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" 486ae27841SAlex Sierra #include "amdgpu.h" 49e596b903SYong Zhao 50af47b390SLaura Abbott #define KFD_MAX_RING_ENTRY_SIZE 8 51af47b390SLaura Abbott 525b5c4e40SEvgeny Pinchuk #define KFD_SYSFS_FILE_MODE 0444 535b5c4e40SEvgeny Pinchuk 54df03ef93SHarish Kasiviswanathan /* GPU ID hash width in bits */ 55df03ef93SHarish Kasiviswanathan #define KFD_GPU_ID_HASH_WIDTH 16 56df03ef93SHarish Kasiviswanathan 57df03ef93SHarish Kasiviswanathan /* Use upper bits of mmap offset to store KFD driver specific information. 58df03ef93SHarish Kasiviswanathan * BITS[63:62] - Encode MMAP type 59df03ef93SHarish Kasiviswanathan * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to 60df03ef93SHarish Kasiviswanathan * BITS[45:0] - MMAP offset value 61df03ef93SHarish Kasiviswanathan * 62df03ef93SHarish Kasiviswanathan * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these 63df03ef93SHarish Kasiviswanathan * defines are w.r.t to PAGE_SIZE 64df03ef93SHarish Kasiviswanathan */ 6529453755SYong Zhao #define KFD_MMAP_TYPE_SHIFT 62 66df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT) 67df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT) 68df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT) 69df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT) 70d33ea570SOak Zeng #define KFD_MMAP_TYPE_MMIO (0x0ULL << KFD_MMAP_TYPE_SHIFT) 71df03ef93SHarish Kasiviswanathan 7229453755SYong Zhao #define KFD_MMAP_GPU_ID_SHIFT 46 73df03ef93SHarish Kasiviswanathan #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \ 74df03ef93SHarish Kasiviswanathan << KFD_MMAP_GPU_ID_SHIFT) 75df03ef93SHarish Kasiviswanathan #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\ 76df03ef93SHarish Kasiviswanathan & KFD_MMAP_GPU_ID_MASK) 7729453755SYong Zhao #define KFD_MMAP_GET_GPU_ID(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \ 78df03ef93SHarish Kasiviswanathan >> KFD_MMAP_GPU_ID_SHIFT) 79df03ef93SHarish Kasiviswanathan 80ed6e6a34SBen Goz /* 81ed6e6a34SBen Goz * When working with cp scheduler we should assign the HIQ manually or via 82e7016d8eSYong Zhao * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot 83ed6e6a34SBen Goz * definitions for Kaveri. In Kaveri only the first ME queues participates 84ed6e6a34SBen Goz * in the cp scheduling taking that in mind we set the HIQ slot in the 85ed6e6a34SBen Goz * second ME. 86ed6e6a34SBen Goz */ 87ed6e6a34SBen Goz #define KFD_CIK_HIQ_PIPE 4 88ed6e6a34SBen Goz #define KFD_CIK_HIQ_QUEUE 0 89ed6e6a34SBen Goz 905b5c4e40SEvgeny Pinchuk /* Macro for allocating structures */ 915b5c4e40SEvgeny Pinchuk #define kfd_alloc_struct(ptr_to_struct) \ 925b5c4e40SEvgeny Pinchuk ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 935b5c4e40SEvgeny Pinchuk 9419f6d2a6SOded Gabbay #define KFD_MAX_NUM_OF_PROCESSES 512 95b8cbab04SOded Gabbay #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 9619f6d2a6SOded Gabbay 9719f6d2a6SOded Gabbay /* 98373d7080SFelix Kuehling * Size of the per-process TBA+TMA buffer: 2 pages 99373d7080SFelix Kuehling * 100373d7080SFelix Kuehling * The first page is the TBA used for the CWSR ISA code. The second 101a4497974SRajneesh Bhardwaj * page is used as TMA for user-mode trap handler setup in daisy-chain mode. 102373d7080SFelix Kuehling */ 103373d7080SFelix Kuehling #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) 104373d7080SFelix Kuehling #define KFD_CWSR_TMA_OFFSET PAGE_SIZE 105373d7080SFelix Kuehling 10674523943SYong Zhao #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 10774523943SYong Zhao (KFD_MAX_NUM_OF_PROCESSES * \ 10874523943SYong Zhao KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 10974523943SYong Zhao 11074523943SYong Zhao #define KFD_KERNEL_QUEUE_SIZE 2048 11174523943SYong Zhao 11214328aa5SPhilip Cox #define KFD_UNMAP_LATENCY_MS (4000) 11314328aa5SPhilip Cox 114373d7080SFelix Kuehling /* 1151f86805aSYong Zhao * 512 = 0x200 1161f86805aSYong Zhao * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the 1171f86805aSYong Zhao * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA. 1181f86805aSYong Zhao * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC 1191f86805aSYong Zhao * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in 1201f86805aSYong Zhao * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE. 1211f86805aSYong Zhao */ 1221f86805aSYong Zhao #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512 1231f86805aSYong Zhao 12436988070SRajneesh Bhardwaj /** 12536988070SRajneesh Bhardwaj * enum kfd_ioctl_flags - KFD ioctl flags 12636988070SRajneesh Bhardwaj * Various flags that can be set in &amdkfd_ioctl_desc.flags to control how 12736988070SRajneesh Bhardwaj * userspace can use a given ioctl. 12836988070SRajneesh Bhardwaj */ 12936988070SRajneesh Bhardwaj enum kfd_ioctl_flags { 13036988070SRajneesh Bhardwaj /* 13136988070SRajneesh Bhardwaj * @KFD_IOC_FLAG_CHECKPOINT_RESTORE: 13236988070SRajneesh Bhardwaj * Certain KFD ioctls such as AMDKFD_IOC_CRIU_OP can potentially 13336988070SRajneesh Bhardwaj * perform privileged operations and load arbitrary data into MQDs and 13436988070SRajneesh Bhardwaj * eventually HQD registers when the queue is mapped by HWS. In order to 13536988070SRajneesh Bhardwaj * prevent this we should perform additional security checks. 13636988070SRajneesh Bhardwaj * 13736988070SRajneesh Bhardwaj * This is equivalent to callers with the CHECKPOINT_RESTORE capability. 13836988070SRajneesh Bhardwaj * 13936988070SRajneesh Bhardwaj * Note: Since earlier versions of docker do not support CHECKPOINT_RESTORE, 14036988070SRajneesh Bhardwaj * we also allow ioctls with SYS_ADMIN capability. 14136988070SRajneesh Bhardwaj */ 14236988070SRajneesh Bhardwaj KFD_IOC_FLAG_CHECKPOINT_RESTORE = BIT(0), 14336988070SRajneesh Bhardwaj }; 1441f86805aSYong Zhao /* 145b8cbab04SOded Gabbay * Kernel module parameter to specify maximum number of supported queues per 146b8cbab04SOded Gabbay * device 14719f6d2a6SOded Gabbay */ 148b8cbab04SOded Gabbay extern int max_num_of_queues_per_device; 14919f6d2a6SOded Gabbay 150ed6e6a34SBen Goz 15131c21fecSBen Goz /* Kernel module parameter to specify the scheduling policy */ 15231c21fecSBen Goz extern int sched_policy; 15331c21fecSBen Goz 154a99c6d4fSFelix Kuehling /* 155a99c6d4fSFelix Kuehling * Kernel module parameter to specify the maximum process 156a99c6d4fSFelix Kuehling * number per HW scheduler 157a99c6d4fSFelix Kuehling */ 158a99c6d4fSFelix Kuehling extern int hws_max_conc_proc; 159a99c6d4fSFelix Kuehling 160373d7080SFelix Kuehling extern int cwsr_enable; 161373d7080SFelix Kuehling 16281663016SOded Gabbay /* 16381663016SOded Gabbay * Kernel module parameter to specify whether to send sigterm to HSA process on 16481663016SOded Gabbay * unhandled exception 16581663016SOded Gabbay */ 16681663016SOded Gabbay extern int send_sigterm; 16781663016SOded Gabbay 168ebcfd1e2SFelix Kuehling /* 169374200b1SFelix Kuehling * This kernel module is used to simulate large bar machine on non-large bar 170374200b1SFelix Kuehling * enabled machines. 171374200b1SFelix Kuehling */ 172374200b1SFelix Kuehling extern int debug_largebar; 173374200b1SFelix Kuehling 174374200b1SFelix Kuehling /* 175ebcfd1e2SFelix Kuehling * Ignore CRAT table during KFD initialization, can be used to work around 176ebcfd1e2SFelix Kuehling * broken CRAT tables on some AMD systems 177ebcfd1e2SFelix Kuehling */ 178ebcfd1e2SFelix Kuehling extern int ignore_crat; 179ebcfd1e2SFelix Kuehling 180a4497974SRajneesh Bhardwaj /* Set sh_mem_config.retry_disable on GFX v9 */ 18175ee6487SFelix Kuehling extern int amdgpu_noretry; 182bed4f110SFelix Kuehling 183a4497974SRajneesh Bhardwaj /* Halt if HWS hang is detected */ 1840e9a860cSYong Zhao extern int halt_if_hws_hang; 1850e9a860cSYong Zhao 186a4497974SRajneesh Bhardwaj /* Whether MEC FW support GWS barriers */ 18729e76462SOak Zeng extern bool hws_gws_support; 18829e76462SOak Zeng 189a4497974SRajneesh Bhardwaj /* Queue preemption timeout in ms */ 19014328aa5SPhilip Cox extern int queue_preemption_timeout_ms; 19114328aa5SPhilip Cox 1926d909c5dSOak Zeng /* 1936d909c5dSOak Zeng * Don't evict process queues on vm fault 1946d909c5dSOak Zeng */ 1956d909c5dSOak Zeng extern int amdgpu_no_queue_eviction_on_vm_fault; 1966d909c5dSOak Zeng 197a4497974SRajneesh Bhardwaj /* Enable eviction debug messages */ 198b2057956SFelix Kuehling extern bool debug_evictions; 199b2057956SFelix Kuehling 200ed6e6a34SBen Goz enum cache_policy { 201ed6e6a34SBen Goz cache_policy_coherent, 202ed6e6a34SBen Goz cache_policy_noncoherent 203ed6e6a34SBen Goz }; 204ed6e6a34SBen Goz 