2 * vfio based device assignment support
4 * Copyright Red Hat, Inc. 2012
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
12 * Based on qemu-kvm device-assignment:
13 * Adapted for KVM by Qumranet.
22 #include <linux/vfio.h>
23 #include <sys/ioctl.h>
26 #include <sys/types.h>
30 #include "exec/address-spaces.h"
31 #include "exec/memory.h"
32 #include "hw/pci/msi.h"
33 #include "hw/pci/msix.h"
34 #include "hw/pci/pci.h"
35 #include "qemu-common.h"
36 #include "qemu/error-report.h"
37 #include "qemu/event_notifier.h"
38 #include "qemu/queue.h"
39 #include "qemu/range.h"
40 #include "sysemu/kvm.h"
41 #include "sysemu/sysemu.h"
43 /* #define DEBUG_VFIO */
45 #define DPRINTF(fmt, ...) \
46 do { fprintf(stderr, "vfio: " fmt, ## __VA_ARGS__); } while (0)
48 #define DPRINTF(fmt, ...) \
52 /* Extra debugging, trap acceleration paths for more logging */
53 #define VFIO_ALLOW_MMAP 1
54 #define VFIO_ALLOW_KVM_INTX 1
58 typedef struct VFIOQuirk {
60 struct VFIODevice *vdev;
61 QLIST_ENTRY(VFIOQuirk) next;
66 typedef struct VFIOBAR {
67 off_t fd_offset; /* offset of BAR within device fd */
68 int fd; /* device fd, allows us to pass VFIOBAR as opaque data */
69 MemoryRegion mem; /* slow, read/write access */
70 MemoryRegion mmap_mem; /* direct mapped access */
73 uint32_t flags; /* VFIO region flags (rd/wr/mmap) */
74 uint8_t nr; /* cache the BAR number for debug */
75 QLIST_HEAD(, VFIOQuirk) quirks;
78 typedef struct VFIOVGARegion {
82 QLIST_HEAD(, VFIOQuirk) quirks;
85 typedef struct VFIOVGA {
88 VFIOVGARegion region[QEMU_PCI_VGA_NUM_REGIONS];
91 typedef struct VFIOINTx {
92 bool pending; /* interrupt pending */
93 bool kvm_accel; /* set when QEMU bypass through KVM enabled */
94 uint8_t pin; /* which pin to pull for qemu_set_irq */
95 EventNotifier interrupt; /* eventfd triggered on interrupt */
96 EventNotifier unmask; /* eventfd for unmask on QEMU bypass */
97 PCIINTxRoute route; /* routing info for QEMU bypass */
98 uint32_t mmap_timeout; /* delay to re-enable mmaps after interrupt */
99 QEMUTimer *mmap_timer; /* enable mmaps after periods w/o interrupts */
102 typedef struct VFIOMSIVector {
103 EventNotifier interrupt; /* eventfd triggered on interrupt */
104 struct VFIODevice *vdev; /* back pointer to device */
105 int virq; /* KVM irqchip route for QEMU bypass */
118 typedef struct VFIOContainer {
119 int fd; /* /dev/vfio/vfio, empowered by the attached groups */
121 /* enable abstraction to support various iommu backends */
123 MemoryListener listener; /* Used by type1 iommu */
125 void (*release)(struct VFIOContainer *);
127 QLIST_HEAD(, VFIOGroup) group_list;
128 QLIST_ENTRY(VFIOContainer) next;
131 /* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
132 typedef struct VFIOMSIXInfo {
136 uint32_t table_offset;
138 MemoryRegion mmap_mem;
142 typedef struct VFIODevice {
146 unsigned int config_size;
147 uint8_t *emulated_config_bits; /* QEMU emulated bits, little-endian */
148 off_t config_offset; /* Offset of config space region within device fd */
149 unsigned int rom_size;
150 off_t rom_offset; /* Offset of ROM region within device fd */
152 VFIOMSIVector *msi_vectors;
154 int nr_vectors; /* Number of MSI/MSIX vectors currently in use */
155 int interrupt; /* Current interrupt type */
156 VFIOBAR bars[PCI_NUM_REGIONS - 1]; /* No ROM */
157 VFIOVGA vga; /* 0xa0000, 0x3b0, 0x3c0 */
158 PCIHostDeviceAddress host;
159 QLIST_ENTRY(VFIODevice) next;
160 struct VFIOGroup *group;
162 #define VFIO_FEATURE_ENABLE_VGA_BIT 0
163 #define VFIO_FEATURE_ENABLE_VGA (1 << VFIO_FEATURE_ENABLE_VGA_BIT)
170 typedef struct VFIOGroup {
173 VFIOContainer *container;
174 QLIST_HEAD(, VFIODevice) device_list;
175 QLIST_ENTRY(VFIOGroup) next;
176 QLIST_ENTRY(VFIOGroup) container_next;
179 #define MSIX_CAP_LENGTH 12
181 static QLIST_HEAD(, VFIOContainer)
182 container_list = QLIST_HEAD_INITIALIZER(container_list);
184 static QLIST_HEAD(, VFIOGroup)
185 group_list = QLIST_HEAD_INITIALIZER(group_list);
187 static void vfio_disable_interrupts(VFIODevice *vdev);
188 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len);
189 static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
190 uint32_t val, int len);
191 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled);
194 * Common VFIO interrupt disable
196 static void vfio_disable_irqindex(VFIODevice *vdev, int index)
198 struct vfio_irq_set irq_set = {
199 .argsz = sizeof(irq_set),
200 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
206 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
212 static void vfio_unmask_intx(VFIODevice *vdev)
214 struct vfio_irq_set irq_set = {
215 .argsz = sizeof(irq_set),
216 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
217 .index = VFIO_PCI_INTX_IRQ_INDEX,
222 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
225 #ifdef CONFIG_KVM /* Unused outside of CONFIG_KVM code */
226 static void vfio_mask_intx(VFIODevice *vdev)
228 struct vfio_irq_set irq_set = {
229 .argsz = sizeof(irq_set),
230 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
231 .index = VFIO_PCI_INTX_IRQ_INDEX,
236 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
241 * Disabling BAR mmaping can be slow, but toggling it around INTx can
242 * also be a huge overhead. We try to get the best of both worlds by
243 * waiting until an interrupt to disable mmaps (subsequent transitions
244 * to the same state are effectively no overhead). If the interrupt has
245 * been serviced and the time gap is long enough, we re-enable mmaps for
246 * performance. This works well for things like graphics cards, which
247 * may not use their interrupt at all and are penalized to an unusable
248 * level by read/write BAR traps. Other devices, like NICs, have more
249 * regular interrupts and see much better latency by staying in non-mmap
250 * mode. We therefore set the default mmap_timeout such that a ping
251 * is just enough to keep the mmap disabled. Users can experiment with
252 * other options with the x-intx-mmap-timeout-ms parameter (a value of
253 * zero disables the timer).
255 static void vfio_intx_mmap_enable(void *opaque)
257 VFIODevice *vdev = opaque;
259 if (vdev->intx.pending) {
260 qemu_mod_timer(vdev->intx.mmap_timer,
261 qemu_get_clock_ms(vm_clock) + vdev->intx.mmap_timeout);
265 vfio_mmap_set_enabled(vdev, true);
268 static void vfio_intx_interrupt(void *opaque)
270 VFIODevice *vdev = opaque;
272 if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
276 DPRINTF("%s(%04x:%02x:%02x.%x) Pin %c\n", __func__, vdev->host.domain,
277 vdev->host.bus, vdev->host.slot, vdev->host.function,
278 'A' + vdev->intx.pin);
280 vdev->intx.pending = true;
281 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 1);
282 vfio_mmap_set_enabled(vdev, false);
283 if (vdev->intx.mmap_timeout) {
284 qemu_mod_timer(vdev->intx.mmap_timer,
285 qemu_get_clock_ms(vm_clock) + vdev->intx.mmap_timeout);
289 static void vfio_eoi(VFIODevice *vdev)
291 if (!vdev->intx.pending) {
295 DPRINTF("%s(%04x:%02x:%02x.%x) EOI\n", __func__, vdev->host.domain,
296 vdev->host.bus, vdev->host.slot, vdev->host.function);
298 vdev->intx.pending = false;
299 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
300 vfio_unmask_intx(vdev);
303 static void vfio_enable_intx_kvm(VFIODevice *vdev)
306 struct kvm_irqfd irqfd = {
307 .fd = event_notifier_get_fd(&vdev->intx.interrupt),
308 .gsi = vdev->intx.route.irq,
309 .flags = KVM_IRQFD_FLAG_RESAMPLE,
311 struct vfio_irq_set *irq_set;
315 if (!VFIO_ALLOW_KVM_INTX || !kvm_irqfds_enabled() ||
316 vdev->intx.route.mode != PCI_INTX_ENABLED ||
317 !kvm_check_extension(kvm_state, KVM_CAP_IRQFD_RESAMPLE)) {
321 /* Get to a known interrupt state */
322 qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
323 vfio_mask_intx(vdev);
324 vdev->intx.pending = false;
325 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
327 /* Get an eventfd for resample/unmask */
328 if (event_notifier_init(&vdev->intx.unmask, 0)) {
329 error_report("vfio: Error: event_notifier_init failed eoi");
333 /* KVM triggers it, VFIO listens for it */
334 irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
336 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
337 error_report("vfio: Error: Failed to setup resample irqfd: %m");
341 argsz = sizeof(*irq_set) + sizeof(*pfd);
343 irq_set = g_malloc0(argsz);
344 irq_set->argsz = argsz;
345 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
346 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
349 pfd = (int32_t *)&irq_set->data;
351 *pfd = irqfd.resamplefd;
353 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
356 error_report("vfio: Error: Failed to setup INTx unmask fd: %m");
361 vfio_unmask_intx(vdev);
363 vdev->intx.kvm_accel = true;
365 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel enabled\n",
366 __func__, vdev->host.domain, vdev->host.bus,
367 vdev->host.slot, vdev->host.function);
372 irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
373 kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
375 event_notifier_cleanup(&vdev->intx.unmask);
377 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
378 vfio_unmask_intx(vdev);
382 static void vfio_disable_intx_kvm(VFIODevice *vdev)
385 struct kvm_irqfd irqfd = {
386 .fd = event_notifier_get_fd(&vdev->intx.interrupt),
387 .gsi = vdev->intx.route.irq,
388 .flags = KVM_IRQFD_FLAG_DEASSIGN,
391 if (!vdev->intx.kvm_accel) {
396 * Get to a known state, hardware masked, QEMU ready to accept new
397 * interrupts, QEMU IRQ de-asserted.
