hw/vfio_pci.c

   1 /*
   2  * vfio based device assignment support
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Alex Williamson <alex.williamson@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Based on qemu-kvm device-assignment:
  13  *  Adapted for KVM by Qumranet.
  14  *  Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com)
  15  *  Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com)
  16  *  Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com)
  17  *  Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com)
  18  *  Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com)
  19  */
  20
  21 #include <dirent.h>
  22 #include <unistd.h>
  23 #include <sys/ioctl.h>
  24 #include <sys/mman.h>
  25 #include <sys/stat.h>
  26 #include <sys/types.h>
  27 #include <linux/vfio.h>
  28
  29 #include "config.h"
  30 #include "event_notifier.h"
  31 #include "exec-memory.h"
  32 #include "kvm.h"
  33 #include "memory.h"
  34 #include "msi.h"
  35 #include "msix.h"
  36 #include "qemu-error.h"
  37 #include "range.h"
  38 #include "vfio_pci_int.h"
  39
  40 /* #define DEBUG_VFIO */
  41 #ifdef DEBUG_VFIO
  42 #define DPRINTF(fmt, ...) \
  43     do { fprintf(stderr, "vfio: " fmt, ## __VA_ARGS__); } while (0)
  44 #else
  45 #define DPRINTF(fmt, ...) \
  46     do { } while (0)
  47 #endif
  48
  49 #define MSIX_CAP_LENGTH 12
  50
  51 static QLIST_HEAD(, VFIOContainer)
  52     container_list = QLIST_HEAD_INITIALIZER(container_list);
  53
  54 static QLIST_HEAD(, VFIOGroup)
  55     group_list = QLIST_HEAD_INITIALIZER(group_list);
  56
  57 static void vfio_disable_interrupts(VFIODevice *vdev);
  58 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len);
  59 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled);
  60
  61 /*
  62  * Common VFIO interrupt disable
  63  */
  64 static void vfio_disable_irqindex(VFIODevice *vdev, int index)
  65 {
  66     struct vfio_irq_set irq_set = {
  67         .argsz = sizeof(irq_set),
  68         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
  69         .index = index,
  70         .start = 0,
  71         .count = 0,
  72     };
  73
  74     ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  75
  76     vdev->interrupt = VFIO_INT_NONE;
  77 }
  78
  79 /*
  80  * INTx
  81  */
  82 static void vfio_unmask_intx(VFIODevice *vdev)
  83 {
  84     struct vfio_irq_set irq_set = {
  85         .argsz = sizeof(irq_set),
  86         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
  87         .index = VFIO_PCI_INTX_IRQ_INDEX,
  88         .start = 0,
  89         .count = 1,
  90     };
  91
  92     ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  93 }
  94
  95 static void vfio_intx_interrupt(void *opaque)
  96 {
  97     VFIODevice *vdev = opaque;
  98
  99     if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
 100         return;
 101     }
 102
 103     DPRINTF("%s(%04x:%02x:%02x.%x) Pin %c\n", __func__, vdev->host.domain,
 104             vdev->host.bus, vdev->host.slot, vdev->host.function,
 105             'A' + vdev->intx.pin);
 106
 107     vdev->intx.pending = true;
 108     qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 1);
 109 }
 110
 111 static void vfio_eoi(VFIODevice *vdev)
 112 {
 113     if (!vdev->intx.pending) {
 114         return;
 115     }
 116
 117     DPRINTF("%s(%04x:%02x:%02x.%x) EOI\n", __func__, vdev->host.domain,
 118             vdev->host.bus, vdev->host.slot, vdev->host.function);
 119
 120     vdev->intx.pending = false;
 121     qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
 122     vfio_unmask_intx(vdev);
 123 }
 124
 125 typedef struct QEMU_PACKED VFIOIRQSetFD {
 126     struct vfio_irq_set irq_set;
 127     int32_t fd;
 128 } VFIOIRQSetFD;
 129
 130 static int vfio_enable_intx(VFIODevice *vdev)
 131 {
 132     VFIOIRQSetFD irq_set_fd = {
 133         .irq_set = {
 134             .argsz = sizeof(irq_set_fd),
 135             .flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER,
 136             .index = VFIO_PCI_INTX_IRQ_INDEX,
 137             .start = 0,
 138             .count = 1,
 139         },
 140     };
 141     uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
 142     int ret;
 143
 144     if (vdev->intx.disabled || !pin) {
 145         return 0;
 146     }
 147
 148     vfio_disable_interrupts(vdev);
 149
 150     vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
 151     ret = event_notifier_init(&vdev->intx.interrupt, 0);
 152     if (ret) {
 153         error_report("vfio: Error: event_notifier_init failed\n");
 154         return ret;
 155     }
 156
 157     irq_set_fd.fd = event_notifier_get_fd(&vdev->intx.interrupt);
 158     qemu_set_fd_handler(irq_set_fd.fd, vfio_intx_interrupt, NULL, vdev);
 159
 160     if (ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set_fd)) {
 161         error_report("vfio: Error: Failed to setup INTx fd: %m\n");
 162         return -errno;
 163     }
 164
 165     /*
 166      * Disable mmaps so we can trap on BAR accesses.  We interpret any
 167      * access as a response to an interrupt and unmask the physical
 168      * device.  The device will re-assert if the interrupt is still
 169      * pending.  We'll likely retrigger on the host multiple times per
 170      * guest interrupt, but without EOI notification it's better than
 171      * nothing.  Acceleration paths through KVM will avoid this.
 172      */
 173     vfio_mmap_set_enabled(vdev, false);
 174
 175     vdev->interrupt = VFIO_INT_INTx;
 176
 177     DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
 178             vdev->host.bus, vdev->host.slot, vdev->host.function);
 179
 180     return 0;
 181 }
 182
 183 static void vfio_disable_intx(VFIODevice *vdev)
 184 {
 185     int fd;
 186
 187     vfio_disable_irqindex(vdev, VFIO_PCI_INTX_IRQ_INDEX);
 188     vdev->intx.pending = false;
 189     qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
 190     vfio_mmap_set_enabled(vdev, true);
 191
 192     fd = event_notifier_get_fd(&vdev->intx.interrupt);
 193     qemu_set_fd_handler(fd, NULL, NULL, vdev);
 194     event_notifier_cleanup(&vdev->intx.interrupt);
 195
 196     vdev->interrupt = VFIO_INT_NONE;
 197
 198     DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
 199             vdev->host.bus, vdev->host.slot, vdev->host.function);
 200 }
 201
 202 /*
 203  * MSI/X
 204  */
 205 static void vfio_msi_interrupt(void *opaque)
 206 {
 207     VFIOMSIVector *vector = opaque;
 208     VFIODevice *vdev = vector->vdev;
 209     int nr = vector - vdev->msi_vectors;
 210
 211     if (!event_notifier_test_and_clear(&vector->interrupt)) {
 212         return;
 213     }
 214
 215     DPRINTF("%s(%04x:%02x:%02x.%x) vector %d\n", __func__,
 216             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 217             vdev->host.function, nr);
 218
 219     if (vdev->interrupt == VFIO_INT_MSIX) {
 220         msix_notify(&vdev->pdev, nr);
 221     } else if (vdev->interrupt == VFIO_INT_MSI) {
 222         msi_notify(&vdev->pdev, nr);
 223     } else {
 224         error_report("vfio: MSI interrupt receieved, but not enabled?\n");
 225     }
 226 }
 227
 228 static int vfio_enable_vectors(VFIODevice *vdev, bool msix)
 229 {
 230     struct vfio_irq_set *irq_set;
 231     int ret = 0, i, argsz;
 232     int32_t *fds;
 233
 234     argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
 235
 236     irq_set = g_malloc0(argsz);
 237     irq_set->argsz = argsz;
 238     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 239     irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
 240     irq_set->start = 0;
 241     irq_set->count = vdev->nr_vectors;
 242     fds = (int32_t *)&irq_set->data;
 243
 244     for (i = 0; i < vdev->nr_vectors; i++) {
 245         if (!vdev->msi_vectors[i].use) {
 246             fds[i] = -1;
 247             continue;
 248         }
 249
 250         fds[i] = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
 251     }
 252
 253     ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
 254
 255     g_free(irq_set);
 256
 257     if (!ret) {
 258         vdev->interrupt = msix ? VFIO_INT_MSIX : VFIO_INT_MSI;
 259     }
 260
 261     return ret;
 262 }
 263
 264 static int vfio_msix_vector_use(PCIDevice *pdev,
 265                                 unsigned int nr, MSIMessage msg)
 266 {
 267     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
 268     VFIOMSIVector *vector;
 269     int ret;
 270
 271     DPRINTF("%s(%04x:%02x:%02x.%x) vector %d used\n", __func__,
 272             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 273             vdev->host.function, nr);
 274
 275     if (vdev->interrupt != VFIO_INT_MSIX) {
 276         vfio_disable_interrupts(vdev);
 277     }
 278
 279     if (!vdev->msi_vectors) {
 280         vdev->msi_vectors = g_malloc0(vdev->msix->entries *
 281                                       sizeof(VFIOMSIVector));
 282     }
 283
 284     vector = &vdev->msi_vectors[nr];
 285     vector->vdev = vdev;
 286     vector->use = true;
 287
 288     msix_vector_use(pdev, nr);
 289
 290     if (event_notifier_init(&vector->interrupt, 0)) {
 291         error_report("vfio: Error: event_notifier_init failed\n");
 292     }
 293
 294     /*
 295      * Attempt to enable route through KVM irqchip,
 296      * default to userspace handling if unavailable.
