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