399 vfio_mask_intx(vdev);
400 vdev->intx.pending = false;
401 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
403 /* Tell KVM to stop listening for an INTx irqfd */
404 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
405 error_report("vfio: Error: Failed to disable INTx irqfd: %m");
408 /* We only need to close the eventfd for VFIO to cleanup the kernel side */
409 event_notifier_cleanup(&vdev->intx.unmask);
411 /* QEMU starts listening for interrupt events. */
412 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
414 vdev->intx.kvm_accel = false;
416 /* If we've missed an event, let it re-fire through QEMU */
417 vfio_unmask_intx(vdev);
419 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel disabled\n",
420 __func__, vdev->host.domain, vdev->host.bus,
421 vdev->host.slot, vdev->host.function);
425 static void vfio_update_irq(PCIDevice *pdev)
427 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
430 if (vdev->interrupt != VFIO_INT_INTx) {
434 route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
436 if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
437 return; /* Nothing changed */
440 DPRINTF("%s(%04x:%02x:%02x.%x) IRQ moved %d -> %d\n", __func__,
441 vdev->host.domain, vdev->host.bus, vdev->host.slot,
442 vdev->host.function, vdev->intx.route.irq, route.irq);
444 vfio_disable_intx_kvm(vdev);
446 vdev->intx.route = route;
448 if (route.mode != PCI_INTX_ENABLED) {
452 vfio_enable_intx_kvm(vdev);
454 /* Re-enable the interrupt in cased we missed an EOI */
458 static int vfio_enable_intx(VFIODevice *vdev)
460 uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
462 struct vfio_irq_set *irq_set;
469 vfio_disable_interrupts(vdev);
471 vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
475 * Only conditional to avoid generating error messages on platforms
476 * where we won't actually use the result anyway.
478 if (kvm_irqfds_enabled() &&
479 kvm_check_extension(kvm_state, KVM_CAP_IRQFD_RESAMPLE)) {
480 vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
485 ret = event_notifier_init(&vdev->intx.interrupt, 0);
487 error_report("vfio: Error: event_notifier_init failed");
491 argsz = sizeof(*irq_set) + sizeof(*pfd);
493 irq_set = g_malloc0(argsz);
494 irq_set->argsz = argsz;
495 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
496 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
499 pfd = (int32_t *)&irq_set->data;
501 *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
502 qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
504 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
507 error_report("vfio: Error: Failed to setup INTx fd: %m");
508 qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
509 event_notifier_cleanup(&vdev->intx.interrupt);
513 vfio_enable_intx_kvm(vdev);
515 vdev->interrupt = VFIO_INT_INTx;
517 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
518 vdev->host.bus, vdev->host.slot, vdev->host.function);
523 static void vfio_disable_intx(VFIODevice *vdev)
527 qemu_del_timer(vdev->intx.mmap_timer);
528 vfio_disable_intx_kvm(vdev);
529 vfio_disable_irqindex(vdev, VFIO_PCI_INTX_IRQ_INDEX);
530 vdev->intx.pending = false;
531 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
532 vfio_mmap_set_enabled(vdev, true);
534 fd = event_notifier_get_fd(&vdev->intx.interrupt);
535 qemu_set_fd_handler(fd, NULL, NULL, vdev);
536 event_notifier_cleanup(&vdev->intx.interrupt);
538 vdev->interrupt = VFIO_INT_NONE;
540 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
541 vdev->host.bus, vdev->host.slot, vdev->host.function);
547 static void vfio_msi_interrupt(void *opaque)
549 VFIOMSIVector *vector = opaque;
550 VFIODevice *vdev = vector->vdev;
551 int nr = vector - vdev->msi_vectors;
553 if (!event_notifier_test_and_clear(&vector->interrupt)) {
557 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d\n", __func__,
558 vdev->host.domain, vdev->host.bus, vdev->host.slot,
559 vdev->host.function, nr);
561 if (vdev->interrupt == VFIO_INT_MSIX) {
562 msix_notify(&vdev->pdev, nr);
563 } else if (vdev->interrupt == VFIO_INT_MSI) {
564 msi_notify(&vdev->pdev, nr);
566 error_report("vfio: MSI interrupt receieved, but not enabled?");
570 static int vfio_enable_vectors(VFIODevice *vdev, bool msix)
572 struct vfio_irq_set *irq_set;
573 int ret = 0, i, argsz;
576 argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
578 irq_set = g_malloc0(argsz);
579 irq_set->argsz = argsz;
580 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
581 irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
583 irq_set->count = vdev->nr_vectors;
584 fds = (int32_t *)&irq_set->data;
586 for (i = 0; i < vdev->nr_vectors; i++) {
587 if (!vdev->msi_vectors[i].use) {
592 fds[i] = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
595 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
602 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
603 MSIMessage *msg, IOHandler *handler)
605 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
606 VFIOMSIVector *vector;
609 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d used\n", __func__,
610 vdev->host.domain, vdev->host.bus, vdev->host.slot,
611 vdev->host.function, nr);
613 vector = &vdev->msi_vectors[nr];
617 msix_vector_use(pdev, nr);
619 if (event_notifier_init(&vector->interrupt, 0)) {
620 error_report("vfio: Error: event_notifier_init failed");
624 * Attempt to enable route through KVM irqchip,
625 * default to userspace handling if unavailable.
627 vector->virq = msg ? kvm_irqchip_add_msi_route(kvm_state, *msg) : -1;
628 if (vector->virq < 0 ||
629 kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
631 if (vector->virq >= 0) {
632 kvm_irqchip_release_virq(kvm_state, vector->virq);
635 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
636 handler, NULL, vector);
640 * We don't want to have the host allocate all possible MSI vectors
641 * for a device if they're not in use, so we shutdown and incrementally
642 * increase them as needed.
644 if (vdev->nr_vectors < nr + 1) {
645 vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
646 vdev->nr_vectors = nr + 1;
647 ret = vfio_enable_vectors(vdev, true);
649 error_report("vfio: failed to enable vectors, %d", ret);
653 struct vfio_irq_set *irq_set;
656 argsz = sizeof(*irq_set) + sizeof(*pfd);
658 irq_set = g_malloc0(argsz);
659 irq_set->argsz = argsz;
660 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
661 VFIO_IRQ_SET_ACTION_TRIGGER;
662 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
665 pfd = (int32_t *)&irq_set->data;
667 *pfd = event_notifier_get_fd(&vector->interrupt);
669 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
672 error_report("vfio: failed to modify vector, %d", ret);
679 static int vfio_msix_vector_use(PCIDevice *pdev,
680 unsigned int nr, MSIMessage msg)
682 return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
685 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
687 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
688 VFIOMSIVector *vector = &vdev->msi_vectors[nr];
690 struct vfio_irq_set *irq_set;
693 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d released\n", __func__,
694 vdev->host.domain, vdev->host.bus, vdev->host.slot,
695 vdev->host.function, nr);
698 * XXX What's the right thing to do here? This turns off the interrupt
699 * completely, but do we really just want to switch the interrupt to
700 * bouncing through userspace and let msix.c drop it? Not sure.
702 msix_vector_unuse(pdev, nr);
704 argsz = sizeof(*irq_set) + sizeof(*pfd);
706 irq_set = g_malloc0(argsz);
707 irq_set->argsz = argsz;
708 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
709 VFIO_IRQ_SET_ACTION_TRIGGER;
710 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
713 pfd = (int32_t *)&irq_set->data;
717 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
721 if (vector->virq < 0) {
722 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
725 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
727 kvm_irqchip_release_virq(kvm_state, vector->virq);
731 event_notifier_cleanup(&vector->interrupt);
735 static void vfio_enable_msix(VFIODevice *vdev)
737 vfio_disable_interrupts(vdev);
739 vdev->msi_vectors = g_malloc0(vdev->msix->entries * sizeof(VFIOMSIVector));
741 vdev->interrupt = VFIO_INT_MSIX;
744 * Some communication channels between VF & PF or PF & fw rely on the
745 * physical state of the device and expect that enabling MSI-X from the
746 * guest enables the same on the host. When our guest is Linux, the
747 * guest driver call to pci_enable_msix() sets the enabling bit in the
748 * MSI-X capability, but leaves the vector table masked. We therefore
749 * can't rely on a vector_use callback (from request_irq() in the guest)
750 * to switch the physical device into MSI-X mode because that may come a
751 * long time after pci_enable_msix(). This code enables vector 0 with
752 * triggering to userspace, then immediately release the vector, leaving
753 * the physical device with no vectors enabled, but MSI-X enabled, just
754 * like the guest view.
756 vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
757 vfio_msix_vector_release(&vdev->pdev, 0);
759 if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
760 vfio_msix_vector_release, NULL)) {
761 error_report("vfio: msix_set_vector_notifiers failed");
764 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
765 vdev->host.bus, vdev->host.slot, vdev->host.function);
768 static void vfio_enable_msi(VFIODevice *vdev)
772 vfio_disable_interrupts(vdev);
774 vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
776 vdev->msi_vectors = g_malloc0(vdev->nr_vectors * sizeof(VFIOMSIVector));
778 for (i = 0; i < vdev->nr_vectors; i++) {
780 VFIOMSIVector *vector = &vdev->msi_vectors[i];
785 if (event_notifier_init(&vector->interrupt, 0)) {
786 error_report("vfio: Error: event_notifier_init failed");
789 msg = msi_get_message(&vdev->pdev, i);
792 * Attempt to enable route through KVM irqchip,
793 * default to userspace handling if unavailable.
795 vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg);
796 if (vector->virq < 0 ||
797 kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
799 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
800 vfio_msi_interrupt, NULL, vector);
804 ret = vfio_enable_vectors(vdev, false);
807 error_report("vfio: Error: Failed to setup MSI fds: %m");
808 } else if (ret != vdev->nr_vectors) {
809 error_report("vfio: Error: Failed to enable %d "
810 "MSI vectors, retry with %d", vdev->nr_vectors, ret);
813 for (i = 0; i < vdev->nr_vectors; i++) {
814 VFIOMSIVector *vector = &vdev->msi_vectors[i];
815 if (vector->virq >= 0) {
816 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
818 kvm_irqchip_release_virq(kvm_state, vector->virq);
821 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
824 event_notifier_cleanup(&vector->interrupt);
827 g_free(vdev->msi_vectors);
829 if (ret > 0 && ret != vdev->nr_vectors) {
830 vdev->nr_vectors = ret;
833 vdev->nr_vectors = 0;
838 vdev->interrupt = VFIO_INT_MSI;
840 DPRINTF("%s(%04x:%02x:%02x.%x) Enabled %d MSI vectors\n", __func__,
841 vdev->host.domain, vdev->host.bus, vdev->host.slot,
842 vdev->host.function, vdev->nr_vectors);
845 static void vfio_disable_msi_common(VFIODevice *vdev)
847 g_free(vdev->msi_vectors);
848 vdev->msi_vectors = NULL;
849 vdev->nr_vectors = 0;
850 vdev->interrupt = VFIO_INT_NONE;
852 vfio_enable_intx(vdev);
855 static void vfio_disable_msix(VFIODevice *vdev)
857 msix_unset_vector_notifiers(&vdev->pdev);
859 if (vdev->nr_vectors) {
860 vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
863 vfio_disable_msi_common(vdev);
865 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
866 vdev->host.bus, vdev->host.slot, vdev->host.function);
869 static void vfio_disable_msi(VFIODevice *vdev)
873 vfio_disable_irqindex(vdev, VFIO_PCI_MSI_IRQ_INDEX);
875 for (i = 0; i < vdev->nr_vectors; i++) {
876 VFIOMSIVector *vector = &vdev->msi_vectors[i];
882 if (vector->virq >= 0) {
883 kvm_irqchip_remove_irqfd_notifier(kvm_state,
884 &vector->interrupt, vector->virq);
885 kvm_irqchip_release_virq(kvm_state, vector->virq);
888 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
892 event_notifier_cleanup(&vector->interrupt);
895 vfio_disable_msi_common(vdev);
897 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
898 vdev->host.bus, vdev->host.slot, vdev->host.function);
902 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
904 static void vfio_bar_write(void *opaque, hwaddr addr,
905 uint64_t data, unsigned size)
907 VFIOBAR *bar = opaque;
920 buf.word = cpu_to_le16(data);
923 buf.dword = cpu_to_le32(data);
926 hw_error("vfio: unsupported write size, %d bytes\n", size);
930 if (pwrite(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
931 error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
932 __func__, addr, data, size);
937 VFIODevice *vdev = container_of(bar, VFIODevice, bars[bar->nr]);
939 DPRINTF("%s(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx", 0x%"PRIx64
940 ", %d)\n", __func__, vdev->host.domain, vdev->host.bus,
941 vdev->host.slot, vdev->host.function, bar->nr, addr,
947 * A read or write to a BAR always signals an INTx EOI. This will
948 * do nothing if not pending (including not in INTx mode). We assume
949 * that a BAR access is in response to an interrupt and that BAR
950 * accesses will service the interrupt. Unfortunately, we don't know
951 * which access will service the interrupt, so we're potentially
952 * getting quite a few host interrupts per guest interrupt.