205dd0ae064SGraham Sider #define KFD_GC_VERSION(dev) ((dev)->adev->ip_versions[GC_HWIP][0]) 206dd0ae064SGraham Sider #define KFD_IS_SOC15(dev) ((KFD_GC_VERSION(dev)) >= (IP_VERSION(9, 0, 1))) 207ef568db7SFelix Kuehling 208f3a39818SAndrew Lewycky struct kfd_event_interrupt_class { 209f3a39818SAndrew Lewycky bool (*interrupt_isr)(struct kfd_dev *dev, 21058e69886SLan Xiao const uint32_t *ih_ring_entry, uint32_t *patched_ihre, 21158e69886SLan Xiao bool *patched_flag); 212f3a39818SAndrew Lewycky void (*interrupt_wq)(struct kfd_dev *dev, 213f3a39818SAndrew Lewycky const uint32_t *ih_ring_entry); 214f3a39818SAndrew Lewycky }; 215f3a39818SAndrew Lewycky 2164a488a7aSOded Gabbay struct kfd_device_info { 2179d6fa9c7SGraham Sider uint32_t gfx_target_version; 218f3a39818SAndrew Lewycky const struct kfd_event_interrupt_class *event_interrupt_class; 2194a488a7aSOded Gabbay unsigned int max_pasid_bits; 220992839adSYair Shachar unsigned int max_no_of_hqd; 221ada2b29cSFelix Kuehling unsigned int doorbell_size; 2224a488a7aSOded Gabbay size_t ih_ring_entry_size; 223f7c826adSAlexey Skidanov uint8_t num_of_watch_points; 22419f6d2a6SOded Gabbay uint16_t mqd_size_aligned; 225373d7080SFelix Kuehling bool supports_cwsr; 22664d1c3a4SFelix Kuehling bool needs_iommu_device; 2273ee2d00cSFelix Kuehling bool needs_pci_atomics; 228fb932dfeSFelix Kuehling uint32_t no_atomic_fw_version; 229d5094189SShaoyun Liu unsigned int num_sdma_queues_per_engine; 2304a488a7aSOded Gabbay }; 2314a488a7aSOded Gabbay 232ee2f17f4SAmber Lin unsigned int kfd_get_num_sdma_engines(struct kfd_dev *kdev); 233ee2f17f4SAmber Lin unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_dev *kdev); 234ee2f17f4SAmber Lin 23536b5c08fSOded Gabbay struct kfd_mem_obj { 23636b5c08fSOded Gabbay uint32_t range_start; 23736b5c08fSOded Gabbay uint32_t range_end; 23836b5c08fSOded Gabbay uint64_t gpu_addr; 23936b5c08fSOded Gabbay uint32_t *cpu_ptr; 240b91d43ddSFelix Kuehling void *gtt_mem; 24136b5c08fSOded Gabbay }; 24236b5c08fSOded Gabbay 24344008d7aSYong Zhao struct kfd_vmid_info { 24444008d7aSYong Zhao uint32_t first_vmid_kfd; 24544008d7aSYong Zhao uint32_t last_vmid_kfd; 24644008d7aSYong Zhao uint32_t vmid_num_kfd; 24744008d7aSYong Zhao }; 24844008d7aSYong Zhao 2494a488a7aSOded Gabbay struct kfd_dev { 250c6c57446SGraham Sider struct amdgpu_device *adev; 2514a488a7aSOded Gabbay 252f0dc99a6SGraham Sider struct kfd_device_info device_info; 2534a488a7aSOded Gabbay struct pci_dev *pdev; 2543a0c3423SHarish Kasiviswanathan struct drm_device *ddev; 2554a488a7aSOded Gabbay 2564a488a7aSOded Gabbay unsigned int id; /* topology stub index */ 2574a488a7aSOded Gabbay 25819f6d2a6SOded Gabbay phys_addr_t doorbell_base; /* Start of actual doorbells used by 25919f6d2a6SOded Gabbay * KFD. It is aligned for mapping 26019f6d2a6SOded Gabbay * into user mode 26119f6d2a6SOded Gabbay */ 262339903faSYong Zhao size_t doorbell_base_dw_offset; /* Offset from the start of the PCI 263339903faSYong Zhao * doorbell BAR to the first KFD 264339903faSYong Zhao * doorbell in dwords. GFX reserves 265339903faSYong Zhao * the segment before this offset. 26619f6d2a6SOded Gabbay */ 26719f6d2a6SOded Gabbay u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 26819f6d2a6SOded Gabbay * page used by kernel queue 26919f6d2a6SOded Gabbay */ 27019f6d2a6SOded Gabbay 2714a488a7aSOded Gabbay struct kgd2kfd_shared_resources shared_resources; 27244008d7aSYong Zhao struct kfd_vmid_info vm_info; 2734a488a7aSOded Gabbay 274cea405b1SXihan Zhang const struct kfd2kgd_calls *kfd2kgd; 275cea405b1SXihan Zhang struct mutex doorbell_mutex; 276f761d8bdSJoe Perches DECLARE_BITMAP(doorbell_available_index, 277f761d8bdSJoe Perches KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 278cea405b1SXihan Zhang 27936b5c08fSOded Gabbay void *gtt_mem; 28036b5c08fSOded Gabbay uint64_t gtt_start_gpu_addr; 28136b5c08fSOded Gabbay void *gtt_start_cpu_ptr; 28236b5c08fSOded Gabbay void *gtt_sa_bitmap; 28336b5c08fSOded Gabbay struct mutex gtt_sa_lock; 28436b5c08fSOded Gabbay unsigned int gtt_sa_chunk_size; 28536b5c08fSOded Gabbay unsigned int gtt_sa_num_of_chunks; 28636b5c08fSOded Gabbay 2872249d558SAndrew Lewycky /* Interrupts */ 28804ad47bdSAndres Rodriguez struct kfifo ih_fifo; 28948e876a2SAndres Rodriguez struct workqueue_struct *ih_wq; 2902249d558SAndrew Lewycky struct work_struct interrupt_work; 2912249d558SAndrew Lewycky spinlock_t interrupt_lock; 2922249d558SAndrew Lewycky 293ed6e6a34SBen Goz /* QCM Device instance */ 294ed6e6a34SBen Goz struct device_queue_manager *dqm; 2954a488a7aSOded Gabbay 296ed6e6a34SBen Goz bool init_complete; 2972249d558SAndrew Lewycky /* 2982249d558SAndrew Lewycky * Interrupts of interest to KFD are copied 2992249d558SAndrew Lewycky * from the HW ring into a SW ring. 3002249d558SAndrew Lewycky */ 3012249d558SAndrew Lewycky bool interrupts_active; 302fbeb661bSYair Shachar 303fbeb661bSYair Shachar /* Debug manager */ 304fbeb661bSYair Shachar struct kfd_dbgmgr *dbgmgr; 305373d7080SFelix Kuehling 3065ade6c9cSFelix Kuehling /* Firmware versions */ 3075ade6c9cSFelix Kuehling uint16_t mec_fw_version; 30829633d0eSJoseph Greathouse uint16_t mec2_fw_version; 3095ade6c9cSFelix Kuehling uint16_t sdma_fw_version; 3105ade6c9cSFelix Kuehling 311a99c6d4fSFelix Kuehling /* Maximum process number mapped to HW scheduler */ 312a99c6d4fSFelix Kuehling unsigned int max_proc_per_quantum; 313a99c6d4fSFelix Kuehling 314373d7080SFelix Kuehling /* CWSR */ 315373d7080SFelix Kuehling bool cwsr_enabled; 316373d7080SFelix Kuehling const void *cwsr_isa; 317373d7080SFelix Kuehling unsigned int cwsr_isa_size; 3180c1690e3SShaoyun Liu 3190c1690e3SShaoyun Liu /* xGMI */ 3200c1690e3SShaoyun Liu uint64_t hive_id; 3210c663695SDivya Shikre 322d35f00d8SEric Huang bool pci_atomic_requested; 3239b54d201SEric Huang 3246127896fSHuang Rui /* Use IOMMU v2 flag */ 3256127896fSHuang Rui bool use_iommu_v2; 3266127896fSHuang Rui 3279b54d201SEric Huang /* SRAM ECC flag */ 3289b54d201SEric Huang atomic_t sram_ecc_flag; 329f756e631SHarish Kasiviswanathan 330f756e631SHarish Kasiviswanathan /* Compute Profile ref. count */ 331f756e631SHarish Kasiviswanathan atomic_t compute_profile; 332e09d4fc8SOak Zeng 333a4497974SRajneesh Bhardwaj /* Global GWS resource shared between processes */ 334e09d4fc8SOak Zeng void *gws; 335938a0650SAmber Lin 336938a0650SAmber Lin /* Clients watching SMI events */ 337938a0650SAmber Lin struct list_head smi_clients; 338938a0650SAmber Lin spinlock_t smi_lock; 33955977744SMukul Joshi 34055977744SMukul Joshi uint32_t reset_seq_num; 34159d7115dSMukul Joshi 34259d7115dSMukul Joshi struct ida doorbell_ida; 34359d7115dSMukul Joshi unsigned int max_doorbell_slices; 3449b498efaSAlex Deucher 3459b498efaSAlex Deucher int noretry; 346814ab993SPhilip Yang 347814ab993SPhilip Yang /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */ 348814ab993SPhilip Yang struct dev_pagemap pgmap; 3494a488a7aSOded Gabbay }; 3504a488a7aSOded Gabbay 35119f6d2a6SOded Gabbay enum kfd_mempool { 35219f6d2a6SOded Gabbay KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 35319f6d2a6SOded Gabbay KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 35419f6d2a6SOded Gabbay KFD_MEMPOOL_FRAMEBUFFER = 3, 35519f6d2a6SOded Gabbay }; 35619f6d2a6SOded Gabbay 3574a488a7aSOded Gabbay /* Character device interface */ 3584a488a7aSOded Gabbay int kfd_chardev_init(void); 3594a488a7aSOded Gabbay void kfd_chardev_exit(void); 3604a488a7aSOded Gabbay struct device *kfd_chardev(void); 3614a488a7aSOded Gabbay 362241f24f8SBen Goz /** 363a4497974SRajneesh Bhardwaj * enum kfd_unmap_queues_filter - Enum for queue filters. 364241f24f8SBen Goz * 3657da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue. 366241f24f8SBen Goz * 3677da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 368241f24f8SBen Goz * running queues list. 369241f24f8SBen Goz * 3707da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 371241f24f8SBen Goz * specific process. 372241f24f8SBen Goz * 373241f24f8SBen Goz */ 3747da2bcf8SYong Zhao enum kfd_unmap_queues_filter { 3757da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE, 3767da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 3777da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 3787da2bcf8SYong Zhao KFD_UNMAP_QUEUES_FILTER_BY_PASID 379241f24f8SBen Goz }; 38019f6d2a6SOded Gabbay 381ed8aab45SBen Goz /** 382a4497974SRajneesh Bhardwaj * enum kfd_queue_type - Enum for various queue types. 