 297      */
 298     vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg);
 299     if (vector->virq < 0 ||
 300         kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
 301                                        vector->virq) < 0) {
 302         if (vector->virq >= 0) {
 303             kvm_irqchip_release_virq(kvm_state, vector->virq);
 304             vector->virq = -1;
 305         }
 306         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 307                             vfio_msi_interrupt, NULL, vector);
 308     }
 309
 310     /*
 311      * We don't want to have the host allocate all possible MSI vectors
 312      * for a device if they're not in use, so we shutdown and incrementally
 313      * increase them as needed.
 314      */
 315     if (vdev->nr_vectors < nr + 1) {
 316         int i;
 317
 318         vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
 319         vdev->nr_vectors = nr + 1;
 320         ret = vfio_enable_vectors(vdev, true);
 321         if (ret) {
 322             error_report("vfio: failed to enable vectors, %d\n", ret);
 323         }
 324
 325         /* We don't know if we've missed interrupts in the interim... */
 326         for (i = 0; i < vdev->msix->entries; i++) {
 327             if (vdev->msi_vectors[i].use) {
 328                 msix_notify(&vdev->pdev, i);
 329             }
 330         }
 331     } else {
 332         VFIOIRQSetFD irq_set_fd = {
 333             .irq_set = {
 334                 .argsz = sizeof(irq_set_fd),
 335                 .flags = VFIO_IRQ_SET_DATA_EVENTFD |
 336                          VFIO_IRQ_SET_ACTION_TRIGGER,
 337                 .index = VFIO_PCI_MSIX_IRQ_INDEX,
 338                 .start = nr,
 339                 .count = 1,
 340             },
 341             .fd = event_notifier_get_fd(&vector->interrupt),
 342         };
 343         ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set_fd);
 344         if (ret) {
 345             error_report("vfio: failed to modify vector, %d\n", ret);
 346         }
 347
 348         /*
 349          * If we were connected to the hardware PBA we could skip this,
 350          * until then, a spurious interrupt is better than starvation.
 351          */
 352         msix_notify(&vdev->pdev, nr);
 353     }
 354
 355     return 0;
 356 }
 357
 358 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
 359 {
 360     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
 361     VFIOMSIVector *vector = &vdev->msi_vectors[nr];
 362     VFIOIRQSetFD irq_set_fd = {
 363         .irq_set = {
 364             .argsz = sizeof(irq_set_fd),
 365             .flags = VFIO_IRQ_SET_DATA_EVENTFD |
 366                      VFIO_IRQ_SET_ACTION_TRIGGER,
 367             .index = VFIO_PCI_MSIX_IRQ_INDEX,
 368             .start = nr,
 369             .count = 1,
 370         },
 371         .fd = -1,
 372     };
 373
 374     DPRINTF("%s(%04x:%02x:%02x.%x) vector %d released\n", __func__,
 375             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 376             vdev->host.function, nr);
 377
 378     /*
 379      * XXX What's the right thing to do here?  This turns off the interrupt
 380      * completely, but do we really just want to switch the interrupt to
 381      * bouncing through userspace and let msix.c drop it?  Not sure.
 382      */
 383     msix_vector_unuse(pdev, nr);
 384     ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set_fd);
 385
 386     if (vector->virq < 0) {
 387         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 388                             NULL, NULL, NULL);
 389     } else {
 390         kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
 391                                           vector->virq);
 392         kvm_irqchip_release_virq(kvm_state, vector->virq);
 393         vector->virq = -1;
 394     }
 395
 396     event_notifier_cleanup(&vector->interrupt);
 397     vector->use = false;
 398 }
 399
 400 /* TODO This should move to msi.c */
 401 static MSIMessage msi_get_msg(PCIDevice *pdev, unsigned int vector)
 402 {
 403     uint16_t flags = pci_get_word(pdev->config + pdev->msi_cap + PCI_MSI_FLAGS);
 404     bool msi64bit = flags & PCI_MSI_FLAGS_64BIT;
 405     MSIMessage msg;
 406
 407     if (msi64bit) {
 408         msg.address = pci_get_quad(pdev->config +
 409                                    pdev->msi_cap + PCI_MSI_ADDRESS_LO);
 410     } else {
 411         msg.address = pci_get_long(pdev->config +
 412                                    pdev->msi_cap + PCI_MSI_ADDRESS_LO);
 413     }
 414
 415     msg.data = pci_get_word(pdev->config + pdev->msi_cap +
 416                             (msi64bit ? PCI_MSI_DATA_64 : PCI_MSI_DATA_32));
 417     msg.data += vector;
 418
 419     return msg;
 420 }
 421
 422 /* So should this */
 423 static void msi_set_qsize(PCIDevice *pdev, uint8_t size)
 424 {
 425     uint8_t *config = pdev->config + pdev->msi_cap;
 426     uint16_t flags;
 427
 428     flags = pci_get_word(config + PCI_MSI_FLAGS);
 429     flags = le16_to_cpu(flags);
 430     flags &= ~PCI_MSI_FLAGS_QSIZE;
 431     flags |= (size & 0x7) << 4;
 432     flags = cpu_to_le16(flags);
 433     pci_set_word(config + PCI_MSI_FLAGS, flags);
 434 }
 435
 436 static void vfio_enable_msi(VFIODevice *vdev)
 437 {
 438     int ret, i;
 439
 440     vfio_disable_interrupts(vdev);
 441
 442     vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
 443 retry:
 444     vdev->msi_vectors = g_malloc0(vdev->nr_vectors * sizeof(VFIOMSIVector));
 445
 446     for (i = 0; i < vdev->nr_vectors; i++) {
 447         MSIMessage msg;
 448         VFIOMSIVector *vector = &vdev->msi_vectors[i];
 449
 450         vector->vdev = vdev;
 451         vector->use = true;
 452
 453         if (event_notifier_init(&vector->interrupt, 0)) {
 454             error_report("vfio: Error: event_notifier_init failed\n");
 455         }
 456
 457         msg = msi_get_msg(&vdev->pdev, i);
 458
 459         /*
 460          * Attempt to enable route through KVM irqchip,
 461          * default to userspace handling if unavailable.
