1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * VFIO PCI interrupt handling 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/device.h> 14 #include <linux/interrupt.h> 15 #include <linux/eventfd.h> 16 #include <linux/msi.h> 17 #include <linux/pci.h> 18 #include <linux/file.h> 19 #include <linux/vfio.h> 20 #include <linux/wait.h> 21 #include <linux/slab.h> 22 23 #include <linux/vfio_pci_core.h> 24 25 /* 26 * INTx 27 */ 28 static void vfio_send_intx_eventfd(void *opaque, void *unused) 29 { 30 struct vfio_pci_core_device *vdev = opaque; 31 32 if (likely(is_intx(vdev) && !vdev->virq_disabled)) 33 eventfd_signal(vdev->ctx[0].trigger, 1); 34 } 35 36 void vfio_pci_intx_mask(struct vfio_pci_core_device *vdev) 37 { 38 struct pci_dev *pdev = vdev->pdev; 39 unsigned long flags; 40 41 spin_lock_irqsave(&vdev->irqlock, flags); 42 43 /* 44 * Masking can come from interrupt, ioctl, or config space 45 * via INTx disable. The latter means this can get called 46 * even when not using intx delivery. In this case, just 47 * try to have the physical bit follow the virtual bit. 48 */ 49 if (unlikely(!is_intx(vdev))) { 50 if (vdev->pci_2_3) 51 pci_intx(pdev, 0); 52 } else if (!vdev->ctx[0].masked) { 53 /* 54 * Can't use check_and_mask here because we always want to 55 * mask, not just when something is pending. 56 */ 57 if (vdev->pci_2_3) 58 pci_intx(pdev, 0); 59 else 60 disable_irq_nosync(pdev->irq); 61 62 vdev->ctx[0].masked = true; 63 } 64 65 spin_unlock_irqrestore(&vdev->irqlock, flags); 66 } 67 68 /* 69 * If this is triggered by an eventfd, we can't call eventfd_signal 70 * or else we'll deadlock on the eventfd wait queue. Return >0 when 71 * a signal is necessary, which can then be handled via a work queue 72 * or directly depending on the caller. 73 */ 74 static int vfio_pci_intx_unmask_handler(void *opaque, void *unused) 75 { 76 struct vfio_pci_core_device *vdev = opaque; 77 struct pci_dev *pdev = vdev->pdev; 78 unsigned long flags; 79 int ret = 0; 80 81 spin_lock_irqsave(&vdev->irqlock, flags); 82 83 /* 84 * Unmasking comes from ioctl or config, so again, have the 85 * physical bit follow the virtual even when not using INTx. 86 */ 87 if (unlikely(!is_intx(vdev))) { 88 if (vdev->pci_2_3) 89 pci_intx(pdev, 1); 90 } else if (vdev->ctx[0].masked && !vdev->virq_disabled) { 91 /* 92 * A pending interrupt here would immediately trigger, 93 * but we can avoid that overhead by just re-sending 94 * the interrupt to the user. 95 */ 96 if (vdev->pci_2_3) { 97 if (!pci_check_and_unmask_intx(pdev)) 98 ret = 1; 99 } else 100 enable_irq(pdev->irq); 101 102 vdev->ctx[0].masked = (ret > 0); 103 } 104 105 spin_unlock_irqrestore(&vdev->irqlock, flags); 106 107 return ret; 108 } 109 110 void vfio_pci_intx_unmask(struct vfio_pci_core_device *vdev) 111 { 112 if (vfio_pci_intx_unmask_handler(vdev, NULL) > 0) 113 vfio_send_intx_eventfd(vdev, NULL); 114 } 115 116 static irqreturn_t vfio_intx_handler(int irq, void *dev_id) 117 { 118 struct vfio_pci_core_device *vdev = dev_id; 119 unsigned long flags; 