xref: /openbmc/qemu/hw/vfio/pci.c (revision d10e05f1)
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/option.h"
30 #include "qemu/range.h"
31 #include "qemu/units.h"
32 #include "sysemu/kvm.h"
33 #include "sysemu/sysemu.h"
34 #include "pci.h"
35 #include "trace.h"
36 #include "qapi/error.h"
37 
38 #define MSIX_CAP_LENGTH 12
39 
40 #define TYPE_VFIO_PCI "vfio-pci"
41 #define PCI_VFIO(obj)    OBJECT_CHECK(VFIOPCIDevice, obj, TYPE_VFIO_PCI)
42 
43 static void vfio_disable_interrupts(VFIOPCIDevice *vdev);
44 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled);
45 
46 /*
47  * Disabling BAR mmaping can be slow, but toggling it around INTx can
48  * also be a huge overhead.  We try to get the best of both worlds by
49  * waiting until an interrupt to disable mmaps (subsequent transitions
50  * to the same state are effectively no overhead).  If the interrupt has
51  * been serviced and the time gap is long enough, we re-enable mmaps for
52  * performance.  This works well for things like graphics cards, which
53  * may not use their interrupt at all and are penalized to an unusable
54  * level by read/write BAR traps.  Other devices, like NICs, have more
55  * regular interrupts and see much better latency by staying in non-mmap
56  * mode.  We therefore set the default mmap_timeout such that a ping
57  * is just enough to keep the mmap disabled.  Users can experiment with
58  * other options with the x-intx-mmap-timeout-ms parameter (a value of
59  * zero disables the timer).
60  */
61 static void vfio_intx_mmap_enable(void *opaque)
62 {
63     VFIOPCIDevice *vdev = opaque;
64 
65     if (vdev->intx.pending) {
66         timer_mod(vdev->intx.mmap_timer,
67                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
68         return;
69     }
70 
71     vfio_mmap_set_enabled(vdev, true);
72 }
73 
74 static void vfio_intx_interrupt(void *opaque)
75 {
76     VFIOPCIDevice *vdev = opaque;
77 
78     if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
79         return;
80     }
81 
82     trace_vfio_intx_interrupt(vdev->vbasedev.name, 'A' + vdev->intx.pin);
83 
84     vdev->intx.pending = true;
85     pci_irq_assert(&vdev->pdev);
86     vfio_mmap_set_enabled(vdev, false);
87     if (vdev->intx.mmap_timeout) {
88         timer_mod(vdev->intx.mmap_timer,
89                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
90     }
91 }
92 
93 static void vfio_intx_eoi(VFIODevice *vbasedev)
94 {
95     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
96 
97     if (!vdev->intx.pending) {
98         return;
99     }
100 
101     trace_vfio_intx_eoi(vbasedev->name);
102 
103     vdev->intx.pending = false;
104     pci_irq_deassert(&vdev->pdev);
105     vfio_unmask_single_irqindex(vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
106 }
107 
108 static void vfio_intx_enable_kvm(VFIOPCIDevice *vdev, Error **errp)
109 {
110 #ifdef CONFIG_KVM
111     struct kvm_irqfd irqfd = {
112         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
113         .gsi = vdev->intx.route.irq,
114         .flags = KVM_IRQFD_FLAG_RESAMPLE,
115     };
116     struct vfio_irq_set *irq_set;
117     int ret, argsz;
118     int32_t *pfd;
119 
120     if (vdev->no_kvm_intx || !kvm_irqfds_enabled() ||
121         vdev->intx.route.mode != PCI_INTX_ENABLED ||
122         !kvm_resamplefds_enabled()) {
123         return;
124     }
125 
126     /* Get to a known interrupt state */
127     qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
128     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
129     vdev->intx.pending = false;
130     pci_irq_deassert(&vdev->pdev);
131 
132     /* Get an eventfd for resample/unmask */
133     if (event_notifier_init(&vdev->intx.unmask, 0)) {
134         error_setg(errp, "event_notifier_init failed eoi");
135         goto fail;
136     }
137 
138     /* KVM triggers it, VFIO listens for it */
139     irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
140 
141     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
142         error_setg_errno(errp, errno, "failed to setup resample irqfd");
143         goto fail_irqfd;
144     }
145 
146     argsz = sizeof(*irq_set) + sizeof(*pfd);
147 
148     irq_set = g_malloc0(argsz);
149     irq_set->argsz = argsz;
150     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
151     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
152     irq_set->start = 0;
153     irq_set->count = 1;
154     pfd = (int32_t *)&irq_set->data;
155 
156     *pfd = irqfd.resamplefd;
157 
158     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
159     g_free(irq_set);
160     if (ret) {
161         error_setg_errno(errp, -ret, "failed to setup INTx unmask fd");
162         goto fail_vfio;
163     }
164 
165     /* Let'em rip */
166     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
167 
168     vdev->intx.kvm_accel = true;
169 
170     trace_vfio_intx_enable_kvm(vdev->vbasedev.name);
171 
172     return;
173 
174 fail_vfio:
175     irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
176     kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
177 fail_irqfd:
178     event_notifier_cleanup(&vdev->intx.unmask);
179 fail:
180     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
181     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
182 #endif
183 }
184 
185 static void vfio_intx_disable_kvm(VFIOPCIDevice *vdev)
186 {
187 #ifdef CONFIG_KVM
188     struct kvm_irqfd irqfd = {
189         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
190         .gsi = vdev->intx.route.irq,
191         .flags = KVM_IRQFD_FLAG_DEASSIGN,
192     };
193 
194     if (!vdev->intx.kvm_accel) {
195         return;
196     }
197 
198     /*
199      * Get to a known state, hardware masked, QEMU ready to accept new
200      * interrupts, QEMU IRQ de-asserted.
201      */
202     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
203     vdev->intx.pending = false;
204     pci_irq_deassert(&vdev->pdev);
205 
206     /* Tell KVM to stop listening for an INTx irqfd */
207     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
208         error_report("vfio: Error: Failed to disable INTx irqfd: %m");
209     }
210 
211     /* We only need to close the eventfd for VFIO to cleanup the kernel side */
212     event_notifier_cleanup(&vdev->intx.unmask);
213 
214     /* QEMU starts listening for interrupt events. */
215     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
216 
217     vdev->intx.kvm_accel = false;
218 
219     /* If we've missed an event, let it re-fire through QEMU */
220     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
221 
222     trace_vfio_intx_disable_kvm(vdev->vbasedev.name);
223 #endif
224 }
225 
226 static void vfio_intx_update(PCIDevice *pdev)
227 {
228     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
229     PCIINTxRoute route;
230     Error *err = NULL;
231 
232     if (vdev->interrupt != VFIO_INT_INTx) {
233         return;
234     }
235 
236     route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
237 
238     if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
239         return; /* Nothing changed */
240     }
241 
242     trace_vfio_intx_update(vdev->vbasedev.name,
243                            vdev->intx.route.irq, route.irq);
244 
245     vfio_intx_disable_kvm(vdev);
246 
247     vdev->intx.route = route;
248 
249     if (route.mode != PCI_INTX_ENABLED) {
250         return;
251     }
252 
253     vfio_intx_enable_kvm(vdev, &err);
254     if (err) {
255         warn_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
256     }
257 
258     /* Re-enable the interrupt in cased we missed an EOI */
259     vfio_intx_eoi(&vdev->vbasedev);
260 }
261 
262 static int vfio_intx_enable(VFIOPCIDevice *vdev, Error **errp)
263 {
264     uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
265     int ret, argsz, retval = 0;
266     struct vfio_irq_set *irq_set;
267     int32_t *pfd;
268     Error *err = NULL;
269 
270     if (!pin) {
271         return 0;
272     }
273 
274     vfio_disable_interrupts(vdev);
275 
276     vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
277     pci_config_set_interrupt_pin(vdev->pdev.config, pin);
278 
279 #ifdef CONFIG_KVM
280     /*
281      * Only conditional to avoid generating error messages on platforms
282      * where we won't actually use the result anyway.
283      */
284     if (kvm_irqfds_enabled() && kvm_resamplefds_enabled()) {
285         vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
286                                                         vdev->intx.pin);
287     }
288 #endif
289 
290     ret = event_notifier_init(&vdev->intx.interrupt, 0);
291     if (ret) {
292         error_setg_errno(errp, -ret, "event_notifier_init failed");
293         return ret;
294     }
295 
296     argsz = sizeof(*irq_set) + sizeof(*pfd);
297 
298     irq_set = g_malloc0(argsz);
299     irq_set->argsz = argsz;
300     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
301     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
302     irq_set->start = 0;
303     irq_set->count = 1;
304     pfd = (int32_t *)&irq_set->data;
305 
306     *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
307     qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
308 
309     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
310     if (ret) {
311         error_setg_errno(errp, -ret, "failed to setup INTx fd");
312         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
313         event_notifier_cleanup(&vdev->intx.interrupt);
314         retval = -errno;
315         goto cleanup;
316     }
317 
318     vfio_intx_enable_kvm(vdev, &err);
319     if (err) {
320         warn_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
321     }
322 
323     vdev->interrupt = VFIO_INT_INTx;
324 
325     trace_vfio_intx_enable(vdev->vbasedev.name);
326 
327 cleanup:
328     g_free(irq_set);
329 
330     return retval;
331 }
332 
333 static void vfio_intx_disable(VFIOPCIDevice *vdev)
334 {
335     int fd;
336 
337     timer_del(vdev->intx.mmap_timer);
338     vfio_intx_disable_kvm(vdev);
339     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
340     vdev->intx.pending = false;
341     pci_irq_deassert(&vdev->pdev);
342     vfio_mmap_set_enabled(vdev, true);
343 
344     fd = event_notifier_get_fd(&vdev->intx.interrupt);
345     qemu_set_fd_handler(fd, NULL, NULL, vdev);
346     event_notifier_cleanup(&vdev->intx.interrupt);
347 
348     vdev->interrupt = VFIO_INT_NONE;
349 
350     trace_vfio_intx_disable(vdev->vbasedev.name);
351 }
352 
353 /*
354  * MSI/X
355  */
356 static void vfio_msi_interrupt(void *opaque)
357 {
358     VFIOMSIVector *vector = opaque;
359     VFIOPCIDevice *vdev = vector->vdev;
360     MSIMessage (*get_msg)(PCIDevice *dev, unsigned vector);
361     void (*notify)(PCIDevice *dev, unsigned vector);
362     MSIMessage msg;
363     int nr = vector - vdev->msi_vectors;
364 
365     if (!event_notifier_test_and_clear(&vector->interrupt)) {
366         return;
367     }
368 
369     if (vdev->interrupt == VFIO_INT_MSIX) {
370         get_msg = msix_get_message;
371         notify = msix_notify;
372 
373         /* A masked vector firing needs to use the PBA, enable it */
374         if (msix_is_masked(&vdev->pdev, nr)) {
375             set_bit(nr, vdev->msix->pending);
376             memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, true);
377             trace_vfio_msix_pba_enable(vdev->vbasedev.name);
378         }
379     } else if (vdev->interrupt == VFIO_INT_MSI) {
380         get_msg = msi_get_message;
381         notify = msi_notify;
382     } else {
383         abort();
384     }
385 
386     msg = get_msg(&vdev->pdev, nr);
387     trace_vfio_msi_interrupt(vdev->vbasedev.name, nr, msg.address, msg.data);
388     notify(&vdev->pdev, nr);
389 }
390 
391 static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool msix)
392 {
393     struct vfio_irq_set *irq_set;
394     int ret = 0, i, argsz;
395     int32_t *fds;
396 
397     argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
398 
399     irq_set = g_malloc0(argsz);
400     irq_set->argsz = argsz;
401     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
402     irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
403     irq_set->start = 0;
404     irq_set->count = vdev->nr_vectors;
405     fds = (int32_t *)&irq_set->data;
406 
407     for (i = 0; i < vdev->nr_vectors; i++) {
408         int fd = -1;
409 
410         /*
411          * MSI vs MSI-X - The guest has direct access to MSI mask and pending
412          * bits, therefore we always use the KVM signaling path when setup.