 462          */
 463         vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg);
 464         if (vector->virq < 0 ||
 465             kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
 466                                            vector->virq) < 0) {
 467             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 468                                 vfio_msi_interrupt, NULL, vector);
 469         }
 470     }
 471
 472     ret = vfio_enable_vectors(vdev, false);
 473     if (ret) {
 474         if (ret < 0) {
 475             error_report("vfio: Error: Failed to setup MSI fds: %m\n");
 476         } else if (ret != vdev->nr_vectors) {
 477             error_report("vfio: Error: Failed to enable %d "
 478                          "MSI vectors, retry with %d\n", vdev->nr_vectors, ret);
 479         }
 480
 481         for (i = 0; i < vdev->nr_vectors; i++) {
 482             VFIOMSIVector *vector = &vdev->msi_vectors[i];
 483             if (vector->virq >= 0) {
 484                 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
 485                                                   vector->virq);
 486                 kvm_irqchip_release_virq(kvm_state, vector->virq);
 487                 vector->virq = -1;
 488             } else {
 489                 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 490                                     NULL, NULL, NULL);
 491             }
 492             event_notifier_cleanup(&vector->interrupt);
 493         }
 494
 495         g_free(vdev->msi_vectors);
 496
 497         if (ret > 0 && ret != vdev->nr_vectors) {
 498             vdev->nr_vectors = ret;
 499             goto retry;
 500         }
 501         vdev->nr_vectors = 0;
 502
 503         return;
 504     }
 505
 506     msi_set_qsize(&vdev->pdev, vdev->nr_vectors);
 507
 508     DPRINTF("%s(%04x:%02x:%02x.%x) Enabled %d MSI vectors\n", __func__,
 509             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 510             vdev->host.function, vdev->nr_vectors);
 511 }
 512
 513 static void vfio_disable_msi_x(VFIODevice *vdev, bool msix)
 514 {
 515     int i;
 516
 517     vfio_disable_irqindex(vdev, msix ? VFIO_PCI_MSIX_IRQ_INDEX :
 518                                        VFIO_PCI_MSI_IRQ_INDEX);
 519
 520     for (i = 0; i < vdev->nr_vectors; i++) {
 521         VFIOMSIVector *vector = &vdev->msi_vectors[i];
 522
 523         if (!vector->use) {
 524             continue;
 525         }
 526
 527         if (vector->virq >= 0) {
 528             kvm_irqchip_remove_irqfd_notifier(kvm_state,
 529                                               &vector->interrupt, vector->virq);
 530             kvm_irqchip_release_virq(kvm_state, vector->virq);
 531             vector->virq = -1;
 532         } else {
 533             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 534                                 NULL, NULL, NULL);
 535         }
 536
 537         if (msix) {
 538             msix_vector_unuse(&vdev->pdev, i);
 539         }
 540
 541         event_notifier_cleanup(&vector->interrupt);
 542     }
 543
 544     g_free(vdev->msi_vectors);
 545     vdev->msi_vectors = NULL;
 546     vdev->nr_vectors = 0;
 547
 548     if (!msix) {
 549         msi_set_qsize(&vdev->pdev, 0); /* Actually still means 1 vector */
 550     }
 551
 552     DPRINTF("%s(%04x:%02x:%02x.%x, msi%s)\n", __func__,
 553             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 554             vdev->host.function, msix ? "x" : "");
 555
 556     vfio_enable_intx(vdev);
 557 }
 558
 559 /*
 560  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
 561  */
 562 static void vfio_bar_write(void *opaque, target_phys_addr_t addr,
 563                            uint64_t data, unsigned size)
 564 {
 565     VFIOBAR *bar = opaque;
 566     union {
 567         uint8_t byte;
 568         uint16_t word;
 569         uint32_t dword;
 570         uint64_t qword;
 571     } buf;
 572
 573     switch (size) {
 574     case 1:
 575         buf.byte = data;
 576         break;
 577     case 2:
 578         buf.word = cpu_to_le16(data);
 579         break;
 580     case 4:
 581         buf.dword = cpu_to_le32(data);
 582         break;
 583     default:
 584         hw_error("vfio: unsupported write size, %d bytes\n", size);
 585         break;
 586     }
 587
 588     if (pwrite(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
 589         error_report("%s(,0x%"TARGET_PRIxPHYS", 0x%"PRIx64", %d) failed: %m\n",
 590                      __func__, addr, data, size);
 591     }
 592
 593     DPRINTF("%s(BAR%d+0x%"TARGET_PRIxPHYS", 0x%"PRIx64", %d)\n",
 594             __func__, bar->nr, addr, data, size);
 595
 596     /*
 597      * A read or write to a BAR always signals an INTx EOI.  This will
 598      * do nothing if not pending (including not in INTx mode).  We assume
 599      * that a BAR access is in response to an interrupt and that BAR
 600      * accesses will service the interrupt.  Unfortunately, we don't know
 601      * which access will service the interrupt, so we're potentially
 602      * getting quite a few host interrupts per guest interrupt.
 603      */
 604     vfio_eoi(DO_UPCAST(VFIODevice, bars[bar->nr], bar));
 605 }
 606
 607 static uint64_t vfio_bar_read(void *opaque,
 608                               target_phys_addr_t addr, unsigned size)
 609 {
 610     VFIOBAR *bar = opaque;
 611     union {
 612         uint8_t byte;
 613         uint16_t word;
 614         uint32_t dword;
 615         uint64_t qword;
 616     } buf;
 617     uint64_t data = 0;
 618
 619     if (pread(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
 620         error_report("%s(,0x%"TARGET_PRIxPHYS", %d) failed: %m\n",
 621                      __func__, addr, size);
 622         return (uint64_t)-1;
 623     }
 624
 625     switch (size) {
 626     case 1:
 627         data = buf.byte;
 628         break;
 629     case 2:
 630         data = le16_to_cpu(buf.word);
 631         break;
 632     case 4:
 633         data = le32_to_cpu(buf.dword);
 634         break;
 635     default:
 636         hw_error("vfio: unsupported read size, %d bytes\n", size);
 637         break;
 638     }
 639
 640     DPRINTF("%s(BAR%d+0x%"TARGET_PRIxPHYS", %d) = 0x%"PRIx64"\n",
 641             __func__, bar->nr, addr, size, data);
 642
 643     /* Same as write above */
 644     vfio_eoi(DO_UPCAST(VFIODevice, bars[bar->nr], bar));
 645
 646     return data;
 647 }
 648
 649 static const MemoryRegionOps vfio_bar_ops = {
 650     .read = vfio_bar_read,
 651     .write = vfio_bar_write,
 652     .endianness = DEVICE_LITTLE_ENDIAN,
 653 };
 654
 655 /*
 656  * PCI config space
 657  */
 658 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
 659 {
 660     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
 661     uint32_t val = 0;
 662
 663     /*
 664      * We only need QEMU PCI config support for the ROM BAR, the MSI and MSIX
 665      * capabilities, and the multifunction bit below.  We let VFIO handle
 666      * virtualizing everything else.  Performance is not a concern here.