120 int ret = IRQ_NONE; 121 122 spin_lock_irqsave(&vdev->irqlock, flags); 123 124 if (!vdev->pci_2_3) { 125 disable_irq_nosync(vdev->pdev->irq); 126 vdev->ctx[0].masked = true; 127 ret = IRQ_HANDLED; 128 } else if (!vdev->ctx[0].masked && /* may be shared */ 129 pci_check_and_mask_intx(vdev->pdev)) { 130 vdev->ctx[0].masked = true; 131 ret = IRQ_HANDLED; 132 } 133 134 spin_unlock_irqrestore(&vdev->irqlock, flags); 135 136 if (ret == IRQ_HANDLED) 137 vfio_send_intx_eventfd(vdev, NULL); 138 139 return ret; 140 } 141 142 static int vfio_intx_enable(struct vfio_pci_core_device *vdev) 143 { 144 if (!is_irq_none(vdev)) 145 return -EINVAL; 146 147 if (!vdev->pdev->irq) 148 return -ENODEV; 149 150 vdev->ctx = kzalloc(sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL); 151 if (!vdev->ctx) 152 return -ENOMEM; 153 154 vdev->num_ctx = 1; 155 156 /* 157 * If the virtual interrupt is masked, restore it. Devices 158 * supporting DisINTx can be masked at the hardware level 159 * here, non-PCI-2.3 devices will have to wait until the 160 * interrupt is enabled. 161 */ 162 vdev->ctx[0].masked = vdev->virq_disabled; 163 if (vdev->pci_2_3) 164 pci_intx(vdev->pdev, !vdev->ctx[0].masked); 165 166 vdev->irq_type = VFIO_PCI_INTX_IRQ_INDEX; 167 168 return 0; 169 } 170 171 static int vfio_intx_set_signal(struct vfio_pci_core_device *vdev, int fd) 172 { 173 struct pci_dev *pdev = vdev->pdev; 174 unsigned long irqflags = IRQF_SHARED; 175 struct eventfd_ctx *trigger; 176 unsigned long flags; 177 int ret; 178 179 if (vdev->ctx[0].trigger) { 180 free_irq(pdev->irq, vdev); 181 kfree(vdev->ctx[0].name); 182 eventfd_ctx_put(vdev->ctx[0].trigger); 183 vdev->ctx[0].trigger = NULL; 184 } 185 186 if (fd < 0) /* Disable only */ 187 return 0; 188 189 vdev->ctx[0].name = kasprintf(GFP_KERNEL, "vfio-intx(%s)", 190 pci_name(pdev)); 191 if (!vdev->ctx[0].name) 192 return -ENOMEM; 193 194 trigger = eventfd_ctx_fdget(fd); 195 if (IS_ERR(trigger)) { 196 kfree(vdev->ctx[0].name); 197 return PTR_ERR(trigger); 198 } 199 200 vdev->ctx[0].trigger = trigger; 201 202 if (!vdev->pci_2_3) 203 irqflags = 0; 204 205 ret = request_irq(pdev->irq, vfio_intx_handler, 206 irqflags, vdev->ctx[0].name, vdev); 207 if (ret) { 208 vdev->ctx[0].trigger = NULL; 209 kfree(vdev->ctx[0].name); 210 eventfd_ctx_put(trigger); 211 return ret; 212 } 213 214 /* 215 * INTx disable will stick across the new irq setup, 216 * disable_irq won't. 217 */ 218 spin_lock_irqsave(&vdev->irqlock, flags); 219 if (!vdev->pci_2_3 && vdev->ctx[0].masked) 220 disable_irq_nosync(pdev->irq); 221 spin_unlock_irqrestore(&vdev->irqlock, flags); 222 223 return 0; 224 } 225 226 static void vfio_intx_disable(struct vfio_pci_core_device *vdev) 227 { 228 vfio_virqfd_disable(&vdev->ctx[0].unmask); 229 vfio_virqfd_disable(&vdev->ctx[0].mask); 230 vfio_intx_set_signal(vdev, -1); 231 vdev->irq_type = VFIO_PCI_NUM_IRQS; 232 vdev->num_ctx = 0; 233 kfree(vdev->ctx); 234 } 235 236 /* 237 * MSI/MSI-X 238 */ 239 static irqreturn_t vfio_msihandler(int irq, void *arg) 240 { 241 struct