413          * MSI-X mask and pending bits are emulated, so we want to use the
414          * KVM signaling path only when configured and unmasked.
415          */
416         if (vdev->msi_vectors[i].use) {
417             if (vdev->msi_vectors[i].virq < 0 ||
418                 (msix && msix_is_masked(&vdev->pdev, i))) {
419                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
420             } else {
421                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].kvm_interrupt);
422             }
423         }
424 
425         fds[i] = fd;
426     }
427 
428     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
429 
430     g_free(irq_set);
431 
432     return ret;
433 }
434 
435 static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
436                                   int vector_n, bool msix)
437 {
438     int virq;
439 
440     if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) {
441         return;
442     }
443 
444     if (event_notifier_init(&vector->kvm_interrupt, 0)) {
445         return;
446     }
447 
448     virq = kvm_irqchip_add_msi_route(kvm_state, vector_n, &vdev->pdev);
449     if (virq < 0) {
450         event_notifier_cleanup(&vector->kvm_interrupt);
451         return;
452     }
453 
454     if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
455                                        NULL, virq) < 0) {
456         kvm_irqchip_release_virq(kvm_state, virq);
457         event_notifier_cleanup(&vector->kvm_interrupt);
458         return;
459     }
460 
461     vector->virq = virq;
462 }
463 
464 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
465 {
466     kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
467                                           vector->virq);
468     kvm_irqchip_release_virq(kvm_state, vector->virq);
469     vector->virq = -1;
470     event_notifier_cleanup(&vector->kvm_interrupt);
471 }
472 
473 static void vfio_update_kvm_msi_virq(VFIOMSIVector *vector, MSIMessage msg,
474                                      PCIDevice *pdev)
475 {
476     kvm_irqchip_update_msi_route(kvm_state, vector->virq, msg, pdev);
477     kvm_irqchip_commit_routes(kvm_state);
478 }
479 
480 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
481                                    MSIMessage *msg, IOHandler *handler)
482 {
483     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
484     VFIOMSIVector *vector;
485     int ret;
486 
487     trace_vfio_msix_vector_do_use(vdev->vbasedev.name, nr);
488 
489     vector = &vdev->msi_vectors[nr];
490 
491     if (!vector->use) {
492         vector->vdev = vdev;
493         vector->virq = -1;
494         if (event_notifier_init(&vector->interrupt, 0)) {
495             error_report("vfio: Error: event_notifier_init failed");
496         }
497         vector->use = true;
498         msix_vector_use(pdev, nr);
499     }
500 
501     qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
502                         handler, NULL, vector);
503 
504     /*
505      * Attempt to enable route through KVM irqchip,
506      * default to userspace handling if unavailable.
507      */
508     if (vector->virq >= 0) {
509         if (!msg) {
510             vfio_remove_kvm_msi_virq(vector);
511         } else {
512             vfio_update_kvm_msi_virq(vector, *msg, pdev);
513         }
514     } else {
515         if (msg) {
516             vfio_add_kvm_msi_virq(vdev, vector, nr, true);
517         }
518     }
519 
520     /*
521      * We don't want to have the host allocate all possible MSI vectors
522      * for a device if they're not in use, so we shutdown and incrementally
523      * increase them as needed.
524      */
525     if (vdev->nr_vectors < nr + 1) {
526         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
527         vdev->nr_vectors = nr + 1;
528         ret = vfio_enable_vectors(vdev, true);
529         if (ret) {
530             error_report("vfio: failed to enable vectors, %d", ret);
531         }
532     } else {
533         int argsz;
534         struct vfio_irq_set *irq_set;
535         int32_t *pfd;
536 
537         argsz = sizeof(*irq_set) + sizeof(*pfd);
538 
539         irq_set = g_malloc0(argsz);
540         irq_set->argsz = argsz;
541         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
542                          VFIO_IRQ_SET_ACTION_TRIGGER;
543         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
544         irq_set->start = nr;
545         irq_set->count = 1;
546         pfd = (int32_t *)&irq_set->data;
547 
548         if (vector->virq >= 0) {
549             *pfd = event_notifier_get_fd(&vector->kvm_interrupt);
550         } else {
551             *pfd = event_notifier_get_fd(&vector->interrupt);
552         }
553 
554         ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
555         g_free(irq_set);
556         if (ret) {
557             error_report("vfio: failed to modify vector, %d", ret);
558         }
559     }
560 
561     /* Disable PBA emulation when nothing more is pending. */
562     clear_bit(nr, vdev->msix->pending);
563     if (find_first_bit(vdev->msix->pending,
564                        vdev->nr_vectors) == vdev->nr_vectors) {
565         memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
566         trace_vfio_msix_pba_disable(vdev->vbasedev.name);
567     }
568 
569     return 0;
570 }
571 
572 static int vfio_msix_vector_use(PCIDevice *pdev,
573                                 unsigned int nr, MSIMessage msg)
574 {
575     return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
576 }
577 
578 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
579 {
580     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
581     VFIOMSIVector *vector = &vdev->msi_vectors[nr];
582 
583     trace_vfio_msix_vector_release(vdev->vbasedev.name, nr);
584 
585     /*
586      * There are still old guests that mask and unmask vectors on every
587      * interrupt.  If we're using QEMU bypass with a KVM irqfd, leave all of
588      * the KVM setup in place, simply switch VFIO to use the non-bypass
589      * eventfd.  We'll then fire the interrupt through QEMU and the MSI-X
590      * core will mask the interrupt and set pending bits, allowing it to
591      * be re-asserted on unmask.  Nothing to do if already using QEMU mode.
592      */
593     if (vector->virq >= 0) {
594         int argsz;
595         struct vfio_irq_set *irq_set;
596         int32_t *pfd;
597 
598         argsz = sizeof(*irq_set) + sizeof(*pfd);
599 
600         irq_set = g_malloc0(argsz);
601         irq_set->argsz = argsz;
602         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
603                          VFIO_IRQ_SET_ACTION_TRIGGER;
604         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
605         irq_set->start = nr;
606         irq_set->count = 1;
607         pfd = (int32_t *)&irq_set->data;
608 
609         *pfd = event_notifier_get_fd(&vector->interrupt);
610 
611         ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
612 
613         g_free(irq_set);
614     }
615 }
616 
617 static void vfio_msix_enable(VFIOPCIDevice *vdev)
618 {
619     vfio_disable_interrupts(vdev);
620 
621     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->msix->entries);
622 
623     vdev->interrupt = VFIO_INT_MSIX;
624 
625     /*
626      * Some communication channels between VF & PF or PF & fw rely on the
627      * physical state of the device and expect that enabling MSI-X from the
628      * guest enables the same on the host.  When our guest is Linux, the
629      * guest driver call to pci_enable_msix() sets the enabling bit in the
630      * MSI-X capability, but leaves the vector table masked.  We therefore
631      * can't rely on a vector_use callback (from request_irq() in the guest)
632      * to switch the physical device into MSI-X mode because that may come a
633      * long time after pci_enable_msix().  This code enables vector 0 with
634      * triggering to userspace, then immediately release the vector, leaving
635      * the physical device with no vectors enabled, but MSI-X enabled, just
636      * like the guest view.
637      */
638     vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
639     vfio_msix_vector_release(&vdev->pdev, 0);
640 
641     if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
642                                   vfio_msix_vector_release, NULL)) {
643         error_report("vfio: msix_set_vector_notifiers failed");
644     }
645 
646     trace_vfio_msix_enable(vdev->vbasedev.name);
647 }
648 
649 static void vfio_msi_enable(VFIOPCIDevice *vdev)
650 {
651     int ret, i;
652 
653     vfio_disable_interrupts(vdev);
654 
655     vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
656 retry:
657     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->nr_vectors);
658 
659     for (i = 0; i < vdev->nr_vectors; i++) {
660         VFIOMSIVector *vector = &vdev->msi_vectors[i];
661 
662         vector->vdev = vdev;
663         vector->virq = -1;
664         vector->use = true;
665 
666         if (event_notifier_init(&vector->interrupt, 0)) {
667             error_report("vfio: Error: event_notifier_init failed");
668         }
669 
670         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
671                             vfio_msi_interrupt, NULL, vector);
672 
673         /*
674          * Attempt to enable route through KVM irqchip,
675          * default to userspace handling if unavailable.
676          */
677         vfio_add_kvm_msi_virq(vdev, vector, i, false);
678     }
679 
680     /* Set interrupt type prior to possible interrupts */
681     vdev->interrupt = VFIO_INT_MSI;
682 
683     ret = vfio_enable_vectors(vdev, false);
684     if (ret) {
685         if (ret < 0) {
686             error_report("vfio: Error: Failed to setup MSI fds: %m");
687         } else if (ret != vdev->nr_vectors) {
688             error_report("vfio: Error: Failed to enable %d "
689                          "MSI vectors, retry with %d", vdev->nr_vectors, ret);
690         }
691 
692         for (i = 0; i < vdev->nr_vectors; i++) {
693             VFIOMSIVector *vector = &vdev->msi_vectors[i];
694             if (vector->virq >= 0) {
695                 vfio_remove_kvm_msi_virq(vector);
696             }
697             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
698                                 NULL, NULL, NULL);
699             event_notifier_cleanup(&vector->interrupt);
700         }
701 
702         g_free(vdev->msi_vectors);
703 
704         if (ret > 0 && ret != vdev->nr_vectors) {
705             vdev->nr_vectors = ret;
706             goto retry;
707         }
708         vdev->nr_vectors = 0;
709 
710         /*
711          * Failing to setup MSI doesn't really fall within any specification.
712          * Let's try leaving interrupts disabled and hope the guest figures
713          * out to fall back to INTx for this device.
714          */
715         error_report("vfio: Error: Failed to enable MSI");
716         vdev->interrupt = VFIO_INT_NONE;
717 
718         return;
719     }
720 
721     trace_vfio_msi_enable(vdev->vbasedev.name, vdev->nr_vectors);
722 }
723 
724 static void vfio_msi_disable_common(VFIOPCIDevice *vdev)
725 {
726     Error *err = NULL;
727     int i;
728 
729     for (i = 0; i < vdev->nr_vectors; i++) {
730         VFIOMSIVector *vector = &vdev->msi_vectors[i];
731         if (vdev->msi_vectors[i].use) {
732             if (vector->virq >= 0) {
733                 vfio_remove_kvm_msi_virq(vector);
734             }
735             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
736                                 NULL, NULL, NULL);
737             event_notifier_cleanup(&vector->interrupt);
738         }
739     }
740 
741     g_free(vdev->msi_vectors);
742     vdev->msi_vectors = NULL;
743     vdev->nr_vectors = 0;
744     vdev->interrupt = VFIO_INT_NONE;
745 
746     vfio_intx_enable(vdev, &err);
747     if (err) {
748         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
749     }
750 }
751 
752 static void vfio_msix_disable(VFIOPCIDevice *vdev)
753 {
754     int i;
755 
756     msix_unset_vector_notifiers(&vdev->pdev);
757 
758     /*
759      * MSI-X will only release vectors if MSI-X is still enabled on the
760      * device, check through the rest and release it ourselves if necessary.
