xref: /openbmc/qemu/hw/vfio/pci.c (revision 406d2aa2)
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/pci/msi.h"
26 #include "hw/pci/msix.h"
27 #include "hw/pci/pci_bridge.h"
28 #include "qemu/error-report.h"
29 #include "qemu/range.h"
30 #include "sysemu/kvm.h"
31 #include "sysemu/sysemu.h"
32 #include "pci.h"
33 #include "trace.h"
34 #include "qapi/error.h"
35 
36 #define MSIX_CAP_LENGTH 12
37 
38 static void vfio_disable_interrupts(VFIOPCIDevice *vdev);
39 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled);
40 
41 /*
42  * Disabling BAR mmaping can be slow, but toggling it around INTx can
43  * also be a huge overhead.  We try to get the best of both worlds by
44  * waiting until an interrupt to disable mmaps (subsequent transitions
45  * to the same state are effectively no overhead).  If the interrupt has
46  * been serviced and the time gap is long enough, we re-enable mmaps for
47  * performance.  This works well for things like graphics cards, which
48  * may not use their interrupt at all and are penalized to an unusable
49  * level by read/write BAR traps.  Other devices, like NICs, have more
50  * regular interrupts and see much better latency by staying in non-mmap
51  * mode.  We therefore set the default mmap_timeout such that a ping
52  * is just enough to keep the mmap disabled.  Users can experiment with
53  * other options with the x-intx-mmap-timeout-ms parameter (a value of
54  * zero disables the timer).
55  */
56 static void vfio_intx_mmap_enable(void *opaque)
57 {
58     VFIOPCIDevice *vdev = opaque;
59 
60     if (vdev->intx.pending) {
61         timer_mod(vdev->intx.mmap_timer,
62                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
63         return;
64     }
65 
66     vfio_mmap_set_enabled(vdev, true);
67 }
68 
69 static void vfio_intx_interrupt(void *opaque)
70 {
71     VFIOPCIDevice *vdev = opaque;
72 
73     if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
74         return;
75     }
76 
77     trace_vfio_intx_interrupt(vdev->vbasedev.name, 'A' + vdev->intx.pin);
78 
79     vdev->intx.pending = true;
80     pci_irq_assert(&vdev->pdev);
81     vfio_mmap_set_enabled(vdev, false);
82     if (vdev->intx.mmap_timeout) {
83         timer_mod(vdev->intx.mmap_timer,
84                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
85     }
86 }
87 
88 static void vfio_intx_eoi(VFIODevice *vbasedev)
89 {
90     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
91 
92     if (!vdev->intx.pending) {
93         return;
94     }
95 
96     trace_vfio_intx_eoi(vbasedev->name);
97 
98     vdev->intx.pending = false;
99     pci_irq_deassert(&vdev->pdev);
100     vfio_unmask_single_irqindex(vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
101 }
102 
103 static void vfio_intx_enable_kvm(VFIOPCIDevice *vdev, Error **errp)
104 {
105 #ifdef CONFIG_KVM
106     struct kvm_irqfd irqfd = {
107         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
108         .gsi = vdev->intx.route.irq,
109         .flags = KVM_IRQFD_FLAG_RESAMPLE,
110     };
111     struct vfio_irq_set *irq_set;
112     int ret, argsz;
113     int32_t *pfd;
114 
115     if (vdev->no_kvm_intx || !kvm_irqfds_enabled() ||
116         vdev->intx.route.mode != PCI_INTX_ENABLED ||
117         !kvm_resamplefds_enabled()) {
118         return;
119     }
120 
121     /* Get to a known interrupt state */
122     qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
123     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
124     vdev->intx.pending = false;
125     pci_irq_deassert(&vdev->pdev);
126 
127     /* Get an eventfd for resample/unmask */
128     if (event_notifier_init(&vdev->intx.unmask, 0)) {
129         error_setg(errp, "event_notifier_init failed eoi");
130         goto fail;
131     }
132 
133     /* KVM triggers it, VFIO listens for it */
134     irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
135 
136     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
137         error_setg_errno(errp, errno, "failed to setup resample irqfd");
138         goto fail_irqfd;
139     }
140 
141     argsz = sizeof(*irq_set) + sizeof(*pfd);
142 
143     irq_set = g_malloc0(argsz);
144     irq_set->argsz = argsz;
145     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
146     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
147     irq_set->start = 0;
148     irq_set->count = 1;
149     pfd = (int32_t *)&irq_set->data;
150 
151     *pfd = irqfd.resamplefd;
152 
153     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
154     g_free(irq_set);
155     if (ret) {
156         error_setg_errno(errp, -ret, "failed to setup INTx unmask fd");
157         goto fail_vfio;
158     }
159 
160     /* Let'em rip */
161     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
162 
163     vdev->intx.kvm_accel = true;
164 
165     trace_vfio_intx_enable_kvm(vdev->vbasedev.name);
166 
167     return;
168 
169 fail_vfio:
170     irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
171     kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
172 fail_irqfd:
173     event_notifier_cleanup(&vdev->intx.unmask);
174 fail:
175     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
176     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
177 #endif
178 }
179 
180 static void vfio_intx_disable_kvm(VFIOPCIDevice *vdev)
181 {
182 #ifdef CONFIG_KVM
183     struct kvm_irqfd irqfd = {
184         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
185         .gsi = vdev->intx.route.irq,
186         .flags = KVM_IRQFD_FLAG_DEASSIGN,
187     };
188 
189     if (!vdev->intx.kvm_accel) {
190         return;
191     }
192 
193     /*
194      * Get to a known state, hardware masked, QEMU ready to accept new
195      * interrupts, QEMU IRQ de-asserted.
196      */
197     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
198     vdev->intx.pending = false;
199     pci_irq_deassert(&vdev->pdev);
200 
201     /* Tell KVM to stop listening for an INTx irqfd */
202     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
203         error_report("vfio: Error: Failed to disable INTx irqfd: %m");
204     }
205 
206     /* We only need to close the eventfd for VFIO to cleanup the kernel side */
207     event_notifier_cleanup(&vdev->intx.unmask);
208 
209     /* QEMU starts listening for interrupt events. */
210     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
211 
212     vdev->intx.kvm_accel = false;
213 
214     /* If we've missed an event, let it re-fire through QEMU */
215     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
216 
217     trace_vfio_intx_disable_kvm(vdev->vbasedev.name);
218 #endif
219 }
220 
221 static void vfio_intx_update(PCIDevice *pdev)
222 {
223     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
224     PCIINTxRoute route;
225     Error *err = NULL;
226 
227     if (vdev->interrupt != VFIO_INT_INTx) {
228         return;
229     }
230 
231     route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
232 
233     if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
234         return; /* Nothing changed */
235     }
236 
237     trace_vfio_intx_update(vdev->vbasedev.name,
238                            vdev->intx.route.irq, route.irq);
239 
240     vfio_intx_disable_kvm(vdev);
241 
242     vdev->intx.route = route;
243 
244     if (route.mode != PCI_INTX_ENABLED) {
245         return;
246     }
247 
248     vfio_intx_enable_kvm(vdev, &err);
249     if (err) {
250         error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name);
251     }
252 
253     /* Re-enable the interrupt in cased we missed an EOI */
254     vfio_intx_eoi(&vdev->vbasedev);
255 }
256 
257 static int vfio_intx_enable(VFIOPCIDevice *vdev, Error **errp)
258 {
259     uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
260     int ret, argsz, retval = 0;
261     struct vfio_irq_set *irq_set;
262     int32_t *pfd;
263     Error *err = NULL;
264 
265     if (!pin) {
266         return 0;
267     }
268 
269     vfio_disable_interrupts(vdev);
270 
271     vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
272     pci_config_set_interrupt_pin(vdev->pdev.config, pin);
273 
274 #ifdef CONFIG_KVM
275     /*
276      * Only conditional to avoid generating error messages on platforms
277      * where we won't actually use the result anyway.
278      */
279     if (kvm_irqfds_enabled() && kvm_resamplefds_enabled()) {
280         vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
281                                                         vdev->intx.pin);
282     }
283 #endif
284 
285     ret = event_notifier_init(&vdev->intx.interrupt, 0);
286     if (ret) {
287         error_setg_errno(errp, -ret, "event_notifier_init failed");
288         return ret;
289     }
290 
291     argsz = sizeof(*irq_set) + sizeof(*pfd);
292 
293     irq_set = g_malloc0(argsz);
294     irq_set->argsz = argsz;
295     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
296     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
297     irq_set->start = 0;
298     irq_set->count = 1;
299     pfd = (int32_t *)&irq_set->data;
300 
301     *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
302     qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
303 
304     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
305     if (ret) {
306         error_setg_errno(errp, -ret, "failed to setup INTx fd");
307         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
308         event_notifier_cleanup(&vdev->intx.interrupt);
309         retval = -errno;
310         goto cleanup;
311     }
312 
313     vfio_intx_enable_kvm(vdev, &err);
314     if (err) {
315         error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name);
316     }
317 
318     vdev->interrupt = VFIO_INT_INTx;
319 
320     trace_vfio_intx_enable(vdev->vbasedev.name);
321 
322 cleanup:
323     g_free(irq_set);
324 
325     return retval;
326 }
327 
328 static void vfio_intx_disable(VFIOPCIDevice *vdev)
329 {
330     int fd;
331 
332     timer_del(vdev->intx.mmap_timer);
333     vfio_intx_disable_kvm(vdev);
334     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
335     vdev->intx.pending = false;
336     pci_irq_deassert(&vdev->pdev);
337     vfio_mmap_set_enabled(vdev, true);
338 
339     fd = event_notifier_get_fd(&vdev->intx.interrupt);
340     qemu_set_fd_handler(fd, NULL, NULL, vdev);
341     event_notifier_cleanup(&vdev->intx.interrupt);
342 
343     vdev->interrupt = VFIO_INT_NONE;
344 
345     trace_vfio_intx_disable(vdev->vbasedev.name);
346 }
347 
348 /*
349  * MSI/X
350  */
351 static void vfio_msi_interrupt(void *opaque)
352 {
353     VFIOMSIVector *vector = opaque;
354     VFIOPCIDevice *vdev = vector->vdev;
355     MSIMessage (*get_msg)(PCIDevice *dev, unsigned vector);
356     void (*notify)(PCIDevice *dev, unsigned vector);
357     MSIMessage msg;
358     int nr = vector - vdev->msi_vectors;
359 
360     if (!event_notifier_test_and_clear(&vector->interrupt)) {
361         return;
362     }
363 
364     if (vdev->interrupt == VFIO_INT_MSIX) {
365         get_msg = msix_get_message;
366         notify = msix_notify;
367 
368         /* A masked vector firing needs to use the PBA, enable it */
369         if (msix_is_masked(&vdev->pdev, nr)) {
370             set_bit(nr, vdev->msix->pending);
371             memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, true);
372             trace_vfio_msix_pba_enable(vdev->vbasedev.name);
373         }
374     } else if (vdev->interrupt == VFIO_INT_MSI) {
375         get_msg = msi_get_message;
376         notify = msi_notify;
377     } else {
378         abort();
379     }
380 
381     msg = get_msg(&vdev->pdev, nr);
382     trace_vfio_msi_interrupt(vdev->vbasedev.name, nr, msg.address, msg.data);
383     notify(&vdev->pdev, nr);
384 }
385 
386 static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool msix)
387 {
388     struct vfio_irq_set *irq_set;
389     int ret = 0, i, argsz;
390     int32_t *fds;
391 
392     argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
393 
394     irq_set = g_malloc0(argsz);
395     irq_set->argsz = argsz;
396     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
397     irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
398     irq_set->start = 0;
399     irq_set->count = vdev->nr_vectors;
400     fds = (int32_t *)&irq_set->data;
401 
402     for (i = 0; i < vdev->nr_vectors; i++) {
403         int fd = -1;
404 
405         /*
406          * MSI vs MSI-X - The guest has direct access to MSI mask and pending
407          * bits, therefore we always use the KVM signaling path when setup.
408          * MSI-X mask and pending bits are emulated, so we want to use the
409          * KVM signaling path only when configured and unmasked.
410          */
411         if (vdev->msi_vectors[i].use) {
412             if (vdev->msi_vectors[i].virq < 0 ||
413                 (msix && msix_is_masked(&vdev->pdev, i))) {
414                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
415             } else {
416                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].kvm_interrupt);
417             }
418         }
419 
420         fds[i] = fd;
421     }
422 
423     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
424 
425     g_free(irq_set);
426 
427     return ret;
428 }
429 
430 static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
431                                   int vector_n, bool msix)
432 {
433     int virq;
434 
435     if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) {
436         return;
437     }
438 
439     if (event_notifier_init(&vector->kvm_interrupt, 0)) {
440         return;
441     }
442 
443     virq = kvm_irqchip_add_msi_route(kvm_state, vector_n, &vdev->pdev);
444     if (virq < 0) {
445         event_notifier_cleanup(&vector->kvm_interrupt);
446         return;
447     }
448 
449     if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
450                                        NULL, virq) < 0) {
451         kvm_irqchip_release_virq(kvm_state, virq);
452         event_notifier_cleanup(&vector->kvm_interrupt);
453         return;
454     }
455 
456     vector->virq = virq;
457 }
458 
459 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
460 {
461     kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
462                                           vector->virq);
463     kvm_irqchip_release_virq(kvm_state, vector->virq);
464     vector->virq = -1;
465     event_notifier_cleanup(&vector->kvm_interrupt);
466 }
467 
468 static void vfio_update_kvm_msi_virq(VFIOMSIVector *vector, MSIMessage msg,
469                                      PCIDevice *pdev)
470 {
471     kvm_irqchip_update_msi_route(kvm_state, vector->virq, msg, pdev);
472     kvm_irqchip_commit_routes(kvm_state);
473 }
474 
475 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
476                                    MSIMessage *msg, IOHandler *handler)
477 {
478     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
479     VFIOMSIVector *vector;
480     int ret;
481 
482     trace_vfio_msix_vector_do_use(vdev->vbasedev.name, nr);
483 
484     vector = &vdev->msi_vectors[nr];
485 
486     if (!vector->use) {
487         vector->vdev = vdev;
488         vector->virq = -1;
489         if (event_notifier_init(&vector->interrupt, 0)) {
490             error_report("vfio: Error: event_notifier_init failed");
491         }
492         vector->use = true;
493         msix_vector_use(pdev, nr);
494     }
495 
496     qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
497                         handler, NULL, vector);
498 
499     /*
500      * Attempt to enable route through KVM irqchip,
501      * default to userspace handling if unavailable.