 667      */
 668     if (ranges_overlap(addr, len, PCI_ROM_ADDRESS, 4) ||
 669         (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
 670          ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) ||
 671         (pdev->cap_present & QEMU_PCI_CAP_MSI &&
 672          ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size))) {
 673
 674         val = pci_default_read_config(pdev, addr, len);
 675     } else {
 676         if (pread(vdev->fd, &val, len, vdev->config_offset + addr) != len) {
 677             error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m\n",
 678                          __func__, vdev->host.domain, vdev->host.bus,
 679                          vdev->host.slot, vdev->host.function, addr, len);
 680             return -errno;
 681         }
 682         val = le32_to_cpu(val);
 683     }
 684
 685     /* Multifunction bit is virualized in QEMU */
 686     if (unlikely(ranges_overlap(addr, len, PCI_HEADER_TYPE, 1))) {
 687         uint32_t mask = PCI_HEADER_TYPE_MULTI_FUNCTION;
 688
 689         if (len == 4) {
 690             mask <<= 16;
 691         }
 692
 693         if (pdev->cap_present & QEMU_PCI_CAP_MULTIFUNCTION) {
 694             val |= mask;
 695         } else {
 696             val &= ~mask;
 697         }
 698     }
 699
 700     DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, len=0x%x) %x\n", __func__,
 701             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 702             vdev->host.function, addr, len, val);
 703
 704     return val;
 705 }
 706
 707 static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
 708                                   uint32_t val, int len)
 709 {
 710     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
 711     uint32_t val_le = cpu_to_le32(val);
 712
 713     DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, 0x%x, len=0x%x)\n", __func__,
 714             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 715             vdev->host.function, addr, val, len);
 716
 717     /* Write everything to VFIO, let it filter out what we can't write */
 718     if (pwrite(vdev->fd, &val_le, len, vdev->config_offset + addr) != len) {
 719         error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m\n",
 720                      __func__, vdev->host.domain, vdev->host.bus,
 721                      vdev->host.slot, vdev->host.function, addr, val, len);
 722     }
 723
 724     /* Write standard header bits to emulation */
 725     if (addr < PCI_CONFIG_HEADER_SIZE) {
 726         pci_default_write_config(pdev, addr, val, len);
 727         return;
 728     }
 729
 730     /* MSI/MSI-X Enabling/Disabling */
 731     if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
 732         ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
 733         int is_enabled, was_enabled = msi_enabled(pdev);
 734
 735         pci_default_write_config(pdev, addr, val, len);
 736
 737         is_enabled = msi_enabled(pdev);
 738
 739         if (!was_enabled && is_enabled) {
 740             vfio_enable_msi(vdev);
 741         } else if (was_enabled && !is_enabled) {
 742             vfio_disable_msi_x(vdev, false);
 743         }
 744     }
 745
 746     if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
 747         ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
 748         int is_enabled, was_enabled = msix_enabled(pdev);
 749
 750         pci_default_write_config(pdev, addr, val, len);
 751
 752         is_enabled = msix_enabled(pdev);
 753
 754         if (!was_enabled && is_enabled) {
 755             /* vfio_msix_vector_use handles this automatically */
 756         } else if (was_enabled && !is_enabled) {
 757             vfio_disable_msi_x(vdev, true);
 758         }
 759     }
 760 }
 761
 762 /*
 763  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
 764  */
 765 static int vfio_dma_map(VFIOContainer *container, target_phys_addr_t iova,
 766                         ram_addr_t size, void *vaddr, bool readonly)
 767 {
 768     struct vfio_iommu_type1_dma_map map = {
 769         .argsz = sizeof(map),
 770         .flags = VFIO_DMA_MAP_FLAG_READ,
 771         .vaddr = (__u64)(intptr_t)vaddr,
 772         .iova = iova,
 773         .size = size,
 774     };
 775
 776     if (!readonly) {
 777         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
 778     }
 779
 780     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map)) {
 781         DPRINTF("VFIO_MAP_DMA: %d\n", -errno);
 782         return -errno;
 783     }
 784
 785     return 0;
 786 }
 787
 788 static int vfio_dma_unmap(VFIOContainer *container,
 789                           target_phys_addr_t iova, ram_addr_t size)
 790 {
 791     struct vfio_iommu_type1_dma_unmap unmap = {
 792         .argsz = sizeof(unmap),
 793         .flags = 0,
 794         .iova = iova,
 795         .size = size,
 796     };
 797
 798     if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
 799         DPRINTF("VFIO_UNMAP_DMA: %d\n", -errno);
 800         return -errno;
 801     }
 802
 803     return 0;
 804 }
 805
 806 static void vfio_listener_dummy1(MemoryListener *listener)
 807 {
 808     /* We don't do batching (begin/commit) or care about logging */
 809 }
 810
 811 static void vfio_listener_dummy2(MemoryListener *listener,
 812                                  MemoryRegionSection *section)
 813 {
 814     /* We don't do logging or care about nops */
 815 }
 816
 817 static void vfio_listener_dummy3(MemoryListener *listener,
 818                                  MemoryRegionSection *section,
 819                                  bool match_data, uint64_t data,
 820                                  EventNotifier *e)
 821 {
 822     /* We don't care about eventfds */
 823 }
 824
 825 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
 826 {
 827     return !memory_region_is_ram(section->mr);
 828 }
 829
 830 static void vfio_listener_region_add(MemoryListener *listener,
 831                                      MemoryRegionSection *section)
 832 {
 833     VFIOContainer *container = container_of(listener, VFIOContainer,
 834                                             iommu_data.listener);
 835     target_phys_addr_t iova, end;
 836     void *vaddr;
 837     int ret;
 838
 839     if (vfio_listener_skipped_section(section)) {
 840         DPRINTF("vfio: SKIPPING region_add %"TARGET_PRIxPHYS" - %"PRIx64"\n",
 841                 section->offset_within_address_space,
 842                 section->offset_within_address_space + section->size - 1);
 843         return;
 844     }
 845
 846     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 847                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 848         error_report("%s received unaligned region\n", __func__);
 849         return;
 850     }
 851
 852     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 853     end = (section->offset_within_address_space + section->size) &
 854           TARGET_PAGE_MASK;
 855
 856     if (iova >= end) {
 857         return;
 858     }
 859
 860     vaddr = memory_region_get_ram_ptr(section->mr) +
 861             section->offset_within_region +
 862             (iova - section->offset_within_address_space);
 863
 864     DPRINTF("vfio: region_add %"TARGET_PRIxPHYS" - %"TARGET_PRIxPHYS" [%p]\n",
 865             iova, end - 1, vaddr);
 866
 867     ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
 868     if (ret) {
 869         error_report("vfio_dma_map(%p, 0x%"TARGET_PRIxPHYS", "
 870                      "0x%"TARGET_PRIxPHYS", %p) = %d (%m)\n",
 871                      container, iova, end - iova, vaddr, ret);
 872     }
 873 }
 874
 875 static void vfio_listener_region_del(MemoryListener *listener,
 876                                      MemoryRegionSection *section)
 877 {
 878     VFIOContainer *container = container_of(listener, VFIOContainer,
 879                                             iommu_data.listener);
 880     target_phys_addr_t iova, end;
 881     int ret;
 882
 883     if (vfio_listener_skipped_section(section)) {
 884         DPRINTF("vfio: SKIPPING region_del %"TARGET_PRIxPHYS" - %"PRIx64"\n",
 885                 section->offset_within_address_space,
 886                 section->offset_within_address_space + section->size - 1);
 887         return;
 888     }
 889
 890     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 891                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 892         error_report("%s received unaligned region\n", __func__);
 893         return;
 894     }
 895
 896     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 897     end = (section->offset_within_address_space + section->size) &
 898           TARGET_PAGE_MASK;
 899
 900     if (iova >= end) {
 901         return;
 902     }
 903
 904     DPRINTF("vfio: region_del %"TARGET_PRIxPHYS" - %"TARGET_PRIxPHYS"\n",
 905             iova, end - 1);
 906
 907     ret = vfio_dma_unmap(container, iova, end - iova);
 908     if (ret) {
 909         error_report("vfio_dma_unmap(%p, 0x%"TARGET_PRIxPHYS", "
 910                      "0x%"TARGET_PRIxPHYS") = %d (%m)\n",
 911                      container, iova, end - iova, ret);
 912     }
 913 }
 914
 915 static MemoryListener vfio_memory_listener = {
 916     .begin = vfio_listener_dummy1,
 917     .commit = vfio_listener_dummy1,
 918     .region_add = vfio_listener_region_add,
 919     .region_del = vfio_listener_region_del,
 920     .region_nop = vfio_listener_dummy2,
 921     .log_start = vfio_listener_dummy2,
 922     .log_stop = vfio_listener_dummy2,
 923     .log_sync = vfio_listener_dummy2,
 924     .log_global_start = vfio_listener_dummy1,
 925     .log_global_stop = vfio_listener_dummy1,
 926     .eventfd_add = vfio_listener_dummy3,
 927     .eventfd_del = vfio_listener_dummy3,
 928 };
 929
 930 static void vfio_listener_release(VFIOContainer *container)
 931 {
 932     memory_listener_unregister(&container->iommu_data.listener);
 933 }
 934
 935 /*
 936  * Interrupt setup
 937  */
 938 static void vfio_disable_interrupts(VFIODevice *vdev)
 939 {
 940     switch (vdev->interrupt) {
 941     case VFIO_INT_INTx:
 942         vfio_disable_intx(vdev);
 943         break;
 944     case VFIO_INT_MSI:
 945         vfio_disable_msi_x(vdev, false);
 946         break;
 947     case VFIO_INT_MSIX:
 948         vfio_disable_msi_x(vdev, true);
 949         break;
 950     }
 951 }
 952
 953 static int vfio_setup_msi(VFIODevice *vdev, int pos)
 954 {
 955     uint16_t ctrl;
 956     bool msi_64bit, msi_maskbit;
 957     int ret, entries;
 958
 959     /*
 960      * TODO: don't peek into msi_supported, let msi_init fail and
 961      * check for ENOTSUP
 962      */
 963     if (!msi_supported) {
 964         return 0;
 965     }
 966
 967     if (pread(vdev->fd, &ctrl, sizeof(ctrl),
 968               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
 969         return -errno;
 970     }
 971     ctrl = le16_to_cpu(ctrl);
 972
 973     msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
 974     msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
 975     entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
 976
 977     DPRINTF("%04x:%02x:%02x.%x PCI MSI CAP @0x%x\n", vdev->host.domain,
 978             vdev->host.bus, vdev->host.slot, vdev->host.function, pos);
 979
 980     ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit);
 981     if (ret < 0) {
 982         error_report("vfio: msi_init failed\n");
 983         return ret;
 984     }
 985     vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
 986
 987     return 0;
 988 }
 989
 990 /*
 991  * We don't have any control over how pci_add_capability() inserts
 992  * capabilities into the chain.  In order to setup MSI-X we need a
 993  * MemoryRegion for the BAR.  In order to setup the BAR and not
 994  * attempt to mmap the MSI-X table area, which VFIO won't allow, we
 995  * need to first look for where the MSI-X table lives.  So we
 996  * unfortunately split MSI-X setup across two functions.