761      */
762     for (i = 0; i < vdev->nr_vectors; i++) {
763         if (vdev->msi_vectors[i].use) {
764             vfio_msix_vector_release(&vdev->pdev, i);
765             msix_vector_unuse(&vdev->pdev, i);
766         }
767     }
768 
769     if (vdev->nr_vectors) {
770         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
771     }
772 
773     vfio_msi_disable_common(vdev);
774 
775     memset(vdev->msix->pending, 0,
776            BITS_TO_LONGS(vdev->msix->entries) * sizeof(unsigned long));
777 
778     trace_vfio_msix_disable(vdev->vbasedev.name);
779 }
780 
781 static void vfio_msi_disable(VFIOPCIDevice *vdev)
782 {
783     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSI_IRQ_INDEX);
784     vfio_msi_disable_common(vdev);
785 
786     trace_vfio_msi_disable(vdev->vbasedev.name);
787 }
788 
789 static void vfio_update_msi(VFIOPCIDevice *vdev)
790 {
791     int i;
792 
793     for (i = 0; i < vdev->nr_vectors; i++) {
794         VFIOMSIVector *vector = &vdev->msi_vectors[i];
795         MSIMessage msg;
796 
797         if (!vector->use || vector->virq < 0) {
798             continue;
799         }
800 
801         msg = msi_get_message(&vdev->pdev, i);
802         vfio_update_kvm_msi_virq(vector, msg, &vdev->pdev);
803     }
804 }
805 
806 static void vfio_pci_load_rom(VFIOPCIDevice *vdev)
807 {
808     struct vfio_region_info *reg_info;
809     uint64_t size;
810     off_t off = 0;
811     ssize_t bytes;
812 
813     if (vfio_get_region_info(&vdev->vbasedev,
814                              VFIO_PCI_ROM_REGION_INDEX, &reg_info)) {
815         error_report("vfio: Error getting ROM info: %m");
816         return;
817     }
818 
819     trace_vfio_pci_load_rom(vdev->vbasedev.name, (unsigned long)reg_info->size,
820                             (unsigned long)reg_info->offset,
821                             (unsigned long)reg_info->flags);
822 
823     vdev->rom_size = size = reg_info->size;
824     vdev->rom_offset = reg_info->offset;
825 
826     g_free(reg_info);
827 
828     if (!vdev->rom_size) {
829         vdev->rom_read_failed = true;
830         error_report("vfio-pci: Cannot read device rom at "
831                     "%s", vdev->vbasedev.name);
832         error_printf("Device option ROM contents are probably invalid "
833                     "(check dmesg).\nSkip option ROM probe with rombar=0, "
834                     "or load from file with romfile=\n");
835         return;
836     }
837 
838     vdev->rom = g_malloc(size);
839     memset(vdev->rom, 0xff, size);
840 
841     while (size) {
842         bytes = pread(vdev->vbasedev.fd, vdev->rom + off,
843                       size, vdev->rom_offset + off);
844         if (bytes == 0) {
845             break;
846         } else if (bytes > 0) {
847             off += bytes;
848             size -= bytes;
849         } else {
850             if (errno == EINTR || errno == EAGAIN) {
851                 continue;
852             }
853             error_report("vfio: Error reading device ROM: %m");
854             break;
855         }
856     }
857 
858     /*
859      * Test the ROM signature against our device, if the vendor is correct
860      * but the device ID doesn't match, store the correct device ID and
861      * recompute the checksum.  Intel IGD devices need this and are known
862      * to have bogus checksums so we can't simply adjust the checksum.
863      */
864     if (pci_get_word(vdev->rom) == 0xaa55 &&
865         pci_get_word(vdev->rom + 0x18) + 8 < vdev->rom_size &&
866         !memcmp(vdev->rom + pci_get_word(vdev->rom + 0x18), "PCIR", 4)) {
867         uint16_t vid, did;
868 
869         vid = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 4);
870         did = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6);
871 
872         if (vid == vdev->vendor_id && did != vdev->device_id) {
873             int i;
874             uint8_t csum, *data = vdev->rom;
875 
876             pci_set_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6,
877                          vdev->device_id);
878             data[6] = 0;
879 
880             for (csum = 0, i = 0; i < vdev->rom_size; i++) {
881                 csum += data[i];
882             }
883 
884             data[6] = -csum;
885         }
886     }
887 }
888 
889 static uint64_t vfio_rom_read(void *opaque, hwaddr addr, unsigned size)
890 {
891     VFIOPCIDevice *vdev = opaque;
892     union {
893         uint8_t byte;
894         uint16_t word;
895         uint32_t dword;
896         uint64_t qword;
897     } val;
898     uint64_t data = 0;
899 
900     /* Load the ROM lazily when the guest tries to read it */
901     if (unlikely(!vdev->rom && !vdev->rom_read_failed)) {
902         vfio_pci_load_rom(vdev);
903     }
904 
905     memcpy(&val, vdev->rom + addr,
906            (addr < vdev->rom_size) ? MIN(size, vdev->rom_size - addr) : 0);
907 
908     switch (size) {
909     case 1:
910         data = val.byte;
911         break;
912     case 2:
913         data = le16_to_cpu(val.word);
914         break;
915     case 4:
916         data = le32_to_cpu(val.dword);
917         break;
918     default:
919         hw_error("vfio: unsupported read size, %d bytes\n", size);
920         break;
921     }
922 
923     trace_vfio_rom_read(vdev->vbasedev.name, addr, size, data);
924 
925     return data;
926 }
927 
928 static void vfio_rom_write(void *opaque, hwaddr addr,
929                            uint64_t data, unsigned size)
930 {
931 }
932 
933 static const MemoryRegionOps vfio_rom_ops = {
934     .read = vfio_rom_read,
935     .write = vfio_rom_write,
936     .endianness = DEVICE_LITTLE_ENDIAN,
937 };
938 
939 static void vfio_pci_size_rom(VFIOPCIDevice *vdev)
940 {
941     uint32_t orig, size = cpu_to_le32((uint32_t)PCI_ROM_ADDRESS_MASK);
942     off_t offset = vdev->config_offset + PCI_ROM_ADDRESS;
943     DeviceState *dev = DEVICE(vdev);
944     char *name;
945     int fd = vdev->vbasedev.fd;
946 
947     if (vdev->pdev.romfile || !vdev->pdev.rom_bar) {
948         /* Since pci handles romfile, just print a message and return */
949         if (vfio_blacklist_opt_rom(vdev) && vdev->pdev.romfile) {
950             error_printf("Warning : Device at %s is known to cause system instability issues during option rom execution. Proceeding anyway since user specified romfile\n",
951                          vdev->vbasedev.name);
952         }
953         return;
954     }
955 
956     /*
957      * Use the same size ROM BAR as the physical device.  The contents
958      * will get filled in later when the guest tries to read it.
959      */
960     if (pread(fd, &orig, 4, offset) != 4 ||
961         pwrite(fd, &size, 4, offset) != 4 ||
962         pread(fd, &size, 4, offset) != 4 ||
963         pwrite(fd, &orig, 4, offset) != 4) {
964         error_report("%s(%s) failed: %m", __func__, vdev->vbasedev.name);
965         return;
966     }
967 
968     size = ~(le32_to_cpu(size) & PCI_ROM_ADDRESS_MASK) + 1;
969 
970     if (!size) {
971         return;
972     }
973 
974     if (vfio_blacklist_opt_rom(vdev)) {
975         if (dev->opts && qemu_opt_get(dev->opts, "rombar")) {
976             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",
977                          vdev->vbasedev.name);
978         } else {
979             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",
980                          vdev->vbasedev.name);
981             return;
982         }
983     }
984 
985     trace_vfio_pci_size_rom(vdev->vbasedev.name, size);
986 
987     name = g_strdup_printf("vfio[%s].rom", vdev->vbasedev.name);
988 
989     memory_region_init_io(&vdev->pdev.rom, OBJECT(vdev),
990                           &vfio_rom_ops, vdev, name, size);
991     g_free(name);
992 
993     pci_register_bar(&vdev->pdev, PCI_ROM_SLOT,
994                      PCI_BASE_ADDRESS_SPACE_MEMORY, &vdev->pdev.rom);
995 
996     vdev->rom_read_failed = false;
997 }
998 
999 void vfio_vga_write(void *opaque, hwaddr addr,
1000                            uint64_t data, unsigned size)
1001 {
1002     VFIOVGARegion *region = opaque;
1003     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1004     union {
1005         uint8_t byte;
1006         uint16_t word;
1007         uint32_t dword;
1008         uint64_t qword;
1009     } buf;
1010     off_t offset = vga->fd_offset + region->offset + addr;
1011 
1012     switch (size) {
1013     case 1:
1014         buf.byte = data;
1015         break;
1016     case 2:
1017         buf.word = cpu_to_le16(data);
1018         break;
1019     case 4:
1020         buf.dword = cpu_to_le32(data);
1021         break;
1022     default:
1023         hw_error("vfio: unsupported write size, %d bytes", size);
1024         break;
1025     }
1026 
1027     if (pwrite(vga->fd, &buf, size, offset) != size) {
1028         error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
1029                      __func__, region->offset + addr, data, size);
1030     }
1031 
1032     trace_vfio_vga_write(region->offset + addr, data, size);
1033 }
1034 
1035 uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size)
1036 {
1037     VFIOVGARegion *region = opaque;
1038     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1039     union {
1040         uint8_t byte;
1041         uint16_t word;
1042         uint32_t dword;
1043         uint64_t qword;
1044     } buf;
1045     uint64_t data = 0;
1046     off_t offset = vga->fd_offset + region->offset + addr;
1047 
1048     if (pread(vga->fd, &buf, size, offset) != size) {
1049         error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
1050                      __func__, region->offset + addr, size);
1051         return (uint64_t)-1;
1052     }
1053 
1054     switch (size) {
1055     case 1:
1056         data = buf.byte;
1057         break;
1058     case 2:
1059         data = le16_to_cpu(buf.word);
1060         break;
1061     case 4:
1062         data = le32_to_cpu(buf.dword);
1063         break;
1064     default:
1065         hw_error("vfio: unsupported read size, %d bytes", size);
1066         break;
1067     }
1068 
1069     trace_vfio_vga_read(region->offset + addr, size, data);
1070 
1071     return data;
1072 }
1073 
1074 static const MemoryRegionOps vfio_vga_ops = {
1075     .read = vfio_vga_read,
1076     .write = vfio_vga_write,
1077     .endianness = DEVICE_LITTLE_ENDIAN,
1078 };
1079 
1080 /*
1081  * Expand memory region of sub-page(size < PAGE_SIZE) MMIO BAR to page
1082  * size if the BAR is in an exclusive page in host so that we could map
1083  * this BAR to guest. But this sub-page BAR may not occupy an exclusive
1084  * page in guest. So we should set the priority of the expanded memory
1085  * region to zero in case of overlap with BARs which share the same page
1086  * with the sub-page BAR in guest. Besides, we should also recover the
1087  * size of this sub-page BAR when its base address is changed in guest
1088  * and not page aligned any more.