502      */
503     if (vector->virq >= 0) {
504         if (!msg) {
505             vfio_remove_kvm_msi_virq(vector);
506         } else {
507             vfio_update_kvm_msi_virq(vector, *msg, pdev);
508         }
509     } else {
510         if (msg) {
511             vfio_add_kvm_msi_virq(vdev, vector, nr, true);
512         }
513     }
514 
515     /*
516      * We don't want to have the host allocate all possible MSI vectors
517      * for a device if they're not in use, so we shutdown and incrementally
518      * increase them as needed.
519      */
520     if (vdev->nr_vectors < nr + 1) {
521         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
522         vdev->nr_vectors = nr + 1;
523         ret = vfio_enable_vectors(vdev, true);
524         if (ret) {
525             error_report("vfio: failed to enable vectors, %d", ret);
526         }
527     } else {
528         int argsz;
529         struct vfio_irq_set *irq_set;
530         int32_t *pfd;
531 
532         argsz = sizeof(*irq_set) + sizeof(*pfd);
533 
534         irq_set = g_malloc0(argsz);
535         irq_set->argsz = argsz;
536         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
537                          VFIO_IRQ_SET_ACTION_TRIGGER;
538         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
539         irq_set->start = nr;
540         irq_set->count = 1;
541         pfd = (int32_t *)&irq_set->data;
542 
543         if (vector->virq >= 0) {
544             *pfd = event_notifier_get_fd(&vector->kvm_interrupt);
545         } else {
546             *pfd = event_notifier_get_fd(&vector->interrupt);
547         }
548 
549         ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
550         g_free(irq_set);
551         if (ret) {
552             error_report("vfio: failed to modify vector, %d", ret);
553         }
554     }
555 
556     /* Disable PBA emulation when nothing more is pending. */
557     clear_bit(nr, vdev->msix->pending);
558     if (find_first_bit(vdev->msix->pending,
559                        vdev->nr_vectors) == vdev->nr_vectors) {
560         memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
561         trace_vfio_msix_pba_disable(vdev->vbasedev.name);
562     }
563 
564     return 0;
565 }
566 
567 static int vfio_msix_vector_use(PCIDevice *pdev,
568                                 unsigned int nr, MSIMessage msg)
569 {
570     return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
571 }
572 
573 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
574 {
575     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
576     VFIOMSIVector *vector = &vdev->msi_vectors[nr];
577 
578     trace_vfio_msix_vector_release(vdev->vbasedev.name, nr);
579 
580     /*
581      * There are still old guests that mask and unmask vectors on every
582      * interrupt.  If we're using QEMU bypass with a KVM irqfd, leave all of
583      * the KVM setup in place, simply switch VFIO to use the non-bypass
584      * eventfd.  We'll then fire the interrupt through QEMU and the MSI-X
585      * core will mask the interrupt and set pending bits, allowing it to
586      * be re-asserted on unmask.  Nothing to do if already using QEMU mode.
587      */
588     if (vector->virq >= 0) {
589         int argsz;
590         struct vfio_irq_set *irq_set;
591         int32_t *pfd;
592 
593         argsz = sizeof(*irq_set) + sizeof(*pfd);
594 
595         irq_set = g_malloc0(argsz);
596         irq_set->argsz = argsz;
597         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
598                          VFIO_IRQ_SET_ACTION_TRIGGER;
599         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
600         irq_set->start = nr;
601         irq_set->count = 1;
602         pfd = (int32_t *)&irq_set->data;
603 
604         *pfd = event_notifier_get_fd(&vector->interrupt);
605 
606         ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
607 
608         g_free(irq_set);
609     }
610 }
611 
612 static void vfio_msix_enable(VFIOPCIDevice *vdev)
613 {
614     vfio_disable_interrupts(vdev);
615 
616     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->msix->entries);
617 
618     vdev->interrupt = VFIO_INT_MSIX;
619 
620     /*
621      * Some communication channels between VF & PF or PF & fw rely on the
622      * physical state of the device and expect that enabling MSI-X from the
623      * guest enables the same on the host.  When our guest is Linux, the
624      * guest driver call to pci_enable_msix() sets the enabling bit in the
625      * MSI-X capability, but leaves the vector table masked.  We therefore
626      * can't rely on a vector_use callback (from request_irq() in the guest)
627      * to switch the physical device into MSI-X mode because that may come a
628      * long time after pci_enable_msix().  This code enables vector 0 with
629      * triggering to userspace, then immediately release the vector, leaving
630      * the physical device with no vectors enabled, but MSI-X enabled, just
631      * like the guest view.
632      */
633     vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
634     vfio_msix_vector_release(&vdev->pdev, 0);
635 
636     if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
637                                   vfio_msix_vector_release, NULL)) {
638         error_report("vfio: msix_set_vector_notifiers failed");
639     }
640 
641     trace_vfio_msix_enable(vdev->vbasedev.name);
642 }
643 
644 static void vfio_msi_enable(VFIOPCIDevice *vdev)
645 {
646     int ret, i;
647 
648     vfio_disable_interrupts(vdev);
649 
650     vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
651 retry:
652     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->nr_vectors);
653 
654     for (i = 0; i < vdev->nr_vectors; i++) {
655         VFIOMSIVector *vector = &vdev->msi_vectors[i];
656 
657         vector->vdev = vdev;
658         vector->virq = -1;
659         vector->use = true;
660 
661         if (event_notifier_init(&vector->interrupt, 0)) {
662             error_report("vfio: Error: event_notifier_init failed");
663         }
664 
665         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
666                             vfio_msi_interrupt, NULL, vector);
667 
668         /*
669          * Attempt to enable route through KVM irqchip,
670          * default to userspace handling if unavailable.
671          */
672         vfio_add_kvm_msi_virq(vdev, vector, i, false);
673     }
674 
675     /* Set interrupt type prior to possible interrupts */
676     vdev->interrupt = VFIO_INT_MSI;
677 
678     ret = vfio_enable_vectors(vdev, false);
679     if (ret) {
680         if (ret < 0) {
681             error_report("vfio: Error: Failed to setup MSI fds: %m");
682         } else if (ret != vdev->nr_vectors) {
683             error_report("vfio: Error: Failed to enable %d "
684                          "MSI vectors, retry with %d", vdev->nr_vectors, ret);
685         }
686 
687         for (i = 0; i < vdev->nr_vectors; i++) {
688             VFIOMSIVector *vector = &vdev->msi_vectors[i];
689             if (vector->virq >= 0) {
690                 vfio_remove_kvm_msi_virq(vector);
691             }
692             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
693                                 NULL, NULL, NULL);
694             event_notifier_cleanup(&vector->interrupt);
695         }
696 
697         g_free(vdev->msi_vectors);
698 
699         if (ret > 0 && ret != vdev->nr_vectors) {
700             vdev->nr_vectors = ret;
701             goto retry;
702         }
703         vdev->nr_vectors = 0;
704 
705         /*
706          * Failing to setup MSI doesn't really fall within any specification.
707          * Let's try leaving interrupts disabled and hope the guest figures
708          * out to fall back to INTx for this device.
709          */
710         error_report("vfio: Error: Failed to enable MSI");
711         vdev->interrupt = VFIO_INT_NONE;
712 
713         return;
714     }
715 
716     trace_vfio_msi_enable(vdev->vbasedev.name, vdev->nr_vectors);
717 }
718 
719 static void vfio_msi_disable_common(VFIOPCIDevice *vdev)
720 {
721     Error *err = NULL;
722     int i;
723 
724     for (i = 0; i < vdev->nr_vectors; i++) {
725         VFIOMSIVector *vector = &vdev->msi_vectors[i];
726         if (vdev->msi_vectors[i].use) {
727             if (vector->virq >= 0) {
728                 vfio_remove_kvm_msi_virq(vector);
729             }
730             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
731                                 NULL, NULL, NULL);
732             event_notifier_cleanup(&vector->interrupt);
733         }
734     }
735 
736     g_free(vdev->msi_vectors);
737     vdev->msi_vectors = NULL;
738     vdev->nr_vectors = 0;
739     vdev->interrupt = VFIO_INT_NONE;
740 
741     vfio_intx_enable(vdev, &err);
742     if (err) {
743         error_reportf_err(err, ERR_PREFIX, vdev->vbasedev.name);
744     }
745 }
746 
747 static void vfio_msix_disable(VFIOPCIDevice *vdev)
748 {
749     int i;
750 
751     msix_unset_vector_notifiers(&vdev->pdev);
752 
753     /*
754      * MSI-X will only release vectors if MSI-X is still enabled on the
755      * device, check through the rest and release it ourselves if necessary.
756      */
757     for (i = 0; i < vdev->nr_vectors; i++) {
758         if (vdev->msi_vectors[i].use) {
759             vfio_msix_vector_release(&vdev->pdev, i);
760             msix_vector_unuse(&vdev->pdev, i);
761         }
762     }
763 
764     if (vdev->nr_vectors) {
765         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
766     }
767 
768     vfio_msi_disable_common(vdev);
769 
770     memset(vdev->msix->pending, 0,
771            BITS_TO_LONGS(vdev->msix->entries) * sizeof(unsigned long));
772 
773     trace_vfio_msix_disable(vdev->vbasedev.name);
774 }
775 
776 static void vfio_msi_disable(VFIOPCIDevice *vdev)
777 {
778     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSI_IRQ_INDEX);
779     vfio_msi_disable_common(vdev);
780 
781     trace_vfio_msi_disable(vdev->vbasedev.name);
782 }
783 
784 static void vfio_update_msi(VFIOPCIDevice *vdev)
785 {
786     int i;
787 
788     for (i = 0; i < vdev->nr_vectors; i++) {
789         VFIOMSIVector *vector = &vdev->msi_vectors[i];
790         MSIMessage msg;
791 
792         if (!vector->use || vector->virq < 0) {
793             continue;
794         }
795 
796         msg = msi_get_message(&vdev->pdev, i);
797         vfio_update_kvm_msi_virq(vector, msg, &vdev->pdev);
798     }
799 }
800 
801 static void vfio_pci_load_rom(VFIOPCIDevice *vdev)
802 {
803     struct vfio_region_info *reg_info;
804     uint64_t size;
805     off_t off = 0;
806     ssize_t bytes;
807 
808     if (vfio_get_region_info(&vdev->vbasedev,
809                              VFIO_PCI_ROM_REGION_INDEX, &reg_info)) {
810         error_report("vfio: Error getting ROM info: %m");
811         return;
812     }
813 
814     trace_vfio_pci_load_rom(vdev->vbasedev.name, (unsigned long)reg_info->size,
815                             (unsigned long)reg_info->offset,
816                             (unsigned long)reg_info->flags);
817 
818     vdev->rom_size = size = reg_info->size;
819     vdev->rom_offset = reg_info->offset;
820 
821     g_free(reg_info);
822 
823     if (!vdev->rom_size) {
824         vdev->rom_read_failed = true;
825         error_report("vfio-pci: Cannot read device rom at "
826                     "%s", vdev->vbasedev.name);
827         error_printf("Device option ROM contents are probably invalid "
828                     "(check dmesg).\nSkip option ROM probe with rombar=0, "
829                     "or load from file with romfile=\n");
830         return;
831     }
832 
833     vdev->rom = g_malloc(size);
834     memset(vdev->rom, 0xff, size);
835 
836     while (size) {
837         bytes = pread(vdev->vbasedev.fd, vdev->rom + off,
838                       size, vdev->rom_offset + off);
839         if (bytes == 0) {
840             break;
841         } else if (bytes > 0) {
842             off += bytes;
843             size -= bytes;
844         } else {
845             if (errno == EINTR || errno == EAGAIN) {
846                 continue;
847             }
848             error_report("vfio: Error reading device ROM: %m");
849             break;
850         }
851     }
852 
853     /*
854      * Test the ROM signature against our device, if the vendor is correct
855      * but the device ID doesn't match, store the correct device ID and
856      * recompute the checksum.  Intel IGD devices need this and are known
857      * to have bogus checksums so we can't simply adjust the checksum.
