1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * VFIO PCI I/O Port & MMIO access 4 * 5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved. 6 * Author: Alex Williamson <alex.williamson@redhat.com> 7 * 8 * Derived from original vfio: 9 * Copyright 2010 Cisco Systems, Inc. All rights reserved. 10 * Author: Tom Lyon, pugs@cisco.com 11 */ 12 13 #include <linux/fs.h> 14 #include <linux/pci.h> 15 #include <linux/uaccess.h> 16 #include <linux/io.h> 17 #include <linux/vfio.h> 18 #include <linux/vgaarb.h> 19 20 #include "vfio_pci_private.h" 21 22 #ifdef __LITTLE_ENDIAN 23 #define vfio_ioread64 ioread64 24 #define vfio_iowrite64 iowrite64 25 #define vfio_ioread32 ioread32 26 #define vfio_iowrite32 iowrite32 27 #define vfio_ioread16 ioread16 28 #define vfio_iowrite16 iowrite16 29 #else 30 #define vfio_ioread64 ioread64be 31 #define vfio_iowrite64 iowrite64be 32 #define vfio_ioread32 ioread32be 33 #define vfio_iowrite32 iowrite32be 34 #define vfio_ioread16 ioread16be 35 #define vfio_iowrite16 iowrite16be 36 #endif 37 #define vfio_ioread8 ioread8 38 #define vfio_iowrite8 iowrite8 39 40 /* 41 * Read or write from an __iomem region (MMIO or I/O port) with an excluded 42 * range which is inaccessible. The excluded range drops writes and fills 43 * reads with -1. This is intended for handling MSI-X vector tables and 44 * leftover space for ROM BARs. 45 */ 46 static ssize_t do_io_rw(void __iomem *io, char __user *buf, 47 loff_t off, size_t count, size_t x_start, 48 size_t x_end, bool iswrite) 49 { 50 ssize_t done = 0; 51 52 while (count) { 53 size_t fillable, filled; 54 55 if (off < x_start) 56 fillable = min(count, (size_t)(x_start - off)); 57 else if (off >= x_end) 58 fillable = count; 59 else 60 fillable = 0; 61 62 if (fillable >= 4 && !(off % 4)) { 63 u32 val; 64 65 if (iswrite) { 66 if (copy_from_user(&val, buf, 4)) 67 return -EFAULT; 68 69 vfio_iowrite32(val, io + off); 70 } else { 71 val = vfio_ioread32(io + off); 72 73 if (copy_to_user(buf, &val, 4)) 74 return -EFAULT; 75 } 76 77 filled = 4; 78 } else if (fillable >= 2 && !(off % 2)) { 79 u16 val; 80 81 if (iswrite) { 82 if (copy_from_user(&val, buf, 2)) 83 return -EFAULT; 84 85 vfio_iowrite16(val, io + off); 86 } else { 87 val = vfio_ioread16(io + off); 88 89 if (copy_to_user(buf, &val, 2)) 90 return -EFAULT; 91 } 92 93 filled = 2; 94 } else if (fillable) { 95 u8 val; 96 97 if (iswrite) { 98 if (copy_from_user(&val, buf, 1)) 99 return -EFAULT; 100 101 vfio_iowrite8(val, io + off); 102 } else { 103 val = vfio_ioread8(io + off); 104 105 if (copy_to_user(buf, &val, 1)) 106 return -EFAULT; 107 } 108 109 filled = 1; 110 } else { 111 /* Fill reads with -1, drop writes */ 112 filled = min(count, (size_t)(x_end - off)); 113 if (!iswrite) { 114 u8 val = 0xFF; 115 size_t i; 116 117 for (i = 0; i < filled; i++) 118 if (copy_to_user(buf + i, &val, 1)) 119 return -EFAULT; 120 } 121 } 122 123 count -= filled; 124 done += filled; 125 off += filled; 126 buf += filled; 127 } 128 129 return done; 130 } 131 132 static int vfio_pci_setup_barmap(struct