954 vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
957 static uint64_t vfio_bar_read(void *opaque,
958 hwaddr addr, unsigned size)
960 VFIOBAR *bar = opaque;
969 if (pread(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
970 error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
971 __func__, addr, size);
980 data = le16_to_cpu(buf.word);
983 data = le32_to_cpu(buf.dword);
986 hw_error("vfio: unsupported read size, %d bytes\n", size);
992 VFIODevice *vdev = container_of(bar, VFIODevice, bars[bar->nr]);
994 DPRINTF("%s(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx
995 ", %d) = 0x%"PRIx64"\n", __func__, vdev->host.domain,
996 vdev->host.bus, vdev->host.slot, vdev->host.function,
997 bar->nr, addr, size, data);
1001 /* Same as write above */
1002 vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
1007 static const MemoryRegionOps vfio_bar_ops = {
1008 .read = vfio_bar_read,
1009 .write = vfio_bar_write,
1010 .endianness = DEVICE_LITTLE_ENDIAN,
1013 static void vfio_vga_write(void *opaque, hwaddr addr,
1014 uint64_t data, unsigned size)
1016 VFIOVGARegion *region = opaque;
1017 VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1024 off_t offset = vga->fd_offset + region->offset + addr;
1031 buf.word = cpu_to_le16(data);
1034 buf.dword = cpu_to_le32(data);
1037 hw_error("vfio: unsupported write size, %d bytes\n", size);
1041 if (pwrite(vga->fd, &buf, size, offset) != size) {
1042 error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
1043 __func__, region->offset + addr, data, size);
1046 DPRINTF("%s(0x%"HWADDR_PRIx", 0x%"PRIx64", %d)\n",
1047 __func__, region->offset + addr, data, size);
1050 static uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size)
1052 VFIOVGARegion *region = opaque;
1053 VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1061 off_t offset = vga->fd_offset + region->offset + addr;
1063 if (pread(vga->fd, &buf, size, offset) != size) {
1064 error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
1065 __func__, region->offset + addr, size);
1066 return (uint64_t)-1;
1074 data = le16_to_cpu(buf.word);
1077 data = le32_to_cpu(buf.dword);
1080 hw_error("vfio: unsupported read size, %d bytes\n", size);
1084 DPRINTF("%s(0x%"HWADDR_PRIx", %d) = 0x%"PRIx64"\n",
1085 __func__, region->offset + addr, size, data);
1090 static const MemoryRegionOps vfio_vga_ops = {
1091 .read = vfio_vga_read,
1092 .write = vfio_vga_write,
1093 .endianness = DEVICE_LITTLE_ENDIAN,
1097 * Device specific quirks
1100 #define PCI_VENDOR_ID_ATI 0x1002
1103 * Device 1002:68f9 (Advanced Micro Devices [AMD] nee ATI Cedar PRO [Radeon
1104 * HD 5450/6350]) reports the upper byte of the physical address of the
1105 * I/O port BAR4 through VGA register 0x3c3. The BAR is 256 bytes, so the
1106 * lower byte is known to be zero. Probing for this quirk reads 0xff from
1107 * port 0x3c3 on some devices so we store the physical address and replace
1108 * reads with the virtual address any time it matches. XXX Research when
1111 static uint64_t vfio_ati_3c3_quirk_read(void *opaque,
1112 hwaddr addr, unsigned size)
1114 VFIOQuirk *quirk = opaque;
1115 VFIODevice *vdev = quirk->vdev;
1116 PCIDevice *pdev = &vdev->pdev;
1117 uint64_t data = vfio_vga_read(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1120 if (data == quirk->data) {
1121 data = pci_get_byte(pdev->config + PCI_BASE_ADDRESS_4 + 1);
1122 DPRINTF("%s(0x3c3, 1) = 0x%"PRIx64"\n", __func__, data);
1128 static const MemoryRegionOps vfio_ati_3c3_quirk = {
1129 .read = vfio_ati_3c3_quirk_read,
1130 .endianness = DEVICE_LITTLE_ENDIAN,
1133 static void vfio_vga_probe_ati_3c3_quirk(VFIODevice *vdev)
1135 PCIDevice *pdev = &vdev->pdev;
1136 off_t physoffset = vdev->config_offset + PCI_BASE_ADDRESS_4;
1140 if (pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI ||
1141 vdev->bars[4].size < 256) {
1145 /* Get I/O port BAR physical address */
1146 if (pread(vdev->fd, &physbar, 4, physoffset) != 4) {
1147 error_report("vfio: probe failed for ATI/AMD 0x3c3 quirk on device "
1148 "%04x:%02x:%02x.%x", vdev->host.domain,
1149 vdev->host.bus, vdev->host.slot, vdev->host.function);
1153 quirk = g_malloc0(sizeof(*quirk));
1155 quirk->data = (physbar >> 8) & 0xff;
1157 memory_region_init_io(&quirk->mem, NULL, &vfio_ati_3c3_quirk, quirk,
1158 "vfio-ati-3c3-quirk", 1);
1159 memory_region_add_subregion(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem, 3,
1162 QLIST_INSERT_HEAD(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks,
1165 DPRINTF("Enabled ATI/AMD quirk 0x3c3 for device %04x:%02x:%02x.%x\n",
1166 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1167 vdev->host.function);
1171 * Device 1002:68f9 (Advanced Micro Devices [AMD] nee ATI Cedar PRO [Radeon
1172 * HD 5450/6350]) reports the physical address of MMIO BAR0 through a
1173 * write/read operation on I/O port BAR4. When uint32_t 0x4010 is written
1174 * to offset 0x0, the subsequent read from offset 0x4 returns the contents
1175 * of BAR0. Test for this quirk on all ATI/AMD devices. XXX - Note that
1176 * 0x10 is the offset of BAR0 in config sapce, is this a window to all of
1179 static uint64_t vfio_ati_4010_quirk_read(void *opaque,
1180 hwaddr addr, unsigned size)
1182 VFIOQuirk *quirk = opaque;
1183 VFIODevice *vdev = quirk->vdev;
1184 PCIDevice *pdev = &vdev->pdev;
1185 uint64_t data = vfio_bar_read(&vdev->bars[4], addr, size);
1187 if (addr == 4 && size == 4 && quirk->data) {
1188 data = pci_get_long(pdev->config + PCI_BASE_ADDRESS_0);
1189 DPRINTF("%s(BAR4+0x4) = 0x%"PRIx64"\n", __func__, data);
1197 static void vfio_ati_4010_quirk_write(void *opaque, hwaddr addr,
1198 uint64_t data, unsigned size)
1200 VFIOQuirk *quirk = opaque;
1201 VFIODevice *vdev = quirk->vdev;
1203 vfio_bar_write(&vdev->bars[4], addr, data, size);
1205 quirk->data = (addr == 0 && size == 4 && data == 0x4010) ? 1 : 0;
1208 static const MemoryRegionOps vfio_ati_4010_quirk = {
1209 .read = vfio_ati_4010_quirk_read,
1210 .write = vfio_ati_4010_quirk_write,
1211 .endianness = DEVICE_LITTLE_ENDIAN,
1214 static void vfio_probe_ati_4010_quirk(VFIODevice *vdev, int nr)
1216 PCIDevice *pdev = &vdev->pdev;
1217 off_t physoffset = vdev->config_offset + PCI_BASE_ADDRESS_0;
1222 if (!vdev->has_vga || nr != 4 || !vdev->bars[0].size ||
1223 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1227 /* Get I/O port BAR physical address */
1228 if (pread(vdev->fd, &physbar0, 4, physoffset) != 4) {
1229 error_report("vfio: probe failed for ATI/AMD 0x4010 quirk on device "
1230 "%04x:%02x:%02x.%x", vdev->host.domain,
1231 vdev->host.bus, vdev->host.slot, vdev->host.function);
1235 /* Write 0x4010 to I/O port BAR offset 0 */
1236 vfio_bar_write(&vdev->bars[4], 0, 0x4010, 4);
1237 /* Read back result */
1238 data = vfio_bar_read(&vdev->bars[4], 4, 4);
1240 /* If the register matches the physical address of BAR0, we need a quirk */
1241 if (data != physbar0) {
1245 quirk = g_malloc0(sizeof(*quirk));
1248 memory_region_init_io(&quirk->mem, NULL, &vfio_ati_4010_quirk, quirk,
1249 "vfio-ati-4010-quirk", 8);
1250 memory_region_add_subregion_overlap(&vdev->bars[nr].mem, 0, &quirk->mem, 1);
1252 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1254 DPRINTF("Enabled ATI/AMD quirk 0x4010 for device %04x:%02x:%02x.%x\n",
1255 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1256 vdev->host.function);
1260 * Device 1002:5b63 (Advanced Micro Devices [AMD] nee ATI RV370 [Radeon X550])
1261 * retrieves the upper half of the MMIO BAR0 physical address by writing
1262 * 0xf10 to I/O port BAR1 offset 0 and reading the result from offset 6.
1263 * XXX - 0x10 is the offset of BAR0 in PCI config space, this could provide
1264 * full access to config space. Config space is little endian, so the data
1265 * register probably starts at 0x4.