858      */
859     if (pci_get_word(vdev->rom) == 0xaa55 &&
860         pci_get_word(vdev->rom + 0x18) + 8 < vdev->rom_size &&
861         !memcmp(vdev->rom + pci_get_word(vdev->rom + 0x18), "PCIR", 4)) {
862         uint16_t vid, did;
863 
864         vid = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 4);
865         did = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6);
866 
867         if (vid == vdev->vendor_id && did != vdev->device_id) {
868             int i;
869             uint8_t csum, *data = vdev->rom;
870 
871             pci_set_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6,
872                          vdev->device_id);
873             data[6] = 0;
874 
875             for (csum = 0, i = 0; i < vdev->rom_size; i++) {
876                 csum += data[i];
877             }
878 
879             data[6] = -csum;
880         }
881     }
882 }
883 
884 static uint64_t vfio_rom_read(void *opaque, hwaddr addr, unsigned size)
885 {
886     VFIOPCIDevice *vdev = opaque;
887     union {
888         uint8_t byte;
889         uint16_t word;
890         uint32_t dword;
891         uint64_t qword;
892     } val;
893     uint64_t data = 0;
894 
895     /* Load the ROM lazily when the guest tries to read it */
896     if (unlikely(!vdev->rom && !vdev->rom_read_failed)) {
897         vfio_pci_load_rom(vdev);
898     }
899 
900     memcpy(&val, vdev->rom + addr,
901            (addr < vdev->rom_size) ? MIN(size, vdev->rom_size - addr) : 0);
902 
903     switch (size) {
904     case 1:
905         data = val.byte;
906         break;
907     case 2:
908         data = le16_to_cpu(val.word);
909         break;
910     case 4:
911         data = le32_to_cpu(val.dword);
912         break;
913     default:
914         hw_error("vfio: unsupported read size, %d bytes\n", size);
915         break;
916     }
917 
918     trace_vfio_rom_read(vdev->vbasedev.name, addr, size, data);
919 
920     return data;
921 }
922 
923 static void vfio_rom_write(void *opaque, hwaddr addr,
924                            uint64_t data, unsigned size)
925 {
926 }
927 
928 static const MemoryRegionOps vfio_rom_ops = {
929     .read = vfio_rom_read,
930     .write = vfio_rom_write,
931     .endianness = DEVICE_LITTLE_ENDIAN,
932 };
933 
934 static void vfio_pci_size_rom(VFIOPCIDevice *vdev)
935 {
936     uint32_t orig, size = cpu_to_le32((uint32_t)PCI_ROM_ADDRESS_MASK);
937     off_t offset = vdev->config_offset + PCI_ROM_ADDRESS;
938     DeviceState *dev = DEVICE(vdev);
939     char *name;
940     int fd = vdev->vbasedev.fd;
941 
942     if (vdev->pdev.romfile || !vdev->pdev.rom_bar) {
943         /* Since pci handles romfile, just print a message and return */
944         if (vfio_blacklist_opt_rom(vdev) && vdev->pdev.romfile) {
945             error_printf("Warning : Device at %s is known to cause system instability issues during option rom execution. Proceeding anyway since user specified romfile\n",
946                          vdev->vbasedev.name);
947         }
948         return;
949     }
950 
951     /*
952      * Use the same size ROM BAR as the physical device.  The contents
953      * will get filled in later when the guest tries to read it.
954      */
955     if (pread(fd, &orig, 4, offset) != 4 ||
956         pwrite(fd, &size, 4, offset) != 4 ||
957         pread(fd, &size, 4, offset) != 4 ||
958         pwrite(fd, &orig, 4, offset) != 4) {
959         error_report("%s(%s) failed: %m", __func__, vdev->vbasedev.name);
960         return;
961     }
962 
963     size = ~(le32_to_cpu(size) & PCI_ROM_ADDRESS_MASK) + 1;
964 
965     if (!size) {
966         return;
967     }
968 
969     if (vfio_blacklist_opt_rom(vdev)) {
970         if (dev->opts && qemu_opt_get(dev->opts, "rombar")) {
971             error_printf("Warning : Device at %s is known to cause system instability issues during option rom execution. Proceeding anyway since user specified non zero value for rombar\n",
972                          vdev->vbasedev.name);
973         } else {
974             error_printf("Warning : Rom loading for device at %s has been disabled due to system instability issues. Specify rombar=1 or romfile to force\n",
975                          vdev->vbasedev.name);
976             return;
977         }
978     }
979 
980     trace_vfio_pci_size_rom(vdev->vbasedev.name, size);
981 
982     name = g_strdup_printf("vfio[%s].rom", vdev->vbasedev.name);
983 
984     memory_region_init_io(&vdev->pdev.rom, OBJECT(vdev),
985                           &vfio_rom_ops, vdev, name, size);
986     g_free(name);
987 
988     pci_register_bar(&vdev->pdev, PCI_ROM_SLOT,
989                      PCI_BASE_ADDRESS_SPACE_MEMORY, &vdev->pdev.rom);
990 
991     vdev->pdev.has_rom = true;
992     vdev->rom_read_failed = false;
993 }
994 
995 void vfio_vga_write(void *opaque, hwaddr addr,
996                            uint64_t data, unsigned size)
997 {
998     VFIOVGARegion *region = opaque;
999     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1000     union {
1001         uint8_t byte;
1002         uint16_t word;
1003         uint32_t dword;
1004         uint64_t qword;
1005     } buf;
1006     off_t offset = vga->fd_offset + region->offset + addr;
1007 
1008     switch (size) {
1009     case 1:
1010         buf.byte = data;
1011         break;
1012     case 2:
1013         buf.word = cpu_to_le16(data);
1014         break;
1015     case 4:
1016         buf.dword = cpu_to_le32(data);
1017         break;
1018     default:
1019         hw_error("vfio: unsupported write size, %d bytes", size);
1020         break;
1021     }
1022 
1023     if (pwrite(vga->fd, &buf, size, offset) != size) {
1024         error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
1025                      __func__, region->offset + addr, data, size);
1026     }
1027 
1028     trace_vfio_vga_write(region->offset + addr, data, size);
1029 }
1030 
1031 uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size)
1032 {
1033     VFIOVGARegion *region = opaque;
1034     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1035     union {
1036         uint8_t byte;
1037         uint16_t word;
1038         uint32_t dword;
1039         uint64_t qword;
1040     } buf;
1041     uint64_t data = 0;
1042     off_t offset = vga->fd_offset + region->offset + addr;
1043 
1044     if (pread(vga->fd, &buf, size, offset) != size) {
1045         error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
1046                      __func__, region->offset + addr, size);
1047         return (uint64_t)-1;
1048     }
1049 
1050     switch (size) {
1051     case 1:
1052         data = buf.byte;
1053         break;
1054     case 2:
1055         data = le16_to_cpu(buf.word);
1056         break;
1057     case 4:
1058         data = le32_to_cpu(buf.dword);
1059         break;
1060     default:
1061         hw_error("vfio: unsupported read size, %d bytes", size);
1062         break;
1063     }
1064 
1065     trace_vfio_vga_read(region->offset + addr, size, data);
1066 
1067     return data;
1068 }
1069 
1070 static const MemoryRegionOps vfio_vga_ops = {
1071     .read = vfio_vga_read,
1072     .write = vfio_vga_write,
1073     .endianness = DEVICE_LITTLE_ENDIAN,
1074 };
1075 
1076 /*
1077  * Expand memory region of sub-page(size < PAGE_SIZE) MMIO BAR to page
1078  * size if the BAR is in an exclusive page in host so that we could map
1079  * this BAR to guest. But this sub-page BAR may not occupy an exclusive
1080  * page in guest. So we should set the priority of the expanded memory
1081  * region to zero in case of overlap with BARs which share the same page
1082  * with the sub-page BAR in guest. Besides, we should also recover the
1083  * size of this sub-page BAR when its base address is changed in guest
1084  * and not page aligned any more.
1085  */
1086 static void vfio_sub_page_bar_update_mapping(PCIDevice *pdev, int bar)
1087 {
1088     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
1089     VFIORegion *region = &vdev->bars[bar].region;
1090     MemoryRegion *mmap_mr, *mr;
1091     PCIIORegion *r;
1092     pcibus_t bar_addr;
1093     uint64_t size = region->size;
1094 
1095     /* Make sure that the whole region is allowed to be mmapped */
1096     if (region->nr_mmaps != 1 || !region->mmaps[0].mmap ||
1097         region->mmaps[0].size != region->size) {
1098         return;
1099     }
1100 
1101     r = &pdev->io_regions[bar];
1102     bar_addr = r->addr;
1103     mr = region->mem;
1104     mmap_mr = &region->mmaps[0].mem;
1105 
1106     /* If BAR is mapped and page aligned, update to fill PAGE_SIZE */
1107     if (bar_addr != PCI_BAR_UNMAPPED &&
1108         !(bar_addr & ~qemu_real_host_page_mask)) {
1109         size = qemu_real_host_page_size;
1110     }
1111 
1112     memory_region_transaction_begin();
1113 
1114     memory_region_set_size(mr, size);
1115     memory_region_set_size(mmap_mr, size);
1116     if (size != region->size && memory_region_is_mapped(mr)) {
1117         memory_region_del_subregion(r->address_space, mr);
1118         memory_region_add_subregion_overlap(r->address_space,
1119                                             bar_addr, mr, 0);
1120     }
1121 
1122     memory_region_transaction_commit();
1123 }
1124 
1125 /*
1126  * PCI config space
1127  */
1128 uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
1129 {
1130     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
1131     uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val;
1132 
1133     memcpy(&emu_bits, vdev->emulated_config_bits + addr, len);
1134     emu_bits = le32_to_cpu(emu_bits);
1135 
1136     if (emu_bits) {
1137         emu_val = pci_default_read_config(pdev, addr, len);
1138     }
1139 
1140     if (~emu_bits & (0xffffffffU >> (32 - len * 8))) {
1141         ssize_t ret;
1142 
1143         ret = pread(vdev->vbasedev.fd, &phys_val, len,
1144                     vdev->config_offset + addr);
1145         if (ret != len) {
1146             error_report("%s(%s, 0x%x, 0x%x) failed: %m",
1147                          __func__, vdev->vbasedev.name, addr, len);
1148             return -errno;
1149         }
1150         phys_val = le32_to_cpu(phys_val);
1151     }
1152 
1153     val = (emu_val & emu_bits) | (phys_val & ~emu_bits);
1154 
1155     trace_vfio_pci_read_config(vdev->vbasedev.name, addr, len, val);
1156 
1157     return val;
1158 }
1159 
1160 void vfio_pci_write_config(PCIDevice *pdev,
1161                            uint32_t addr, uint32_t val, int len)
1162 {
1163     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
1164     uint32_t val_le = cpu_to_le32(val);
1165 
1166     trace_vfio_pci_write_config(vdev->vbasedev.name, addr, val, len);
1167 
1168     /* Write everything to VFIO, let it filter out what we can't write */
1169     if (pwrite(vdev->vbasedev.fd, &val_le, len, vdev->config_offset + addr)
1170                 != len) {
1171         error_report("%s(%s, 0x%x, 0x%x, 0x%x) failed: %m",
1172                      __func__, vdev->vbasedev.name, addr, val, len);
1173     }
1174 
1175     /* MSI/MSI-X Enabling/Disabling */
1176     if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
1177         ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
1178         int is_enabled, was_enabled = msi_enabled(pdev);
1179 
1180         pci_default_write_config(pdev, addr, val, len);
1181 
1182         is_enabled = msi_enabled(pdev);
1183 
1184         if (!was_enabled) {
1185             if (is_enabled) {
1186                 vfio_msi_enable(vdev);
1187             }
1188         } else {
1189             if (!is_enabled) {
1190                 vfio_msi_disable(vdev);
1191             } else {
1192                 vfio_update_msi(vdev);
1193             }
1194         }
1195     } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
1196         ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
1197         int is_enabled, was_enabled = msix_enabled(pdev);
1198 
1199         pci_default_write_config(pdev, addr, val, len);
1200 
1201         is_enabled = msix_enabled(pdev);
1202 
1203         if (!was_enabled && is_enabled) {
1204             vfio_msix_enable(vdev);
1205         } else if (was_enabled && !is_enabled) {
1206             vfio_msix_disable(vdev);
1207         }
1208     } else if (ranges_overlap(addr, len, PCI_BASE_ADDRESS_0, 24) ||
1209         range_covers_byte(addr, len, PCI_COMMAND)) {
1210         pcibus_t old_addr[PCI_NUM_REGIONS - 1];
1211         int bar;
1212 
1213         for (bar = 0; bar < PCI_ROM_SLOT; bar++) {
1214             old_addr[bar] = pdev->io_regions[bar].addr;
1215         }
1216 
1217         pci_default_write_config(pdev, addr, val, len);
1218 
1219         for (bar = 0; bar < PCI_ROM_SLOT; bar++) {
1220             if (old_addr[bar] != pdev->io_regions[bar].addr &&
1221                 pdev->io_regions[bar].size > 0 &&
1222                 pdev->io_regions[bar].size < qemu_real_host_page_size) {
1223                 vfio_sub_page_bar_update_mapping(pdev, bar);
1224             }
1225         }
1226     } else {
1227         /* Write everything to QEMU to keep emulated bits correct */
1228         pci_default_write_config(pdev, addr, val, len);
1229     }
1230 }
1231 
1232 /*
1233  * Interrupt setup
1234  */
1235 static void vfio_disable_interrupts(VFIOPCIDevice *vdev)
1236 {
1237     /*
1238      * More complicated than it looks.  Disabling MSI/X transitions the
1239      * device to INTx mode (if supported).  Therefore we need to first
1240      * disable MSI/X and then cleanup by disabling INTx.
