen Goz }; 95445102048SBen Goz 9559fd3f1bfSFelix Kuehling void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 9569fd3f1bfSFelix Kuehling void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 95745102048SBen Goz int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 95845102048SBen Goz void pqm_uninit(struct process_queue_manager *pqm); 95945102048SBen Goz int pqm_create_queue(struct process_queue_manager *pqm, 96045102048SBen Goz struct kfd_dev *dev, 96145102048SBen Goz struct file *f, 96245102048SBen Goz struct queue_properties *properties, 963e47a8b52SYong Zhao unsigned int *qid, 964e47a8b52SYong Zhao uint32_t *p_doorbell_offset_in_process); 96545102048SBen Goz int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 96645102048SBen Goz int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid, 96745102048SBen Goz struct queue_properties *p); 96839e7f331SFelix Kuehling int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid, 96939e7f331SFelix Kuehling struct queue_properties *p); 970eb82da1dSOak Zeng int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, 971eb82da1dSOak Zeng void *gws); 972fbeb661bSYair Shachar struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 973fbeb661bSYair Shachar unsigned int qid); 9745bb4b78bSOak Zeng struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, 9755bb4b78bSOak Zeng unsigned int qid); 9765df099e8SJay Cornwall int pqm_get_wave_state(struct process_queue_manager *pqm, 9775df099e8SJay Cornwall unsigned int qid, 9785df099e8SJay Cornwall void __user *ctl_stack, 9795df099e8SJay Cornwall u32 *ctl_stack_used_size, 9805df099e8SJay Cornwall u32 *save_area_used_size); 98145102048SBen Goz 982788bf83dSYair Shachar int amdkfd_fence_wait_timeout(unsigned int *fence_addr, 983788bf83dSYair Shachar unsigned int fence_value, 9848c72c3d7SYong Zhao unsigned int timeout_ms); 985788bf83dSYair Shachar 986ed6e6a34SBen Goz /* Packet Manager */ 987ed6e6a34SBen Goz 98864c7f8cfSBen Goz #define KFD_FENCE_COMPLETED (100) 98964c7f8cfSBen Goz #define KFD_FENCE_INIT (10) 990241f24f8SBen Goz 991ed6e6a34SBen Goz struct packet_manager { 992ed6e6a34SBen Goz struct device_queue_manager *dqm; 993ed6e6a34SBen Goz struct kernel_queue *priv_queue; 994ed6e6a34SBen Goz struct mutex lock; 995ed6e6a34SBen Goz bool allocated; 996ed6e6a34SBen Goz struct kfd_mem_obj *ib_buffer_obj; 997851a645eSFelix Kuehling unsigned int ib_size_bytes; 998819ec5acSFelix Kuehling bool is_over_subscription; 999f6e27ff1SFelix Kuehling 1000f6e27ff1SFelix Kuehling const struct packet_manager_funcs *pmf; 1001ed6e6a34SBen Goz }; 1002ed6e6a34SBen Goz 1003f6e27ff1SFelix Kuehling struct packet_manager_funcs { 1004f6e27ff1SFelix Kuehling /* Support ASIC-specific packet formats for PM4 packets */ 1005f6e27ff1SFelix Kuehling int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 1006f6e27ff1SFelix Kuehling struct qcm_process_device *qpd); 1007f6e27ff1SFelix Kuehling int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 1008f6e27ff1SFelix Kuehling uint64_t ib, size_t ib_size_in_dwords, bool chain); 1009f6e27ff1SFelix Kuehling int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 1010f6e27ff1SFelix Kuehling struct scheduling_resources *res); 1011f6e27ff1SFelix Kuehling int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 1012f6e27ff1SFelix Kuehling struct queue *q, bool is_static); 1013f6e27ff1SFelix Kuehling int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 1014f6e27ff1SFelix Kuehling enum kfd_queue_type type, 1015f6e27ff1SFelix Kuehling enum kfd_unmap_queues_filter mode, 1016f6e27ff1SFelix Kuehling uint32_t filter_param, bool reset, 1017f6e27ff1SFelix Kuehling unsigned