1d87f36a0SRajneesh Bhardwaj /* SPDX-License-Identifier: GPL-2.0 OR MIT */ 24a488a7aSOded Gabbay /* 3d87f36a0SRajneesh Bhardwaj * Copyright 2014-2022 Advanced Micro Devices, Inc. 44a488a7aSOded Gabbay * 54a488a7aSOded Gabbay * Permission is hereby granted, free of charge, to any person obtaining a 64a488a7aSOded Gabbay * copy of this software and associated documentation files (the "Software"), 74a488a7aSOded Gabbay * to deal in the Software without restriction, including without limitation 84a488a7aSOded Gabbay * the rights to use, copy, modify, merge, publish, distribute, sublicense, 94a488a7aSOded Gabbay * and/or sell copies of the Software, and to permit persons to whom the 104a488a7aSOded Gabbay * Software is furnished to do so, subject to the following conditions: 114a488a7aSOded Gabbay * 124a488a7aSOded Gabbay * The above copyright notice and this permission notice shall be included in 134a488a7aSOded Gabbay * all copies or substantial portions of the Software. 144a488a7aSOded Gabbay * 154a488a7aSOded Gabbay * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 164a488a7aSOded Gabbay * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 174a488a7aSOded Gabbay * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 184a488a7aSOded Gabbay * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 194a488a7aSOded Gabbay * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 204a488a7aSOded Gabbay * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 214a488a7aSOded Gabbay * OTHER DEALINGS IN THE SOFTWARE. 224a488a7aSOded Gabbay */ 234a488a7aSOded Gabbay 244a488a7aSOded Gabbay #ifndef KFD_PRIV_H_INCLUDED 254a488a7aSOded Gabbay #define KFD_PRIV_H_INCLUDED 264a488a7aSOded Gabbay 274a488a7aSOded Gabbay #include <linux/hashtable.h> 284a488a7aSOded Gabbay #include <linux/mmu_notifier.h> 294a488a7aSOded Gabbay #include <linux/mutex.h> 304a488a7aSOded Gabbay #include <linux/types.h> 314a488a7aSOded Gabbay #include <linux/atomic.h> 324a488a7aSOded Gabbay #include <linux/workqueue.h> 334a488a7aSOded Gabbay #include <linux/spinlock.h> 3419f6d2a6SOded Gabbay #include <linux/kfd_ioctl.h> 35482f0777SFelix Kuehling #include <linux/idr.h> 3604ad47bdSAndres Rodriguez #include <linux/kfifo.h> 37851a645eSFelix Kuehling #include <linux/seq_file.h> 385ce10687SFelix Kuehling #include <linux/kref.h> 39de9f26bbSKent Russell #include <linux/sysfs.h> 406b855f7bSHarish Kasiviswanathan #include <linux/device_cgroup.h> 411cd4d9eeSStephen Rothwell #include <drm/drm_file.h> 421cd4d9eeSStephen Rothwell #include <drm/drm_drv.h> 431cd4d9eeSStephen Rothwell #include <drm/drm_device.h> 4499c7b309SLorenz Brun #include <drm/drm_ioctl.h> 454a488a7aSOded Gabbay #include <kgd_kfd_interface.h> 466d220a7eSAmber Lin #include <linux/swap.h> 474a488a7aSOded Gabbay 48e596b903SYong Zhao #include "amd_shared.h" 496ae27841SAlex Sierra #include "amdgpu.h" 50e596b903SYong Zhao 51af47b390SLaura Abbott #define KFD_MAX_RING_ENTRY_SIZE 8 52af47b390SLaura Abbott 535b5c4e40SEvgeny Pinchuk #define KFD_SYSFS_FILE_MODE 0444 545b5c4e40SEvgeny Pinchuk 55df03ef93SHarish Kasiviswanathan /* GPU ID hash width in bits */ 56df03ef93SHarish Kasiviswanathan #define KFD_GPU_ID_HASH_WIDTH 16 57df03ef93SHarish Kasiviswanathan 58df03ef93SHarish Kasiviswanathan /* Use upper bits of mmap offset to store KFD driver specific information. 59df03ef93SHarish Kasiviswanathan * BITS[63:62] - Encode MMAP type 60df03ef93SHarish Kasiviswanathan * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to 61df03ef93SHarish Kasiviswanathan * BITS[45:0] - MMAP offset value 62df03ef93SHarish Kasiviswanathan * 63df03ef93SHarish Kasiviswanathan * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these 64df03ef93SHarish Kasiviswanathan * defines are w.r.t to PAGE_SIZE 65df03ef93SHarish Kasiviswanathan */ 6629453755SYong Zhao #define KFD_MMAP_TYPE_SHIFT 62 67df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT) 68df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT) 69df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT) 70df03ef93SHarish Kasiviswanathan #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT) 71d33ea570SOak Zeng #define KFD_MMAP_TYPE_MMIO (0x0ULL << KFD_MMAP_TYPE_SHIFT) 72df03ef93SHarish Kasiviswanathan 7329453755SYong Zhao #define KFD_MMAP_GPU_ID_SHIFT 46 74df03ef93SHarish Kasiviswanathan #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \ 75df03ef93SHarish Kasiviswanathan << KFD_MMAP_GPU_ID_SHIFT) 76df03ef93SHarish Kasiviswanathan #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\ 77df03ef93SHarish Kasiviswanathan & KFD_MMAP_GPU_ID_MASK) 7829453755SYong Zhao #define KFD_MMAP_GET_GPU_ID(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \ 79df03ef93SHarish Kasiviswanathan >> KFD_MMAP_GPU_ID_SHIFT) 80df03ef93SHarish Kasiviswanathan 81ed6e6a34SBen Goz /* 82ed6e6a34SBen Goz * When working with cp scheduler we should assign the HIQ manually or via 83e7016d8eSYong Zhao * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot 84ed6e6a34SBen Goz * definitions for Kaveri. In Kaveri only the first ME queues participates 85ed6e6a34SBen Goz * in the cp scheduling taking that in mind we set the HIQ slot in the 86ed6e6a34SBen Goz * second ME. 87ed6e6a34SBen Goz */ 88ed6e6a34SBen Goz #define KFD_CIK_HIQ_PIPE 4 89ed6e6a34SBen Goz #define KFD_CIK_HIQ_QUEUE 0 90ed6e6a34SBen Goz 915b5c4e40SEvgeny Pinchuk /* Macro for allocating structures */ 925b5c4e40SEvgeny Pinchuk #define kfd_alloc_struct(ptr_to_struct) \ 935b5c4e40SEvgeny Pinchuk ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL)) 945b5c4e40SEvgeny Pinchuk 9519f6d2a6SOded Gabbay #define KFD_MAX_NUM_OF_PROCESSES 512 96b8cbab04SOded Gabbay #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024 9719f6d2a6SOded Gabbay 9819f6d2a6SOded Gabbay /* 99373d7080SFelix Kuehling * Size of the per-process TBA+TMA buffer: 2 pages 100373d7080SFelix Kuehling * 101373d7080SFelix Kuehling * The first page is the TBA used for the CWSR ISA code. The second 102a4497974SRajneesh Bhardwaj * page is used as TMA for user-mode trap handler setup in daisy-chain mode. 103373d7080SFelix Kuehling */ 104373d7080SFelix Kuehling #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2) 105373d7080SFelix Kuehling #define KFD_CWSR_TMA_OFFSET PAGE_SIZE 106373d7080SFelix Kuehling 10774523943SYong Zhao #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \ 10874523943SYong Zhao (KFD_MAX_NUM_OF_PROCESSES * \ 10974523943SYong Zhao KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 11074523943SYong Zhao 11174523943SYong Zhao #define KFD_KERNEL_QUEUE_SIZE 2048 11274523943SYong Zhao 11314328aa5SPhilip Cox #define KFD_UNMAP_LATENCY_MS (4000) 11414328aa5SPhilip Cox 115373d7080SFelix Kuehling /* 1161f86805aSYong Zhao * 512 = 0x200 1171f86805aSYong Zhao * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the 1181f86805aSYong Zhao * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA. 1191f86805aSYong Zhao * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC 1201f86805aSYong Zhao * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in 1211f86805aSYong Zhao * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE. 1221f86805aSYong Zhao */ 1231f86805aSYong Zhao #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512 1241f86805aSYong Zhao 12536988070SRajneesh Bhardwaj /** 12636988070SRajneesh Bhardwaj * enum kfd_ioctl_flags - KFD ioctl flags 12736988070SRajneesh Bhardwaj * Various flags that can be set in &amdkfd_ioctl_desc.flags to control how 12836988070SRajneesh Bhardwaj * userspace can use a given ioctl. 12936988070SRajneesh Bhardwaj */ 13036988070SRajneesh Bhardwaj enum kfd_ioctl_flags { 13136988070SRajneesh Bhardwaj /* 13236988070SRajneesh Bhardwaj * @KFD_IOC_FLAG_CHECKPOINT_RESTORE: 13336988070SRajneesh Bhardwaj * Certain KFD ioctls such as AMDKFD_IOC_CRIU_OP can potentially 13436988070SRajneesh Bhardwaj * perform privileged operations and load arbitrary data into MQDs and 13536988070SRajneesh Bhardwaj * eventually HQD registers when the queue is mapped by HWS. In order to 13636988070SRajneesh Bhardwaj * prevent this we should perform additional security checks. 13736988070SRajneesh Bhardwaj * 13836988070SRajneesh Bhardwaj * This is equivalent to callers with the CHECKPOINT_RESTORE capability. 13936988070SRajneesh Bhardwaj * 14036988070SRajneesh Bhardwaj * Note: Since earlier versions of docker do not support CHECKPOINT_RESTORE, 14136988070SRajneesh Bhardwaj * we also allow ioctls with SYS_ADMIN capability. 14236988070SRajneesh Bhardwaj */ 14336988070SRajneesh Bhardwaj KFD_IOC_FLAG_CHECKPOINT_RESTORE = BIT(0), 14436988070SRajneesh Bhardwaj }; 1451f86805aSYong Zhao /* 146b8cbab04SOded Gabbay * Kernel module parameter to specify maximum number of supported queues per 147b8cbab04SOded Gabbay * device 14819f6d2a6SOded Gabbay */ 149b8cbab04SOded Gabbay extern int max_num_of_queues_per_device; 15019f6d2a6SOded Gabbay 151ed6e6a34SBen Goz 15231c21fecSBen Goz /* Kernel module parameter to specify the scheduling policy */ 15331c21fecSBen Goz extern int sched_policy; 15431c21fecSBen Goz 155a99c6d4fSFelix Kuehling /* 156a99c6d4fSFelix Kuehling * Kernel module parameter to specify the maximum process 157a99c6d4fSFelix Kuehling * number per HW scheduler 158a99c6d4fSFelix Kuehling */ 159a99c6d4fSFelix Kuehling extern int hws_max_conc_proc; 160a99c6d4fSFelix Kuehling 161373d7080SFelix Kuehling extern int cwsr_enable; 162373d7080SFelix Kuehling 16381663016SOded Gabbay /* 16481663016SOded Gabbay * Kernel module parameter to specify whether to send sigterm to HSA process on 16581663016SOded Gabbay * unhandled exception 16681663016SOded Gabbay */ 16781663016SOded Gabbay extern int send_sigterm; 16881663016SOded Gabbay 169ebcfd1e2SFelix Kuehling /* 170374200b1SFelix Kuehling * This kernel module is used to simulate large bar machine on non-large bar 171374200b1SFelix Kuehling * enabled machines. 