1267 static uint64_t vfio_ati_f10_quirk_read(void *opaque,
1268 hwaddr addr, unsigned size)
1270 VFIOQuirk *quirk = opaque;
1271 VFIODevice *vdev = quirk->vdev;
1272 PCIDevice *pdev = &vdev->pdev;
1273 uint64_t data = vfio_bar_read(&vdev->bars[1], addr, size);
1275 if (addr == 6 && size == 2 && quirk->data) {
1276 data = pci_get_word(pdev->config + PCI_BASE_ADDRESS_0 + 2);
1277 DPRINTF("%s(BAR1+0x6) = 0x%"PRIx64"\n", __func__, data);
1285 static void vfio_ati_f10_quirk_write(void *opaque, hwaddr addr,
1286 uint64_t data, unsigned size)
1288 VFIOQuirk *quirk = opaque;
1289 VFIODevice *vdev = quirk->vdev;
1291 vfio_bar_write(&vdev->bars[1], addr, data, size);
1293 quirk->data = (addr == 0 && size == 4 && data == 0xf10) ? 1 : 0;
1296 static const MemoryRegionOps vfio_ati_f10_quirk = {
1297 .read = vfio_ati_f10_quirk_read,
1298 .write = vfio_ati_f10_quirk_write,
1299 .endianness = DEVICE_LITTLE_ENDIAN,
1302 static void vfio_probe_ati_f10_quirk(VFIODevice *vdev, int nr)
1304 PCIDevice *pdev = &vdev->pdev;
1305 off_t physoffset = vdev->config_offset + PCI_BASE_ADDRESS_0;
1310 if (!vdev->has_vga || nr != 1 || !vdev->bars[0].size ||
1311 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1315 /* Get I/O port BAR physical address */
1316 if (pread(vdev->fd, &physbar0, 4, physoffset) != 4) {
1317 error_report("vfio: probe failed for ATI/AMD 0xf10 quirk on device "
1318 "%04x:%02x:%02x.%x", vdev->host.domain,
1319 vdev->host.bus, vdev->host.slot, vdev->host.function);
1323 vfio_bar_write(&vdev->bars[1], 0, 0xf10, 4);
1324 data = vfio_bar_read(&vdev->bars[1], 0x6, 2);
1326 /* If the register matches the physical address of BAR0, we need a quirk */
1327 if (data != (le32_to_cpu(physbar0) >> 16)) {
1331 quirk = g_malloc0(sizeof(*quirk));
1334 memory_region_init_io(&quirk->mem, NULL, &vfio_ati_f10_quirk, quirk,
1335 "vfio-ati-f10-quirk", 8);
1336 memory_region_add_subregion_overlap(&vdev->bars[nr].mem, 0, &quirk->mem, 1);
1338 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1340 DPRINTF("Enabled ATI/AMD quirk 0xf10 for device %04x:%02x:%02x.%x\n",
1341 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1342 vdev->host.function);
1345 #define PCI_VENDOR_ID_NVIDIA 0x10de
1348 * Nvidia has several different methods to get to config space, the
1349 * nouveu project has several of these documented here:
1350 * https://github.com/pathscale/envytools/tree/master/hwdocs
1352 * The first quirk is actually not documented in envytools and is found
1353 * on 10de:01d1 (NVIDIA Corporation G72 [GeForce 7300 LE]). This is an
1354 * NV46 chipset. The backdoor uses the legacy VGA I/O ports to access
1355 * the mirror of PCI config space found at BAR0 offset 0x1800. The access
1356 * sequence first writes 0x338 to I/O port 0x3d4. The target offset is
1357 * then written to 0x3d0. Finally 0x538 is written for a read and 0x738
1358 * is written for a write to 0x3d4. The BAR0 offset is then accessible
1359 * through 0x3d0. This quirk doesn't seem to be necessary on newer cards
1360 * that use the I/O port BAR5 window but it doesn't hurt to leave it.
1370 static uint64_t vfio_nvidia_3d0_quirk_read(void *opaque,
1371 hwaddr addr, unsigned size)
1373 VFIOQuirk *quirk = opaque;
1374 VFIODevice *vdev = quirk->vdev;
1375 PCIDevice *pdev = &vdev->pdev;
1376 uint64_t data = vfio_vga_read(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1379 if (quirk->data == NV_3D0_READ && addr == 0) {
1380 data = vfio_pci_read_config(pdev, quirk->data2, size);
1381 DPRINTF("%s(0x3d0, %d) = 0x%"PRIx64"\n", __func__, size, data);
1384 quirk->data = NV_3D0_NONE;
1389 static void vfio_nvidia_3d0_quirk_write(void *opaque, hwaddr addr,
1390 uint64_t data, unsigned size)
1392 VFIOQuirk *quirk = opaque;
1393 VFIODevice *vdev = quirk->vdev;
1394 PCIDevice *pdev = &vdev->pdev;
1396 switch (quirk->data) {
1398 if (addr == 4 && data == 0x338) {
1399 quirk->data = NV_3D0_SELECT;
1403 quirk->data = NV_3D0_NONE;
1404 if (addr == 0 && (data & ~0xff) == 0x1800) {
1405 quirk->data = NV_3D0_WINDOW;
1406 quirk->data2 = data & 0xff;
1410 quirk->data = NV_3D0_NONE;
1412 if (data == 0x538) {
1413 quirk->data = NV_3D0_READ;
1414 } else if (data == 0x738) {
1415 quirk->data = NV_3D0_WRITE;
1420 quirk->data = NV_3D0_NONE;
1422 vfio_pci_write_config(pdev, quirk->data2, data, size);
1423 DPRINTF("%s(0x3d0, 0x%"PRIx64", %d)\n", __func__, data, size);
1428 quirk->data = NV_3D0_NONE;
1431 vfio_vga_write(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1432 addr + 0x10, data, size);
1435 static const MemoryRegionOps vfio_nvidia_3d0_quirk = {
1436 .read = vfio_nvidia_3d0_quirk_read,
1437 .write = vfio_nvidia_3d0_quirk_write,
1438 .endianness = DEVICE_LITTLE_ENDIAN,
1441 static void vfio_vga_probe_nvidia_3d0_quirk(VFIODevice *vdev)
1443 PCIDevice *pdev = &vdev->pdev;
1446 if (pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA ||
1447 !vdev->bars[1].size) {
1451 quirk = g_malloc0(sizeof(*quirk));
1454 memory_region_init_io(&quirk->mem, NULL, &vfio_nvidia_3d0_quirk, quirk,
1455 "vfio-nvidia-3d0-quirk", 6);
1456 memory_region_add_subregion(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem,
1459 QLIST_INSERT_HEAD(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks,
1462 DPRINTF("Enabled NVIDIA VGA 0x3d0 quirk for device %04x:%02x:%02x.%x\n",
1463 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1464 vdev->host.function);
1468 * The second quirk is documented in envytools. The I/O port BAR5 is just
1469 * a set of address/data ports to the MMIO BARs. The BAR we care about is
1470 * again BAR0. This backdoor is apparently a bit newer than the one above
1471 * so we need to not only trap 256 bytes @0x1800, but all of PCI config
1472 * space, including extended space is available at the 4k @0x88000.
1475 NV_BAR5_ADDRESS = 0x1,
1476 NV_BAR5_ENABLE = 0x2,
1477 NV_BAR5_MASTER = 0x4,
1478 NV_BAR5_VALID = 0x7,
1481 static uint64_t vfio_nvidia_bar5_window_quirk_read(void *opaque,
1482 hwaddr addr, unsigned size)
1484 VFIOQuirk *quirk = opaque;
1485 VFIODevice *vdev = quirk->vdev;
1486 uint64_t data = vfio_bar_read(&vdev->bars[5], addr, size);
1488 if (addr == 0xc && quirk->data == NV_BAR5_VALID) {
1489 data = vfio_pci_read_config(&vdev->pdev, quirk->data2, size);
1490 DPRINTF("%s(%04x:%02x:%02x.%x:BAR5+0x%"HWADDR_PRIx", %d) = 0x%"
1491 PRIx64"\n", __func__, vdev->host.domain, vdev->host.bus,
1492 vdev->host.slot, vdev->host.function, addr, size, data);
1498 static void vfio_nvidia_bar5_window_quirk_write(void *opaque, hwaddr addr,
1499 uint64_t data, unsigned size)
1501 VFIOQuirk *quirk = opaque;
1502 VFIODevice *vdev = quirk->vdev;
1505 * Use quirk->data to track enables and quirk->data2 for the offset
1510 quirk->data |= NV_BAR5_MASTER;
1512 quirk->data &= ~NV_BAR5_MASTER;
1517 quirk->data |= NV_BAR5_ENABLE;
1519 quirk->data &= ~NV_BAR5_ENABLE;
1523 if (quirk->data & NV_BAR5_MASTER) {
1524 if ((data & ~0xfff) == 0x88000) {
1525 quirk->data |= NV_BAR5_ADDRESS;
1526 quirk->data2 = data & 0xfff;
1527 } else if ((data & ~0xff) == 0x1800) {
1528 quirk->data |= NV_BAR5_ADDRESS;
1529 quirk->data2 = data & 0xff;
1531 quirk->data &= ~NV_BAR5_ADDRESS;
1536 if (quirk->data == NV_BAR5_VALID) {
1537 vfio_pci_write_config(&vdev->pdev, quirk->data2, data, size);
1538 DPRINTF("%s(%04x:%02x:%02x.%x:BAR5+0x%"HWADDR_PRIx", 0x%"
1539 PRIx64", %d)\n", __func__, vdev->host.domain,
1540 vdev->host.bus, vdev->host.slot, vdev->host.function,
1546 vfio_bar_write(&vdev->bars[5], addr, data, size);
1549 static const MemoryRegionOps vfio_nvidia_bar5_window_quirk = {
1550 .read = vfio_nvidia_bar5_window_quirk_read,
1551 .write = vfio_nvidia_bar5_window_quirk_write,
1552 .valid.min_access_size = 4,
1553 .endianness = DEVICE_LITTLE_ENDIAN,
1556 static void vfio_probe_nvidia_bar5_window_quirk(VFIODevice *vdev, int nr)
1558 PCIDevice *pdev = &vdev->pdev;
1561 if (!vdev->has_vga || nr != 5 ||
1562 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1566 quirk = g_malloc0(sizeof(*quirk));
1569 memory_region_init_io(&quirk->mem, NULL, &vfio_nvidia_bar5_window_quirk, quirk,
1570 "vfio-nvidia-bar5-window-quirk", 16);
1571 memory_region_add_subregion_overlap(&vdev->bars[nr].mem, 0, &quirk->mem, 1);
1573 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1575 DPRINTF("Enabled NVIDIA BAR5 window quirk for device %04x:%02x:%02x.%x\n",
1576 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1577 vdev->host.function);
1581 * Finally, BAR0 itself. We want to redirect any accesses to either
1582 * 0x1800 or 0x88000 through the PCI config space access functions.
1584 * NB - quirk at a page granularity or else they don't seem to work when
1587 * Here's offset 0x88000...