1089  */
1090 static void vfio_sub_page_bar_update_mapping(PCIDevice *pdev, int bar)
1091 {
1092     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
1093     VFIORegion *region = &vdev->bars[bar].region;
1094     MemoryRegion *mmap_mr, *region_mr, *base_mr;
1095     PCIIORegion *r;
1096     pcibus_t bar_addr;
1097     uint64_t size = region->size;
1098 
1099     /* Make sure that the whole region is allowed to be mmapped */
1100     if (region->nr_mmaps != 1 || !region->mmaps[0].mmap ||
1101         region->mmaps[0].size != region->size) {
1102         return;
1103     }
1104 
1105     r = &pdev->io_regions[bar];
1106     bar_addr = r->addr;
1107     base_mr = vdev->bars[bar].mr;
1108     region_mr = region->mem;
1109     mmap_mr = &region->mmaps[0].mem;
1110 
1111     /* If BAR is mapped and page aligned, update to fill PAGE_SIZE */
1112     if (bar_addr != PCI_BAR_UNMAPPED &&
1113         !(bar_addr & ~qemu_real_host_page_mask)) {
1114         size = qemu_real_host_page_size;
1115     }
1116 
1117     memory_region_transaction_begin();
1118 
1119     if (vdev->bars[bar].size < size) {
1120         memory_region_set_size(base_mr, size);
1121     }
1122     memory_region_set_size(region_mr, size);
1123     memory_region_set_size(mmap_mr, size);
1124     if (size != vdev->bars[bar].size && memory_region_is_mapped(base_mr)) {
1125         memory_region_del_subregion(r->address_space, base_mr);
1126         memory_region_add_subregion_overlap(r->address_space,
1127                                             bar_addr, base_mr, 0);
1128     }
1129 
1130     memory_region_transaction_commit();
1131 }
1132 
1133 /*
1134  * PCI config space
1135  */
1136 uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
1137 {
1138     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
1139     uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val;
1140 
1141     memcpy(&emu_bits, vdev->emulated_config_bits + addr, len);
1142     emu_bits = le32_to_cpu(emu_bits);
1143 
1144     if (emu_bits) {
1145         emu_val = pci_default_read_config(pdev, addr, len);
1146     }
1147 
1148     if (~emu_bits & (0xffffffffU >> (32 - len * 8))) {
1149         ssize_t ret;
1150 
1151         ret = pread(vdev->vbasedev.fd, &phys_val, len,
1152                     vdev->config_offset + addr);
1153         if (ret != len) {
1154             error_report("%s(%s, 0x%x, 0x%x) failed: %m",
1155                          __func__, vdev->vbasedev.name, addr, len);
1156             return -errno;
1157         }
1158         phys_val = le32_to_cpu(phys_val);
1159     }
1160 
1161     val = (emu_val & emu_bits) | (phys_val & ~emu_bits);
1162 
1163     trace_vfio_pci_read_config(vdev->vbasedev.name, addr, len, val);
1164 
1165     return val;
1166 }
1167 
1168 void vfio_pci_write_config(PCIDevice *pdev,
1169                            uint32_t addr, uint32_t val, int len)
1170 {
1171     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
1172     uint32_t val_le = cpu_to_le32(val);
1173 
1174     trace_vfio_pci_write_config(vdev->vbasedev.name, addr, val, len);
1175 
1176     /* Write everything to VFIO, let it filter out what we can't write */
1177     if (pwrite(vdev->vbasedev.fd, &val_le, len, vdev->config_offset + addr)
1178                 != len) {
1179         error_report("%s(%s, 0x%x, 0x%x, 0x%x) failed: %m",
1180                      __func__, vdev->vbasedev.name, addr, val, len);
1181     }
1182 
1183     /* MSI/MSI-X Enabling/Disabling */
1184     if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
1185         ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
1186         int is_enabled, was_enabled = msi_enabled(pdev);
1187 
1188         pci_default_write_config(pdev, addr, val, len);
1189 
1190         is_enabled = msi_enabled(pdev);
1191 
1192         if (!was_enabled) {
1193             if (is_enabled) {
1194                 vfio_msi_enable(vdev);
1195             }
1196         } else {
1197             if (!is_enabled) {
1198                 vfio_msi_disable(vdev);
1199             } else {
1200                 vfio_update_msi(vdev);
1201             }
1202         }
1203     } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
1204         ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
1205         int is_enabled, was_enabled = msix_enabled(pdev);
1206 
1207         pci_default_write_config(pdev, addr, val, len);
1208 
1209         is_enabled = msix_enabled(pdev);
1210 
1211         if (!was_enabled && is_enabled) {
1212             vfio_msix_enable(vdev);
1213         } else if (was_enabled && !is_enabled) {
1214             vfio_msix_disable(vdev);
1215         }
1216     } else if (ranges_overlap(addr, len, PCI_BASE_ADDRESS_0, 24) ||
1217         range_covers_byte(addr, len, PCI_COMMAND)) {
1218         pcibus_t old_addr[PCI_NUM_REGIONS - 1];
1219         int bar;
1220 
1221         for (bar = 0; bar < PCI_ROM_SLOT; bar++) {
1222             old_addr[bar] = pdev->io_regions[bar].addr;
1223         }
1224 
1225         pci_default_write_config(pdev, addr, val, len);
1226 
1227         for (bar = 0; bar < PCI_ROM_SLOT; bar++) {
1228             if (old_addr[bar] != pdev->io_regions[bar].addr &&
1229                 vdev->bars[bar].region.size > 0 &&
1230                 vdev->bars[bar].region.size < qemu_real_host_page_size) {
1231                 vfio_sub_page_bar_update_mapping(pdev, bar);
1232             }
1233         }
1234     } else {
1235         /* Write everything to QEMU to keep emulated bits correct */
1236         pci_default_write_config(pdev, addr, val, len);
1237     }
1238 }
1239 
1240 /*
1241  * Interrupt setup
1242  */
1243 static void vfio_disable_interrupts(VFIOPCIDevice *vdev)
1244 {
1245     /*
1246      * More complicated than it looks.  Disabling MSI/X transitions the
1247      * device to INTx mode (if supported).  Therefore we need to first
1248      * disable MSI/X and then cleanup by disabling INTx.
1249      */
1250     if (vdev->interrupt == VFIO_INT_MSIX) {
1251         vfio_msix_disable(vdev);
1252     } else if (vdev->interrupt == VFIO_INT_MSI) {
1253         vfio_msi_disable(vdev);
1254     }
1255 
1256     if (vdev->interrupt == VFIO_INT_INTx) {
1257         vfio_intx_disable(vdev);
1258     }
1259 }
1260 
1261 static int vfio_msi_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1262 {
1263     uint16_t ctrl;
1264     bool msi_64bit, msi_maskbit;
1265     int ret, entries;
1266     Error *err = NULL;
1267 
1268     if (pread(vdev->vbasedev.fd, &ctrl, sizeof(ctrl),
1269               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1270         error_setg_errno(errp, errno, "failed reading MSI PCI_CAP_FLAGS");
1271         return -errno;
1272     }
1273     ctrl = le16_to_cpu(ctrl);
1274 
1275     msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
1276     msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
1277     entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
1278 
1279     trace_vfio_msi_setup(vdev->vbasedev.name, pos);
1280 
1281     ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit, &err);
1282     if (ret < 0) {
1283         if (ret == -ENOTSUP) {
1284             return 0;
1285         }
1286         error_propagate_prepend(errp, err, "msi_init failed: ");
1287         return ret;
1288     }
1289     vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
1290 
1291     return 0;
1292 }
1293 
1294 static void vfio_pci_fixup_msix_region(VFIOPCIDevice *vdev)
1295 {
1296     off_t start, end;
1297     VFIORegion *region = &vdev->bars[vdev->msix->table_bar].region;
1298 
1299     /*
1300      * If the host driver allows mapping of a MSIX data, we are going to
1301      * do map the entire BAR and emulate MSIX table on top of that.
1302      */
1303     if (vfio_has_region_cap(&vdev->vbasedev, region->nr,
1304                             VFIO_REGION_INFO_CAP_MSIX_MAPPABLE)) {
1305         return;
1306     }
1307 
1308     /*
1309      * We expect to find a single mmap covering the whole BAR, anything else
1310      * means it's either unsupported or already setup.
1311      */
1312     if (region->nr_mmaps != 1 || region->mmaps[0].offset ||
1313         region->size != region->mmaps[0].size) {
1314         return;
1315     }
1316 
1317     /* MSI-X table start and end aligned to host page size */
1318     start = vdev->msix->table_offset & qemu_real_host_page_mask;
1319     end = REAL_HOST_PAGE_ALIGN((uint64_t)vdev->msix->table_offset +
1320                                (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
1321 
1322     /*
1323      * Does the MSI-X table cover the beginning of the BAR?  The whole BAR?
1324      * NB - Host page size is necessarily a power of two and so is the PCI
1325      * BAR (not counting EA yet), therefore if we have host page aligned
1326      * @start and @end, then any remainder of the BAR before or after those
1327      * must be at least host page sized and therefore mmap'able.
1328      */
1329     if (!start) {
1330         if (end >= region->size) {
1331             region->nr_mmaps = 0;
1332             g_free(region->mmaps);
1333             region->mmaps = NULL;
1334             trace_vfio_msix_fixup(vdev->vbasedev.name,
1335                                   vdev->msix->table_bar, 0, 0);
1336         } else {
1337             region->mmaps[0].offset = end;
1338             region->mmaps[0].size = region->size - end;
1339             trace_vfio_msix_fixup(vdev->vbasedev.name,
1340                               vdev->msix->table_bar, region->mmaps[0].offset,
1341                               region->mmaps[0].offset + region->mmaps[0].size);
1342         }
1343 
1344     /* Maybe it's aligned at the end of the BAR */
1345     } else if (end >= region->size) {
1346         region->mmaps[0].size = start;
1347         trace_vfio_msix_fixup(vdev->vbasedev.name,
1348                               vdev->msix->table_bar, region->mmaps[0].offset,
1349                               region->mmaps[0].offset + region->mmaps[0].size);
1350 
1351     /* Otherwise it must split the BAR */
1352     } else {
1353         region->nr_mmaps = 2;
1354         region->mmaps = g_renew(VFIOMmap, region->mmaps, 2);
1355 
1356         memcpy(&region->mmaps[1], &region->mmaps[0], sizeof(VFIOMmap));
1357 
1358         region->mmaps[0].size = start;
1359         trace_vfio_msix_fixup(vdev->vbasedev.name,
1360                               vdev->msix->table_bar, region->mmaps[0].offset,
1361                               region->mmaps[0].offset + region->mmaps[0].size);
1362 
1363         region->mmaps[1].offset = end;
1364         region->mmaps[1].size = region->size - end;
1365         trace_vfio_msix_fixup(vdev->vbasedev.name,
1366                               vdev->msix->table_bar, region->mmaps[1].offset,
1367                               region->mmaps[1].offset + region->mmaps[1].size);
1368     }
1369 }
1370 
1371 static void vfio_pci_relocate_msix(VFIOPCIDevice *vdev, Error **errp)
1372 {
1373     int target_bar = -1;
1374     size_t msix_sz;
1375 
1376     if (!vdev->msix || vdev->msix_relo == OFF_AUTOPCIBAR_OFF) {
1377         return;
1378     }
1379 
1380     /* The actual minimum size of MSI-X structures */
1381     msix_sz = (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE) +
1382               (QEMU_ALIGN_UP(vdev->msix->entries, 64) / 8);
1383     /* Round up to host pages, we don't want to share a page */
1384     msix_sz = REAL_HOST_PAGE_ALIGN(msix_sz);
1385     /* PCI BARs must be a power of 2 */
1386     msix_sz = pow2ceil(msix_sz);
1387 
1388     if (vdev->msix_relo == OFF_AUTOPCIBAR_AUTO) {
1389         /*
1390          * TODO: Lookup table for known devices.
1391          *
1392          * Logically we might use an algorithm here to select the BAR adding
1393          * the least additional MMIO space, but we cannot programatically
1394          * predict the driver dependency on BAR ordering or sizing, therefore
1395          * 'auto' becomes a lookup for combinations reported to work.
1396          */
1397         if (target_bar < 0) {
1398             error_setg(errp, "No automatic MSI-X relocation available for "
1399                        "device %04x:%04x", vdev->vendor_id, vdev->device_id);
1400             return;
1401         }
1402     } else {
1403         target_bar = (int)(vdev->msix_relo - OFF_AUTOPCIBAR_BAR0);
1404     }
1405 
1406     /* I/O port BARs cannot host MSI-X structures */
1407     if (vdev->bars[target_bar].ioport) {
1408         error_setg(errp, "Invalid MSI-X relocation BAR %d, "
1409                    "I/O port BAR", target_bar);
1410         return;
1411     }
1412 
1413     /* Cannot use a BAR in the "shadow" of a 64-bit BAR */
1414     if (!vdev->bars[target_bar].size &&
1415          target_bar > 0 && vdev->bars[target_bar - 1].mem64) {
1416         error_setg(errp, "Invalid MSI-X relocation BAR %d, "
1417                    "consumed by 64-bit BAR %d", target_bar, target_bar - 1);
1418         return;
1419     }
1420 
1421     /* 2GB max size for 32-bit BARs, cannot double if already > 1G */
1422     if (vdev->bars[target_bar].size > 1 * GiB &&
1423         !vdev->bars[target_bar].mem64) {
1424         error_setg(errp, "Invalid MSI-X relocation BAR %d, "
1425                    "no space to extend 32-bit BAR", target_bar);
1426         return;
1427     }
1428 
1429     /*
1430      * If adding a new BAR, test if we can make it 64bit.  We make it
1431      * prefetchable since QEMU MSI-X emulation has no read side effects
1432      * and doing so makes mapping more flexible.