int sdma_engine); 1018f6e27ff1SFelix Kuehling int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 1019f6e27ff1SFelix Kuehling uint64_t fence_address, uint32_t fence_value); 1020f6e27ff1SFelix Kuehling int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 1021f6e27ff1SFelix Kuehling 1022f6e27ff1SFelix Kuehling /* Packet sizes */ 1023f6e27ff1SFelix Kuehling int map_process_size; 1024f6e27ff1SFelix Kuehling int runlist_size; 1025f6e27ff1SFelix Kuehling int set_resources_size; 1026f6e27ff1SFelix Kuehling int map_queues_size; 1027f6e27ff1SFelix Kuehling int unmap_queues_size; 1028f6e27ff1SFelix Kuehling int query_status_size; 1029f6e27ff1SFelix Kuehling int release_mem_size; 1030f6e27ff1SFelix Kuehling }; 1031f6e27ff1SFelix Kuehling 1032f6e27ff1SFelix Kuehling extern const struct packet_manager_funcs kfd_vi_pm_funcs; 1033454150b1SFelix Kuehling extern const struct packet_manager_funcs kfd_v9_pm_funcs; 1034f6e27ff1SFelix Kuehling 103564c7f8cfSBen Goz int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 1036c2a77fdeSFelix Kuehling void pm_uninit(struct packet_manager *pm, bool hanging); 103764c7f8cfSBen Goz int pm_send_set_resources(struct packet_manager *pm, 103864c7f8cfSBen Goz struct scheduling_resources *res); 103964c7f8cfSBen Goz int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 104064c7f8cfSBen Goz int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 104164c7f8cfSBen Goz uint32_t fence_value); 104264c7f8cfSBen Goz 104364c7f8cfSBen Goz int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 10447da2bcf8SYong Zhao enum kfd_unmap_queues_filter mode, 104564c7f8cfSBen Goz uint32_t filter_param, bool reset, 104664c7f8cfSBen Goz unsigned int sdma_engine); 104764c7f8cfSBen Goz 1048241f24f8SBen Goz void pm_release_ib(struct packet_manager *pm); 1049241f24f8SBen Goz 1050454150b1SFelix Kuehling /* Following PM funcs can be shared among VI and AI */ 1051454150b1SFelix Kuehling unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 105214328aa5SPhilip Cox 105319f6d2a6SOded Gabbay uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 105419f6d2a6SOded Gabbay 1055f3a39818SAndrew Lewycky /* Events */ 1056f3a39818SAndrew Lewycky extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 1057ca750681SFelix Kuehling extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 1058ca750681SFelix Kuehling 1059930c5ff4SAlexey Skidanov extern const struct kfd_device_global_init_class device_global_init_class_cik; 1060f3a39818SAndrew Lewycky 1061f3a39818SAndrew Lewycky void kfd_event_init_process(struct kfd_process *p); 1062f3a39818SAndrew Lewycky void kfd_event_free_process(struct kfd_process *p); 1063f3a39818SAndrew Lewycky int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 1064f3a39818SAndrew Lewycky int kfd_wait_on_events(struct kfd_process *p, 106559d3e8beSAlexey Skidanov uint32_t num_events, void __user *data, 1066f3a39818SAndrew Lewycky bool all, uint32_t user_timeout_ms, 1067fdf0c833SFelix Kuehling uint32_t *wait_result); 1068f3a39818SAndrew Lewycky void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id, 1069f3a39818SAndrew Lewycky uint32_t valid_id_bits); 107059d3e8beSAlexey Skidanov void kfd_signal_iommu_event(struct kfd_dev *dev, 107159d3e8beSAlexey Skidanov unsigned int pasid, unsigned long address, 107259d3e8beSAlexey Skidanov bool is_write_requested, bool is_execute_requested); 1073930c5ff4SAlexey Skidanov void kfd_signal_hw_exception_event(unsigned int pasid); 1074f3a39818SAndrew Lewycky int kfd_set_event(struct kfd_process *p, uint32_t event_id); 1075f3a39818SAndrew Lewycky int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 10760fc8011fSFelix