xref: /openbmc/qemu/hw/vfio/pci.c (revision 7678b74a)
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 "qemu/osdep.h"
22 #include <linux/vfio.h>
23 #include <sys/ioctl.h>
24 
25 #include "hw/hw.h"
26 #include "hw/pci/msi.h"
27 #include "hw/pci/msix.h"
28 #include "hw/pci/pci_bridge.h"
29 #include "hw/qdev-properties.h"
30 #include "migration/vmstate.h"
31 #include "qemu/error-report.h"
32 #include "qemu/main-loop.h"
33 #include "qemu/module.h"
34 #include "qemu/option.h"
35 #include "qemu/range.h"
36 #include "qemu/units.h"
37 #include "sysemu/kvm.h"
38 #include "sysemu/runstate.h"
39 #include "sysemu/sysemu.h"
40 #include "pci.h"
41 #include "trace.h"
42 #include "qapi/error.h"
43 
44 #define TYPE_VFIO_PCI "vfio-pci"
45 #define PCI_VFIO(obj)    OBJECT_CHECK(VFIOPCIDevice, obj, TYPE_VFIO_PCI)
46 
47 #define TYPE_VFIO_PCI_NOHOTPLUG "vfio-pci-nohotplug"
48 
49 static void vfio_disable_interrupts(VFIOPCIDevice *vdev);
50 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled);
51 
52 /*
53  * Disabling BAR mmaping can be slow, but toggling it around INTx can
54  * also be a huge overhead.  We try to get the best of both worlds by
55  * waiting until an interrupt to disable mmaps (subsequent transitions
56  * to the same state are effectively no overhead).  If the interrupt has
57  * been serviced and the time gap is long enough, we re-enable mmaps for
58  * performance.  This works well for things like graphics cards, which
59  * may not use their interrupt at all and are penalized to an unusable
60  * level by read/write BAR traps.  Other devices, like NICs, have more
61  * regular interrupts and see much better latency by staying in non-mmap
62  * mode.  We therefore set the default mmap_timeout such that a ping
63  * is just enough to keep the mmap disabled.  Users can experiment with
64  * other options with the x-intx-mmap-timeout-ms parameter (a value of
65  * zero disables the timer).
66  */
67 static void vfio_intx_mmap_enable(void *opaque)
68 {
69     VFIOPCIDevice *vdev = opaque;
70 
71     if (vdev->intx.pending) {
72         timer_mod(vdev->intx.mmap_timer,
73                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
74         return;
75     }
76 
77     vfio_mmap_set_enabled(vdev, true);
78 }
79 
80 static void vfio_intx_interrupt(void *opaque)
81 {
82     VFIOPCIDevice *vdev = opaque;
83 
84     if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
85         return;
86     }
87 
88     trace_vfio_intx_interrupt(vdev->vbasedev.name, 'A' + vdev->intx.pin);
89 
90     vdev->intx.pending = true;
91     pci_irq_assert(&vdev->pdev);
92     vfio_mmap_set_enabled(vdev, false);
93     if (vdev->intx.mmap_timeout) {
94         timer_mod(vdev->intx.mmap_timer,
95                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
96     }
97 }
98 
99 static void vfio_intx_eoi(VFIODevice *vbasedev)
100 {
101     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
102 
103     if (!vdev->intx.pending) {
104         return;
105     }
106 
107     trace_vfio_intx_eoi(vbasedev->name);
108 
109     vdev->intx.pending = false;
110     pci_irq_deassert(&vdev->pdev);
111     vfio_unmask_single_irqindex(vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
112 }
113 
114 static void vfio_intx_enable_kvm(VFIOPCIDevice *vdev, Error **errp)
115 {
116 #ifdef CONFIG_KVM
117     struct kvm_irqfd irqfd = {
118         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
119         .gsi = vdev->intx.route.irq,
120         .flags = KVM_IRQFD_FLAG_RESAMPLE,
121     };
122     Error *err = NULL;
123 
124     if (vdev->no_kvm_intx || !kvm_irqfds_enabled() ||
125         vdev->intx.route.mode != PCI_INTX_ENABLED ||
126         !kvm_resamplefds_enabled()) {
127         return;
128     }
129 
130     /* Get to a known interrupt state */
131     qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
132     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
133     vdev->intx.pending = false;
134     pci_irq_deassert(&vdev->pdev);
135 
136     /* Get an eventfd for resample/unmask */
137     if (event_notifier_init(&vdev->intx.unmask, 0)) {
138         error_setg(errp, "event_notifier_init failed eoi");
139         goto fail;
140     }
141 
142     /* KVM triggers it, VFIO listens for it */
143     irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
144 
145     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
146         error_setg_errno(errp, errno, "failed to setup resample irqfd");
147         goto fail_irqfd;
148     }
149 
150     if (vfio_set_irq_signaling(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX, 0,
151                                VFIO_IRQ_SET_ACTION_UNMASK,
152                                irqfd.resamplefd, &err)) {
153         error_propagate(errp, err);
154         goto fail_vfio;
155     }
156 
157     /* Let'em rip */
158     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
159 
160     vdev->intx.kvm_accel = true;
161 
162     trace_vfio_intx_enable_kvm(vdev->vbasedev.name);
163 
164     return;
165 
166 fail_vfio:
167     irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
168     kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
169 fail_irqfd:
170     event_notifier_cleanup(&vdev->intx.unmask);
171 fail:
172     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
173     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
174 #endif
175 }
176 
177 static void vfio_intx_disable_kvm(VFIOPCIDevice *vdev)
178 {
179 #ifdef CONFIG_KVM
180     struct kvm_irqfd irqfd = {
181         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
182         .gsi = vdev->intx.route.irq,
183         .flags = KVM_IRQFD_FLAG_DEASSIGN,
184     };
185 
186     if (!vdev->intx.kvm_accel) {
187         return;
188     }
189 
190     /*
191      * Get to a known state, hardware masked, QEMU ready to accept new
192      * interrupts, QEMU IRQ de-asserted.
193      */
194     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
195     vdev->intx.pending = false;
196     pci_irq_deassert(&vdev->pdev);
197 
198     /* Tell KVM to stop listening for an INTx irqfd */
199     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
200         error_report("vfio: Error: Failed to disable INTx irqfd: %m");
201     }
202 
203     /* We only need to close the eventfd for VFIO to cleanup the kernel side */
204     event_notifier_cleanup(&vdev->intx.unmask);
205 
206     /* QEMU starts listening for interrupt events. */
207     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
208 
209     vdev->intx.kvm_accel = false;
210 
211     /* If we've missed an event, let it re-fire through QEMU */
212     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
213 
214     trace_vfio_intx_disable_kvm(vdev->vbasedev.name);
215 #endif
216 }
217 
218 static void vfio_intx_update(PCIDevice *pdev)
219 {
220     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
221     PCIINTxRoute route;
222     Error *err = NULL;
223 
224     if (vdev->interrupt != VFIO_INT_INTx) {
225         return;
226     }
227 
228     route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
229 
230     if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
231         return; /* Nothing changed */
232     }
233 
234     trace_vfio_intx_update(vdev->vbasedev.name,
235                            vdev->intx.route.irq, route.irq);
236 
237     vfio_intx_disable_kvm(vdev);
238 
239     vdev->intx.route = route;
240 
241     if (route.mode != PCI_INTX_ENABLED) {
242         return;
243     }
244 
245     vfio_intx_enable_kvm(vdev, &err);
246     if (err) {
247         warn_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
248     }
249 
250     /* Re-enable the interrupt in cased we missed an EOI */
251     vfio_intx_eoi(&vdev->vbasedev);
252 }
253 
254 static int vfio_intx_enable(VFIOPCIDevice *vdev, Error **errp)
255 {
256     uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
257     Error *err = NULL;
258     int32_t fd;
259     int ret;
260 
261 
262     if (!pin) {
263         return 0;
264     }
265 
266     vfio_disable_interrupts(vdev);
267 
268     vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
269     pci_config_set_interrupt_pin(vdev->pdev.config, pin);
270 
271 #ifdef CONFIG_KVM
272     /*
273      * Only conditional to avoid generating error messages on platforms
274      * where we won't actually use the result anyway.
275      */
276     if (kvm_irqfds_enabled() && kvm_resamplefds_enabled()) {
277         vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
278                                                         vdev->intx.pin);
279     }
280 #endif
281 
282     ret = event_notifier_init(&vdev->intx.interrupt, 0);
283     if (ret) {
284         error_setg_errno(errp, -ret, "event_notifier_init failed");
285         return ret;
286     }
287     fd = event_notifier_get_fd(&vdev->intx.interrupt);
288     qemu_set_fd_handler(fd, vfio_intx_interrupt, NULL, vdev);
289 
290     if (vfio_set_irq_signaling(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX, 0,
291                                VFIO_IRQ_SET_ACTION_TRIGGER, fd, &err)) {
292         error_propagate(errp, err);
293         qemu_set_fd_handler(fd, NULL, NULL, vdev);
294         event_notifier_cleanup(&vdev->intx.interrupt);
295         return -errno;
296     }
297 
298     vfio_intx_enable_kvm(vdev, &err);
299     if (err) {
300         warn_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
301     }
302 
303     vdev->interrupt = VFIO_INT_INTx;
304 
305     trace_vfio_intx_enable(vdev->vbasedev.name);
306     return 0;
307 }
308 
309 static void vfio_intx_disable(VFIOPCIDevice *vdev)
310 {
311     int fd;
312 
313     timer_del(vdev->intx.mmap_timer);
314     vfio_intx_disable_kvm(vdev);
315     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
316     vdev->intx.pending = false;
317     pci_irq_deassert(&vdev->pdev);
318     vfio_mmap_set_enabled(vdev, true);
319 
320     fd = event_notifier_get_fd(&vdev->intx.interrupt);
321     qemu_set_fd_handler(fd, NULL, NULL, vdev);
322     event_notifier_cleanup(&vdev->intx.interrupt);
323 
324     vdev->interrupt = VFIO_INT_NONE;
325 
326     trace_vfio_intx_disable(vdev->vbasedev.name);
327 }
328 
329 /*
330  * MSI/X
331  */
332 static void vfio_msi_interrupt(void *opaque)
333 {
334     VFIOMSIVector *vector = opaque;
335     VFIOPCIDevice *vdev = vector->vdev;
336     MSIMessage (*get_msg)(PCIDevice *dev, unsigned vector);
337     void (*notify)(PCIDevice *dev, unsigned vector);
338     MSIMessage msg;
339     int nr = vector - vdev->msi_vectors;
340 
341     if (!event_notifier_test_and_clear(&vector->interrupt)) {
342         return;
343     }
344 
345     if (vdev->interrupt == VFIO_INT_MSIX) {
346         get_msg = msix_get_message;
347         notify = msix_notify;
348 
349         /* A masked vector firing needs to use the PBA, enable it */
350         if (msix_is_masked(&vdev->pdev, nr)) {
351             set_bit(nr, vdev->msix->pending);
352             memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, true);
353             trace_vfio_msix_pba_enable(vdev->vbasedev.name);
354         }
355     } else if (vdev->interrupt == VFIO_INT_MSI) {
356         get_msg = msi_get_message;
357         notify = msi_notify;
358     } else {
359         abort();
360     }
361 
362     msg = get_msg(&vdev->pdev, nr);
363     trace_vfio_msi_interrupt(vdev->vbasedev.name, nr, msg.address, msg.data);
364     notify(&vdev->pdev, nr);
365 }
366 
367 static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool msix)
368 {
369     struct vfio_irq_set *irq_set;
370     int ret = 0, i, argsz;
371     int32_t *fds;
372 
373     argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
374 
375     irq_set = g_malloc0(argsz);
376     irq_set->argsz = argsz;
377     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
378     irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
379     irq_set->start = 0;
380     irq_set->count = vdev->nr_vectors;
381     fds = (int32_t *)&irq_set->data;
382 
383     for (i = 0; i < vdev->nr_vectors; i++) {
384         int fd = -1;
385 
386         /*
387          * MSI vs MSI-X - The guest has direct access to MSI mask and pending
388          * bits, therefore we always use the KVM signaling path when setup.
389          * MSI-X mask and pending bits are emulated, so we want to use the
390          * KVM signaling path only when configured and unmasked.
391          */
392         if (vdev->msi_vectors[i].use) {
393             if (vdev->msi_vectors[i].virq < 0 ||
394                 (msix && msix_is_masked(&vdev->pdev, i))) {
395                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
396             } else {
397                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].kvm_interrupt);
398             }
399         }
400 
401         fds[i] = fd;
402     }
403 
404     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
405 
406     g_free(irq_set);
407 
408     return ret;
409 }
410 
411 static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
412                                   int vector_n, bool msix)
413 {
414     int virq;
415 
416     if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) {
417         return;
418     }
419 
420     if (event_notifier_init(&vector->kvm_interrupt, 0)) {
421         return;
422     }
423 
424     virq = kvm_irqchip_add_msi_route(kvm_state, vector_n, &vdev->pdev);
425     if (virq < 0) {
426         event_notifier_cleanup(&vector->kvm_interrupt);
427         return;
428     }
429 
430     if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
431                                        NULL, virq) < 0) {
432         kvm_irqchip_release_virq(kvm_state, virq);
433         event_notifier_cleanup(&vector->kvm_interrupt);
434         return;
435     }
436 
437     vector->virq = virq;
438 }
439 
440 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
441 {
442     kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
443                                           vector->virq);
444     kvm_irqchip_release_virq(kvm_state, vector->virq);
445     vector->virq = -1;
446     event_notifier_cleanup(&vector->kvm_interrupt);
447 }
448 
449 static void vfio_update_kvm_msi_virq(VFIOMSIVector *vector, MSIMessage msg,
450                                      PCIDevice *pdev)
451 {
452     kvm_irqchip_update_msi_route(kvm_state, vector->virq, msg, pdev);
453     kvm_irqchip_commit_routes(kvm_state);
454 }
455 
456 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
457                                    MSIMessage *msg, IOHandler *handler)
458 {
459     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
460     VFIOMSIVector *vector;
461     int ret;
462 
463     trace_vfio_msix_vector_do_use(vdev->vbasedev.name, nr);
464 
465     vector = &vdev->msi_vectors[nr];
466 
467     if (!vector->use) {
468         vector->vdev = vdev;
469         vector->virq = -1;
470         if (event_notifier_init(&vector->interrupt, 0)) {
471             error_report("vfio: Error: event_notifier_init failed");
472         }
473         vector->use = true;
474         msix_vector_use(pdev, nr);
475     }
476 
477     qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
478                         handler, NULL, vector);
479 
480     /*
481      * Attempt to enable route through KVM irqchip,
482      * default to userspace handling if unavailable.
483      */
484     if (vector->virq >= 0) {
485         if (!msg) {
486             vfio_remove_kvm_msi_virq(vector);
487         } else {
488             vfio_update_kvm_msi_virq(vector, *msg, pdev);
489         }
490     } else {
491         if (msg) {
492             vfio_add_kvm_msi_virq(vdev, vector, nr, true);
493         }
494     }
495 
496     /*
497      * We don't want to have the host allocate all possible MSI vectors
498      * for a device if they're not in use, so we shutdown and incrementally
499      * increase them as needed.