172374200b1SFelix Kuehling */ 173374200b1SFelix Kuehling extern int debug_largebar; 174374200b1SFelix Kuehling 175374200b1SFelix Kuehling /* 176ebcfd1e2SFelix Kuehling * Ignore CRAT table during KFD initialization, can be used to work around 177ebcfd1e2SFelix Kuehling * broken CRAT tables on some AMD systems 178ebcfd1e2SFelix Kuehling */ 179ebcfd1e2SFelix Kuehling extern int ignore_crat; 180ebcfd1e2SFelix Kuehling 181a4497974SRajneesh Bhardwaj /* Set sh_mem_config.retry_disable on GFX v9 */ 18275ee6487SFelix Kuehling extern int amdgpu_noretry; 183bed4f110SFelix Kuehling 184a4497974SRajneesh Bhardwaj /* Halt if HWS hang is detected */ 1850e9a860cSYong Zhao extern int halt_if_hws_hang; 1860e9a860cSYong Zhao 187a4497974SRajneesh Bhardwaj /* Whether MEC FW support GWS barriers */ 18829e76462SOak Zeng extern bool hws_gws_support; 18929e76462SOak Zeng 190a4497974SRajneesh Bhardwaj /* Queue preemption timeout in ms */ 19114328aa5SPhilip Cox extern int queue_preemption_timeout_ms; 19214328aa5SPhilip Cox 1936d909c5dSOak Zeng /* 1946d909c5dSOak Zeng * Don't evict process queues on vm fault 1956d909c5dSOak Zeng */ 1966d909c5dSOak Zeng extern int amdgpu_no_queue_eviction_on_vm_fault; 1976d909c5dSOak Zeng 198a4497974SRajneesh Bhardwaj /* Enable eviction debug messages */ 199b2057956SFelix Kuehling extern bool debug_evictions; 200b2057956SFelix Kuehling 201ed6e6a34SBen Goz enum cache_policy { 202ed6e6a34SBen Goz cache_policy_coherent, 203ed6e6a34SBen Goz cache_policy_noncoherent 204ed6e6a34SBen Goz }; 205ed6e6a34SBen Goz 206dd0ae064SGraham Sider #define KFD_GC_VERSION(dev) ((dev)->adev->ip_versions[GC_HWIP][0]) 207dd0ae064SGraham Sider #define KFD_IS_SOC15(dev) ((KFD_GC_VERSION(dev)) >= (IP_VERSION(9, 0, 1))) 208ef568db7SFelix Kuehling 209f3a39818SAndrew Lewycky struct kfd_event_interrupt_class { 210f3a39818SAndrew Lewycky bool (*interrupt_isr)(struct kfd_dev *dev, 21158e69886SLan Xiao const uint32_t *ih_ring_entry, uint32_t *patched_ihre, 21258e69886SLan Xiao bool *patched_flag); 213f3a39818SAndrew Lewycky void (*interrupt_wq)(struct kfd_dev *dev, 214f3a39818SAndrew Lewycky const uint32_t *ih_ring_entry); 215f3a39818SAndrew Lewycky }; 216f3a39818SAndrew Lewycky 2174a488a7aSOded Gabbay struct kfd_device_info { 2189d6fa9c7SGraham Sider uint32_t gfx_target_version; 219f3a39818SAndrew Lewycky const struct kfd_event_interrupt_class *event_interrupt_class; 2204a488a7aSOded Gabbay unsigned int max_pasid_bits; 221992839adSYair Shachar unsigned int max_no_of_hqd; 222ada2b29cSFelix Kuehling unsigned int doorbell_size; 2234a488a7aSOded Gabbay size_t ih_ring_entry_size; 224f7c826adSAlexey Skidanov uint8_t num_of_watch_points; 22519f6d2a6SOded Gabbay uint16_t mqd_size_aligned; 226373d7080SFelix Kuehling bool supports_cwsr; 22764d1c3a4SFelix Kuehling bool needs_iommu_device; 2283ee2d00cSFelix Kuehling bool needs_pci_atomics; 229fb932dfeSFelix Kuehling uint32_t no_atomic_fw_version; 230d5094189SShaoyun Liu unsigned int num_sdma_queues_per_engine; 2314a488a7aSOded Gabbay }; 2324a488a7aSOded Gabbay 233ee2f17f4SAmber Lin unsigned int kfd_get_num_sdma_engines(struct kfd_dev *kdev); 234ee2f17f4SAmber Lin unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_dev *kdev); 235ee2f17f4SAmber Lin 23636b5c08fSOded Gabbay struct kfd_mem_obj { 23736b5c08fSOded Gabbay uint32_t range_start; 23836b5c08fSOded Gabbay uint32_t range_end; 23936b5c08fSOded Gabbay uint64_t gpu_addr; 24036b5c08fSOded Gabbay uint32_t *cpu_ptr; 241b91d43ddSFelix Kuehling void *gtt_mem; 24236b5c08fSOded Gabbay }; 24336b5c08fSOded Gabbay 24444008d7aSYong Zhao struct kfd_vmid_info { 24544008d7aSYong Zhao uint32_t first_vmid_kfd; 24644008d7aSYong Zhao uint32_t last_vmid_kfd; 24744008d7aSYong Zhao uint32_t vmid_num_kfd; 24844008d7aSYong Zhao }; 24944008d7aSYong Zhao 2504a488a7aSOded Gabbay struct kfd_dev { 251c6c57446SGraham Sider struct amdgpu_device *adev; 2524a488a7aSOded Gabbay 253f0dc99a6SGraham Sider struct kfd_device_info device_info; 2544a488a7aSOded Gabbay struct pci_dev *pdev; 2553a0c3423SHarish Kasiviswanathan struct drm_device *ddev; 2564a488a7aSOded Gabbay 2574a488a7aSOded Gabbay unsigned int id; /* topology stub index */ 2584a488a7aSOded Gabbay 25919f6d2a6SOded Gabbay phys_addr_t doorbell_base; /* Start of actual doorbells used by 26019f6d2a6SOded Gabbay * KFD. It is aligned for mapping 26119f6d2a6SOded Gabbay * into user mode 26219f6d2a6SOded Gabbay */ 263339903faSYong Zhao size_t doorbell_base_dw_offset; /* Offset from the start of the PCI 264339903faSYong Zhao * doorbell BAR to the first KFD 265339903faSYong Zhao * doorbell in dwords. GFX reserves 266339903faSYong Zhao * the segment before this offset. 26719f6d2a6SOded Gabbay */ 26819f6d2a6SOded Gabbay u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells 26919f6d2a6SOded Gabbay * page used by kernel queue 27019f6d2a6SOded Gabbay */ 27119f6d2a6SOded Gabbay 2724a488a7aSOded Gabbay struct kgd2kfd_shared_resources shared_resources; 27344008d7aSYong Zhao struct kfd_vmid_info vm_info; 2744a488a7aSOded Gabbay 275cea405b1SXihan Zhang const struct kfd2kgd_calls *kfd2kgd; 276cea405b1SXihan Zhang struct mutex doorbell_mutex; 277f761d8bdSJoe Perches DECLARE_BITMAP(doorbell_available_index, 278f761d8bdSJoe Perches KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 279cea405b1SXihan Zhang 28036b5c08fSOded Gabbay void *gtt_mem; 28136b5c08fSOded Gabbay uint64_t gtt_start_gpu_addr; 28236b5c08fSOded Gabbay void *gtt_start_cpu_ptr; 28336b5c08fSOded Gabbay void *gtt_sa_bitmap; 28436b5c08fSOded Gabbay struct mutex gtt_sa_lock; 28536b5c08fSOded Gabbay unsigned int gtt_sa_chunk_size; 28636b5c08fSOded Gabbay unsigned int gtt_sa_num_of_chunks; 28736b5c08fSOded Gabbay 2882249d558SAndrew Lewycky /* Interrupts */ 28904ad47bdSAndres Rodriguez struct kfifo ih_fifo; 29048e876a2SAndres Rodriguez struct workqueue_struct *ih_wq; 2912249d558SAndrew Lewycky struct work_struct interrupt_work; 2922249d558SAndrew Lewycky spinlock_t interrupt_lock; 2932249d558SAndrew Lewycky 294ed6e6a34SBen Goz /* QCM Device instance */ 295ed6e6a34SBen Goz struct device_queue_manager *dqm; 2964a488a7aSOded Gabbay 297ed6e6a34SBen Goz bool init_complete; 2982249d558SAndrew Lewycky /* 2992249d558SAndrew Lewycky * Interrupts of interest to KFD are copied 3002249d558SAndrew Lewycky * from the HW ring into a SW ring. 3012249d558SAndrew Lewycky */ 3022249d558SAndrew Lewycky bool interrupts_active; 303fbeb661bSYair Shachar 3045ade6c9cSFelix Kuehling /* Firmware versions */ 3055ade6c9cSFelix Kuehling uint16_t mec_fw_version; 30629633d0eSJoseph Greathouse uint16_t mec2_fw_version; 3075ade6c9cSFelix Kuehling uint16_t sdma_fw_version; 3085ade6c9cSFelix Kuehling 309a99c6d4fSFelix Kuehling /* Maximum process number mapped to HW scheduler */ 310a99c6d4fSFelix Kuehling unsigned int max_proc_per_quantum; 311a99c6d4fSFelix Kuehling 312373d7080SFelix Kuehling /* CWSR */ 313373d7080SFelix Kuehling bool cwsr_enabled; 314373d7080SFelix Kuehling const void *cwsr_isa; 315373d7080SFelix Kuehling unsigned int cwsr_isa_size; 3160c1690e3SShaoyun Liu 3170c1690e3SShaoyun Liu /* xGMI */ 3180c1690e3SShaoyun Liu uint64_t hive_id; 3190c663695SDivya Shikre 320d35f00d8SEric Huang bool pci_atomic_requested; 3219b54d201SEric Huang 3226127896fSHuang Rui /* Use IOMMU v2 flag */ 3236127896fSHuang Rui bool use_iommu_v2; 3246127896fSHuang Rui 3259b54d201SEric Huang /* SRAM ECC flag */ 3269b54d201SEric Huang atomic_t sram_ecc_flag; 327f756e631SHarish Kasiviswanathan 328f756e631SHarish Kasiviswanathan /* Compute Profile ref. count */ 329f756e631SHarish Kasiviswanathan atomic_t compute_profile; 330e09d4fc8SOak Zeng 331a4497974SRajneesh Bhardwaj /* Global GWS resource shared between processes */ 332e09d4fc8SOak Zeng void *gws; 333938a0650SAmber Lin 334938a0650SAmber Lin /* Clients watching SMI events */ 335938a0650SAmber Lin struct list_head smi_clients; 336938a0650SAmber Lin spinlock_t smi_lock; 33755977744SMukul Joshi 33855977744SMukul Joshi uint32_t reset_seq_num; 33959d7115dSMukul Joshi 34059d7115dSMukul Joshi struct ida doorbell_ida; 34159d7115dSMukul Joshi unsigned int max_doorbell_slices; 3429b498efaSAlex Deucher 3439b498efaSAlex Deucher int noretry; 344814ab993SPhilip Yang 345814ab993SPhilip Yang /* HMM page migration MEMORY_DEVICE_PRIVATE mapping */ 346814ab993SPhilip Yang struct dev_pagemap pgmap; 3474a488a7aSOded Gabbay }; 3484a488a7aSOded Gabbay 34919f6d2a6SOded Gabbay enum kfd_mempool { 35019f6d2a6SOded Gabbay KFD_MEMPOOL_SYSTEM_CACHEABLE = 1, 35119f6d2a6SOded Gabbay KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2, 35219f6d2a6SOded Gabbay KFD_MEMPOOL_FRAMEBUFFER = 3, 35319f6d2a6SOded Gabbay }; 35419f6d2a6SOded Gabbay 3554a488a7aSOded Gabbay /* Character device interface */ 3564a488a7aSOded Gabbay int kfd_chardev_init(void); 3574a488a7aSOded Gabbay void kfd_chardev_exit(void); 3584a488a7aSOded Gabbay 359241f24f8SBen Goz /** 360a4497974SRajneesh Bhardwaj * enum kfd_unmap_queues_filter - Enum for queue filters. 