1 /* 2 * VFIO PCI interrupt handling 3 * 4 * Copyright (C) 2012 Red Hat, Inc. All rights reserved. 5 * Author: Alex Williamson <alex.williamson@redhat.com> 6 * 7 * This program is free software; you can redistribute it and/or modify 8 * it under the terms of the GNU General Public License version 2 as 9 * published by the Free Software Foundation. 10 * 11 * Derived from original vfio: 12 * Copyright 2010 Cisco Systems, Inc. All rights reserved. 13 * Author: Tom Lyon, pugs@cisco.com 14 */ 15 16 #include <linux/device.h> 17 #include <linux/interrupt.h> 18 #include <linux/eventfd.h> 19 #include <linux/msi.h> 20 #include <linux/pci.h> 21 #include <linux/file.h> 22 #include <linux/vfio.h> 23 #include <linux/wait.h> 24 #include <linux/slab.h> 25 26 #include "vfio_pci_private.h" 27 28 /* 29 * INTx 30 */ 31 static void vfio_send_intx_eventfd(void *opaque, void *unused) 32 { 33 struct vfio_pci_device *vdev = opaque; 34 35 if (likely(is_intx(vdev) && !vdev->virq_disabled)) 36 eventfd_signal(vdev->ctx[0].trigger, 1); 37 } 38 39 void vfio_pci_intx_mask(struct vfio_pci_device *vdev) 40 { 41 struct pci_dev *pdev = vdev->pdev; 42 unsigned long flags; 43 44 spin_lock_irqsave(&vdev->irqlock, flags); 45 46 /* 47 * Masking can come from interrupt, ioctl, or config space 48 * via INTx disable. The latter means this can get called 49 * even when not using intx delivery. In this case, just 50 * try to have the physical bit follow the virtual bit. 51 */ 52 if (unlikely(!is_intx(vdev))) { 53 if (vdev->pci_2_3) 54 pci_intx(pdev, 0); 55 } else if (!vdev->ctx[0].masked) { 56 /* 57 * Can't use check_and_mask here because we always want to 58 * mask, not just when something is pending. 59 */ 60 if (vdev->pci_2_3) 61 pci_intx(pdev, 0); 62 else 63 disable_irq_nosync(pdev->irq); 64 65 vdev->ctx[0].masked = true; 66 } 67 68 spin_unlock_irqrestore(&vdev->irqlock, flags); 69 } 70 71 /* 72 * If this is triggered by an eventfd, we can't call eventfd_signal 73 * or else we'll deadlock on the eventfd wait queue. Return >0 when 74 * a signal is necessary, which can then be handled via a work queue 75 * or directly depending on the caller. 76 */ 77 static int vfio_pci_intx_unmask_handler(void *opaque, void *unused) 78 { 79 struct vfio_pci_device *vdev = opaque; 80 struct pci_dev *pdev = vdev->pdev; 81 unsigned long flags; 82 int ret = 0; 83 84 spin_lock_irqsave(&vdev->irqlock, flags); 85 86 /* 87 * Unmasking comes from ioctl or config, so again, have the 88 * physical bit follow the virtual even when not using INTx. 89 */ 90 if (unlikely(!is_intx(vdev))) { 91 if (vdev->pci_2_3) 92 pci_intx(pdev, 1); 93 } else if (vdev->ctx[0].masked && !vdev->virq_disabled) { 94 /* 95 * A pending interrupt here would immediately trigger, 96 * but we can avoid that overhead by just re-sending 97 * the interrupt to the user. 98 */ 99 if (vdev->pci_2_3) { 100 if (!pci_check_and_unmask_intx(pdev)) 101 ret = 1; 102 } else 103 enable_irq(pdev->irq); 104 105 vdev->ctx[0].masked = (ret > 0); 106 } 107 108 