1241      */
1242     if (vdev->interrupt == VFIO_INT_MSIX) {
1243         vfio_msix_disable(vdev);
1244     } else if (vdev->interrupt == VFIO_INT_MSI) {
1245         vfio_msi_disable(vdev);
1246     }
1247 
1248     if (vdev->interrupt == VFIO_INT_INTx) {
1249         vfio_intx_disable(vdev);
1250     }
1251 }
1252 
1253 static int vfio_msi_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1254 {
1255     uint16_t ctrl;
1256     bool msi_64bit, msi_maskbit;
1257     int ret, entries;
1258     Error *err = NULL;
1259 
1260     if (pread(vdev->vbasedev.fd, &ctrl, sizeof(ctrl),
1261               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1262         error_setg_errno(errp, errno, "failed reading MSI PCI_CAP_FLAGS");
1263         return -errno;
1264     }
1265     ctrl = le16_to_cpu(ctrl);
1266 
1267     msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
1268     msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
1269     entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
1270 
1271     trace_vfio_msi_setup(vdev->vbasedev.name, pos);
1272 
1273     ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit, &err);
1274     if (ret < 0) {
1275         if (ret == -ENOTSUP) {
1276             return 0;
1277         }
1278         error_prepend(&err, "msi_init failed: ");
1279         error_propagate(errp, err);
1280         return ret;
1281     }
1282     vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
1283 
1284     return 0;
1285 }
1286 
1287 static void vfio_pci_fixup_msix_region(VFIOPCIDevice *vdev)
1288 {
1289     off_t start, end;
1290     VFIORegion *region = &vdev->bars[vdev->msix->table_bar].region;
1291 
1292     /*
1293      * We expect to find a single mmap covering the whole BAR, anything else
1294      * means it's either unsupported or already setup.
1295      */
1296     if (region->nr_mmaps != 1 || region->mmaps[0].offset ||
1297         region->size != region->mmaps[0].size) {
1298         return;
1299     }
1300 
1301     /* MSI-X table start and end aligned to host page size */
1302     start = vdev->msix->table_offset & qemu_real_host_page_mask;
1303     end = REAL_HOST_PAGE_ALIGN((uint64_t)vdev->msix->table_offset +
1304                                (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
1305 
1306     /*
1307      * Does the MSI-X table cover the beginning of the BAR?  The whole BAR?
1308      * NB - Host page size is necessarily a power of two and so is the PCI
1309      * BAR (not counting EA yet), therefore if we have host page aligned
1310      * @start and @end, then any remainder of the BAR before or after those
1311      * must be at least host page sized and therefore mmap'able.
1312      */
1313     if (!start) {
1314         if (end >= region->size) {
1315             region->nr_mmaps = 0;
1316             g_free(region->mmaps);
1317             region->mmaps = NULL;
1318             trace_vfio_msix_fixup(vdev->vbasedev.name,
1319                                   vdev->msix->table_bar, 0, 0);
1320         } else {
1321             region->mmaps[0].offset = end;
1322             region->mmaps[0].size = region->size - end;
1323             trace_vfio_msix_fixup(vdev->vbasedev.name,
1324                               vdev->msix->table_bar, region->mmaps[0].offset,
1325                               region->mmaps[0].offset + region->mmaps[0].size);
1326         }
1327 
1328     /* Maybe it's aligned at the end of the BAR */
1329     } else if (end >= region->size) {
1330         region->mmaps[0].size = start;
1331         trace_vfio_msix_fixup(vdev->vbasedev.name,
1332                               vdev->msix->table_bar, region->mmaps[0].offset,
1333                               region->mmaps[0].offset + region->mmaps[0].size);
1334 
1335     /* Otherwise it must split the BAR */
1336     } else {
1337         region->nr_mmaps = 2;
1338         region->mmaps = g_renew(VFIOMmap, region->mmaps, 2);
1339 
1340         memcpy(&region->mmaps[1], &region->mmaps[0], sizeof(VFIOMmap));
1341 
1342         region->mmaps[0].size = start;
1343         trace_vfio_msix_fixup(vdev->vbasedev.name,
1344                               vdev->msix->table_bar, region->mmaps[0].offset,
1345                               region->mmaps[0].offset + region->mmaps[0].size);
1346 
1347         region->mmaps[1].offset = end;
1348         region->mmaps[1].size = region->size - end;
1349         trace_vfio_msix_fixup(vdev->vbasedev.name,
1350                               vdev->msix->table_bar, region->mmaps[1].offset,
1351                               region->mmaps[1].offset + region->mmaps[1].size);
1352     }
1353 }
1354 
1355 /*
1356  * We don't have any control over how pci_add_capability() inserts
1357  * capabilities into the chain.  In order to setup MSI-X we need a
1358  * MemoryRegion for the BAR.  In order to setup the BAR and not
1359  * attempt to mmap the MSI-X table area, which VFIO won't allow, we
1360  * need to first look for where the MSI-X table lives.  So we
1361  * unfortunately split MSI-X setup across two functions.
1362  */
1363 static void vfio_msix_early_setup(VFIOPCIDevice *vdev, Error **errp)
1364 {
1365     uint8_t pos;
1366     uint16_t ctrl;
1367     uint32_t table, pba;
1368     int fd = vdev->vbasedev.fd;
1369     VFIOMSIXInfo *msix;
1370 
1371     pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
1372     if (!pos) {
1373         return;
1374     }
1375 
1376     if (pread(fd, &ctrl, sizeof(ctrl),
1377               vdev->config_offset + pos + PCI_MSIX_FLAGS) != sizeof(ctrl)) {
1378         error_setg_errno(errp, errno, "failed to read PCI MSIX FLAGS");
1379         return;
1380     }
1381 
1382     if (pread(fd, &table, sizeof(table),
1383               vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
1384         error_setg_errno(errp, errno, "failed to read PCI MSIX TABLE");
1385         return;
1386     }
1387 
1388     if (pread(fd, &pba, sizeof(pba),
1389               vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
1390         error_setg_errno(errp, errno, "failed to read PCI MSIX PBA");
1391         return;
1392     }
1393 
1394     ctrl = le16_to_cpu(ctrl);
1395     table = le32_to_cpu(table);
1396     pba = le32_to_cpu(pba);
1397 
1398     msix = g_malloc0(sizeof(*msix));
1399     msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
1400     msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
1401     msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
1402     msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
1403     msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
1404 
1405     /*
1406      * Test the size of the pba_offset variable and catch if it extends outside
1407      * of the specified BAR. If it is the case, we need to apply a hardware
1408      * specific quirk if the device is known or we have a broken configuration.
1409      */
1410     if (msix->pba_offset >= vdev->bars[msix->pba_bar].region.size) {
1411         /*
1412          * Chelsio T5 Virtual Function devices are encoded as 0x58xx for T5
1413          * adapters. The T5 hardware returns an incorrect value of 0x8000 for
1414          * the VF PBA offset while the BAR itself is only 8k. The correct value
1415          * is 0x1000, so we hard code that here.
1416          */
1417         if (vdev->vendor_id == PCI_VENDOR_ID_CHELSIO &&
1418             (vdev->device_id & 0xff00) == 0x5800) {
1419             msix->pba_offset = 0x1000;
1420         } else {
1421             error_setg(errp, "hardware reports invalid configuration, "
1422                        "MSIX PBA outside of specified BAR");
1423             g_free(msix);
1424             return;
1425         }
1426     }
1427 
1428     trace_vfio_msix_early_setup(vdev->vbasedev.name, pos, msix->table_bar,
1429                                 msix->table_offset, msix->entries);
1430     vdev->msix = msix;
1431 
1432     vfio_pci_fixup_msix_region(vdev);
1433 }
1434 
1435 static int vfio_msix_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1436 {
1437     int ret;
1438     Error *err = NULL;
1439 
1440     vdev->msix->pending = g_malloc0(BITS_TO_LONGS(vdev->msix->entries) *
1441                                     sizeof(unsigned long));
1442     ret = msix_init(&vdev->pdev, vdev->msix->entries,
1443                     vdev->bars[vdev->msix->table_bar].region.mem,
1444                     vdev->msix->table_bar, vdev->msix->table_offset,
1445                     vdev->bars[vdev->msix->pba_bar].region.mem,
1446                     vdev->msix->pba_bar, vdev->msix->pba_offset, pos,
1447                     &err);
1448     if (ret < 0) {
1449         if (ret == -ENOTSUP) {
1450             error_report_err(err);
1451             return 0;
1452         }
1453 
1454         error_propagate(errp, err);
1455         return ret;
1456     }
1457 
1458     /*
1459      * The PCI spec suggests that devices provide additional alignment for
1460      * MSI-X structures and avoid overlapping non-MSI-X related registers.
1461      * For an assigned device, this hopefully means that emulation of MSI-X
1462      * structures does not affect the performance of the device.  If devices
1463      * fail to provide that alignment, a significant performance penalty may
1464      * result, for instance Mellanox MT27500 VFs:
1465      * http://www.spinics.net/lists/kvm/msg125881.html
1466      *
1467      * The PBA is simply not that important for such a serious regression and
1468      * most drivers do not appear to look at it.  The solution for this is to
1469      * disable the PBA MemoryRegion unless it's being used.  We disable it
1470      * here and only enable it if a masked vector fires through QEMU.  As the
1471      * vector-use notifier is called, which occurs on unmask, we test whether
1472      * PBA emulation is needed and again disable if not.
1473      */
1474     memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
1475 
1476     return 0;
1477 }
1478 
1479 static void vfio_teardown_msi(VFIOPCIDevice *vdev)
1480 {
1481     msi_uninit(&vdev->pdev);
1482 
1483     if (vdev->msix) {
1484         msix_uninit(&vdev->pdev,
1485                     vdev->bars[vdev->msix->table_bar].region.mem,
1486                     vdev->bars[vdev->msix->pba_bar].region.mem);
1487         g_free(vdev->msix->pending);
1488     }
1489 }
1490 
1491 /*
1492  * Resource setup
1493  */
1494 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled)
1495 {
1496     int i;
1497 
1498     for (i = 0; i < PCI_ROM_SLOT; i++) {
1499         vfio_region_mmaps_set_enabled(&vdev->bars[i].region, enabled);
1500     }
1501 }
1502 
1503 static void vfio_bar_setup(VFIOPCIDevice *vdev, int nr)
1504 {
1505     VFIOBAR *bar = &vdev->bars[nr];
1506 
1507     uint32_t pci_bar;
1508     uint8_t type;
1509     int ret;
1510 
1511     /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1512     if (!bar->region.size) {
1513         return;
1514     }
1515 
1516     /* Determine what type of BAR this is for registration */
1517     ret = pread(vdev->vbasedev.fd, &pci_bar, sizeof(pci_bar),
1518                 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
1519     if (ret != sizeof(pci_bar)) {
1520         error_report("vfio: Failed to read BAR %d (%m)", nr);
1521         return;
1522     }
1523 
1524     pci_bar = le32_to_cpu(pci_bar);
1525     bar->ioport = (pci_bar & PCI_BASE_ADDRESS_SPACE_IO);
1526     bar->mem64 = bar->ioport ? 0 : (pci_bar & PCI_BASE_ADDRESS_MEM_TYPE_64);
1527     type = pci_bar & (bar->ioport ? ~PCI_BASE_ADDRESS_IO_MASK :
1528                                     ~PCI_BASE_ADDRESS_MEM_MASK);
1529 
1530     if (vfio_region_mmap(&bar->region)) {
1531         error_report("Failed to mmap %s BAR %d. Performance may be slow",
1532                      vdev->vbasedev.name, nr);
1533     }
1534 
1535     pci_register_bar(&vdev->pdev, nr, type, bar->region.mem);
1536 }
1537 
1538 static void vfio_bars_setup(VFIOPCIDevice *vdev)
1539 {
1540     int i;
1541 
1542     for (i = 0; i < PCI_ROM_SLOT; i++) {
1543         vfio_bar_setup(vdev, i);
1544     }
1545 }
1546 
1547 static void vfio_bars_exit(VFIOPCIDevice *vdev)
1548 {
1549     int i;
1550 
1551     for (i = 0; i < PCI_ROM_SLOT; i++) {
1552         vfio_bar_quirk_exit(vdev, i);
1553         vfio_region_exit(&vdev->bars[i].region);
1554     }
1555 
1556     if (vdev->vga) {
1557         pci_unregister_vga(&vdev->pdev);
1558         vfio_vga_quirk_exit(vdev);
1559     }
1560 }
1561 
1562 static void vfio_bars_finalize(VFIOPCIDevice *vdev)
1563 {
1564     int i;
1565 
1566     for (i = 0; i < PCI_ROM_SLOT; i++) {
1567         vfio_bar_quirk_finalize(vdev, i);
1568         vfio_region_finalize(&vdev->bars[i].region);
1569     }
1570 
1571     if (vdev->vga) {
1572         vfio_vga_quirk_finalize(vdev);
1573         for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) {
1574             object_unparent(OBJECT(&vdev->vga->region[i].mem));
1575         }
1576         g_free(vdev->vga);
1577     }
1578 }
1579 
1580 /*
1581  * General setup
1582  */
1583 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
1584 {
1585     uint8_t tmp;
1586     uint16_t next = PCI_CONFIG_SPACE_SIZE;
1587 
1588     for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
1589          tmp = pdev->config[tmp + PCI_CAP_LIST_NEXT]) {
1590         if (tmp > pos && tmp < next) {
1591             next = tmp;
1592         }
1593     }
1594 
1595     return next - pos;
1596 }
1597 
1598 
1599 static uint16_t vfio_ext_cap_max_size(const uint8_t *config, uint16_t pos)
1600 {
1601     uint16_t tmp, next = PCIE_CONFIG_SPACE_SIZE;
1602 
1603     for (tmp = PCI_CONFIG_SPACE_SIZE; tmp;
1604         tmp = PCI_EXT_CAP_NEXT(pci_get_long(config + tmp))) {
1605         if (tmp > pos && tmp < next) {
1606             next = tmp;
1607         }
1608     }
1609 
1610     return next - pos;
1611 }
1612 
1613 static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask)
1614 {
1615     pci_set_word(buf, (pci_get_word(buf) & ~mask) | val);
1616 }
1617 
1618 static void vfio_add_emulated_word(VFIOPCIDevice *vdev, int pos,
1619                                    uint16_t val, uint16_t mask)
1620 {
1621     vfio_set_word_bits(vdev->pdev.config + pos, val, mask);
1622     vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask);
1623     vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask);
1624 }
1625 
1626 static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask)
1627 {
1628     pci_set_long(buf, (pci_get_long(buf) & ~mask) | val);
1629 }
1630 
1631 static void vfio_add_emulated_long(VFIOPCIDevice *vdev, int pos,
1632                                    uint32_t val, uint32_t mask)
1633 {
1634     vfio_set_long_bits(vdev->pdev.config + pos, val, mask);
1635     vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask);
1636     vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask);
1637 }
1638 
1639 static int vfio_setup_pcie_cap(VFIOPCIDevice *vdev, int pos, uint8_t size,
1640                                Error **errp)
1641 {
1642     uint16_t flags;
1643     uint8_t type;
1644 
1645     flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS);
1646     type = (flags & PCI_EXP_FLAGS_TYPE) >> 4;
1647 
1648     if (type != PCI_EXP_TYPE_ENDPOINT &&
1649         type != PCI_EXP_TYPE_LEG_END &&
1650         type != PCI_EXP_TYPE_RC_END) {
1651 
1652         error_setg(errp, "assignment of PCIe type 0x%x "
1653                    "devices is not currently supported", type);
1654         return -EINVAL;
1655     }
1656 
1657     if (!pci_bus_is_express(pci_get_bus(&vdev->pdev))) {
1658         PCIBus *bus = pci_get_bus(&vdev->pdev);
1659         PCIDevice *bridge;
1660 
1661         /*
1662          * Traditionally PCI device assignment exposes the PCIe capability
1663          * as-is on non-express buses.  The reason being that some drivers
1664          * simply assume that it's there, for example tg3.  However when
1665          * we're running on a native PCIe machine type, like Q35, we need
1666          * to hide the PCIe capability.  The reason for this is twofold;
1667          * first Windows guests get a Code 10 error when the PCIe capability
1668          * is exposed in this configuration.  Therefore express devices won't
1669          * work at all unless they're attached to express buses in the VM.
