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 /* Calculate physical address of doorbell */ 141 address = kfd_get_process_doorbells(dev, process); 142 143 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE | 144 VM_DONTDUMP | VM_PFNMAP; 145 146 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 147 148 pr_debug("kfd: mapping doorbell page in kfd_doorbell_mmap\n" 149 " target user address == 0x%08llX\n" 150 " physical address == 0x%08llX\n" 151 " vm_flags == 0x%04lX\n" 152 " size == 0x%04lX\n", 153 (unsigned long long) vma->vm_start, address, vma->vm_flags, 154 doorbell_process_allocation()); 155 156 157 return io_remap_pfn_range(vma, 158 vma->vm_start, 159 address >> PAGE_SHIFT, 160 doorbell_process_allocation(), 161 vma->vm_page_prot); 162 } 163 164 165 /* get kernel iomem pointer for a doorbell */ 166 u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd, 167 unsigned int *doorbell_off) 168 { 169 u32 inx; 170 171 BUG_ON(!kfd || !doorbell_off); 172 173 mutex_lock(&doorbell_mutex); 174 inx = find_first_zero_bit(doorbell_available_index, 175 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); 176 177 __set_bit(inx, doorbell_available_index); 178 mutex_unlock(&doorbell_mutex); 179 180 if (inx >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) 181 return NULL; 182 183 /* 184 * Calculating the kernel doorbell offset using "faked" kernel 185 * pasid that allocated for kernel queues only 186 */ 187 *doorbell_off = KERNEL_DOORBELL_PASID * (doorbell_process_allocation() / 188 sizeof(u32)) + inx; 189 190 pr_debug("kfd: get kernel queue doorbell\n" 191 " doorbell offset == 0x%08d\n" 192 " kernel address == 0x%08lX\n", 193 *doorbell_off, (uintptr_t)(kfd->doorbell_kernel_ptr + inx)); 194 195 return kfd->doorbell_kernel_ptr + inx; 196 } 197 198 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr) 199 { 200 unsigned int inx; 201 202 BUG_ON(!kfd || !db_addr); 203 204 inx = (unsigned int)(db_addr - kfd->doorbell_kernel_ptr); 205 206 mutex_lock(&doorbell_mutex); 207 __clear_bit(inx, doorbell_available_index); 208 mutex_unlock(&doorbell_mutex); 209 } 210 211 inline void write_kernel_doorbell(u32 __iomem *db, u32 value) 212 { 213 if (db) { 214 writel(value, db); 215 pr_debug("writing %d to doorbell address 0x%p\n", value, db); 216 } 217 } 218 219 /* 220 * queue_ids are in the range [0,MAX_PROCESS_QUEUES) and are mapped 1:1 221 * to doorbells with the process's doorbell page 222 */ 223 unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd, 224 struct kfd_process *process, 225 unsigned int queue_id) 226 { 227 /* 228 * doorbell_id_offset accounts for doorbells taken by KGD. 229 * pasid * doorbell_process_allocation/sizeof(u32) adjusts 230 * to the process's doorbells 231 */ 232 return kfd->doorbell_id_offset + 233 process->pasid * (doorbell_process_allocation()/sizeof(u32)) + 234 queue_id; 235 } 236 237 uint64_t kfd_get_number_elems(struct kfd_dev *kfd) 238 { 239 uint64_t num_of_elems = (kfd->shared_resources.doorbell_aperture_size - 240 kfd->shared_resources.doorbell_start_offset) / 241 doorbell_process_allocation() + 1; 242 243 return num_of_elems; 244 245 } 246 247 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev, 248 struct kfd_process *process) 249 { 250 return dev->doorbell_base + 251 process->pasid * doorbell_process_allocation(); 252 } 253