1589 static uint64_t vfio_nvidia_bar0_88000_quirk_read(void *opaque,
1590 hwaddr addr, unsigned size)
1592 VFIOQuirk *quirk = opaque;
1593 VFIODevice *vdev = quirk->vdev;
1594 hwaddr base = 0x88000 & TARGET_PAGE_MASK;
1595 hwaddr offset = 0x88000 & ~TARGET_PAGE_MASK;
1596 uint64_t data = vfio_bar_read(&vdev->bars[0], addr + base, size);
1598 if (ranges_overlap(addr, size, offset, PCI_CONFIG_SPACE_SIZE)) {
1599 data = vfio_pci_read_config(&vdev->pdev, addr - offset, size);
1601 DPRINTF("%s(%04x:%02x:%02x.%x:BAR0+0x%"HWADDR_PRIx", %d) = 0x%"
1602 PRIx64"\n", __func__, vdev->host.domain, vdev->host.bus,
1603 vdev->host.slot, vdev->host.function, addr + base, size, data);
1609 static void vfio_nvidia_bar0_88000_quirk_write(void *opaque, hwaddr addr,
1610 uint64_t data, unsigned size)
1612 VFIOQuirk *quirk = opaque;
1613 VFIODevice *vdev = quirk->vdev;
1614 hwaddr base = 0x88000 & TARGET_PAGE_MASK;
1615 hwaddr offset = 0x88000 & ~TARGET_PAGE_MASK;
1617 if (ranges_overlap(addr, size, offset, PCI_CONFIG_SPACE_SIZE)) {
1618 vfio_pci_write_config(&vdev->pdev, addr - offset, data, size);
1620 DPRINTF("%s(%04x:%02x:%02x.%x:BAR0+0x%"HWADDR_PRIx", 0x%"
1621 PRIx64", %d)\n", __func__, vdev->host.domain, vdev->host.bus,
1622 vdev->host.slot, vdev->host.function, addr + base, data, size);
1624 vfio_bar_write(&vdev->bars[0], addr + base, data, size);
1628 static const MemoryRegionOps vfio_nvidia_bar0_88000_quirk = {
1629 .read = vfio_nvidia_bar0_88000_quirk_read,
1630 .write = vfio_nvidia_bar0_88000_quirk_write,
1631 .endianness = DEVICE_LITTLE_ENDIAN,
1634 static void vfio_probe_nvidia_bar0_88000_quirk(VFIODevice *vdev, int nr)
1636 PCIDevice *pdev = &vdev->pdev;
1639 if (!vdev->has_vga || nr != 0 ||
1640 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1644 quirk = g_malloc0(sizeof(*quirk));
1647 memory_region_init_io(&quirk->mem, NULL, &vfio_nvidia_bar0_88000_quirk, quirk,
1648 "vfio-nvidia-bar0-88000-quirk",
1649 TARGET_PAGE_ALIGN(PCIE_CONFIG_SPACE_SIZE));
1650 memory_region_add_subregion_overlap(&vdev->bars[nr].mem,
1651 0x88000 & TARGET_PAGE_MASK,
1654 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1656 DPRINTF("Enabled NVIDIA BAR0 0x88000 quirk for device %04x:%02x:%02x.%x\n",
1657 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1658 vdev->host.function);
1662 * And here's the same for BAR0 offset 0x1800...
1664 static uint64_t vfio_nvidia_bar0_1800_quirk_read(void *opaque,
1665 hwaddr addr, unsigned size)
1667 VFIOQuirk *quirk = opaque;
1668 VFIODevice *vdev = quirk->vdev;
1669 hwaddr base = 0x1800 & TARGET_PAGE_MASK;
1670 hwaddr offset = 0x1800 & ~TARGET_PAGE_MASK;
1671 uint64_t data = vfio_bar_read(&vdev->bars[0], addr + base, size);
1673 if (ranges_overlap(addr, size, offset, PCI_CONFIG_SPACE_SIZE)) {
1674 data = vfio_pci_read_config(&vdev->pdev, addr - offset, size);
1676 DPRINTF("%s(%04x:%02x:%02x.%x:BAR0+0x%"HWADDR_PRIx", %d) = 0x%"
1677 PRIx64"\n", __func__, vdev->host.domain, vdev->host.bus,
1678 vdev->host.slot, vdev->host.function, addr + base, size, data);
1684 static void vfio_nvidia_bar0_1800_quirk_write(void *opaque, hwaddr addr,
1685 uint64_t data, unsigned size)
1687 VFIOQuirk *quirk = opaque;
1688 VFIODevice *vdev = quirk->vdev;
1689 hwaddr base = 0x1800 & TARGET_PAGE_MASK;
1690 hwaddr offset = 0x1800 & ~TARGET_PAGE_MASK;
1692 if (ranges_overlap(addr, size, offset, PCI_CONFIG_SPACE_SIZE)) {
1693 vfio_pci_write_config(&vdev->pdev, addr - offset, data, size);
1695 DPRINTF("%s(%04x:%02x:%02x.%x:BAR0+0x%"HWADDR_PRIx", 0x%"
1696 PRIx64", %d)\n", __func__, vdev->host.domain, vdev->host.bus,
1697 vdev->host.slot, vdev->host.function, addr + base, data, size);
1699 vfio_bar_write(&vdev->bars[0], addr + base, data, size);
1703 static const MemoryRegionOps vfio_nvidia_bar0_1800_quirk = {
1704 .read = vfio_nvidia_bar0_1800_quirk_read,
1705 .write = vfio_nvidia_bar0_1800_quirk_write,
1706 .endianness = DEVICE_LITTLE_ENDIAN,
1709 static void vfio_probe_nvidia_bar0_1800_quirk(VFIODevice *vdev, int nr)
1711 PCIDevice *pdev = &vdev->pdev;
1714 if (!vdev->has_vga || nr != 0 ||
1715 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1719 /* Log the chipset ID */
1720 DPRINTF("Nvidia NV%02x\n",
1721 (unsigned int)(vfio_bar_read(&vdev->bars[0], 0, 4) >> 20) & 0xff);
1723 quirk = g_malloc0(sizeof(*quirk));
1726 memory_region_init_io(&quirk->mem, NULL, &vfio_nvidia_bar0_1800_quirk, quirk,
1727 "vfio-nvidia-bar0-1800-quirk",
1728 TARGET_PAGE_ALIGN(PCI_CONFIG_SPACE_SIZE));
1729 memory_region_add_subregion_overlap(&vdev->bars[nr].mem,
1730 0x1800 & TARGET_PAGE_MASK,
1733 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1735 DPRINTF("Enabled NVIDIA BAR0 0x1800 quirk for device %04x:%02x:%02x.%x\n",
1736 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1737 vdev->host.function);
1741 * TODO - Some Nvidia devices provide config access to their companion HDA
1742 * device and even to their parent bridge via these config space mirrors.
1743 * Add quirks for those regions.
1747 * Common quirk probe entry points.
1749 static void vfio_vga_quirk_setup(VFIODevice *vdev)
1751 vfio_vga_probe_ati_3c3_quirk(vdev);
1752 vfio_vga_probe_nvidia_3d0_quirk(vdev);
1755 static void vfio_vga_quirk_teardown(VFIODevice *vdev)
1759 for (i = 0; i < ARRAY_SIZE(vdev->vga.region); i++) {
1760 while (!QLIST_EMPTY(&vdev->vga.region[i].quirks)) {
1761 VFIOQuirk *quirk = QLIST_FIRST(&vdev->vga.region[i].quirks);
1762 memory_region_del_subregion(&vdev->vga.region[i].mem, &quirk->mem);
1763 QLIST_REMOVE(quirk, next);
1769 static void vfio_bar_quirk_setup(VFIODevice *vdev, int nr)
1771 vfio_probe_ati_4010_quirk(vdev, nr);
1772 vfio_probe_ati_f10_quirk(vdev, nr);
1773 vfio_probe_nvidia_bar5_window_quirk(vdev, nr);
1774 vfio_probe_nvidia_bar0_88000_quirk(vdev, nr);
1775 vfio_probe_nvidia_bar0_1800_quirk(vdev, nr);
1778 static void vfio_bar_quirk_teardown(VFIODevice *vdev, int nr)
1780 VFIOBAR *bar = &vdev->bars[nr];
1782 while (!QLIST_EMPTY(&bar->quirks)) {
1783 VFIOQuirk *quirk = QLIST_FIRST(&bar->quirks);
1784 memory_region_del_subregion(&bar->mem, &quirk->mem);
1785 QLIST_REMOVE(quirk, next);
1793 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
1795 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1796 uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val;
1798 memcpy(&emu_bits, vdev->emulated_config_bits + addr, len);
1799 emu_bits = le32_to_cpu(emu_bits);
1802 emu_val = pci_default_read_config(pdev, addr, len);
1805 if (~emu_bits & (0xffffffffU >> (32 - len * 8))) {
1808 ret = pread(vdev->fd, &phys_val, len, vdev->config_offset + addr);
1810 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m",
1811 __func__, vdev->host.domain, vdev->host.bus,
1812 vdev->host.slot, vdev->host.function, addr, len);
1815 phys_val = le32_to_cpu(phys_val);
1818 val = (emu_val & emu_bits) | (phys_val & ~emu_bits);
1820 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, len=0x%x) %x\n", __func__,
1821 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1822 vdev->host.function, addr, len, val);
1827 static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
1828 uint32_t val, int len)
1830 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1831 uint32_t val_le = cpu_to_le32(val);
1833 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, 0x%x, len=0x%x)\n", __func__,
1834 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1835 vdev->host.function, addr, val, len);
1837 /* Write everything to VFIO, let it filter out what we can't write */
1838 if (pwrite(vdev->fd, &val_le, len, vdev->config_offset + addr) != len) {
1839 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m",
1840 __func__, vdev->host.domain, vdev->host.bus,
1841 vdev->host.slot, vdev->host.function, addr, val, len);
1844 /* MSI/MSI-X Enabling/Disabling */
1845 if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
1846 ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
1847 int is_enabled, was_enabled = msi_enabled(pdev);
1849 pci_default_write_config(pdev, addr, val, len);
1851 is_enabled = msi_enabled(pdev);
1853 if (!was_enabled && is_enabled) {
1854 vfio_enable_msi(vdev);
1855 } else if (was_enabled && !is_enabled) {
1856 vfio_disable_msi(vdev);
1858 } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
1859 ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
1860 int is_enabled, was_enabled = msix_enabled(pdev);
1862 pci_default_write_config(pdev, addr, val, len);
1864 is_enabled = msix_enabled(pdev);
1866 if (!was_enabled && is_enabled) {
1867 vfio_enable_msix(vdev);
1868 } else if (was_enabled && !is_enabled) {
1869 vfio_disable_msix(vdev);
1872 /* Write everything to QEMU to keep emulated bits correct */
1873 pci_default_write_config(pdev, addr, val, len);
1878 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
1880 static int vfio_dma_unmap(VFIOContainer *container,
1881 hwaddr iova, ram_addr_t size)
1883 struct vfio_iommu_type1_dma_unmap unmap = {
1884 .argsz = sizeof(unmap),
1890 if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
1891 DPRINTF("VFIO_UNMAP_DMA: %d\n", -errno);
1898 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
1899 ram_addr_t size, void *vaddr, bool readonly)
1901 struct vfio_iommu_type1_dma_map map = {
1902 .argsz = sizeof(map),
1903 .flags = VFIO_DMA_MAP_FLAG_READ,
1904 .vaddr = (__u64)(uintptr_t)vaddr,
1910 map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
1914 * Try the mapping, if it fails with EBUSY, unmap the region and try
1915 * again. This shouldn't be necessary, but we sometimes see it in
1916 * the the VGA ROM space.