383ed8aab45SBen Goz * 384ed8aab45SBen Goz * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 385ed8aab45SBen Goz * 386a4497974SRajneesh Bhardwaj * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type. 387ed8aab45SBen Goz * 388ed8aab45SBen Goz * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 389ed8aab45SBen Goz * 390ed8aab45SBen Goz * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 391a4497974SRajneesh Bhardwaj * 392a4497974SRajneesh Bhardwaj * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface. 393ed8aab45SBen Goz */ 394ed8aab45SBen Goz enum kfd_queue_type { 395ed8aab45SBen Goz KFD_QUEUE_TYPE_COMPUTE, 396ed8aab45SBen Goz KFD_QUEUE_TYPE_SDMA, 397ed8aab45SBen Goz KFD_QUEUE_TYPE_HIQ, 3981b4670f6SOak Zeng KFD_QUEUE_TYPE_DIQ, 3991b4670f6SOak Zeng KFD_QUEUE_TYPE_SDMA_XGMI 400ed8aab45SBen Goz }; 401ed8aab45SBen Goz 4026e99df57SBen Goz enum kfd_queue_format { 4036e99df57SBen Goz KFD_QUEUE_FORMAT_PM4, 4046e99df57SBen Goz KFD_QUEUE_FORMAT_AQL 4056e99df57SBen Goz }; 4066e99df57SBen Goz 4070ccbc7cdSOak Zeng enum KFD_QUEUE_PRIORITY { 4080ccbc7cdSOak Zeng KFD_QUEUE_PRIORITY_MINIMUM = 0, 4090ccbc7cdSOak Zeng KFD_QUEUE_PRIORITY_MAXIMUM = 15 4100ccbc7cdSOak Zeng }; 4110ccbc7cdSOak Zeng 412ed8aab45SBen Goz /** 413ed8aab45SBen Goz * struct queue_properties 414ed8aab45SBen Goz * 415ed8aab45SBen Goz * @type: The queue type. 416ed8aab45SBen Goz * 417ed8aab45SBen Goz * @queue_id: Queue identifier. 418ed8aab45SBen Goz * 419ed8aab45SBen Goz * @queue_address: Queue ring buffer address. 420ed8aab45SBen Goz * 421ed8aab45SBen Goz * @queue_size: Queue ring buffer size. 422ed8aab45SBen Goz * 423ed8aab45SBen Goz * @priority: Defines the queue priority relative to other queues in the 424ed8aab45SBen Goz * process. 425ed8aab45SBen Goz * This is just an indication and HW scheduling may override the priority as 426ed8aab45SBen Goz * necessary while keeping the relative prioritization. 427ed8aab45SBen Goz * the priority granularity is from 0 to f which f is the highest priority. 428ed8aab45SBen Goz * currently all queues are initialized with the highest priority. 429ed8aab45SBen Goz * 430ed8aab45SBen Goz * @queue_percent: This field is partially implemented and currently a zero in 431ed8aab45SBen Goz * this field defines that the queue is non active. 432ed8aab45SBen Goz * 433ed8aab45SBen Goz * @read_ptr: User space address which points to the number of dwords the 434ed8aab45SBen Goz * cp read from the ring buffer. This field updates automatically by the H/W. 435ed8aab45SBen Goz * 436ed8aab45SBen Goz * @write_ptr: Defines the number of dwords written to the ring buffer. 437ed8aab45SBen Goz * 438a4497974SRajneesh Bhardwaj * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring 439a4497974SRajneesh Bhardwaj * buffer. This field should be similar to write_ptr and the user should 440a4497974SRajneesh Bhardwaj * update this field after updating the write_ptr. 441ed8aab45SBen Goz * 442ed8aab45SBen Goz * @doorbell_off: The doorbell offset in the doorbell pci-bar. 443ed8aab45SBen Goz * 4448eabaf54SKent Russell * @is_interop: Defines if this is a interop queue. Interop queue means that 4458eabaf54SKent Russell * the queue can access both graphics and compute resources. 446ed8aab45SBen Goz * 44726103436SFelix Kuehling * @is_evicted: Defines if the queue is evicted. Only active queues 44826103436SFelix Kuehling * are evicted, rendering them inactive. 44926103436SFelix Kuehling * 45026103436SFelix Kuehling * @is_active: Defines if the queue is active or not. @is_active and 45126103436SFelix Kuehling * @is_evicted are protected by the DQM lock. 452ed8aab45SBen Goz * 453b8020b03SJoseph Greathouse * @is_gws: Defines if the queue has been updated to be GWS-capable or not. 454b8020b03SJoseph Greathouse * @is_gws should be protected by the DQM lock, since changing it can yield the 455b8020b03SJoseph Greathouse * possibility of updating DQM state on number of GWS queues. 456b8020b03SJoseph Greathouse * 457ed8aab45SBen Goz * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 458ed8aab45SBen Goz * of the queue. 459ed8aab45SBen Goz * 460ed8aab45SBen Goz * This structure represents the queue properties for each queue no matter if 461ed8aab45SBen Goz * it's user mode or kernel mode queue. 462ed8aab45SBen Goz * 463ed8aab45SBen Goz */ 4648668dfc3SDavid Yat Sin 465ed8aab45SBen Goz struct queue_properties { 466ed8aab45SBen Goz enum kfd_queue_type type; 4676e99df57SBen Goz enum kfd_queue_format format; 468ed8aab45SBen Goz unsigned int queue_id; 469ed8aab45SBen Goz uint64_t queue_address; 470ed8aab45SBen Goz uint64_t queue_size; 471ed8aab45SBen Goz uint32_t priority; 472ed8aab45SBen Goz uint32_t queue_percent; 473ed8aab45SBen Goz uint32_t *read_ptr; 474ed8aab45SBen Goz uint32_t *write_ptr; 475ada2b29cSFelix Kuehling void __iomem *doorbell_ptr; 476ed8aab45SBen Goz uint32_t doorbell_off; 477ed8aab45SBen Goz bool is_interop; 47826103436SFelix Kuehling bool is_evicted; 479ed8aab45SBen Goz bool is_active; 480b8020b03SJoseph Greathouse bool is_gws; 481ed8aab45SBen Goz /* Not relevant for user mode queues in cp scheduling */ 482ed8aab45SBen Goz unsigned int vmid; 48377669eb8SBen Goz /* Relevant only for sdma queues*/ 48477669eb8SBen Goz uint32_t sdma_engine_id; 48577669eb8SBen Goz uint32_t sdma_queue_id; 48677669eb8SBen Goz uint32_t sdma_vm_addr; 487ff3d04a1SBen Goz /* Relevant only for VI */ 488ff3d04a1SBen Goz uint64_t eop_ring_buffer_address; 489ff3d04a1SBen Goz uint32_t eop_ring_buffer_size; 490ff3d04a1SBen Goz uint64_t ctx_save_restore_area_address; 491ff3d04a1SBen Goz uint32_t ctx_save_restore_area_size; 492373d7080SFelix Kuehling uint32_t ctl_stack_size; 493373d7080SFelix Kuehling uint64_t tba_addr; 494373d7080SFelix Kuehling uint64_t tma_addr; 495ed8aab45SBen Goz }; 496ed8aab45SBen Goz 497bb2d2128SFelix Kuehling #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ 498bb2d2128SFelix Kuehling (q).queue_address != 0 && \ 499bb2d2128SFelix Kuehling (q).queue_percent > 0 && \ 500bb2d2128SFelix Kuehling !(q).is_evicted) 501bb2d2128SFelix Kuehling 5027c695a2cSLang Yu enum mqd_update_flag { 5037c695a2cSLang Yu UPDATE_FLAG_CU_MASK = 0, 5047c695a2cSLang Yu }; 5057c695a2cSLang Yu 5067c695a2cSLang Yu struct mqd_update_info { 5077c695a2cSLang Yu union { 5087c695a2cSLang Yu struct { 5097c695a2cSLang Yu uint32_t count; /* Must be a multiple of 32 */ 5107c695a2cSLang Yu uint32_t *ptr; 5117c695a2cSLang Yu } cu_mask; 5127c695a2cSLang Yu }; 5137c695a2cSLang Yu enum mqd_update_flag update_flag; 5147c695a2cSLang Yu }; 515c6e559ebSLang Yu 516ed8aab45SBen Goz /** 517ed8aab45SBen Goz * struct queue 518ed8aab45SBen Goz * 519ed8aab45SBen Goz * @list: Queue linked list. 520ed8aab45SBen Goz * 521a4497974SRajneesh Bhardwaj * @mqd: The queue MQD (memory queue descriptor). 522ed8aab45SBen Goz * 523ed8aab45SBen Goz * @mqd_mem_obj: The MQD local gpu memory object. 524ed8aab45SBen Goz * 525ed8aab45SBen Goz * @gart_mqd_addr: The MQD gart mc address. 526ed8aab45SBen Goz * 527ed8aab45SBen Goz * @properties: The queue properties. 528ed8aab45SBen Goz * 529ed8aab45SBen Goz * @mec: Used only in no cp scheduling mode and identifies to micro engine id 530a4497974SRajneesh Bhardwaj * that the queue should be executed on. 531ed8aab45SBen Goz * 5328eabaf54SKent Russell * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 5338eabaf54SKent Russell * id. 534ed8aab45SBen Goz * 535ed8aab45SBen Goz * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 536ed8aab45SBen Goz * 537ed8aab45SBen Goz * @process: The kfd process that created this queue. 538ed8aab45SBen Goz * 539ed8aab45SBen Goz * @device: The kfd device that created this queue. 540ed8aab45SBen Goz * 541eb82da1dSOak Zeng * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL 542eb82da1dSOak Zeng * otherwise. 543eb82da1dSOak Zeng * 544ed8aab45SBen Goz * This structure represents user mode compute queues. 545ed8aab45SBen Goz * It contains all the necessary data to handle such queues. 