500      */
501     if (vdev->nr_vectors < nr + 1) {
502         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
503         vdev->nr_vectors = nr + 1;
504         ret = vfio_enable_vectors(vdev, true);
505         if (ret) {
506             error_report("vfio: failed to enable vectors, %d", ret);
507         }
508     } else {
509         Error *err = NULL;
510         int32_t fd;
511 
512         if (vector->virq >= 0) {
513             fd = event_notifier_get_fd(&vector->kvm_interrupt);
514         } else {
515             fd = event_notifier_get_fd(&vector->interrupt);
516         }
517 
518         if (vfio_set_irq_signaling(&vdev->vbasedev,
519                                      VFIO_PCI_MSIX_IRQ_INDEX, nr,
520                                      VFIO_IRQ_SET_ACTION_TRIGGER, fd, &err)) {
521             error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
522         }
523     }
524 
525     /* Disable PBA emulation when nothing more is pending. */
526     clear_bit(nr, vdev->msix->pending);
527     if (find_first_bit(vdev->msix->pending,
528                        vdev->nr_vectors) == vdev->nr_vectors) {
529         memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
530         trace_vfio_msix_pba_disable(vdev->vbasedev.name);
531     }
532 
533     return 0;
534 }
535 
536 static int vfio_msix_vector_use(PCIDevice *pdev,
537                                 unsigned int nr, MSIMessage msg)
538 {
539     return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
540 }
541 
542 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
543 {
544     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
545     VFIOMSIVector *vector = &vdev->msi_vectors[nr];
546 
547     trace_vfio_msix_vector_release(vdev->vbasedev.name, nr);
548 
549     /*
550      * There are still old guests that mask and unmask vectors on every
551      * interrupt.  If we're using QEMU bypass with a KVM irqfd, leave all of
552      * the KVM setup in place, simply switch VFIO to use the non-bypass
553      * eventfd.  We'll then fire the interrupt through QEMU and the MSI-X
554      * core will mask the interrupt and set pending bits, allowing it to
555      * be re-asserted on unmask.  Nothing to do if already using QEMU mode.
556      */
557     if (vector->virq >= 0) {
558         int32_t fd = event_notifier_get_fd(&vector->interrupt);
559         Error *err = NULL;
560 
561         if (vfio_set_irq_signaling(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX, nr,
562                                    VFIO_IRQ_SET_ACTION_TRIGGER, fd, &err)) {
563             error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
564         }
565     }
566 }
567 
568 static void vfio_msix_enable(VFIOPCIDevice *vdev)
569 {
570     vfio_disable_interrupts(vdev);
571 
572     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->msix->entries);
573 
574     vdev->interrupt = VFIO_INT_MSIX;
575 
576     /*
577      * Some communication channels between VF & PF or PF & fw rely on the
578      * physical state of the device and expect that enabling MSI-X from the
579      * guest enables the same on the host.  When our guest is Linux, the
580      * guest driver call to pci_enable_msix() sets the enabling bit in the
581      * MSI-X capability, but leaves the vector table masked.  We therefore
582      * can't rely on a vector_use callback (from request_irq() in the guest)
583      * to switch the physical device into MSI-X mode because that may come a
584      * long time after pci_enable_msix().  This code enables vector 0 with
585      * triggering to userspace, then immediately release the vector, leaving
586      * the physical device with no vectors enabled, but MSI-X enabled, just
587      * like the guest view.
588      */
589     vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
590     vfio_msix_vector_release(&vdev->pdev, 0);
591 
592     if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
593                                   vfio_msix_vector_release, NULL)) {
594         error_report("vfio: msix_set_vector_notifiers failed");
595     }
596 
597     trace_vfio_msix_enable(vdev->vbasedev.name);
598 }
599 
600 static void vfio_msi_enable(VFIOPCIDevice *vdev)
601 {
602     int ret, i;
603 
604     vfio_disable_interrupts(vdev);
605 
606     vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
607 retry:
608     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->nr_vectors);
609 
610     for (i = 0; i < vdev->nr_vectors; i++) {
611         VFIOMSIVector *vector = &vdev->msi_vectors[i];
612 
613         vector->vdev = vdev;
614         vector->virq = -1;
615         vector->use = true;
616 
617         if (event_notifier_init(&vector->interrupt, 0)) {
618             error_report("vfio: Error: event_notifier_init failed");
619         }
620 
621         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
622                             vfio_msi_interrupt, NULL, vector);
623 
624         /*
625          * Attempt to enable route through KVM irqchip,
626          * default to userspace handling if unavailable.
627          */
628         vfio_add_kvm_msi_virq(vdev, vector, i, false);
629     }
630 
631     /* Set interrupt type prior to possible interrupts */
632     vdev->interrupt = VFIO_INT_MSI;
633 
634     ret = vfio_enable_vectors(vdev, false);
635     if (ret) {
636         if (ret < 0) {
637             error_report("vfio: Error: Failed to setup MSI fds: %m");
638         } else if (ret != vdev->nr_vectors) {
639             error_report("vfio: Error: Failed to enable %d "
640                          "MSI vectors, retry with %d", vdev->nr_vectors, ret);
641         }
642 
643         for (i = 0; i < vdev->nr_vectors; i++) {
644             VFIOMSIVector *vector = &vdev->msi_vectors[i];
645             if (vector->virq >= 0) {
646                 vfio_remove_kvm_msi_virq(vector);
647             }
648             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
649                                 NULL, NULL, NULL);
650             event_notifier_cleanup(&vector->interrupt);
651         }
652 
653         g_free(vdev->msi_vectors);
654 
655         if (ret > 0 && ret != vdev->nr_vectors) {
656             vdev->nr_vectors = ret;
657             goto retry;
658         }
659         vdev->nr_vectors = 0;
660 
661         /*
662          * Failing to setup MSI doesn't really fall within any specification.
663          * Let's try leaving interrupts disabled and hope the guest figures
664          * out to fall back to INTx for this device.
665          */
666         error_report("vfio: Error: Failed to enable MSI");
667         vdev->interrupt = VFIO_INT_NONE;
668 
669         return;
670     }
671 
672     trace_vfio_msi_enable(vdev->vbasedev.name, vdev->nr_vectors);
673 }
674 
675 static void vfio_msi_disable_common(VFIOPCIDevice *vdev)
676 {
677     Error *err = NULL;
678     int i;
679 
680     for (i = 0; i < vdev->nr_vectors; i++) {
681         VFIOMSIVector *vector = &vdev->msi_vectors[i];
682         if (vdev->msi_vectors[i].use) {
683             if (vector->virq >= 0) {
684                 vfio_remove_kvm_msi_virq(vector);
685             }
686             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
687                                 NULL, NULL, NULL);
688             event_notifier_cleanup(&vector->interrupt);
689         }
690     }
691 
692     g_free(vdev->msi_vectors);
693     vdev->msi_vectors = NULL;
694     vdev->nr_vectors = 0;
695     vdev->interrupt = VFIO_INT_NONE;
696 
697     vfio_intx_enable(vdev, &err);
698     if (err) {
699         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
700     }
701 }
702 
703 static void vfio_msix_disable(VFIOPCIDevice *vdev)
704 {
705     int i;
706 
707     msix_unset_vector_notifiers(&vdev->pdev);
708 
709     /*
710      * MSI-X will only release vectors if MSI-X is still enabled on the
711      * device, check through the rest and release it ourselves if necessary.
712      */
713     for (i = 0; i < vdev->nr_vectors; i++) {
714         if (vdev->msi_vectors[i].use) {
715             vfio_msix_vector_release(&vdev->pdev, i);
716             msix_vector_unuse(&vdev->pdev, i);
717         }
718     }
719 
720     if (vdev->nr_vectors) {
721         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
722     }
723 
724     vfio_msi_disable_common(vdev);
725 
726     memset(vdev->msix->pending, 0,
727            BITS_TO_LONGS(vdev->msix->entries) * sizeof(unsigned long));
728 
729     trace_vfio_msix_disable(vdev->vbasedev.name);
730 }
731 
732 static void vfio_msi_disable(VFIOPCIDevice *vdev)
733 {
734     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSI_IRQ_INDEX);
735     vfio_msi_disable_common(vdev);
736 
737     trace_vfio_msi_disable(vdev->vbasedev.name);
738 }
739 
740 static void vfio_update_msi(VFIOPCIDevice *vdev)
741 {
742     int i;
743 
744     for (i = 0; i < vdev->nr_vectors; i++) {
745         VFIOMSIVector *vector = &vdev->msi_vectors[i];
746         MSIMessage msg;
747 
748         if (!vector->use || vector->virq < 0) {
749             continue;
750         }
751 
752         msg = msi_get_message(&vdev->pdev, i);
753         vfio_update_kvm_msi_virq(vector, msg, &vdev->pdev);
754     }
755 }
756 
757 static void vfio_pci_load_rom(VFIOPCIDevice *vdev)
758 {
759     struct vfio_region_info *reg_info;
760     uint64_t size;
761     off_t off = 0;
762     ssize_t bytes;
763 
764     if (vfio_get_region_info(&vdev->vbasedev,
765                              VFIO_PCI_ROM_REGION_INDEX, &reg_info)) {
766         error_report("vfio: Error getting ROM info: %m");
767         return;
768     }
769 
770     trace_vfio_pci_load_rom(vdev->vbasedev.name, (unsigned long)reg_info->size,
771                             (unsigned long)reg_info->offset,
772                             (unsigned long)reg_info->flags);
773 
774     vdev->rom_size = size = reg_info->size;
775     vdev->rom_offset = reg_info->offset;
776 
777     g_free(reg_info);
778 
779     if (!vdev->rom_size) {
780         vdev->rom_read_failed = true;
781         error_report("vfio-pci: Cannot read device rom at "
782                     "%s", vdev->vbasedev.name);
783         error_printf("Device option ROM contents are probably invalid "
784                     "(check dmesg).\nSkip option ROM probe with rombar=0, "
785                     "or load from file with romfile=\n");
786         return;
787     }
788 
789     vdev->rom = g_malloc(size);
790     memset(vdev->rom, 0xff, size);
791 
792     while (size) {
793         bytes = pread(vdev->vbasedev.fd, vdev->rom + off,
794                       size, vdev->rom_offset + off);
795         if (bytes == 0) {
796             break;
797         } else if (bytes > 0) {
798             off += bytes;
799             size -= bytes;
800         } else {
801             if (errno == EINTR || errno == EAGAIN) {
802                 continue;
803             }
804             error_report("vfio: Error reading device ROM: %m");
805             break;
806         }
807     }
808 
809     /*
810      * Test the ROM signature against our device, if the vendor is correct
811      * but the device ID doesn't match, store the correct device ID and
812      * recompute the checksum.  Intel IGD devices need this and are known
813      * to have bogus checksums so we can't simply adjust the checksum.
814      */
815     if (pci_get_word(vdev->rom) == 0xaa55 &&
816         pci_get_word(vdev->rom + 0x18) + 8 < vdev->rom_size &&
817         !memcmp(vdev->rom + pci_get_word(vdev->rom + 0x18), "PCIR", 4)) {
818         uint16_t vid, did;
819 
820         vid = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 4);
821         did = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6);
822 
823         if (vid == vdev->vendor_id && did != vdev->device_id) {
824             int i;
825             uint8_t csum, *data = vdev->rom;
826 
827             pci_set_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6,
828                          vdev->device_id);
829             data[6] = 0;
830 
831             for (csum = 0, i = 0; i < vdev->rom_size; i++) {
832                 csum += data[i];
833             }
834 
835             data[6] = -csum;
836         }
837     }
838 }
839 
840 static uint64_t vfio_rom_read(void *opaque, hwaddr addr, unsigned size)
841 {
842     VFIOPCIDevice *vdev = opaque;
843     union {
844         uint8_t byte;
845         uint16_t word;
846         uint32_t dword;
847         uint64_t qword;
848     } val;
849     uint64_t data = 0;
850 
851     /* Load the ROM lazily when the guest tries to read it */
852     if (unlikely(!vdev->rom && !vdev->rom_read_failed)) {
853         vfio_pci_load_rom(vdev);
854     }
855 
856     memcpy(&val, vdev->rom + addr,
857            (addr < vdev->rom_size) ? MIN(size, vdev->rom_size - addr) : 0);
858 
859     switch (size) {
860     case 1:
861         data = val.byte;
862         break;
863     case 2:
864         data = le16_to_cpu(val.word);
865         break;
866     case 4:
867         data = le32_to_cpu(val.dword);
868         break;
869     default:
870         hw_error("vfio: unsupported read size, %d bytes\n", size);
871         break;
872     }
873 
874     trace_vfio_rom_read(vdev->vbasedev.name, addr, size, data);
875 
876     return data;
877 }
878 
879 static void vfio_rom_write(void *opaque, hwaddr addr,
880                            uint64_t data, unsigned size)
881 {
882 }
883 
884 static const MemoryRegionOps vfio_rom_ops = {
885     .read = vfio_rom_read,
886     .write = vfio_rom_write,
887     .endianness = DEVICE_LITTLE_ENDIAN,
888 };
889 
890 static void vfio_pci_size_rom(VFIOPCIDevice *vdev)
891 {
892     uint32_t orig, size = cpu_to_le32((uint32_t)PCI_ROM_ADDRESS_MASK);
893     off_t offset = vdev->config_offset + PCI_ROM_ADDRESS;
894     DeviceState *dev = DEVICE(vdev);
895     char *name;
896     int fd = vdev->vbasedev.fd;
897 
898     if (vdev->pdev.romfile || !vdev->pdev.rom_bar) {
899         /* Since pci handles romfile, just print a message and return */
900         if (vfio_blacklist_opt_rom(vdev) && vdev->pdev.romfile) {
901             warn_report("Device at %s is known to cause system instability"
902                         " issues during option rom execution",
903                         vdev->vbasedev.name);
904             error_printf("Proceeding anyway since user specified romfile\n");
905         }
906         return;
907     }
908 
909     /*
910      * Use the same size ROM BAR as the physical device.  The contents
911      * will get filled in later when the guest tries to read it.