1918 if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
1919 (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
1920 ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
1924 DPRINTF("VFIO_MAP_DMA: %d\n", -errno);
1928 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
1930 return !memory_region_is_ram(section->mr);
1933 static void vfio_listener_region_add(MemoryListener *listener,
1934 MemoryRegionSection *section)
1936 VFIOContainer *container = container_of(listener, VFIOContainer,
1937 iommu_data.listener);
1942 assert(!memory_region_is_iommu(section->mr));
1944 if (vfio_listener_skipped_section(section)) {
1945 DPRINTF("SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
1946 section->offset_within_address_space,
1947 section->offset_within_address_space + section->size - 1);
1951 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
1952 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
1953 error_report("%s received unaligned region", __func__);
1957 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
1958 end = (section->offset_within_address_space + int128_get64(section->size)) &
1965 vaddr = memory_region_get_ram_ptr(section->mr) +
1966 section->offset_within_region +
1967 (iova - section->offset_within_address_space);
1969 DPRINTF("region_add %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
1970 iova, end - 1, vaddr);
1972 memory_region_ref(section->mr);
1973 ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
1975 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
1976 "0x%"HWADDR_PRIx", %p) = %d (%m)",
1977 container, iova, end - iova, vaddr, ret);
1981 static void vfio_listener_region_del(MemoryListener *listener,
1982 MemoryRegionSection *section)
1984 VFIOContainer *container = container_of(listener, VFIOContainer,
1985 iommu_data.listener);
1989 if (vfio_listener_skipped_section(section)) {
1990 DPRINTF("SKIPPING region_del %"HWADDR_PRIx" - %"PRIx64"\n",
1991 section->offset_within_address_space,
1992 section->offset_within_address_space + section->size - 1);
1996 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
1997 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
1998 error_report("%s received unaligned region", __func__);
2002 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
2003 end = (section->offset_within_address_space + int128_get64(section->size)) &
2010 DPRINTF("region_del %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
2013 ret = vfio_dma_unmap(container, iova, end - iova);
2014 memory_region_unref(section->mr);
2016 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
2017 "0x%"HWADDR_PRIx") = %d (%m)",
2018 container, iova, end - iova, ret);
2022 static MemoryListener vfio_memory_listener = {
2023 .region_add = vfio_listener_region_add,
2024 .region_del = vfio_listener_region_del,
2027 static void vfio_listener_release(VFIOContainer *container)
2029 memory_listener_unregister(&container->iommu_data.listener);
2035 static void vfio_disable_interrupts(VFIODevice *vdev)
2037 switch (vdev->interrupt) {
2039 vfio_disable_intx(vdev);
2042 vfio_disable_msi(vdev);
2045 vfio_disable_msix(vdev);
2050 static int vfio_setup_msi(VFIODevice *vdev, int pos)
2053 bool msi_64bit, msi_maskbit;
2056 if (pread(vdev->fd, &ctrl, sizeof(ctrl),
2057 vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
2060 ctrl = le16_to_cpu(ctrl);
2062 msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
2063 msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
2064 entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
2066 DPRINTF("%04x:%02x:%02x.%x PCI MSI CAP @0x%x\n", vdev->host.domain,
2067 vdev->host.bus, vdev->host.slot, vdev->host.function, pos);
2069 ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit);
2071 if (ret == -ENOTSUP) {
2074 error_report("vfio: msi_init failed");
2077 vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
2083 * We don't have any control over how pci_add_capability() inserts
2084 * capabilities into the chain. In order to setup MSI-X we need a
2085 * MemoryRegion for the BAR. In order to setup the BAR and not
2086 * attempt to mmap the MSI-X table area, which VFIO won't allow, we
2087 * need to first look for where the MSI-X table lives. So we
2088 * unfortunately split MSI-X setup across two functions.
2090 static int vfio_early_setup_msix(VFIODevice *vdev)
2094 uint32_t table, pba;
2096 pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
2101 if (pread(vdev->fd, &ctrl, sizeof(ctrl),
2102 vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
2106 if (pread(vdev->fd, &table, sizeof(table),
2107 vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
2111 if (pread(vdev->fd, &pba, sizeof(pba),
2112 vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
2116 ctrl = le16_to_cpu(ctrl);
2117 table = le32_to_cpu(table);
2118 pba = le32_to_cpu(pba);
2120 vdev->msix = g_malloc0(sizeof(*(vdev->msix)));
2121 vdev->msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
2122 vdev->msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
2123 vdev->msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
2124 vdev->msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
2125 vdev->msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
2127 DPRINTF("%04x:%02x:%02x.%x "
2128 "PCI MSI-X CAP @0x%x, BAR %d, offset 0x%x, entries %d\n",
2129 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2130 vdev->host.function, pos, vdev->msix->table_bar,
2131 vdev->msix->table_offset, vdev->msix->entries);
2136 static int vfio_setup_msix(VFIODevice *vdev, int pos)
2140 ret = msix_init(&vdev->pdev, vdev->msix->entries,
2141 &vdev->bars[vdev->msix->table_bar].mem,
2142 vdev->msix->table_bar, vdev->msix->table_offset,
2143 &vdev->bars[vdev->msix->pba_bar].mem,
2144 vdev->msix->pba_bar, vdev->msix->pba_offset, pos);
2146 if (ret == -ENOTSUP) {
2149 error_report("vfio: msix_init failed");
2156 static void vfio_teardown_msi(VFIODevice *vdev)
2158 msi_uninit(&vdev->pdev);
2161 msix_uninit(&vdev->pdev, &vdev->bars[vdev->msix->table_bar].mem,
2162 &vdev->bars[vdev->msix->pba_bar].mem);
2169 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled)
2173 for (i = 0; i < PCI_ROM_SLOT; i++) {
2174 VFIOBAR *bar = &vdev->bars[i];
2180 memory_region_set_enabled(&bar->mmap_mem, enabled);
2181 if (vdev->msix && vdev->msix->table_bar == i) {
2182 memory_region_set_enabled(&vdev->msix->mmap_mem, enabled);
2187 static void vfio_unmap_bar(VFIODevice *vdev, int nr)
2189 VFIOBAR *bar = &vdev->bars[nr];
2195 vfio_bar_quirk_teardown(vdev, nr);
2197 memory_region_del_subregion(&bar->mem, &bar->mmap_mem);
2198 munmap(bar->mmap, memory_region_size(&bar->mmap_mem));
2200 if (vdev->msix && vdev->msix->table_bar == nr) {
2201 memory_region_del_subregion(&bar->mem, &vdev->msix->mmap_mem);
2202 munmap(vdev->msix->mmap, memory_region_size(&vdev->msix->mmap_mem));
2205 memory_region_destroy(&bar->mem);
2208 static int vfio_mmap_bar(VFIOBAR *bar, MemoryRegion *mem, MemoryRegion *submem,
2209 void **map, size_t size, off_t offset,
2214 if (VFIO_ALLOW_MMAP && size && bar->flags & VFIO_REGION_INFO_FLAG_MMAP) {
2217 if (bar->flags & VFIO_REGION_INFO_FLAG_READ) {
2221 if (bar->flags & VFIO_REGION_INFO_FLAG_WRITE) {
2225 *map = mmap(NULL, size, prot, MAP_SHARED,
2226 bar->fd, bar->fd_offset + offset);
2227 if (*map == MAP_FAILED) {
2233 memory_region_init_ram_ptr(submem, NULL, name, size, *map);
2236 /* Create a zero sized sub-region to make cleanup easy. */
2237 memory_region_init(submem, NULL, name, 0);
2240 memory_region_add_subregion(mem, offset, submem);
2245 static void vfio_map_bar(VFIODevice *vdev, int nr)
2247 VFIOBAR *bar = &vdev->bars[nr];
2248 unsigned size = bar->size;
2254 /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
2259 snprintf(name, sizeof(name), "VFIO %04x:%02x:%02x.%x BAR %d",
2260 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2261 vdev->host.function, nr);
2263 /* Determine what type of BAR this is for registration */
2264 ret = pread(vdev->fd, &pci_bar, sizeof(pci_bar),
2265 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
2266 if (ret != sizeof(pci_bar)) {
2267 error_report("vfio: Failed to read BAR %d (%m)", nr);
2271 pci_bar = le32_to_cpu(pci_bar);
2272 type = pci_bar & (pci_bar & PCI_BASE_ADDRESS_SPACE_IO ?
2273 ~PCI_BASE_ADDRESS_IO_MASK : ~PCI_BASE_ADDRESS_MEM_MASK);
2275 /* A "slow" read/write mapping underlies all BARs */
2276 memory_region_init_io(&bar->mem, NULL, &vfio_bar_ops, bar, name, size);
2277 pci_register_bar(&vdev->pdev, nr, type, &bar->mem);
2280 * We can't mmap areas overlapping the MSIX vector table, so we
2281 * potentially insert a direct-mapped subregion before and after it.
2283 if (vdev->msix && vdev->msix->table_bar == nr) {
2284 size = vdev->msix->table_offset & TARGET_PAGE_MASK;
2287 strncat(name, " mmap", sizeof(name) - strlen(name) - 1);
2288 if (vfio_mmap_bar(bar, &bar->mem,
2289 &bar->mmap_mem, &bar->mmap, size, 0, name)) {
2290 error_report("%s unsupported. Performance may be slow", name);
2293 if (vdev->msix && vdev->msix->table_bar == nr) {
2296 start = TARGET_PAGE_ALIGN(vdev->msix->table_offset +
2297 (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
2299 size = start < bar->size ? bar->size - start : 0;
2300 strncat(name, " msix-hi", sizeof(name) - strlen(name) - 1);
2301 /* VFIOMSIXInfo contains another MemoryRegion for this mapping */
2302 if (vfio_mmap_bar(bar, &bar->mem, &vdev->msix->mmap_mem,
2303 &vdev->msix->mmap, size, start, name)) {
2304 error_report("%s unsupported. Performance may be slow", name);
2308 vfio_bar_quirk_setup(vdev, nr);
2311 static void vfio_map_bars(VFIODevice *vdev)
2315 for (i = 0; i < PCI_ROM_SLOT; i++) {
2316 vfio_map_bar(vdev, i);
2319 if (vdev->has_vga) {
2320 memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_MEM].mem, NULL,
2322 &vdev->vga.region[QEMU_PCI_VGA_MEM],
2323 "vfio-vga-mmio@0xa0000",
2324 QEMU_PCI_VGA_MEM_SIZE);
2325 memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem, NULL,
2327 &vdev->vga.region[QEMU_PCI_VGA_IO_LO],
2328 "vfio-vga-io@0x3b0",
2329 QEMU_PCI_VGA_IO_LO_SIZE);
2330 memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem, NULL,
2332 &vdev->vga.region[QEMU_PCI_VGA_IO_HI],
2333 "vfio-vga-io@0x3c0",
2334 QEMU_PCI_VGA_IO_HI_SIZE);
2336 pci_register_vga(&vdev->pdev, &vdev->vga.region[QEMU_PCI_VGA_MEM].mem,
2337 &vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem,
2338 &vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem);
2339 vfio_vga_quirk_setup(vdev);
2343 static void vfio_unmap_bars(VFIODevice *vdev)
2347 for (i = 0; i < PCI_ROM_SLOT; i++) {
2348 vfio_unmap_bar(vdev, i);
2351 if (vdev->has_vga) {
2352 vfio_vga_quirk_teardown(vdev);
2353 pci_unregister_vga(&vdev->pdev);
2354 memory_region_destroy(&vdev->vga.region[QEMU_PCI_VGA_MEM].mem);
2355 memory_region_destroy(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem);
2356 memory_region_destroy(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem);
2363 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
2365 uint8_t tmp, next = 0xff;
2367 for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
2368 tmp = pdev->config[tmp + 1]) {
2369 if (tmp > pos && tmp < next) {
2377 static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask)
2379 pci_set_word(buf, (pci_get_word(buf) & ~mask) | val);
2382 static void vfio_add_emulated_word(VFIODevice *vdev, int pos,
2383 uint16_t val, uint16_t mask)
2385 vfio_set_word_bits(vdev->pdev.config + pos, val, mask);
2386 vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask);
2387 vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask);
2390 static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask)
2392 pci_set_long(buf, (pci_get_long(buf) & ~mask) | val);
2395 static void vfio_add_emulated_long(VFIODevice *vdev, int pos,
2396 uint32_t val, uint32_t mask)
2398 vfio_set_long_bits(vdev->pdev.config + pos, val, mask);
2399 vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask);
2400 vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask);
2403 static int vfio_setup_pcie_cap(VFIODevice *vdev, int pos, uint8_t size)
2408 flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS);
2409 type = (flags & PCI_EXP_FLAGS_TYPE) >> 4;
2411 if (type != PCI_EXP_TYPE_ENDPOINT &&
2412 type != PCI_EXP_TYPE_LEG_END &&
2413 type != PCI_EXP_TYPE_RC_END) {
2415 error_report("vfio: Assignment of PCIe type 0x%x "
2416 "devices is not currently supported", type);
2420 if (!pci_bus_is_express(vdev->pdev.bus)) {
2422 * Use express capability as-is on PCI bus. It doesn't make much
2423 * sense to even expose, but some drivers (ex. tg3) depend on it
2424 * and guests don't seem to be particular about it. We'll need
2425 * to revist this or force express devices to express buses if we
2426 * ever expose an IOMMU to the guest.