546ed8aab45SBen Goz * 547ed8aab45SBen Goz */ 548ed8aab45SBen Goz 549ed8aab45SBen Goz struct queue { 550ed8aab45SBen Goz struct list_head list; 551ed8aab45SBen Goz void *mqd; 552ed8aab45SBen Goz struct kfd_mem_obj *mqd_mem_obj; 553ed8aab45SBen Goz uint64_t gart_mqd_addr; 554ed8aab45SBen Goz struct queue_properties properties; 555ed8aab45SBen Goz 556ed8aab45SBen Goz uint32_t mec; 557ed8aab45SBen Goz uint32_t pipe; 558ed8aab45SBen Goz uint32_t queue; 559ed8aab45SBen Goz 56077669eb8SBen Goz unsigned int sdma_id; 561ef568db7SFelix Kuehling unsigned int doorbell_id; 56277669eb8SBen Goz 563ed8aab45SBen Goz struct kfd_process *process; 564ed8aab45SBen Goz struct kfd_dev *device; 565eb82da1dSOak Zeng void *gws; 5666d220a7eSAmber Lin 5676d220a7eSAmber Lin /* procfs */ 5686d220a7eSAmber Lin struct kobject kobj; 569ed8aab45SBen Goz }; 570ed8aab45SBen Goz 5716e99df57SBen Goz enum KFD_MQD_TYPE { 572d7c0b047SYong Zhao KFD_MQD_TYPE_HIQ = 0, /* for hiq */ 57385d258f9SBen Goz KFD_MQD_TYPE_CP, /* for cp queues and diq */ 57485d258f9SBen Goz KFD_MQD_TYPE_SDMA, /* for sdma queues */ 57559f650a0SOak Zeng KFD_MQD_TYPE_DIQ, /* for diq */ 5766e99df57SBen Goz KFD_MQD_TYPE_MAX 5776e99df57SBen Goz }; 5786e99df57SBen Goz 5790ccbc7cdSOak Zeng enum KFD_PIPE_PRIORITY { 5800ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_LOW = 0, 5810ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_MEDIUM, 5820ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_HIGH 5830ccbc7cdSOak Zeng }; 5840ccbc7cdSOak Zeng 585241f24f8SBen Goz struct scheduling_resources { 586241f24f8SBen Goz unsigned int vmid_mask; 587241f24f8SBen Goz enum kfd_queue_type type; 588241f24f8SBen Goz uint64_t queue_mask; 589241f24f8SBen Goz uint64_t gws_mask; 590241f24f8SBen Goz uint32_t oac_mask; 591241f24f8SBen Goz uint32_t gds_heap_base; 592241f24f8SBen Goz uint32_t gds_heap_size; 593241f24f8SBen Goz }; 594241f24f8SBen Goz 595241f24f8SBen Goz struct process_queue_manager { 596241f24f8SBen Goz /* data */ 597241f24f8SBen Goz struct kfd_process *process; 598241f24f8SBen Goz struct list_head queues; 599241f24f8SBen Goz unsigned long *queue_slot_bitmap; 600241f24f8SBen Goz }; 601241f24f8SBen Goz 602241f24f8SBen Goz struct qcm_process_device { 603241f24f8SBen Goz /* The Device Queue Manager that owns this data */ 604241f24f8SBen Goz struct device_queue_manager *dqm; 605241f24f8SBen Goz struct process_queue_manager *pqm; 606241f24f8SBen Goz /* Queues list */ 607241f24f8SBen Goz struct list_head queues_list; 608241f24f8SBen Goz struct list_head priv_queue_list; 609241f24f8SBen Goz 610241f24f8SBen Goz unsigned int queue_count; 611241f24f8SBen Goz unsigned int vmid; 612241f24f8SBen Goz bool is_debug; 61326103436SFelix Kuehling unsigned int evicted; /* eviction counter, 0=active */ 6149fd3f1bfSFelix Kuehling 6159fd3f1bfSFelix Kuehling /* This flag tells if we should reset all wavefronts on 6169fd3f1bfSFelix Kuehling * process termination 6179fd3f1bfSFelix Kuehling */ 6189fd3f1bfSFelix Kuehling bool reset_wavefronts; 6199fd3f1bfSFelix Kuehling 620b8020b03SJoseph Greathouse /* This flag tells us if this process has a GWS-capable 621b8020b03SJoseph Greathouse * queue that will be mapped into the runlist. It's 622b8020b03SJoseph Greathouse * possible to request a GWS BO, but not have the queue 623b8020b03SJoseph Greathouse * currently mapped, and this changes how the MAP_PROCESS 624b8020b03SJoseph Greathouse * PM4 packet is configured. 625b8020b03SJoseph Greathouse */ 626b8020b03SJoseph Greathouse bool mapped_gws_queue; 627b8020b03SJoseph Greathouse 628a4497974SRajneesh Bhardwaj /* All the memory management data should be here too */ 629241f24f8SBen Goz uint64_t gds_context_area; 630435e2f97SYong Zhao /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */ 631e715c6d0SShaoyun Liu uint64_t page_table_base; 632241f24f8SBen Goz uint32_t sh_mem_config; 633241f24f8SBen Goz uint32_t sh_mem_bases; 634241f24f8SBen Goz uint32_t sh_mem_ape1_base; 635241f24f8SBen Goz uint32_t sh_mem_ape1_limit; 636241f24f8SBen Goz uint32_t gds_size; 637241f24f8SBen Goz uint32_t num_gws; 638241f24f8SBen Goz uint32_t num_oac; 6396a1c9510SMoses Reuben uint32_t sh_hidden_private_base; 640373d7080SFelix Kuehling 641373d7080SFelix Kuehling /* CWSR memory */ 64268df0f19SLang Yu struct kgd_mem *cwsr_mem; 643373d7080SFelix Kuehling void *cwsr_kaddr; 644d01994c2SFelix Kuehling uint64_t cwsr_base; 645373d7080SFelix Kuehling uint64_t tba_addr; 646373d7080SFelix Kuehling uint64_t tma_addr; 647d01994c2SFelix Kuehling 648d01994c2SFelix Kuehling /* IB memory */ 64968df0f19SLang Yu struct kgd_mem *ib_mem; 650d01994c2SFelix Kuehling uint64_t ib_base; 651552764b6SFelix Kuehling void *ib_kaddr; 652ef568db7SFelix Kuehling 653ef568db7SFelix Kuehling /* doorbell resources per process per device */ 654ef568db7SFelix Kuehling unsigned long *doorbell_bitmap; 655241f24f8SBen Goz }; 656241f24f8SBen Goz 65726103436SFelix Kuehling /* KFD Memory Eviction */ 65826103436SFelix Kuehling 65926103436SFelix Kuehling /* Approx. wait time before attempting to restore evicted BOs */ 66026103436SFelix Kuehling #define PROCESS_RESTORE_TIME_MS 100 66126103436SFelix Kuehling /* Approx. back off time if restore fails due to lack of memory */ 66226103436SFelix Kuehling #define PROCESS_BACK_OFF_TIME_MS 100 66326103436SFelix Kuehling /* Approx. time before evicting the process again */ 66426103436SFelix Kuehling #define PROCESS_ACTIVE_TIME_MS 10 66526103436SFelix Kuehling 6665ec7e028SFelix Kuehling /* 8 byte handle containing GPU ID in the most significant 4 bytes and 6675ec7e028SFelix Kuehling * idr_handle in the least significant 4 bytes 6685ec7e028SFelix Kuehling */ 6695ec7e028SFelix Kuehling #define MAKE_HANDLE(gpu_id, idr_handle) \ 6705ec7e028SFelix Kuehling (((uint64_t)(gpu_id) << 32) + idr_handle) 6715ec7e028SFelix Kuehling #define GET_GPU_ID(handle) (handle >> 32) 6725ec7e028SFelix Kuehling #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 6735ec7e028SFelix Kuehling 674733fa1f7SYong Zhao enum kfd_pdd_bound { 675733fa1f7SYong Zhao PDD_UNBOUND = 0, 676733fa1f7SYong Zhao PDD_BOUND, 677733fa1f7SYong Zhao PDD_BOUND_SUSPENDED, 678733fa1f7SYong Zhao }; 679733fa1f7SYong Zhao 6804327bed2SPhilip Cox #define MAX_SYSFS_FILENAME_LEN 15 68132cb59f3SMukul Joshi 68232cb59f3SMukul Joshi /* 68332cb59f3SMukul Joshi * SDMA counter runs at 100MHz frequency. 68432cb59f3SMukul Joshi * We display SDMA activity in microsecond granularity in sysfs. 68532cb59f3SMukul Joshi * As a result, the divisor is 100. 68632cb59f3SMukul Joshi */ 68732cb59f3SMukul Joshi #define SDMA_ACTIVITY_DIVISOR 100 688d4566deeSMukul Joshi 68919f6d2a6SOded Gabbay /* Data that is per-process-per device. */ 69019f6d2a6SOded Gabbay struct kfd_process_device { 69119f6d2a6SOded Gabbay /* The device that owns this data. */ 69219f6d2a6SOded Gabbay struct kfd_dev *dev; 69319f6d2a6SOded Gabbay 6949fd3f1bfSFelix Kuehling /* The process that owns this kfd_process_device. */ 6959fd3f1bfSFelix Kuehling struct kfd_process *process; 69619f6d2a6SOded Gabbay 69745102048SBen Goz /* per-process-per device QCM data structure */ 69845102048SBen Goz struct qcm_process_device qpd; 69945102048SBen Goz 70019f6d2a6SOded Gabbay /*Apertures*/ 70119f6d2a6SOded Gabbay uint64_t lds_base; 70219f6d2a6SOded Gabbay uint64_t lds_limit; 70319f6d2a6SOded Gabbay uint64_t gpuvm_base; 70419f6d2a6SOded Gabbay uint64_t gpuvm_limit; 70519f6d2a6SOded Gabbay uint64_t scratch_base; 70619f6d2a6SOded Gabbay uint64_t scratch_limit; 70719f6d2a6SOded Gabbay 708403575c4SFelix Kuehling /* VM context for GPUVM allocations */ 709b84394e2SFelix Kuehling struct file *drm_file; 710b40a6ab2SFelix Kuehling void *drm_priv; 711403575c4SFelix Kuehling 71252b29d73SFelix Kuehling /* GPUVM allocations storage */ 71352b29d73SFelix Kuehling struct idr alloc_idr; 71452b29d73SFelix Kuehling 7159fd3f1bfSFelix Kuehling /* Flag used to tell the pdd has dequeued from the dqm. 7169fd3f1bfSFelix Kuehling * This is used to prevent dev->dqm->ops.process_termination() from 7179fd3f1bfSFelix Kuehling * being called twice when it is already called in IOMMU callback 7189fd3f1bfSFelix Kuehling * function. 719a82918f1SBen Goz */ 7209fd3f1bfSFelix Kuehling bool already_dequeued; 7219593f4d6SRajneesh Bhardwaj bool runtime_inuse; 72264d1c3a4SFelix Kuehling 72364d1c3a4SFelix Kuehling /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 72464d1c3a4SFelix Kuehling enum kfd_pdd_bound bound; 725d4566deeSMukul Joshi 726d4566deeSMukul Joshi /* VRAM usage */ 727d4566deeSMukul Joshi uint64_t vram_usage; 728d4566deeSMukul Joshi struct attribute attr_vram; 72932cb59f3SMukul Joshi char vram_filename[MAX_SYSFS_FILENAME_LEN]; 73032cb59f3SMukul Joshi 73132cb59f3SMukul Joshi /* SDMA activity tracking */ 73232cb59f3SMukul Joshi uint64_t sdma_past_activity_counter; 73332cb59f3SMukul Joshi struct attribute attr_sdma; 73432cb59f3SMukul Joshi char sdma_filename[MAX_SYSFS_FILENAME_LEN]; 7354327bed2SPhilip Cox 7364327bed2SPhilip Cox /* Eviction activity tracking */ 7374327bed2SPhilip Cox uint64_t last_evict_timestamp; 7384327bed2SPhilip Cox atomic64_t evict_duration_counter; 7394327bed2SPhilip Cox struct attribute attr_evict; 7404327bed2SPhilip Cox 7414327bed2SPhilip Cox struct kobject *kobj_stats; 74259d7115dSMukul Joshi unsigned int doorbell_index; 743f2fa07b3SRamesh Errabolu 744f2fa07b3SRamesh Errabolu /* 745f2fa07b3SRamesh Errabolu * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process 746f2fa07b3SRamesh Errabolu * that is associated with device encoded by "this" struct instance. The 747f2fa07b3SRamesh Errabolu * value reflects CU usage by all of the waves launched by this process 748f2fa07b3SRamesh Errabolu * on this device. A very important property of occupancy parameter is 749f2fa07b3SRamesh Errabolu * that its value is a snapshot of current use. 750f2fa07b3SRamesh Errabolu * 751f2fa07b3SRamesh Errabolu * Following is to be noted regarding how this parameter is reported: 752f2fa07b3SRamesh Errabolu * 753f2fa07b3SRamesh Errabolu * The number of waves that a CU can launch is limited by couple of 754f2fa07b3SRamesh Errabolu * parameters. These are encoded by struct amdgpu_cu_info instance 755f2fa07b3SRamesh Errabolu * that is part of every device definition. For GFX9 devices this 756f2fa07b3SRamesh Errabolu * translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves 757f2fa07b3SRamesh Errabolu * do not use scratch memory and 32 waves (max_scratch_slots_per_cu) 758f2fa07b3SRamesh Errabolu * when they do use scratch memory. This could change for future 759f2fa07b3SRamesh Errabolu * devices and therefore this example should be considered as a guide. 