eventfd_ctx *trigger = arg; 242 243 eventfd_signal(trigger, 1); 244 return IRQ_HANDLED; 245 } 246 247 static int vfio_msi_enable(struct vfio_pci_core_device *vdev, int nvec, bool msix) 248 { 249 struct pci_dev *pdev = vdev->pdev; 250 unsigned int flag = msix ? PCI_IRQ_MSIX : PCI_IRQ_MSI; 251 int ret; 252 u16 cmd; 253 254 if (!is_irq_none(vdev)) 255 return -EINVAL; 256 257 vdev->ctx = kcalloc(nvec, sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL); 258 if (!vdev->ctx) 259 return -ENOMEM; 260 261 /* return the number of supported vectors if we can't get all: */ 262 cmd = vfio_pci_memory_lock_and_enable(vdev); 263 ret = pci_alloc_irq_vectors(pdev, 1, nvec, flag); 264 if (ret < nvec) { 265 if (ret > 0) 266 pci_free_irq_vectors(pdev); 267 vfio_pci_memory_unlock_and_restore(vdev, cmd); 268 kfree(vdev->ctx); 269 return ret; 270 } 271 vfio_pci_memory_unlock_and_restore(vdev, cmd); 272 273 vdev->num_ctx = nvec; 274 vdev->irq_type = msix ? VFIO_PCI_MSIX_IRQ_INDEX : 275 VFIO_PCI_MSI_IRQ_INDEX; 276 277 if (!msix) { 278 /* 279 * Compute the virtual hardware field for max msi vectors - 280 * it is the log base 2 of the number of vectors. 281 */ 282 vdev->msi_qmax = fls(nvec * 2 - 1) - 1; 283 } 284 285 return 0; 286 } 287 288 static int vfio_msi_set_vector_signal(struct vfio_pci_core_device *vdev, 289 int vector, int fd, bool msix) 290 { 291 struct pci_dev *pdev = vdev->pdev; 292 struct eventfd_ctx *trigger; 293 int irq, ret; 294 u16 cmd; 295 296 if (vector < 0 || vector >= vdev->num_ctx) 297 return -EINVAL; 298 299 irq = pci_irq_vector(pdev, vector); 300 301 if (vdev->ctx[vector].trigger) { 302 irq_bypass_unregister_producer(&vdev->ctx[vector].producer); 303 304 cmd = vfio_pci_memory_lock_and_enable(vdev); 305 free_irq(irq, vdev->ctx[vector].trigger); 306 vfio_pci_memory_unlock_and_restore(vdev, cmd); 307 308 kfree(vdev->ctx[vector].name); 309 eventfd_ctx_put(vdev->ctx[vector].trigger); 310 vdev->ctx[vector].trigger = NULL; 311 } 312 313 if (fd < 0) 314 return 0; 315 316 vdev->ctx[vector].name = kasprintf(GFP_KERNEL, "vfio-msi%s[%d](%s)", 317 msix ? "x" : "", vector, 318 pci_name(pdev)); 319 if (!vdev->ctx[vector].name) 320 return -ENOMEM; 321 322 trigger = eventfd_ctx_fdget(fd); 323 if (IS_ERR(trigger)) { 324 kfree(vdev->ctx[vector].name); 325 return PTR_ERR(trigger); 326 } 327 328 /* 329 * The MSIx vector table resides in device memory which may be cleared 330 * via backdoor resets. We don't allow direct access to the vector 331 * table so even if a userspace driver attempts to save/restore around 332 * such a reset it would be unsuccessful. To avoid this, restore the 333 * cached value of the message prior to enabling. 