vfio_pci_device *vdev, int bar) 133 { 134 struct pci_dev *pdev = vdev->pdev; 135 int ret; 136 void __iomem *io; 137 138 if (vdev->barmap[bar]) 139 return 0; 140 141 ret = pci_request_selected_regions(pdev, 1 << bar, "vfio"); 142 if (ret) 143 return ret; 144 145 io = pci_iomap(pdev, bar, 0); 146 if (!io) { 147 pci_release_selected_regions(pdev, 1 << bar); 148 return -ENOMEM; 149 } 150 151 vdev->barmap[bar] = io; 152 153 return 0; 154 } 155 156 ssize_t vfio_pci_bar_rw(struct vfio_pci_device *vdev, char __user *buf, 157 size_t count, loff_t *ppos, bool iswrite) 158 { 159 struct pci_dev *pdev = vdev->pdev; 160 loff_t pos = *ppos & VFIO_PCI_OFFSET_MASK; 161 int bar = VFIO_PCI_OFFSET_TO_INDEX(*ppos); 162 size_t x_start = 0, x_end = 0; 163 resource_size_t end; 164 void __iomem *io; 165 ssize_t done; 166 167 if (pci_resource_start(pdev, bar)) 168 end = pci_resource_len(pdev, bar); 169 else if (bar == PCI_ROM_RESOURCE && 170 pdev->resource[bar].flags & IORESOURCE_ROM_SHADOW) 171 end = 0x20000; 172 else 173 return -EINVAL; 174 175 if (pos >= end) 176 return -EINVAL; 177 178 count = min(count, (size_t)(end - pos)); 179 180 if (bar == PCI_ROM_RESOURCE) { 181 /* 182 * The ROM can fill less space than the BAR, so we start the 183 * excluded range at the end of the actual ROM. This makes 184 * filling large ROM BARs much faster. 185 */ 186 io = pci_map_rom(pdev, &x_start); 187 if (!io) 188 return -ENOMEM; 189 x_end = end; 190 } else { 191 int ret = vfio_pci_setup_barmap(vdev, bar); 192 if (ret) 193 return ret; 194 195 io = vdev->barmap[bar]; 196 } 197 198 if (bar == vdev->msix_bar) { 199 x_start = vdev->msix_offset; 200 x_end = vdev->msix_offset + vdev->msix_size; 201 } 202 203 done = do_io_rw(io, buf, pos, count, x_start, x_end, iswrite); 204 205 if (done >= 0) 206 *ppos += done; 207 208 if (bar == PCI_ROM_RESOURCE) 209 pci_unmap_rom(pdev, io); 210 211 return done; 212 } 213 214 ssize_t vfio_pci_vga_rw(struct vfio_pci_device *vdev, char __user *buf, 215 size_t count, loff_t *ppos, bool iswrite) 216 { 217 int ret; 218 loff_t off, pos = *ppos & VFIO_PCI_OFFSET_MASK; 219 void __iomem *iomem = NULL; 220 unsigned int rsrc; 221 bool is_ioport; 222 ssize_t done; 223 224 if (!vdev->has_vga) 225 return -EINVAL; 226 227 if (pos > 0xbfffful) 228 return -EINVAL; 229 230 switch ((u32)pos) { 231 case 0xa0000 ... 0xbffff: 232 count = min(count, (size_t)(0xc0000 - pos)); 233 iomem = ioremap(0xa0000, 0xbffff - 0xa0000 + 1); 234 off = pos - 0xa0000; 235 rsrc = VGA_RSRC_LEGACY_MEM; 236 is_ioport = false; 237 break; 238 case 0x3b0 ... 0x3bb: 239 count = min(count, (size_t)(0x3bc - pos)); 240 iomem = ioport_map(0x3b0, 0x3bb - 0x3b0 + 1); 241 off = pos - 0x3b0; 242 rsrc = VGA_RSRC_LEGACY_IO; 243 is_ioport = true; 244 break; 245 case 0x3c0 ... 