1433      */
1434     if (!vdev->bars[target_bar].size) {
1435         if (target_bar < (PCI_ROM_SLOT - 1) &&
1436             !vdev->bars[target_bar + 1].size) {
1437             vdev->bars[target_bar].mem64 = true;
1438             vdev->bars[target_bar].type = PCI_BASE_ADDRESS_MEM_TYPE_64;
1439         }
1440         vdev->bars[target_bar].type |= PCI_BASE_ADDRESS_MEM_PREFETCH;
1441         vdev->bars[target_bar].size = msix_sz;
1442         vdev->msix->table_offset = 0;
1443     } else {
1444         vdev->bars[target_bar].size = MAX(vdev->bars[target_bar].size * 2,
1445                                           msix_sz * 2);
1446         /*
1447          * Due to above size calc, MSI-X always starts halfway into the BAR,
1448          * which will always be a separate host page.
1449          */
1450         vdev->msix->table_offset = vdev->bars[target_bar].size / 2;
1451     }
1452 
1453     vdev->msix->table_bar = target_bar;
1454     vdev->msix->pba_bar = target_bar;
1455     /* Requires 8-byte alignment, but PCI_MSIX_ENTRY_SIZE guarantees that */
1456     vdev->msix->pba_offset = vdev->msix->table_offset +
1457                                   (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE);
1458 
1459     trace_vfio_msix_relo(vdev->vbasedev.name,
1460                          vdev->msix->table_bar, vdev->msix->table_offset);
1461 }
1462 
1463 /*
1464  * We don't have any control over how pci_add_capability() inserts
1465  * capabilities into the chain.  In order to setup MSI-X we need a
1466  * MemoryRegion for the BAR.  In order to setup the BAR and not
1467  * attempt to mmap the MSI-X table area, which VFIO won't allow, we
1468  * need to first look for where the MSI-X table lives.  So we
1469  * unfortunately split MSI-X setup across two functions.
1470  */
1471 static void vfio_msix_early_setup(VFIOPCIDevice *vdev, Error **errp)
1472 {
1473     uint8_t pos;
1474     uint16_t ctrl;
1475     uint32_t table, pba;
1476     int fd = vdev->vbasedev.fd;
1477     VFIOMSIXInfo *msix;
1478 
1479     pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
1480     if (!pos) {
1481         return;
1482     }
1483 
1484     if (pread(fd, &ctrl, sizeof(ctrl),
1485               vdev->config_offset + pos + PCI_MSIX_FLAGS) != sizeof(ctrl)) {
1486         error_setg_errno(errp, errno, "failed to read PCI MSIX FLAGS");
1487         return;
1488     }
1489 
1490     if (pread(fd, &table, sizeof(table),
1491               vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
1492         error_setg_errno(errp, errno, "failed to read PCI MSIX TABLE");
1493         return;
1494     }
1495 
1496     if (pread(fd, &pba, sizeof(pba),
1497               vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
1498         error_setg_errno(errp, errno, "failed to read PCI MSIX PBA");
1499         return;
1500     }
1501 
1502     ctrl = le16_to_cpu(ctrl);
1503     table = le32_to_cpu(table);
1504     pba = le32_to_cpu(pba);
1505 
1506     msix = g_malloc0(sizeof(*msix));
1507     msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
1508     msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
1509     msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
1510     msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
1511     msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
1512 
1513     /*
1514      * Test the size of the pba_offset variable and catch if it extends outside
1515      * of the specified BAR. If it is the case, we need to apply a hardware
1516      * specific quirk if the device is known or we have a broken configuration.
1517      */
1518     if (msix->pba_offset >= vdev->bars[msix->pba_bar].region.size) {
1519         /*
1520          * Chelsio T5 Virtual Function devices are encoded as 0x58xx for T5
1521          * adapters. The T5 hardware returns an incorrect value of 0x8000 for
1522          * the VF PBA offset while the BAR itself is only 8k. The correct value
1523          * is 0x1000, so we hard code that here.
1524          */
1525         if (vdev->vendor_id == PCI_VENDOR_ID_CHELSIO &&
1526             (vdev->device_id & 0xff00) == 0x5800) {
1527             msix->pba_offset = 0x1000;
1528         } else {
1529             error_setg(errp, "hardware reports invalid configuration, "
1530                        "MSIX PBA outside of specified BAR");
1531             g_free(msix);
1532             return;
1533         }
1534     }
1535 
1536     trace_vfio_msix_early_setup(vdev->vbasedev.name, pos, msix->table_bar,
1537                                 msix->table_offset, msix->entries);
1538     vdev->msix = msix;
1539 
1540     vfio_pci_fixup_msix_region(vdev);
1541 
1542     vfio_pci_relocate_msix(vdev, errp);
1543 }
1544 
1545 static int vfio_msix_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1546 {
1547     int ret;
1548     Error *err = NULL;
1549 
1550     vdev->msix->pending = g_malloc0(BITS_TO_LONGS(vdev->msix->entries) *
1551                                     sizeof(unsigned long));
1552     ret = msix_init(&vdev->pdev, vdev->msix->entries,
1553                     vdev->bars[vdev->msix->table_bar].mr,
1554                     vdev->msix->table_bar, vdev->msix->table_offset,
1555                     vdev->bars[vdev->msix->pba_bar].mr,
1556                     vdev->msix->pba_bar, vdev->msix->pba_offset, pos,
1557                     &err);
1558     if (ret < 0) {
1559         if (ret == -ENOTSUP) {
1560             warn_report_err(err);
1561             return 0;
1562         }
1563 
1564         error_propagate(errp, err);
1565         return ret;
1566     }
1567 
1568     /*
1569      * The PCI spec suggests that devices provide additional alignment for
1570      * MSI-X structures and avoid overlapping non-MSI-X related registers.
1571      * For an assigned device, this hopefully means that emulation of MSI-X
1572      * structures does not affect the performance of the device.  If devices
1573      * fail to provide that alignment, a significant performance penalty may
1574      * result, for instance Mellanox MT27500 VFs:
1575      * http://www.spinics.net/lists/kvm/msg125881.html
1576      *
1577      * The PBA is simply not that important for such a serious regression and
1578      * most drivers do not appear to look at it.  The solution for this is to
1579      * disable the PBA MemoryRegion unless it's being used.  We disable it
1580      * here and only enable it if a masked vector fires through QEMU.  As the
1581      * vector-use notifier is called, which occurs on unmask, we test whether
1582      * PBA emulation is needed and again disable if not.
1583      */
1584     memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
1585 
1586     /*
1587      * The emulated machine may provide a paravirt interface for MSIX setup
1588      * so it is not strictly necessary to emulate MSIX here. This becomes
1589      * helpful when frequently accessed MMIO registers are located in
1590      * subpages adjacent to the MSIX table but the MSIX data containing page
1591      * cannot be mapped because of a host page size bigger than the MSIX table
1592      * alignment.
1593      */
1594     if (object_property_get_bool(OBJECT(qdev_get_machine()),
1595                                  "vfio-no-msix-emulation", NULL)) {
1596         memory_region_set_enabled(&vdev->pdev.msix_table_mmio, false);
1597     }
1598 
1599     return 0;
1600 }
1601 
1602 static void vfio_teardown_msi(VFIOPCIDevice *vdev)
1603 {
1604     msi_uninit(&vdev->pdev);
1605 
1606     if (vdev->msix) {
1607         msix_uninit(&vdev->pdev,
1608                     vdev->bars[vdev->msix->table_bar].mr,
1609                     vdev->bars[vdev->msix->pba_bar].mr);
1610         g_free(vdev->msix->pending);
1611     }
1612 }
1613 
1614 /*
1615  * Resource setup
1616  */
1617 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled)
1618 {
1619     int i;
1620 
1621     for (i = 0; i < PCI_ROM_SLOT; i++) {
1622         vfio_region_mmaps_set_enabled(&vdev->bars[i].region, enabled);
1623     }
1624 }
1625 
1626 static void vfio_bar_prepare(VFIOPCIDevice *vdev, int nr)
1627 {
1628     VFIOBAR *bar = &vdev->bars[nr];
1629 
1630     uint32_t pci_bar;
1631     int ret;
1632 
1633     /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1634     if (!bar->region.size) {
1635         return;
1636     }
1637 
1638     /* Determine what type of BAR this is for registration */
1639     ret = pread(vdev->vbasedev.fd, &pci_bar, sizeof(pci_bar),
1640                 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
1641     if (ret != sizeof(pci_bar)) {
1642         error_report("vfio: Failed to read BAR %d (%m)", nr);
1643         return;
1644     }
1645 
1646     pci_bar = le32_to_cpu(pci_bar);
1647     bar->ioport = (pci_bar & PCI_BASE_ADDRESS_SPACE_IO);
1648     bar->mem64 = bar->ioport ? 0 : (pci_bar & PCI_BASE_ADDRESS_MEM_TYPE_64);
1649     bar->type = pci_bar & (bar->ioport ? ~PCI_BASE_ADDRESS_IO_MASK :
1650                                          ~PCI_BASE_ADDRESS_MEM_MASK);
1651     bar->size = bar->region.size;
1652 }
1653 
1654 static void vfio_bars_prepare(VFIOPCIDevice *vdev)
1655 {
1656     int i;
1657 
1658     for (i = 0; i < PCI_ROM_SLOT; i++) {
1659         vfio_bar_prepare(vdev, i);
1660     }
1661 }
1662 
1663 static void vfio_bar_register(VFIOPCIDevice *vdev, int nr)
1664 {
1665     VFIOBAR *bar = &vdev->bars[nr];
1666     char *name;
1667 
1668     if (!bar->size) {
1669         return;
1670     }
1671 
1672     bar->mr = g_new0(MemoryRegion, 1);
1673     name = g_strdup_printf("%s base BAR %d", vdev->vbasedev.name, nr);
1674     memory_region_init_io(bar->mr, OBJECT(vdev), NULL, NULL, name, bar->size);
1675     g_free(name);
1676 
1677     if (bar->region.size) {
1678         memory_region_add_subregion(bar->mr, 0, bar->region.mem);
1679 
1680         if (vfio_region_mmap(&bar->region)) {
1681             error_report("Failed to mmap %s BAR %d. Performance may be slow",
1682                          vdev->vbasedev.name, nr);
1683         }
1684     }
1685 
1686     pci_register_bar(&vdev->pdev, nr, bar->type, bar->mr);
1687 }
1688 
1689 static void vfio_bars_register(VFIOPCIDevice *vdev)
1690 {
1691     int i;
1692 
1693     for (i = 0; i < PCI_ROM_SLOT; i++) {
1694         vfio_bar_register(vdev, i);
1695     }
1696 }
1697 
1698 static void vfio_bars_exit(VFIOPCIDevice *vdev)
1699 {
1700     int i;
1701 
1702     for (i = 0; i < PCI_ROM_SLOT; i++) {
1703         VFIOBAR *bar = &vdev->bars[i];
1704 
1705         vfio_bar_quirk_exit(vdev, i);
1706         vfio_region_exit(&bar->region);
1707         if (bar->region.size) {
1708             memory_region_del_subregion(bar->mr, bar->region.mem);
1709         }
1710     }
1711 
1712     if (vdev->vga) {
1713         pci_unregister_vga(&vdev->pdev);
1714         vfio_vga_quirk_exit(vdev);
1715     }
1716 }
1717 
1718 static void vfio_bars_finalize(VFIOPCIDevice *vdev)
1719 {
1720     int i;
1721 
1722     for (i = 0; i < PCI_ROM_SLOT; i++) {
1723         VFIOBAR *bar = &vdev->bars[i];
1724 
1725         vfio_bar_quirk_finalize(vdev, i);
1726         vfio_region_finalize(&bar->region);
1727         if (bar->size) {
1728             object_unparent(OBJECT(bar->mr));
1729             g_free(bar->mr);
1730         }
1731     }
1732 
1733     if (vdev->vga) {
1734         vfio_vga_quirk_finalize(vdev);
1735         for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) {
1736             object_unparent(OBJECT(&vdev->vga->region[i].mem));
1737         }
1738         g_free(vdev->vga);
1739     }
1740 }
1741 
1742 /*
1743  * General setup
1744  */
1745 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
1746 {
1747     uint8_t tmp;
1748     uint16_t next = PCI_CONFIG_SPACE_SIZE;
1749 
1750     for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
1751          tmp = pdev->config[tmp + PCI_CAP_LIST_NEXT]) {
1752         if (tmp > pos && tmp < next) {
1753             next = tmp;
1754         }
1755     }
1756 
1757     return next - pos;
1758 }
1759 
1760 
1761 static uint16_t vfio_ext_cap_max_size(const uint8_t *config, uint16_t pos)
1762 {
1763     uint16_t tmp, next = PCIE_CONFIG_SPACE_SIZE;
1764 
1765     for (tmp = PCI_CONFIG_SPACE_SIZE; tmp;
1766         tmp = PCI_EXT_CAP_NEXT(pci_get_long(config + tmp))) {
1767         if (tmp > pos && tmp < next) {
1768             next = tmp;
1769         }
1770     }
1771 
1772     return next - pos;
1773 }
1774 
1775 static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask)
1776 {
1777     pci_set_word(buf, (pci_get_word(buf) & ~mask) | val);
1778 }
1779 
1780 static void vfio_add_emulated_word(VFIOPCIDevice *vdev, int pos,
1781                                    uint16_t val, uint16_t mask)
1782 {
1783     vfio_set_word_bits(vdev->pdev.config + pos, val, mask);
1784     vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask);
1785     vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask);
1786 }
1787 
1788 static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask)
1789 {
1790     pci_set_long(buf, (pci_get_long(buf) & ~mask) | val);
1791 }
1792 
1793 static void vfio_add_emulated_long(VFIOPCIDevice *vdev, int pos,
1794                                    uint32_t val, uint32_t mask)
1795 {
1796     vfio_set_long_bits(vdev->pdev.config + pos, val, mask);
1797     vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask);
1798     vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask);
1799 }
1800 
1801 static int vfio_setup_pcie_cap(VFIOPCIDevice *vdev, int pos, uint8_t size,
1802                                Error **errp)
1803 {
1804     uint16_t flags;
1805     uint8_t type;
1806 
1807     flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS);
1808     type = (flags & PCI_EXP_FLAGS_TYPE) >> 4;
1809 
1810     if (type != PCI_EXP_TYPE_ENDPOINT &&
1811         type != PCI_EXP_TYPE_LEG_END &&
1812         type != PCI_EXP_TYPE_RC_END) {
1813 
1814         error_setg(errp, "assignment of PCIe type 0x%x "
1815                    "devices is not currently supported", type);
1816         return -EINVAL;
1817     }
1818 
1819     if (!pci_bus_is_express(pci_get_bus(&vdev->pdev))) {
1820         PCIBus *bus = pci_get_bus(&vdev->pdev);
1821         PCIDevice *bridge;
1822 
1823         /*
1824          * Traditionally PCI device assignment exposes the PCIe capability
1825          * as-is on non-express buses.  The reason being that some drivers
1826          * simply assume that it's there, for example tg3.  However when
1827          * we're running on a native PCIe machine type, like Q35, we need
1828          * to hide the PCIe capability.  The reason for this is twofold;
1829          * first Windows guests get a Code 10 error when the PCIe capability
1830          * is exposed in this configuration.  Therefore express devices won't
1831          * work at all unless they're attached to express buses in the VM.