912      */
913     if (pread(fd, &orig, 4, offset) != 4 ||
914         pwrite(fd, &size, 4, offset) != 4 ||
915         pread(fd, &size, 4, offset) != 4 ||
916         pwrite(fd, &orig, 4, offset) != 4) {
917         error_report("%s(%s) failed: %m", __func__, vdev->vbasedev.name);
918         return;
919     }
920 
921     size = ~(le32_to_cpu(size) & PCI_ROM_ADDRESS_MASK) + 1;
922 
923     if (!size) {
924         return;
925     }
926 
927     if (vfio_blacklist_opt_rom(vdev)) {
928         if (dev->opts && qemu_opt_get(dev->opts, "rombar")) {
929             warn_report("Device at %s is known to cause system instability"
930                         " issues during option rom execution",
931                         vdev->vbasedev.name);
932             error_printf("Proceeding anyway since user specified"
933                          " non zero value for rombar\n");
934         } else {
935             warn_report("Rom loading for device at %s has been disabled"
936                         " due to system instability issues",
937                         vdev->vbasedev.name);
938             error_printf("Specify rombar=1 or romfile to force\n");
939             return;
940         }
941     }
942 
943     trace_vfio_pci_size_rom(vdev->vbasedev.name, size);
944 
945     name = g_strdup_printf("vfio[%s].rom", vdev->vbasedev.name);
946 
947     memory_region_init_io(&vdev->pdev.rom, OBJECT(vdev),
948                           &vfio_rom_ops, vdev, name, size);
949     g_free(name);
950 
951     pci_register_bar(&vdev->pdev, PCI_ROM_SLOT,
952                      PCI_BASE_ADDRESS_SPACE_MEMORY, &vdev->pdev.rom);
953 
954     vdev->rom_read_failed = false;
955 }
956 
957 void vfio_vga_write(void *opaque, hwaddr addr,
958                            uint64_t data, unsigned size)
959 {
960     VFIOVGARegion *region = opaque;
961     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
962     union {
963         uint8_t byte;
964         uint16_t word;
965         uint32_t dword;
966         uint64_t qword;
967     } buf;
968     off_t offset = vga->fd_offset + region->offset + addr;
969 
970     switch (size) {
971     case 1:
972         buf.byte = data;
973         break;
974     case 2:
975         buf.word = cpu_to_le16(data);
976         break;
977     case 4:
978         buf.dword = cpu_to_le32(data);
979         break;
980     default:
981         hw_error("vfio: unsupported write size, %d bytes", size);
982         break;
983     }
984 
985     if (pwrite(vga->fd, &buf, size, offset) != size) {
986         error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
987                      __func__, region->offset + addr, data, size);
988     }
989 
990     trace_vfio_vga_write(region->offset + addr, data, size);
991 }
992 
993 uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size)
994 {
995     VFIOVGARegion *region = opaque;
996     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
997     union {
998         uint8_t byte;
999         uint16_t word;
1000         uint32_t dword;
1001         uint64_t qword;
1002     } buf;
1003     uint64_t data = 0;
1004     off_t offset = vga->fd_offset + region->offset + addr;
1005 
1006     if (pread(vga->fd, &buf, size, offset) != size) {
1007         error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
1008                      __func__, region->offset + addr, size);
1009         return (uint64_t)-1;
1010     }
1011 
1012     switch (size) {
1013     case 1:
1014         data = buf.byte;
1015         break;
1016     case 2:
1017         data = le16_to_cpu(buf.word);
1018         break;
1019     case 4:
1020         data = le32_to_cpu(buf.dword);
1021         break;
1022     default:
1023         hw_error("vfio: unsupported read size, %d bytes", size);
1024         break;
1025     }
1026 
1027     trace_vfio_vga_read(region->offset + addr, size, data);
1028 
1029     return data;
1030 }
1031 
1032 static const MemoryRegionOps vfio_vga_ops = {
1033     .read = vfio_vga_read,
1034     .write = vfio_vga_write,
1035     .endianness = DEVICE_LITTLE_ENDIAN,
1036 };
1037 
1038 /*
1039  * Expand memory region of sub-page(size < PAGE_SIZE) MMIO BAR to page
1040  * size if the BAR is in an exclusive page in host so that we could map
1041  * this BAR to guest. But this sub-page BAR may not occupy an exclusive
1042  * page in guest. So we should set the priority of the expanded memory
1043  * region to zero in case of overlap with BARs which share the same page
1044  * with the sub-page BAR in guest. Besides, we should also recover the
1045  * size of this sub-page BAR when its base address is changed in guest
1046  * and not page aligned any more.
1047  */
1048 static void vfio_sub_page_bar_update_mapping(PCIDevice *pdev, int bar)
1049 {
1050     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
1051     VFIORegion *region = &vdev->bars[bar].region;
1052     MemoryRegion *mmap_mr, *region_mr, *base_mr;
1053     PCIIORegion *r;
1054     pcibus_t bar_addr;
1055     uint64_t size = region->size;
1056 
1057     /* Make sure that the whole region is allowed to be mmapped */
1058     if (region->nr_mmaps != 1 || !region->mmaps[0].mmap ||
1059         region->mmaps[0].size != region->size) {
1060         return;
1061     }
1062 
1063     r = &pdev->io_regions[bar];
1064     bar_addr = r->addr;
1065     base_mr = vdev->bars[bar].mr;
1066     region_mr = region->mem;
1067     mmap_mr = &region->mmaps[0].mem;
1068 
1069     /* If BAR is mapped and page aligned, update to fill PAGE_SIZE */
1070     if (bar_addr != PCI_BAR_UNMAPPED &&
1071         !(bar_addr & ~qemu_real_host_page_mask)) {
1072         size = qemu_real_host_page_size;
1073     }
1074 
1075     memory_region_transaction_begin();
1076 
1077     if (vdev->bars[bar].size < size) {
1078         memory_region_set_size(base_mr, size);
1079     }
1080     memory_region_set_size(region_mr, size);
1081     memory_region_set_size(mmap_mr, size);
1082     if (size != vdev->bars[bar].size && memory_region_is_mapped(base_mr)) {
1083         memory_region_del_subregion(r->address_space, base_mr);
1084         memory_region_add_subregion_overlap(r->address_space,
1085                                             bar_addr, base_mr, 0);
1086     }
1087 
1088     memory_region_transaction_commit();
1089 }
1090 
1091 /*
1092  * PCI config space
1093  */
1094 uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
1095 {
1096     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
1097     uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val;
1098 
1099     memcpy(&emu_bits, vdev->emulated_config_bits + addr, len);
1100     emu_bits = le32_to_cpu(emu_bits);
1101 
1102     if (emu_bits) {
1103         emu_val = pci_default_read_config(pdev, addr, len);
1104     }
1105 
1106     if (~emu_bits & (0xffffffffU >> (32 - len * 8))) {
1107         ssize_t ret;
1108 
1109         ret = pread(vdev->vbasedev.fd, &phys_val, len,
1110                     vdev->config_offset + addr);
1111         if (ret != len) {
1112             error_report("%s(%s, 0x%x, 0x%x) failed: %m",
1113                          __func__, vdev->vbasedev.name, addr, len);
1114             return -errno;
1115         }
1116         phys_val = le32_to_cpu(phys_val);
1117     }
1118 
1119     val = (emu_val & emu_bits) | (phys_val & ~emu_bits);
1120 
1121     trace_vfio_pci_read_config(vdev->vbasedev.name, addr, len, val);
1122 
1123     return val;
1124 }
1125 
1126 void vfio_pci_write_config(PCIDevice *pdev,
1127                            uint32_t addr, uint32_t val, int len)
1128 {
1129     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
1130     uint32_t val_le = cpu_to_le32(val);
1131 
1132     trace_vfio_pci_write_config(vdev->vbasedev.name, addr, val, len);
1133 
1134     /* Write everything to VFIO, let it filter out what we can't write */
1135     if (pwrite(vdev->vbasedev.fd, &val_le, len, vdev->config_offset + addr)
1136                 != len) {
1137         error_report("%s(%s, 0x%x, 0x%x, 0x%x) failed: %m",
1138                      __func__, vdev->vbasedev.name, addr, val, len);
1139     }
1140 
1141     /* MSI/MSI-X Enabling/Disabling */
1142     if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
1143         ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
1144         int is_enabled, was_enabled = msi_enabled(pdev);
1145 
1146         pci_default_write_config(pdev, addr, val, len);
1147 
1148         is_enabled = msi_enabled(pdev);
1149 
1150         if (!was_enabled) {
1151             if (is_enabled) {
1152                 vfio_msi_enable(vdev);
1153             }
1154         } else {
1155             if (!is_enabled) {
1156                 vfio_msi_disable(vdev);
1157             } else {
1158                 vfio_update_msi(vdev);
1159             }
1160         }
1161     } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
1162         ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
1163         int is_enabled, was_enabled = msix_enabled(pdev);
1164 
1165         pci_default_write_config(pdev, addr, val, len);
1166 
1167         is_enabled = msix_enabled(pdev);
1168 
1169         if (!was_enabled && is_enabled) {
1170             vfio_msix_enable(vdev);
1171         } else if (was_enabled && !is_enabled) {
1172             vfio_msix_disable(vdev);
1173         }
1174     } else if (ranges_overlap(addr, len, PCI_BASE_ADDRESS_0, 24) ||
1175         range_covers_byte(addr, len, PCI_COMMAND)) {
1176         pcibus_t old_addr[PCI_NUM_REGIONS - 1];
1177         int bar;
1178 
1179         for (bar = 0; bar < PCI_ROM_SLOT; bar++) {
1180             old_addr[bar] = pdev->io_regions[bar].addr;
1181         }
1182 
1183         pci_default_write_config(pdev, addr, val, len);
1184 
1185         for (bar = 0; bar < PCI_ROM_SLOT; bar++) {
1186             if (old_addr[bar] != pdev->io_regions[bar].addr &&
1187                 vdev->bars[bar].region.size > 0 &&
1188                 vdev->bars[bar].region.size < qemu_real_host_page_size) {
1189                 vfio_sub_page_bar_update_mapping(pdev, bar);
1190             }
1191         }
1192     } else {
1193         /* Write everything to QEMU to keep emulated bits correct */
1194         pci_default_write_config(pdev, addr, val, len);
1195     }
1196 }
1197 
1198 /*
1199  * Interrupt setup
1200  */
1201 static void vfio_disable_interrupts(VFIOPCIDevice *vdev)
1202 {
1203     /*
1204      * More complicated than it looks.  Disabling MSI/X transitions the
1205      * device to INTx mode (if supported).  Therefore we need to first
1206      * disable MSI/X and then cleanup by disabling INTx.
1207      */
1208     if (vdev->interrupt == VFIO_INT_MSIX) {
1209         vfio_msix_disable(vdev);
1210     } else if (vdev->interrupt == VFIO_INT_MSI) {
1211         vfio_msi_disable(vdev);
1212     }
1213 
1214     if (vdev->interrupt == VFIO_INT_INTx) {
1215         vfio_intx_disable(vdev);
1216     }
1217 }
1218 
1219 static int vfio_msi_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1220 {
1221     uint16_t ctrl;
1222     bool msi_64bit, msi_maskbit;
1223     int ret, entries;
1224     Error *err = NULL;
1225 
1226     if (pread(vdev->vbasedev.fd, &ctrl, sizeof(ctrl),
1227               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1228         error_setg_errno(errp, errno, "failed reading MSI PCI_CAP_FLAGS");
1229         return -errno;
1230     }
1231     ctrl = le16_to_cpu(ctrl);
1232 
1233     msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
1234     msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
1235     entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
1236 
1237     trace_vfio_msi_setup(vdev->vbasedev.name, pos);
1238 
1239     ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit, &err);
1240     if (ret < 0) {
1241         if (ret == -ENOTSUP) {
1242             return 0;
1243         }
1244         error_propagate_prepend(errp, err, "msi_init failed: ");
1245         return ret;
1246     }
1247     vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
1248 
1249     return 0;
1250 }
1251 
1252 static void vfio_pci_fixup_msix_region(VFIOPCIDevice *vdev)
1253 {
1254     off_t start, end;
1255     VFIORegion *region = &vdev->bars[vdev->msix->table_bar].region;
1256 
1257     /*
1258      * If the host driver allows mapping of a MSIX data, we are going to
1259      * do map the entire BAR and emulate MSIX table on top of that.
1260      */
1261     if (vfio_has_region_cap(&vdev->vbasedev, region->nr,
1262                             VFIO_REGION_INFO_CAP_MSIX_MAPPABLE)) {
1263         return;
1264     }
1265 
1266     /*
1267      * We expect to find a single mmap covering the whole BAR, anything else
1268      * means it's either unsupported or already setup.
1269      */
1270     if (region->nr_mmaps != 1 || region->mmaps[0].offset ||
1271         region->size != region->mmaps[0].size) {
1272         return;
1273     }
1274 
1275     /* MSI-X table start and end aligned to host page size */
1276     start = vdev->msix->table_offset & qemu_real_host_page_mask;
1277     end = REAL_HOST_PAGE_ALIGN((uint64_t)vdev->msix->table_offset +
1278                                (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
1279 
1280     /*
1281      * Does the MSI-X table cover the beginning of the BAR?  The whole BAR?
1282      * NB - Host page size is necessarily a power of two and so is the PCI
1283      * BAR (not counting EA yet), therefore if we have host page aligned
1284      * @start and @end, then any remainder of the BAR before or after those
1285      * must be at least host page sized and therefore mmap'able.
1286      */
1287     if (!start) {
1288         if (end >= region->size) {
1289             region->nr_mmaps = 0;
1290             g_free(region->mmaps);
1291             region->mmaps = NULL;
1292             trace_vfio_msix_fixup(vdev->vbasedev.name,
1293                                   vdev->msix->table_bar, 0, 0);
1294         } else {
1295             region->mmaps[0].offset = end;
1296             region->mmaps[0].size = region->size - end;
1297             trace_vfio_msix_fixup(vdev->vbasedev.name,
1298                               vdev->msix->table_bar, region->mmaps[0].offset,
1299                               region->mmaps[0].offset + region->mmaps[0].size);
1300         }
1301 
1302     /* Maybe it's aligned at the end of the BAR */
1303     } else if (end >= region->size) {
1304         region->mmaps[0].size = start;
1305         trace_vfio_msix_fixup(vdev->vbasedev.name,
1306                               vdev->msix->table_bar, region->mmaps[0].offset,
1307                               region->mmaps[0].offset + region->mmaps[0].size);
1308 
1309     /* Otherwise it must split the BAR */
1310     } else {
1311         region->nr_mmaps = 2;
1312         region->mmaps = g_renew(VFIOMmap, region->mmaps, 2);
1313 
1314         memcpy(&region->mmaps[1], &region->mmaps[0], sizeof(VFIOMmap));
1315 
1316         region->mmaps[0].size = start;
1317         trace_vfio_msix_fixup(vdev->vbasedev.name,
1318                               vdev->msix->table_bar, region->mmaps[0].offset,
1319                               region->mmaps[0].offset + region->mmaps[0].size);
1320 
1321         region->mmaps[1].offset = end;
1322         region->mmaps[1].size = region->size - end;
1323         trace_vfio_msix_fixup(vdev->vbasedev.name,
1324                               vdev->msix->table_bar, region->mmaps[1].offset,
1325                               region->mmaps[1].offset + region->mmaps[1].size);
1326     }
1327 }
1328 
1329 static void vfio_pci_relocate_msix(VFIOPCIDevice *vdev, Error **errp)
1330 {
1331     int target_bar = -1;
1332     size_t msix_sz;
1333 
1334     if (!vdev->msix || vdev->msix_relo == OFF_AUTOPCIBAR_OFF) {
1335         return;
1336     }
1337 
1338     /* The actual minimum size of MSI-X structures */
1339     msix_sz = (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE) +
1340               (QEMU_ALIGN_UP(vdev->msix->entries, 64) / 8);
1341     /* Round up to host pages, we don't want to share a page */
1342     msix_sz = REAL_HOST_PAGE_ALIGN(msix_sz);
1343     /* PCI BARs must be a power of 2 */
1344     msix_sz = pow2ceil(msix_sz);
1345 
1346     if (vdev->msix_relo == OFF_AUTOPCIBAR_AUTO) {
1347         /*
1348          * TODO: Lookup table for known devices.
