1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright 2014-2022 Advanced Micro Devices, Inc. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice shall be included in 13 * all copies or substantial portions of the Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 21 * OTHER DEALINGS IN THE SOFTWARE. 22 */ 23 #include "kfd_priv.h" 24 #include <linux/mm.h> 25 #include <linux/mman.h> 26 #include <linux/slab.h> 27 #include <linux/io.h> 28 #include <linux/idr.h> 29 30 /* 31 * This extension supports a kernel level doorbells management for the 32 * kernel queues using the first doorbell page reserved for the kernel. 33 */ 34 35 /* 36 * Each device exposes a doorbell aperture, a PCI MMIO aperture that 37 * receives 32-bit writes that are passed to queues as wptr values. 38 * The doorbells are intended to be written by applications as part 39 * of queueing work on user-mode queues. 40 * We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks. 41 * We map the doorbell address space into user-mode when a process creates 42 * its first queue on each device. 43 * Although the mapping is done by KFD, it is equivalent to an mmap of 44 * the /dev/kfd with the particular device encoded in the mmap offset. 45 * There will be other uses for mmap of /dev/kfd, so only a range of 46 * offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells. 47 */ 48 49 /* # of doorbell bytes allocated for each process. */ 50 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd) 51 { 52 if (!kfd->shared_resources.enable_mes) 53 return roundup(kfd->device_info.doorbell_size * 54 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, 55 PAGE_SIZE); 56 else 57 return amdgpu_mes_doorbell_process_slice( 58 (struct amdgpu_device *)kfd->adev); 59 } 60 61 /* Doorbell calculations for device init. */ 62 int kfd_doorbell_init(struct kfd_dev *kfd) 63 { 64 int size = PAGE_SIZE; 65 int r; 66 67 /* 68 * Todo: KFD kernel level operations need only one doorbell for 69 * ring test/HWS. So instead of reserving a whole page here for 70 * kernel, reserve and consume a doorbell from existing KGD kernel 71 * doorbell page. 72 */ 73 74 /* Bitmap to dynamically allocate doorbells from kernel page */ 75 kfd->doorbell_bitmap = bitmap_zalloc(size / sizeof(u32), GFP_KERNEL); 76 if (!kfd->doorbell_bitmap) { 77 DRM_ERROR("Failed to allocate kernel doorbell bitmap\n"); 78 return -ENOMEM; 79 } 80 81 /* Alloc a doorbell page for KFD kernel usages */ 82 r = amdgpu_bo_create_kernel(kfd->adev, 83 size, 84 PAGE_SIZE, 85 AMDGPU_GEM_DOMAIN_DOORBELL, 86 &kfd->doorbells, 87 NULL, 88 (void **)&kfd->doorbell_kernel_ptr); 89 if (r) { 90 pr_err("failed to allocate kernel doorbells\n"); 91 bitmap_free(kfd->doorbell_bitmap); 92 return r; 93 } 94 95 pr_debug("Doorbell kernel address == %p\n", kfd->doorbell_kernel_ptr); 96 return 0; 97 } 98 99 void kfd_doorbell_fini(struct kfd_dev *kfd) 100 { 101 bitmap_free(kfd->doorbell_bitmap); 102 amdgpu_bo_free_kernel(&kfd->doorbells, NULL, 103 (void **)&kfd->doorbell_kernel_ptr); 104 } 105 106 int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process, 107 struct vm_area_struct *vma) 108 { 109 phys_addr_t address; 110 struct kfd_process_device *pdd; 111 112 /* 113 * For simplicitly we only allow mapping of the entire doorbell 114 * allocation of a single device & process. 115 */ 116 if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev->kfd)) 117 return -EINVAL; 118 119 pdd = kfd_get_process_device_data(dev, process); 120 if (!pdd) 121 return -EINVAL; 122 123 /* Calculate physical address of doorbell */ 124 address = kfd_get_process_doorbells(pdd); 125 if (!address) 126 return -ENOMEM; 127 vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE | 128 VM_DONTDUMP | VM_PFNMAP); 129 130 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 131 132 pr_debug("Mapping doorbell page\n" 133 " target user address == 0x%08llX\n" 134 " physical address == 0x%08llX\n" 135 " vm_flags == 0x%04lX\n" 136 " size == 0x%04lX\n", 137 (unsigned long long) vma->vm_start, address, vma->vm_flags, 138 kfd_doorbell_process_slice(dev->kfd)); 139 140 141 return io_remap_pfn_range(vma, 142 vma->vm_start, 143 address >> PAGE_SHIFT, 144 kfd_doorbell_process_slice(dev->kfd), 145 vma->vm_page_prot); 146 } 147 148 149 /* get kernel iomem pointer for a doorbell */ 150 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 151 unsigned int *doorbell_off) 152 { 153 u32 inx; 154 155 mutex_lock(&kfd->doorbell_mutex); 156 inx = find_first_zero_bit(kfd->doorbell_bitmap, PAGE_SIZE / sizeof(u32)); 157 158 __set_bit(inx, kfd->doorbell_bitmap); 159 mutex_unlock(&kfd->doorbell_mutex); 160 161 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 162 return NULL; 163 164 *doorbell_off = amdgpu_doorbell_index_on_bar(kfd->adev, kfd->doorbells, inx); 165 166 pr_debug("Get kernel queue doorbell\n" 167 " doorbell offset == 0x%08X\n" 168 " doorbell index == 0x%x\n", 169 *doorbell_off, inx); 170 171 return kfd->doorbell_kernel_ptr + inx; 172 } 173 174 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr) 175 { 176 unsigned int inx; 177 178 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr); 179 180 mutex_lock(&kfd->doorbell_mutex); 181 __clear_bit(inx, kfd->doorbell_bitmap); 182 mutex_unlock(&kfd->doorbell_mutex); 183 } 184 185 void write_kernel_doorbell(void __iomem *db, u32 value) 186 { 187 if (db) { 188 writel(value, db); 189 pr_debug("Writing %d to doorbell address %p\n", value, db); 190 } 191 } 192 193 void write_kernel_doorbell64(void __iomem *db, u64 value) 194 { 195 if (db) { 196 WARN(((unsigned long)db & 7) != 0, 197 "Unaligned 64-bit doorbell"); 198 writeq(value, (u64 __iomem *)db); 199 pr_debug("writing %llu to doorbell address %p\n", value, db); 200 } 201 } 202 203 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, 204 struct kfd_process_device *pdd, 205 unsigned int doorbell_id) 206 { 207 /* 208 * doorbell_base_dw_offset accounts for doorbells taken by KGD. 209 * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to 210 * the process's doorbells. The offset returned is in dword 211 * units regardless of the ASIC-dependent doorbell size. 212 */ 213 if (!kfd->shared_resources.enable_mes) 214 return kfd->doorbell_base_dw_offset + 215 pdd->doorbell_index 216 * kfd_doorbell_process_slice(kfd) / sizeof(u32) + 217 doorbell_id * 218 kfd->device_info.doorbell_size / sizeof(u32); 219 else 220 return amdgpu_mes_get_doorbell_dw_offset_in_bar( 221 (struct amdgpu_device *)kfd->adev, 222 pdd->doorbell_index, doorbell_id); 223 } 224 225 uint64_t kfd_get_number_elems(struct kfd_dev *kfd) 226 { 227 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size - 228 kfd->shared_resources.doorbell_start_offset) / 229 kfd_doorbell_process_slice(kfd) + 1; 230 231 return num_of_elems; 232 233 } 234 235 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd) 236 { 237 if (!pdd->doorbell_index) { 238 int r = kfd_alloc_process_doorbells(pdd->dev->kfd, 239 &pdd->doorbell_index); 240 if (r < 0) 241 return 0; 242 } 243 244 return pdd->dev->kfd->doorbell_base + 245 pdd->doorbell_index * kfd_doorbell_process_slice(pdd->dev->kfd); 246 } 247 248 int kfd_alloc_process_doorbells(struct kfd_dev *kfd, unsigned int *doorbell_index) 249 { 250 int r = 0; 251 252 if (!kfd->shared_resources.enable_mes) 253 r = ida_simple_get(&kfd->doorbell_ida, 1, 254 kfd->max_doorbell_slices, GFP_KERNEL); 255 else 256 r = amdgpu_mes_alloc_process_doorbells( 257 (struct amdgpu_device *)kfd->adev, 258 doorbell_index); 259 260 if (r > 0) 261 *doorbell_index = r; 262 263 if (r < 0) 264 pr_err("Failed to allocate process doorbells\n"); 265 266 return r; 267 } 268 269 void kfd_free_process_doorbells(struct kfd_dev *kfd, unsigned int doorbell_index) 270 { 271 if (doorbell_index) { 272 if (!kfd->shared_resources.enable_mes) 273 ida_simple_remove(&kfd->doorbell_ida, doorbell_index); 274 else 275 amdgpu_mes_free_process_doorbells( 276 (struct amdgpu_device *)kfd->adev, 277 doorbell_index); 278 } 279 } 280