361241f24f8SBen Goz * 3627da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the 363241f24f8SBen Goz * running queues list. 364241f24f8SBen Goz * 365d2cb0b21SJonathan Kim * @KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: Preempts all non-static queues 366d2cb0b21SJonathan Kim * in the run list. 367d2cb0b21SJonathan Kim * 3687da2bcf8SYong Zhao * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to 369241f24f8SBen Goz * specific process. 370241f24f8SBen Goz * 371241f24f8SBen Goz */ 3727da2bcf8SYong Zhao enum kfd_unmap_queues_filter { 373d2cb0b21SJonathan Kim KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES = 1, 374d2cb0b21SJonathan Kim KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES = 2, 375d2cb0b21SJonathan Kim KFD_UNMAP_QUEUES_FILTER_BY_PASID = 3 376241f24f8SBen Goz }; 37719f6d2a6SOded Gabbay 378ed8aab45SBen Goz /** 379a4497974SRajneesh Bhardwaj * enum kfd_queue_type - Enum for various queue types. 380ed8aab45SBen Goz * 381ed8aab45SBen Goz * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type. 382ed8aab45SBen Goz * 383a4497974SRajneesh Bhardwaj * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type. 384ed8aab45SBen Goz * 385ed8aab45SBen Goz * @KFD_QUEUE_TYPE_HIQ: HIQ queue type. 386ed8aab45SBen Goz * 387ed8aab45SBen Goz * @KFD_QUEUE_TYPE_DIQ: DIQ queue type. 388a4497974SRajneesh Bhardwaj * 389a4497974SRajneesh Bhardwaj * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface. 390ed8aab45SBen Goz */ 391ed8aab45SBen Goz enum kfd_queue_type { 392ed8aab45SBen Goz KFD_QUEUE_TYPE_COMPUTE, 393ed8aab45SBen Goz KFD_QUEUE_TYPE_SDMA, 394ed8aab45SBen Goz KFD_QUEUE_TYPE_HIQ, 3951b4670f6SOak Zeng KFD_QUEUE_TYPE_DIQ, 3961b4670f6SOak Zeng KFD_QUEUE_TYPE_SDMA_XGMI 397ed8aab45SBen Goz }; 398ed8aab45SBen Goz 3996e99df57SBen Goz enum kfd_queue_format { 4006e99df57SBen Goz KFD_QUEUE_FORMAT_PM4, 4016e99df57SBen Goz KFD_QUEUE_FORMAT_AQL 4026e99df57SBen Goz }; 4036e99df57SBen Goz 4040ccbc7cdSOak Zeng enum KFD_QUEUE_PRIORITY { 4050ccbc7cdSOak Zeng KFD_QUEUE_PRIORITY_MINIMUM = 0, 4060ccbc7cdSOak Zeng KFD_QUEUE_PRIORITY_MAXIMUM = 15 4070ccbc7cdSOak Zeng }; 4080ccbc7cdSOak Zeng 409ed8aab45SBen Goz /** 410ed8aab45SBen Goz * struct queue_properties 411ed8aab45SBen Goz * 412ed8aab45SBen Goz * @type: The queue type. 413ed8aab45SBen Goz * 414ed8aab45SBen Goz * @queue_id: Queue identifier. 415ed8aab45SBen Goz * 416ed8aab45SBen Goz * @queue_address: Queue ring buffer address. 417ed8aab45SBen Goz * 418ed8aab45SBen Goz * @queue_size: Queue ring buffer size. 419ed8aab45SBen Goz * 420ed8aab45SBen Goz * @priority: Defines the queue priority relative to other queues in the 421ed8aab45SBen Goz * process. 422ed8aab45SBen Goz * This is just an indication and HW scheduling may override the priority as 423ed8aab45SBen Goz * necessary while keeping the relative prioritization. 424ed8aab45SBen Goz * the priority granularity is from 0 to f which f is the highest priority. 425ed8aab45SBen Goz * currently all queues are initialized with the highest priority. 426ed8aab45SBen Goz * 427ed8aab45SBen Goz * @queue_percent: This field is partially implemented and currently a zero in 428ed8aab45SBen Goz * this field defines that the queue is non active. 429ed8aab45SBen Goz * 430ed8aab45SBen Goz * @read_ptr: User space address which points to the number of dwords the 431ed8aab45SBen Goz * cp read from the ring buffer. This field updates automatically by the H/W. 432ed8aab45SBen Goz * 433ed8aab45SBen Goz * @write_ptr: Defines the number of dwords written to the ring buffer. 434ed8aab45SBen Goz * 435a4497974SRajneesh Bhardwaj * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring 436a4497974SRajneesh Bhardwaj * buffer. This field should be similar to write_ptr and the user should 437a4497974SRajneesh Bhardwaj * update this field after updating the write_ptr. 438ed8aab45SBen Goz * 439ed8aab45SBen Goz * @doorbell_off: The doorbell offset in the doorbell pci-bar. 440ed8aab45SBen Goz * 4418eabaf54SKent Russell * @is_interop: Defines if this is a interop queue. Interop queue means that 4428eabaf54SKent Russell * the queue can access both graphics and compute resources. 443ed8aab45SBen Goz * 44426103436SFelix Kuehling * @is_evicted: Defines if the queue is evicted. Only active queues 44526103436SFelix Kuehling * are evicted, rendering them inactive. 44626103436SFelix Kuehling * 44726103436SFelix Kuehling * @is_active: Defines if the queue is active or not. @is_active and 44826103436SFelix Kuehling * @is_evicted are protected by the DQM lock. 449ed8aab45SBen Goz * 450b8020b03SJoseph Greathouse * @is_gws: Defines if the queue has been updated to be GWS-capable or not. 451b8020b03SJoseph Greathouse * @is_gws should be protected by the DQM lock, since changing it can yield the 452b8020b03SJoseph Greathouse * possibility of updating DQM state on number of GWS queues. 453b8020b03SJoseph Greathouse * 454ed8aab45SBen Goz * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid 455ed8aab45SBen Goz * of the queue. 456ed8aab45SBen Goz * 457ed8aab45SBen Goz * This structure represents the queue properties for each queue no matter if 458ed8aab45SBen Goz * it's user mode or kernel mode queue. 459ed8aab45SBen Goz * 460ed8aab45SBen Goz */ 4618668dfc3SDavid Yat Sin 462ed8aab45SBen Goz struct queue_properties { 463ed8aab45SBen Goz enum kfd_queue_type type; 4646e99df57SBen Goz enum kfd_queue_format format; 465ed8aab45SBen Goz unsigned int queue_id; 466ed8aab45SBen Goz uint64_t queue_address; 467ed8aab45SBen Goz uint64_t queue_size; 468ed8aab45SBen Goz uint32_t priority; 469ed8aab45SBen Goz uint32_t queue_percent; 470ed8aab45SBen Goz uint32_t *read_ptr; 471ed8aab45SBen Goz uint32_t *write_ptr; 472ada2b29cSFelix Kuehling void __iomem *doorbell_ptr; 473ed8aab45SBen Goz uint32_t doorbell_off; 474ed8aab45SBen Goz bool is_interop; 47526103436SFelix Kuehling bool is_evicted; 476ed8aab45SBen Goz bool is_active; 477b8020b03SJoseph Greathouse bool is_gws; 478ed8aab45SBen Goz /* Not relevant for user mode queues in cp scheduling */ 479ed8aab45SBen Goz unsigned int vmid; 48077669eb8SBen Goz /* Relevant only for sdma queues*/ 48177669eb8SBen Goz uint32_t sdma_engine_id; 48277669eb8SBen Goz uint32_t sdma_queue_id; 48377669eb8SBen Goz uint32_t sdma_vm_addr; 484ff3d04a1SBen Goz /* Relevant only for VI */ 485ff3d04a1SBen Goz uint64_t eop_ring_buffer_address; 486ff3d04a1SBen Goz uint32_t eop_ring_buffer_size; 487ff3d04a1SBen Goz uint64_t ctx_save_restore_area_address; 488ff3d04a1SBen Goz uint32_t ctx_save_restore_area_size; 489373d7080SFelix Kuehling uint32_t ctl_stack_size; 490373d7080SFelix Kuehling uint64_t tba_addr; 491373d7080SFelix Kuehling uint64_t tma_addr; 492ed8aab45SBen Goz }; 493ed8aab45SBen Goz 494bb2d2128SFelix Kuehling #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 && \ 495bb2d2128SFelix Kuehling (q).queue_address != 0 && \ 496bb2d2128SFelix Kuehling (q).queue_percent > 0 && \ 497bb2d2128SFelix Kuehling !(q).is_evicted) 498bb2d2128SFelix Kuehling 4997c695a2cSLang Yu enum mqd_update_flag { 5007c695a2cSLang Yu UPDATE_FLAG_CU_MASK = 0, 5017c695a2cSLang Yu }; 5027c695a2cSLang Yu 5037c695a2cSLang Yu struct mqd_update_info { 5047c695a2cSLang Yu union { 5057c695a2cSLang Yu struct { 5067c695a2cSLang Yu uint32_t count; /* Must be a multiple of 32 */ 5077c695a2cSLang Yu uint32_t *ptr; 5087c695a2cSLang Yu } cu_mask; 5097c695a2cSLang Yu }; 5107c695a2cSLang Yu enum mqd_update_flag update_flag; 5117c695a2cSLang Yu }; 512c6e559ebSLang Yu 513ed8aab45SBen Goz /** 514ed8aab45SBen Goz * struct queue 515ed8aab45SBen Goz * 516ed8aab45SBen Goz * @list: Queue linked list. 517ed8aab45SBen Goz * 518a4497974SRajneesh Bhardwaj * @mqd: The queue MQD (memory queue descriptor). 519ed8aab45SBen Goz * 520ed8aab45SBen Goz * @mqd_mem_obj: The MQD local gpu memory object. 521ed8aab45SBen Goz * 522ed8aab45SBen Goz * @gart_mqd_addr: The MQD gart mc address. 523ed8aab45SBen Goz * 524ed8aab45SBen Goz * @properties: The queue properties. 525ed8aab45SBen Goz * 526ed8aab45SBen Goz * @mec: Used only in no cp scheduling mode and identifies to micro engine id 527a4497974SRajneesh Bhardwaj * that the queue should be executed on. 528ed8aab45SBen Goz * 5298eabaf54SKent Russell * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe 5308eabaf54SKent Russell * id. 531ed8aab45SBen Goz * 532ed8aab45SBen Goz * @queue: Used only in no cp scheduliong mode and identifies the queue's slot. 533ed8aab45SBen Goz * 534ed8aab45SBen Goz * @process: The kfd process that created this queue. 535ed8aab45SBen Goz * 536ed8aab45SBen Goz * @device: The kfd device that created this queue. 537ed8aab45SBen Goz * 538eb82da1dSOak Zeng * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL 539eb82da1dSOak Zeng * otherwise. 540eb82da1dSOak Zeng * 541ed8aab45SBen Goz * This structure represents user mode compute queues. 542ed8aab45SBen Goz * It contains all the necessary data to handle such queues. 