spin_unlock_irqrestore(&vdev->irqlock, flags); 109 110 return ret; 111 } 112 113 void vfio_pci_intx_unmask(struct vfio_pci_device *vdev) 114 { 115 if (vfio_pci_intx_unmask_handler(vdev, NULL) > 0) 116 vfio_send_intx_eventfd(vdev, NULL); 117 } 118 119 static irqreturn_t vfio_intx_handler(int irq, void *dev_id) 120 { 121 struct vfio_pci_device *vdev = dev_id; 122 unsigned long flags; 123 int ret = IRQ_NONE; 124 125 spin_lock_irqsave(&vdev->irqlock, flags); 126 127 if (!vdev->pci_2_3) { 128 disable_irq_nosync(vdev->pdev->irq); 129 vdev->ctx[0].masked = true; 130 ret = IRQ_HANDLED; 131 } else if (!vdev->ctx[0].masked && /* may be shared */ 132 pci_check_and_mask_intx(vdev->pdev)) { 133 vdev->ctx[0].masked = true; 134 ret = IRQ_HANDLED; 135 } 136 137 spin_unlock_irqrestore(&vdev->irqlock, flags); 138 139 if (ret == IRQ_HANDLED) 140 vfio_send_intx_eventfd(vdev, NULL); 141 142 return ret; 143 } 144 145 static int vfio_intx_enable(struct vfio_pci_device *vdev) 146 { 147 if (!is_irq_none(vdev)) 148 return -EINVAL; 149 150 if (!vdev->pdev->irq) 151 return -ENODEV; 152 153 vdev->ctx = kzalloc(sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL); 154 if (!vdev->ctx) 155 return -ENOMEM; 156 157 vdev->num_ctx = 1; 158 159 /* 160 * If the virtual interrupt is masked, restore it. Devices 161 * supporting DisINTx can be masked at the hardware level 162 * here, non-PCI-2.3 devices will have to wait until the 163 * interrupt is enabled. 164 */ 165 vdev->ctx[0].masked = vdev->virq_disabled; 166 if (vdev->pci_2_3) 167 pci_intx(vdev->pdev, !vdev->ctx[0].masked); 168 169 vdev->irq_type = VFIO_PCI_INTX_IRQ_INDEX; 170 171 return 0; 172 } 173 174 static int vfio_intx_set_signal(struct vfio_pci_device *vdev, int fd) 175 { 176 struct pci_dev *pdev = vdev->pdev; 177 unsigned long irqflags = IRQF_SHARED; 178 struct eventfd_ctx *trigger; 179 unsigned long flags; 180 int ret; 181 182 if (vdev->ctx[0].trigger) { 183 free_irq(pdev->irq, vdev); 184 kfree(vdev->ctx[0].name); 185 eventfd_ctx_put(vdev->ctx[0].trigger); 186 vdev->ctx[0].trigger = NULL; 187 } 188 189 if (fd < 0) /* Disable only */ 190 return 0; 191 192 vdev->ctx[0].name = kasprintf(GFP_KERNEL, "vfio-intx(%s)", 193 pci_name(pdev)); 194 if (!vdev->ctx[0].name) 195 return -ENOMEM; 196 197 trigger = eventfd_ctx_fdget(fd); 198 if (IS_ERR(trigger)) { 199 kfree(vdev->ctx[0].name); 200 return PTR_ERR(trigger); 201 } 202 203 vdev->ctx[0].trigger = trigger; 204 205 if (!vdev->pci_2_3) 206 irqflags = 0; 207 208 ret = request_irq(pdev->irq, vfio_intx_handler, 209 irqflags, vdev->ctx[0].name, vdev); 210 if (ret) { 211 vdev->ctx[0].trigger = NULL; 212 kfree(vdev->ctx[0].name); 213 eventfd_ctx_put(trigger); 214 return ret; 215 } 216 217 /* 218 * INTx disable will stick across the new irq setup, 219 * disable_irq won't. 220 */ 221 spin_lock_irqsave(&vdev->irqlock, flags); 222 if (!vdev->pci_2_3 && vdev->ctx[0].masked) 223 disable_irq_nosync(pdev->irq); 224 spin_unlock_irqrestore(&vdev->irqlock, flags); 225 226 return 0; 227 } 228 229 static void