1832          * Second, a native PCIe machine introduces the possibility of fine
1833          * granularity IOMMUs supporting both translation and isolation.
1834          * Guest code to discover the IOMMU visibility of a device, such as
1835          * IOMMU grouping code on Linux, is very aware of device types and
1836          * valid transitions between bus types.  An express device on a non-
1837          * express bus is not a valid combination on bare metal systems.
1838          *
1839          * Drivers that require a PCIe capability to make the device
1840          * functional are simply going to need to have their devices placed
1841          * on a PCIe bus in the VM.
1842          */
1843         while (!pci_bus_is_root(bus)) {
1844             bridge = pci_bridge_get_device(bus);
1845             bus = pci_get_bus(bridge);
1846         }
1847 
1848         if (pci_bus_is_express(bus)) {
1849             return 0;
1850         }
1851 
1852     } else if (pci_bus_is_root(pci_get_bus(&vdev->pdev))) {
1853         /*
1854          * On a Root Complex bus Endpoints become Root Complex Integrated
1855          * Endpoints, which changes the type and clears the LNK & LNK2 fields.
1856          */
1857         if (type == PCI_EXP_TYPE_ENDPOINT) {
1858             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1859                                    PCI_EXP_TYPE_RC_END << 4,
1860                                    PCI_EXP_FLAGS_TYPE);
1861 
1862             /* Link Capabilities, Status, and Control goes away */
1863             if (size > PCI_EXP_LNKCTL) {
1864                 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0);
1865                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1866                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0);
1867 
1868 #ifndef PCI_EXP_LNKCAP2
1869 #define PCI_EXP_LNKCAP2 44
1870 #endif
1871 #ifndef PCI_EXP_LNKSTA2
1872 #define PCI_EXP_LNKSTA2 50
1873 #endif
1874                 /* Link 2 Capabilities, Status, and Control goes away */
1875                 if (size > PCI_EXP_LNKCAP2) {
1876                     vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0);
1877                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0);
1878                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0);
1879                 }
1880             }
1881 
1882         } else if (type == PCI_EXP_TYPE_LEG_END) {
1883             /*
1884              * Legacy endpoints don't belong on the root complex.  Windows
1885              * seems to be happier with devices if we skip the capability.
1886              */
1887             return 0;
1888         }
1889 
1890     } else {
1891         /*
1892          * Convert Root Complex Integrated Endpoints to regular endpoints.
1893          * These devices don't support LNK/LNK2 capabilities, so make them up.
1894          */
1895         if (type == PCI_EXP_TYPE_RC_END) {
1896             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1897                                    PCI_EXP_TYPE_ENDPOINT << 4,
1898                                    PCI_EXP_FLAGS_TYPE);
1899             vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP,
1900                                    PCI_EXP_LNK_MLW_1 | PCI_EXP_LNK_LS_25, ~0);
1901             vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1902         }
1903 
1904         /* Mark the Link Status bits as emulated to allow virtual negotiation */
1905         vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA,
1906                                pci_get_word(vdev->pdev.config + pos +
1907                                             PCI_EXP_LNKSTA),
1908                                PCI_EXP_LNKCAP_MLW | PCI_EXP_LNKCAP_SLS);
1909     }
1910 
1911     /*
1912      * Intel 82599 SR-IOV VFs report an invalid PCIe capability version 0
1913      * (Niantic errate #35) causing Windows to error with a Code 10 for the
1914      * device on Q35.  Fixup any such devices to report version 1.  If we
1915      * were to remove the capability entirely the guest would lose extended
1916      * config space.
1917      */
1918     if ((flags & PCI_EXP_FLAGS_VERS) == 0) {
1919         vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1920                                1, PCI_EXP_FLAGS_VERS);
1921     }
1922 
1923     pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size,
1924                              errp);
1925     if (pos < 0) {
1926         return pos;
1927     }
1928 
1929     vdev->pdev.exp.exp_cap = pos;
1930 
1931     return pos;
1932 }
1933 
1934 static void vfio_check_pcie_flr(VFIOPCIDevice *vdev, uint8_t pos)
1935 {
1936     uint32_t cap = pci_get_long(vdev->pdev.config + pos + PCI_EXP_DEVCAP);
1937 
1938     if (cap & PCI_EXP_DEVCAP_FLR) {
1939         trace_vfio_check_pcie_flr(vdev->vbasedev.name);
1940         vdev->has_flr = true;
1941     }
1942 }
1943 
1944 static void vfio_check_pm_reset(VFIOPCIDevice *vdev, uint8_t pos)
1945 {
1946     uint16_t csr = pci_get_word(vdev->pdev.config + pos + PCI_PM_CTRL);
1947 
1948     if (!(csr & PCI_PM_CTRL_NO_SOFT_RESET)) {
1949         trace_vfio_check_pm_reset(vdev->vbasedev.name);
1950         vdev->has_pm_reset = true;
1951     }
1952 }
1953 
1954 static void vfio_check_af_flr(VFIOPCIDevice *vdev, uint8_t pos)
1955 {
1956     uint8_t cap = pci_get_byte(vdev->pdev.config + pos + PCI_AF_CAP);
1957 
1958     if ((cap & PCI_AF_CAP_TP) && (cap & PCI_AF_CAP_FLR)) {
1959         trace_vfio_check_af_flr(vdev->vbasedev.name);
1960         vdev->has_flr = true;
1961     }
1962 }
1963 
1964 static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos, Error **errp)
1965 {
1966     PCIDevice *pdev = &vdev->pdev;
1967     uint8_t cap_id, next, size;
1968     int ret;
1969 
1970     cap_id = pdev->config[pos];
1971     next = pdev->config[pos + PCI_CAP_LIST_NEXT];
1972 
1973     /*
1974      * If it becomes important to configure capabilities to their actual
1975      * size, use this as the default when it's something we don't recognize.
1976      * Since QEMU doesn't actually handle many of the config accesses,
1977      * exact size doesn't seem worthwhile.
1978      */
1979     size = vfio_std_cap_max_size(pdev, pos);
1980 
1981     /*
1982      * pci_add_capability always inserts the new capability at the head
1983      * of the chain.  Therefore to end up with a chain that matches the
1984      * physical device, we insert from the end by making this recursive.
1985      * This is also why we pre-calculate size above as cached config space
1986      * will be changed as we unwind the stack.
1987      */
1988     if (next) {
1989         ret = vfio_add_std_cap(vdev, next, errp);
1990         if (ret) {
1991             return ret;
1992         }
1993     } else {
1994         /* Begin the rebuild, use QEMU emulated list bits */
1995         pdev->config[PCI_CAPABILITY_LIST] = 0;
1996         vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
1997         vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
1998 
1999         ret = vfio_add_virt_caps(vdev, errp);
2000         if (ret) {
2001             return ret;
2002         }
2003     }
2004 
2005     /* Scale down size, esp in case virt caps were added above */
2006     size = MIN(size, vfio_std_cap_max_size(pdev, pos));
2007 
2008     /* Use emulated next pointer to allow dropping caps */
2009     pci_set_byte(vdev->emulated_config_bits + pos + PCI_CAP_LIST_NEXT, 0xff);
2010 
2011     switch (cap_id) {
2012     case PCI_CAP_ID_MSI:
2013         ret = vfio_msi_setup(vdev, pos, errp);
2014         break;
2015     case PCI_CAP_ID_EXP:
2016         vfio_check_pcie_flr(vdev, pos);
2017         ret = vfio_setup_pcie_cap(vdev, pos, size, errp);
2018         break;
2019     case PCI_CAP_ID_MSIX:
2020         ret = vfio_msix_setup(vdev, pos, errp);
2021         break;
2022     case PCI_CAP_ID_PM:
2023         vfio_check_pm_reset(vdev, pos);
2024         vdev->pm_cap = pos;
2025         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
2026         break;
2027     case PCI_CAP_ID_AF:
2028         vfio_check_af_flr(vdev, pos);
2029         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
2030         break;
2031     default:
2032         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
2033         break;
2034     }
2035 
2036     if (ret < 0) {
2037         error_prepend(errp,
2038                       "failed to add PCI capability 0x%x[0x%x]@0x%x: ",
2039                       cap_id, size, pos);
2040         return ret;
2041     }
2042 
2043     return 0;
2044 }
2045 
2046 static void vfio_add_ext_cap(VFIOPCIDevice *vdev)
2047 {
2048     PCIDevice *pdev = &vdev->pdev;
2049     uint32_t header;
2050     uint16_t cap_id, next, size;
2051     uint8_t cap_ver;
2052     uint8_t *config;
2053 
2054     /* Only add extended caps if we have them and the guest can see them */
2055     if (!pci_is_express(pdev) || !pci_bus_is_express(pci_get_bus(pdev)) ||
2056         !pci_get_long(pdev->config + PCI_CONFIG_SPACE_SIZE)) {
2057         return;
2058     }
2059 
2060     /*
2061      * pcie_add_capability always inserts the new capability at the tail
2062      * of the chain.  Therefore to end up with a chain that matches the
2063      * physical device, we cache the config space to avoid overwriting
2064      * the original config space when we parse the extended capabilities.