334 */ 335 cmd = vfio_pci_memory_lock_and_enable(vdev); 336 if (msix) { 337 struct msi_msg msg; 338 339 get_cached_msi_msg(irq, &msg); 340 pci_write_msi_msg(irq, &msg); 341 } 342 343 ret = request_irq(irq, vfio_msihandler, 0, 344 vdev->ctx[vector].name, trigger); 345 vfio_pci_memory_unlock_and_restore(vdev, cmd); 346 if (ret) { 347 kfree(vdev->ctx[vector].name); 348 eventfd_ctx_put(trigger); 349 return ret; 350 } 351 352 vdev->ctx[vector].producer.token = trigger; 353 vdev->ctx[vector].producer.irq = irq; 354 ret = irq_bypass_register_producer(&vdev->ctx[vector].producer); 355 if (unlikely(ret)) { 356 dev_info(&pdev->dev, 357 "irq bypass producer (token %p) registration fails: %d\n", 358 vdev->ctx[vector].producer.token, ret); 359 360 vdev->ctx[vector].producer.token = NULL; 361 } 362 vdev->ctx[vector].trigger = trigger; 363 364 return 0; 365 } 366 367 static int vfio_msi_set_block(struct vfio_pci_core_device *vdev, unsigned start, 368 unsigned count, int32_t *fds, bool msix) 369 { 370 int i, j, ret = 0; 371 372 if (start >= vdev->num_ctx || start + count > vdev->num_ctx) 373 return -EINVAL; 374 375 for (i = 0, j = start; i < count && !ret; i++, j++) { 376 int fd = fds ? fds[i] : -1; 377 ret = vfio_msi_set_vector_signal(vdev, j, fd, msix); 378 } 379 380 if (ret) { 381 for (--j; j >= (int)start; j--) 382 vfio_msi_set_vector_signal(vdev, j, -1, msix); 383 } 384 385 return ret; 386 } 387 388 static void vfio_msi_disable(struct vfio_pci_core_device *vdev, bool msix) 389 { 390 struct pci_dev *pdev = vdev->pdev; 391 int i; 392 u16 cmd; 393 394 for (i = 0; i < vdev->num_ctx; i++) { 395 vfio_virqfd_disable(&vdev->ctx[i].unmask); 396 vfio_virqfd_disable(&vdev->ctx[i].mask); 397 } 398 399 vfio_msi_set_block(vdev, 0, vdev->num_ctx, NULL, msix); 400 401 cmd = vfio_pci_memory_lock_and_enable(vdev); 402 pci_free_irq_vectors(pdev); 403 vfio_pci_memory_unlock_and_restore(vdev, cmd); 404 405 /* 406 * Both disable paths above use pci_intx_for_msi() to clear DisINTx 407 * via their shutdown paths. Restore for NoINTx devices. 408 */ 409 if (vdev->nointx) 410 pci_intx(pdev, 0); 411 412 vdev->irq_type = VFIO_PCI_NUM_IRQS; 413 vdev->num_ctx = 0; 414 kfree(vdev->ctx); 415 } 416 417 /* 418 * IOCTL support 419 */ 420 static int vfio_pci_set_intx_unmask(struct vfio_pci_core_device *vdev, 421 unsigned index, unsigned start, 422 unsigned count, uint32_t flags, void *data) 423 { 424 if (!is_intx(vdev) || start != 0 || count != 1) 425 return -EINVAL; 426 427 if (flags & VFIO_IRQ_SET_DATA_NONE) { 428 vfio_pci_intx_unmask(vdev); 429 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 430 uint8_t unmask = *(uint8_t *)data; 431 if (unmask) 432 vfio_pci_intx_unmask(vdev); 433 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) { 434 int32_t fd = *(int32_t *)data; 435 if (fd >= 0) 436 return vfio_virqfd_enable((void *) vdev, 437 vfio_pci_intx_unmask_handler, 438 vfio_send_intx_eventfd, NULL, 439 &vdev->ctx[0].unmask, fd); 440 441 vfio_virqfd_disable(&vdev->ctx[0].unmask); 442 } 443 444 return 0; 445 } 446 447 static int vfio_pci_set_intx_mask(struct vfio_pci_core_device *vdev, 448 unsigned index, unsigned start, 449 unsigned count, uint32_t flags, void *data) 450 { 451 if (!is_intx(vdev) || start != 0 || count != 1) 452 return -EINVAL; 453 454 if (flags & VFIO_IRQ_SET_DATA_NONE) { 455 vfio_pci_intx_mask(vdev); 456 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 