 997  */
 998 static int vfio_early_setup_msix(VFIODevice *vdev)
 999 {
1000     uint8_t pos;
1001     uint16_t ctrl;
1002     uint32_t table, pba;
1003
1004     pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
1005     if (!pos) {
1006         return 0;
1007     }
1008
1009     if (pread(vdev->fd, &ctrl, sizeof(ctrl),
1010               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1011         return -errno;
1012     }
1013
1014     if (pread(vdev->fd, &table, sizeof(table),
1015               vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
1016         return -errno;
1017     }
1018
1019     if (pread(vdev->fd, &pba, sizeof(pba),
1020               vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
1021         return -errno;
1022     }
1023
1024     ctrl = le16_to_cpu(ctrl);
1025     table = le32_to_cpu(table);
1026     pba = le32_to_cpu(pba);
1027
1028     vdev->msix = g_malloc0(sizeof(*(vdev->msix)));
1029     vdev->msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
1030     vdev->msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
1031     vdev->msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
1032     vdev->msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
1033     vdev->msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
1034
1035     DPRINTF("%04x:%02x:%02x.%x "
1036             "PCI MSI-X CAP @0x%x, BAR %d, offset 0x%x, entries %d\n",
1037             vdev->host.domain, vdev->host.bus, vdev->host.slot,
1038             vdev->host.function, pos, vdev->msix->table_bar,
1039             vdev->msix->table_offset, vdev->msix->entries);
1040
1041     return 0;
1042 }
1043
1044 static int vfio_setup_msix(VFIODevice *vdev, int pos)
1045 {
1046     int ret;
1047
1048     /*
1049      * TODO: don't peek into msi_supported, let msix_init fail and
1050      * check for ENOTSUP
1051      */
1052     if (!msi_supported) {
1053         return 0;
1054     }
1055
1056     ret = msix_init(&vdev->pdev, vdev->msix->entries,
1057                     &vdev->bars[vdev->msix->table_bar].mem,
1058                     vdev->msix->table_bar, vdev->msix->table_offset,
1059                     &vdev->bars[vdev->msix->pba_bar].mem,
1060                     vdev->msix->pba_bar, vdev->msix->pba_offset, pos);
1061     if (ret < 0) {
1062         error_report("vfio: msix_init failed\n");
1063         return ret;
1064     }
1065
1066     ret = msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
1067                                     vfio_msix_vector_release);
1068     if (ret) {
1069         error_report("vfio: msix_set_vector_notifiers failed %d\n", ret);
1070         msix_uninit(&vdev->pdev, &vdev->bars[vdev->msix->table_bar].mem,
1071                     &vdev->bars[vdev->msix->pba_bar].mem);
1072         return ret;
1073     }
1074
1075     return 0;
1076 }
1077
1078 static void vfio_teardown_msi(VFIODevice *vdev)
1079 {
1080     msi_uninit(&vdev->pdev);
1081
1082     if (vdev->msix) {
1083         /* FIXME: Why can't unset just silently do nothing?? */
1084         if (vdev->pdev.msix_vector_use_notifier &&
1085             vdev->pdev.msix_vector_release_notifier) {
1086             msix_unset_vector_notifiers(&vdev->pdev);
1087         }
1088
1089         msix_uninit(&vdev->pdev, &vdev->bars[vdev->msix->table_bar].mem,
1090                     &vdev->bars[vdev->msix->pba_bar].mem);
1091     }
1092 }
1093
1094 /*
1095  * Resource setup
1096  */
1097 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled)
1098 {
1099     int i;
1100
1101     for (i = 0; i < PCI_ROM_SLOT; i++) {
1102         VFIOBAR *bar = &vdev->bars[i];
1103
1104         if (!bar->size) {
1105             continue;
1106         }
1107
1108         memory_region_set_enabled(&bar->mmap_mem, enabled);
1109         if (vdev->msix && vdev->msix->table_bar == i) {
1110             memory_region_set_enabled(&vdev->msix->mmap_mem, enabled);
1111         }
1112     }
1113 }
1114
1115 static void vfio_unmap_bar(VFIODevice *vdev, int nr)
1116 {
1117     VFIOBAR *bar = &vdev->bars[nr];
1118
1119     if (!bar->size) {
1120         return;
1121     }
1122
1123     memory_region_del_subregion(&bar->mem, &bar->mmap_mem);
1124     munmap(bar->mmap, memory_region_size(&bar->mmap_mem));
1125
1126     if (vdev->msix && vdev->msix->table_bar == nr) {
1127         memory_region_del_subregion(&bar->mem, &vdev->msix->mmap_mem);
1128         munmap(vdev->msix->mmap, memory_region_size(&vdev->msix->mmap_mem));
1129     }
1130
1131     memory_region_destroy(&bar->mem);
1132 }
1133
1134 static int vfio_mmap_bar(VFIOBAR *bar, MemoryRegion *mem, MemoryRegion *submem,
1135                          void **map, size_t size, off_t offset,
1136                          const char *name)
1137 {
1138     int ret = 0;
1139
1140     if (size && bar->flags & VFIO_REGION_INFO_FLAG_MMAP) {
1141         int prot = 0;
1142
1143         if (bar->flags & VFIO_REGION_INFO_FLAG_READ) {
1144             prot |= PROT_READ;
1145         }
1146
1147         if (bar->flags & VFIO_REGION_INFO_FLAG_WRITE) {
1148             prot |= PROT_WRITE;
1149         }
1150
1151         *map = mmap(NULL, size, prot, MAP_SHARED,
1152                     bar->fd, bar->fd_offset + offset);
1153         if (*map == MAP_FAILED) {
1154             *map = NULL;
1155             ret = -errno;
1156             goto empty_region;
1157         }
1158
1159         memory_region_init_ram_ptr(submem, name, size, *map);
1160     } else {
1161 empty_region:
1162         /* Create a zero sized sub-region to make cleanup easy. */
1163         memory_region_init(submem, name, 0);
1164     }
1165
1166     memory_region_add_subregion(mem, offset, submem);
1167
1168     return ret;
1169 }
1170
1171 static void vfio_map_bar(VFIODevice *vdev, int nr)
1172 {
1173     VFIOBAR *bar = &vdev->bars[nr];
1174     unsigned size = bar->size;
1175     char name[64];
1176     uint32_t pci_bar;
1177     uint8_t type;
1178     int ret;
1179
1180     /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1181     if (!size) {
1182         return;
1183     }
1184
1185     snprintf(name, sizeof(name), "VFIO %04x:%02x:%02x.%x BAR %d",
1186              vdev->host.domain, vdev->host.bus, vdev->host.slot,
1187              vdev->host.function, nr);
1188
1189     /* Determine what type of BAR this is for registration */
1190     ret = pread(vdev->fd, &pci_bar, sizeof(pci_bar),
1191                 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
1192     if (ret != sizeof(pci_bar)) {
1193         error_report("vfio: Failed to read BAR %d (%m)\n", nr);
1194         return;
1195     }
1196
1197     pci_bar = le32_to_cpu(pci_bar);
1198     type = pci_bar & (pci_bar & PCI_BASE_ADDRESS_SPACE_IO ?