1670          * Second, a native PCIe machine introduces the possibility of fine
1671          * granularity IOMMUs supporting both translation and isolation.
1672          * Guest code to discover the IOMMU visibility of a device, such as
1673          * IOMMU grouping code on Linux, is very aware of device types and
1674          * valid transitions between bus types.  An express device on a non-
1675          * express bus is not a valid combination on bare metal systems.
1676          *
1677          * Drivers that require a PCIe capability to make the device
1678          * functional are simply going to need to have their devices placed
1679          * on a PCIe bus in the VM.
1680          */
1681         while (!pci_bus_is_root(bus)) {
1682             bridge = pci_bridge_get_device(bus);
1683             bus = pci_get_bus(bridge);
1684         }
1685 
1686         if (pci_bus_is_express(bus)) {
1687             return 0;
1688         }
1689 
1690     } else if (pci_bus_is_root(pci_get_bus(&vdev->pdev))) {
1691         /*
1692          * On a Root Complex bus Endpoints become Root Complex Integrated
1693          * Endpoints, which changes the type and clears the LNK & LNK2 fields.
1694          */
1695         if (type == PCI_EXP_TYPE_ENDPOINT) {
1696             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1697                                    PCI_EXP_TYPE_RC_END << 4,
1698                                    PCI_EXP_FLAGS_TYPE);
1699 
1700             /* Link Capabilities, Status, and Control goes away */
1701             if (size > PCI_EXP_LNKCTL) {
1702                 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0);
1703                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1704                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0);
1705 
1706 #ifndef PCI_EXP_LNKCAP2
1707 #define PCI_EXP_LNKCAP2 44
1708 #endif
1709 #ifndef PCI_EXP_LNKSTA2
1710 #define PCI_EXP_LNKSTA2 50
1711 #endif
1712                 /* Link 2 Capabilities, Status, and Control goes away */
1713                 if (size > PCI_EXP_LNKCAP2) {
1714                     vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0);
1715                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0);
1716                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0);
1717                 }
1718             }
1719 
1720         } else if (type == PCI_EXP_TYPE_LEG_END) {
1721             /*
1722              * Legacy endpoints don't belong on the root complex.  Windows
1723              * seems to be happier with devices if we skip the capability.
1724              */
1725             return 0;
1726         }
1727 
1728     } else {
1729         /*
1730          * Convert Root Complex Integrated Endpoints to regular endpoints.
1731          * These devices don't support LNK/LNK2 capabilities, so make them up.
1732          */
1733         if (type == PCI_EXP_TYPE_RC_END) {
1734             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1735                                    PCI_EXP_TYPE_ENDPOINT << 4,
1736                                    PCI_EXP_FLAGS_TYPE);
1737             vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP,
1738                                    PCI_EXP_LNK_MLW_1 | PCI_EXP_LNK_LS_25, ~0);
1739             vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1740         }
1741 
1742         /* Mark the Link Status bits as emulated to allow virtual negotiation */
1743         vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA,
1744                                pci_get_word(vdev->pdev.config + pos +
1745                                             PCI_EXP_LNKSTA),
1746                                PCI_EXP_LNKCAP_MLW | PCI_EXP_LNKCAP_SLS);
1747     }
1748 
1749     /*
1750      * Intel 82599 SR-IOV VFs report an invalid PCIe capability version 0
1751      * (Niantic errate #35) causing Windows to error with a Code 10 for the
1752      * device on Q35.  Fixup any such devices to report version 1.  If we
1753      * were to remove the capability entirely the guest would lose extended
1754      * config space.
1755      */
1756     if ((flags & PCI_EXP_FLAGS_VERS) == 0) {
1757         vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1758                                1, PCI_EXP_FLAGS_VERS);
1759     }
1760 
1761     pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size,
1762                              errp);
1763     if (pos < 0) {
1764         return pos;
1765     }
1766 
1767     vdev->pdev.exp.exp_cap = pos;
1768 
1769     return pos;
1770 }
1771 
1772 static void vfio_check_pcie_flr(VFIOPCIDevice *vdev, uint8_t pos)
1773 {
1774     uint32_t cap = pci_get_long(vdev->pdev.config + pos + PCI_EXP_DEVCAP);
1775 
1776     if (cap & PCI_EXP_DEVCAP_FLR) {
1777         trace_vfio_check_pcie_flr(vdev->vbasedev.name);
1778         vdev->has_flr = true;
1779     }
1780 }
1781 
1782 static void vfio_check_pm_reset(VFIOPCIDevice *vdev, uint8_t pos)
1783 {
1784     uint16_t csr = pci_get_word(vdev->pdev.config + pos + PCI_PM_CTRL);
1785 
1786     if (!(csr & PCI_PM_CTRL_NO_SOFT_RESET)) {
1787         trace_vfio_check_pm_reset(vdev->vbasedev.name);
1788         vdev->has_pm_reset = true;
1789     }
1790 }
1791 
1792 static void vfio_check_af_flr(VFIOPCIDevice *vdev, uint8_t pos)
1793 {
1794     uint8_t cap = pci_get_byte(vdev->pdev.config + pos + PCI_AF_CAP);
1795 
1796     if ((cap & PCI_AF_CAP_TP) && (cap & PCI_AF_CAP_FLR)) {
1797         trace_vfio_check_af_flr(vdev->vbasedev.name);
1798         vdev->has_flr = true;
1799     }
1800 }
1801 
1802 static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos, Error **errp)
1803 {
1804     PCIDevice *pdev = &vdev->pdev;
1805     uint8_t cap_id, next, size;
1806     int ret;
1807 
1808     cap_id = pdev->config[pos];
1809     next = pdev->config[pos + PCI_CAP_LIST_NEXT];
1810 
1811     /*
1812      * If it becomes important to configure capabilities to their actual
1813      * size, use this as the default when it's something we don't recognize.
1814      * Since QEMU doesn't actually handle many of the config accesses,
1815      * exact size doesn't seem worthwhile.
1816      */
1817     size = vfio_std_cap_max_size(pdev, pos);
1818 
1819     /*
1820      * pci_add_capability always inserts the new capability at the head
1821      * of the chain.  Therefore to end up with a chain that matches the
1822      * physical device, we insert from the end by making this recursive.
1823      * This is also why we pre-calculate size above as cached config space
1824      * will be changed as we unwind the stack.
1825      */
1826     if (next) {
1827         ret = vfio_add_std_cap(vdev, next, errp);
1828         if (ret) {
1829             return ret;
1830         }
1831     } else {
1832         /* Begin the rebuild, use QEMU emulated list bits */
1833         pdev->config[PCI_CAPABILITY_LIST] = 0;
1834         vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
1835         vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
1836 
1837         ret = vfio_add_virt_caps(vdev, errp);
1838         if (ret) {
1839             return ret;
1840         }
1841     }
1842 
1843     /* Scale down size, esp in case virt caps were added above */
1844     size = MIN(size, vfio_std_cap_max_size(pdev, pos));
1845 
1846     /* Use emulated next pointer to allow dropping caps */
1847     pci_set_byte(vdev->emulated_config_bits + pos + PCI_CAP_LIST_NEXT, 0xff);
1848 
1849     switch (cap_id) {
1850     case PCI_CAP_ID_MSI:
1851         ret = vfio_msi_setup(vdev, pos, errp);
1852         break;
1853     case PCI_CAP_ID_EXP:
1854         vfio_check_pcie_flr(vdev, pos);
1855         ret = vfio_setup_pcie_cap(vdev, pos, size, errp);
1856         break;
1857     case PCI_CAP_ID_MSIX:
1858         ret = vfio_msix_setup(vdev, pos, errp);
1859         break;
1860     case PCI_CAP_ID_PM:
1861         vfio_check_pm_reset(vdev, pos);
1862         vdev->pm_cap = pos;
1863         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
1864         break;
1865     case PCI_CAP_ID_AF:
1866         vfio_check_af_flr(vdev, pos);
1867         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
1868         break;
1869     default:
1870         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
1871         break;
1872     }
1873 
1874     if (ret < 0) {
1875         error_prepend(errp,
1876                       "failed to add PCI capability 0x%x[0x%x]@0x%x: ",
1877                       cap_id, size, pos);
1878         return ret;
1879     }
1880 
1881     return 0;
1882 }
1883 
1884 static void vfio_add_ext_cap(VFIOPCIDevice *vdev)
1885 {
1886     PCIDevice *pdev = &vdev->pdev;
1887     uint32_t header;
1888     uint16_t cap_id, next, size;
1889     uint8_t cap_ver;
1890     uint8_t *config;
1891 
1892     /* Only add extended caps if we have them and the guest can see them */
1893     if (!pci_is_express(pdev) || !pci_bus_is_express(pci_get_bus(pdev)) ||
1894         !pci_get_long(pdev->config + PCI_CONFIG_SPACE_SIZE)) {
1895         return;
1896     }
1897 
1898     /*
1899      * pcie_add_capability always inserts the new capability at the tail
1900      * of the chain.  Therefore to end up with a chain that matches the
1901      * physical device, we cache the config space to avoid overwriting
1902      * the original config space when we parse the extended capabilities.
1903      */
1904     config = g_memdup(pdev->config, vdev->config_size);
1905 
1906     /*
1907      * Extended capabilities are chained with each pointing to the next, so we
1908      * can drop anything other than the head of the chain simply by modifying
1909      * the previous next pointer.  Seed the head of the chain here such that
1910      * we can simply skip any capabilities we want to drop below, regardless
1911      * of their position in the chain.  If this stub capability still exists
1912      * after we add the capabilities we want to expose, update the capability
1913      * ID to zero.  Note that we cannot seed with the capability header being
1914      * zero as this conflicts with definition of an absent capability chain
1915      * and prevents capabilities beyond the head of the list from being added.
1916      * By replacing the dummy capability ID with zero after walking the device
1917      * chain, we also transparently mark extended capabilities as absent if
1918      * no capabilities were added.  Note that the PCIe spec defines an absence
1919      * of extended capabilities to be determined by a value of zero for the
1920      * capability ID, version, AND next pointer.  A non-zero next pointer
1921      * should be sufficient to indicate additional capabilities are present,
1922      * which will occur if we call pcie_add_capability() below.  The entire
1923      * first dword is emulated to support this.
1924      *
1925      * NB. The kernel side does similar masking, so be prepared that our
1926      * view of the device may also contain a capability ID zero in the head
1927      * of the chain.  Skip it for the same reason that we cannot seed the
1928      * chain with a zero capability.
1929      */
1930     pci_set_long(pdev->config + PCI_CONFIG_SPACE_SIZE,
1931                  PCI_EXT_CAP(0xFFFF, 0, 0));
1932     pci_set_long(pdev->wmask + PCI_CONFIG_SPACE_SIZE, 0);
1933     pci_set_long(vdev->emulated_config_bits + PCI_CONFIG_SPACE_SIZE, ~0);
1934 
1935     for (next = PCI_CONFIG_SPACE_SIZE; next;
1936          next = PCI_EXT_CAP_NEXT(pci_get_long(config + next))) {
1937         header = pci_get_long(config + next);
1938         cap_id = PCI_EXT_CAP_ID(header);
1939         cap_ver = PCI_EXT_CAP_VER(header);
1940 
1941         /*
1942          * If it becomes important to configure extended capabilities to their
1943          * actual size, use this as the default when it's something we don't
1944          * recognize. Since QEMU doesn't actually handle many of the config
1945          * accesses, exact size doesn't seem worthwhile.