457 uint8_t mask = *(uint8_t *)data; 458 if (mask) 459 vfio_pci_intx_mask(vdev); 460 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) { 461 return -ENOTTY; /* XXX implement me */ 462 } 463 464 return 0; 465 } 466 467 static int vfio_pci_set_intx_trigger(struct vfio_pci_core_device *vdev, 468 unsigned index, unsigned start, 469 unsigned count, uint32_t flags, void *data) 470 { 471 if (is_intx(vdev) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) { 472 vfio_intx_disable(vdev); 473 return 0; 474 } 475 476 if (!(is_intx(vdev) || is_irq_none(vdev)) || start != 0 || count != 1) 477 return -EINVAL; 478 479 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) { 480 int32_t fd = *(int32_t *)data; 481 int ret; 482 483 if (is_intx(vdev)) 484 return vfio_intx_set_signal(vdev, fd); 485 486 ret = vfio_intx_enable(vdev); 487 if (ret) 488 return ret; 489 490 ret = vfio_intx_set_signal(vdev, fd); 491 if (ret) 492 vfio_intx_disable(vdev); 493 494 return ret; 495 } 496 497 if (!is_intx(vdev)) 498 return -EINVAL; 499 500 if (flags & VFIO_IRQ_SET_DATA_NONE) { 501 vfio_send_intx_eventfd(vdev, NULL); 502 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 503 uint8_t trigger = *(uint8_t *)data; 504 if (trigger) 505 vfio_send_intx_eventfd(vdev, NULL); 506 } 507 return 0; 508 } 509 510 static int vfio_pci_set_msi_trigger(struct vfio_pci_core_device *vdev, 511 unsigned index, unsigned start, 512 unsigned count, uint32_t flags, void *data) 513 { 514 int i; 515 bool msix = (index == VFIO_PCI_MSIX_IRQ_INDEX) ? true : false; 516 517 if (irq_is(vdev, index) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) { 518 vfio_msi_disable(vdev, msix); 519 return 0; 520 } 521 522 if (!(irq_is(vdev, index) || is_irq_none(vdev))) 523 return -EINVAL; 524 525 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) { 526 int32_t *fds = data; 527 int ret; 528 529 if (vdev->irq_type == index) 530 return vfio_msi_set_block(vdev, start, count, 531 fds, msix); 532 533 ret = vfio_msi_enable(vdev, start + count, msix); 534 if (ret) 535 return ret; 536 537 ret = vfio_msi_set_block(vdev, start, count, fds, msix); 538 if (ret) 539 vfio_msi_disable(vdev, msix); 540 541 return ret; 542 } 543 544 if (!irq_is(vdev, index) || start + count > vdev->num_ctx) 545 return -EINVAL; 546 547 for (i = start; i < start + count; i++) { 548 if (!vdev->ctx[i].trigger) 549 continue; 550 if (flags & VFIO_IRQ_SET_DATA_NONE) { 551 eventfd_signal(vdev->ctx[i].trigger, 1); 552 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 553 uint8_t *bools = data; 554 if (bools[i - start]) 555 eventfd_signal(vdev->ctx[i].trigger, 1); 556 } 557 } 558 return 0; 559 } 560 561 static int vfio_pci_set_ctx_trigger_single(struct eventfd_ctx **ctx, 562 unsigned int count, uint32_t flags, 563 void *data) 564 { 565 /* DATA_NONE/DATA_BOOL enables loopback testing */ 566 if (flags & VFIO_IRQ_SET_DATA_NONE) { 567 if (*ctx) { 568 if (count) { 569 eventfd_signal(*ctx, 1); 570 } else { 571 eventfd_ctx_put(*ctx); 572 *ctx = NULL; 573 } 574 return 0; 575 } 576 