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 28 /* 29 * This extension supports a kernel level doorbells management for 30 * the kernel queues. 31 * Basically the last doorbells page is devoted to kernel queues 32 * and that's assures that any user process won't get access to the 33 * kernel doorbells page 34 */ 35 static DEFINE_MUTEX(doorbell_mutex); 36 static unsigned long doorbell_available_index[ 37 DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, BITS_PER_LONG)] = { 0 }; 38 39 #define KERNEL_DOORBELL_PASID 1 40 #define KFD_SIZE_OF_DOORBELL_IN_BYTES 4 41 42 /* 43 * Each device exposes a doorbell aperture, a PCI MMIO aperture that 44 * receives 32-bit writes that are passed to queues as wptr values. 45 * The doorbells are intended to be written by applications as part 46 * of queueing work on user-mode queues. 47 * We assign doorbells to applications in PAGE_SIZE-sized and aligned chunks. 48 * We map the doorbell address space into user-mode when a process creates 49 * its first queue on each device. 50 * Although the mapping is done by KFD, it is equivalent to an mmap of 51 * the /dev/kfd with the particular device encoded in the mmap offset. 52 * There will be other uses for mmap of /dev/kfd, so only a range of 53 * offsets (KFD_MMAP_DOORBELL_START-END) is used for doorbells. 54 */ 55 56 /* # of doorbell bytes allocated for each process. */ 57 static inline size_t doorbell_process_allocation(void) 58 { 59 return roundup(KFD_SIZE_OF_DOORBELL_IN_BYTES * 60 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, 61 PAGE_SIZE); 62 } 63 64 /* Doorbell calculations for device init. */ 65 void kfd_doorbell_init(struct kfd_dev *kfd) 66 { 67 size_t doorbell_start_offset; 68 size_t doorbell_aperture_size; 69 size_t doorbell_process_limit; 70 71 /* 72 * We start with calculations in bytes because the input data might 73 * only be byte-aligned. 74 * Only after we have done the rounding can we assume any alignment. 75 */ 76 77 doorbell_start_offset = 78 roundup(kfd->shared_resources.doorbell_start_offset, 79 doorbell_process_allocation()); 80 81 doorbell_aperture_size = 82 rounddown(kfd->shared_resources.doorbell_aperture_size, 83 doorbell_process_allocation()); 84 85 if (doorbell_aperture_size > doorbell_start_offset) 86 doorbell_process_limit = 87 (doorbell_aperture_size - doorbell_start_offset) / 88 doorbell_process_allocation(); 89 else 90 doorbell_process_limit = 0; 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 kfd->doorbell_process_limit = doorbell_process_limit - 1; 97 98 kfd->doorbell_kernel_ptr = ioremap(kfd->doorbell_base, 99 doorbell_process_allocation()); 100 101 BUG_ON(!kfd->doorbell_kernel_ptr); 102 103 pr_debug("kfd: doorbell initialization:\n"); 104 pr_debug("kfd: doorbell base == 0x%08lX\n", 105 (uintptr_t)kfd->doorbell_base); 106 107 pr_debug("kfd: doorbell_id_offset == 0x%08lX\n", 108 kfd->doorbell_id_offset); 109 110 pr_debug("kfd: doorbell_process_limit == 0x%08lX\n", 111 doorbell_process_limit); 112 113 pr_debug("kfd: doorbell_kernel_offset == 0x%08lX\n", 114 (uintptr_t)kfd->doorbell_base); 115 116 pr_debug("kfd: doorbell aperture size == 0x%08lX\n", 117 kfd->shared_resources.doorbell_aperture_size); 118 119 pr_debug("kfd: doorbell kernel address == 0x%08lX\n", 120 (uintptr_t)kfd->doorbell_kernel_ptr); 121 } 122 123 int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma) 124 { 125 phys_addr_t address; 126 struct kfd_dev *dev; 127 128 /* 129 * For simplicitly we only allow mapping of the entire doorbell 130 * allocation of a single device & process. 131 */ 132 if (vma->vm_end - vma->vm_start != doorbell_process_allocation()) 133 return -EINVAL; 134 135 /* Find kfd device according to gpu id */ 136 dev = kfd_device_by_id(vma->vm_pgoff); 137 if (dev == NULL) 138 return -EINVAL; 139 140 /* Find if pdd exists for combination of process and gpu id */ 141 if (!kfd_get_process_device_data(dev, process, 0)) 142 return -EINVAL; 143 144 /* Calculate physical address of doorbell */ 145 address = kfd_get_process_doorbells(dev, process); 146 147 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE | 148 VM_DONTDUMP | VM_PFNMAP; 149 150 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 151 152 pr_debug("kfd: mapping doorbell page in kfd_doorbell_mmap\n" 153 " target user address == 0x%08llX\n" 154 " physical address == 0x%08llX\n" 155 " vm_flags == 0x%04lX\n" 156 " size == 0x%04lX\n", 157 (unsigned long long) vma->vm_start, address, vma->vm_flags, 158 doorbell_process_allocation()); 159 160 161 return io_remap_pfn_range(vma, 162 vma->vm_start, 163 address >> PAGE_SHIFT, 164 doorbell_process_allocation(), 165 vma->vm_page_prot); 166 } 167 168 169 /* get kernel iomem pointer for a doorbell */ 170 u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 171 unsigned int *doorbell_off) 172 { 173 u32 inx; 174 175 BUG_ON(!kfd || !doorbell_off); 176 177 mutex_lock(&doorbell_mutex); 178 inx = find_first_zero_bit(doorbell_available_index, 179 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 180 181 __set_bit(inx, doorbell_available_index); 182 mutex_unlock(&doorbell_mutex); 183 184 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 185 return NULL; 186 187 /* 188 * Calculating the kernel doorbell offset using "faked" kernel 189 * pasid that allocated for kernel queues only 190 */ 191 *doorbell_off = KERNEL_DOORBELL_PASID * (doorbell_process_allocation() / 192 sizeof(u32)) + inx; 193 194 pr_debug("kfd: get kernel queue doorbell\n" 195 " doorbell offset == 0x%08d\n" 196 " kernel address == 0x%08lX\n", 197 *doorbell_off, (uintptr_t)(kfd->doorbell_kernel_ptr + inx)); 198 199 return kfd->doorbell_kernel_ptr + inx; 200 } 201 202 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr) 203 { 204 unsigned int inx; 205 206 BUG_ON(!kfd || !db_addr); 207 208 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr); 209 210 mutex_lock(&doorbell_mutex); 211 __clear_bit(inx, doorbell_available_index); 212 mutex_unlock(&doorbell_mutex); 213 } 214 215 inline void write_kernel_doorbell(u32 __iomem *db, u32 value) 216 { 217 if (db) { 218 writel(value, db); 219 pr_debug("writing %d to doorbell address 0x%p\n", value, db); 220 } 221 } 222 223 /* 224 * queue_ids are in the range [0,MAX_PROCESS_QUEUES) and are mapped 1:1 225 * to doorbells with the process's doorbell page 226 */ 227 unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd, 228 struct kfd_process *process, 229 unsigned int queue_id) 230 { 231 /* 232 * doorbell_id_offset accounts for doorbells taken by KGD. 233 * pasid * doorbell_process_allocation/sizeof(u32) adjusts 234 * to the process's doorbells 235 */ 236 return kfd->doorbell_id_offset + 237 process->pasid * (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->pasid * doorbell_process_allocation(); 256 } 257