1946          */
1947         size = vfio_ext_cap_max_size(config, next);
1948 
1949         /* Use emulated next pointer to allow dropping extended caps */
1950         pci_long_test_and_set_mask(vdev->emulated_config_bits + next,
1951                                    PCI_EXT_CAP_NEXT_MASK);
1952 
1953         switch (cap_id) {
1954         case 0: /* kernel masked capability */
1955         case PCI_EXT_CAP_ID_SRIOV: /* Read-only VF BARs confuse OVMF */
1956         case PCI_EXT_CAP_ID_ARI: /* XXX Needs next function virtualization */
1957             trace_vfio_add_ext_cap_dropped(vdev->vbasedev.name, cap_id, next);
1958             break;
1959         default:
1960             pcie_add_capability(pdev, cap_id, cap_ver, next, size);
1961         }
1962 
1963     }
1964 
1965     /* Cleanup chain head ID if necessary */
1966     if (pci_get_word(pdev->config + PCI_CONFIG_SPACE_SIZE) == 0xFFFF) {
1967         pci_set_word(pdev->config + PCI_CONFIG_SPACE_SIZE, 0);
1968     }
1969 
1970     g_free(config);
1971     return;
1972 }
1973 
1974 static int vfio_add_capabilities(VFIOPCIDevice *vdev, Error **errp)
1975 {
1976     PCIDevice *pdev = &vdev->pdev;
1977     int ret;
1978 
1979     if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
1980         !pdev->config[PCI_CAPABILITY_LIST]) {
1981         return 0; /* Nothing to add */
1982     }
1983 
1984     ret = vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST], errp);
1985     if (ret) {
1986         return ret;
1987     }
1988 
1989     vfio_add_ext_cap(vdev);
1990     return 0;
1991 }
1992 
1993 static void vfio_pci_pre_reset(VFIOPCIDevice *vdev)
1994 {
1995     PCIDevice *pdev = &vdev->pdev;
1996     uint16_t cmd;
1997 
1998     vfio_disable_interrupts(vdev);
1999 
2000     /* Make sure the device is in D0 */
2001     if (vdev->pm_cap) {
2002         uint16_t pmcsr;
2003         uint8_t state;
2004 
2005         pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2006         state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2007         if (state) {
2008             pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
2009             vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2);
2010             /* vfio handles the necessary delay here */
2011             pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2012             state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2013             if (state) {
2014                 error_report("vfio: Unable to power on device, stuck in D%d",
2015                              state);
2016             }
2017         }
2018     }
2019 
2020     /*
2021      * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
2022      * Also put INTx Disable in known state.
2023      */
2024     cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
2025     cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
2026              PCI_COMMAND_INTX_DISABLE);
2027     vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
2028 }
2029 
2030 static void vfio_pci_post_reset(VFIOPCIDevice *vdev)
2031 {
2032     Error *err = NULL;
2033     int nr;
2034 
2035     vfio_intx_enable(vdev, &err);
2036     if (err) {
2037         error_reportf_err(err, ERR_PREFIX, vdev->vbasedev.name);
2038     }
2039 
2040     for (nr = 0; nr < PCI_NUM_REGIONS - 1; ++nr) {
2041         off_t addr = vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr);
2042         uint32_t val = 0;
2043         uint32_t len = sizeof(val);
2044 
2045         if (pwrite(vdev->vbasedev.fd, &val, len, addr) != len) {
2046             error_report("%s(%s) reset bar %d failed: %m", __func__,
2047                          vdev->vbasedev.name, nr);
2048         }
2049     }
2050 }
2051 
2052 static bool vfio_pci_host_match(PCIHostDeviceAddress *addr, const char *name)
2053 {
2054     char tmp[13];
2055 
2056     sprintf(tmp, "%04x:%02x:%02x.%1x", addr->domain,
2057             addr->bus, addr->slot, addr->function);
2058 
2059     return (strcmp(tmp, name) == 0);
2060 }
2061 
2062 static int vfio_pci_hot_reset(VFIOPCIDevice *vdev, bool single)
2063 {
2064     VFIOGroup *group;
2065     struct vfio_pci_hot_reset_info *info;
2066     struct vfio_pci_dependent_device *devices;
2067     struct vfio_pci_hot_reset *reset;
2068     int32_t *fds;
2069     int ret, i, count;
2070     bool multi = false;
2071 
2072     trace_vfio_pci_hot_reset(vdev->vbasedev.name, single ? "one" : "multi");
2073 
2074     if (!single) {
2075         vfio_pci_pre_reset(vdev);
2076     }
2077     vdev->vbasedev.needs_reset = false;
2078 
2079     info = g_malloc0(sizeof(*info));
2080     info->argsz = sizeof(*info);
2081 
2082     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2083     if (ret && errno != ENOSPC) {
2084         ret = -errno;
2085         if (!vdev->has_pm_reset) {
2086             error_report("vfio: Cannot reset device %s, "
2087                          "no available reset mechanism.", vdev->vbasedev.name);
2088         }
2089         goto out_single;
2090     }
2091 
2092     count = info->count;
2093     info = g_realloc(info, sizeof(*info) + (count * sizeof(*devices)));
2094     info->argsz = sizeof(*info) + (count * sizeof(*devices));
2095     devices = &info->devices[0];
2096 
2097     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2098     if (ret) {
2099         ret = -errno;
2100         error_report("vfio: hot reset info failed: %m");
2101         goto out_single;
2102     }
2103 
2104     trace_vfio_pci_hot_reset_has_dep_devices(vdev->vbasedev.name);
2105 
2106     /* Verify that we have all the groups required */
2107     for (i = 0; i < info->count; i++) {
2108         PCIHostDeviceAddress host;
2109         VFIOPCIDevice *tmp;
2110         VFIODevice *vbasedev_iter;
2111 
2112         host.domain = devices[i].segment;
2113         host.bus = devices[i].bus;
2114         host.slot = PCI_SLOT(devices[i].devfn);
2115         host.function = PCI_FUNC(devices[i].devfn);
2116 
2117         trace_vfio_pci_hot_reset_dep_devices(host.domain,
2118                 host.bus, host.slot, host.function, devices[i].group_id);
2119 
2120         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2121             continue;
2122         }
2123 
2124         QLIST_FOREACH(group, &vfio_group_list, next) {
2125             if (group->groupid == devices[i].group_id) {
2126                 break;
2127             }
2128         }
2129 
2130         if (!group) {
2131             if (!vdev->has_pm_reset) {
2132                 error_report("vfio: Cannot reset device %s, "
2133                              "depends on group %d which is not owned.",
2134                              vdev->vbasedev.name, devices[i].group_id);
2135             }
2136             ret = -EPERM;
2137             goto out;
2138         }
2139 
2140         /* Prep dependent devices for reset and clear our marker. */
2141         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2142             if (!vbasedev_iter->dev->realized ||
2143                 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2144                 continue;
2145             }
2146             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2147             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2148                 if (single) {
2149                     ret = -EINVAL;
2150                     goto out_single;
2151                 }
2152                 vfio_pci_pre_reset(tmp);
2153                 tmp->vbasedev.needs_reset = false;
2154                 multi = true;
2155                 break;
2156             }
2157         }
2158     }
2159 
2160     if (!single && !multi) {
2161         ret = -EINVAL;
2162         goto out_single;
2163     }
2164 
2165     /* Determine how many group fds need to be passed */
2166     count = 0;
2167     QLIST_FOREACH(group, &vfio_group_list, next) {
2168         for (i = 0; i < info->count; i++) {
2169             if (group->groupid == devices[i].group_id) {
2170                 count++;
2171                 break;
2172             }
2173         }
2174     }
2175 
2176     reset = g_malloc0(sizeof(*reset) + (count * sizeof(*fds)));
2177     reset->argsz = sizeof(*reset) + (count * sizeof(*fds));
2178     fds = &reset->group_fds[0];
2179 
2180     /* Fill in group fds */
2181     QLIST_FOREACH(group, &vfio_group_list, next) {
2182         for (i = 0; i < info->count; i++) {
2183             if (group->groupid == devices[i].group_id) {
2184                 fds[reset->count++] = group->fd;
2185                 break;
2186             }
2187         }
2188     }
2189 
2190     /* Bus reset! */
2191     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_PCI_HOT_RESET, reset);
2192     g_free(reset);
2193 
2194     trace_vfio_pci_hot_reset_result(vdev->vbasedev.name,
2195                                     ret ? "%m" : "Success");
2196 
2197 out:
2198     /* Re-enable INTx on affected devices */
2199     for (i = 0; i < info->count; i++) {
2200         PCIHostDeviceAddress host;
2201         VFIOPCIDevice *tmp;
2202         VFIODevice *vbasedev_iter;
2203 
2204         host.domain = devices[i].segment;
2205         host.bus = devices[i].bus;
2206         host.slot = PCI_SLOT(devices[i].devfn);
2207         host.function = PCI_FUNC(devices[i].devfn);
2208 
2209         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2210             continue;
2211         }
2212 
2213         QLIST_FOREACH(group, &vfio_group_list, next) {
2214             if (group->groupid == devices[i].group_id) {
2215                 break;
2216             }
2217         }
2218 
2219         if (!group) {
2220             break;
2221         }
2222 
2223         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2224             if (!vbasedev_iter->dev->realized ||
2225                 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2226                 continue;
2227             }
2228             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2229             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2230                 vfio_pci_post_reset(tmp);
2231                 break;
2232             }
2233         }
2234     }
2235 out_single:
2236     if (!single) {
2237         vfio_pci_post_reset(vdev);
2238     }
2239     g_free(info);
2240 
2241     return ret;
2242 }
2243 
2244 /*
2245  * We want to differentiate hot reset of mulitple in-use devices vs hot reset
2246  * of a single in-use device.  VFIO_DEVICE_RESET will already handle the case
2247  * of doing hot resets when there is only a single device per bus.  The in-use
2248  * here refers to how many VFIODevices are affected.  A hot reset that affects
2249  * multiple devices, but only a single in-use device, means that we can call
2250  * it from our bus ->reset() callback since the extent is effectively a single
2251  * device.  This allows us to make use of it in the hotplug path.  When there
2252  * are multiple in-use devices, we can only trigger the hot reset during a
2253  * system reset and thus from our reset handler.  We separate _one vs _multi
2254  * here so that we don't overlap and do a double reset on the system reset
2255  * path where both our reset handler and ->reset() callback are used.  Calling
2256  * _one() will only do a hot reset for the one in-use devices case, calling
2257  * _multi() will do nothing if a _one() would have been sufficient.