1199            ~PCI_BASE_ADDRESS_IO_MASK : ~PCI_BASE_ADDRESS_MEM_MASK);
1200
1201     /* A "slow" read/write mapping underlies all BARs */
1202     memory_region_init_io(&bar->mem, &vfio_bar_ops, bar, name, size);
1203     pci_register_bar(&vdev->pdev, nr, type, &bar->mem);
1204
1205     /*
1206      * We can't mmap areas overlapping the MSIX vector table, so we
1207      * potentially insert a direct-mapped subregion before and after it.
1208      */
1209     if (vdev->msix && vdev->msix->table_bar == nr) {
1210         size = vdev->msix->table_offset & TARGET_PAGE_MASK;
1211     }
1212
1213     strncat(name, " mmap", sizeof(name) - strlen(name) - 1);
1214     if (vfio_mmap_bar(bar, &bar->mem,
1215                       &bar->mmap_mem, &bar->mmap, size, 0, name)) {
1216         error_report("%s unsupported. Performance may be slow\n", name);
1217     }
1218
1219     if (vdev->msix && vdev->msix->table_bar == nr) {
1220         unsigned start;
1221
1222         start = TARGET_PAGE_ALIGN(vdev->msix->table_offset +
1223                                   (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
1224
1225         size = start < bar->size ? bar->size - start : 0;
1226         strncat(name, " msix-hi", sizeof(name) - strlen(name) - 1);
1227         /* VFIOMSIXInfo contains another MemoryRegion for this mapping */
1228         if (vfio_mmap_bar(bar, &bar->mem, &vdev->msix->mmap_mem,
1229                           &vdev->msix->mmap, size, start, name)) {
1230             error_report("%s unsupported. Performance may be slow\n", name);
1231         }
1232     }
1233 }
1234
1235 static void vfio_map_bars(VFIODevice *vdev)
1236 {
1237     int i;
1238
1239     for (i = 0; i < PCI_ROM_SLOT; i++) {
1240         vfio_map_bar(vdev, i);
1241     }
1242 }
1243
1244 static void vfio_unmap_bars(VFIODevice *vdev)
1245 {
1246     int i;
1247
1248     for (i = 0; i < PCI_ROM_SLOT; i++) {
1249         vfio_unmap_bar(vdev, i);
1250     }
1251 }
1252
1253 /*
1254  * General setup
1255  */
1256 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
1257 {
1258     uint8_t tmp, next = 0xff;
1259
1260     for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
1261          tmp = pdev->config[tmp + 1]) {
1262         if (tmp > pos && tmp < next) {
1263             next = tmp;
1264         }
1265     }
1266
1267     return next - pos;
1268 }
1269
1270 static int vfio_add_std_cap(VFIODevice *vdev, uint8_t pos)
1271 {
1272     PCIDevice *pdev = &vdev->pdev;
1273     uint8_t cap_id, next, size;
1274     int ret;
1275
1276     cap_id = pdev->config[pos];
1277     next = pdev->config[pos + 1];
1278
1279     /*
1280      * If it becomes important to configure capabilities to their actual
1281      * size, use this as the default when it's something we don't recognize.
1282      * Since QEMU doesn't actually handle many of the config accesses,
1283      * exact size doesn't seem worthwhile.
1284      */
1285     size = vfio_std_cap_max_size(pdev, pos);
1286
1287     /*
1288      * pci_add_capability always inserts the new capability at the head
1289      * of the chain.  Therefore to end up with a chain that matches the
1290      * physical device, we insert from the end by making this recursive.
1291      * This is also why we pre-caclulate size above as cached config space
1292      * will be changed as we unwind the stack.
1293      */
1294     if (next) {
1295         ret = vfio_add_std_cap(vdev, next);
1296         if (ret) {
1297             return ret;
1298         }
1299     } else {
1300         pdev->config[PCI_CAPABILITY_LIST] = 0; /* Begin the rebuild */
1301     }
1302
1303     switch (cap_id) {
1304     case PCI_CAP_ID_MSI:
1305         ret = vfio_setup_msi(vdev, pos);
1306         break;
1307     case PCI_CAP_ID_MSIX:
1308         ret = vfio_setup_msix(vdev, pos);
1309         break;
1310     default:
1311         ret = pci_add_capability(pdev, cap_id, pos, size);
1312         break;
1313     }
1314
1315     if (ret < 0) {
1316         error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability "
1317                      "0x%x[0x%x]@0x%x: %d\n", vdev->host.domain,
1318                      vdev->host.bus, vdev->host.slot, vdev->host.function,
1319                      cap_id, size, pos, ret);
1320         return ret;
1321     }
1322
1323     return 0;
1324 }
1325
1326 static int vfio_add_capabilities(VFIODevice *vdev)
1327 {
1328     PCIDevice *pdev = &vdev->pdev;
1329
1330     if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
1331         !pdev->config[PCI_CAPABILITY_LIST]) {
1332         return 0; /* Nothing to add */
1333     }
1334
1335     return vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST]);
1336 }
1337
1338 static int vfio_load_rom(VFIODevice *vdev)
1339 {
1340     uint64_t size = vdev->rom_size;
1341     char name[32];
1342     off_t off = 0, voff = vdev->rom_offset;
1343     ssize_t bytes;
1344     void *ptr;
1345
1346     /* If loading ROM from file, pci handles it */
1347     if (vdev->pdev.romfile || !vdev->pdev.rom_bar || !size) {
1348         return 0;
1349     }
1350
1351     DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
1352             vdev->host.bus, vdev->host.slot, vdev->host.function);
1353
1354     snprintf(name, sizeof(name), "vfio[%04x:%02x:%02x.%x].rom",
1355              vdev->host.domain, vdev->host.bus, vdev->host.slot,
1356              vdev->host.function);
1357     memory_region_init_ram(&vdev->pdev.rom, name, size);
1358     ptr = memory_region_get_ram_ptr(&vdev->pdev.rom);
1359     memset(ptr, 0xff, size);
1360
1361     while (size) {
1362         bytes = pread(vdev->fd, ptr + off, size, voff + off);
1363         if (bytes == 0) {
1364             break; /* expect that we could get back less than the ROM BAR */
1365         } else if (bytes > 0) {
1366             off += bytes;
1367             size -= bytes;
1368         } else {
1369             if (errno == EINTR || errno == EAGAIN) {
1370                 continue;
1371             }
1372             error_report("vfio: Error reading device ROM: %m\n");
1373             memory_region_destroy(&vdev->pdev.rom);
1374             return -errno;
1375         }
1376     }
1377
1378     pci_register_bar(&vdev->pdev, PCI_ROM_SLOT, 0, &vdev->pdev.rom);
1379     vdev->pdev.has_rom = true;
1380     return 0;
1381 }
1382
1383 static int vfio_connect_container(VFIOGroup *group)
1384 {
1385     VFIOContainer *container;
1386     int ret, fd;
1387
1388     if (group->container) {
1389         return 0;
1390     }
1391
1392     QLIST_FOREACH(container, &container_list, next) {
1393         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
1394             group->container = container;
1395             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1396             return 0;
1397         }
1398     }
1399
1400     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
1401     if (fd < 0) {
1402         error_report("vfio: failed to open /dev/vfio/vfio: %m\n");
1403         return -errno;
1404     }
1405
1406     ret = ioctl(fd, VFIO_GET_API_VERSION);
1407     if (ret != VFIO_API_VERSION) {
1408         error_report("vfio: supported vfio version: %d, "
1409                      "reported version: %d\n", VFIO_API_VERSION, ret);
1410         close(fd);
1411         return -EINVAL;
1412     }
1413
1414     container = g_malloc0(sizeof(*container));
1415     container->fd = fd;
1416
1417     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) {
1418         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1419         if (ret) {
1420             error_report("vfio: failed to set group container: %m\n");