543ed8aab45SBen Goz * 544ed8aab45SBen Goz */ 545ed8aab45SBen Goz 546ed8aab45SBen Goz struct queue { 547ed8aab45SBen Goz struct list_head list; 548ed8aab45SBen Goz void *mqd; 549ed8aab45SBen Goz struct kfd_mem_obj *mqd_mem_obj; 550ed8aab45SBen Goz uint64_t gart_mqd_addr; 551ed8aab45SBen Goz struct queue_properties properties; 552ed8aab45SBen Goz 553ed8aab45SBen Goz uint32_t mec; 554ed8aab45SBen Goz uint32_t pipe; 555ed8aab45SBen Goz uint32_t queue; 556ed8aab45SBen Goz 55777669eb8SBen Goz unsigned int sdma_id; 558ef568db7SFelix Kuehling unsigned int doorbell_id; 55977669eb8SBen Goz 560ed8aab45SBen Goz struct kfd_process *process; 561ed8aab45SBen Goz struct kfd_dev *device; 562eb82da1dSOak Zeng void *gws; 5636d220a7eSAmber Lin 5646d220a7eSAmber Lin /* procfs */ 5656d220a7eSAmber Lin struct kobject kobj; 566ed8aab45SBen Goz }; 567ed8aab45SBen Goz 5686e99df57SBen Goz enum KFD_MQD_TYPE { 569d7c0b047SYong Zhao KFD_MQD_TYPE_HIQ = 0, /* for hiq */ 57085d258f9SBen Goz KFD_MQD_TYPE_CP, /* for cp queues and diq */ 57185d258f9SBen Goz KFD_MQD_TYPE_SDMA, /* for sdma queues */ 57259f650a0SOak Zeng KFD_MQD_TYPE_DIQ, /* for diq */ 5736e99df57SBen Goz KFD_MQD_TYPE_MAX 5746e99df57SBen Goz }; 5756e99df57SBen Goz 5760ccbc7cdSOak Zeng enum KFD_PIPE_PRIORITY { 5770ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_LOW = 0, 5780ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_MEDIUM, 5790ccbc7cdSOak Zeng KFD_PIPE_PRIORITY_CS_HIGH 5800ccbc7cdSOak Zeng }; 5810ccbc7cdSOak Zeng 582241f24f8SBen Goz struct scheduling_resources { 583241f24f8SBen Goz unsigned int vmid_mask; 584241f24f8SBen Goz enum kfd_queue_type type; 585241f24f8SBen Goz uint64_t queue_mask; 586241f24f8SBen Goz uint64_t gws_mask; 587241f24f8SBen Goz uint32_t oac_mask; 588241f24f8SBen Goz uint32_t gds_heap_base; 589241f24f8SBen Goz uint32_t gds_heap_size; 590241f24f8SBen Goz }; 591241f24f8SBen Goz 592241f24f8SBen Goz struct process_queue_manager { 593241f24f8SBen Goz /* data */ 594241f24f8SBen Goz struct kfd_process *process; 595241f24f8SBen Goz struct list_head queues; 596241f24f8SBen Goz unsigned long *queue_slot_bitmap; 597241f24f8SBen Goz }; 598241f24f8SBen Goz 599241f24f8SBen Goz struct qcm_process_device { 600241f24f8SBen Goz /* The Device Queue Manager that owns this data */ 601241f24f8SBen Goz struct device_queue_manager *dqm; 602241f24f8SBen Goz struct process_queue_manager *pqm; 603241f24f8SBen Goz /* Queues list */ 604241f24f8SBen Goz struct list_head queues_list; 605241f24f8SBen Goz struct list_head priv_queue_list; 606241f24f8SBen Goz 607241f24f8SBen Goz unsigned int queue_count; 608241f24f8SBen Goz unsigned int vmid; 609241f24f8SBen Goz bool is_debug; 61026103436SFelix Kuehling unsigned int evicted; /* eviction counter, 0=active */ 6119fd3f1bfSFelix Kuehling 6129fd3f1bfSFelix Kuehling /* This flag tells if we should reset all wavefronts on 6139fd3f1bfSFelix Kuehling * process termination 6149fd3f1bfSFelix Kuehling */ 6159fd3f1bfSFelix Kuehling bool reset_wavefronts; 6169fd3f1bfSFelix Kuehling 617b8020b03SJoseph Greathouse /* This flag tells us if this process has a GWS-capable 618b8020b03SJoseph Greathouse * queue that will be mapped into the runlist. It's 619b8020b03SJoseph Greathouse * possible to request a GWS BO, but not have the queue 620b8020b03SJoseph Greathouse * currently mapped, and this changes how the MAP_PROCESS 621b8020b03SJoseph Greathouse * PM4 packet is configured. 622b8020b03SJoseph Greathouse */ 623b8020b03SJoseph Greathouse bool mapped_gws_queue; 624b8020b03SJoseph Greathouse 625a4497974SRajneesh Bhardwaj /* All the memory management data should be here too */ 626241f24f8SBen Goz uint64_t gds_context_area; 627435e2f97SYong Zhao /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */ 628e715c6d0SShaoyun Liu uint64_t page_table_base; 629241f24f8SBen Goz uint32_t sh_mem_config; 630241f24f8SBen Goz uint32_t sh_mem_bases; 631241f24f8SBen Goz uint32_t sh_mem_ape1_base; 632241f24f8SBen Goz uint32_t sh_mem_ape1_limit; 633241f24f8SBen Goz uint32_t gds_size; 634241f24f8SBen Goz uint32_t num_gws; 635241f24f8SBen Goz uint32_t num_oac; 6366a1c9510SMoses Reuben uint32_t sh_hidden_private_base; 637373d7080SFelix Kuehling 638373d7080SFelix Kuehling /* CWSR memory */ 63968df0f19SLang Yu struct kgd_mem *cwsr_mem; 640373d7080SFelix Kuehling void *cwsr_kaddr; 641d01994c2SFelix Kuehling uint64_t cwsr_base; 642373d7080SFelix Kuehling uint64_t tba_addr; 643373d7080SFelix Kuehling uint64_t tma_addr; 644d01994c2SFelix Kuehling 645d01994c2SFelix Kuehling /* IB memory */ 64668df0f19SLang Yu struct kgd_mem *ib_mem; 647d01994c2SFelix Kuehling uint64_t ib_base; 648552764b6SFelix Kuehling void *ib_kaddr; 649ef568db7SFelix Kuehling 650ef568db7SFelix Kuehling /* doorbell resources per process per device */ 651ef568db7SFelix Kuehling unsigned long *doorbell_bitmap; 652241f24f8SBen Goz }; 653241f24f8SBen Goz 65426103436SFelix Kuehling /* KFD Memory Eviction */ 65526103436SFelix Kuehling 65626103436SFelix Kuehling /* Approx. wait time before attempting to restore evicted BOs */ 65726103436SFelix Kuehling #define PROCESS_RESTORE_TIME_MS 100 65826103436SFelix Kuehling /* Approx. back off time if restore fails due to lack of memory */ 65926103436SFelix Kuehling #define PROCESS_BACK_OFF_TIME_MS 100 66026103436SFelix Kuehling /* Approx. time before evicting the process again */ 66126103436SFelix Kuehling #define PROCESS_ACTIVE_TIME_MS 10 66226103436SFelix Kuehling 6635ec7e028SFelix Kuehling /* 8 byte handle containing GPU ID in the most significant 4 bytes and 6645ec7e028SFelix Kuehling * idr_handle in the least significant 4 bytes 6655ec7e028SFelix Kuehling */ 6665ec7e028SFelix Kuehling #define MAKE_HANDLE(gpu_id, idr_handle) \ 6675ec7e028SFelix Kuehling (((uint64_t)(gpu_id) << 32) + idr_handle) 6685ec7e028SFelix Kuehling #define GET_GPU_ID(handle) (handle >> 32) 6695ec7e028SFelix Kuehling #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF) 6705ec7e028SFelix Kuehling 671733fa1f7SYong Zhao enum kfd_pdd_bound { 672733fa1f7SYong Zhao PDD_UNBOUND = 0, 673733fa1f7SYong Zhao PDD_BOUND, 674733fa1f7SYong Zhao PDD_BOUND_SUSPENDED, 675733fa1f7SYong Zhao }; 676733fa1f7SYong Zhao 6774327bed2SPhilip Cox #define MAX_SYSFS_FILENAME_LEN 15 67832cb59f3SMukul Joshi 67932cb59f3SMukul Joshi /* 68032cb59f3SMukul Joshi * SDMA counter runs at 100MHz frequency. 68132cb59f3SMukul Joshi * We display SDMA activity in microsecond granularity in sysfs. 68232cb59f3SMukul Joshi * As a result, the divisor is 100. 68332cb59f3SMukul Joshi */ 68432cb59f3SMukul Joshi #define SDMA_ACTIVITY_DIVISOR 100 685d4566deeSMukul Joshi 68619f6d2a6SOded Gabbay /* Data that is per-process-per device. */ 68719f6d2a6SOded Gabbay struct kfd_process_device { 68819f6d2a6SOded Gabbay /* The device that owns this data. */ 68919f6d2a6SOded Gabbay struct kfd_dev *dev; 69019f6d2a6SOded Gabbay 6919fd3f1bfSFelix Kuehling /* The process that owns this kfd_process_device. */ 6929fd3f1bfSFelix Kuehling struct kfd_process *process; 69319f6d2a6SOded Gabbay 69445102048SBen Goz /* per-process-per device QCM data structure */ 69545102048SBen Goz struct qcm_process_device qpd; 69645102048SBen Goz 69719f6d2a6SOded Gabbay /*Apertures*/ 69819f6d2a6SOded Gabbay uint64_t lds_base; 69919f6d2a6SOded Gabbay uint64_t lds_limit; 70019f6d2a6SOded Gabbay uint64_t gpuvm_base; 70119f6d2a6SOded Gabbay uint64_t gpuvm_limit; 70219f6d2a6SOded Gabbay uint64_t scratch_base; 70319f6d2a6SOded Gabbay uint64_t scratch_limit; 70419f6d2a6SOded Gabbay 705403575c4SFelix Kuehling /* VM context for GPUVM allocations */ 706b84394e2SFelix Kuehling struct file *drm_file; 707b40a6ab2SFelix Kuehling void *drm_priv; 7088fde0248SPhilip Yang atomic64_t tlb_seq; 709403575c4SFelix Kuehling 71052b29d73SFelix Kuehling /* GPUVM allocations storage */ 71152b29d73SFelix Kuehling struct idr alloc_idr; 71252b29d73SFelix Kuehling 7139fd3f1bfSFelix Kuehling /* Flag used to tell the pdd has dequeued from the dqm. 7149fd3f1bfSFelix Kuehling * This is used to prevent dev->dqm->ops.process_termination() from 7159fd3f1bfSFelix Kuehling * being called twice when it is already called in IOMMU callback 7169fd3f1bfSFelix Kuehling * function. 717a82918f1SBen Goz */ 7189fd3f1bfSFelix Kuehling bool already_dequeued; 7199593f4d6SRajneesh Bhardwaj bool runtime_inuse; 72064d1c3a4SFelix Kuehling 72164d1c3a4SFelix Kuehling /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */ 72264d1c3a4SFelix Kuehling enum kfd_pdd_bound bound; 723d4566deeSMukul Joshi 724d4566deeSMukul Joshi /* VRAM usage */ 725d4566deeSMukul Joshi uint64_t vram_usage; 726d4566deeSMukul Joshi struct attribute attr_vram; 72732cb59f3SMukul Joshi char vram_filename[MAX_SYSFS_FILENAME_LEN]; 72832cb59f3SMukul Joshi 72932cb59f3SMukul Joshi /* SDMA activity tracking */ 73032cb59f3SMukul Joshi uint64_t sdma_past_activity_counter; 73132cb59f3SMukul Joshi struct attribute attr_sdma; 73232cb59f3SMukul Joshi char sdma_filename[MAX_SYSFS_FILENAME_LEN]; 7334327bed2SPhilip Cox 7344327bed2SPhilip Cox /* Eviction activity tracking */ 7354327bed2SPhilip Cox uint64_t last_evict_timestamp; 7364327bed2SPhilip Cox atomic64_t evict_duration_counter; 7374327bed2SPhilip Cox struct attribute attr_evict; 7384327bed2SPhilip Cox 7394327bed2SPhilip Cox struct kobject *kobj_stats; 74059d7115dSMukul Joshi unsigned int doorbell_index; 741f2fa07b3SRamesh Errabolu 742f2fa07b3SRamesh Errabolu /* 743f2fa07b3SRamesh Errabolu * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process 744f2fa07b3SRamesh Errabolu * that is associated with device encoded by "this" struct instance. The 745f2fa07b3SRamesh Errabolu * value reflects CU usage by all of the waves launched by this process 746f2fa07b3SRamesh Errabolu * on this device. A very important property of occupancy parameter is 747f2fa07b3SRamesh Errabolu * that its value is a snapshot of current use. 748f2fa07b3SRamesh Errabolu * 749f2fa07b3SRamesh Errabolu * Following is to be noted regarding how this parameter is reported: 750f2fa07b3SRamesh Errabolu * 751f2fa07b3SRamesh Errabolu * The number of waves that a CU can launch is limited by couple of 752f2fa07b3SRamesh Errabolu * parameters. These are encoded by struct amdgpu_cu_info instance 753f2fa07b3SRamesh Errabolu * that is part of every device definition. For GFX9 devices this 754f2fa07b3SRamesh Errabolu * translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves 755f2fa07b3SRamesh Errabolu * do not use scratch memory and 32 waves (max_scratch_slots_per_cu) 756f2fa07b3SRamesh Errabolu * when they do use scratch memory. This could change for future 757f2fa07b3SRamesh Errabolu * devices and therefore this example should be considered as a guide. 