2065      */
2066     config = g_memdup(pdev->config, vdev->config_size);
2067 
2068     /*
2069      * Extended capabilities are chained with each pointing to the next, so we
2070      * can drop anything other than the head of the chain simply by modifying
2071      * the previous next pointer.  Seed the head of the chain here such that
2072      * we can simply skip any capabilities we want to drop below, regardless
2073      * of their position in the chain.  If this stub capability still exists
2074      * after we add the capabilities we want to expose, update the capability
2075      * ID to zero.  Note that we cannot seed with the capability header being
2076      * zero as this conflicts with definition of an absent capability chain
2077      * and prevents capabilities beyond the head of the list from being added.
2078      * By replacing the dummy capability ID with zero after walking the device
2079      * chain, we also transparently mark extended capabilities as absent if
2080      * no capabilities were added.  Note that the PCIe spec defines an absence
2081      * of extended capabilities to be determined by a value of zero for the
2082      * capability ID, version, AND next pointer.  A non-zero next pointer
2083      * should be sufficient to indicate additional capabilities are present,
2084      * which will occur if we call pcie_add_capability() below.  The entire
2085      * first dword is emulated to support this.
2086      *
2087      * NB. The kernel side does similar masking, so be prepared that our
2088      * view of the device may also contain a capability ID zero in the head
2089      * of the chain.  Skip it for the same reason that we cannot seed the
2090      * chain with a zero capability.
2091      */
2092     pci_set_long(pdev->config + PCI_CONFIG_SPACE_SIZE,
2093                  PCI_EXT_CAP(0xFFFF, 0, 0));
2094     pci_set_long(pdev->wmask + PCI_CONFIG_SPACE_SIZE, 0);
2095     pci_set_long(vdev->emulated_config_bits + PCI_CONFIG_SPACE_SIZE, ~0);
2096 
2097     for (next = PCI_CONFIG_SPACE_SIZE; next;
2098          next = PCI_EXT_CAP_NEXT(pci_get_long(config + next))) {
2099         header = pci_get_long(config + next);
2100         cap_id = PCI_EXT_CAP_ID(header);
2101         cap_ver = PCI_EXT_CAP_VER(header);
2102 
2103         /*
2104          * If it becomes important to configure extended capabilities to their
2105          * actual size, use this as the default when it's something we don't
2106          * recognize. Since QEMU doesn't actually handle many of the config
2107          * accesses, exact size doesn't seem worthwhile.
2108          */
2109         size = vfio_ext_cap_max_size(config, next);
2110 
2111         /* Use emulated next pointer to allow dropping extended caps */
2112         pci_long_test_and_set_mask(vdev->emulated_config_bits + next,
2113                                    PCI_EXT_CAP_NEXT_MASK);
2114 
2115         switch (cap_id) {
2116         case 0: /* kernel masked capability */
2117         case PCI_EXT_CAP_ID_SRIOV: /* Read-only VF BARs confuse OVMF */
2118         case PCI_EXT_CAP_ID_ARI: /* XXX Needs next function virtualization */
2119             trace_vfio_add_ext_cap_dropped(vdev->vbasedev.name, cap_id, next);
2120             break;
2121         default:
2122             pcie_add_capability(pdev, cap_id, cap_ver, next, size);
2123         }
2124 
2125     }
2126 
2127     /* Cleanup chain head ID if necessary */
2128     if (pci_get_word(pdev->config + PCI_CONFIG_SPACE_SIZE) == 0xFFFF) {
2129         pci_set_word(pdev->config + PCI_CONFIG_SPACE_SIZE, 0);
2130     }
2131 
2132     g_free(config);
2133     return;
2134 }
2135 
2136 static int vfio_add_capabilities(VFIOPCIDevice *vdev, Error **errp)
2137 {
2138     PCIDevice *pdev = &vdev->pdev;
2139     int ret;
2140 
2141     if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
2142         !pdev->config[PCI_CAPABILITY_LIST]) {
2143         return 0; /* Nothing to add */
2144     }
2145 
2146     ret = vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST], errp);
2147     if (ret) {
2148         return ret;
2149     }
2150 
2151     vfio_add_ext_cap(vdev);
2152     return 0;
2153 }
2154 
2155 static void vfio_pci_pre_reset(VFIOPCIDevice *vdev)
2156 {
2157     PCIDevice *pdev = &vdev->pdev;
2158     uint16_t cmd;
2159 
2160     vfio_disable_interrupts(vdev);
2161 
2162     /* Make sure the device is in D0 */
2163     if (vdev->pm_cap) {
2164         uint16_t pmcsr;
2165         uint8_t state;
2166 
2167         pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2168         state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2169         if (state) {
2170             pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
2171             vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2);
2172             /* vfio handles the necessary delay here */
2173             pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2174             state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2175             if (state) {
2176                 error_report("vfio: Unable to power on device, stuck in D%d",
2177                              state);
2178             }
2179         }
2180     }
2181 
2182     /*
2183      * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
2184      * Also put INTx Disable in known state.
2185      */
2186     cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
2187     cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
2188              PCI_COMMAND_INTX_DISABLE);
2189     vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
2190 }
2191 
2192 static void vfio_pci_post_reset(VFIOPCIDevice *vdev)
2193 {
2194     Error *err = NULL;
2195     int nr;
2196 
2197     vfio_intx_enable(vdev, &err);
2198     if (err) {
2199         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
2200     }
2201 
2202     for (nr = 0; nr < PCI_NUM_REGIONS - 1; ++nr) {
2203         off_t addr = vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr);
2204         uint32_t val = 0;
2205         uint32_t len = sizeof(val);
2206 
2207         if (pwrite(vdev->vbasedev.fd, &val, len, addr) != len) {
2208             error_report("%s(%s) reset bar %d failed: %m", __func__,
2209                          vdev->vbasedev.name, nr);
2210         }
2211     }
2212 
2213     vfio_quirk_reset(vdev);
2214 }
2215 
2216 static bool vfio_pci_host_match(PCIHostDeviceAddress *addr, const char *name)
2217 {
2218     char tmp[13];
2219 
2220     sprintf(tmp, "%04x:%02x:%02x.%1x", addr->domain,
2221             addr->bus, addr->slot, addr->function);
2222 
2223     return (strcmp(tmp, name) == 0);
2224 }
2225 
2226 static int vfio_pci_hot_reset(VFIOPCIDevice *vdev, bool single)
2227 {
2228     VFIOGroup *group;
2229     struct vfio_pci_hot_reset_info *info;
2230     struct vfio_pci_dependent_device *devices;
2231     struct vfio_pci_hot_reset *reset;
2232     int32_t *fds;
2233     int ret, i, count;
2234     bool multi = false;
2235 
2236     trace_vfio_pci_hot_reset(vdev->vbasedev.name, single ? "one" : "multi");
2237 
2238     if (!single) {
2239         vfio_pci_pre_reset(vdev);
2240     }
2241     vdev->vbasedev.needs_reset = false;
2242 
2243     info = g_malloc0(sizeof(*info));
2244     info->argsz = sizeof(*info);
2245 
2246     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2247     if (ret && errno != ENOSPC) {
2248         ret = -errno;
2249         if (!vdev->has_pm_reset) {
2250             error_report("vfio: Cannot reset device %s, "
2251                          "no available reset mechanism.", vdev->vbasedev.name);
2252         }
2253         goto out_single;
2254     }
2255 
2256     count = info->count;
2257     info = g_realloc(info, sizeof(*info) + (count * sizeof(*devices)));
2258     info->argsz = sizeof(*info) + (count * sizeof(*devices));
2259     devices = &info->devices[0];
2260 
2261     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2262     if (ret) {
2263         ret = -errno;
2264         error_report("vfio: hot reset info failed: %m");
2265         goto out_single;
2266     }
2267 
2268     trace_vfio_pci_hot_reset_has_dep_devices(vdev->vbasedev.name);
2269 
2270     /* Verify that we have all the groups required */
2271     for (i = 0; i < info->count; i++) {
2272         PCIHostDeviceAddress host;
2273         VFIOPCIDevice *tmp;
2274         VFIODevice *vbasedev_iter;
2275 
2276         host.domain = devices[i].segment;
2277         host.bus = devices[i].bus;
2278         host.slot = PCI_SLOT(devices[i].devfn);
2279         host.function = PCI_FUNC(devices[i].devfn);
2280 
2281         trace_vfio_pci_hot_reset_dep_devices(host.domain,
2282                 host.bus, host.slot, host.function, devices[i].group_id);
2283 
2284         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2285             continue;
2286         }
2287 
2288         QLIST_FOREACH(group, &vfio_group_list, next) {
2289             if (group->groupid == devices[i].group_id) {
2290                 break;
2291             }
2292         }
2293 
2294         if (!group) {
2295             if (!vdev->has_pm_reset) {
2296                 error_report("vfio: Cannot reset device %s, "
2297                              "depends on group %d which is not owned.",
2298                              vdev->vbasedev.name, devices[i].group_id);
2299             }
2300             ret = -EPERM;
2301             goto out;
2302         }
2303 
2304         /* Prep dependent devices for reset and clear our marker. */
2305         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2306             if (!vbasedev_iter->dev->realized ||
2307                 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2308                 continue;
2309             }
2310             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2311             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2312                 if (single) {
2313                     ret = -EINVAL;
2314                     goto out_single;
2315                 }
2316                 vfio_pci_pre_reset(tmp);
2317                 tmp->vbasedev.needs_reset = false;
2318                 multi = true;
2319                 break;
2320             }
2321         }
2322     }
2323 
2324     if (!single && !multi) {
2325         ret = -EINVAL;
2326         goto out_single;
2327     }
2328 
2329     /* Determine how many group fds need to be passed */
2330     count = 0;
2331     QLIST_FOREACH(group, &vfio_group_list, next) {
2332         for (i = 0; i < info->count; i++) {
2333             if (group->groupid == devices[i].group_id) {
2334                 count++;
2335                 break;
2336             }
2337         }
2338     }
2339 
2340     reset = g_malloc0(sizeof(*reset) + (count * sizeof(*fds)));
2341     reset->argsz = sizeof(*reset) + (count * sizeof(*fds));
2342     fds = &reset->group_fds[0];
2343 
2344     /* Fill in group fds */
2345     QLIST_FOREACH(group, &vfio_group_list, next) {
2346         for (i = 0; i < info->count; i++) {
2347             if (group->groupid == devices[i].group_id) {
2348                 fds[reset->count++] = group->fd;
2349                 break;
2350             }
2351         }
2352     }
2353 
2354     /* Bus reset! */
2355     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_PCI_HOT_RESET, reset);
2356     g_free(reset);
2357 
2358     trace_vfio_pci_hot_reset_result(vdev->vbasedev.name,
2359                                     ret ? "%m" : "Success");
2360 
2361 out:
2362     /* Re-enable INTx on affected devices */
2363     for (i = 0; i < info->count; i++) {
2364         PCIHostDeviceAddress host;
2365         VFIOPCIDevice *tmp;
2366         VFIODevice *vbasedev_iter;
2367 
2368         host.domain = devices[i].segment;
2369         host.bus = devices[i].bus;
2370         host.slot = PCI_SLOT(devices[i].devfn);
2371         host.function = PCI_FUNC(devices[i].devfn);
2372 
2373         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2374             continue;
2375         }
2376 
2377         QLIST_FOREACH(group, &vfio_group_list, next) {
2378             if (group->groupid == devices[i].group_id) {
2379                 break;
2380             }
2381         }
2382 
2383         if (!group) {
2384             break;
2385         }
2386 
2387         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2388             if (!vbasedev_iter->dev->realized ||
2389                 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2390                 continue;
2391             }
2392             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2393             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2394                 vfio_pci_post_reset(tmp);
2395                 break;
2396             }
2397         }
2398     }
2399 out_single:
2400     if (!single) {
2401         vfio_pci_post_reset(vdev);
2402     }
2403     g_free(info);
2404 
2405     return ret;
2406 }
2407 
2408 /*
2409  * We want to differentiate hot reset of mulitple in-use devices vs hot reset
2410  * of a single in-use device.  VFIO_DEVICE_RESET will already handle the case
2411  * of doing hot resets when there is only a single device per bus.  The in-use
2412  * here refers to how many VFIODevices are affected.  A hot reset that affects
2413  * multiple devices, but only a single in-use device, means that we can call
2414  * it from our bus ->reset() callback since the extent is effectively a single
2415  * device.  This allows us to make use of it in the hotplug path.  When there
2416  * are multiple in-use devices, we can only trigger the hot reset during a
2417  * system reset and thus from our reset handler.  We separate _one vs _multi
2418  * here so that we don't overlap and do a double reset on the system reset
2419  * path where both our reset handler and ->reset() callback are used.  Calling
2420  * _one() will only do a hot reset for the one in-use devices case, calling
2421  * _multi() will do nothing if a _one() would have been sufficient.
