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 return roundup(kfd->device_info.doorbell_size * 53 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, 54 PAGE_SIZE); 55 } 56 57 /* Doorbell calculations for device init. */ 58 int kfd_doorbell_init(struct kfd_dev *kfd) 59 { 60 size_t doorbell_start_offset; 61 size_t doorbell_aperture_size; 62 size_t doorbell_process_limit; 63 64 /* 65 * We start with calculations in bytes because the input data might 66 * only be byte-aligned. 67 * Only after we have done the rounding can we assume any alignment. 68 */ 69 70 doorbell_start_offset = 71 roundup(kfd->shared_resources.doorbell_start_offset, 72 kfd_doorbell_process_slice(kfd)); 73 74 doorbell_aperture_size = 75 rounddown(kfd->shared_resources.doorbell_aperture_size, 76 kfd_doorbell_process_slice(kfd)); 77 78 if (doorbell_aperture_size > doorbell_start_offset) 79 doorbell_process_limit = 80 (doorbell_aperture_size - doorbell_start_offset) / 81 kfd_doorbell_process_slice(kfd); 82 else 83 return -ENOSPC; 84 85 if (!kfd->max_doorbell_slices || 86 doorbell_process_limit < kfd->max_doorbell_slices) 87 kfd->max_doorbell_slices = doorbell_process_limit; 88 89 kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address + 90 doorbell_start_offset; 91 92 kfd->doorbell_base_dw_offset = doorbell_start_offset / sizeof(u32); 93 94 kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base, 95 kfd_doorbell_process_slice(kfd)); 96 97 if (!kfd->doorbell_kernel_ptr) 98 return -ENOMEM; 99 100 pr_debug("Doorbell initialization:\n"); 101 pr_debug("doorbell base == 0x%08lX\n", 102 (uintptr_t)kfd->doorbell_base); 103 104 pr_debug("doorbell_base_dw_offset == 0x%08lX\n", 105 kfd->doorbell_base_dw_offset); 106 107 pr_debug("doorbell_process_limit == 0x%08lX\n", 108 doorbell_process_limit); 109 110 pr_debug("doorbell_kernel_offset == 0x%08lX\n", 111 (uintptr_t)kfd->doorbell_base); 112 113 pr_debug("doorbell aperture size == 0x%08lX\n", 114 kfd->shared_resources.doorbell_aperture_size); 115 116 pr_debug("doorbell kernel address == %p\n", kfd->doorbell_kernel_ptr); 117 118 return 0; 119 } 120 121 void kfd_doorbell_fini(struct kfd_dev *kfd) 122 { 123 if (kfd->doorbell_kernel_ptr) 124 iounmap(kfd->doorbell_kernel_ptr); 125 } 126 127 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process, 128 struct vm_area_struct *vma) 129 { 130 phys_addr_t address; 131 struct kfd_process_device *pdd; 132 133 /* 134 * For simplicitly we only allow mapping of the entire doorbell 135 * allocation of a single device & process. 136 */ 137 if (vma->vm_end - vma->vm_start != kfd_doorbell_process_slice(dev)) 138 return -EINVAL; 139 140 pdd = kfd_get_process_device_data(dev, process); 141 if (!pdd) 142 return -EINVAL; 143 144 /* Calculate physical address of doorbell */ 145 address = kfd_get_process_doorbells(pdd); 146 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE | 147 VM_DONTDUMP | VM_PFNMAP; 148 149 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 150 151 pr_debug("Mapping doorbell page\n" 152 " target user address == 0x%08llX\n" 153 " physical address == 0x%08llX\n" 154 " vm_flags == 0x%04lX\n" 155 " size == 0x%04lX\n", 156 (unsigned long long) vma->vm_start, address, vma->vm_flags, 157 kfd_doorbell_process_slice(dev)); 158 159 160 return io_remap_pfn_range(vma, 161 vma->vm_start, 162 address >> PAGE_SHIFT, 163 kfd_doorbell_process_slice(dev), 164 vma->vm_page_prot); 165 } 166 167 168 /* get kernel iomem pointer for a doorbell */ 169 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 170 unsigned int *doorbell_off) 171 { 172 u32 inx; 173 174 mutex_lock(&kfd->doorbell_mutex); 175 inx = find_first_zero_bit(kfd->doorbell_available_index, 176 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 177 178 __set_bit(inx, kfd->doorbell_available_index); 179 mutex_unlock(&kfd->doorbell_mutex); 180 181 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 182 return NULL; 183 184 inx *= kfd->device_info.doorbell_size / sizeof(u32); 185 186 /* 187 * Calculating the kernel doorbell offset using the first 188 * doorbell page. 189 */ 190 *doorbell_off = kfd->doorbell_base_dw_offset + inx; 191 192 pr_debug("Get kernel queue doorbell\n" 193 " doorbell offset == 0x%08X\n" 194 " doorbell index == 0x%x\n", 195 *doorbell_off, inx); 196 197 return kfd->doorbell_kernel_ptr + inx; 198 } 199 200 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr) 201 { 202 unsigned int inx; 203 204 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr) 205 * sizeof(u32) / kfd->device_info.doorbell_size; 206 207 mutex_lock(&kfd->doorbell_mutex); 208 __clear_bit(inx, kfd->doorbell_available_index); 209 mutex_unlock(&kfd->doorbell_mutex); 210 } 211 212 void write_kernel_doorbell(void __iomem *db, u32 value) 213 { 214 if (db) { 215 writel(value, db); 216 pr_debug("Writing %d to doorbell address %p\n", value, db); 217 } 218 } 219 220 void write_kernel_doorbell64(void __iomem *db, u64 value) 221 { 222 if (db) { 223 WARN(((unsigned long)db & 7) != 0, 224 "Unaligned 64-bit doorbell"); 225 writeq(value, (u64 __iomem *)db); 226 pr_debug("writing %llu to doorbell address %p\n", value, db); 227 } 228 } 229 230 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd, 231 struct kfd_process_device *pdd, 232 unsigned int doorbell_id) 233 { 234 /* 235 * doorbell_base_dw_offset accounts for doorbells taken by KGD. 236 * index * kfd_doorbell_process_slice/sizeof(u32) adjusts to 237 * the process's doorbells. The offset returned is in dword 238 * units regardless of the ASIC-dependent doorbell size. 239 */ 240 return kfd->doorbell_base_dw_offset + 241 pdd->doorbell_index 242 * kfd_doorbell_process_slice(kfd) / sizeof(u32) + 243 doorbell_id * kfd->device_info.doorbell_size / sizeof(u32); 244 } 245 246 uint64_t kfd_get_number_elems(struct kfd_dev *kfd) 247 { 248 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size - 249 kfd->shared_resources.doorbell_start_offset) / 250 kfd_doorbell_process_slice(kfd) + 1; 251 252 return num_of_elems; 253 254 } 255 256 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd) 257 { 258 return pdd->dev->doorbell_base + 259 pdd->doorbell_index * kfd_doorbell_process_slice(pdd->dev); 260 } 261 262 int kfd_alloc_process_doorbells(struct kfd_dev *kfd, unsigned int *doorbell_index) 263 { 264 int r = ida_simple_get(&kfd->doorbell_ida, 1, kfd->max_doorbell_slices, 265 GFP_KERNEL); 266 if (r > 0) 267 *doorbell_index = r; 268 269 return r; 270 } 271 272 void kfd_free_process_doorbells(struct kfd_dev *kfd, unsigned int doorbell_index) 273 { 274 if (doorbell_index) 275 ida_simple_remove(&kfd->doorbell_ida, doorbell_index); 276 } 277