760f2fa07b3SRamesh Errabolu * 761f2fa07b3SRamesh Errabolu * All CU's of a device are available for the process. This may not be true 762f2fa07b3SRamesh Errabolu * under certain conditions - e.g. CU masking. 763f2fa07b3SRamesh Errabolu * 764f2fa07b3SRamesh Errabolu * Finally number of CU's that are occupied by a process is affected by both 765f2fa07b3SRamesh Errabolu * number of CU's a device has along with number of other competing processes 766f2fa07b3SRamesh Errabolu */ 767f2fa07b3SRamesh Errabolu struct attribute attr_cu_occupancy; 768751580b3SPhilip Yang 769751580b3SPhilip Yang /* sysfs counters for GPU retry fault and page migration tracking */ 770751580b3SPhilip Yang struct kobject *kobj_counters; 771751580b3SPhilip Yang struct attribute attr_faults; 772751580b3SPhilip Yang struct attribute attr_page_in; 773751580b3SPhilip Yang struct attribute attr_page_out; 774751580b3SPhilip Yang uint64_t faults; 775751580b3SPhilip Yang uint64_t page_in; 776751580b3SPhilip Yang uint64_t page_out; 77719f6d2a6SOded Gabbay }; 77819f6d2a6SOded Gabbay 77952a5fdceSAlexey Skidanov #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 78052a5fdceSAlexey Skidanov 78142de677fSPhilip Yang struct svm_range_list { 78242de677fSPhilip Yang struct mutex lock; 78342de677fSPhilip Yang struct rb_root_cached objects; 78442de677fSPhilip Yang struct list_head list; 7854683cfecSPhilip Yang struct work_struct deferred_list_work; 7864683cfecSPhilip Yang struct list_head deferred_range_list; 7874683cfecSPhilip Yang spinlock_t deferred_list_lock; 7888a7c184aSFelix Kuehling atomic_t evicted_ranges; 7892e447728SPhilip Yang atomic_t drain_pagefaults; 7908a7c184aSFelix Kuehling struct delayed_work restore_work; 7915a75ea56SFelix Kuehling DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE); 792a6283010SAlex Sierra struct task_struct *faulting_task; 79342de677fSPhilip Yang }; 79442de677fSPhilip Yang 7954a488a7aSOded Gabbay /* Process data */ 7964a488a7aSOded Gabbay struct kfd_process { 79719f6d2a6SOded Gabbay /* 79819f6d2a6SOded Gabbay * kfd_process are stored in an mm_struct*->kfd_process* 79919f6d2a6SOded Gabbay * hash table (kfd_processes in kfd_process.c) 80019f6d2a6SOded Gabbay */ 80119f6d2a6SOded Gabbay struct hlist_node kfd_processes; 80219f6d2a6SOded Gabbay 8039b56bb11SFelix Kuehling /* 8049b56bb11SFelix Kuehling * Opaque pointer to mm_struct. We don't hold a reference to 8059b56bb11SFelix Kuehling * it so it should never be dereferenced from here. This is 8069b56bb11SFelix Kuehling * only used for looking up processes by their mm. 8079b56bb11SFelix Kuehling */ 8089b56bb11SFelix Kuehling void *mm; 80919f6d2a6SOded Gabbay 8105ce10687SFelix Kuehling struct kref ref; 8115ce10687SFelix Kuehling struct work_struct release_work; 8125ce10687SFelix Kuehling 81319f6d2a6SOded Gabbay struct mutex mutex; 81419f6d2a6SOded Gabbay 81519f6d2a6SOded Gabbay /* 81619f6d2a6SOded Gabbay * In any process, the thread that started main() is the lead 81719f6d2a6SOded Gabbay * thread and outlives the rest. 81819f6d2a6SOded Gabbay * It is here because amd_iommu_bind_pasid wants a task_struct. 819894a8293SFelix Kuehling * It can also be used for safely getting a reference to the 820894a8293SFelix Kuehling * mm_struct of the process. 82119f6d2a6SOded Gabbay */ 82219f6d2a6SOded Gabbay struct task_struct *lead_thread; 82319f6d2a6SOded Gabbay 82419f6d2a6SOded Gabbay /* We want to receive a notification when the mm_struct is destroyed */ 82519f6d2a6SOded Gabbay struct mmu_notifier mmu_notifier; 82619f6d2a6SOded Gabbay 827c7b6bac9SFenghua Yu u32 pasid; 82819f6d2a6SOded Gabbay 82919f6d2a6SOded Gabbay /* 8306ae27841SAlex Sierra * Array of kfd_process_device pointers, 83119f6d2a6SOded Gabbay * one for each device the process is using. 83219f6d2a6SOded Gabbay */ 8336ae27841SAlex Sierra struct kfd_process_device *pdds[MAX_GPU_INSTANCE]; 8346ae27841SAlex Sierra uint32_t n_pdds; 83519f6d2a6SOded Gabbay 83645102048SBen Goz struct process_queue_manager pqm; 83745102048SBen Goz 83819f6d2a6SOded Gabbay /*Is the user space process 32 bit?*/ 83919f6d2a6SOded Gabbay bool is_32bit_user_mode; 840f3a39818SAndrew Lewycky 841f3a39818SAndrew Lewycky /* Event-related data */ 842f3a39818SAndrew Lewycky struct mutex event_mutex; 843482f0777SFelix Kuehling /* Event ID allocator and lookup */ 844482f0777SFelix Kuehling struct idr event_idr; 84550cb7dd9SFelix Kuehling /* Event page */ 84668df0f19SLang Yu u64 signal_handle; 84750cb7dd9SFelix Kuehling struct kfd_signal_page *signal_page; 848b9a5d0a5SFelix Kuehling size_t signal_mapped_size; 849f3a39818SAndrew Lewycky size_t signal_event_count; 850c986169fSFelix Kuehling bool signal_event_limit_reached; 851403575c4SFelix Kuehling 852403575c4SFelix Kuehling /* Information used for memory eviction */ 853403575c4SFelix Kuehling void *kgd_process_info; 854403575c4SFelix Kuehling /* Eviction fence that is attached to all the BOs of this process. The 855403575c4SFelix Kuehling * fence will be triggered during eviction and new one will be created 856403575c4SFelix Kuehling * during restore 857403575c4SFelix Kuehling */ 858403575c4SFelix Kuehling struct dma_fence *ef; 85926103436SFelix Kuehling 86026103436SFelix Kuehling /* Work items for evicting and restoring BOs */ 86126103436SFelix Kuehling struct delayed_work eviction_work; 86226103436SFelix Kuehling struct delayed_work restore_work; 86326103436SFelix Kuehling /* seqno of the last scheduled eviction */ 86426103436SFelix Kuehling unsigned int last_eviction_seqno; 86526103436SFelix Kuehling /* Approx. the last timestamp (in jiffies) when the process was 86626103436SFelix Kuehling * restored after an eviction 86726103436SFelix Kuehling */ 86826103436SFelix Kuehling unsigned long last_restore_timestamp; 869de9f26bbSKent Russell 870de9f26bbSKent Russell /* Kobj for our procfs */ 871de9f26bbSKent Russell struct kobject *kobj; 8726d220a7eSAmber Lin struct kobject *kobj_queues; 873de9f26bbSKent Russell struct attribute attr_pasid; 87440ce74d1SPhilip Yang 87542de677fSPhilip Yang /* shared virtual memory registered by this process */ 87642de677fSPhilip Yang struct svm_range_list svms; 877063e33c5SAlex Sierra 878063e33c5SAlex Sierra bool xnack_enabled; 879b6485bedSTao Zhou 880b6485bedSTao Zhou atomic_t poison; 881cd9f7910SDavid Yat Sin /* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */ 882cd9f7910SDavid Yat Sin bool queues_paused; 8834a488a7aSOded Gabbay }; 8844a488a7aSOded Gabbay 88564d1c3a4SFelix Kuehling #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 88664d1c3a4SFelix Kuehling extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 88764d1c3a4SFelix Kuehling extern struct srcu_struct kfd_processes_srcu; 88864d1c3a4SFelix Kuehling 88976baee6cSOded Gabbay /** 890a4497974SRajneesh Bhardwaj * typedef amdkfd_ioctl_t - typedef for ioctl function pointer. 89176baee6cSOded Gabbay * 892a4497974SRajneesh Bhardwaj * @filep: pointer to file structure. 893a4497974SRajneesh Bhardwaj * @p: amdkfd process pointer. 