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) { 577 uint8_t trigger; 578 579 if (!count) 580 return -EINVAL; 581 582 trigger = *(uint8_t *)data; 583 if (trigger && *ctx) 584 eventfd_signal(*ctx, 1); 585 586 return 0; 587 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) { 588 int32_t fd; 589 590 if (!count) 591 return -EINVAL; 592 593 fd = *(int32_t *)data; 594 if (fd == -1) { 595 if (*ctx) 596 eventfd_ctx_put(*ctx); 597 *ctx = NULL; 598 } else if (fd >= 0) { 599 struct eventfd_ctx *efdctx; 600 601 efdctx = eventfd_ctx_fdget(fd); 602 if (IS_ERR(efdctx)) 603 return PTR_ERR(efdctx); 604 605 if (*ctx) 606 eventfd_ctx_put(*ctx); 607 608 *ctx = efdctx; 609 } 610 return 0; 611 } 612 613 return -EINVAL; 614 } 615 616 static int vfio_pci_set_err_trigger(struct vfio_pci_core_device *vdev, 617 unsigned index, unsigned start, 618 unsigned count, uint32_t flags, void *data) 619 { 620 if (index != VFIO_PCI_ERR_IRQ_INDEX || start != 0 || count > 1) 621 return -EINVAL; 622 623 return vfio_pci_set_ctx_trigger_single(&vdev->err_trigger, 624 count, flags, data); 625 } 626 627 static int vfio_pci_set_req_trigger(struct vfio_pci_core_device *vdev, 628 unsigned index, unsigned start, 629 unsigned count, uint32_t flags, void *data) 630 { 631 if (index != VFIO_PCI_REQ_IRQ_INDEX || start != 0 || count > 1) 632 return -EINVAL; 633 634 return vfio_pci_set_ctx_trigger_single(&vdev->req_trigger, 635 count, flags, data); 636 } 637 638 int vfio_pci_set_irqs_ioctl(struct vfio_pci_core_device *vdev, uint32_t flags, 639 unsigned index, unsigned start, unsigned count, 640 void *data) 641 { 642 int (*func)(struct vfio_pci_core_device *vdev, unsigned index, 643 unsigned start, unsigned count, uint32_t flags, 644 void *data) = NULL; 645 646 switch (index) { 647 case VFIO_PCI_INTX_IRQ_INDEX: 648 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) { 649 case VFIO_IRQ_SET_ACTION_MASK: 650 func = vfio_pci_set_intx_mask; 651 break; 652 case VFIO_IRQ_SET_ACTION_UNMASK: 653 func = vfio_pci_set_intx_unmask; 654 break; 655 case VFIO_IRQ_SET_ACTION_TRIGGER: 656 func = vfio_pci_set_intx_trigger; 657 break; 658 } 659 break; 660 case VFIO_PCI_MSI_IRQ_INDEX: 661 case VFIO_PCI_MSIX_IRQ_INDEX: 662 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) { 663 case VFIO_IRQ_SET_ACTION_MASK: 664 case VFIO_IRQ_SET_ACTION_UNMASK: 665 /* XXX Need masking support exported */ 666 break; 667 case VFIO_IRQ_SET_ACTION_TRIGGER: 668 func = vfio_pci_set_msi_trigger; 669 break; 670 } 671 break; 672 case VFIO_PCI_ERR_IRQ_INDEX: 673 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) { 674 case VFIO_IRQ_SET_ACTION_TRIGGER: 675 if (pci_is_pcie(vdev->pdev)) 676 func = vfio_pci_set_err_trigger; 677 break; 678 } 679 break; 680 case VFIO_PCI_REQ_IRQ_INDEX: 681 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) { 682 case VFIO_IRQ_SET_ACTION_TRIGGER: 683 func = vfio_pci_set_req_trigger; 684 break; 685 } 686 break; 687 } 688 689 if (!func) 690 return -ENOTTY; 691 692 return func(vdev, index, start, count, flags, data); 693 } 694