Kuehling int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 10770fc8011fSFelix Kuehling uint64_t size); 1078f3a39818SAndrew Lewycky int kfd_event_create(struct file *devkfd, struct kfd_process *p, 1079f3a39818SAndrew Lewycky uint32_t event_type, bool auto_reset, uint32_t node_id, 1080f3a39818SAndrew Lewycky uint32_t *event_id, uint32_t *event_trigger_data, 1081f3a39818SAndrew Lewycky uint64_t *event_page_offset, uint32_t *event_slot_index); 1082f3a39818SAndrew Lewycky int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 1083f3a39818SAndrew Lewycky 10842640c3faSshaoyunl void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid, 10852640c3faSshaoyunl struct kfd_vm_fault_info *info); 10862640c3faSshaoyunl 1087e42051d2SShaoyun Liu void kfd_signal_reset_event(struct kfd_dev *dev); 1088e42051d2SShaoyun Liu 1089403575c4SFelix Kuehling void kfd_flush_tlb(struct kfd_process_device *pdd); 1090403575c4SFelix Kuehling 1091c3447e81SBen Goz int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 1092c3447e81SBen Goz 1093e42051d2SShaoyun Liu bool kfd_is_locked(void); 1094e42051d2SShaoyun Liu 1095f756e631SHarish Kasiviswanathan /* Compute profile */ 1096f756e631SHarish Kasiviswanathan void kfd_inc_compute_active(struct kfd_dev *dev); 1097f756e631SHarish Kasiviswanathan void kfd_dec_compute_active(struct kfd_dev *dev); 1098f756e631SHarish Kasiviswanathan 10996b855f7bSHarish Kasiviswanathan /* Cgroup Support */ 11006b855f7bSHarish Kasiviswanathan /* Check with device cgroup if @kfd device is accessible */ 11016b855f7bSHarish Kasiviswanathan static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) 11026b855f7bSHarish Kasiviswanathan { 1103eec8fd02SOdin Ugedal #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF) 11046b855f7bSHarish Kasiviswanathan struct drm_device *ddev = kfd->ddev; 11056b855f7bSHarish Kasiviswanathan 110699c7b309SLorenz Brun return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR, 11076b855f7bSHarish Kasiviswanathan ddev->render->index, 11086b855f7bSHarish Kasiviswanathan DEVCG_ACC_WRITE | DEVCG_ACC_READ); 11096b855f7bSHarish Kasiviswanathan #else 11106b855f7bSHarish Kasiviswanathan return 0; 11116b855f7bSHarish Kasiviswanathan #endif 11126b855f7bSHarish Kasiviswanathan } 11136b855f7bSHarish Kasiviswanathan 1114851a645eSFelix Kuehling /* Debugfs */ 1115851a645eSFelix Kuehling #if defined(CONFIG_DEBUG_FS) 1116851a645eSFelix Kuehling 1117851a645eSFelix Kuehling void kfd_debugfs_init(void); 1118851a645eSFelix Kuehling void kfd_debugfs_fini(void); 1119851a645eSFelix Kuehling int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 1120851a645eSFelix Kuehling int pqm_debugfs_mqds(struct seq_file *m, void *data); 1121851a645eSFelix Kuehling int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 1122851a645eSFelix Kuehling int dqm_debugfs_hqds(struct seq_file *m, void *data); 1123851a645eSFelix Kuehling int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 1124851a645eSFelix Kuehling int pm_debugfs_runlist(struct seq_file *m, void *data); 1125851a645eSFelix Kuehling 1126a29ec470SShaoyun Liu int kfd_debugfs_hang_hws(struct kfd_dev *dev); 1127a29ec470SShaoyun Liu int pm_debugfs_hang_hws(struct packet_manager *pm); 1128a29ec470SShaoyun Liu int dqm_debugfs_execute_queues(struct device_queue_manager *dqm); 1129a29ec470SShaoyun Liu 1130851a645eSFelix Kuehling #else 1131851a645eSFelix Kuehling 1132851a645eSFelix Kuehling static inline void kfd_debugfs_init(void) {} 1133851a645eSFelix Kuehling static inline void kfd_debugfs_fini(void) {} 1134851a645eSFelix Kuehling 1135851a645eSFelix Kuehling #endif 1136851a645eSFelix Kuehling 11374a488a7aSOded Gabbay #endif 1138