vfio_intx_disable(struct vfio_pci_device *vdev) 230 { 231 vfio_virqfd_disable(&vdev->ctx[0].unmask); 232 vfio_virqfd_disable(&vdev->ctx[0].mask); 233 vfio_intx_set_signal(vdev, -1); 234 vdev->irq_type = VFIO_PCI_NUM_IRQS; 235 vdev->num_ctx = 0; 236 kfree(vdev->ctx); 237 } 238 239 /* 240 * MSI/MSI-X 241 */ 242 static irqreturn_t vfio_msihandler(int irq, void *arg) 243 { 244 struct eventfd_ctx *trigger = arg; 245 246 eventfd_signal(trigger, 1); 247 return IRQ_HANDLED; 248 } 249 250 static int vfio_msi_enable(struct vfio_pci_device *vdev, int nvec, bool msix) 251 { 252 struct pci_dev *pdev = vdev->pdev; 253 int ret; 254 255 if (!is_irq_none(vdev)) 256 return -EINVAL; 257 258 vdev->ctx = kzalloc(nvec * sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL); 259 if (!vdev->ctx) 260 return -ENOMEM; 261 262 if (msix) { 263 int i; 264 265 vdev->msix = kzalloc(nvec * sizeof(struct msix_entry), 266 GFP_KERNEL); 267 if (!vdev->msix) { 268 kfree(vdev->ctx); 269 return -ENOMEM; 270 } 271 272 for (i = 0; i < nvec; i++) 273 vdev->msix[i].entry = i; 274 275 ret = pci_enable_msix_range(pdev, vdev->msix, 1, nvec); 276 if (ret < nvec) { 277 if (ret > 0) 278 pci_disable_msix(pdev); 279 kfree(vdev->msix); 280 kfree(vdev->ctx); 281 return ret; 282 } 283 } else { 284 ret = pci_enable_msi_range(pdev, 1, nvec); 285 if (ret < nvec) { 286 if (ret > 0) 287 pci_disable_msi(pdev); 288 kfree(vdev->ctx); 289 return ret; 290 } 291 } 292 293 vdev->num_ctx = nvec; 294 vdev->irq_type = msix ? VFIO_PCI_MSIX_IRQ_INDEX : 295 VFIO_PCI_MSI_IRQ_INDEX; 296 297 if (!msix) { 298 /* 299 * Compute the virtual hardware field for max msi vectors - 300 * it is the log base 2 of the number of vectors. 301 */ 302 vdev->msi_qmax = fls(nvec * 2 - 1) - 1; 303 } 304 305 return 0; 306 } 307 308 static int vfio_msi_set_vector_signal(struct vfio_pci_device *vdev, 309 int vector, int fd, bool msix) 310 { 311 struct pci_dev *pdev = vdev->pdev; 312 struct eventfd_ctx *trigger; 313 int irq, ret; 314 315 if (vector < 0 || vector >= vdev->num_ctx) 316 return -EINVAL; 317 318 irq = msix ? vdev->msix[vector].vector : pdev->irq + vector; 319 320 if (vdev->ctx[vector].trigger) { 321 free_irq(irq, vdev->ctx[vector].trigger); 322 irq_bypass_unregister_producer(&vdev->ctx[vector].producer); 323 kfree(vdev->ctx[vector].name); 324 eventfd_ctx_put(vdev->ctx[vector].trigger); 325 vdev->ctx[vector].trigger = NULL; 326 } 327 328 if (fd < 0) 329 return 0; 330 331 vdev->ctx[vector].name = kasprintf(GFP_KERNEL, "vfio-msi%s[%d](%s)", 332 msix ? "x" : "", vector, 333 pci_name(pdev)); 334 if (!vdev->ctx[vector].name) 335 return -ENOMEM; 336 337 trigger = eventfd_ctx_fdget(fd); 338 if (IS_ERR(trigger)) { 339 kfree(vdev->ctx[vector].name); 340 return PTR_ERR(trigger); 341 } 342 343 /* 344 * The MSIx vector table resides in device memory which may be cleared 345 * via backdoor resets. We don't allow direct access to the vector 346 * table so even if a userspace driver attempts to save/restore around 347 * such a reset it would be unsuccessful. To avoid this, restore the 348 * cached value of the message prior to enabling. 