1349          *
1350          * Logically we might use an algorithm here to select the BAR adding
1351          * the least additional MMIO space, but we cannot programatically
1352          * predict the driver dependency on BAR ordering or sizing, therefore
1353          * 'auto' becomes a lookup for combinations reported to work.
1354          */
1355         if (target_bar < 0) {
1356             error_setg(errp, "No automatic MSI-X relocation available for "
1357                        "device %04x:%04x", vdev->vendor_id, vdev->device_id);
1358             return;
1359         }
1360     } else {
1361         target_bar = (int)(vdev->msix_relo - OFF_AUTOPCIBAR_BAR0);
1362     }
1363 
1364     /* I/O port BARs cannot host MSI-X structures */
1365     if (vdev->bars[target_bar].ioport) {
1366         error_setg(errp, "Invalid MSI-X relocation BAR %d, "
1367                    "I/O port BAR", target_bar);
1368         return;
1369     }
1370 
1371     /* Cannot use a BAR in the "shadow" of a 64-bit BAR */
1372     if (!vdev->bars[target_bar].size &&
1373          target_bar > 0 && vdev->bars[target_bar - 1].mem64) {
1374         error_setg(errp, "Invalid MSI-X relocation BAR %d, "
1375                    "consumed by 64-bit BAR %d", target_bar, target_bar - 1);
1376         return;
1377     }
1378 
1379     /* 2GB max size for 32-bit BARs, cannot double if already > 1G */
1380     if (vdev->bars[target_bar].size > 1 * GiB &&
1381         !vdev->bars[target_bar].mem64) {
1382         error_setg(errp, "Invalid MSI-X relocation BAR %d, "
1383                    "no space to extend 32-bit BAR", target_bar);
1384         return;
1385     }
1386 
1387     /*
1388      * If adding a new BAR, test if we can make it 64bit.  We make it
1389      * prefetchable since QEMU MSI-X emulation has no read side effects
1390      * and doing so makes mapping more flexible.
1391      */
1392     if (!vdev->bars[target_bar].size) {
1393         if (target_bar < (PCI_ROM_SLOT - 1) &&
1394             !vdev->bars[target_bar + 1].size) {
1395             vdev->bars[target_bar].mem64 = true;
1396             vdev->bars[target_bar].type = PCI_BASE_ADDRESS_MEM_TYPE_64;
1397         }
1398         vdev->bars[target_bar].type |= PCI_BASE_ADDRESS_MEM_PREFETCH;
1399         vdev->bars[target_bar].size = msix_sz;
1400         vdev->msix->table_offset = 0;
1401     } else {
1402         vdev->bars[target_bar].size = MAX(vdev->bars[target_bar].size * 2,
1403                                           msix_sz * 2);
1404         /*
1405          * Due to above size calc, MSI-X always starts halfway into the BAR,
1406          * which will always be a separate host page.
1407          */
1408         vdev->msix->table_offset = vdev->bars[target_bar].size / 2;
1409     }
1410 
1411     vdev->msix->table_bar = target_bar;
1412     vdev->msix->pba_bar = target_bar;
1413     /* Requires 8-byte alignment, but PCI_MSIX_ENTRY_SIZE guarantees that */
1414     vdev->msix->pba_offset = vdev->msix->table_offset +
1415                                   (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE);
1416 
1417     trace_vfio_msix_relo(vdev->vbasedev.name,
1418                          vdev->msix->table_bar, vdev->msix->table_offset);
1419 }
1420 
1421 /*
1422  * We don't have any control over how pci_add_capability() inserts
1423  * capabilities into the chain.  In order to setup MSI-X we need a
1424  * MemoryRegion for the BAR.  In order to setup the BAR and not
1425  * attempt to mmap the MSI-X table area, which VFIO won't allow, we
1426  * need to first look for where the MSI-X table lives.  So we
1427  * unfortunately split MSI-X setup across two functions.
1428  */
1429 static void vfio_msix_early_setup(VFIOPCIDevice *vdev, Error **errp)
1430 {
1431     uint8_t pos;
1432     uint16_t ctrl;
1433     uint32_t table, pba;
1434     int fd = vdev->vbasedev.fd;
1435     VFIOMSIXInfo *msix;
1436 
1437     pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
1438     if (!pos) {
1439         return;
1440     }
1441 
1442     if (pread(fd, &ctrl, sizeof(ctrl),
1443               vdev->config_offset + pos + PCI_MSIX_FLAGS) != sizeof(ctrl)) {
1444         error_setg_errno(errp, errno, "failed to read PCI MSIX FLAGS");
1445         return;
1446     }
1447 
1448     if (pread(fd, &table, sizeof(table),
1449               vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
1450         error_setg_errno(errp, errno, "failed to read PCI MSIX TABLE");
1451         return;
1452     }
1453 
1454     if (pread(fd, &pba, sizeof(pba),
1455               vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
1456         error_setg_errno(errp, errno, "failed to read PCI MSIX PBA");
1457         return;
1458     }
1459 
1460     ctrl = le16_to_cpu(ctrl);
1461     table = le32_to_cpu(table);
1462     pba = le32_to_cpu(pba);
1463 
1464     msix = g_malloc0(sizeof(*msix));
1465     msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
1466     msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
1467     msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
1468     msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
1469     msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
1470 
1471     /*
1472      * Test the size of the pba_offset variable and catch if it extends outside
1473      * of the specified BAR. If it is the case, we need to apply a hardware
1474      * specific quirk if the device is known or we have a broken configuration.
1475      */
1476     if (msix->pba_offset >= vdev->bars[msix->pba_bar].region.size) {
1477         /*
1478          * Chelsio T5 Virtual Function devices are encoded as 0x58xx for T5
1479          * adapters. The T5 hardware returns an incorrect value of 0x8000 for
1480          * the VF PBA offset while the BAR itself is only 8k. The correct value
1481          * is 0x1000, so we hard code that here.
1482          */
1483         if (vdev->vendor_id == PCI_VENDOR_ID_CHELSIO &&
1484             (vdev->device_id & 0xff00) == 0x5800) {
1485             msix->pba_offset = 0x1000;
1486         } else if (vdev->msix_relo == OFF_AUTOPCIBAR_OFF) {
1487             error_setg(errp, "hardware reports invalid configuration, "
1488                        "MSIX PBA outside of specified BAR");
1489             g_free(msix);
1490             return;
1491         }
1492     }
1493 
1494     trace_vfio_msix_early_setup(vdev->vbasedev.name, pos, msix->table_bar,
1495                                 msix->table_offset, msix->entries);
1496     vdev->msix = msix;
1497 
1498     vfio_pci_fixup_msix_region(vdev);
1499 
1500     vfio_pci_relocate_msix(vdev, errp);
1501 }
1502 
1503 static int vfio_msix_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1504 {
1505     int ret;
1506     Error *err = NULL;
1507 
1508     vdev->msix->pending = g_malloc0(BITS_TO_LONGS(vdev->msix->entries) *
1509                                     sizeof(unsigned long));
1510     ret = msix_init(&vdev->pdev, vdev->msix->entries,
1511                     vdev->bars[vdev->msix->table_bar].mr,
1512                     vdev->msix->table_bar, vdev->msix->table_offset,
1513                     vdev->bars[vdev->msix->pba_bar].mr,
1514                     vdev->msix->pba_bar, vdev->msix->pba_offset, pos,
1515                     &err);
1516     if (ret < 0) {
1517         if (ret == -ENOTSUP) {
1518             warn_report_err(err);
1519             return 0;
1520         }
1521 
1522         error_propagate(errp, err);
1523         return ret;
1524     }
1525 
1526     /*
1527      * The PCI spec suggests that devices provide additional alignment for
1528      * MSI-X structures and avoid overlapping non-MSI-X related registers.
1529      * For an assigned device, this hopefully means that emulation of MSI-X
1530      * structures does not affect the performance of the device.  If devices
1531      * fail to provide that alignment, a significant performance penalty may
1532      * result, for instance Mellanox MT27500 VFs:
1533      * http://www.spinics.net/lists/kvm/msg125881.html
1534      *
1535      * The PBA is simply not that important for such a serious regression and
1536      * most drivers do not appear to look at it.  The solution for this is to
1537      * disable the PBA MemoryRegion unless it's being used.  We disable it
1538      * here and only enable it if a masked vector fires through QEMU.  As the
1539      * vector-use notifier is called, which occurs on unmask, we test whether
1540      * PBA emulation is needed and again disable if not.
1541      */
1542     memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
1543 
1544     /*
1545      * The emulated machine may provide a paravirt interface for MSIX setup
1546      * so it is not strictly necessary to emulate MSIX here. This becomes
1547      * helpful when frequently accessed MMIO registers are located in
1548      * subpages adjacent to the MSIX table but the MSIX data containing page
1549      * cannot be mapped because of a host page size bigger than the MSIX table
1550      * alignment.
1551      */
1552     if (object_property_get_bool(OBJECT(qdev_get_machine()),
1553                                  "vfio-no-msix-emulation", NULL)) {
1554         memory_region_set_enabled(&vdev->pdev.msix_table_mmio, false);
1555     }
1556 
1557     return 0;
1558 }
1559 
1560 static void vfio_teardown_msi(VFIOPCIDevice *vdev)
1561 {
1562     msi_uninit(&vdev->pdev);
1563 
1564     if (vdev->msix) {
1565         msix_uninit(&vdev->pdev,
1566                     vdev->bars[vdev->msix->table_bar].mr,
1567                     vdev->bars[vdev->msix->pba_bar].mr);
1568         g_free(vdev->msix->pending);
1569     }
1570 }
1571 
1572 /*
1573  * Resource setup
1574  */
1575 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled)
1576 {
1577     int i;
1578 
1579     for (i = 0; i < PCI_ROM_SLOT; i++) {
1580         vfio_region_mmaps_set_enabled(&vdev->bars[i].region, enabled);
1581     }
1582 }
1583 
1584 static void vfio_bar_prepare(VFIOPCIDevice *vdev, int nr)
1585 {
1586     VFIOBAR *bar = &vdev->bars[nr];
1587 
1588     uint32_t pci_bar;
1589     int ret;
1590 
1591     /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1592     if (!bar->region.size) {
1593         return;
1594     }
1595 
1596     /* Determine what type of BAR this is for registration */
1597     ret = pread(vdev->vbasedev.fd, &pci_bar, sizeof(pci_bar),
1598                 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
1599     if (ret != sizeof(pci_bar)) {
1600         error_report("vfio: Failed to read BAR %d (%m)", nr);
1601         return;
1602     }
1603 
1604     pci_bar = le32_to_cpu(pci_bar);
1605     bar->ioport = (pci_bar & PCI_BASE_ADDRESS_SPACE_IO);
1606     bar->mem64 = bar->ioport ? 0 : (pci_bar & PCI_BASE_ADDRESS_MEM_TYPE_64);
1607     bar->type = pci_bar & (bar->ioport ? ~PCI_BASE_ADDRESS_IO_MASK :
1608                                          ~PCI_BASE_ADDRESS_MEM_MASK);
1609     bar->size = bar->region.size;
1610 }
1611 
1612 static void vfio_bars_prepare(VFIOPCIDevice *vdev)
1613 {
1614     int i;
1615 
1616     for (i = 0; i < PCI_ROM_SLOT; i++) {
1617         vfio_bar_prepare(vdev, i);
1618     }
1619 }
1620 
1621 static void vfio_bar_register(VFIOPCIDevice *vdev, int nr)
1622 {
1623     VFIOBAR *bar = &vdev->bars[nr];
1624     char *name;
1625 
1626     if (!bar->size) {
1627         return;
1628     }
1629 
1630     bar->mr = g_new0(MemoryRegion, 1);
1631     name = g_strdup_printf("%s base BAR %d", vdev->vbasedev.name, nr);
1632     memory_region_init_io(bar->mr, OBJECT(vdev), NULL, NULL, name, bar->size);
1633     g_free(name);
1634 
1635     if (bar->region.size) {
1636         memory_region_add_subregion(bar->mr, 0, bar->region.mem);
1637 
1638         if (vfio_region_mmap(&bar->region)) {
1639             error_report("Failed to mmap %s BAR %d. Performance may be slow",
1640                          vdev->vbasedev.name, nr);
1641         }
1642     }
1643 
1644     pci_register_bar(&vdev->pdev, nr, bar->type, bar->mr);
1645 }
1646 
1647 static void vfio_bars_register(VFIOPCIDevice *vdev)
1648 {
1649     int i;
1650 
1651     for (i = 0; i < PCI_ROM_SLOT; i++) {
1652         vfio_bar_register(vdev, i);
1653     }
1654 }
1655 
1656 static void vfio_bars_exit(VFIOPCIDevice *vdev)
1657 {
1658     int i;
1659 
1660     for (i = 0; i < PCI_ROM_SLOT; i++) {
1661         VFIOBAR *bar = &vdev->bars[i];
1662 
1663         vfio_bar_quirk_exit(vdev, i);
1664         vfio_region_exit(&bar->region);
1665         if (bar->region.size) {
1666             memory_region_del_subregion(bar->mr, bar->region.mem);
1667         }
1668     }
1669 
1670     if (vdev->vga) {
1671         pci_unregister_vga(&vdev->pdev);
1672         vfio_vga_quirk_exit(vdev);
1673     }
1674 }
1675 
1676 static void vfio_bars_finalize(VFIOPCIDevice *vdev)
1677 {
1678     int i;
1679 
1680     for (i = 0; i < PCI_ROM_SLOT; i++) {
1681         VFIOBAR *bar = &vdev->bars[i];
1682 
1683         vfio_bar_quirk_finalize(vdev, i);
1684         vfio_region_finalize(&bar->region);
1685         if (bar->size) {
1686             object_unparent(OBJECT(bar->mr));
1687             g_free(bar->mr);
1688         }
1689     }
1690 
1691     if (vdev->vga) {
1692         vfio_vga_quirk_finalize(vdev);
1693         for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) {
1694             object_unparent(OBJECT(&vdev->vga->region[i].mem));
1695         }
1696         g_free(vdev->vga);
1697     }
1698 }
1699 
1700 /*
1701  * General setup
1702  */
1703 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
1704 {
1705     uint8_t tmp;
1706     uint16_t next = PCI_CONFIG_SPACE_SIZE;
1707 
1708     for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
1709          tmp = pdev->config[tmp + PCI_CAP_LIST_NEXT]) {
1710         if (tmp > pos && tmp < next) {
1711             next = tmp;
1712         }
1713     }
1714 
1715     return next - pos;
1716 }
1717 
1718 
1719 static uint16_t vfio_ext_cap_max_size(const uint8_t *config, uint16_t pos)
1720 {
1721     uint16_t tmp, next = PCIE_CONFIG_SPACE_SIZE;
1722 
1723     for (tmp = PCI_CONFIG_SPACE_SIZE; tmp;
1724         tmp = PCI_EXT_CAP_NEXT(pci_get_long(config + tmp))) {
1725         if (tmp > pos && tmp < next) {
1726             next = tmp;
1727         }
1728     }
1729 
1730     return next - pos;
1731 }
1732 
1733 static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask)
1734 {
1735     pci_set_word(buf, (pci_get_word(buf) & ~mask) | val);
1736 }
1737 
1738 static void vfio_add_emulated_word(VFIOPCIDevice *vdev, int pos,
1739                                    uint16_t val, uint16_t mask)
1740 {
1741     vfio_set_word_bits(vdev->pdev.config + pos, val, mask);
1742     vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask);
1743     vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask);
1744 }
1745 
1746 static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask)
1747 {
1748     pci_set_long(buf, (pci_get_long(buf) & ~mask) | val);
1749 }
1750 
1751 static void vfio_add_emulated_long(VFIOPCIDevice *vdev, int pos,
1752                                    uint32_t val, uint32_t mask)
1753 {
1754     vfio_set_long_bits(vdev->pdev.config + pos, val, mask);
1755     vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask);
1756     vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask);
1757 }
1758 
1759 static int vfio_setup_pcie_cap(VFIOPCIDevice *vdev, int pos, uint8_t size,
1760                                Error **errp)
1761 {
1762     uint16_t flags;
1763     uint8_t type;
1764 
1765     flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS);
1766     type = (flags & PCI_EXP_FLAGS_TYPE) >> 4;
1767 
1768     if (type != PCI_EXP_TYPE_ENDPOINT &&
1769         type != PCI_EXP_TYPE_LEG_END &&
1770         type != PCI_EXP_TYPE_RC_END) {
1771 
1772         error_setg(errp, "assignment of PCIe type 0x%x "
1773                    "devices is not currently supported", type);
1774         return -EINVAL;
1775     }
1776 
1777     if (!pci_bus_is_express(pci_get_bus(&vdev->pdev))) {
1778         PCIBus *bus = pci_get_bus(&vdev->pdev);
1779         PCIDevice *bridge;
1780 
1781         /*
1782          * Traditionally PCI device assignment exposes the PCIe capability
1783          * as-is on non-express buses.  The reason being that some drivers
1784          * simply assume that it's there, for example tg3.  However when
1785          * we're running on a native PCIe machine type, like Q35, we need
1786          * to hide the PCIe capability.  The reason for this is twofold;
1787          * first Windows guests get a Code 10 error when the PCIe capability
1788          * is exposed in this configuration.  Therefore express devices won't
1789          * work at all unless they're attached to express buses in the VM.