2428 } else if (pci_bus_is_root(vdev->pdev.bus)) {
2430 * On a Root Complex bus Endpoints become Root Complex Integrated
2431 * Endpoints, which changes the type and clears the LNK & LNK2 fields.
2433 if (type == PCI_EXP_TYPE_ENDPOINT) {
2434 vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
2435 PCI_EXP_TYPE_RC_END << 4,
2436 PCI_EXP_FLAGS_TYPE);
2438 /* Link Capabilities, Status, and Control goes away */
2439 if (size > PCI_EXP_LNKCTL) {
2440 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0);
2441 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
2442 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0);
2444 #ifndef PCI_EXP_LNKCAP2
2445 #define PCI_EXP_LNKCAP2 44
2447 #ifndef PCI_EXP_LNKSTA2
2448 #define PCI_EXP_LNKSTA2 50
2450 /* Link 2 Capabilities, Status, and Control goes away */
2451 if (size > PCI_EXP_LNKCAP2) {
2452 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0);
2453 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0);
2454 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0);
2458 } else if (type == PCI_EXP_TYPE_LEG_END) {
2460 * Legacy endpoints don't belong on the root complex. Windows
2461 * seems to be happier with devices if we skip the capability.
2468 * Convert Root Complex Integrated Endpoints to regular endpoints.
2469 * These devices don't support LNK/LNK2 capabilities, so make them up.
2471 if (type == PCI_EXP_TYPE_RC_END) {
2472 vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
2473 PCI_EXP_TYPE_ENDPOINT << 4,
2474 PCI_EXP_FLAGS_TYPE);
2475 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP,
2476 PCI_EXP_LNK_MLW_1 | PCI_EXP_LNK_LS_25, ~0);
2477 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
2480 /* Mark the Link Status bits as emulated to allow virtual negotiation */
2481 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA,
2482 pci_get_word(vdev->pdev.config + pos +
2484 PCI_EXP_LNKCAP_MLW | PCI_EXP_LNKCAP_SLS);
2487 pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size);
2489 vdev->pdev.exp.exp_cap = pos;
2495 static int vfio_add_std_cap(VFIODevice *vdev, uint8_t pos)
2497 PCIDevice *pdev = &vdev->pdev;
2498 uint8_t cap_id, next, size;
2501 cap_id = pdev->config[pos];
2502 next = pdev->config[pos + 1];
2505 * If it becomes important to configure capabilities to their actual
2506 * size, use this as the default when it's something we don't recognize.
2507 * Since QEMU doesn't actually handle many of the config accesses,
2508 * exact size doesn't seem worthwhile.
2510 size = vfio_std_cap_max_size(pdev, pos);
2513 * pci_add_capability always inserts the new capability at the head
2514 * of the chain. Therefore to end up with a chain that matches the
2515 * physical device, we insert from the end by making this recursive.
2516 * This is also why we pre-caclulate size above as cached config space
2517 * will be changed as we unwind the stack.
2520 ret = vfio_add_std_cap(vdev, next);
2525 /* Begin the rebuild, use QEMU emulated list bits */
2526 pdev->config[PCI_CAPABILITY_LIST] = 0;
2527 vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
2528 vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
2531 /* Use emulated next pointer to allow dropping caps */
2532 pci_set_byte(vdev->emulated_config_bits + pos + 1, 0xff);
2535 case PCI_CAP_ID_MSI:
2536 ret = vfio_setup_msi(vdev, pos);
2538 case PCI_CAP_ID_EXP:
2539 ret = vfio_setup_pcie_cap(vdev, pos, size);
2541 case PCI_CAP_ID_MSIX:
2542 ret = vfio_setup_msix(vdev, pos);
2547 ret = pci_add_capability(pdev, cap_id, pos, size);
2552 error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability "
2553 "0x%x[0x%x]@0x%x: %d", vdev->host.domain,
2554 vdev->host.bus, vdev->host.slot, vdev->host.function,
2555 cap_id, size, pos, ret);
2562 static int vfio_add_capabilities(VFIODevice *vdev)
2564 PCIDevice *pdev = &vdev->pdev;
2566 if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
2567 !pdev->config[PCI_CAPABILITY_LIST]) {
2568 return 0; /* Nothing to add */
2571 return vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST]);
2574 static int vfio_load_rom(VFIODevice *vdev)
2576 uint64_t size = vdev->rom_size;
2578 off_t off = 0, voff = vdev->rom_offset;
2582 /* If loading ROM from file, pci handles it */
2583 if (vdev->pdev.romfile || !vdev->pdev.rom_bar || !size) {
2587 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
2588 vdev->host.bus, vdev->host.slot, vdev->host.function);
2590 snprintf(name, sizeof(name), "vfio[%04x:%02x:%02x.%x].rom",
2591 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2592 vdev->host.function);
2593 memory_region_init_ram(&vdev->pdev.rom, NULL, name, size);
2594 ptr = memory_region_get_ram_ptr(&vdev->pdev.rom);
2595 memset(ptr, 0xff, size);
2598 bytes = pread(vdev->fd, ptr + off, size, voff + off);
2600 break; /* expect that we could get back less than the ROM BAR */
2601 } else if (bytes > 0) {
2605 if (errno == EINTR || errno == EAGAIN) {
2608 error_report("vfio: Error reading device ROM: %m");
2609 memory_region_destroy(&vdev->pdev.rom);
2614 pci_register_bar(&vdev->pdev, PCI_ROM_SLOT, 0, &vdev->pdev.rom);
2615 vdev->pdev.has_rom = true;
2619 static int vfio_connect_container(VFIOGroup *group)
2621 VFIOContainer *container;
2624 if (group->container) {
2628 QLIST_FOREACH(container, &container_list, next) {
2629 if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
2630 group->container = container;
2631 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2636 fd = qemu_open("/dev/vfio/vfio", O_RDWR);
2638 error_report("vfio: failed to open /dev/vfio/vfio: %m");
2642 ret = ioctl(fd, VFIO_GET_API_VERSION);
2643 if (ret != VFIO_API_VERSION) {
2644 error_report("vfio: supported vfio version: %d, "
2645 "reported version: %d", VFIO_API_VERSION, ret);
2650 container = g_malloc0(sizeof(*container));
2653 if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) {
2654 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
2656 error_report("vfio: failed to set group container: %m");
2662 ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU);
2664 error_report("vfio: failed to set iommu for container: %m");
2670 container->iommu_data.listener = vfio_memory_listener;
2671 container->iommu_data.release = vfio_listener_release;
2673 memory_listener_register(&container->iommu_data.listener, &address_space_memory);
2675 error_report("vfio: No available IOMMU models");
2681 QLIST_INIT(&container->group_list);
2682 QLIST_INSERT_HEAD(&container_list, container, next);
2684 group->container = container;
2685 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2690 static void vfio_disconnect_container(VFIOGroup *group)
2692 VFIOContainer *container = group->container;
2694 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
2695 error_report("vfio: error disconnecting group %d from container",
2699 QLIST_REMOVE(group, container_next);
2700 group->container = NULL;
2702 if (QLIST_EMPTY(&container->group_list)) {
2703 if (container->iommu_data.release) {
2704 container->iommu_data.release(container);
2706 QLIST_REMOVE(container, next);
2707 DPRINTF("vfio_disconnect_container: close container->fd\n");
2708 close(container->fd);
2713 static VFIOGroup *vfio_get_group(int groupid)
2717 struct vfio_group_status status = { .argsz = sizeof(status) };
2719 QLIST_FOREACH(group, &group_list, next) {
2720 if (group->groupid == groupid) {
2725 group = g_malloc0(sizeof(*group));
2727 snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
2728 group->fd = qemu_open(path, O_RDWR);
2729 if (group->fd < 0) {
2730 error_report("vfio: error opening %s: %m", path);
2735 if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
2736 error_report("vfio: error getting group status: %m");
2742 if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
2743 error_report("vfio: error, group %d is not viable, please ensure "
2744 "all devices within the iommu_group are bound to their "
2745 "vfio bus driver.", groupid);
2751 group->groupid = groupid;
2752 QLIST_INIT(&group->device_list);
2754 if (vfio_connect_container(group)) {
2755 error_report("vfio: failed to setup container for group %d", groupid);
2761 QLIST_INSERT_HEAD(&group_list, group, next);
2766 static void vfio_put_group(VFIOGroup *group)
2768 if (!QLIST_EMPTY(&group->device_list)) {
2772 vfio_disconnect_container(group);
2773 QLIST_REMOVE(group, next);
2774 DPRINTF("vfio_put_group: close group->fd\n");
2779 static int vfio_get_device(VFIOGroup *group, const char *name, VFIODevice *vdev)
2781 struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
2782 struct vfio_region_info reg_info = { .argsz = sizeof(reg_info) };
2785 ret = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
2787 error_report("vfio: error getting device %s from group %d: %m",
2788 name, group->groupid);
2789 error_printf("Verify all devices in group %d are bound to vfio-pci "
2790 "or pci-stub and not already in use\n", group->groupid);
2795 vdev->group = group;
2796 QLIST_INSERT_HEAD(&group->device_list, vdev, next);
2798 /* Sanity check device */
2799 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_INFO, &dev_info);
2801 error_report("vfio: error getting device info: %m");
2805 DPRINTF("Device %s flags: %u, regions: %u, irgs: %u\n", name,
2806 dev_info.flags, dev_info.num_regions, dev_info.num_irqs);
2808 if (!(dev_info.flags & VFIO_DEVICE_FLAGS_PCI)) {
2809 error_report("vfio: Um, this isn't a PCI device");
2813 vdev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
2814 if (!vdev->reset_works) {
2815 error_report("Warning, device %s does not support reset", name);
2818 if (dev_info.num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
2819 error_report("vfio: unexpected number of io regions %u",
2820 dev_info.num_regions);
2824 if (dev_info.num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
2825 error_report("vfio: unexpected number of irqs %u", dev_info.num_irqs);
2829 for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
2832 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, ®_info);
2834 error_report("vfio: Error getting region %d info: %m", i);
2838 DPRINTF("Device %s region %d:\n", name, i);
2839 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