758f2fa07b3SRamesh Errabolu * 759f2fa07b3SRamesh Errabolu * All CU's of a device are available for the process. This may not be true 760f2fa07b3SRamesh Errabolu * under certain conditions - e.g. CU masking. 761f2fa07b3SRamesh Errabolu * 762f2fa07b3SRamesh Errabolu * Finally number of CU's that are occupied by a process is affected by both 763f2fa07b3SRamesh Errabolu * number of CU's a device has along with number of other competing processes 764f2fa07b3SRamesh Errabolu */ 765f2fa07b3SRamesh Errabolu struct attribute attr_cu_occupancy; 766751580b3SPhilip Yang 767751580b3SPhilip Yang /* sysfs counters for GPU retry fault and page migration tracking */ 768751580b3SPhilip Yang struct kobject *kobj_counters; 769751580b3SPhilip Yang struct attribute attr_faults; 770751580b3SPhilip Yang struct attribute attr_page_in; 771751580b3SPhilip Yang struct attribute attr_page_out; 772751580b3SPhilip Yang uint64_t faults; 773751580b3SPhilip Yang uint64_t page_in; 774751580b3SPhilip Yang uint64_t page_out; 775bef153b7SDavid Yat Sin /* 776bef153b7SDavid Yat Sin * If this process has been checkpointed before, then the user 777bef153b7SDavid Yat Sin * application will use the original gpu_id on the 778bef153b7SDavid Yat Sin * checkpointed node to refer to this device. 779bef153b7SDavid Yat Sin */ 780bef153b7SDavid Yat Sin uint32_t user_gpu_id; 78119f6d2a6SOded Gabbay }; 78219f6d2a6SOded Gabbay 78352a5fdceSAlexey Skidanov #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd) 78452a5fdceSAlexey Skidanov 78542de677fSPhilip Yang struct svm_range_list { 78642de677fSPhilip Yang struct mutex lock; 78742de677fSPhilip Yang struct rb_root_cached objects; 78842de677fSPhilip Yang struct list_head list; 7894683cfecSPhilip Yang struct work_struct deferred_list_work; 7904683cfecSPhilip Yang struct list_head deferred_range_list; 791c2db32ceSRajneesh Bhardwaj struct list_head criu_svm_metadata_list; 7924683cfecSPhilip Yang spinlock_t deferred_list_lock; 7938a7c184aSFelix Kuehling atomic_t evicted_ranges; 7942e447728SPhilip Yang atomic_t drain_pagefaults; 7958a7c184aSFelix Kuehling struct delayed_work restore_work; 7965a75ea56SFelix Kuehling DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE); 797a6283010SAlex Sierra struct task_struct *faulting_task; 79842de677fSPhilip Yang }; 79942de677fSPhilip Yang 8004a488a7aSOded Gabbay /* Process data */ 8014a488a7aSOded Gabbay struct kfd_process { 80219f6d2a6SOded Gabbay /* 80319f6d2a6SOded Gabbay * kfd_process are stored in an mm_struct*->kfd_process* 80419f6d2a6SOded Gabbay * hash table (kfd_processes in kfd_process.c) 80519f6d2a6SOded Gabbay */ 80619f6d2a6SOded Gabbay struct hlist_node kfd_processes; 80719f6d2a6SOded Gabbay 8089b56bb11SFelix Kuehling /* 8099b56bb11SFelix Kuehling * Opaque pointer to mm_struct. We don't hold a reference to 8109b56bb11SFelix Kuehling * it so it should never be dereferenced from here. This is 8119b56bb11SFelix Kuehling * only used for looking up processes by their mm. 8129b56bb11SFelix Kuehling */ 8139b56bb11SFelix Kuehling void *mm; 81419f6d2a6SOded Gabbay 8155ce10687SFelix Kuehling struct kref ref; 8165ce10687SFelix Kuehling struct work_struct release_work; 8175ce10687SFelix Kuehling 81819f6d2a6SOded Gabbay struct mutex mutex; 81919f6d2a6SOded Gabbay 82019f6d2a6SOded Gabbay /* 82119f6d2a6SOded Gabbay * In any process, the thread that started main() is the lead 82219f6d2a6SOded Gabbay * thread and outlives the rest. 82319f6d2a6SOded Gabbay * It is here because amd_iommu_bind_pasid wants a task_struct. 824894a8293SFelix Kuehling * It can also be used for safely getting a reference to the 825894a8293SFelix Kuehling * mm_struct of the process. 82619f6d2a6SOded Gabbay */ 82719f6d2a6SOded Gabbay struct task_struct *lead_thread; 82819f6d2a6SOded Gabbay 82919f6d2a6SOded Gabbay /* We want to receive a notification when the mm_struct is destroyed */ 83019f6d2a6SOded Gabbay struct mmu_notifier mmu_notifier; 83119f6d2a6SOded Gabbay 832c7b6bac9SFenghua Yu u32 pasid; 83319f6d2a6SOded Gabbay 83419f6d2a6SOded Gabbay /* 8356ae27841SAlex Sierra * Array of kfd_process_device pointers, 83619f6d2a6SOded Gabbay * one for each device the process is using. 83719f6d2a6SOded Gabbay */ 8386ae27841SAlex Sierra struct kfd_process_device *pdds[MAX_GPU_INSTANCE]; 8396ae27841SAlex Sierra uint32_t n_pdds; 84019f6d2a6SOded Gabbay 84145102048SBen Goz struct process_queue_manager pqm; 84245102048SBen Goz 84319f6d2a6SOded Gabbay /*Is the user space process 32 bit?*/ 84419f6d2a6SOded Gabbay bool is_32bit_user_mode; 845f3a39818SAndrew Lewycky 846f3a39818SAndrew Lewycky /* Event-related data */ 847f3a39818SAndrew Lewycky struct mutex event_mutex; 848482f0777SFelix Kuehling /* Event ID allocator and lookup */ 849482f0777SFelix Kuehling struct idr event_idr; 85050cb7dd9SFelix Kuehling /* Event page */ 85168df0f19SLang Yu u64 signal_handle; 85250cb7dd9SFelix Kuehling struct kfd_signal_page *signal_page; 853b9a5d0a5SFelix Kuehling size_t signal_mapped_size; 854f3a39818SAndrew Lewycky size_t signal_event_count; 855c986169fSFelix Kuehling bool signal_event_limit_reached; 856403575c4SFelix Kuehling 857403575c4SFelix Kuehling /* Information used for memory eviction */ 858403575c4SFelix Kuehling void *kgd_process_info; 859403575c4SFelix Kuehling /* Eviction fence that is attached to all the BOs of this process. The 860403575c4SFelix Kuehling * fence will be triggered during eviction and new one will be created 861403575c4SFelix Kuehling * during restore 862403575c4SFelix Kuehling */ 863403575c4SFelix Kuehling struct dma_fence *ef; 86426103436SFelix Kuehling 86526103436SFelix Kuehling /* Work items for evicting and restoring BOs */ 86626103436SFelix Kuehling struct delayed_work eviction_work; 86726103436SFelix Kuehling struct delayed_work restore_work; 86826103436SFelix Kuehling /* seqno of the last scheduled eviction */ 86926103436SFelix Kuehling unsigned int last_eviction_seqno; 87026103436SFelix Kuehling /* Approx. the last timestamp (in jiffies) when the process was 87126103436SFelix Kuehling * restored after an eviction 87226103436SFelix Kuehling */ 87326103436SFelix Kuehling unsigned long last_restore_timestamp; 874de9f26bbSKent Russell 875de9f26bbSKent Russell /* Kobj for our procfs */ 876de9f26bbSKent Russell struct kobject *kobj; 8776d220a7eSAmber Lin struct kobject *kobj_queues; 878de9f26bbSKent Russell struct attribute attr_pasid; 87940ce74d1SPhilip Yang 88042de677fSPhilip Yang /* shared virtual memory registered by this process */ 88142de677fSPhilip Yang struct svm_range_list svms; 882063e33c5SAlex Sierra 883063e33c5SAlex Sierra bool xnack_enabled; 884b6485bedSTao Zhou 885b6485bedSTao Zhou atomic_t poison; 886cd9f7910SDavid Yat Sin /* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */ 887cd9f7910SDavid Yat Sin bool queues_paused; 8884a488a7aSOded Gabbay }; 8894a488a7aSOded Gabbay 89064d1c3a4SFelix Kuehling #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */ 89164d1c3a4SFelix Kuehling extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE); 89264d1c3a4SFelix Kuehling extern struct srcu_struct kfd_processes_srcu; 89364d1c3a4SFelix Kuehling 89476baee6cSOded Gabbay /** 895a4497974SRajneesh Bhardwaj * typedef amdkfd_ioctl_t - typedef for ioctl function pointer. 89676baee6cSOded Gabbay * 897a4497974SRajneesh Bhardwaj * @filep: pointer to file structure. 898a4497974SRajneesh Bhardwaj * @p: amdkfd process pointer. 