894a4497974SRajneesh Bhardwaj * @data: pointer to arg that was copied from user. 895a4497974SRajneesh Bhardwaj * 896a4497974SRajneesh Bhardwaj * Return: returns ioctl completion code. 89776baee6cSOded Gabbay */ 89876baee6cSOded Gabbay typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 89976baee6cSOded Gabbay void *data); 90076baee6cSOded Gabbay 90176baee6cSOded Gabbay struct amdkfd_ioctl_desc { 90276baee6cSOded Gabbay unsigned int cmd; 90376baee6cSOded Gabbay int flags; 90476baee6cSOded Gabbay amdkfd_ioctl_t *func; 90576baee6cSOded Gabbay unsigned int cmd_drv; 90676baee6cSOded Gabbay const char *name; 90776baee6cSOded Gabbay }; 90867f7cf9fSshaoyunl bool kfd_dev_is_large_bar(struct kfd_dev *dev); 90976baee6cSOded Gabbay 9101679ae8fSFelix Kuehling int kfd_process_create_wq(void); 91119f6d2a6SOded Gabbay void kfd_process_destroy_wq(void); 912373d7080SFelix Kuehling struct kfd_process *kfd_create_process(struct file *filep); 91319f6d2a6SOded Gabbay struct kfd_process *kfd_get_process(const struct task_struct *); 914c7b6bac9SFenghua Yu struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid); 91526103436SFelix Kuehling struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 9162aeb742bSAlex Sierra 9172aeb742bSAlex Sierra int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id); 91856c5977eSGraham Sider int kfd_process_gpuid_from_adev(struct kfd_process *p, 919cda0f85bSFelix Kuehling struct amdgpu_device *adev, uint32_t *gpuid, 920cda0f85bSFelix Kuehling uint32_t *gpuidx); 9212aeb742bSAlex Sierra static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p, 9222aeb742bSAlex Sierra uint32_t gpuidx, uint32_t *gpuid) { 9232aeb742bSAlex Sierra return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL; 9242aeb742bSAlex Sierra } 9252aeb742bSAlex Sierra static inline struct kfd_process_device *kfd_process_device_from_gpuidx( 9262aeb742bSAlex Sierra struct kfd_process *p, uint32_t gpuidx) { 9272aeb742bSAlex Sierra return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL; 9282aeb742bSAlex Sierra } 9292aeb742bSAlex Sierra 930abb208a8SFelix Kuehling void kfd_unref_process(struct kfd_process *p); 9316b95e797SFelix Kuehling int kfd_process_evict_queues(struct kfd_process *p); 9326b95e797SFelix Kuehling int kfd_process_restore_queues(struct kfd_process *p); 93326103436SFelix Kuehling void kfd_suspend_all_processes(void); 93426103436SFelix Kuehling int kfd_resume_all_processes(void); 93519f6d2a6SOded Gabbay 936b84394e2SFelix Kuehling int kfd_process_device_init_vm(struct kfd_process_device *pdd, 937b84394e2SFelix Kuehling struct file *drm_file); 93864c7f8cfSBen Goz struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 93964c7f8cfSBen Goz struct kfd_process *p); 94019f6d2a6SOded Gabbay struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 941093c7d8cSAlexey Skidanov struct kfd_process *p); 942093c7d8cSAlexey Skidanov struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 943093c7d8cSAlexey Skidanov struct kfd_process *p); 94419f6d2a6SOded Gabbay 945063e33c5SAlex Sierra bool kfd_process_xnack_mode(struct kfd_process *p, bool supported); 946063e33c5SAlex Sierra 947df03ef93SHarish Kasiviswanathan int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 948373d7080SFelix Kuehling struct vm_area_struct *vma); 949373d7080SFelix Kuehling 95052b29d73SFelix Kuehling /* KFD process API for creating and translating handles */ 95152b29d73SFelix Kuehling int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 95252b29d73SFelix Kuehling void *mem); 95352b29d73SFelix Kuehling void *kfd_process_device_translate_handle(struct kfd_process_device *p, 95452b29d73SFelix Kuehling int handle); 95552b29d73SFelix Kuehling void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 95652b29d73SFelix Kuehling int handle); 957011bbb03SRajneesh Bhardwaj struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid); 95852b29d73SFelix Kuehling 95919f6d2a6SOded Gabbay /* PASIDs */ 96019f6d2a6SOded Gabbay int kfd_pasid_init(void); 96119f6d2a6SOded Gabbay void kfd_pasid_exit(void); 96219f6d2a6SOded Gabbay bool kfd_set_pasid_limit(unsigned int new_limit); 96319f6d2a6SOded Gabbay unsigned int kfd_get_pasid_limit(void); 964c7b6bac9SFenghua Yu u32 kfd_pasid_alloc(void); 965c7b6bac9SFenghua Yu void kfd_pasid_free(u32 pasid); 96619f6d2a6SOded Gabbay 96719f6d2a6SOded Gabbay /* Doorbells */ 968ef568db7SFelix Kuehling size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 969735df2baSFelix Kuehling int kfd_doorbell_init(struct kfd_dev *kfd); 970735df2baSFelix Kuehling void kfd_doorbell_fini(struct kfd_dev *kfd); 971df03ef93SHarish Kasiviswanathan int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 972df03ef93SHarish Kasiviswanathan struct vm_area_struct *vma); 973ada2b29cSFelix Kuehling void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 97419f6d2a6SOded Gabbay unsigned int *doorbell_off); 97519f6d2a6SOded Gabbay void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 97619f6d2a6SOded Gabbay u32 read_kernel_doorbell(u32 __iomem *db); 977ada2b29cSFelix Kuehling void write_kernel_doorbell(void __iomem *db, u32 value); 9789d7d0248SFelix Kuehling void write_kernel_doorbell64(void __iomem *db, u64 value); 979339903faSYong Zhao unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, 98059d7115dSMukul Joshi struct kfd_process_device *pdd, 981ef568db7SFelix Kuehling unsigned int doorbell_id); 98259d7115dSMukul Joshi phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd); 98359d7115dSMukul Joshi int kfd_alloc_process_doorbells(struct kfd_dev *kfd, 98459d7115dSMukul Joshi unsigned int *doorbell_index); 98559d7115dSMukul Joshi void kfd_free_process_doorbells(struct kfd_dev *kfd, 98659d7115dSMukul Joshi unsigned int doorbell_index); 9876e81090bSOded Gabbay /* GTT Sub-Allocator */ 9886e81090bSOded Gabbay 9896e81090bSOded Gabbay int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 9906e81090bSOded Gabbay struct kfd_mem_obj **mem_obj); 9916e81090bSOded Gabbay 9926e81090bSOded Gabbay int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 9936e81090bSOded Gabbay 9944a488a7aSOded Gabbay extern struct device *kfd_device; 9954a488a7aSOded Gabbay 996de9f26bbSKent Russell /* KFD's procfs */ 997de9f26bbSKent Russell void kfd_procfs_init(void); 998de9f26bbSKent Russell void kfd_procfs_shutdown(void); 9996d220a7eSAmber Lin int kfd_procfs_add_queue(struct queue *q); 10006d220a7eSAmber Lin void kfd_procfs_del_queue(struct queue *q); 1001de9f26bbSKent Russell 10025b5c4e40SEvgeny Pinchuk /* Topology */ 10035b5c4e40SEvgeny Pinchuk int kfd_topology_init(void); 10045b5c4e40SEvgeny Pinchuk void kfd_topology_shutdown(void); 10055b5c4e40SEvgeny Pinchuk int kfd_topology_add_device(struct kfd_dev *gpu); 10065b5c4e40SEvgeny Pinchuk int kfd_topology_remove_device(struct kfd_dev *gpu); 10073a87177eSHarish Kasiviswanathan struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 10083a87177eSHarish Kasiviswanathan uint32_t proximity_domain); 100944d8cc6fSYong Zhao struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); 10105b5c4e40SEvgeny Pinchuk struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 10115b5c4e40SEvgeny Pinchuk struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 1012574c4183SGraham Sider struct kfd_dev *kfd_device_by_adev(const struct amdgpu_device *adev); 10136d82eb0eSHarish Kasiviswanathan int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 1014520b8fb7SFelix Kuehling int kfd_numa_node_to_apic_id(int numa_node_id); 10156127896fSHuang Rui void kfd_double_confirm_iommu_support(struct kfd_dev *gpu); 10165b5c4e40SEvgeny Pinchuk 10174a488a7aSOded Gabbay /* Interrupts */ 10182249d558SAndrew Lewycky int kfd_interrupt_init(struct kfd_dev *dev); 10192249d558SAndrew Lewycky void kfd_interrupt_exit(struct kfd_dev *dev); 10202249d558SAndrew Lewycky bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 102158e69886SLan Xiao bool interrupt_is_wanted(struct kfd_dev *dev, 102258e69886SLan Xiao const uint32_t *ih_ring_entry, 102358e69886SLan Xiao uint32_t *patched_ihre, bool *flag); 10244a488a7aSOded Gabbay 102519f6d2a6SOded Gabbay /* amdkfd Apertures */ 102619f6d2a6SOded Gabbay int kfd_init_apertures(struct kfd_process *process); 102719f6d2a6SOded Gabbay 10287c9631afSJay Cornwall void kfd_process_set_trap_handler(struct qcm_process_device *qpd, 10297c9631afSJay Cornwall uint64_t tba_addr, 10307c9631afSJay Cornwall uint64_t tma_addr); 10317c9631afSJay Cornwall 103236988070SRajneesh Bhardwaj /* CRIU */ 103336988070SRajneesh Bhardwaj /* 103436988070SRajneesh Bhardwaj * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private 103536988070SRajneesh Bhardwaj * structures: 103636988070SRajneesh Bhardwaj * kfd_criu_process_priv_data 103736988070SRajneesh Bhardwaj * kfd_criu_device_priv_data 103836988070SRajneesh Bhardwaj * kfd_criu_bo_priv_data 103936988070SRajneesh Bhardwaj * kfd_criu_queue_priv_data 104036988070SRajneesh Bhardwaj * kfd_criu_event_priv_data 104136988070SRajneesh Bhardwaj * kfd_criu_svm_range_priv_data 104236988070SRajneesh