0x3df: 246 count = min(count, (size_t)(0x3e0 - pos)); 247 iomem = ioport_map(0x3c0, 0x3df - 0x3c0 + 1); 248 off = pos - 0x3c0; 249 rsrc = VGA_RSRC_LEGACY_IO; 250 is_ioport = true; 251 break; 252 default: 253 return -EINVAL; 254 } 255 256 if (!iomem) 257 return -ENOMEM; 258 259 ret = vga_get_interruptible(vdev->pdev, rsrc); 260 if (ret) { 261 is_ioport ? ioport_unmap(iomem) : iounmap(iomem); 262 return ret; 263 } 264 265 done = do_io_rw(iomem, buf, off, count, 0, 0, iswrite); 266 267 vga_put(vdev->pdev, rsrc); 268 269 is_ioport ? ioport_unmap(iomem) : iounmap(iomem); 270 271 if (done >= 0) 272 *ppos += done; 273 274 return done; 275 } 276 277 static int vfio_pci_ioeventfd_handler(void *opaque, void *unused) 278 { 279 struct vfio_pci_ioeventfd *ioeventfd = opaque; 280 281 switch (ioeventfd->count) { 282 case 1: 283 vfio_iowrite8(ioeventfd->data, ioeventfd->addr); 284 break; 285 case 2: 286 vfio_iowrite16(ioeventfd->data, ioeventfd->addr); 287 break; 288 case 4: 289 vfio_iowrite32(ioeventfd->data, ioeventfd->addr); 290 break; 291 #ifdef iowrite64 292 case 8: 293 vfio_iowrite64(ioeventfd->data, ioeventfd->addr); 294 break; 295 #endif 296 } 297 298 return 0; 299 } 300 301 long vfio_pci_ioeventfd(struct vfio_pci_device *vdev, loff_t offset, 302 uint64_t data, int count, int fd) 303 { 304 struct pci_dev *pdev = vdev->pdev; 305 loff_t pos = offset & VFIO_PCI_OFFSET_MASK; 306 int ret, bar = VFIO_PCI_OFFSET_TO_INDEX(offset); 307 struct vfio_pci_ioeventfd *ioeventfd; 308 309 /* Only support ioeventfds into BARs */ 310 if (bar > VFIO_PCI_BAR5_REGION_INDEX) 311 return -EINVAL; 312 313 if (pos + count > pci_resource_len(pdev, bar)) 314 return -EINVAL; 315 316 /* Disallow ioeventfds working around MSI-X table writes */ 317 if (bar == vdev->msix_bar && 318 !(pos + count <= vdev->msix_offset || 319 pos >= vdev->msix_offset + vdev->msix_size)) 320 return -EINVAL; 321 322 #ifndef iowrite64 323 if (count == 8) 324 return -EINVAL; 325 #endif 326 327 ret = vfio_pci_setup_barmap(vdev, bar); 328 if (ret) 329 return ret; 330 331 mutex_lock(&vdev->ioeventfds_lock); 332 333 list_for_each_entry(ioeventfd, &vdev->ioeventfds_list, next) { 334 if (ioeventfd->pos == pos && ioeventfd->bar == bar && 335 ioeventfd->data == data && ioeventfd->count == count) { 336 if (fd == -1) { 337 vfio_virqfd_disable(&ioeventfd->virqfd); 338 list_del(&ioeventfd->next); 339 vdev->ioeventfds_nr--; 340 kfree(ioeventfd); 341 ret = 0; 342 } else 343 ret = -EEXIST; 344 345 goto out_unlock; 346 } 347 } 348 349 if (fd < 0) { 350 ret = -ENODEV; 351 goto out_unlock; 352 } 353 354 if (vdev->ioeventfds_nr >= VFIO_PCI_IOEVENTFD_MAX) { 355 ret = -ENOSPC; 356 goto out_unlock; 357 } 358 359 ioeventfd = kzalloc(sizeof(*ioeventfd), GFP_KERNEL); 360 if (!ioeventfd) { 361 ret = -ENOMEM; 362 goto out_unlock; 363 } 364 365 ioeventfd->addr = vdev->barmap[bar] + pos; 366 ioeventfd->data = data; 367 ioeventfd->pos = pos; 368 ioeventfd->bar = bar; 369 ioeventfd->count = count; 370 371 ret = vfio_virqfd_enable(ioeventfd, vfio_pci_ioeventfd_handler, 372 NULL, NULL, &ioeventfd->virqfd, fd); 373 if (ret) { 374 kfree(ioeventfd); 375 goto out_unlock; 376 } 377 378 list_add(&ioeventfd->next, &vdev->ioeventfds_list); 379 vdev->ioeventfds_nr++; 380 381 out_unlock: 382 mutex_unlock(&vdev->ioeventfds_lock); 383 384 return ret; 385 } 386