1790          * Second, a native PCIe machine introduces the possibility of fine
1791          * granularity IOMMUs supporting both translation and isolation.
1792          * Guest code to discover the IOMMU visibility of a device, such as
1793          * IOMMU grouping code on Linux, is very aware of device types and
1794          * valid transitions between bus types.  An express device on a non-
1795          * express bus is not a valid combination on bare metal systems.
1796          *
1797          * Drivers that require a PCIe capability to make the device
1798          * functional are simply going to need to have their devices placed
1799          * on a PCIe bus in the VM.
1800          */
1801         while (!pci_bus_is_root(bus)) {
1802             bridge = pci_bridge_get_device(bus);
1803             bus = pci_get_bus(bridge);
1804         }
1805 
1806         if (pci_bus_is_express(bus)) {
1807             return 0;
1808         }
1809 
1810     } else if (pci_bus_is_root(pci_get_bus(&vdev->pdev))) {
1811         /*
1812          * On a Root Complex bus Endpoints become Root Complex Integrated
1813          * Endpoints, which changes the type and clears the LNK & LNK2 fields.
1814          */
1815         if (type == PCI_EXP_TYPE_ENDPOINT) {
1816             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1817                                    PCI_EXP_TYPE_RC_END << 4,
1818                                    PCI_EXP_FLAGS_TYPE);
1819 
1820             /* Link Capabilities, Status, and Control goes away */
1821             if (size > PCI_EXP_LNKCTL) {
1822                 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0);
1823                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1824                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0);
1825 
1826 #ifndef PCI_EXP_LNKCAP2
1827 #define PCI_EXP_LNKCAP2 44
1828 #endif
1829 #ifndef PCI_EXP_LNKSTA2
1830 #define PCI_EXP_LNKSTA2 50
1831 #endif
1832                 /* Link 2 Capabilities, Status, and Control goes away */
1833                 if (size > PCI_EXP_LNKCAP2) {
1834                     vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0);
1835                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0);
1836                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0);
1837                 }
1838             }
1839 
1840         } else if (type == PCI_EXP_TYPE_LEG_END) {
1841             /*
1842              * Legacy endpoints don't belong on the root complex.  Windows
1843              * seems to be happier with devices if we skip the capability.
1844              */
1845             return 0;
1846         }
1847 
1848     } else {
1849         /*
1850          * Convert Root Complex Integrated Endpoints to regular endpoints.
1851          * These devices don't support LNK/LNK2 capabilities, so make them up.
1852          */
1853         if (type == PCI_EXP_TYPE_RC_END) {
1854             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1855                                    PCI_EXP_TYPE_ENDPOINT << 4,
1856                                    PCI_EXP_FLAGS_TYPE);
1857             vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP,
1858                            QEMU_PCI_EXP_LNKCAP_MLW(QEMU_PCI_EXP_LNK_X1) |
1859                            QEMU_PCI_EXP_LNKCAP_MLS(QEMU_PCI_EXP_LNK_2_5GT), ~0);
1860             vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1861         }
1862     }
1863 
1864     /*
1865      * Intel 82599 SR-IOV VFs report an invalid PCIe capability version 0
1866      * (Niantic errate #35) causing Windows to error with a Code 10 for the
1867      * device on Q35.  Fixup any such devices to report version 1.  If we
1868      * were to remove the capability entirely the guest would lose extended
1869      * config space.
1870      */
1871     if ((flags & PCI_EXP_FLAGS_VERS) == 0) {
1872         vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1873                                1, PCI_EXP_FLAGS_VERS);
1874     }
1875 
1876     pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size,
1877                              errp);
1878     if (pos < 0) {
1879         return pos;
1880     }
1881 
1882     vdev->pdev.exp.exp_cap = pos;
1883 
1884     return pos;
1885 }
1886 
1887 static void vfio_check_pcie_flr(VFIOPCIDevice *vdev, uint8_t pos)
1888 {
1889     uint32_t cap = pci_get_long(vdev->pdev.config + pos + PCI_EXP_DEVCAP);
1890 
1891     if (cap & PCI_EXP_DEVCAP_FLR) {
1892         trace_vfio_check_pcie_flr(vdev->vbasedev.name);
1893         vdev->has_flr = true;
1894     }
1895 }
1896 
1897 static void vfio_check_pm_reset(VFIOPCIDevice *vdev, uint8_t pos)
1898 {
1899     uint16_t csr = pci_get_word(vdev->pdev.config + pos + PCI_PM_CTRL);
1900 
1901     if (!(csr & PCI_PM_CTRL_NO_SOFT_RESET)) {
1902         trace_vfio_check_pm_reset(vdev->vbasedev.name);
1903         vdev->has_pm_reset = true;
1904     }
1905 }
1906 
1907 static void vfio_check_af_flr(VFIOPCIDevice *vdev, uint8_t pos)
1908 {
1909     uint8_t cap = pci_get_byte(vdev->pdev.config + pos + PCI_AF_CAP);
1910 
1911     if ((cap & PCI_AF_CAP_TP) && (cap & PCI_AF_CAP_FLR)) {
1912         trace_vfio_check_af_flr(vdev->vbasedev.name);
1913         vdev->has_flr = true;
1914     }
1915 }
1916 
1917 static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos, Error **errp)
1918 {
1919     PCIDevice *pdev = &vdev->pdev;
1920     uint8_t cap_id, next, size;
1921     int ret;
1922 
1923     cap_id = pdev->config[pos];
1924     next = pdev->config[pos + PCI_CAP_LIST_NEXT];
1925 
1926     /*
1927      * If it becomes important to configure capabilities to their actual
1928      * size, use this as the default when it's something we don't recognize.
1929      * Since QEMU doesn't actually handle many of the config accesses,
1930      * exact size doesn't seem worthwhile.
1931      */
1932     size = vfio_std_cap_max_size(pdev, pos);
1933 
1934     /*
1935      * pci_add_capability always inserts the new capability at the head
1936      * of the chain.  Therefore to end up with a chain that matches the
1937      * physical device, we insert from the end by making this recursive.
1938      * This is also why we pre-calculate size above as cached config space
1939      * will be changed as we unwind the stack.
1940      */
1941     if (next) {
1942         ret = vfio_add_std_cap(vdev, next, errp);
1943         if (ret) {
1944             return ret;
1945         }
1946     } else {
1947         /* Begin the rebuild, use QEMU emulated list bits */
1948         pdev->config[PCI_CAPABILITY_LIST] = 0;
1949         vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
1950         vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
1951 
1952         ret = vfio_add_virt_caps(vdev, errp);
1953         if (ret) {
1954             return ret;
1955         }
1956     }
1957 
1958     /* Scale down size, esp in case virt caps were added above */
1959     size = MIN(size, vfio_std_cap_max_size(pdev, pos));
1960 
1961     /* Use emulated next pointer to allow dropping caps */
1962     pci_set_byte(vdev->emulated_config_bits + pos + PCI_CAP_LIST_NEXT, 0xff);
1963 
1964     switch (cap_id) {
1965     case PCI_CAP_ID_MSI:
1966         ret = vfio_msi_setup(vdev, pos, errp);
1967         break;
1968     case PCI_CAP_ID_EXP:
1969         vfio_check_pcie_flr(vdev, pos);
1970         ret = vfio_setup_pcie_cap(vdev, pos, size, errp);
1971         break;
1972     case PCI_CAP_ID_MSIX:
1973         ret = vfio_msix_setup(vdev, pos, errp);
1974         break;
1975     case PCI_CAP_ID_PM:
1976         vfio_check_pm_reset(vdev, pos);
1977         vdev->pm_cap = pos;
1978         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
1979         break;
1980     case PCI_CAP_ID_AF:
1981         vfio_check_af_flr(vdev, pos);
1982         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
1983         break;
1984     default:
1985         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
1986         break;
1987     }
1988 
1989     if (ret < 0) {
1990         error_prepend(errp,
1991                       "failed to add PCI capability 0x%x[0x%x]@0x%x: ",
1992                       cap_id, size, pos);
1993         return ret;
1994     }
1995 
1996     return 0;
1997 }
1998 
1999 static void vfio_add_ext_cap(VFIOPCIDevice *vdev)
2000 {
2001     PCIDevice *pdev = &vdev->pdev;
2002     uint32_t header;
2003     uint16_t cap_id, next, size;
2004     uint8_t cap_ver;
2005     uint8_t *config;
2006 
2007     /* Only add extended caps if we have them and the guest can see them */
2008     if (!pci_is_express(pdev) || !pci_bus_is_express(pci_get_bus(pdev)) ||
2009         !pci_get_long(pdev->config + PCI_CONFIG_SPACE_SIZE)) {
2010         return;
2011     }
2012 
2013     /*
2014      * pcie_add_capability always inserts the new capability at the tail
2015      * of the chain.  Therefore to end up with a chain that matches the
2016      * physical device, we cache the config space to avoid overwriting
2017      * the original config space when we parse the extended capabilities.
2018      */
2019     config = g_memdup(pdev->config, vdev->config_size);
2020 
2021     /*
2022      * Extended capabilities are chained with each pointing to the next, so we
2023      * can drop anything other than the head of the chain simply by modifying
2024      * the previous next pointer.  Seed the head of the chain here such that
2025      * we can simply skip any capabilities we want to drop below, regardless
2026      * of their position in the chain.  If this stub capability still exists
2027      * after we add the capabilities we want to expose, update the capability
2028      * ID to zero.  Note that we cannot seed with the capability header being
2029      * zero as this conflicts with definition of an absent capability chain
2030      * and prevents capabilities beyond the head of the list from being added.
2031      * By replacing the dummy capability ID with zero after walking the device
2032      * chain, we also transparently mark extended capabilities as absent if
2033      * no capabilities were added.  Note that the PCIe spec defines an absence
2034      * of extended capabilities to be determined by a value of zero for the
2035      * capability ID, version, AND next pointer.  A non-zero next pointer
2036      * should be sufficient to indicate additional capabilities are present,
2037      * which will occur if we call pcie_add_capability() below.  The entire
2038      * first dword is emulated to support this.
2039      *
2040      * NB. The kernel side does similar masking, so be prepared that our
2041      * view of the device may also contain a capability ID zero in the head
2042      * of the chain.  Skip it for the same reason that we cannot seed the
2043      * chain with a zero capability.
2044      */
2045     pci_set_long(pdev->config + PCI_CONFIG_SPACE_SIZE,
2046                  PCI_EXT_CAP(0xFFFF, 0, 0));
2047     pci_set_long(pdev->wmask + PCI_CONFIG_SPACE_SIZE, 0);
2048     pci_set_long(vdev->emulated_config_bits + PCI_CONFIG_SPACE_SIZE, ~0);
2049 
2050     for (next = PCI_CONFIG_SPACE_SIZE; next;
2051          next = PCI_EXT_CAP_NEXT(pci_get_long(config + next))) {
2052         header = pci_get_long(config + next);
2053         cap_id = PCI_EXT_CAP_ID(header);
2054         cap_ver = PCI_EXT_CAP_VER(header);
2055 
2056         /*
2057          * If it becomes important to configure extended capabilities to their
2058          * actual size, use this as the default when it's something we don't
2059          * recognize. Since QEMU doesn't actually handle many of the config
2060          * accesses, exact size doesn't seem worthwhile.