Bhardwaj */ 104336988070SRajneesh Bhardwaj 104436988070SRajneesh Bhardwaj #define KFD_CRIU_PRIV_VERSION 1 104536988070SRajneesh Bhardwaj 104636988070SRajneesh Bhardwaj struct kfd_criu_process_priv_data { 104736988070SRajneesh Bhardwaj uint32_t version; 104836988070SRajneesh Bhardwaj }; 104936988070SRajneesh Bhardwaj 105036988070SRajneesh Bhardwaj struct kfd_criu_device_priv_data { 105136988070SRajneesh Bhardwaj /* For future use */ 105236988070SRajneesh Bhardwaj uint64_t reserved; 105336988070SRajneesh Bhardwaj }; 105436988070SRajneesh Bhardwaj 105536988070SRajneesh Bhardwaj struct kfd_criu_bo_priv_data { 10565ccbb057SRajneesh Bhardwaj uint64_t user_addr; 10575ccbb057SRajneesh Bhardwaj uint32_t idr_handle; 10585ccbb057SRajneesh Bhardwaj uint32_t mapped_gpuids[MAX_GPU_INSTANCE]; 105936988070SRajneesh Bhardwaj }; 106036988070SRajneesh Bhardwaj 1061626f7b31SDavid Yat Sin /* 1062626f7b31SDavid Yat Sin * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data, 1063626f7b31SDavid Yat Sin * kfd_criu_svm_range_priv_data is the object type 1064626f7b31SDavid Yat Sin */ 1065626f7b31SDavid Yat Sin enum kfd_criu_object_type { 1066626f7b31SDavid Yat Sin KFD_CRIU_OBJECT_TYPE_QUEUE, 1067626f7b31SDavid Yat Sin KFD_CRIU_OBJECT_TYPE_EVENT, 1068626f7b31SDavid Yat Sin KFD_CRIU_OBJECT_TYPE_SVM_RANGE, 1069626f7b31SDavid Yat Sin }; 1070626f7b31SDavid Yat Sin 107136988070SRajneesh Bhardwaj struct kfd_criu_svm_range_priv_data { 107236988070SRajneesh Bhardwaj uint32_t object_type; 107336988070SRajneesh Bhardwaj uint32_t reserved; 107436988070SRajneesh Bhardwaj }; 107536988070SRajneesh Bhardwaj 107636988070SRajneesh Bhardwaj struct kfd_criu_queue_priv_data { 107736988070SRajneesh Bhardwaj uint32_t object_type; 1078626f7b31SDavid Yat Sin uint64_t q_address; 1079626f7b31SDavid Yat Sin uint64_t q_size; 1080626f7b31SDavid Yat Sin uint64_t read_ptr_addr; 1081626f7b31SDavid Yat Sin uint64_t write_ptr_addr; 1082626f7b31SDavid Yat Sin uint64_t doorbell_off; 1083626f7b31SDavid Yat Sin uint64_t eop_ring_buffer_address; 1084626f7b31SDavid Yat Sin uint64_t ctx_save_restore_area_address; 1085626f7b31SDavid Yat Sin uint32_t gpu_id; 1086626f7b31SDavid Yat Sin uint32_t type; 1087626f7b31SDavid Yat Sin uint32_t format; 1088626f7b31SDavid Yat Sin uint32_t q_id; 1089626f7b31SDavid Yat Sin uint32_t priority; 1090626f7b31SDavid Yat Sin uint32_t q_percent; 1091626f7b31SDavid Yat Sin uint32_t doorbell_id; 1092626f7b31SDavid Yat Sin uint32_t is_gws; 1093626f7b31SDavid Yat Sin uint32_t sdma_id; 1094626f7b31SDavid Yat Sin uint32_t eop_ring_buffer_size; 1095626f7b31SDavid Yat Sin uint32_t ctx_save_restore_area_size; 1096626f7b31SDavid Yat Sin uint32_t ctl_stack_size; 1097626f7b31SDavid Yat Sin uint32_t mqd_size; 109836988070SRajneesh Bhardwaj }; 109936988070SRajneesh Bhardwaj 110036988070SRajneesh Bhardwaj struct kfd_criu_event_priv_data { 110136988070SRajneesh Bhardwaj uint32_t object_type; 110236988070SRajneesh Bhardwaj uint32_t reserved; 110336988070SRajneesh Bhardwaj }; 110436988070SRajneesh Bhardwaj 1105626f7b31SDavid Yat Sin int kfd_process_get_queue_info(struct kfd_process *p, 1106626f7b31SDavid Yat Sin uint32_t *num_queues, 1107626f7b31SDavid Yat Sin uint64_t *priv_data_sizes); 1108626f7b31SDavid Yat Sin 1109626f7b31SDavid Yat Sin int kfd_criu_checkpoint_queues(struct kfd_process *p, 1110626f7b31SDavid Yat Sin uint8_t __user *user_priv_data, 1111626f7b31SDavid Yat Sin uint64_t *priv_data_offset); 1112626f7b31SDavid Yat Sin 1113626f7b31SDavid Yat Sin int kfd_criu_restore_queue(struct kfd_process *p, 1114626f7b31SDavid Yat Sin uint8_t __user *user_priv_data, 1115626f7b31SDavid Yat Sin uint64_t *priv_data_offset, 1116626f7b31SDavid Yat Sin uint64_t max_priv_data_size); 111736988070SRajneesh Bhardwaj /* CRIU - End */ 111836988070SRajneesh Bhardwaj 1119ed6e6a34SBen Goz /* Queue Context Management */ 1120e88a614cSEdward O'Callaghan int init_queue(struct queue **q, const struct queue_properties *properties); 1121ed6e6a34SBen Goz void uninit_queue(struct queue *q); 112245102048SBen Goz void print_queue_properties(struct queue_properties *q); 1123ed6e6a34SBen Goz void print_queue(struct queue *q); 1124ed6e6a34SBen Goz 11254b8f589bSBen Goz struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 11264b8f589bSBen Goz struct kfd_dev *dev); 1127ee04955aSFelix Kuehling struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 1128ee04955aSFelix Kuehling struct kfd_dev *dev); 11294b8f589bSBen Goz struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 11304b8f589bSBen Goz struct kfd_dev *dev); 1131ee04955aSFelix Kuehling struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 1132ee04955aSFelix Kuehling struct kfd_dev *dev); 1133b91d43ddSFelix Kuehling struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 1134b91d43ddSFelix Kuehling struct kfd_dev *dev); 113514328aa5SPhilip Cox struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, 113614328aa5SPhilip Cox struct kfd_dev *dev); 113764c7f8cfSBen Goz struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 113864c7f8cfSBen Goz void device_queue_manager_uninit(struct device_queue_manager *dqm); 1139241f24f8SBen Goz struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 1140241f24f8SBen Goz enum kfd_queue_type type); 1141c2a77fdeSFelix Kuehling void kernel_queue_uninit(struct kernel_queue *kq, bool hanging); 1142c7b6bac9SFenghua Yu int kfd_process_vm_fault(struct device_queue_manager *dqm, u32 pasid); 1143241f24f8SBen Goz 114445102048SBen Goz /* Process Queue Manager */ 114545102048SBen Goz struct process_queue_node { 114645102048SBen Goz struct queue *q; 114745102048SBen Goz struct kernel_queue *kq; 114845102048SBen Goz struct list_head process_queue_list; 114945102048SBen Goz }; 115045102048SBen Goz 11519fd3f1bfSFelix Kuehling void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 11529fd3f1bfSFelix Kuehling void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 115345102048SBen Goz int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 115445102048SBen Goz void pqm_uninit(struct process_queue_manager *pqm); 115545102048SBen Goz int pqm_create_queue(struct process_queue_manager *pqm, 115645102048SBen Goz struct kfd_dev *dev, 115745102048SBen Goz struct file *f, 115845102048SBen Goz struct queue_properties *properties, 1159e47a8b52SYong Zhao unsigned int *qid, 11608668dfc3SDavid Yat Sin const struct kfd_criu_queue_priv_data *q_data, 1161*42c6c482SDavid Yat Sin const void *restore_mqd, 1162e47a8b52SYong Zhao uint32_t *p_doorbell_offset_in_process); 116345102048SBen Goz int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 11647c695a2cSLang Yu int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid, 116545102048SBen Goz struct queue_properties *p); 11667c695a2cSLang Yu int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid, 11677c695a2cSLang Yu struct mqd_update_info *minfo); 1168eb82da1dSOak Zeng int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, 1169eb82da1dSOak Zeng void *gws); 1170fbeb661bSYair Shachar struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 1171fbeb661bSYair Shachar unsigned int qid); 11725bb4b78bSOak Zeng struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, 11735bb4b78bSOak Zeng unsigned int qid); 11745df099e8SJay Cornwall int pqm_get_wave_state(struct process_queue_manager *pqm, 11755df099e8SJay Cornwall unsigned int qid, 11765df099e8SJay Cornwall void __user *ctl_stack, 11775df099e8SJay Cornwall u32 *ctl_stack_used_size, 11785df099e8SJay Cornwall u32 *save_area_used_size); 117945102048SBen Goz 1180b010affeSQu Huang int amdkfd_fence_wait_timeout(uint64_t *fence_addr, 1181b010affeSQu Huang uint64_t fence_value, 11828c72c3d7SYong Zhao unsigned int timeout_ms); 1183788bf83dSYair Shachar 1184*42c6c482SDavid Yat Sin int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm, 1185*42c6c482SDavid Yat Sin unsigned int qid, 1186*42c6c482SDavid Yat Sin u32 *mqd_size); 1187*42c6c482SDavid Yat Sin 1188ed6e6a34SBen Goz /* Packet Manager */ 1189ed6e6a34SBen Goz 119064c7f8cfSBen Goz #define KFD_FENCE_COMPLETED (100) 119164c7f8cfSBen Goz #define KFD_FENCE_INIT (10) 1192241f24f8SBen Goz 1193ed6e6a34SBen Goz struct packet_manager { 1194ed6e6a34SBen Goz struct device_queue_manager *dqm; 1195ed6e6a34SBen Goz struct kernel_queue *priv_queue; 1196ed6e6a34SBen Goz struct mutex lock; 1197ed6e6a34SBen Goz bool allocated; 1198ed6e6a34SBen Goz struct kfd_mem_obj *ib_buffer_obj; 1199851a645eSFelix Kuehling unsigned int ib_size_bytes; 1200819ec5acSFelix Kuehling bool is_over_subscription; 1201f6e27ff1SFelix Kuehling 1202f6e27ff1SFelix Kuehling const struct packet_manager_funcs *pmf; 1203ed6e6a34SBen Goz }; 1204ed6e6a34SBen Goz 1205f6e27ff1SFelix Kuehling struct packet_manager_funcs { 1206f6e27ff1SFelix Kuehling /* Support ASIC-specific packet formats for PM4 packets */ 1207f6e27ff1SFelix Kuehling int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 1208f6e27ff1SFelix Kuehling struct qcm_process_device *qpd); 1209f6e27ff1SFelix Kuehling int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 1210f6e27ff1SFelix Kuehling uint64_t ib, size_t ib_size_in_dwords, bool chain); 1211f6e27ff1SFelix Kuehling int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 1212f6e27ff1SFelix Kuehling struct scheduling_resources *res); 1213f6e27ff1SFelix Kuehling int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 1214f6e27ff1SFelix Kuehling struct queue *q, bool is_static); 1215f6e27ff1SFelix Kuehling int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 1216f6e27ff1SFelix Kuehling enum kfd_queue_type type, 1217f6e27ff1SFelix Kuehling enum kfd_unmap_queues_filter mode, 1218f6e27ff1SFelix Kuehling uint32_t filter_param, bool reset, 1219f6e27ff1SFelix Kuehling unsigned int sdma_engine); 1220f6e27ff1SFelix Kuehling int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 1221b010affeSQu Huang uint64_t fence_address, uint64_t fence_value); 1222f6e27ff1SFelix Kuehling int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 1223f6e27ff1SFelix Kuehling 1224f6e27ff1SFelix Kuehling /* Packet sizes */ 1225f6e27ff1SFelix Kuehling int map_process_size; 1226f6e27ff1SFelix Kuehling int runlist_size; 1227f6e27ff1SFelix Kuehling int set_resources_size; 1228f6e27ff1SFelix Kuehling int map_queues_size; 1229f6e27ff1SFelix Kuehling int unmap_queues_size; 1230f6e27ff1SFelix Kuehling int query_status_size; 1231f6e27ff1SFelix Kuehling int release_mem_size; 1232f6e27ff1SFelix Kuehling }; 1233f6e27ff1SFelix Kuehling 1234f6e27ff1SFelix Kuehling extern const struct packet_manager_funcs kfd_vi_pm_funcs; 1235454150b1SFelix Kuehling extern const struct packet_manager_funcs kfd_v9_pm_funcs; 1236fd6a440eSJonathan Kim extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs; 1237f6e27ff1SFelix Kuehling 123864c7f8cfSBen Goz int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 1239c2a77fdeSFelix Kuehling void pm_uninit(struct packet_manager *pm, bool hanging); 124064c7f8cfSBen Goz int pm_send_set_resources(struct packet_manager *pm, 124164c7f8cfSBen Goz struct scheduling_resources *res); 124264c7f8cfSBen Goz int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 124364c7f8cfSBen Goz int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 1244b010affeSQu Huang uint64_t fence_value); 124564c7f8cfSBen Goz 124664c7f8cfSBen Goz int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, 12477da2bcf8SYong Zhao enum kfd_unmap_queues_filter mode, 124864c7f8cfSBen Goz uint32_t filter_param, bool reset, 124964c7f8cfSBen Goz unsigned int sdma_engine); 125064c7f8cfSBen Goz 1251241f24f8SBen Goz void pm_release_ib(struct packet_manager *pm); 1252241f24f8SBen Goz 1253454150b1SFelix Kuehling /* Following PM funcs can be shared among VI and AI */ 1254454150b1SFelix Kuehling unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 125514328aa5SPhilip Cox 125619f6d2a6SOded Gabbay uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 125719f6d2a6SOded Gabbay 1258f3a39818SAndrew Lewycky /* Events */ 1259f3a39818SAndrew Lewycky extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 1260ca750681SFelix Kuehling extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 1261ca750681SFelix Kuehling 1262930c5ff4SAlexey Skidanov extern const struct kfd_device_global_init_class device_global_init_class_cik; 1263f3a39818SAndrew Lewycky 1264f3a39818SAndrew Lewycky void kfd_event_init_process(struct kfd_process *p); 1265f3a39818SAndrew Lewycky void kfd_event_free_process(struct kfd_process *p); 1266f3a39818SAndrew Lewycky int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 1267f3a39818SAndrew Lewycky int kfd_wait_on_events(struct kfd_process *p, 126859d3e8beSAlexey Skidanov uint32_t num_events, void __user *data, 1269f3a39818SAndrew Lewycky bool all, uint32_t user_timeout_ms, 1270fdf0c833SFelix Kuehling uint32_t *wait_result); 1271c7b6bac9SFenghua Yu void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, 1272f3a39818SAndrew Lewycky uint32_t valid_id_bits); 127359d3e8beSAlexey Skidanov void kfd_signal_iommu_event(struct kfd_dev *dev, 1274c7b6bac9SFenghua Yu u32 pasid, unsigned long address, 127559d3e8beSAlexey Skidanov bool is_write_requested, bool is_execute_requested); 1276c7b6bac9SFenghua Yu void kfd_signal_hw_exception_event(u32 pasid); 1277f3a39818SAndrew Lewycky int kfd_set_event(struct kfd_process *p, uint32_t event_id); 1278f3a39818SAndrew Lewycky int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 12790fc8011fSFelix Kuehling int kfd_event_page_set(struct kfd_process *p, void *kernel_address, 12800fc8011fSFelix Kuehling uint64_t size); 1281f3a39818SAndrew Lewycky int kfd_event_create(struct file *devkfd, struct kfd_process *p, 1282f3a39818SAndrew Lewycky uint32_t event_type, bool auto_reset, uint32_t node_id, 1283f3a39818SAndrew Lewycky uint32_t *event_id, uint32_t *event_trigger_data, 1284f3a39818SAndrew Lewycky uint64_t *event_page_offset, uint32_t *event_slot_index); 1285f3a39818SAndrew Lewycky int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 1286f3a39818SAndrew Lewycky 1287c7b6bac9SFenghua Yu void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, 12882640c3faSshaoyunl struct kfd_vm_fault_info *info); 12892640c3faSshaoyunl 1290e42051d2SShaoyun Liu void kfd_signal_reset_event(struct kfd_dev *dev); 1291e42051d2SShaoyun Liu 1292e2b1f9f5SDennis Li void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid); 1293e2b1f9f5SDennis Li 12943543b055SEric Huang void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type); 1295403575c4SFelix Kuehling 1296c3447e81SBen Goz int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p); 1297c3447e81SBen Goz 1298e42051d2SShaoyun Liu bool kfd_is_locked(void); 1299e42051d2SShaoyun Liu 1300f756e631SHarish Kasiviswanathan /* Compute profile */ 1301f756e631SHarish Kasiviswanathan void kfd_inc_compute_active(struct kfd_dev *dev); 1302f756e631SHarish Kasiviswanathan void kfd_dec_compute_active(struct kfd_dev *dev); 1303f756e631SHarish Kasiviswanathan 13046b855f7bSHarish Kasiviswanathan /* Cgroup Support */ 13056b855f7bSHarish Kasiviswanathan /* Check with device cgroup if @kfd device is accessible */ 13066b855f7bSHarish Kasiviswanathan static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) 13076b855f7bSHarish Kasiviswanathan { 1308eec8fd02SOdin Ugedal #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF) 13096b855f7bSHarish Kasiviswanathan struct drm_device *ddev = kfd->ddev; 13106b855f7bSHarish Kasiviswanathan 131199c7b309SLorenz Brun return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR, 13126b855f7bSHarish Kasiviswanathan ddev->render->index, 13136b855f7bSHarish Kasiviswanathan DEVCG_ACC_WRITE | DEVCG_ACC_READ); 13146b855f7bSHarish Kasiviswanathan #else 13156b855f7bSHarish Kasiviswanathan return 0; 13166b855f7bSHarish Kasiviswanathan #endif 13176b855f7bSHarish Kasiviswanathan } 13186b855f7bSHarish Kasiviswanathan 1319851a645eSFelix Kuehling /* Debugfs */ 1320851a645eSFelix Kuehling #if defined(CONFIG_DEBUG_FS) 1321851a645eSFelix Kuehling 1322851a645eSFelix Kuehling void kfd_debugfs_init(void); 1323851a645eSFelix Kuehling void kfd_debugfs_fini(void); 1324851a645eSFelix Kuehling int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 1325851a645eSFelix Kuehling int pqm_debugfs_mqds(struct seq_file *m, void *data); 1326851a645eSFelix Kuehling int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 1327851a645eSFelix Kuehling int dqm_debugfs_hqds(struct seq_file *m, void *data); 1328851a645eSFelix Kuehling int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 1329851a645eSFelix Kuehling int pm_debugfs_runlist(struct seq_file *m, void *data); 1330851a645eSFelix Kuehling 1331a29ec470SShaoyun Liu int kfd_debugfs_hang_hws(struct kfd_dev *dev); 1332a29ec470SShaoyun Liu int pm_debugfs_hang_hws(struct packet_manager *pm); 13334f942aaeSOak Zeng int dqm_debugfs_hang_hws(struct device_queue_manager *dqm); 1334a29ec470SShaoyun Liu 1335851a645eSFelix Kuehling #else 1336851a645eSFelix Kuehling 1337851a645eSFelix Kuehling static inline void kfd_debugfs_init(void) {} 1338851a645eSFelix Kuehling static inline void kfd_debugfs_fini(void) {} 1339851a645eSFelix Kuehling 1340851a645eSFelix Kuehling #endif 1341851a645eSFelix Kuehling 13424a488a7aSOded Gabbay #endif 1343