1421             g_free(container);
1422             close(fd);
1423             return -errno;
1424         }
1425
1426         ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU);
1427         if (ret) {
1428             error_report("vfio: failed to set iommu for container: %m\n");
1429             g_free(container);
1430             close(fd);
1431             return -errno;
1432         }
1433
1434         container->iommu_data.listener = vfio_memory_listener;
1435         container->iommu_data.release = vfio_listener_release;
1436
1437         memory_listener_register(&container->iommu_data.listener,
1438                                  get_system_memory());
1439     } else {
1440         error_report("vfio: No available IOMMU models\n");
1441         g_free(container);
1442         close(fd);
1443         return -EINVAL;
1444     }
1445
1446     QLIST_INIT(&container->group_list);
1447     QLIST_INSERT_HEAD(&container_list, container, next);
1448
1449     group->container = container;
1450     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1451
1452     return 0;
1453 }
1454
1455 static void vfio_disconnect_container(VFIOGroup *group)
1456 {
1457     VFIOContainer *container = group->container;
1458
1459     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1460         error_report("vfio: error disconnecting group %d from container\n",
1461                      group->groupid);
1462     }
1463
1464     QLIST_REMOVE(group, container_next);
1465     group->container = NULL;
1466
1467     if (QLIST_EMPTY(&container->group_list)) {
1468         if (container->iommu_data.release) {
1469             container->iommu_data.release(container);
1470         }
1471         QLIST_REMOVE(container, next);
1472         DPRINTF("vfio_disconnect_container: close container->fd\n");
1473         close(container->fd);
1474         g_free(container);
1475     }
1476 }
1477
1478 static VFIOGroup *vfio_get_group(int groupid)
1479 {
1480     VFIOGroup *group;
1481     char path[32];
1482     struct vfio_group_status status = { .argsz = sizeof(status) };
1483
1484     QLIST_FOREACH(group, &group_list, next) {
1485         if (group->groupid == groupid) {
1486             return group;
1487         }
1488     }
1489
1490     group = g_malloc0(sizeof(*group));
1491
1492     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1493     group->fd = qemu_open(path, O_RDWR);
1494     if (group->fd < 0) {
1495         error_report("vfio: error opening %s: %m\n", path);
1496         g_free(group);
1497         return NULL;
1498     }
1499
1500     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1501         error_report("vfio: error getting group status: %m\n");
1502         close(group->fd);
1503         g_free(group);
1504         return NULL;
1505     }
1506
1507     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1508         error_report("vfio: error, group %d is not viable, please ensure "
1509                      "all devices within the iommu_group are bound to their "
1510                      "vfio bus driver.\n", groupid);
1511         close(group->fd);
1512         g_free(group);
1513         return NULL;
1514     }
1515
1516     group->groupid = groupid;
1517     QLIST_INIT(&group->device_list);
1518
1519     if (vfio_connect_container(group)) {
1520         error_report("vfio: failed to setup container for group %d\n", groupid);
1521         close(group->fd);
1522         g_free(group);
1523         return NULL;
1524     }
1525
1526     QLIST_INSERT_HEAD(&group_list, group, next);
1527
1528     return group;
1529 }
1530
1531 static void vfio_put_group(VFIOGroup *group)
1532 {
1533     if (!QLIST_EMPTY(&group->device_list)) {
1534         return;
1535     }
1536
1537     vfio_disconnect_container(group);
1538     QLIST_REMOVE(group, next);
1539     DPRINTF("vfio_put_group: close group->fd\n");
1540     close(group->fd);
1541     g_free(group);
1542 }
1543
1544 static int vfio_get_device(VFIOGroup *group, const char *name, VFIODevice *vdev)
1545 {
1546     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1547     struct vfio_region_info reg_info = { .argsz = sizeof(reg_info) };
1548     int ret, i;
1549
1550     ret = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1551     if (ret < 0) {
1552         error_report("vfio: error getting device %s from group %d: %m\n",
1553                      name, group->groupid);
1554         error_report("Verify all devices in group %d are bound to vfio-pci "
1555                      "or pci-stub and not already in use\n", group->groupid);
1556         return ret;
1557     }
1558
1559     vdev->fd = ret;
1560     vdev->group = group;
1561     QLIST_INSERT_HEAD(&group->device_list, vdev, next);
1562
1563     /* Sanity check device */
1564     ret = ioctl(vdev->fd, VFIO_DEVICE_GET_INFO, &dev_info);
1565     if (ret) {
1566         error_report("vfio: error getting device info: %m\n");
1567         goto error;
1568     }
1569
1570     DPRINTF("Device %s flags: %u, regions: %u, irgs: %u\n", name,
1571             dev_info.flags, dev_info.num_regions, dev_info.num_irqs);
1572
1573     if (!(dev_info.flags & VFIO_DEVICE_FLAGS_PCI)) {
1574         error_report("vfio: Um, this isn't a PCI device\n");
1575         goto error;
1576     }
1577
1578     vdev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1579     if (!vdev->reset_works) {
1580         error_report("Warning, device %s does not support reset\n", name);
1581     }
1582
1583     if (dev_info.num_regions != VFIO_PCI_NUM_REGIONS) {
1584         error_report("vfio: unexpected number of io regions %u\n",
1585                      dev_info.num_regions);
1586         goto error;
1587     }
1588
1589     if (dev_info.num_irqs != VFIO_PCI_NUM_IRQS) {
1590         error_report("vfio: unexpected number of irqs %u\n", dev_info.num_irqs);
1591         goto error;
1592     }
1593
1594     for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
1595         reg_info.index = i;
1596
1597         ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
1598         if (ret) {
1599             error_report("vfio: Error getting region %d info: %m\n", i);
1600             goto error;
1601         }
1602
1603         DPRINTF("Device %s region %d:\n", name, i);
1604         DPRINTF("  size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1605                 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
1606                 (unsigned long)reg_info.flags);
1607
1608         vdev->bars[i].flags = reg_info.flags;
1609         vdev->bars[i].size = reg_info.size;
1610         vdev->bars[i].fd_offset = reg_info.offset;
1611         vdev->bars[i].fd = vdev->fd;
1612         vdev->bars[i].nr = i;
1613     }
1614
1615     reg_info.index = VFIO_PCI_ROM_REGION_INDEX;
1616
1617     ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
1618     if (ret) {
1619         error_report("vfio: Error getting ROM info: %m\n");
1620         goto error;
1621     }
1622
1623     DPRINTF("Device %s ROM:\n", name);
1624     DPRINTF("  size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1625             (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
1626             (unsigned long)reg_info.flags);
1627
1628     vdev->rom_size = reg_info.size;
1629     vdev->rom_offset = reg_info.offset;
1630
1631     reg_info.index = VFIO_PCI_CONFIG_REGION_INDEX;
1632
1633     ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
1634     if (ret) {
1635         error_report("vfio: Error getting config info: %m\n");
1636         goto error;
1637     }
1638
1639     DPRINTF("Device %s config:\n", name);