2061          */
2062         size = vfio_ext_cap_max_size(config, next);
2063 
2064         /* Use emulated next pointer to allow dropping extended caps */
2065         pci_long_test_and_set_mask(vdev->emulated_config_bits + next,
2066                                    PCI_EXT_CAP_NEXT_MASK);
2067 
2068         switch (cap_id) {
2069         case 0: /* kernel masked capability */
2070         case PCI_EXT_CAP_ID_SRIOV: /* Read-only VF BARs confuse OVMF */
2071         case PCI_EXT_CAP_ID_ARI: /* XXX Needs next function virtualization */
2072         case PCI_EXT_CAP_ID_REBAR: /* Can't expose read-only */
2073             trace_vfio_add_ext_cap_dropped(vdev->vbasedev.name, cap_id, next);
2074             break;
2075         default:
2076             pcie_add_capability(pdev, cap_id, cap_ver, next, size);
2077         }
2078 
2079     }
2080 
2081     /* Cleanup chain head ID if necessary */
2082     if (pci_get_word(pdev->config + PCI_CONFIG_SPACE_SIZE) == 0xFFFF) {
2083         pci_set_word(pdev->config + PCI_CONFIG_SPACE_SIZE, 0);
2084     }
2085 
2086     g_free(config);
2087     return;
2088 }
2089 
2090 static int vfio_add_capabilities(VFIOPCIDevice *vdev, Error **errp)
2091 {
2092     PCIDevice *pdev = &vdev->pdev;
2093     int ret;
2094 
2095     if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
2096         !pdev->config[PCI_CAPABILITY_LIST]) {
2097         return 0; /* Nothing to add */
2098     }
2099 
2100     ret = vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST], errp);
2101     if (ret) {
2102         return ret;
2103     }
2104 
2105     vfio_add_ext_cap(vdev);
2106     return 0;
2107 }
2108 
2109 static void vfio_pci_pre_reset(VFIOPCIDevice *vdev)
2110 {
2111     PCIDevice *pdev = &vdev->pdev;
2112     uint16_t cmd;
2113 
2114     vfio_disable_interrupts(vdev);
2115 
2116     /* Make sure the device is in D0 */
2117     if (vdev->pm_cap) {
2118         uint16_t pmcsr;
2119         uint8_t state;
2120 
2121         pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2122         state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2123         if (state) {
2124             pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
2125             vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2);
2126             /* vfio handles the necessary delay here */
2127             pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2128             state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2129             if (state) {
2130                 error_report("vfio: Unable to power on device, stuck in D%d",
2131                              state);
2132             }
2133         }
2134     }
2135 
2136     /*
2137      * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
2138      * Also put INTx Disable in known state.
2139      */
2140     cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
2141     cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
2142              PCI_COMMAND_INTX_DISABLE);
2143     vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
2144 }
2145 
2146 static void vfio_pci_post_reset(VFIOPCIDevice *vdev)
2147 {
2148     Error *err = NULL;
2149     int nr;
2150 
2151     vfio_intx_enable(vdev, &err);
2152     if (err) {
2153         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
2154     }
2155 
2156     for (nr = 0; nr < PCI_NUM_REGIONS - 1; ++nr) {
2157         off_t addr = vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr);
2158         uint32_t val = 0;
2159         uint32_t len = sizeof(val);
2160 
2161         if (pwrite(vdev->vbasedev.fd, &val, len, addr) != len) {
2162             error_report("%s(%s) reset bar %d failed: %m", __func__,
2163                          vdev->vbasedev.name, nr);
2164         }
2165     }
2166 
2167     vfio_quirk_reset(vdev);
2168 }
2169 
2170 static bool vfio_pci_host_match(PCIHostDeviceAddress *addr, const char *name)
2171 {
2172     char tmp[13];
2173 
2174     sprintf(tmp, "%04x:%02x:%02x.%1x", addr->domain,
2175             addr->bus, addr->slot, addr->function);
2176 
2177     return (strcmp(tmp, name) == 0);
2178 }
2179 
2180 static int vfio_pci_hot_reset(VFIOPCIDevice *vdev, bool single)
2181 {
2182     VFIOGroup *group;
2183     struct vfio_pci_hot_reset_info *info;
2184     struct vfio_pci_dependent_device *devices;
2185     struct vfio_pci_hot_reset *reset;
2186     int32_t *fds;
2187     int ret, i, count;
2188     bool multi = false;
2189 
2190     trace_vfio_pci_hot_reset(vdev->vbasedev.name, single ? "one" : "multi");
2191 
2192     if (!single) {
2193         vfio_pci_pre_reset(vdev);
2194     }
2195     vdev->vbasedev.needs_reset = false;
2196 
2197     info = g_malloc0(sizeof(*info));
2198     info->argsz = sizeof(*info);
2199 
2200     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2201     if (ret && errno != ENOSPC) {
2202         ret = -errno;
2203         if (!vdev->has_pm_reset) {
2204             error_report("vfio: Cannot reset device %s, "
2205                          "no available reset mechanism.", vdev->vbasedev.name);
2206         }
2207         goto out_single;
2208     }
2209 
2210     count = info->count;
2211     info = g_realloc(info, sizeof(*info) + (count * sizeof(*devices)));
2212     info->argsz = sizeof(*info) + (count * sizeof(*devices));
2213     devices = &info->devices[0];
2214 
2215     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2216     if (ret) {
2217         ret = -errno;
2218         error_report("vfio: hot reset info failed: %m");
2219         goto out_single;
2220     }
2221 
2222     trace_vfio_pci_hot_reset_has_dep_devices(vdev->vbasedev.name);
2223 
2224     /* Verify that we have all the groups required */
2225     for (i = 0; i < info->count; i++) {
2226         PCIHostDeviceAddress host;
2227         VFIOPCIDevice *tmp;
2228         VFIODevice *vbasedev_iter;
2229 
2230         host.domain = devices[i].segment;
2231         host.bus = devices[i].bus;
2232         host.slot = PCI_SLOT(devices[i].devfn);
2233         host.function = PCI_FUNC(devices[i].devfn);
2234 
2235         trace_vfio_pci_hot_reset_dep_devices(host.domain,
2236                 host.bus, host.slot, host.function, devices[i].group_id);
2237 
2238         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2239             continue;
2240         }
2241 
2242         QLIST_FOREACH(group, &vfio_group_list, next) {
2243             if (group->groupid == devices[i].group_id) {
2244                 break;
2245             }
2246         }
2247 
2248         if (!group) {
2249             if (!vdev->has_pm_reset) {
2250                 error_report("vfio: Cannot reset device %s, "
2251                              "depends on group %d which is not owned.",
2252                              vdev->vbasedev.name, devices[i].group_id);
2253             }
2254             ret = -EPERM;
2255             goto out;
2256         }
2257 
2258         /* Prep dependent devices for reset and clear our marker. */
2259         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2260             if (!vbasedev_iter->dev->realized ||
2261                 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2262                 continue;
2263             }
2264             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2265             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2266                 if (single) {
2267                     ret = -EINVAL;
2268                     goto out_single;
2269                 }
2270                 vfio_pci_pre_reset(tmp);
2271                 tmp->vbasedev.needs_reset = false;
2272                 multi = true;
2273                 break;
2274             }
2275         }
2276     }
2277 
2278     if (!single && !multi) {
2279         ret = -EINVAL;
2280         goto out_single;
2281     }
2282 
2283     /* Determine how many group fds need to be passed */
2284     count = 0;
2285     QLIST_FOREACH(group, &vfio_group_list, next) {
2286         for (i = 0; i < info->count; i++) {
2287             if (group->groupid == devices[i].group_id) {
2288                 count++;
2289                 break;
2290             }
2291         }
2292     }
2293 
2294     reset = g_malloc0(sizeof(*reset) + (count * sizeof(*fds)));
2295     reset->argsz = sizeof(*reset) + (count * sizeof(*fds));
2296     fds = &reset->group_fds[0];
2297 
2298     /* Fill in group fds */
2299     QLIST_FOREACH(group, &vfio_group_list, next) {
2300         for (i = 0; i < info->count; i++) {
2301             if (group->groupid == devices[i].group_id) {
2302                 fds[reset->count++] = group->fd;
2303                 break;
2304             }
2305         }
2306     }
2307 
2308     /* Bus reset! */
2309     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_PCI_HOT_RESET, reset);
2310     g_free(reset);
2311 
2312     trace_vfio_pci_hot_reset_result(vdev->vbasedev.name,
2313                                     ret ? "%m" : "Success");
2314 
2315 out:
2316     /* Re-enable INTx on affected devices */
2317     for (i = 0; i < info->count; i++) {
2318         PCIHostDeviceAddress host;
2319         VFIOPCIDevice *tmp;
2320         VFIODevice *vbasedev_iter;
2321 
2322         host.domain = devices[i].segment;
2323         host.bus = devices[i].bus;
2324         host.slot = PCI_SLOT(devices[i].devfn);
2325         host.function = PCI_FUNC(devices[i].devfn);
2326 
2327         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2328             continue;
2329         }
2330 
2331         QLIST_FOREACH(group, &vfio_group_list, next) {
2332             if (group->groupid == devices[i].group_id) {
2333                 break;
2334             }
2335         }
2336 
2337         if (!group) {
2338             break;
2339         }
2340 
2341         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2342             if (!vbasedev_iter->dev->realized ||
2343                 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2344                 continue;
2345             }
2346             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2347             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2348                 vfio_pci_post_reset(tmp);
2349                 break;
2350             }
2351         }
2352     }
2353 out_single:
2354     if (!single) {
2355         vfio_pci_post_reset(vdev);
2356     }
2357     g_free(info);
2358 
2359     return ret;
2360 }
2361 
2362 /*
2363  * We want to differentiate hot reset of mulitple in-use devices vs hot reset
2364  * of a single in-use device.  VFIO_DEVICE_RESET will already handle the case
2365  * of doing hot resets when there is only a single device per bus.  The in-use
2366  * here refers to how many VFIODevices are affected.  A hot reset that affects
2367  * multiple devices, but only a single in-use device, means that we can call
2368  * it from our bus ->reset() callback since the extent is effectively a single
2369  * device.  This allows us to make use of it in the hotplug path.  When there
2370  * are multiple in-use devices, we can only trigger the hot reset during a
2371  * system reset and thus from our reset handler.  We separate _one vs _multi
2372  * here so that we don't overlap and do a double reset on the system reset
2373  * path where both our reset handler and ->reset() callback are used.  Calling
2374  * _one() will only do a hot reset for the one in-use devices case, calling
2375  * _multi() will do nothing if a _one() would have been sufficient.
2376  */
2377 static int vfio_pci_hot_reset_one(VFIOPCIDevice *vdev)
2378 {
2379     return vfio_pci_hot_reset(vdev, true);
2380 }
2381 
2382 static int vfio_pci_hot_reset_multi(VFIODevice *vbasedev)
2383 {
2384     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2385     return vfio_pci_hot_reset(vdev, false);
2386 }
2387 
2388 static void vfio_pci_compute_needs_reset(VFIODevice *vbasedev)
2389 {
2390     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2391     if (!vbasedev->reset_works || (!vdev->has_flr && vdev->has_pm_reset)) {
2392         vbasedev->needs_reset = true;
2393     }
2394 }
2395 
2396 static VFIODeviceOps vfio_pci_ops = {
2397     .vfio_compute_needs_reset = vfio_pci_compute_needs_reset,
2398     .vfio_hot_reset_multi = vfio_pci_hot_reset_multi,
2399     .vfio_eoi = vfio_intx_eoi,
2400 };
2401 
2402 int vfio_populate_vga(VFIOPCIDevice *vdev, Error **errp)
2403 {
2404     VFIODevice *vbasedev = &vdev->vbasedev;
2405     struct vfio_region_info *reg_info;
2406     int ret;
2407 
2408     ret = vfio_get_region_info(vbasedev, VFIO_PCI_VGA_REGION_INDEX, &reg_info);
2409     if (ret) {
2410         error_setg_errno(errp, -ret,
2411                          "failed getting region info for VGA region index %d",
2412                          VFIO_PCI_VGA_REGION_INDEX);
2413         return ret;
2414     }
2415 
2416     if (!(reg_info->flags & VFIO_REGION_INFO_FLAG_READ) ||
2417         !(reg_info->flags & VFIO_REGION_INFO_FLAG_WRITE) ||
2418         reg_info->size < 0xbffff + 1) {
2419         error_setg(errp, "unexpected VGA info, flags 0x%lx, size 0x%lx",
2420                    (unsigned long)reg_info->flags,
2421                    (unsigned long)reg_info->size);
2422         g_free(reg_info);
2423         return -EINVAL;
2424     }
2425 
2426     vdev->vga = g_new0(VFIOVGA, 1);
2427 
2428     vdev->vga->fd_offset = reg_info->offset;
2429     vdev->vga->fd = vdev->vbasedev.fd;
2430 
2431     g_free(reg_info);
2432 
2433     vdev->vga->region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE;
2434     vdev->vga->region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM;
2435     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_MEM].quirks);
2436 
2437     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2438                           OBJECT(vdev), &vfio_vga_ops,
2439                           &vdev->vga->region[QEMU_PCI_VGA_MEM],
2440                           "vfio-vga-mmio@0xa0000",
2441                           QEMU_PCI_VGA_MEM_SIZE);
2442 
2443     vdev->vga->region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE;
2444     vdev->vga->region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO;
2445     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].quirks);
2446 
2447     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2448                           OBJECT(vdev), &vfio_vga_ops,
2449                           &vdev->vga->region[QEMU_PCI_VGA_IO_LO],
2450                           "vfio-vga-io@0x3b0",
2451                           QEMU_PCI_VGA_IO_LO_SIZE);
2452 
2453     vdev->vga->region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE;
2454     vdev->vga->region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI;
2455     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks);
2456 
2457     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
2458                           OBJECT(vdev), &vfio_vga_ops,
2459                           &vdev->vga->region[QEMU_PCI_VGA_IO_HI],
2460                           "vfio-vga-io@0x3c0",
2461                           QEMU_PCI_VGA_IO_HI_SIZE);
2462 
2463     pci_register_vga(&vdev->pdev, &vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2464                      &vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2465                      &vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem);
2466 
2467     return 0;
2468 }
2469 
2470 static void vfio_populate_device(VFIOPCIDevice *vdev, Error **errp)
2471 {
2472     VFIODevice *vbasedev = &vdev->vbasedev;
2473     struct vfio_region_info *reg_info;
2474     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) };
2475     int i, ret = -1;
2476 
2477     /* Sanity check device */
2478     if (!(vbasedev->flags & VFIO_DEVICE_FLAGS_PCI)) {
2479         error_setg(errp, "this isn't a PCI device");
2480         return;
2481     }
2482 
2483     if (vbasedev->num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
2484         error_setg(errp, "unexpected number of io regions %u",
2485                    vbasedev->num_regions);
2486         return;
2487     }
2488 
2489     if (vbasedev->num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
2490         error_setg(errp, "unexpected number of irqs %u", vbasedev->num_irqs);
2491         return;
2492     }
2493 
2494     for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
2495         char *name = g_strdup_printf("%s BAR %d", vbasedev->name, i);
2496 
2497         ret = vfio_region_setup(OBJECT(vdev), vbasedev,
2498                                 &vdev->bars[i].region, i, name);
2499         g_free(name);
2500 
2501         if (ret) {
2502             error_setg_errno(errp, -ret, "failed to get region %d info", i);
2503             return;
2504         }
2505 
2506         QLIST_INIT(&vdev->bars[i].quirks);
2507     }
2508 
2509     ret = vfio_get_region_info(vbasedev,
2510                                VFIO_PCI_CONFIG_REGION_INDEX, &reg_info);
2511     if (ret) {
2512         error_setg_errno(errp, -ret, "failed to get config info");
2513         return;
2514     }
2515 
2516     trace_vfio_populate_device_config(vdev->vbasedev.name,
2517                                       (unsigned long)reg_info->size,
2518                                       (unsigned long)reg_info->offset,
2519                                       (unsigned long)reg_info->flags);
2520 
2521     vdev->config_size = reg_info->size;
2522     if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
2523         vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
2524     }
2525     vdev->config_offset = reg_info->offset;
2526 
2527     g_free(reg_info);
2528 
2529     if (vdev->features & VFIO_FEATURE_ENABLE_VGA) {
2530         ret = vfio_populate_vga(vdev, errp);
2531         if (ret) {
2532             error_append_hint(errp, "device does not support "
2533                               "requested feature x-vga\n");
2534             return;
2535         }
2536     }
2537 
2538     irq_info.index = VFIO_PCI_ERR_IRQ_INDEX;
2539 
2540     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info);
2541     if (ret) {
2542         /* This can fail for an old kernel or legacy PCI dev */
2543         trace_vfio_populate_device_get_irq_info_failure(strerror(errno));
2544     } else if (irq_info.count == 1) {
2545         vdev->pci_aer = true;
2546     } else {
2547         warn_report(VFIO_MSG_PREFIX
2548                     "Could not enable error recovery for the device",
2549                     vbasedev->name);
2550     }
2551 }
2552 
2553 static void vfio_put_device(VFIOPCIDevice *vdev)
2554 {
2555     g_free(vdev->vbasedev.name);
2556     g_free(vdev->msix);
2557 
2558     vfio_put_base_device(&vdev->vbasedev);
2559 }
2560 
2561 static void vfio_err_notifier_handler(void *opaque)
2562 {
2563     VFIOPCIDevice *vdev = opaque;
2564 
2565     if (!event_notifier_test_and_clear(&vdev->err_notifier)) {
2566         return;
2567     }
2568 
2569     /*
2570      * TBD. Retrieve the error details and decide what action
2571      * needs to be taken. One of the actions could be to pass
2572      * the error to the guest and have the guest driver recover
2573      * from the error. This requires that PCIe capabilities be
2574      * exposed to the guest. For now, we just terminate the
2575      * guest to contain the error.