899a4497974SRajneesh Bhardwaj * @data: pointer to arg that was copied from user. 900a4497974SRajneesh Bhardwaj * 901a4497974SRajneesh Bhardwaj * Return: returns ioctl completion code. 90276baee6cSOded Gabbay */ 90376baee6cSOded Gabbay typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p, 90476baee6cSOded Gabbay void *data); 90576baee6cSOded Gabbay 90676baee6cSOded Gabbay struct amdkfd_ioctl_desc { 90776baee6cSOded Gabbay unsigned int cmd; 90876baee6cSOded Gabbay int flags; 90976baee6cSOded Gabbay amdkfd_ioctl_t *func; 91076baee6cSOded Gabbay unsigned int cmd_drv; 91176baee6cSOded Gabbay const char *name; 91276baee6cSOded Gabbay }; 91367f7cf9fSshaoyunl bool kfd_dev_is_large_bar(struct kfd_dev *dev); 91476baee6cSOded Gabbay 9151679ae8fSFelix Kuehling int kfd_process_create_wq(void); 91619f6d2a6SOded Gabbay void kfd_process_destroy_wq(void); 917373d7080SFelix Kuehling struct kfd_process *kfd_create_process(struct file *filep); 9182243f493SRajneesh Bhardwaj struct kfd_process *kfd_get_process(const struct task_struct *task); 919c7b6bac9SFenghua Yu struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid); 92026103436SFelix Kuehling struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm); 9212aeb742bSAlex Sierra 9222aeb742bSAlex Sierra int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id); 92356c5977eSGraham Sider int kfd_process_gpuid_from_adev(struct kfd_process *p, 924cda0f85bSFelix Kuehling struct amdgpu_device *adev, uint32_t *gpuid, 925cda0f85bSFelix Kuehling uint32_t *gpuidx); 9262aeb742bSAlex Sierra static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p, 9272aeb742bSAlex Sierra uint32_t gpuidx, uint32_t *gpuid) { 9282aeb742bSAlex Sierra return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL; 9292aeb742bSAlex Sierra } 9302aeb742bSAlex Sierra static inline struct kfd_process_device *kfd_process_device_from_gpuidx( 9312aeb742bSAlex Sierra struct kfd_process *p, uint32_t gpuidx) { 9322aeb742bSAlex Sierra return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL; 9332aeb742bSAlex Sierra } 9342aeb742bSAlex Sierra 935abb208a8SFelix Kuehling void kfd_unref_process(struct kfd_process *p); 9366b95e797SFelix Kuehling int kfd_process_evict_queues(struct kfd_process *p); 9376b95e797SFelix Kuehling int kfd_process_restore_queues(struct kfd_process *p); 93826103436SFelix Kuehling void kfd_suspend_all_processes(void); 93926103436SFelix Kuehling int kfd_resume_all_processes(void); 94019f6d2a6SOded Gabbay 941bef153b7SDavid Yat Sin struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *process, 942bef153b7SDavid Yat Sin uint32_t gpu_id); 943bef153b7SDavid Yat Sin 944bef153b7SDavid Yat Sin int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id); 945bef153b7SDavid Yat Sin 946b84394e2SFelix Kuehling int kfd_process_device_init_vm(struct kfd_process_device *pdd, 947b84394e2SFelix Kuehling struct file *drm_file); 94864c7f8cfSBen Goz struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev, 94964c7f8cfSBen Goz struct kfd_process *p); 95019f6d2a6SOded Gabbay struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev, 951093c7d8cSAlexey Skidanov struct kfd_process *p); 952093c7d8cSAlexey Skidanov struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev, 953093c7d8cSAlexey Skidanov struct kfd_process *p); 95419f6d2a6SOded Gabbay 955063e33c5SAlex Sierra bool kfd_process_xnack_mode(struct kfd_process *p, bool supported); 956063e33c5SAlex Sierra 957df03ef93SHarish Kasiviswanathan int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process, 958373d7080SFelix Kuehling struct vm_area_struct *vma); 959373d7080SFelix Kuehling 96052b29d73SFelix Kuehling /* KFD process API for creating and translating handles */ 96152b29d73SFelix Kuehling int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd, 96252b29d73SFelix Kuehling void *mem); 96352b29d73SFelix Kuehling void *kfd_process_device_translate_handle(struct kfd_process_device *p, 96452b29d73SFelix Kuehling int handle); 96552b29d73SFelix Kuehling void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd, 96652b29d73SFelix Kuehling int handle); 967011bbb03SRajneesh Bhardwaj struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid); 96852b29d73SFelix Kuehling 96919f6d2a6SOded Gabbay /* PASIDs */ 97019f6d2a6SOded Gabbay int kfd_pasid_init(void); 97119f6d2a6SOded Gabbay void kfd_pasid_exit(void); 97219f6d2a6SOded Gabbay bool kfd_set_pasid_limit(unsigned int new_limit); 97319f6d2a6SOded Gabbay unsigned int kfd_get_pasid_limit(void); 974c7b6bac9SFenghua Yu u32 kfd_pasid_alloc(void); 975c7b6bac9SFenghua Yu void kfd_pasid_free(u32 pasid); 97619f6d2a6SOded Gabbay 97719f6d2a6SOded Gabbay /* Doorbells */ 978ef568db7SFelix Kuehling size_t kfd_doorbell_process_slice(struct kfd_dev *kfd); 979735df2baSFelix Kuehling int kfd_doorbell_init(struct kfd_dev *kfd); 980735df2baSFelix Kuehling void kfd_doorbell_fini(struct kfd_dev *kfd); 981df03ef93SHarish Kasiviswanathan int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 982df03ef93SHarish Kasiviswanathan struct vm_area_struct *vma); 983ada2b29cSFelix Kuehling void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 98419f6d2a6SOded Gabbay unsigned int *doorbell_off); 98519f6d2a6SOded Gabbay void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr); 98619f6d2a6SOded Gabbay u32 read_kernel_doorbell(u32 __iomem *db); 987ada2b29cSFelix Kuehling void write_kernel_doorbell(void __iomem *db, u32 value); 9889d7d0248SFelix Kuehling void write_kernel_doorbell64(void __iomem *db, u64 value); 989339903faSYong Zhao unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, 99059d7115dSMukul Joshi struct kfd_process_device *pdd, 991ef568db7SFelix Kuehling unsigned int doorbell_id); 99259d7115dSMukul Joshi phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd); 99359d7115dSMukul Joshi int kfd_alloc_process_doorbells(struct kfd_dev *kfd, 99459d7115dSMukul Joshi unsigned int *doorbell_index); 99559d7115dSMukul Joshi void kfd_free_process_doorbells(struct kfd_dev *kfd, 99659d7115dSMukul Joshi unsigned int doorbell_index); 9976e81090bSOded Gabbay /* GTT Sub-Allocator */ 9986e81090bSOded Gabbay 9996e81090bSOded Gabbay int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size, 10006e81090bSOded Gabbay struct kfd_mem_obj **mem_obj); 10016e81090bSOded Gabbay 10026e81090bSOded Gabbay int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj); 10036e81090bSOded Gabbay 10044a488a7aSOded Gabbay extern struct device *kfd_device; 10054a488a7aSOded Gabbay 1006de9f26bbSKent Russell /* KFD's procfs */ 1007de9f26bbSKent Russell void kfd_procfs_init(void); 1008de9f26bbSKent Russell void kfd_procfs_shutdown(void); 10096d220a7eSAmber Lin int kfd_procfs_add_queue(struct queue *q); 10106d220a7eSAmber Lin void kfd_procfs_del_queue(struct queue *q); 1011de9f26bbSKent Russell 10125b5c4e40SEvgeny Pinchuk /* Topology */ 10135b5c4e40SEvgeny Pinchuk int kfd_topology_init(void); 10145b5c4e40SEvgeny Pinchuk void kfd_topology_shutdown(void); 10155b5c4e40SEvgeny Pinchuk int kfd_topology_add_device(struct kfd_dev *gpu); 10165b5c4e40SEvgeny Pinchuk int kfd_topology_remove_device(struct kfd_dev *gpu); 10173a87177eSHarish Kasiviswanathan struct kfd_topology_device *kfd_topology_device_by_proximity_domain( 10183a87177eSHarish Kasiviswanathan uint32_t proximity_domain); 101946d18d51SMukul Joshi struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock( 102046d18d51SMukul Joshi uint32_t proximity_domain); 102144d8cc6fSYong Zhao struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id); 10225b5c4e40SEvgeny Pinchuk struct kfd_dev *kfd_device_by_id(uint32_t gpu_id); 10235b5c4e40SEvgeny Pinchuk struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev); 1024574c4183SGraham Sider struct kfd_dev *kfd_device_by_adev(const struct amdgpu_device *adev); 10256d82eb0eSHarish Kasiviswanathan int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev); 1026520b8fb7SFelix Kuehling int kfd_numa_node_to_apic_id(int numa_node_id); 10276127896fSHuang Rui void kfd_double_confirm_iommu_support(struct kfd_dev *gpu); 10285b5c4e40SEvgeny Pinchuk 10294a488a7aSOded Gabbay /* Interrupts */ 10302249d558SAndrew Lewycky int kfd_interrupt_init(struct kfd_dev *dev); 10312249d558SAndrew Lewycky void kfd_interrupt_exit(struct kfd_dev *dev); 10322249d558SAndrew Lewycky bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry); 103358e69886SLan Xiao bool interrupt_is_wanted(struct kfd_dev *dev, 103458e69886SLan Xiao const uint32_t *ih_ring_entry, 103558e69886SLan Xiao uint32_t *patched_ihre, bool *flag); 10364a488a7aSOded Gabbay 103719f6d2a6SOded Gabbay /* amdkfd Apertures */ 103819f6d2a6SOded Gabbay int kfd_init_apertures(struct kfd_process *process); 103919f6d2a6SOded Gabbay 10407c9631afSJay Cornwall void kfd_process_set_trap_handler(struct qcm_process_device *qpd, 10417c9631afSJay Cornwall uint64_t tba_addr, 10427c9631afSJay Cornwall uint64_t tma_addr); 10437c9631afSJay Cornwall 104436988070SRajneesh Bhardwaj /* CRIU */ 104536988070SRajneesh Bhardwaj /* 104636988070SRajneesh Bhardwaj * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private 104736988070SRajneesh Bhardwaj * structures: 104836988070SRajneesh Bhardwaj * kfd_criu_process_priv_data 104936988070SRajneesh Bhardwaj * kfd_criu_device_priv_data 105036988070SRajneesh Bhardwaj * kfd_criu_bo_priv_data 105136988070SRajneesh Bhardwaj * kfd_criu_queue_priv_data 105236988070SRajneesh Bhardwaj * kfd_criu_event_priv_data 105336988070SRajneesh Bhardwaj * kfd_criu_svm_range_priv_data 105436988070SRajneesh Bhardwaj */ 105536988070SRajneesh Bhardwaj 105636988070SRajneesh Bhardwaj #define KFD_CRIU_PRIV_VERSION 1 105736988070SRajneesh Bhardwaj 105836988070SRajneesh Bhardwaj struct kfd_criu_process_priv_data { 105936988070SRajneesh Bhardwaj uint32_t version; 10604717fe3dSRajneesh Bhardwaj uint32_t xnack_mode; 106136988070SRajneesh Bhardwaj }; 106236988070SRajneesh Bhardwaj 106336988070SRajneesh Bhardwaj