2258  */
2259 static int vfio_pci_hot_reset_one(VFIOPCIDevice *vdev)
2260 {
2261     return vfio_pci_hot_reset(vdev, true);
2262 }
2263 
2264 static int vfio_pci_hot_reset_multi(VFIODevice *vbasedev)
2265 {
2266     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2267     return vfio_pci_hot_reset(vdev, false);
2268 }
2269 
2270 static void vfio_pci_compute_needs_reset(VFIODevice *vbasedev)
2271 {
2272     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2273     if (!vbasedev->reset_works || (!vdev->has_flr && vdev->has_pm_reset)) {
2274         vbasedev->needs_reset = true;
2275     }
2276 }
2277 
2278 static VFIODeviceOps vfio_pci_ops = {
2279     .vfio_compute_needs_reset = vfio_pci_compute_needs_reset,
2280     .vfio_hot_reset_multi = vfio_pci_hot_reset_multi,
2281     .vfio_eoi = vfio_intx_eoi,
2282 };
2283 
2284 int vfio_populate_vga(VFIOPCIDevice *vdev, Error **errp)
2285 {
2286     VFIODevice *vbasedev = &vdev->vbasedev;
2287     struct vfio_region_info *reg_info;
2288     int ret;
2289 
2290     ret = vfio_get_region_info(vbasedev, VFIO_PCI_VGA_REGION_INDEX, &reg_info);
2291     if (ret) {
2292         error_setg_errno(errp, -ret,
2293                          "failed getting region info for VGA region index %d",
2294                          VFIO_PCI_VGA_REGION_INDEX);
2295         return ret;
2296     }
2297 
2298     if (!(reg_info->flags & VFIO_REGION_INFO_FLAG_READ) ||
2299         !(reg_info->flags & VFIO_REGION_INFO_FLAG_WRITE) ||
2300         reg_info->size < 0xbffff + 1) {
2301         error_setg(errp, "unexpected VGA info, flags 0x%lx, size 0x%lx",
2302                    (unsigned long)reg_info->flags,
2303                    (unsigned long)reg_info->size);
2304         g_free(reg_info);
2305         return -EINVAL;
2306     }
2307 
2308     vdev->vga = g_new0(VFIOVGA, 1);
2309 
2310     vdev->vga->fd_offset = reg_info->offset;
2311     vdev->vga->fd = vdev->vbasedev.fd;
2312 
2313     g_free(reg_info);
2314 
2315     vdev->vga->region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE;
2316     vdev->vga->region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM;
2317     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_MEM].quirks);
2318 
2319     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2320                           OBJECT(vdev), &vfio_vga_ops,
2321                           &vdev->vga->region[QEMU_PCI_VGA_MEM],
2322                           "vfio-vga-mmio@0xa0000",
2323                           QEMU_PCI_VGA_MEM_SIZE);
2324 
2325     vdev->vga->region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE;
2326     vdev->vga->region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO;
2327     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].quirks);
2328 
2329     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2330                           OBJECT(vdev), &vfio_vga_ops,
2331                           &vdev->vga->region[QEMU_PCI_VGA_IO_LO],
2332                           "vfio-vga-io@0x3b0",
2333                           QEMU_PCI_VGA_IO_LO_SIZE);
2334 
2335     vdev->vga->region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE;
2336     vdev->vga->region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI;
2337     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks);
2338 
2339     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
2340                           OBJECT(vdev), &vfio_vga_ops,
2341                           &vdev->vga->region[QEMU_PCI_VGA_IO_HI],
2342                           "vfio-vga-io@0x3c0",
2343                           QEMU_PCI_VGA_IO_HI_SIZE);
2344 
2345     pci_register_vga(&vdev->pdev, &vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2346                      &vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2347                      &vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem);
2348 
2349     return 0;
2350 }
2351 
2352 static void vfio_populate_device(VFIOPCIDevice *vdev, Error **errp)
2353 {
2354     VFIODevice *vbasedev = &vdev->vbasedev;
2355     struct vfio_region_info *reg_info;
2356     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) };
2357     int i, ret = -1;
2358 
2359     /* Sanity check device */
2360     if (!(vbasedev->flags & VFIO_DEVICE_FLAGS_PCI)) {
2361         error_setg(errp, "this isn't a PCI device");
2362         return;
2363     }
2364 
2365     if (vbasedev->num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
2366         error_setg(errp, "unexpected number of io regions %u",
2367                    vbasedev->num_regions);
2368         return;
2369     }
2370 
2371     if (vbasedev->num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
2372         error_setg(errp, "unexpected number of irqs %u", vbasedev->num_irqs);
2373         return;
2374     }
2375 
2376     for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
2377         char *name = g_strdup_printf("%s BAR %d", vbasedev->name, i);
2378 
2379         ret = vfio_region_setup(OBJECT(vdev), vbasedev,
2380                                 &vdev->bars[i].region, i, name);
2381         g_free(name);
2382 
2383         if (ret) {
2384             error_setg_errno(errp, -ret, "failed to get region %d info", i);
2385             return;
2386         }
2387 
2388         QLIST_INIT(&vdev->bars[i].quirks);
2389     }
2390 
2391     ret = vfio_get_region_info(vbasedev,
2392                                VFIO_PCI_CONFIG_REGION_INDEX, &reg_info);
2393     if (ret) {
2394         error_setg_errno(errp, -ret, "failed to get config info");
2395         return;
2396     }
2397 
2398     trace_vfio_populate_device_config(vdev->vbasedev.name,
2399                                       (unsigned long)reg_info->size,
2400                                       (unsigned long)reg_info->offset,
2401                                       (unsigned long)reg_info->flags);
2402 
2403     vdev->config_size = reg_info->size;
2404     if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
2405         vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
2406     }
2407     vdev->config_offset = reg_info->offset;
2408 
2409     g_free(reg_info);
2410 
2411     if (vdev->features & VFIO_FEATURE_ENABLE_VGA) {
2412         ret = vfio_populate_vga(vdev, errp);
2413         if (ret) {
2414             error_append_hint(errp, "device does not support "
2415                               "requested feature x-vga\n");
2416             return;
2417         }
2418     }
2419 
2420     irq_info.index = VFIO_PCI_ERR_IRQ_INDEX;
2421 
2422     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info);
2423     if (ret) {
2424         /* This can fail for an old kernel or legacy PCI dev */
2425         trace_vfio_populate_device_get_irq_info_failure();
2426     } else if (irq_info.count == 1) {
2427         vdev->pci_aer = true;
2428     } else {
2429         error_report(WARN_PREFIX
2430                      "Could not enable error recovery for the device",
2431                      vbasedev->name);
2432     }
2433 }
2434 
2435 static void vfio_put_device(VFIOPCIDevice *vdev)
2436 {
2437     g_free(vdev->vbasedev.name);
2438     g_free(vdev->msix);
2439 
2440     vfio_put_base_device(&vdev->vbasedev);
2441 }
2442 
2443 static void vfio_err_notifier_handler(void *opaque)
2444 {
2445     VFIOPCIDevice *vdev = opaque;
2446 
2447     if (!event_notifier_test_and_clear(&vdev->err_notifier)) {
2448         return;
2449     }
2450 
2451     /*
2452      * TBD. Retrieve the error details and decide what action
2453      * needs to be taken. One of the actions could be to pass
2454      * the error to the guest and have the guest driver recover
2455      * from the error. This requires that PCIe capabilities be
2456      * exposed to the guest. For now, we just terminate the
2457      * guest to contain the error.
2458      */
2459 
2460     error_report("%s(%s) Unrecoverable error detected. Please collect any data possible and then kill the guest", __func__, vdev->vbasedev.name);
2461 
2462     vm_stop(RUN_STATE_INTERNAL_ERROR);
2463 }
2464 
2465 /*
2466  * Registers error notifier for devices supporting error recovery.
2467  * If we encounter a failure in this function, we report an error
2468  * and continue after disabling error recovery support for the
2469  * device.
2470  */
2471 static void vfio_register_err_notifier(VFIOPCIDevice *vdev)
2472 {
2473     int ret;
2474     int argsz;
2475     struct vfio_irq_set *irq_set;
2476     int32_t *pfd;
2477 
2478     if (!vdev->pci_aer) {
2479         return;
2480     }
2481 
2482     if (event_notifier_init(&vdev->err_notifier, 0)) {
2483         error_report("vfio: Unable to init event notifier for error detection");
2484         vdev->pci_aer = false;
2485         return;
2486     }
2487 
2488     argsz = sizeof(*irq_set) + sizeof(*pfd);
2489 
2490     irq_set = g_malloc0(argsz);
2491     irq_set->argsz = argsz;
2492     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2493                      VFIO_IRQ_SET_ACTION_TRIGGER;
2494     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
2495     irq_set->start = 0;
2496     irq_set->count = 1;
2497     pfd = (int32_t *)&irq_set->data;
2498 
2499     *pfd = event_notifier_get_fd(&vdev->err_notifier);
2500     qemu_set_fd_handler(*pfd, vfio_err_notifier_handler, NULL, vdev);
2501 
2502     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
2503     if (ret) {
2504         error_report("vfio: Failed to set up error notification");
2505         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
2506         event_notifier_cleanup(&vdev->err_notifier);
2507         vdev->pci_aer = false;
2508     }
2509     g_free(irq_set);
2510 }
2511 
2512 static void vfio_unregister_err_notifier(VFIOPCIDevice *vdev)
2513 {
2514     int argsz;
2515     struct vfio_irq_set *irq_set;
2516     int32_t *pfd;
2517     int ret;
2518 
2519     if (!vdev->pci_aer) {
2520         return;
2521     }
2522 
2523     argsz = sizeof(*irq_set) + sizeof(*pfd);
2524 
2525     irq_set = g_malloc0(argsz);
2526     irq_set->argsz = argsz;
2527     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2528                      VFIO_IRQ_SET_ACTION_TRIGGER;
2529     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
2530     irq_set->start = 0;
2531     irq_set->count = 1;
2532     pfd = (int32_t *)&irq_set->data;
2533     *pfd = -1;
2534 
2535     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
2536     if (ret) {
2537         error_report("vfio: Failed to de-assign error fd: %m");
2538     }
2539     g_free(irq_set);
2540     qemu_set_fd_handler(event_notifier_get_fd(&vdev->err_notifier),
2541                         NULL, NULL, vdev);
2542     event_notifier_cleanup(&vdev->err_notifier);
2543 }
2544 
2545 static void vfio_req_notifier_handler(void *opaque)
2546 {
2547     VFIOPCIDevice *vdev = opaque;
2548     Error *err = NULL;
2549 
2550     if (!event_notifier_test_and_clear(&vdev->req_notifier)) {
2551         return;
2552     }
2553 
2554     qdev_unplug(&vdev->pdev.qdev, &err);
2555     if (err) {
2556         error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name);
2557     }
2558 }
2559 
2560 static void vfio_register_req_notifier(VFIOPCIDevice *vdev)
2561 {
2562     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info),
2563                                       .index = VFIO_PCI_REQ_IRQ_INDEX };
2564     int argsz;
2565     struct vfio_irq_set *irq_set;
2566     int32_t *pfd;
2567 
2568     if (!(vdev->features & VFIO_FEATURE_ENABLE_REQ)) {
2569         return;
2570     }
2571 
2572     if (ioctl(vdev->vbasedev.fd,
2573               VFIO_DEVICE_GET_IRQ_INFO, &irq_info) < 0 || irq_info.count < 1) {
2574         return;
2575     }
2576 
2577     if (event_notifier_init(&vdev->req_notifier, 0)) {
2578         error_report("vfio: Unable to init event notifier for device request");
2579         return;
2580     }
2581 
2582     argsz = sizeof(*irq_set) + sizeof(*pfd);
2583 
2584     irq_set = g_malloc0(argsz);
2585     irq_set->argsz = argsz;
2586     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2587                      VFIO_IRQ_SET_ACTION_TRIGGER;
2588     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
2589     irq_set->start = 0;
2590     irq_set->count = 1;
2591     pfd = (int32_t *)&irq_set->data;
2592 
2593     *pfd = event_notifier_get_fd(&vdev->req_notifier);
2594     qemu_set_fd_handler(*pfd, vfio_req_notifier_handler, NULL, vdev);
2595 
2596     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
2597         error_report("vfio: Failed to set up device request notification");
2598         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
2599         event_notifier_cleanup(&vdev->req_notifier);
2600     } else {
2601         vdev->req_enabled = true;
2602     }
2603 
2604     g_free(irq_set);
2605 }
2606 
2607 static void vfio_unregister_req_notifier(VFIOPCIDevice *vdev)
2608 {
2609     int argsz;
2610     struct vfio_irq_set *irq_set;
2611     int32_t *pfd;
2612 
2613     if (!vdev->req_enabled) {
2614         return;
2615     }
2616 
2617     argsz = sizeof(*irq_set) + sizeof(*pfd);
2618 
2619     irq_set = g_malloc0(argsz);
2620     irq_set->argsz = argsz;
2621     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2622                      VFIO_IRQ_SET_ACTION_TRIGGER;
2623     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
2624     irq_set->start = 0;
2625     irq_set->count = 1;
2626     pfd = (int32_t *)&irq_set->data;
2627     *pfd = -1;
2628 
2629     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
2630         error_report("vfio: Failed to de-assign device request fd: %m");
2631     }
2632     g_free(irq_set);
2633     qemu_set_fd_handler(event_notifier_get_fd(&vdev->req_notifier),
2634                         NULL, NULL, vdev);
2635     event_notifier_cleanup(&vdev->req_notifier);
2636 
2637     vdev->req_enabled = false;
2638 }
2639 
2640 static void vfio_realize(PCIDevice *pdev, Error **errp)
2641 {
2642     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
2643     VFIODevice *vbasedev_iter;
2644     VFIOGroup *group;
2645     char *tmp, group_path[PATH_MAX], *group_name;
2646     Error *err = NULL;
2647     ssize_t len;
2648     struct stat st;
2649     int groupid;
2650     int i, ret;
2651 
2652     if (!vdev->vbasedev.sysfsdev) {
2653         if (!(~vdev->host.domain || ~vdev->host.bus ||
2654               ~vdev->host.slot || ~vdev->host.function)) {
2655             error_setg(errp, "No provided host device");
2656             error_append_hint(errp, "Use -device vfio-pci,host=DDDD:BB:DD.F "
2657                               "or -device vfio-pci,sysfsdev=PATH_TO_DEVICE\n");
2658             return;
2659         }
2660         vdev->vbasedev.sysfsdev =
2661             g_strdup_printf("/sys/bus/pci/devices/%04x:%02x:%02x.%01x",
2662                             vdev->host.domain, vdev->host.bus,
2663                             vdev->host.slot, vdev->host.function);
2664     }
2665 
2666     if (stat(vdev->vbasedev.sysfsdev, &st) < 0) {
2667         error_setg_errno(errp, errno, "no such host device");
2668         error_prepend(errp, ERR_PREFIX, vdev->vbasedev.sysfsdev);
2669         return;
2670     }
2671 
2672     vdev->vbasedev.name = g_strdup(basename(vdev->vbasedev.sysfsdev));
2673     vdev->vbasedev.ops = &vfio_pci_ops;
2674     vdev->vbasedev.type = VFIO_DEVICE_TYPE_PCI;
2675     vdev->vbasedev.dev = &vdev->pdev.qdev;
2676 
2677     tmp = g_strdup_printf("%s/iommu_group", vdev->vbasedev.sysfsdev);
2678     len = readlink(tmp, group_path, sizeof(group_path));
2679     g_free(tmp);
2680 
2681     if (len <= 0 || len >= sizeof(group_path)) {
2682         error_setg_errno(errp, len < 0 ? errno : ENAMETOOLONG,
2683                          "no iommu_group found");
2684         goto error;
2685     }
2686 
2687     group_path[len] = 0;
2688 
2689     group_name = basename(group_path);
2690     if (sscanf(group_name, "%d", &groupid) != 1) {
2691         error_setg_errno(errp, errno, "failed to read %s", group_path);
2692         goto error;
2693     }
2694 
2695     trace_vfio_realize(vdev->vbasedev.name, groupid);
2696 
2697     group = vfio_get_group(groupid, pci_device_iommu_address_space(pdev), errp);
2698     if (!group) {
2699         goto error;
2700     }
2701 
2702     QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2703         if (strcmp(vbasedev_iter->name, vdev->vbasedev.name) == 0) {
2704             error_setg(errp, "device is already attached");
2705             vfio_put_group(group);
2706             goto error;
2707         }
2708     }
2709 
2710     ret = vfio_get_device(group, vdev->vbasedev.name, &vdev->vbasedev, errp);
2711     if (ret) {
2712         vfio_put_group(group);
2713         goto error;
2714     }
2715 
2716     vfio_populate_device(vdev, &err);
2717     if (err) {
2718         error_propagate(errp, err);
2719         goto error;
2720     }
2721 
2722     /* Get a copy of config space */
2723     ret = pread(vdev->vbasedev.fd, vdev->pdev.config,
2724                 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
2725                 vdev->config_offset);
2726     if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
2727         ret = ret < 0 ? -errno : -EFAULT;
2728         error_setg_errno(errp, -ret, "failed to read device config space");
2729         goto error;
2730     }
2731 
2732     /* vfio emulates a lot for us, but some bits need extra love */
2733     vdev->emulated_config_bits = g_malloc0(vdev->config_size);
2734 
2735     /* QEMU can choose to expose the ROM or not */
2736     memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4);
2737 
2738     /*
2739      * The PCI spec reserves vendor ID 0xffff as an invalid value.  The
2740      * device ID is managed by the vendor and need only be a 16-bit value.