1640     DPRINTF("  size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1641             (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
1642             (unsigned long)reg_info.flags);
1643
1644     vdev->config_size = reg_info.size;
1645     vdev->config_offset = reg_info.offset;
1646
1647 error:
1648     if (ret) {
1649         QLIST_REMOVE(vdev, next);
1650         vdev->group = NULL;
1651         close(vdev->fd);
1652     }
1653     return ret;
1654 }
1655
1656 static void vfio_put_device(VFIODevice *vdev)
1657 {
1658     QLIST_REMOVE(vdev, next);
1659     vdev->group = NULL;
1660     DPRINTF("vfio_put_device: close vdev->fd\n");
1661     close(vdev->fd);
1662     if (vdev->msix) {
1663         g_free(vdev->msix);
1664         vdev->msix = NULL;
1665     }
1666 }
1667
1668 static int vfio_initfn(PCIDevice *pdev)
1669 {
1670     VFIODevice *pvdev, *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1671     VFIOGroup *group;
1672     char path[PATH_MAX], iommu_group_path[PATH_MAX], *group_name;
1673     ssize_t len;
1674     struct stat st;
1675     int groupid;
1676     int ret;
1677
1678     /* Check that the host device exists */
1679     snprintf(path, sizeof(path),
1680              "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/",
1681              vdev->host.domain, vdev->host.bus, vdev->host.slot,
1682              vdev->host.function);
1683     if (stat(path, &st) < 0) {
1684         error_report("vfio: error: no such host device: %s\n", path);
1685         return -errno;
1686     }
1687
1688     strncat(path, "iommu_group", sizeof(path) - strlen(path) - 1);
1689
1690     len = readlink(path, iommu_group_path, PATH_MAX);
1691     if (len <= 0) {
1692         error_report("vfio: error no iommu_group for device\n");
1693         return -errno;
1694     }
1695
1696     iommu_group_path[len] = 0;
1697     group_name = basename(iommu_group_path);
1698
1699     if (sscanf(group_name, "%d", &groupid) != 1) {
1700         error_report("vfio: error reading %s: %m\n", path);
1701         return -errno;
1702     }
1703
1704     DPRINTF("%s(%04x:%02x:%02x.%x) group %d\n", __func__, vdev->host.domain,
1705             vdev->host.bus, vdev->host.slot, vdev->host.function, groupid);
1706
1707     group = vfio_get_group(groupid);
1708     if (!group) {
1709         error_report("vfio: failed to get group %d\n", groupid);
1710         return -ENOENT;
1711     }
1712
1713     snprintf(path, sizeof(path), "%04x:%02x:%02x.%01x",
1714             vdev->host.domain, vdev->host.bus, vdev->host.slot,
1715             vdev->host.function);
1716
1717     QLIST_FOREACH(pvdev, &group->device_list, next) {
1718         if (pvdev->host.domain == vdev->host.domain &&
1719             pvdev->host.bus == vdev->host.bus &&
1720             pvdev->host.slot == vdev->host.slot &&
1721             pvdev->host.function == vdev->host.function) {
1722
1723             error_report("vfio: error: device %s is already attached\n", path);
1724             vfio_put_group(group);
1725             return -EBUSY;
1726         }
1727     }
1728
1729     ret = vfio_get_device(group, path, vdev);
1730     if (ret) {
1731         error_report("vfio: failed to get device %s\n", path);
1732         vfio_put_group(group);
1733         return ret;
1734     }
1735
1736     /* Get a copy of config space */
1737     ret = pread(vdev->fd, vdev->pdev.config,
1738                 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
1739                 vdev->config_offset);
1740     if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
1741         ret = ret < 0 ? -errno : -EFAULT;
1742         error_report("vfio: Failed to read device config space\n");
1743         goto out_put;
1744     }
1745
1746     /*
1747      * Clear host resource mapping info.  If we choose not to register a
1748      * BAR, such as might be the case with the option ROM, we can get
1749      * confusing, unwritable, residual addresses from the host here.
1750      */
1751     memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
1752     memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
1753
1754     vfio_load_rom(vdev);
1755
1756     ret = vfio_early_setup_msix(vdev);
1757     if (ret) {
1758         goto out_put;
1759     }
1760
1761     vfio_map_bars(vdev);
1762
1763     ret = vfio_add_capabilities(vdev);
1764     if (ret) {
1765         goto out_teardown;
1766     }
1767
1768     if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
1769         if (vdev->intx.intx && strcmp(vdev->intx.intx, "off")) {
1770             error_report("vfio: Unknown option x-intx=%s, "
1771                          "valid options: \"off\".\n", vdev->intx.intx);
1772             ret = -EINVAL;
1773             goto out_teardown;
1774         }
1775
1776         if (vdev->intx.intx && !strcmp(vdev->intx.intx, "off")) {
1777             vdev->intx.disabled = true;
1778         }
1779
1780         ret = vfio_enable_intx(vdev);
1781         if (ret) {
1782             goto out_teardown;
1783         }
1784     }
1785
1786     return 0;
1787
1788 out_teardown:
1789     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
1790     vfio_teardown_msi(vdev);
1791     vfio_unmap_bars(vdev);
1792 out_put:
1793     vfio_put_device(vdev);
1794     vfio_put_group(group);
1795     return ret;
1796 }
1797
1798 static void vfio_exitfn(PCIDevice *pdev)
1799 {
1800     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1801     VFIOGroup *group = vdev->group;
1802
1803     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
1804     vfio_disable_interrupts(vdev);
1805     vfio_teardown_msi(vdev);
1806     vfio_unmap_bars(vdev);
1807     vfio_put_device(vdev);
1808     vfio_put_group(group);
1809 }
1810
1811 static void vfio_pci_reset(DeviceState *dev)
1812 {
1813     PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
1814     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1815
1816     if (!vdev->reset_works) {
1817         return;
1818     }
1819
1820     if (ioctl(vdev->fd, VFIO_DEVICE_RESET)) {
1821         error_report("vfio: Error unable to reset physical device "
1822                      "(%04x:%02x:%02x.%x): %m\n", vdev->host.domain,
1823                      vdev->host.bus, vdev->host.slot, vdev->host.function);
1824     }
1825 }
1826
1827 static Property vfio_pci_dev_properties[] = {
1828     DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIODevice, host),
1829     DEFINE_PROP_STRING("x-intx", VFIODevice, intx.intx),
1830     /*
1831      * TODO - support passed fds... is this necessary?
1832      * DEFINE_PROP_STRING("vfiofd", VFIODevice, vfiofd_name),
1833      * DEFINE_PROP_STRING("vfiogroupfd, VFIODevice, vfiogroupfd_name),
1834      */
1835     DEFINE_PROP_END_OF_LIST(),
1836 };
1837
1838
1839 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
1840 {
1841     DeviceClass *dc = DEVICE_CLASS(klass);
1842     PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
1843
1844     dc->reset = vfio_pci_reset;
1845     dc->props = vfio_pci_dev_properties;
1846     pdc->init = vfio_initfn;
1847     pdc->exit = vfio_exitfn;
1848     pdc->config_read = vfio_pci_read_config;
1849     pdc->config_write = vfio_pci_write_config;
1850 }
1851
1852 static const TypeInfo vfio_pci_dev_info = {
1853     .name = "vfio-pci",
1854     .parent = TYPE_PCI_DEVICE,
1855     .instance_size = sizeof(VFIODevice),
1856     .class_init = vfio_pci_dev_class_init,
1857 };
1858
1859 static void register_vfio_pci_dev_type(void)
1860 {
1861     type_register_static(&vfio_pci_dev_info);
1862 }
1863
1864 type_init(register_vfio_pci_dev_type)