2840 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
2841 (unsigned long)reg_info.flags);
2843 vdev->bars[i].flags = reg_info.flags;
2844 vdev->bars[i].size = reg_info.size;
2845 vdev->bars[i].fd_offset = reg_info.offset;
2846 vdev->bars[i].fd = vdev->fd;
2847 vdev->bars[i].nr = i;
2848 QLIST_INIT(&vdev->bars[i].quirks);
2851 reg_info.index = VFIO_PCI_ROM_REGION_INDEX;
2853 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, ®_info);
2855 error_report("vfio: Error getting ROM info: %m");
2859 DPRINTF("Device %s ROM:\n", name);
2860 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
2861 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
2862 (unsigned long)reg_info.flags);
2864 vdev->rom_size = reg_info.size;
2865 vdev->rom_offset = reg_info.offset;
2867 reg_info.index = VFIO_PCI_CONFIG_REGION_INDEX;
2869 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, ®_info);
2871 error_report("vfio: Error getting config info: %m");
2875 DPRINTF("Device %s config:\n", name);
2876 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
2877 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
2878 (unsigned long)reg_info.flags);
2880 vdev->config_size = reg_info.size;
2881 if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
2882 vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
2884 vdev->config_offset = reg_info.offset;
2886 if ((vdev->features & VFIO_FEATURE_ENABLE_VGA) &&
2887 dev_info.num_regions > VFIO_PCI_VGA_REGION_INDEX) {
2888 struct vfio_region_info vga_info = {
2889 .argsz = sizeof(vga_info),
2890 .index = VFIO_PCI_VGA_REGION_INDEX,
2893 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &vga_info);
2896 "vfio: Device does not support requested feature x-vga");
2900 if (!(vga_info.flags & VFIO_REGION_INFO_FLAG_READ) ||
2901 !(vga_info.flags & VFIO_REGION_INFO_FLAG_WRITE) ||
2902 vga_info.size < 0xbffff + 1) {
2903 error_report("vfio: Unexpected VGA info, flags 0x%lx, size 0x%lx",
2904 (unsigned long)vga_info.flags,
2905 (unsigned long)vga_info.size);
2909 vdev->vga.fd_offset = vga_info.offset;
2910 vdev->vga.fd = vdev->fd;
2912 vdev->vga.region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE;
2913 vdev->vga.region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM;
2914 QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_MEM].quirks);
2916 vdev->vga.region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE;
2917 vdev->vga.region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO;
2918 QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].quirks);
2920 vdev->vga.region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE;
2921 vdev->vga.region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI;
2922 QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks);
2924 vdev->has_vga = true;
2929 QLIST_REMOVE(vdev, next);
2936 static void vfio_put_device(VFIODevice *vdev)
2938 QLIST_REMOVE(vdev, next);
2940 DPRINTF("vfio_put_device: close vdev->fd\n");
2948 static int vfio_initfn(PCIDevice *pdev)
2950 VFIODevice *pvdev, *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
2952 char path[PATH_MAX], iommu_group_path[PATH_MAX], *group_name;
2958 /* Check that the host device exists */
2959 snprintf(path, sizeof(path),
2960 "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/",
2961 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2962 vdev->host.function);
2963 if (stat(path, &st) < 0) {
2964 error_report("vfio: error: no such host device: %s", path);
2968 strncat(path, "iommu_group", sizeof(path) - strlen(path) - 1);
2970 len = readlink(path, iommu_group_path, PATH_MAX);
2972 error_report("vfio: error no iommu_group for device");
2976 iommu_group_path[len] = 0;
2977 group_name = basename(iommu_group_path);
2979 if (sscanf(group_name, "%d", &groupid) != 1) {
2980 error_report("vfio: error reading %s: %m", path);
2984 DPRINTF("%s(%04x:%02x:%02x.%x) group %d\n", __func__, vdev->host.domain,
2985 vdev->host.bus, vdev->host.slot, vdev->host.function, groupid);
2987 group = vfio_get_group(groupid);
2989 error_report("vfio: failed to get group %d", groupid);
2993 snprintf(path, sizeof(path), "%04x:%02x:%02x.%01x",
2994 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2995 vdev->host.function);
2997 QLIST_FOREACH(pvdev, &group->device_list, next) {
2998 if (pvdev->host.domain == vdev->host.domain &&
2999 pvdev->host.bus == vdev->host.bus &&
3000 pvdev->host.slot == vdev->host.slot &&
3001 pvdev->host.function == vdev->host.function) {
3003 error_report("vfio: error: device %s is already attached", path);
3004 vfio_put_group(group);
3009 ret = vfio_get_device(group, path, vdev);
3011 error_report("vfio: failed to get device %s", path);
3012 vfio_put_group(group);
3016 /* Get a copy of config space */
3017 ret = pread(vdev->fd, vdev->pdev.config,
3018 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
3019 vdev->config_offset);
3020 if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
3021 ret = ret < 0 ? -errno : -EFAULT;
3022 error_report("vfio: Failed to read device config space");
3026 /* vfio emulates a lot for us, but some bits need extra love */
3027 vdev->emulated_config_bits = g_malloc0(vdev->config_size);
3029 /* QEMU can choose to expose the ROM or not */
3030 memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4);
3032 /* QEMU can change multi-function devices to single function, or reverse */
3033 vdev->emulated_config_bits[PCI_HEADER_TYPE] =
3034 PCI_HEADER_TYPE_MULTI_FUNCTION;
3037 * Clear host resource mapping info. If we choose not to register a
3038 * BAR, such as might be the case with the option ROM, we can get
3039 * confusing, unwritable, residual addresses from the host here.
3041 memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
3042 memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
3044 vfio_load_rom(vdev);
3046 ret = vfio_early_setup_msix(vdev);
3051 vfio_map_bars(vdev);
3053 ret = vfio_add_capabilities(vdev);
3058 /* QEMU emulates all of MSI & MSIX */
3059 if (pdev->cap_present & QEMU_PCI_CAP_MSIX) {
3060 memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff,
3064 if (pdev->cap_present & QEMU_PCI_CAP_MSI) {
3065 memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff,
3066 vdev->msi_cap_size);
3069 if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
3070 vdev->intx.mmap_timer = qemu_new_timer_ms(vm_clock,
3071 vfio_intx_mmap_enable, vdev);
3072 pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_update_irq);
3073 ret = vfio_enable_intx(vdev);
3079 add_boot_device_path(vdev->bootindex, &pdev->qdev, NULL);
3084 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3085 vfio_teardown_msi(vdev);
3086 vfio_unmap_bars(vdev);
3088 g_free(vdev->emulated_config_bits);
3089 vfio_put_device(vdev);
3090 vfio_put_group(group);
3094 static void vfio_exitfn(PCIDevice *pdev)
3096 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
3097 VFIOGroup *group = vdev->group;
3099 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3100 vfio_disable_interrupts(vdev);
3101 if (vdev->intx.mmap_timer) {
3102 qemu_free_timer(vdev->intx.mmap_timer);
3104 vfio_teardown_msi(vdev);
3105 vfio_unmap_bars(vdev);
3106 g_free(vdev->emulated_config_bits);
3107 vfio_put_device(vdev);
3108 vfio_put_group(group);
3111 static void vfio_pci_reset(DeviceState *dev)
3113 PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
3114 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
3117 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
3118 vdev->host.bus, vdev->host.slot, vdev->host.function);
3120 vfio_disable_interrupts(vdev);
3122 /* Make sure the device is in D0 */
3127 pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
3128 state = pmcsr & PCI_PM_CTRL_STATE_MASK;
3130 pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
3131 vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2);
3132 /* vfio handles the necessary delay here */
3133 pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
3134 state = pmcsr & PCI_PM_CTRL_STATE_MASK;
3136 error_report("vfio: Unable to power on device, stuck in D%d\n",
3143 * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
3144 * Also put INTx Disable in known state.
3146 cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
3147 cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
3148 PCI_COMMAND_INTX_DISABLE);
3149 vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
3151 if (vdev->reset_works) {
3152 if (ioctl(vdev->fd, VFIO_DEVICE_RESET)) {
3153 error_report("vfio: Error unable to reset physical device "
3154 "(%04x:%02x:%02x.%x): %m", vdev->host.domain,
3155 vdev->host.bus, vdev->host.slot, vdev->host.function);
3159 vfio_enable_intx(vdev);
3162 static Property vfio_pci_dev_properties[] = {
3163 DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIODevice, host),
3164 DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIODevice,
3165 intx.mmap_timeout, 1100),
3166 DEFINE_PROP_BIT("x-vga", VFIODevice, features,
3167 VFIO_FEATURE_ENABLE_VGA_BIT, false),
3168 DEFINE_PROP_INT32("bootindex", VFIODevice, bootindex, -1),
3170 * TODO - support passed fds... is this necessary?
3171 * DEFINE_PROP_STRING("vfiofd", VFIODevice, vfiofd_name),
3172 * DEFINE_PROP_STRING("vfiogroupfd, VFIODevice, vfiogroupfd_name),
3174 DEFINE_PROP_END_OF_LIST(),
3177 static const VMStateDescription vfio_pci_vmstate = {
3182 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
3184 DeviceClass *dc = DEVICE_CLASS(klass);
3185 PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
3187 dc->reset = vfio_pci_reset;
3188 dc->props = vfio_pci_dev_properties;
3189 dc->vmsd = &vfio_pci_vmstate;
3190 dc->desc = "VFIO-based PCI device assignment";
3191 pdc->init = vfio_initfn;
3192 pdc->exit = vfio_exitfn;
3193 pdc->config_read = vfio_pci_read_config;
3194 pdc->config_write = vfio_pci_write_config;
3195 pdc->is_express = 1; /* We might be */
3198 static const TypeInfo vfio_pci_dev_info = {
3200 .parent = TYPE_PCI_DEVICE,
3201 .instance_size = sizeof(VFIODevice),
3202 .class_init = vfio_pci_dev_class_init,
3205 static void register_vfio_pci_dev_type(void)
3207 type_register_static(&vfio_pci_dev_info);
3210 type_init(register_vfio_pci_dev_type)