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 #define KFD_SIZE_OF_DOORBELL_IN_BYTES 4 37 38 /* 39 * Each device exposes a doorbell aperture, a PCI MMIO aperture that 40 * receives 32-bit writes that are passed to queues as wptr values. 41 * The doorbells are intended to be written by applications as part 42 * of queueing work on user-mode queues. 43 * We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks. 44 * We map the doorbell address space into user-mode when a process creates 45 * its first queue on each device. 46 * Although the mapping is done by KFD, it is equivalent to an mmap of 47 * the /dev/kfd with the particular device encoded in the mmap offset. 48 * There will be other uses for mmap of /dev/kfd, so only a range of 49 * offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells. 50 */ 51 52 /* # of doorbell bytes allocated for each process. */ 53 static inline size_t doorbell_process_allocation(void) 54 { 55 return roundup(KFD_SIZE_OF_DOORBELL_IN_BYTES * 56 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, 57 PAGE_SIZE); 58 } 59 60 /* Doorbell calculations for device init. */ 61 int kfd_doorbell_init(struct kfd_dev *kfd) 62 { 63 size_t doorbell_start_offset; 64 size_t doorbell_aperture_size; 65 size_t doorbell_process_limit; 66 67 /* 68 * We start with calculations in bytes because the input data might 69 * only be byte-aligned. 70 * Only after we have done the rounding can we assume any alignment. 71 */ 72 73 doorbell_start_offset = 74 roundup(kfd->shared_resources.doorbell_start_offset, 75 doorbell_process_allocation()); 76 77 doorbell_aperture_size = 78 rounddown(kfd->shared_resources.doorbell_aperture_size, 79 doorbell_process_allocation()); 80 81 if (doorbell_aperture_size > doorbell_start_offset) 82 doorbell_process_limit = 83 (doorbell_aperture_size - doorbell_start_offset) / 84 doorbell_process_allocation(); 85 else 86 return -ENOSPC; 87 88 if (!max_doorbell_slices || 89 doorbell_process_limit < max_doorbell_slices) 90 max_doorbell_slices = doorbell_process_limit; 91 92 kfd->doorbell_base = kfd->shared_resources.doorbell_physical_address + 93 doorbell_start_offset; 94 95 kfd->doorbell_id_offset = doorbell_start_offset / sizeof(u32); 96 97 kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base, 98 doorbell_process_allocation()); 99 100 if (!kfd->doorbell_kernel_ptr) 101 return -ENOMEM; 102 103 pr_debug("Doorbell initialization:\n"); 104 pr_debug("doorbell base == 0x%08lX\n", 105 (uintptr_t)kfd->doorbell_base); 106 107 pr_debug("doorbell_id_offset == 0x%08lX\n", 108 kfd->doorbell_id_offset); 109 110 pr_debug("doorbell_process_limit == 0x%08lX\n", 111 doorbell_process_limit); 112 113 pr_debug("doorbell_kernel_offset == 0x%08lX\n", 114 (uintptr_t)kfd->doorbell_base); 115 116 pr_debug("doorbell aperture size == 0x%08lX\n", 117 kfd->shared_resources.doorbell_aperture_size); 118 119 pr_debug("doorbell kernel address == 0x%08lX\n", 120 (uintptr_t)kfd->doorbell_kernel_ptr); 121 122 return 0; 123 } 124 125 void kfd_doorbell_fini(struct kfd_dev *kfd) 126 { 127 if (kfd->doorbell_kernel_ptr) 128 iounmap(kfd->doorbell_kernel_ptr); 129 } 130 131 int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma) 132 { 133 phys_addr_t address; 134 struct kfd_dev *dev; 135 136 /* 137 * For simplicitly we only allow mapping of the entire doorbell 138 * allocation of a single device & process. 139 */ 140 if (vma->vm_end - vma->vm_start != doorbell_process_allocation()) 141 return -EINVAL; 142 143 /* Find kfd device according to gpu id */ 144 dev = kfd_device_by_id(vma->vm_pgoff); 145 if (!dev) 146 return -EINVAL; 147 148 /* Calculate physical address of doorbell */ 149 address = kfd_get_process_doorbells(dev, process); 150 151 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE | 152 VM_DONTDUMP | VM_PFNMAP; 153 154 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 155 156 pr_debug("Mapping doorbell page\n" 157 " target user address == 0x%08llX\n" 158 " physical address == 0x%08llX\n" 159 " vm_flags == 0x%04lX\n" 160 " size == 0x%04lX\n", 161 (unsigned long long) vma->vm_start, address, vma->vm_flags, 162 doorbell_process_allocation()); 163 164 165 return io_remap_pfn_range(vma, 166 vma->vm_start, 167 address >> PAGE_SHIFT, 168 doorbell_process_allocation(), 169 vma->vm_page_prot); 170 } 171 172 173 /* get kernel iomem pointer for a doorbell */ 174 u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 175 unsigned int *doorbell_off) 176 { 177 u32 inx; 178 179 mutex_lock(&kfd->doorbell_mutex); 180 inx = find_first_zero_bit(kfd->doorbell_available_index, 181 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 182 183 __set_bit(inx, kfd->doorbell_available_index); 184 mutex_unlock(&kfd->doorbell_mutex); 185 186 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 187 return NULL; 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 == 0x%08lX\n", 198 *doorbell_off, (uintptr_t)(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 inline void write_kernel_doorbell(u32 __iomem *db, u32 value) 215 { 216 if (db) { 217 writel(value, db); 218 pr_debug("Writing %d to doorbell address 0x%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 * doorbell_process_allocation/sizeof(u32) adjusts to 233 * the process's doorbells. 234 */ 235 return kfd->doorbell_id_offset + 236 process->doorbell_index 237 * doorbell_process_allocation() / sizeof(u32) + 238 queue_id; 239 } 240 241 uint64_t kfd_get_number_elems(struct kfd_dev *kfd) 242 { 243 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size - 244 kfd->shared_resources.doorbell_start_offset) / 245 doorbell_process_allocation() + 1; 246 247 return num_of_elems; 248 249 } 250 251 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 252 struct kfd_process *process) 253 { 254 return dev->doorbell_base + 255 process->doorbell_index * doorbell_process_allocation(); 256 } 257 258 int kfd_alloc_process_doorbells(struct kfd_process *process) 259 { 260 int r = ida_simple_get(&doorbell_ida, 1, max_doorbell_slices, 261 GFP_KERNEL); 262 if (r > 0) 263 process->doorbell_index = r; 264 265 return r; 266 } 267 268 void kfd_free_process_doorbells(struct kfd_process *process) 269 { 270 if (process->doorbell_index) 271 ida_simple_remove(&doorbell_ida, process->doorbell_index); 272 } 273