struct kfd_criu_device_priv_data { 106436988070SRajneesh Bhardwaj /* For future use */ 106536988070SRajneesh Bhardwaj uint64_t reserved; 106636988070SRajneesh Bhardwaj }; 106736988070SRajneesh Bhardwaj 106836988070SRajneesh Bhardwaj struct kfd_criu_bo_priv_data { 10695ccbb057SRajneesh Bhardwaj uint64_t user_addr; 10705ccbb057SRajneesh Bhardwaj uint32_t idr_handle; 10715ccbb057SRajneesh Bhardwaj uint32_t mapped_gpuids[MAX_GPU_INSTANCE]; 107236988070SRajneesh Bhardwaj }; 107336988070SRajneesh Bhardwaj 1074626f7b31SDavid Yat Sin /* 1075626f7b31SDavid Yat Sin * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data, 1076626f7b31SDavid Yat Sin * kfd_criu_svm_range_priv_data is the object type 1077626f7b31SDavid Yat Sin */ 1078626f7b31SDavid Yat Sin enum kfd_criu_object_type { 1079626f7b31SDavid Yat Sin KFD_CRIU_OBJECT_TYPE_QUEUE, 1080626f7b31SDavid Yat Sin KFD_CRIU_OBJECT_TYPE_EVENT, 1081626f7b31SDavid Yat Sin KFD_CRIU_OBJECT_TYPE_SVM_RANGE, 1082626f7b31SDavid Yat Sin }; 1083626f7b31SDavid Yat Sin 108436988070SRajneesh Bhardwaj struct kfd_criu_svm_range_priv_data { 108536988070SRajneesh Bhardwaj uint32_t object_type; 108608a987a8SRajneesh Bhardwaj uint64_t start_addr; 108708a987a8SRajneesh Bhardwaj uint64_t size; 108808a987a8SRajneesh Bhardwaj /* Variable length array of attributes */ 1089d5c83156SChangcheng Deng struct kfd_ioctl_svm_attribute attrs[]; 109036988070SRajneesh Bhardwaj }; 109136988070SRajneesh Bhardwaj 109236988070SRajneesh Bhardwaj struct kfd_criu_queue_priv_data { 109336988070SRajneesh Bhardwaj uint32_t object_type; 1094626f7b31SDavid Yat Sin uint64_t q_address; 1095626f7b31SDavid Yat Sin uint64_t q_size; 1096626f7b31SDavid Yat Sin uint64_t read_ptr_addr; 1097626f7b31SDavid Yat Sin uint64_t write_ptr_addr; 1098626f7b31SDavid Yat Sin uint64_t doorbell_off; 1099626f7b31SDavid Yat Sin uint64_t eop_ring_buffer_address; 1100626f7b31SDavid Yat Sin uint64_t ctx_save_restore_area_address; 1101626f7b31SDavid Yat Sin uint32_t gpu_id; 1102626f7b31SDavid Yat Sin uint32_t type; 1103626f7b31SDavid Yat Sin uint32_t format; 1104626f7b31SDavid Yat Sin uint32_t q_id; 1105626f7b31SDavid Yat Sin uint32_t priority; 1106626f7b31SDavid Yat Sin uint32_t q_percent; 1107626f7b31SDavid Yat Sin uint32_t doorbell_id; 1108747eea07SDavid Yat Sin uint32_t gws; 1109626f7b31SDavid Yat Sin uint32_t sdma_id; 1110626f7b31SDavid Yat Sin uint32_t eop_ring_buffer_size; 1111626f7b31SDavid Yat Sin uint32_t ctx_save_restore_area_size; 1112626f7b31SDavid Yat Sin uint32_t ctl_stack_size; 1113626f7b31SDavid Yat Sin uint32_t mqd_size; 111436988070SRajneesh Bhardwaj }; 111536988070SRajneesh Bhardwaj 111636988070SRajneesh Bhardwaj struct kfd_criu_event_priv_data { 111736988070SRajneesh Bhardwaj uint32_t object_type; 111840e8a766SDavid Yat Sin uint64_t user_handle; 111940e8a766SDavid Yat Sin uint32_t event_id; 112040e8a766SDavid Yat Sin uint32_t auto_reset; 112140e8a766SDavid Yat Sin uint32_t type; 112240e8a766SDavid Yat Sin uint32_t signaled; 112340e8a766SDavid Yat Sin 112440e8a766SDavid Yat Sin union { 112540e8a766SDavid Yat Sin struct kfd_hsa_memory_exception_data memory_exception_data; 112640e8a766SDavid Yat Sin struct kfd_hsa_hw_exception_data hw_exception_data; 112740e8a766SDavid Yat Sin }; 112836988070SRajneesh Bhardwaj }; 112936988070SRajneesh Bhardwaj 1130626f7b31SDavid Yat Sin int kfd_process_get_queue_info(struct kfd_process *p, 1131626f7b31SDavid Yat Sin uint32_t *num_queues, 1132626f7b31SDavid Yat Sin uint64_t *priv_data_sizes); 1133626f7b31SDavid Yat Sin 1134626f7b31SDavid Yat Sin int kfd_criu_checkpoint_queues(struct kfd_process *p, 1135626f7b31SDavid Yat Sin uint8_t __user *user_priv_data, 1136626f7b31SDavid Yat Sin uint64_t *priv_data_offset); 1137626f7b31SDavid Yat Sin 1138626f7b31SDavid Yat Sin int kfd_criu_restore_queue(struct kfd_process *p, 1139626f7b31SDavid Yat Sin uint8_t __user *user_priv_data, 1140626f7b31SDavid Yat Sin uint64_t *priv_data_offset, 1141626f7b31SDavid Yat Sin uint64_t max_priv_data_size); 114240e8a766SDavid Yat Sin 114340e8a766SDavid Yat Sin int kfd_criu_checkpoint_events(struct kfd_process *p, 114440e8a766SDavid Yat Sin uint8_t __user *user_priv_data, 114540e8a766SDavid Yat Sin uint64_t *priv_data_offset); 114640e8a766SDavid Yat Sin 114740e8a766SDavid Yat Sin int kfd_criu_restore_event(struct file *devkfd, 114840e8a766SDavid Yat Sin struct kfd_process *p, 114940e8a766SDavid Yat Sin uint8_t __user *user_priv_data, 115040e8a766SDavid Yat Sin uint64_t *priv_data_offset, 115140e8a766SDavid Yat Sin uint64_t max_priv_data_size); 115236988070SRajneesh Bhardwaj /* CRIU - End */ 115336988070SRajneesh Bhardwaj 1154ed6e6a34SBen Goz /* Queue Context Management */ 1155e88a614cSEdward O'Callaghan int init_queue(struct queue **q, const struct queue_properties *properties); 1156ed6e6a34SBen Goz void uninit_queue(struct queue *q); 115745102048SBen Goz void print_queue_properties(struct queue_properties *q); 1158ed6e6a34SBen Goz void print_queue(struct queue *q); 1159ed6e6a34SBen Goz 11604b8f589bSBen Goz struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type, 11614b8f589bSBen Goz struct kfd_dev *dev); 1162ee04955aSFelix Kuehling struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type, 1163ee04955aSFelix Kuehling struct kfd_dev *dev); 11644b8f589bSBen Goz struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type, 11654b8f589bSBen Goz struct kfd_dev *dev); 1166ee04955aSFelix Kuehling struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type, 1167ee04955aSFelix Kuehling struct kfd_dev *dev); 1168b91d43ddSFelix Kuehling struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type, 1169b91d43ddSFelix Kuehling struct kfd_dev *dev); 117014328aa5SPhilip Cox struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type, 117114328aa5SPhilip Cox struct kfd_dev *dev); 117264c7f8cfSBen Goz struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev); 117364c7f8cfSBen Goz void device_queue_manager_uninit(struct device_queue_manager *dqm); 1174241f24f8SBen Goz struct kernel_queue *kernel_queue_init(struct kfd_dev *dev, 1175241f24f8SBen Goz enum kfd_queue_type type); 1176c2a77fdeSFelix Kuehling void kernel_queue_uninit(struct kernel_queue *kq, bool hanging); 117703e5b167STao Zhou int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid); 1178241f24f8SBen Goz 117945102048SBen Goz /* Process Queue Manager */ 118045102048SBen Goz struct process_queue_node { 118145102048SBen Goz struct queue *q; 118245102048SBen Goz struct kernel_queue *kq; 118345102048SBen Goz struct list_head process_queue_list; 118445102048SBen Goz }; 118545102048SBen Goz 11869fd3f1bfSFelix Kuehling void kfd_process_dequeue_from_device(struct kfd_process_device *pdd); 11879fd3f1bfSFelix Kuehling void kfd_process_dequeue_from_all_devices(struct kfd_process *p); 118845102048SBen Goz int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p); 118945102048SBen Goz void pqm_uninit(struct process_queue_manager *pqm); 119045102048SBen Goz int pqm_create_queue(struct process_queue_manager *pqm, 119145102048SBen Goz struct kfd_dev *dev, 119245102048SBen Goz struct file *f, 119345102048SBen Goz struct queue_properties *properties, 1194e47a8b52SYong Zhao unsigned int *qid, 11958668dfc3SDavid Yat Sin const struct kfd_criu_queue_priv_data *q_data, 119642c6c482SDavid Yat Sin const void *restore_mqd, 11973a9822d7SDavid Yat Sin const void *restore_ctl_stack, 1198e47a8b52SYong Zhao uint32_t *p_doorbell_offset_in_process); 119945102048SBen Goz int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid); 12007c695a2cSLang Yu int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid, 120145102048SBen Goz struct queue_properties *p); 12027c695a2cSLang Yu int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid, 12037c695a2cSLang Yu struct mqd_update_info *minfo); 1204eb82da1dSOak Zeng int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, 1205eb82da1dSOak Zeng void *gws); 1206fbeb661bSYair Shachar struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm, 1207fbeb661bSYair Shachar unsigned int qid); 12085bb4b78bSOak Zeng struct queue *pqm_get_user_queue(struct process_queue_manager *pqm, 12095bb4b78bSOak Zeng unsigned int qid); 12105df099e8SJay Cornwall int pqm_get_wave_state(struct process_queue_manager *pqm, 12115df099e8SJay Cornwall unsigned int qid, 12125df099e8SJay Cornwall void __user *ctl_stack, 12135df099e8SJay Cornwall u32 *ctl_stack_used_size, 12145df099e8SJay Cornwall u32 *save_area_used_size); 121545102048SBen Goz 1216b010affeSQu Huang int amdkfd_fence_wait_timeout(uint64_t *fence_addr, 1217b010affeSQu Huang uint64_t fence_value, 12188c72c3d7SYong Zhao unsigned int timeout_ms); 1219788bf83dSYair Shachar 122042c6c482SDavid Yat Sin int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm, 122142c6c482SDavid Yat Sin unsigned int qid, 12223a9822d7SDavid Yat Sin u32 *mqd_size, 12233a9822d7SDavid Yat Sin u32 *ctl_stack_size); 1224ed6e6a34SBen Goz /* Packet Manager */ 1225ed6e6a34SBen Goz 122664c7f8cfSBen Goz #define KFD_FENCE_COMPLETED (100) 122764c7f8cfSBen Goz #define KFD_FENCE_INIT (10) 1228241f24f8SBen Goz 1229ed6e6a34SBen Goz struct packet_manager { 1230ed6e6a34SBen Goz struct device_queue_manager *dqm; 1231ed6e6a34SBen Goz struct kernel_queue *priv_queue; 1232ed6e6a34SBen Goz struct mutex lock; 