2576      */
2577 
2578     error_report("%s(%s) Unrecoverable error detected. Please collect any data possible and then kill the guest", __func__, vdev->vbasedev.name);
2579 
2580     vm_stop(RUN_STATE_INTERNAL_ERROR);
2581 }
2582 
2583 /*
2584  * Registers error notifier for devices supporting error recovery.
2585  * If we encounter a failure in this function, we report an error
2586  * and continue after disabling error recovery support for the
2587  * device.
2588  */
2589 static void vfio_register_err_notifier(VFIOPCIDevice *vdev)
2590 {
2591     Error *err = NULL;
2592     int32_t fd;
2593 
2594     if (!vdev->pci_aer) {
2595         return;
2596     }
2597 
2598     if (event_notifier_init(&vdev->err_notifier, 0)) {
2599         error_report("vfio: Unable to init event notifier for error detection");
2600         vdev->pci_aer = false;
2601         return;
2602     }
2603 
2604     fd = event_notifier_get_fd(&vdev->err_notifier);
2605     qemu_set_fd_handler(fd, vfio_err_notifier_handler, NULL, vdev);
2606 
2607     if (vfio_set_irq_signaling(&vdev->vbasedev, VFIO_PCI_ERR_IRQ_INDEX, 0,
2608                                VFIO_IRQ_SET_ACTION_TRIGGER, fd, &err)) {
2609         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
2610         qemu_set_fd_handler(fd, NULL, NULL, vdev);
2611         event_notifier_cleanup(&vdev->err_notifier);
2612         vdev->pci_aer = false;
2613     }
2614 }
2615 
2616 static void vfio_unregister_err_notifier(VFIOPCIDevice *vdev)
2617 {
2618     Error *err = NULL;
2619 
2620     if (!vdev->pci_aer) {
2621         return;
2622     }
2623 
2624     if (vfio_set_irq_signaling(&vdev->vbasedev, VFIO_PCI_ERR_IRQ_INDEX, 0,
2625                                VFIO_IRQ_SET_ACTION_TRIGGER, -1, &err)) {
2626         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
2627     }
2628     qemu_set_fd_handler(event_notifier_get_fd(&vdev->err_notifier),
2629                         NULL, NULL, vdev);
2630     event_notifier_cleanup(&vdev->err_notifier);
2631 }
2632 
2633 static void vfio_req_notifier_handler(void *opaque)
2634 {
2635     VFIOPCIDevice *vdev = opaque;
2636     Error *err = NULL;
2637 
2638     if (!event_notifier_test_and_clear(&vdev->req_notifier)) {
2639         return;
2640     }
2641 
2642     qdev_unplug(DEVICE(vdev), &err);
2643     if (err) {
2644         warn_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
2645     }
2646 }
2647 
2648 static void vfio_register_req_notifier(VFIOPCIDevice *vdev)
2649 {
2650     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info),
2651                                       .index = VFIO_PCI_REQ_IRQ_INDEX };
2652     Error *err = NULL;
2653     int32_t fd;
2654 
2655     if (!(vdev->features & VFIO_FEATURE_ENABLE_REQ)) {
2656         return;
2657     }
2658 
2659     if (ioctl(vdev->vbasedev.fd,
2660               VFIO_DEVICE_GET_IRQ_INFO, &irq_info) < 0 || irq_info.count < 1) {
2661         return;
2662     }
2663 
2664     if (event_notifier_init(&vdev->req_notifier, 0)) {
2665         error_report("vfio: Unable to init event notifier for device request");
2666         return;
2667     }
2668 
2669     fd = event_notifier_get_fd(&vdev->req_notifier);
2670     qemu_set_fd_handler(fd, vfio_req_notifier_handler, NULL, vdev);
2671 
2672     if (vfio_set_irq_signaling(&vdev->vbasedev, VFIO_PCI_REQ_IRQ_INDEX, 0,
2673                            VFIO_IRQ_SET_ACTION_TRIGGER, fd, &err)) {
2674         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
2675         qemu_set_fd_handler(fd, NULL, NULL, vdev);
2676         event_notifier_cleanup(&vdev->req_notifier);
2677     } else {
2678         vdev->req_enabled = true;
2679     }
2680 }
2681 
2682 static void vfio_unregister_req_notifier(VFIOPCIDevice *vdev)
2683 {
2684     Error *err = NULL;
2685 
2686     if (!vdev->req_enabled) {
2687         return;
2688     }
2689 
2690     if (vfio_set_irq_signaling(&vdev->vbasedev, VFIO_PCI_REQ_IRQ_INDEX, 0,
2691                                VFIO_IRQ_SET_ACTION_TRIGGER, -1, &err)) {
2692         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
2693     }
2694     qemu_set_fd_handler(event_notifier_get_fd(&vdev->req_notifier),
2695                         NULL, NULL, vdev);
2696     event_notifier_cleanup(&vdev->req_notifier);
2697 
2698     vdev->req_enabled = false;
2699 }
2700 
2701 static void vfio_realize(PCIDevice *pdev, Error **errp)
2702 {
2703     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
2704     VFIODevice *vbasedev_iter;
2705     VFIOGroup *group;
2706     char *tmp, *subsys, group_path[PATH_MAX], *group_name;
2707     Error *err = NULL;
2708     ssize_t len;
2709     struct stat st;
2710     int groupid;
2711     int i, ret;
2712     bool is_mdev;
2713 
2714     if (!vdev->vbasedev.sysfsdev) {
2715         if (!(~vdev->host.domain || ~vdev->host.bus ||
2716               ~vdev->host.slot || ~vdev->host.function)) {
2717             error_setg(errp, "No provided host device");
2718             error_append_hint(errp, "Use -device vfio-pci,host=DDDD:BB:DD.F "
2719                               "or -device vfio-pci,sysfsdev=PATH_TO_DEVICE\n");
2720             return;
2721         }
2722         vdev->vbasedev.sysfsdev =
2723             g_strdup_printf("/sys/bus/pci/devices/%04x:%02x:%02x.%01x",
2724                             vdev->host.domain, vdev->host.bus,
2725                             vdev->host.slot, vdev->host.function);
2726     }
2727 
2728     if (stat(vdev->vbasedev.sysfsdev, &st) < 0) {
2729         error_setg_errno(errp, errno, "no such host device");
2730         error_prepend(errp, VFIO_MSG_PREFIX, vdev->vbasedev.sysfsdev);
2731         return;
2732     }
2733 
2734     vdev->vbasedev.name = g_path_get_basename(vdev->vbasedev.sysfsdev);
2735     vdev->vbasedev.ops = &vfio_pci_ops;
2736     vdev->vbasedev.type = VFIO_DEVICE_TYPE_PCI;
2737     vdev->vbasedev.dev = DEVICE(vdev);
2738 
2739     tmp = g_strdup_printf("%s/iommu_group", vdev->vbasedev.sysfsdev);
2740     len = readlink(tmp, group_path, sizeof(group_path));
2741     g_free(tmp);
2742 
2743     if (len <= 0 || len >= sizeof(group_path)) {
2744         error_setg_errno(errp, len < 0 ? errno : ENAMETOOLONG,
2745                          "no iommu_group found");
2746         goto error;
2747     }
2748 
2749     group_path[len] = 0;
2750 
2751     group_name = basename(group_path);
2752     if (sscanf(group_name, "%d", &groupid) != 1) {
2753         error_setg_errno(errp, errno, "failed to read %s", group_path);
2754         goto error;
2755     }
2756 
2757     trace_vfio_realize(vdev->vbasedev.name, groupid);
2758 
2759     group = vfio_get_group(groupid, pci_device_iommu_address_space(pdev), errp);
2760     if (!group) {
2761         goto error;
2762     }
2763 
2764     QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2765         if (strcmp(vbasedev_iter->name, vdev->vbasedev.name) == 0) {
2766             error_setg(errp, "device is already attached");
2767             vfio_put_group(group);
2768             goto error;
2769         }
2770     }
2771 
2772     /*
2773      * Mediated devices *might* operate compatibly with memory ballooning, but
2774      * we cannot know for certain, it depends on whether the mdev vendor driver
2775      * stays in sync with the active working set of the guest driver.  Prevent
2776      * the x-balloon-allowed option unless this is minimally an mdev device.
2777      */
2778     tmp = g_strdup_printf("%s/subsystem", vdev->vbasedev.sysfsdev);
2779     subsys = realpath(tmp, NULL);
2780     g_free(tmp);
2781     is_mdev = subsys && (strcmp(subsys, "/sys/bus/mdev") == 0);
2782     free(subsys);
2783 
2784     trace_vfio_mdev(vdev->vbasedev.name, is_mdev);
2785 
2786     if (vdev->vbasedev.balloon_allowed && !is_mdev) {
2787         error_setg(errp, "x-balloon-allowed only potentially compatible "
2788                    "with mdev devices");
2789         vfio_put_group(group);
2790         goto error;
2791     }
2792 
2793     ret = vfio_get_device(group, vdev->vbasedev.name, &vdev->vbasedev, errp);
2794     if (ret) {
2795         vfio_put_group(group);
2796         goto error;
2797     }
2798 
2799     vfio_populate_device(vdev, &err);
2800     if (err) {
2801         error_propagate(errp, err);
2802         goto error;
2803     }
2804 
2805     /* Get a copy of config space */
2806     ret = pread(vdev->vbasedev.fd, vdev->pdev.config,
2807                 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
2808                 vdev->config_offset);
2809     if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
2810         ret = ret < 0 ? -errno : -EFAULT;
2811         error_setg_errno(errp, -ret, "failed to read device config space");
2812         goto error;
2813     }
2814 
2815     /* vfio emulates a lot for us, but some bits need extra love */
2816     vdev->emulated_config_bits = g_malloc0(vdev->config_size);
2817 
2818     /* QEMU can choose to expose the ROM or not */
2819     memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4);
2820     /* QEMU can also add or extend BARs */
2821     memset(vdev->emulated_config_bits + PCI_BASE_ADDRESS_0, 0xff, 6 * 4);
2822 
2823     /*
2824      * The PCI spec reserves vendor ID 0xffff as an invalid value.  The
2825      * device ID is managed by the vendor and need only be a 16-bit value.
2826      * Allow any 16-bit value for subsystem so they can be hidden or changed.
2827      */
2828     if (vdev->vendor_id != PCI_ANY_ID) {
2829         if (vdev->vendor_id >= 0xffff) {
2830             error_setg(errp, "invalid PCI vendor ID provided");
2831             goto error;
2832         }
2833         vfio_add_emulated_word(vdev, PCI_VENDOR_ID, vdev->vendor_id, ~0);
2834         trace_vfio_pci_emulated_vendor_id(vdev->vbasedev.name, vdev->vendor_id);
2835     } else {
2836         vdev->vendor_id = pci_get_word(pdev->config + PCI_VENDOR_ID);
2837     }
2838 
2839     if (vdev->device_id != PCI_ANY_ID) {
2840         if (vdev->device_id > 0xffff) {
2841             error_setg(errp, "invalid PCI device ID provided");
2842             goto error;
2843         }
2844         vfio_add_emulated_word(vdev, PCI_DEVICE_ID, vdev->device_id, ~0);
2845         trace_vfio_pci_emulated_device_id(vdev->vbasedev.name, vdev->device_id);
2846     } else {
2847         vdev->device_id = pci_get_word(pdev->config + PCI_DEVICE_ID);
2848     }
2849 
2850     if (vdev->sub_vendor_id != PCI_ANY_ID) {
2851         if (vdev->sub_vendor_id > 0xffff) {
2852             error_setg(errp, "invalid PCI subsystem vendor ID provided");
2853             goto error;
2854         }
2855         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_VENDOR_ID,
2856                                vdev->sub_vendor_id, ~0);
2857         trace_vfio_pci_emulated_sub_vendor_id(vdev->vbasedev.name,
2858                                               vdev->sub_vendor_id);
2859     }
2860 
2861     if (vdev->sub_device_id != PCI_ANY_ID) {
2862         if (vdev->sub_device_id > 0xffff) {
2863             error_setg(errp, "invalid PCI subsystem device ID provided");
2864             goto error;
2865         }
2866         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_ID, vdev->sub_device_id, ~0);
2867         trace_vfio_pci_emulated_sub_device_id(vdev->vbasedev.name,
2868                                               vdev->sub_device_id);
2869     }
2870 
2871     /* QEMU can change multi-function devices to single function, or reverse */
2872     vdev->emulated_config_bits[PCI_HEADER_TYPE] =
2873                                               PCI_HEADER_TYPE_MULTI_FUNCTION;
2874 
2875     /* Restore or clear multifunction, this is always controlled by QEMU */
2876     if (vdev->pdev.cap_present & QEMU_PCI_CAP_MULTIFUNCTION) {
2877         vdev->pdev.config[PCI_HEADER_TYPE] |= PCI_HEADER_TYPE_MULTI_FUNCTION;
2878     } else {
2879         vdev->pdev.config[PCI_HEADER_TYPE] &= ~PCI_HEADER_TYPE_MULTI_FUNCTION;
2880     }
2881 
2882     /*
2883      * Clear host resource mapping info.  If we choose not to register a
2884      * BAR, such as might be the case with the option ROM, we can get
2885      * confusing, unwritable, residual addresses from the host here.