2741      * Allow any 16-bit value for subsystem so they can be hidden or changed.
2742      */
2743     if (vdev->vendor_id != PCI_ANY_ID) {
2744         if (vdev->vendor_id >= 0xffff) {
2745             error_setg(errp, "invalid PCI vendor ID provided");
2746             goto error;
2747         }
2748         vfio_add_emulated_word(vdev, PCI_VENDOR_ID, vdev->vendor_id, ~0);
2749         trace_vfio_pci_emulated_vendor_id(vdev->vbasedev.name, vdev->vendor_id);
2750     } else {
2751         vdev->vendor_id = pci_get_word(pdev->config + PCI_VENDOR_ID);
2752     }
2753 
2754     if (vdev->device_id != PCI_ANY_ID) {
2755         if (vdev->device_id > 0xffff) {
2756             error_setg(errp, "invalid PCI device ID provided");
2757             goto error;
2758         }
2759         vfio_add_emulated_word(vdev, PCI_DEVICE_ID, vdev->device_id, ~0);
2760         trace_vfio_pci_emulated_device_id(vdev->vbasedev.name, vdev->device_id);
2761     } else {
2762         vdev->device_id = pci_get_word(pdev->config + PCI_DEVICE_ID);
2763     }
2764 
2765     if (vdev->sub_vendor_id != PCI_ANY_ID) {
2766         if (vdev->sub_vendor_id > 0xffff) {
2767             error_setg(errp, "invalid PCI subsystem vendor ID provided");
2768             goto error;
2769         }
2770         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_VENDOR_ID,
2771                                vdev->sub_vendor_id, ~0);
2772         trace_vfio_pci_emulated_sub_vendor_id(vdev->vbasedev.name,
2773                                               vdev->sub_vendor_id);
2774     }
2775 
2776     if (vdev->sub_device_id != PCI_ANY_ID) {
2777         if (vdev->sub_device_id > 0xffff) {
2778             error_setg(errp, "invalid PCI subsystem device ID provided");
2779             goto error;
2780         }
2781         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_ID, vdev->sub_device_id, ~0);
2782         trace_vfio_pci_emulated_sub_device_id(vdev->vbasedev.name,
2783                                               vdev->sub_device_id);
2784     }
2785 
2786     /* QEMU can change multi-function devices to single function, or reverse */
2787     vdev->emulated_config_bits[PCI_HEADER_TYPE] =
2788                                               PCI_HEADER_TYPE_MULTI_FUNCTION;
2789 
2790     /* Restore or clear multifunction, this is always controlled by QEMU */
2791     if (vdev->pdev.cap_present & QEMU_PCI_CAP_MULTIFUNCTION) {
2792         vdev->pdev.config[PCI_HEADER_TYPE] |= PCI_HEADER_TYPE_MULTI_FUNCTION;
2793     } else {
2794         vdev->pdev.config[PCI_HEADER_TYPE] &= ~PCI_HEADER_TYPE_MULTI_FUNCTION;
2795     }
2796 
2797     /*
2798      * Clear host resource mapping info.  If we choose not to register a
2799      * BAR, such as might be the case with the option ROM, we can get
2800      * confusing, unwritable, residual addresses from the host here.
2801      */
2802     memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
2803     memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
2804 
2805     vfio_pci_size_rom(vdev);
2806 
2807     vfio_msix_early_setup(vdev, &err);
2808     if (err) {
2809         error_propagate(errp, err);
2810         goto error;
2811     }
2812 
2813     vfio_bars_setup(vdev);
2814 
2815     ret = vfio_add_capabilities(vdev, errp);
2816     if (ret) {
2817         goto out_teardown;
2818     }
2819 
2820     if (vdev->vga) {
2821         vfio_vga_quirk_setup(vdev);
2822     }
2823 
2824     for (i = 0; i < PCI_ROM_SLOT; i++) {
2825         vfio_bar_quirk_setup(vdev, i);
2826     }
2827 
2828     if (!vdev->igd_opregion &&
2829         vdev->features & VFIO_FEATURE_ENABLE_IGD_OPREGION) {
2830         struct vfio_region_info *opregion;
2831 
2832         if (vdev->pdev.qdev.hotplugged) {
2833             error_setg(errp,
2834                        "cannot support IGD OpRegion feature on hotplugged "
2835                        "device");
2836             goto out_teardown;
2837         }
2838 
2839         ret = vfio_get_dev_region_info(&vdev->vbasedev,
2840                         VFIO_REGION_TYPE_PCI_VENDOR_TYPE | PCI_VENDOR_ID_INTEL,
2841                         VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION, &opregion);
2842         if (ret) {
2843             error_setg_errno(errp, -ret,
2844                              "does not support requested IGD OpRegion feature");
2845             goto out_teardown;
2846         }
2847 
2848         ret = vfio_pci_igd_opregion_init(vdev, opregion, errp);
2849         g_free(opregion);
2850         if (ret) {
2851             goto out_teardown;
2852         }
2853     }
2854 
2855     /* QEMU emulates all of MSI & MSIX */
2856     if (pdev->cap_present & QEMU_PCI_CAP_MSIX) {
2857         memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff,
2858                MSIX_CAP_LENGTH);
2859     }
2860 
2861     if (pdev->cap_present & QEMU_PCI_CAP_MSI) {
2862         memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff,
2863                vdev->msi_cap_size);
2864     }
2865 
2866     if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
2867         vdev->intx.mmap_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL,
2868                                                   vfio_intx_mmap_enable, vdev);
2869         pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_intx_update);
2870         ret = vfio_intx_enable(vdev, errp);
2871         if (ret) {
2872             goto out_teardown;
2873         }
2874     }
2875 
2876     vfio_register_err_notifier(vdev);
2877     vfio_register_req_notifier(vdev);
2878     vfio_setup_resetfn_quirk(vdev);
2879 
2880     return;
2881 
2882 out_teardown:
2883     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
2884     vfio_teardown_msi(vdev);
2885     vfio_bars_exit(vdev);
2886 error:
2887     error_prepend(errp, ERR_PREFIX, vdev->vbasedev.name);
2888 }
2889 
2890 static void vfio_instance_finalize(Object *obj)
2891 {
2892     PCIDevice *pci_dev = PCI_DEVICE(obj);
2893     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pci_dev);
2894     VFIOGroup *group = vdev->vbasedev.group;
2895 
2896     vfio_bars_finalize(vdev);
2897     g_free(vdev->emulated_config_bits);
2898     g_free(vdev->rom);
2899     /*
2900      * XXX Leaking igd_opregion is not an oversight, we can't remove the
2901      * fw_cfg entry therefore leaking this allocation seems like the safest
2902      * option.
2903      *
2904      * g_free(vdev->igd_opregion);
2905      */
2906     vfio_put_device(vdev);
2907     vfio_put_group(group);
2908 }
2909 
2910 static void vfio_exitfn(PCIDevice *pdev)
2911 {
2912     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
2913 
2914     vfio_unregister_req_notifier(vdev);
2915     vfio_unregister_err_notifier(vdev);
2916     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
2917     vfio_disable_interrupts(vdev);
2918     if (vdev->intx.mmap_timer) {
2919         timer_free(vdev->intx.mmap_timer);
2920     }
2921     vfio_teardown_msi(vdev);
2922     vfio_bars_exit(vdev);
2923 }
2924 
2925 static void vfio_pci_reset(DeviceState *dev)
2926 {
2927     PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
2928     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
2929 
2930     trace_vfio_pci_reset(vdev->vbasedev.name);
2931 
2932     vfio_pci_pre_reset(vdev);
2933 
2934     if (vdev->resetfn && !vdev->resetfn(vdev)) {
2935         goto post_reset;
2936     }
2937 
2938     if (vdev->vbasedev.reset_works &&
2939         (vdev->has_flr || !vdev->has_pm_reset) &&
2940         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
2941         trace_vfio_pci_reset_flr(vdev->vbasedev.name);
2942         goto post_reset;
2943     }
2944 
2945     /* See if we can do our own bus reset */
2946     if (!vfio_pci_hot_reset_one(vdev)) {
2947         goto post_reset;
2948     }
2949 
2950     /* If nothing else works and the device supports PM reset, use it */
2951     if (vdev->vbasedev.reset_works && vdev->has_pm_reset &&
2952         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
2953         trace_vfio_pci_reset_pm(vdev->vbasedev.name);
2954         goto post_reset;
2955     }
2956 
2957 post_reset:
2958     vfio_pci_post_reset(vdev);
2959 }
2960 
2961 static void vfio_instance_init(Object *obj)
2962 {
2963     PCIDevice *pci_dev = PCI_DEVICE(obj);
2964     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, PCI_DEVICE(obj));
2965 
2966     device_add_bootindex_property(obj, &vdev->bootindex,
2967                                   "bootindex", NULL,
2968                                   &pci_dev->qdev, NULL);
2969     vdev->host.domain = ~0U;
2970     vdev->host.bus = ~0U;
2971     vdev->host.slot = ~0U;
2972     vdev->host.function = ~0U;
2973 
2974     vdev->nv_gpudirect_clique = 0xFF;
2975 }
2976 
2977 static Property vfio_pci_dev_properties[] = {
2978     DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIOPCIDevice, host),
2979     DEFINE_PROP_STRING("sysfsdev", VFIOPCIDevice, vbasedev.sysfsdev),
2980     DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIOPCIDevice,
2981                        intx.mmap_timeout, 1100),
2982     DEFINE_PROP_BIT("x-vga", VFIOPCIDevice, features,
2983                     VFIO_FEATURE_ENABLE_VGA_BIT, false),
2984     DEFINE_PROP_BIT("x-req", VFIOPCIDevice, features,
2985                     VFIO_FEATURE_ENABLE_REQ_BIT, true),
2986     DEFINE_PROP_BIT("x-igd-opregion", VFIOPCIDevice, features,
2987                     VFIO_FEATURE_ENABLE_IGD_OPREGION_BIT, false),
2988     DEFINE_PROP_BOOL("x-no-mmap", VFIOPCIDevice, vbasedev.no_mmap, false),
2989     DEFINE_PROP_BOOL("x-no-kvm-intx", VFIOPCIDevice, no_kvm_intx, false),
2990     DEFINE_PROP_BOOL("x-no-kvm-msi", VFIOPCIDevice, no_kvm_msi, false),
2991     DEFINE_PROP_BOOL("x-no-kvm-msix", VFIOPCIDevice, no_kvm_msix, false),
2992     DEFINE_PROP_UINT32("x-pci-vendor-id", VFIOPCIDevice, vendor_id, PCI_ANY_ID),
2993     DEFINE_PROP_UINT32("x-pci-device-id", VFIOPCIDevice, device_id, PCI_ANY_ID),
2994     DEFINE_PROP_UINT32("x-pci-sub-vendor-id", VFIOPCIDevice,
2995                        sub_vendor_id, PCI_ANY_ID),
2996     DEFINE_PROP_UINT32("x-pci-sub-device-id", VFIOPCIDevice,
2997                        sub_device_id, PCI_ANY_ID),
2998     DEFINE_PROP_UINT32("x-igd-gms", VFIOPCIDevice, igd_gms, 0),
2999     DEFINE_PROP_UNSIGNED_NODEFAULT("x-nv-gpudirect-clique", VFIOPCIDevice,
3000                                    nv_gpudirect_clique,
3001                                    qdev_prop_nv_gpudirect_clique, uint8_t),
3002     /*
3003      * TODO - support passed fds... is this necessary?
3004      * DEFINE_PROP_STRING("vfiofd", VFIOPCIDevice, vfiofd_name),
3005      * DEFINE_PROP_STRING("vfiogroupfd, VFIOPCIDevice, vfiogroupfd_name),
3006      */
3007     DEFINE_PROP_END_OF_LIST(),
3008 };
3009 
3010 static const VMStateDescription vfio_pci_vmstate = {
3011     .name = "vfio-pci",
3012     .unmigratable = 1,
3013 };
3014 
3015 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
3016 {
3017     DeviceClass *dc = DEVICE_CLASS(klass);
3018     PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
3019 
3020     dc->reset = vfio_pci_reset;
3021     dc->props = vfio_pci_dev_properties;
3022     dc->vmsd = &vfio_pci_vmstate;
3023     dc->desc = "VFIO-based PCI device assignment";
3024     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
3025     pdc->realize = vfio_realize;
3026     pdc->exit = vfio_exitfn;
3027     pdc->config_read = vfio_pci_read_config;
3028     pdc->config_write = vfio_pci_write_config;
3029     pdc->is_express = 1; /* We might be */
3030 }
3031 
3032 static const TypeInfo vfio_pci_dev_info = {
3033     .name = "vfio-pci",
3034     .parent = TYPE_PCI_DEVICE,
3035     .instance_size = sizeof(VFIOPCIDevice),
3036     .class_init = vfio_pci_dev_class_init,
3037     .instance_init = vfio_instance_init,
3038     .instance_finalize = vfio_instance_finalize,
3039     .interfaces = (InterfaceInfo[]) {
3040         { INTERFACE_PCIE_DEVICE },
3041         { INTERFACE_CONVENTIONAL_PCI_DEVICE },
3042         { }
3043     },
3044 };
3045 
3046 static void register_vfio_pci_dev_type(void)
3047 {
3048     type_register_static(&vfio_pci_dev_info);
3049 }
3050 
3051 type_init(register_vfio_pci_dev_type)
3052