349 */ 350 if (msix) { 351 struct msi_msg msg; 352 353 get_cached_msi_msg(irq, &msg); 354 pci_write_msi_msg(irq, &msg); 355 } 356 357 ret = request_irq(irq, vfio_msihandler, 0, 358 vdev->ctx[vector].name, trigger); 359 if (ret) { 360 kfree(vdev->ctx[vector].name); 361 eventfd_ctx_put(trigger); 362 return ret; 363 } 364 365 vdev->ctx[vector].producer.token = trigger; 366 vdev->ctx[vector].producer.irq = irq; 367 ret = irq_bypass_register_producer(&vdev->ctx[vector].producer); 368 if (unlikely(ret)) 369 dev_info(&pdev->dev, 370 "irq bypass producer (token %p) registration fails: %d\n", 371 vdev->ctx[vector].producer.token, ret); 372 373 vdev->ctx[vector].trigger = trigger; 374 375 return 0; 376 } 377 378 static int vfio_msi_set_block(struct vfio_pci_device *vdev, unsigned start, 379 unsigned count, int32_t *fds, bool msix) 380 { 381 int i, j, ret = 0; 382 383 if (start >= vdev->num_ctx || start + count > vdev->num_ctx) 384 return -EINVAL; 385 386 for (i = 0, j = start; i < count && !ret; i++, j++) { 387 int fd = fds ? fds[i] : -1; 388 ret = vfio_msi_set_vector_signal(vdev, j, fd, msix); 389 } 390 391 if (ret) { 392 for (--j; j >= (int)start; j--) 393 vfio_msi_set_vector_signal(vdev, j, -1, msix); 394 } 395 396 return ret; 397 } 398 399 static void vfio_msi_disable(struct vfio_pci_device *vdev, bool msix) 400 { 401 struct pci_dev *pdev = vdev->pdev; 402 int i; 403 404 for (i = 0; i < vdev->num_ctx; i++) { 405 vfio_virqfd_disable(&vdev->ctx[i].unmask); 406 vfio_virqfd_disable(&vdev->ctx[i].mask); 407 } 408 409 vfio_msi_set_block(vdev, 0, vdev->num_ctx, NULL, msix); 410 411 if (msix) { 412 pci_disable_msix(vdev->pdev); 413 kfree(vdev->msix); 414 } else 415 pci_disable_msi(pdev); 416 417 vdev->irq_type = VFIO_PCI_NUM_IRQS; 418 vdev->num_ctx = 0; 419 kfree(vdev->ctx); 420 } 421 422 /* 423 * IOCTL support 424 */ 425 static int vfio_pci_set_intx_unmask(struct vfio_pci_device *vdev, 426 unsigned index, unsigned start, 427 unsigned count, uint32_t flags, void *data) 428 { 429 if (!is_intx(vdev) || start != 0 || count != 1) 430 return -EINVAL; 431 432 if (flags & VFIO_IRQ_SET_DATA_NONE) { 433 vfio_pci_intx_unmask(vdev); 434 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 435 uint8_t unmask = *(uint8_t *)data; 436 if (unmask) 437 vfio_pci_intx_unmask(vdev); 438 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) { 439 int32_t fd = *(int32_t *)data; 440 if (fd >= 0) 441 return vfio_virqfd_enable((void *) vdev, 442 vfio_pci_intx_unmask_handler, 443 vfio_send_intx_eventfd, NULL, 444 &vdev->ctx[0].unmask, fd); 445 446 vfio_virqfd_disable(&vdev->ctx[0].unmask); 447 } 448 449 return 0; 450 } 451 452 static int vfio_pci_set_intx_mask(struct vfio_pci_device *vdev, 453 unsigned index, unsigned start, 454 unsigned count, uint32_t flags, void *data) 455 { 456 if (!is_intx(vdev) || start != 0 || count != 1) 457 return -EINVAL; 458 459 if (flags & VFIO_IRQ_SET_DATA_NONE) { 460 vfio_pci_intx_mask(vdev); 