1233ed6e6a34SBen Goz bool allocated; 1234ed6e6a34SBen Goz struct kfd_mem_obj *ib_buffer_obj; 1235851a645eSFelix Kuehling unsigned int ib_size_bytes; 1236819ec5acSFelix Kuehling bool is_over_subscription; 1237f6e27ff1SFelix Kuehling 1238f6e27ff1SFelix Kuehling const struct packet_manager_funcs *pmf; 1239ed6e6a34SBen Goz }; 1240ed6e6a34SBen Goz 1241f6e27ff1SFelix Kuehling struct packet_manager_funcs { 1242f6e27ff1SFelix Kuehling /* Support ASIC-specific packet formats for PM4 packets */ 1243f6e27ff1SFelix Kuehling int (*map_process)(struct packet_manager *pm, uint32_t *buffer, 1244f6e27ff1SFelix Kuehling struct qcm_process_device *qpd); 1245f6e27ff1SFelix Kuehling int (*runlist)(struct packet_manager *pm, uint32_t *buffer, 1246f6e27ff1SFelix Kuehling uint64_t ib, size_t ib_size_in_dwords, bool chain); 1247f6e27ff1SFelix Kuehling int (*set_resources)(struct packet_manager *pm, uint32_t *buffer, 1248f6e27ff1SFelix Kuehling struct scheduling_resources *res); 1249f6e27ff1SFelix Kuehling int (*map_queues)(struct packet_manager *pm, uint32_t *buffer, 1250f6e27ff1SFelix Kuehling struct queue *q, bool is_static); 1251f6e27ff1SFelix Kuehling int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, 1252f6e27ff1SFelix Kuehling enum kfd_unmap_queues_filter mode, 1253d2cb0b21SJonathan Kim uint32_t filter_param, bool reset); 1254f6e27ff1SFelix Kuehling int (*query_status)(struct packet_manager *pm, uint32_t *buffer, 1255b010affeSQu Huang uint64_t fence_address, uint64_t fence_value); 1256f6e27ff1SFelix Kuehling int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); 1257f6e27ff1SFelix Kuehling 1258f6e27ff1SFelix Kuehling /* Packet sizes */ 1259f6e27ff1SFelix Kuehling int map_process_size; 1260f6e27ff1SFelix Kuehling int runlist_size; 1261f6e27ff1SFelix Kuehling int set_resources_size; 1262f6e27ff1SFelix Kuehling int map_queues_size; 1263f6e27ff1SFelix Kuehling int unmap_queues_size; 1264f6e27ff1SFelix Kuehling int query_status_size; 1265f6e27ff1SFelix Kuehling int release_mem_size; 1266f6e27ff1SFelix Kuehling }; 1267f6e27ff1SFelix Kuehling 1268f6e27ff1SFelix Kuehling extern const struct packet_manager_funcs kfd_vi_pm_funcs; 1269454150b1SFelix Kuehling extern const struct packet_manager_funcs kfd_v9_pm_funcs; 1270fd6a440eSJonathan Kim extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs; 1271f6e27ff1SFelix Kuehling 127264c7f8cfSBen Goz int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm); 1273c2a77fdeSFelix Kuehling void pm_uninit(struct packet_manager *pm, bool hanging); 127464c7f8cfSBen Goz int pm_send_set_resources(struct packet_manager *pm, 127564c7f8cfSBen Goz struct scheduling_resources *res); 127664c7f8cfSBen Goz int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues); 127764c7f8cfSBen Goz int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, 1278b010affeSQu Huang uint64_t fence_value); 127964c7f8cfSBen Goz 1280d2cb0b21SJonathan Kim int pm_send_unmap_queue(struct packet_manager *pm, 12817da2bcf8SYong Zhao enum kfd_unmap_queues_filter mode, 1282d2cb0b21SJonathan Kim uint32_t filter_param, bool reset); 128364c7f8cfSBen Goz 1284241f24f8SBen Goz void pm_release_ib(struct packet_manager *pm); 1285241f24f8SBen Goz 1286454150b1SFelix Kuehling /* Following PM funcs can be shared among VI and AI */ 1287454150b1SFelix Kuehling unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); 128814328aa5SPhilip Cox 128919f6d2a6SOded Gabbay uint64_t kfd_get_number_elems(struct kfd_dev *kfd); 129019f6d2a6SOded Gabbay 1291f3a39818SAndrew Lewycky /* Events */ 1292f3a39818SAndrew Lewycky extern const struct kfd_event_interrupt_class event_interrupt_class_cik; 1293ca750681SFelix Kuehling extern const struct kfd_event_interrupt_class event_interrupt_class_v9; 1294ca750681SFelix Kuehling 1295930c5ff4SAlexey Skidanov extern const struct kfd_device_global_init_class device_global_init_class_cik; 1296f3a39818SAndrew Lewycky 1297*c3eb12dfSFelix Kuehling int kfd_event_init_process(struct kfd_process *p); 1298f3a39818SAndrew Lewycky void kfd_event_free_process(struct kfd_process *p); 1299f3a39818SAndrew Lewycky int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma); 1300f3a39818SAndrew Lewycky int kfd_wait_on_events(struct kfd_process *p, 130159d3e8beSAlexey Skidanov uint32_t num_events, void __user *data, 1302f3a39818SAndrew Lewycky bool all, uint32_t user_timeout_ms, 1303fdf0c833SFelix Kuehling uint32_t *wait_result); 1304c7b6bac9SFenghua Yu void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, 1305f3a39818SAndrew Lewycky uint32_t valid_id_bits); 130659d3e8beSAlexey Skidanov void kfd_signal_iommu_event(struct kfd_dev *dev, 1307c7b6bac9SFenghua Yu u32 pasid, unsigned long address, 130859d3e8beSAlexey Skidanov bool is_write_requested, bool is_execute_requested); 1309c7b6bac9SFenghua Yu void kfd_signal_hw_exception_event(u32 pasid); 1310f3a39818SAndrew Lewycky int kfd_set_event(struct kfd_process *p, uint32_t event_id); 1311f3a39818SAndrew Lewycky int kfd_reset_event(struct kfd_process *p, uint32_t event_id); 131240e8a766SDavid Yat Sin int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset); 131340e8a766SDavid Yat Sin 1314f3a39818SAndrew Lewycky int kfd_event_create(struct file *devkfd, struct kfd_process *p, 1315f3a39818SAndrew Lewycky uint32_t event_type, bool auto_reset, uint32_t node_id, 1316f3a39818SAndrew Lewycky uint32_t *event_id, uint32_t *event_trigger_data, 1317f3a39818SAndrew Lewycky uint64_t *event_page_offset, uint32_t *event_slot_index); 131840e8a766SDavid Yat Sin 131940e8a766SDavid Yat Sin int kfd_get_num_events(struct kfd_process *p); 1320f3a39818SAndrew Lewycky int kfd_event_destroy(struct kfd_process *p, uint32_t event_id); 1321f3a39818SAndrew Lewycky 1322c7b6bac9SFenghua Yu void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, 13232640c3faSshaoyunl struct kfd_vm_fault_info *info); 13242640c3faSshaoyunl 1325e42051d2SShaoyun Liu void kfd_signal_reset_event(struct kfd_dev *dev); 1326e42051d2SShaoyun Liu 1327e2b1f9f5SDennis Li void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid); 1328e2b1f9f5SDennis Li 13293543b055SEric Huang void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type); 1330403575c4SFelix Kuehling 1331459ccca5SLang Yu static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev) 1332459ccca5SLang Yu { 1333459ccca5SLang Yu return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2) || 1334459ccca5SLang Yu (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && 1335459ccca5SLang Yu dev->adev->sdma.instance[0].fw_version >= 18) || 1336459ccca5SLang Yu KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0); 1337459ccca5SLang Yu } 1338459ccca5SLang Yu 1339e42051d2SShaoyun Liu bool kfd_is_locked(void); 1340e42051d2SShaoyun Liu 1341f756e631SHarish Kasiviswanathan /* Compute profile */ 1342f756e631SHarish Kasiviswanathan void kfd_inc_compute_active(struct kfd_dev *dev); 1343f756e631SHarish Kasiviswanathan void kfd_dec_compute_active(struct kfd_dev *dev); 1344f756e631SHarish Kasiviswanathan 13456b855f7bSHarish Kasiviswanathan /* Cgroup Support */ 13466b855f7bSHarish Kasiviswanathan /* Check with device cgroup if @kfd device is accessible */ 13476b855f7bSHarish Kasiviswanathan static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd) 13486b855f7bSHarish Kasiviswanathan { 1349eec8fd02SOdin Ugedal #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF) 13506b855f7bSHarish Kasiviswanathan struct drm_device *ddev = kfd->ddev; 13516b855f7bSHarish Kasiviswanathan 135299c7b309SLorenz Brun return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR, 13536b855f7bSHarish Kasiviswanathan ddev->render->index, 13546b855f7bSHarish Kasiviswanathan DEVCG_ACC_WRITE | DEVCG_ACC_READ); 13556b855f7bSHarish Kasiviswanathan #else 13566b855f7bSHarish Kasiviswanathan return 0; 13576b855f7bSHarish Kasiviswanathan #endif 13586b855f7bSHarish Kasiviswanathan } 13596b855f7bSHarish Kasiviswanathan 1360851a645eSFelix Kuehling /* Debugfs */ 1361851a645eSFelix Kuehling #if defined(CONFIG_DEBUG_FS) 1362851a645eSFelix Kuehling 1363851a645eSFelix Kuehling void kfd_debugfs_init(void); 1364851a645eSFelix Kuehling void kfd_debugfs_fini(void); 1365851a645eSFelix Kuehling int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data); 1366851a645eSFelix Kuehling int pqm_debugfs_mqds(struct seq_file *m, void *data); 1367851a645eSFelix Kuehling int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data); 1368851a645eSFelix Kuehling int dqm_debugfs_hqds(struct seq_file *m, void *data); 1369851a645eSFelix Kuehling int kfd_debugfs_rls_by_device(struct seq_file *m, void *data); 1370851a645eSFelix Kuehling int pm_debugfs_runlist(struct seq_file *m, void *data); 1371851a645eSFelix Kuehling 1372a29ec470SShaoyun Liu int kfd_debugfs_hang_hws(struct kfd_dev *dev); 1373a29ec470SShaoyun Liu int pm_debugfs_hang_hws(struct packet_manager *pm); 13744f942aaeSOak Zeng int dqm_debugfs_hang_hws(struct device_queue_manager *dqm); 1375a29ec470SShaoyun Liu 1376851a645eSFelix Kuehling #else 1377851a645eSFelix Kuehling 1378851a645eSFelix Kuehling static inline void kfd_debugfs_init(void) {} 1379851a645eSFelix Kuehling static inline void kfd_debugfs_fini(void) {} 1380851a645eSFelix Kuehling 1381851a645eSFelix Kuehling #endif 1382851a645eSFelix Kuehling 13834a488a7aSOded Gabbay #endif 1384