2886      */
2887     memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
2888     memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
2889 
2890     vfio_pci_size_rom(vdev);
2891 
2892     vfio_bars_prepare(vdev);
2893 
2894     vfio_msix_early_setup(vdev, &err);
2895     if (err) {
2896         error_propagate(errp, err);
2897         goto error;
2898     }
2899 
2900     vfio_bars_register(vdev);
2901 
2902     ret = vfio_add_capabilities(vdev, errp);
2903     if (ret) {
2904         goto out_teardown;
2905     }
2906 
2907     if (vdev->vga) {
2908         vfio_vga_quirk_setup(vdev);
2909     }
2910 
2911     for (i = 0; i < PCI_ROM_SLOT; i++) {
2912         vfio_bar_quirk_setup(vdev, i);
2913     }
2914 
2915     if (!vdev->igd_opregion &&
2916         vdev->features & VFIO_FEATURE_ENABLE_IGD_OPREGION) {
2917         struct vfio_region_info *opregion;
2918 
2919         if (vdev->pdev.qdev.hotplugged) {
2920             error_setg(errp,
2921                        "cannot support IGD OpRegion feature on hotplugged "
2922                        "device");
2923             goto out_teardown;
2924         }
2925 
2926         ret = vfio_get_dev_region_info(&vdev->vbasedev,
2927                         VFIO_REGION_TYPE_PCI_VENDOR_TYPE | PCI_VENDOR_ID_INTEL,
2928                         VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION, &opregion);
2929         if (ret) {
2930             error_setg_errno(errp, -ret,
2931                              "does not support requested IGD OpRegion feature");
2932             goto out_teardown;
2933         }
2934 
2935         ret = vfio_pci_igd_opregion_init(vdev, opregion, errp);
2936         g_free(opregion);
2937         if (ret) {
2938             goto out_teardown;
2939         }
2940     }
2941 
2942     /* QEMU emulates all of MSI & MSIX */
2943     if (pdev->cap_present & QEMU_PCI_CAP_MSIX) {
2944         memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff,
2945                MSIX_CAP_LENGTH);
2946     }
2947 
2948     if (pdev->cap_present & QEMU_PCI_CAP_MSI) {
2949         memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff,
2950                vdev->msi_cap_size);
2951     }
2952 
2953     if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
2954         vdev->intx.mmap_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL,
2955                                                   vfio_intx_mmap_enable, vdev);
2956         pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_intx_update);
2957         ret = vfio_intx_enable(vdev, errp);
2958         if (ret) {
2959             goto out_teardown;
2960         }
2961     }
2962 
2963     if (vdev->display != ON_OFF_AUTO_OFF) {
2964         ret = vfio_display_probe(vdev, errp);
2965         if (ret) {
2966             goto out_teardown;
2967         }
2968     }
2969     if (vdev->enable_ramfb && vdev->dpy == NULL) {
2970         error_setg(errp, "ramfb=on requires display=on");
2971         goto out_teardown;
2972     }
2973     if (vdev->display_xres || vdev->display_yres) {
2974         if (vdev->dpy == NULL) {
2975             error_setg(errp, "xres and yres properties require display=on");
2976             goto out_teardown;
2977         }
2978         if (vdev->dpy->edid_regs == NULL) {
2979             error_setg(errp, "xres and yres properties need edid support");
2980             goto out_teardown;
2981         }
2982     }
2983 
2984     if (vdev->vendor_id == PCI_VENDOR_ID_NVIDIA) {
2985         ret = vfio_pci_nvidia_v100_ram_init(vdev, errp);
2986         if (ret && ret != -ENODEV) {
2987             error_report("Failed to setup NVIDIA V100 GPU RAM");
2988         }
2989     }
2990 
2991     if (vdev->vendor_id == PCI_VENDOR_ID_IBM) {
2992         ret = vfio_pci_nvlink2_init(vdev, errp);
2993         if (ret && ret != -ENODEV) {
2994             error_report("Failed to setup NVlink2 bridge");
2995         }
2996     }
2997 
2998     vfio_register_err_notifier(vdev);
2999     vfio_register_req_notifier(vdev);
3000     vfio_setup_resetfn_quirk(vdev);
3001 
3002     return;
3003 
3004 out_teardown:
3005     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3006     vfio_teardown_msi(vdev);
3007     vfio_bars_exit(vdev);
3008 error:
3009     error_prepend(errp, VFIO_MSG_PREFIX, vdev->vbasedev.name);
3010 }
3011 
3012 static void vfio_instance_finalize(Object *obj)
3013 {
3014     VFIOPCIDevice *vdev = PCI_VFIO(obj);
3015     VFIOGroup *group = vdev->vbasedev.group;
3016 
3017     vfio_display_finalize(vdev);
3018     vfio_bars_finalize(vdev);
3019     g_free(vdev->emulated_config_bits);
3020     g_free(vdev->rom);
3021     /*
3022      * XXX Leaking igd_opregion is not an oversight, we can't remove the
3023      * fw_cfg entry therefore leaking this allocation seems like the safest
3024      * option.
3025      *
3026      * g_free(vdev->igd_opregion);
3027      */
3028     vfio_put_device(vdev);
3029     vfio_put_group(group);
3030 }
3031 
3032 static void vfio_exitfn(PCIDevice *pdev)
3033 {
3034     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
3035 
3036     vfio_unregister_req_notifier(vdev);
3037     vfio_unregister_err_notifier(vdev);
3038     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3039     vfio_disable_interrupts(vdev);
3040     if (vdev->intx.mmap_timer) {
3041         timer_free(vdev->intx.mmap_timer);
3042     }
3043     vfio_teardown_msi(vdev);
3044     vfio_bars_exit(vdev);
3045 }
3046 
3047 static void vfio_pci_reset(DeviceState *dev)
3048 {
3049     VFIOPCIDevice *vdev = PCI_VFIO(dev);
3050 
3051     trace_vfio_pci_reset(vdev->vbasedev.name);
3052 
3053     vfio_pci_pre_reset(vdev);
3054 
3055     if (vdev->display != ON_OFF_AUTO_OFF) {
3056         vfio_display_reset(vdev);
3057     }
3058 
3059     if (vdev->resetfn && !vdev->resetfn(vdev)) {
3060         goto post_reset;
3061     }
3062 
3063     if (vdev->vbasedev.reset_works &&
3064         (vdev->has_flr || !vdev->has_pm_reset) &&
3065         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
3066         trace_vfio_pci_reset_flr(vdev->vbasedev.name);
3067         goto post_reset;
3068     }
3069 
3070     /* See if we can do our own bus reset */
3071     if (!vfio_pci_hot_reset_one(vdev)) {
3072         goto post_reset;
3073     }
3074 
3075     /* If nothing else works and the device supports PM reset, use it */
3076     if (vdev->vbasedev.reset_works && vdev->has_pm_reset &&
3077         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
3078         trace_vfio_pci_reset_pm(vdev->vbasedev.name);
3079         goto post_reset;
3080     }
3081 
3082 post_reset:
3083     vfio_pci_post_reset(vdev);
3084 }
3085 
3086 static void vfio_instance_init(Object *obj)
3087 {
3088     PCIDevice *pci_dev = PCI_DEVICE(obj);
3089     VFIOPCIDevice *vdev = PCI_VFIO(obj);
3090 
3091     device_add_bootindex_property(obj, &vdev->bootindex,
3092                                   "bootindex", NULL,
3093                                   &pci_dev->qdev, NULL);
3094     vdev->host.domain = ~0U;
3095     vdev->host.bus = ~0U;
3096     vdev->host.slot = ~0U;
3097     vdev->host.function = ~0U;
3098 
3099     vdev->nv_gpudirect_clique = 0xFF;
3100 
3101     /* QEMU_PCI_CAP_EXPRESS initialization does not depend on QEMU command
3102      * line, therefore, no need to wait to realize like other devices */
3103     pci_dev->cap_present |= QEMU_PCI_CAP_EXPRESS;
3104 }
3105 
3106 static Property vfio_pci_dev_properties[] = {
3107     DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIOPCIDevice, host),
3108     DEFINE_PROP_STRING("sysfsdev", VFIOPCIDevice, vbasedev.sysfsdev),
3109     DEFINE_PROP_ON_OFF_AUTO("display", VFIOPCIDevice,
3110                             display, ON_OFF_AUTO_OFF),
3111     DEFINE_PROP_UINT32("xres", VFIOPCIDevice, display_xres, 0),
3112     DEFINE_PROP_UINT32("yres", VFIOPCIDevice, display_yres, 0),
3113     DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIOPCIDevice,
3114                        intx.mmap_timeout, 1100),
3115     DEFINE_PROP_BIT("x-vga", VFIOPCIDevice, features,
3116                     VFIO_FEATURE_ENABLE_VGA_BIT, false),
3117     DEFINE_PROP_BIT("x-req", VFIOPCIDevice, features,
3118                     VFIO_FEATURE_ENABLE_REQ_BIT, true),
3119     DEFINE_PROP_BIT("x-igd-opregion", VFIOPCIDevice, features,
3120                     VFIO_FEATURE_ENABLE_IGD_OPREGION_BIT, false),
3121     DEFINE_PROP_BOOL("x-no-mmap", VFIOPCIDevice, vbasedev.no_mmap, false),
3122     DEFINE_PROP_BOOL("x-balloon-allowed", VFIOPCIDevice,
3123                      vbasedev.balloon_allowed, false),
3124     DEFINE_PROP_BOOL("x-no-kvm-intx", VFIOPCIDevice, no_kvm_intx, false),
3125     DEFINE_PROP_BOOL("x-no-kvm-msi", VFIOPCIDevice, no_kvm_msi, false),
3126     DEFINE_PROP_BOOL("x-no-kvm-msix", VFIOPCIDevice, no_kvm_msix, false),
3127     DEFINE_PROP_BOOL("x-no-geforce-quirks", VFIOPCIDevice,
3128                      no_geforce_quirks, false),
3129     DEFINE_PROP_BOOL("x-no-kvm-ioeventfd", VFIOPCIDevice, no_kvm_ioeventfd,
3130                      false),
3131     DEFINE_PROP_BOOL("x-no-vfio-ioeventfd", VFIOPCIDevice, no_vfio_ioeventfd,
3132                      false),
3133     DEFINE_PROP_UINT32("x-pci-vendor-id", VFIOPCIDevice, vendor_id, PCI_ANY_ID),
3134     DEFINE_PROP_UINT32("x-pci-device-id", VFIOPCIDevice, device_id, PCI_ANY_ID),
3135     DEFINE_PROP_UINT32("x-pci-sub-vendor-id", VFIOPCIDevice,
3136                        sub_vendor_id, PCI_ANY_ID),
3137     DEFINE_PROP_UINT32("x-pci-sub-device-id", VFIOPCIDevice,
3138                        sub_device_id, PCI_ANY_ID),
3139     DEFINE_PROP_UINT32("x-igd-gms", VFIOPCIDevice, igd_gms, 0),
3140     DEFINE_PROP_UNSIGNED_NODEFAULT("x-nv-gpudirect-clique", VFIOPCIDevice,
3141                                    nv_gpudirect_clique,
3142                                    qdev_prop_nv_gpudirect_clique, uint8_t),
3143     DEFINE_PROP_OFF_AUTO_PCIBAR("x-msix-relocation", VFIOPCIDevice, msix_relo,
3144                                 OFF_AUTOPCIBAR_OFF),
3145     /*
3146      * TODO - support passed fds... is this necessary?
3147      * DEFINE_PROP_STRING("vfiofd", VFIOPCIDevice, vfiofd_name),
3148      * DEFINE_PROP_STRING("vfiogroupfd, VFIOPCIDevice, vfiogroupfd_name),
3149      */
3150     DEFINE_PROP_END_OF_LIST(),
3151 };
3152 
3153 static const VMStateDescription vfio_pci_vmstate = {
3154     .name = "vfio-pci",
3155     .unmigratable = 1,
3156 };
3157 
3158 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
3159 {
3160     DeviceClass *dc = DEVICE_CLASS(klass);
3161     PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
3162 
3163     dc->reset = vfio_pci_reset;
3164     dc->props = vfio_pci_dev_properties;
3165     dc->vmsd = &vfio_pci_vmstate;
3166     dc->desc = "VFIO-based PCI device assignment";
3167     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
3168     pdc->realize = vfio_realize;
3169     pdc->exit = vfio_exitfn;
3170     pdc->config_read = vfio_pci_read_config;
3171     pdc->config_write = vfio_pci_write_config;
3172 }
3173 
3174 static const TypeInfo vfio_pci_dev_info = {
3175     .name = TYPE_VFIO_PCI,
3176     .parent = TYPE_PCI_DEVICE,
3177     .instance_size = sizeof(VFIOPCIDevice),
3178     .class_init = vfio_pci_dev_class_init,
3179     .instance_init = vfio_instance_init,
3180     .instance_finalize = vfio_instance_finalize,
3181     .interfaces = (InterfaceInfo[]) {
3182         { INTERFACE_PCIE_DEVICE },
3183         { INTERFACE_CONVENTIONAL_PCI_DEVICE },
3184         { }
3185     },
3186 };
3187 
3188 static Property vfio_pci_dev_nohotplug_properties[] = {
3189     DEFINE_PROP_BOOL("ramfb", VFIOPCIDevice, enable_ramfb, false),
3190     DEFINE_PROP_END_OF_LIST(),
3191 };
3192 
3193 static void vfio_pci_nohotplug_dev_class_init(ObjectClass *klass, void *data)
3194 {
3195     DeviceClass *dc = DEVICE_CLASS(klass);
3196 
3197     dc->props = vfio_pci_dev_nohotplug_properties;
3198     dc->hotpluggable = false;
3199 }
3200 
3201 static const TypeInfo vfio_pci_nohotplug_dev_info = {
3202     .name = TYPE_VFIO_PCI_NOHOTPLUG,
3203     .parent = TYPE_VFIO_PCI,
3204     .instance_size = sizeof(VFIOPCIDevice),
3205     .class_init = vfio_pci_nohotplug_dev_class_init,
3206 };
3207 
3208 static void register_vfio_pci_dev_type(void)
3209 {
3210     type_register_static(&vfio_pci_dev_info);
3211     type_register_static(&vfio_pci_nohotplug_dev_info);
3212 }
3213 
3214 type_init(register_vfio_pci_dev_type)
3215