461 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 462 uint8_t mask = *(uint8_t *)data; 463 if (mask) 464 vfio_pci_intx_mask(vdev); 465 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) { 466 return -ENOTTY; /* XXX implement me */ 467 } 468 469 return 0; 470 } 471 472 static int vfio_pci_set_intx_trigger(struct vfio_pci_device *vdev, 473 unsigned index, unsigned start, 474 unsigned count, uint32_t flags, void *data) 475 { 476 if (is_intx(vdev) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) { 477 vfio_intx_disable(vdev); 478 return 0; 479 } 480 481 if (!(is_intx(vdev) || is_irq_none(vdev)) || start != 0 || count != 1) 482 return -EINVAL; 483 484 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) { 485 int32_t fd = *(int32_t *)data; 486 int ret; 487 488 if (is_intx(vdev)) 489 return vfio_intx_set_signal(vdev, fd); 490 491 ret = vfio_intx_enable(vdev); 492 if (ret) 493 return ret; 494 495 ret = vfio_intx_set_signal(vdev, fd); 496 if (ret) 497 vfio_intx_disable(vdev); 498 499 return ret; 500 } 501 502 if (!is_intx(vdev)) 503 return -EINVAL; 504 505 if (flags & VFIO_IRQ_SET_DATA_NONE) { 506 vfio_send_intx_eventfd(vdev, NULL); 507 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 508 uint8_t trigger = *(uint8_t *)data; 509 if (trigger) 510 vfio_send_intx_eventfd(vdev, NULL); 511 } 512 return 0; 513 } 514 515 static int vfio_pci_set_msi_trigger(struct vfio_pci_device *vdev, 516 unsigned index, unsigned start, 517 unsigned count, uint32_t flags, void *data) 518 { 519 int i; 520 bool msix = (index == VFIO_PCI_MSIX_IRQ_INDEX) ? true : false; 521 522 if (irq_is(vdev, index) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) { 523 vfio_msi_disable(vdev, msix); 524 return 0; 525 } 526 527 if (!(irq_is(vdev, index) || is_irq_none(vdev))) 528 return -EINVAL; 529 530 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) { 531 int32_t *fds = data; 532 int ret; 533 534 if (vdev->irq_type == index) 535 return vfio_msi_set_block(vdev, start, count, 536 fds, msix); 537 538 ret = vfio_msi_enable(vdev, start + count, msix); 539 if (ret) 540 return ret; 541 542 ret = vfio_msi_set_block(vdev, start, count, fds, msix); 543 if (ret) 544 vfio_msi_disable(vdev, msix); 545 546 return ret; 547 } 548 549 if (!irq_is(vdev, index) || start + count > vdev->num_ctx) 550 return -EINVAL; 551 552 for (i = start; i < start + count; i++) { 553 if (!vdev->ctx[i].trigger) 554 continue; 555 if (flags & VFIO_IRQ_SET_DATA_NONE) { 556 eventfd_signal(vdev->ctx[i].trigger, 1); 557 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 558 uint8_t *bools = data; 559 if (bools[i - start]) 560 eventfd_signal(vdev->ctx[i].trigger, 1); 561 } 562 } 563 return 0; 564 } 565 566 static int vfio_pci_set_ctx_trigger_single(struct eventfd_ctx **ctx, 567 uint32_t flags, void *data) 568 { 569 int32_t fd = *(int32_t *)data; 570 571 if (!(flags & VFIO_IRQ_SET_DATA_TYPE_MASK)) 572 return -EINVAL; 573 574 /* DATA_NONE/DATA_BOOL enables loopback testing */ 575 if (flags & VFIO_IRQ_SET_DATA_NONE) { 576 if (*ctx) 577 eventfd_signal(*ctx, 1); 578 return 0; 579 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 580 uint8_t trigger = *(uint8_t *)data; 581 if (trigger && *ctx) 582 eventfd_signal(*ctx, 1); 583 return 0; 584 } 585 586 /* Handle SET_DATA_EVENTFD */ 587 if (fd == -1) { 588 if (*ctx) 589 eventfd_ctx_put(*ctx); 590 *ctx = NULL; 591 return 0; 592 } else if (fd >= 0) { 593 struct eventfd_ctx *efdctx; 594 efdctx = eventfd_ctx_fdget(fd); 595 if (IS_ERR(efdctx)) 596 return PTR_ERR(efdctx); 597 if (*ctx) 598 eventfd_ctx_put(*ctx); 599 *ctx = efdctx; 600 return 0; 601 } else 602 return -EINVAL; 603 } 604 605 static int vfio_pci_set_err_trigger(struct vfio_pci_device *vdev, 606 unsigned index, unsigned start, 607 unsigned count, uint32_t flags, void *data) 608 { 609 if (index != VFIO_PCI_ERR_IRQ_INDEX) 610 return -EINVAL; 611 612 /* 613 * We should sanitize start & count, but that wasn't caught 614 * originally, so this IRQ index must forever ignore them :-( 615 */ 616 617 return vfio_pci_set_ctx_trigger_single(&vdev->err_trigger, flags, data); 618 } 619 620 static int vfio_pci_set_req_trigger(struct vfio_pci_device *vdev, 621 unsigned index, unsigned start, 622 unsigned count, uint32_t flags, void *data) 623 { 624 if (index != VFIO_PCI_REQ_IRQ_INDEX || start != 0 || count != 1) 625 return -EINVAL; 626 627 return vfio_pci_set_ctx_trigger_single(&vdev->req_trigger, flags, data); 628 } 629 630 int vfio_pci_set_irqs_ioctl(struct vfio_pci_device *vdev, uint32_t flags, 631 unsigned index, unsigned start, unsigned count, 632 void *data) 633 { 634 int (*func)(struct vfio_pci_device *vdev, unsigned index, 635 unsigned start, unsigned count, uint32_t flags, 636 void *data) = NULL; 637 638 switch (index) { 639 case VFIO_PCI_INTX_IRQ_INDEX: 640 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) { 641 case VFIO_IRQ_SET_ACTION_MASK: 642 func = vfio_pci_set_intx_mask; 643 break; 644 case VFIO_IRQ_SET_ACTION_UNMASK: 645 func = vfio_pci_set_intx_unmask; 646 break; 647 case VFIO_IRQ_SET_ACTION_TRIGGER: 648 func = vfio_pci_set_intx_trigger; 649 break; 650 } 651 break; 652 case VFIO_PCI_MSI_IRQ_INDEX: 653 case VFIO_PCI_MSIX_IRQ_INDEX: 654 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) { 655 case VFIO_IRQ_SET_ACTION_MASK: 656 case VFIO_IRQ_SET_ACTION_UNMASK: 657 /* XXX Need masking support exported */ 658 break; 659 case VFIO_IRQ_SET_ACTION_TRIGGER: 660 func = vfio_pci_set_msi_trigger; 661 break; 662 } 663 break; 664 case VFIO_PCI_ERR_IRQ_INDEX: 665 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) { 666 case VFIO_IRQ_SET_ACTION_TRIGGER: 667 if (pci_is_pcie(vdev->pdev)) 668 func = vfio_pci_set_err_trigger; 669 break; 670 } 671 break; 672 case VFIO_PCI_REQ_IRQ_INDEX: 673 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) { 674 case VFIO_IRQ_SET_ACTION_TRIGGER: 675 func = vfio_pci_set_req_trigger; 676 break; 677 } 678 break; 679 } 680 681 if (!func) 682 return -ENOTTY; 683 684 return func(vdev, index, start, count, flags, data); 685 } 686