2422  */
2423 static int vfio_pci_hot_reset_one(VFIOPCIDevice *vdev)
2424 {
2425     return vfio_pci_hot_reset(vdev, true);
2426 }
2427 
2428 static int vfio_pci_hot_reset_multi(VFIODevice *vbasedev)
2429 {
2430     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2431     return vfio_pci_hot_reset(vdev, false);
2432 }
2433 
2434 static void vfio_pci_compute_needs_reset(VFIODevice *vbasedev)
2435 {
2436     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2437     if (!vbasedev->reset_works || (!vdev->has_flr && vdev->has_pm_reset)) {
2438         vbasedev->needs_reset = true;
2439     }
2440 }
2441 
2442 static VFIODeviceOps vfio_pci_ops = {
2443     .vfio_compute_needs_reset = vfio_pci_compute_needs_reset,
2444     .vfio_hot_reset_multi = vfio_pci_hot_reset_multi,
2445     .vfio_eoi = vfio_intx_eoi,
2446 };
2447 
2448 int vfio_populate_vga(VFIOPCIDevice *vdev, Error **errp)
2449 {
2450     VFIODevice *vbasedev = &vdev->vbasedev;
2451     struct vfio_region_info *reg_info;
2452     int ret;
2453 
2454     ret = vfio_get_region_info(vbasedev, VFIO_PCI_VGA_REGION_INDEX, &reg_info);
2455     if (ret) {
2456         error_setg_errno(errp, -ret,
2457                          "failed getting region info for VGA region index %d",
2458                          VFIO_PCI_VGA_REGION_INDEX);
2459         return ret;
2460     }
2461 
2462     if (!(reg_info->flags & VFIO_REGION_INFO_FLAG_READ) ||
2463         !(reg_info->flags & VFIO_REGION_INFO_FLAG_WRITE) ||
2464         reg_info->size < 0xbffff + 1) {
2465         error_setg(errp, "unexpected VGA info, flags 0x%lx, size 0x%lx",
2466                    (unsigned long)reg_info->flags,
2467                    (unsigned long)reg_info->size);
2468         g_free(reg_info);
2469         return -EINVAL;
2470     }
2471 
2472     vdev->vga = g_new0(VFIOVGA, 1);
2473 
2474     vdev->vga->fd_offset = reg_info->offset;
2475     vdev->vga->fd = vdev->vbasedev.fd;
2476 
2477     g_free(reg_info);
2478 
2479     vdev->vga->region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE;
2480     vdev->vga->region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM;
2481     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_MEM].quirks);
2482 
2483     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2484                           OBJECT(vdev), &vfio_vga_ops,
2485                           &vdev->vga->region[QEMU_PCI_VGA_MEM],
2486                           "vfio-vga-mmio@0xa0000",
2487                           QEMU_PCI_VGA_MEM_SIZE);
2488 
2489     vdev->vga->region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE;
2490     vdev->vga->region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO;
2491     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].quirks);
2492 
2493     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2494                           OBJECT(vdev), &vfio_vga_ops,
2495                           &vdev->vga->region[QEMU_PCI_VGA_IO_LO],
2496                           "vfio-vga-io@0x3b0",
2497                           QEMU_PCI_VGA_IO_LO_SIZE);
2498 
2499     vdev->vga->region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE;
2500     vdev->vga->region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI;
2501     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks);
2502 
2503     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
2504                           OBJECT(vdev), &vfio_vga_ops,
2505                           &vdev->vga->region[QEMU_PCI_VGA_IO_HI],
2506                           "vfio-vga-io@0x3c0",
2507                           QEMU_PCI_VGA_IO_HI_SIZE);
2508 
2509     pci_register_vga(&vdev->pdev, &vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2510                      &vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2511                      &vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem);
2512 
2513     return 0;
2514 }
2515 
2516 static void vfio_populate_device(VFIOPCIDevice *vdev, Error **errp)
2517 {
2518     VFIODevice *vbasedev = &vdev->vbasedev;
2519     struct vfio_region_info *reg_info;
2520     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) };
2521     int i, ret = -1;
2522 
2523     /* Sanity check device */
2524     if (!(vbasedev->flags & VFIO_DEVICE_FLAGS_PCI)) {
2525         error_setg(errp, "this isn't a PCI device");
2526         return;
2527     }
2528 
2529     if (vbasedev->num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
2530         error_setg(errp, "unexpected number of io regions %u",
2531                    vbasedev->num_regions);
2532         return;
2533     }
2534 
2535     if (vbasedev->num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
2536         error_setg(errp, "unexpected number of irqs %u", vbasedev->num_irqs);
2537         return;
2538     }
2539 
2540     for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
2541         char *name = g_strdup_printf("%s BAR %d", vbasedev->name, i);
2542 
2543         ret = vfio_region_setup(OBJECT(vdev), vbasedev,
2544                                 &vdev->bars[i].region, i, name);
2545         g_free(name);
2546 
2547         if (ret) {
2548             error_setg_errno(errp, -ret, "failed to get region %d info", i);
2549             return;
2550         }
2551 
2552         QLIST_INIT(&vdev->bars[i].quirks);
2553     }
2554 
2555     ret = vfio_get_region_info(vbasedev,
2556                                VFIO_PCI_CONFIG_REGION_INDEX, &reg_info);
2557     if (ret) {
2558         error_setg_errno(errp, -ret, "failed to get config info");
2559         return;
2560     }
2561 
2562     trace_vfio_populate_device_config(vdev->vbasedev.name,
2563                                       (unsigned long)reg_info->size,
2564                                       (unsigned long)reg_info->offset,
2565                                       (unsigned long)reg_info->flags);
2566 
2567     vdev->config_size = reg_info->size;
2568     if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
2569         vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
2570     }
2571     vdev->config_offset = reg_info->offset;
2572 
2573     g_free(reg_info);
2574 
2575     if (vdev->features & VFIO_FEATURE_ENABLE_VGA) {
2576         ret = vfio_populate_vga(vdev, errp);
2577         if (ret) {
2578             error_append_hint(errp, "device does not support "
2579                               "requested feature x-vga\n");
2580             return;
2581         }
2582     }
2583 
2584     irq_info.index = VFIO_PCI_ERR_IRQ_INDEX;
2585 
2586     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info);
2587     if (ret) {
2588         /* This can fail for an old kernel or legacy PCI dev */
2589         trace_vfio_populate_device_get_irq_info_failure();
2590     } else if (irq_info.count == 1) {
2591         vdev->pci_aer = true;
2592     } else {
2593         warn_report(VFIO_MSG_PREFIX
2594                     "Could not enable error recovery for the device",
2595                     vbasedev->name);
2596     }
2597 }
2598 
2599 static void vfio_put_device(VFIOPCIDevice *vdev)
2600 {
2601     g_free(vdev->vbasedev.name);
2602     g_free(vdev->msix);
2603 
2604     vfio_put_base_device(&vdev->vbasedev);
2605 }
2606 
2607 static void vfio_err_notifier_handler(void *opaque)
2608 {
2609     VFIOPCIDevice *vdev = opaque;
2610 
2611     if (!event_notifier_test_and_clear(&vdev->err_notifier)) {
2612         return;
2613     }
2614 
2615     /*
2616      * TBD. Retrieve the error details and decide what action
2617      * needs to be taken. One of the actions could be to pass
2618      * the error to the guest and have the guest driver recover
2619      * from the error. This requires that PCIe capabilities be
2620      * exposed to the guest. For now, we just terminate the
2621      * guest to contain the error.
2622      */
2623 
2624     error_report("%s(%s) Unrecoverable error detected. Please collect any data possible and then kill the guest", __func__, vdev->vbasedev.name);
2625 
2626     vm_stop(RUN_STATE_INTERNAL_ERROR);
2627 }
2628 
2629 /*
2630  * Registers error notifier for devices supporting error recovery.
2631  * If we encounter a failure in this function, we report an error
2632  * and continue after disabling error recovery support for the
2633  * device.
2634  */
2635 static void vfio_register_err_notifier(VFIOPCIDevice *vdev)
2636 {
2637     int ret;
2638     int argsz;
2639     struct vfio_irq_set *irq_set;
2640     int32_t *pfd;
2641 
2642     if (!vdev->pci_aer) {
2643         return;
2644     }
2645 
2646     if (event_notifier_init(&vdev->err_notifier, 0)) {
2647         error_report("vfio: Unable to init event notifier for error detection");
2648         vdev->pci_aer = false;
2649         return;
2650     }
2651 
2652     argsz = sizeof(*irq_set) + sizeof(*pfd);
2653 
2654     irq_set = g_malloc0(argsz);
2655     irq_set->argsz = argsz;
2656     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2657                      VFIO_IRQ_SET_ACTION_TRIGGER;
2658     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
2659     irq_set->start = 0;
2660     irq_set->count = 1;
2661     pfd = (int32_t *)&irq_set->data;
2662 
2663     *pfd = event_notifier_get_fd(&vdev->err_notifier);
2664     qemu_set_fd_handler(*pfd, vfio_err_notifier_handler, NULL, vdev);
2665 
2666     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
2667     if (ret) {
2668         error_report("vfio: Failed to set up error notification");
2669         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
2670         event_notifier_cleanup(&vdev->err_notifier);
2671         vdev->pci_aer = false;
2672     }
2673     g_free(irq_set);
2674 }
2675 
2676 static void vfio_unregister_err_notifier(VFIOPCIDevice *vdev)
2677 {
2678     int argsz;
2679     struct vfio_irq_set *irq_set;
2680     int32_t *pfd;
2681     int ret;
2682 
2683     if (!vdev->pci_aer) {
2684         return;
2685     }
2686 
2687     argsz = sizeof(*irq_set) + sizeof(*pfd);
2688 
2689     irq_set = g_malloc0(argsz);
2690     irq_set->argsz = argsz;
2691     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2692                      VFIO_IRQ_SET_ACTION_TRIGGER;
2693     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
2694     irq_set->start = 0;
2695     irq_set->count = 1;
2696     pfd = (int32_t *)&irq_set->data;
2697     *pfd = -1;
2698 
2699     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
2700     if (ret) {
2701         error_report("vfio: Failed to de-assign error fd: %m");
2702     }
2703     g_free(irq_set);
2704     qemu_set_fd_handler(event_notifier_get_fd(&vdev->err_notifier),
2705                         NULL, NULL, vdev);
2706     event_notifier_cleanup(&vdev->err_notifier);
2707 }
2708 
2709 static void vfio_req_notifier_handler(void *opaque)
2710 {
2711     VFIOPCIDevice *vdev = opaque;
2712     Error *err = NULL;
2713 
2714     if (!event_notifier_test_and_clear(&vdev->req_notifier)) {
2715         return;
2716     }
2717 
2718     qdev_unplug(&vdev->pdev.qdev, &err);
2719     if (err) {
2720         warn_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
2721     }
2722 }
2723 
2724 static void vfio_register_req_notifier(VFIOPCIDevice *vdev)
2725 {
2726     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info),
2727                                       .index = VFIO_PCI_REQ_IRQ_INDEX };
2728     int argsz;
2729     struct vfio_irq_set *irq_set;
2730     int32_t *pfd;
2731 
2732     if (!(vdev->features & VFIO_FEATURE_ENABLE_REQ)) {
2733         return;
2734     }
2735 
2736     if (ioctl(vdev->vbasedev.fd,
2737               VFIO_DEVICE_GET_IRQ_INFO, &irq_info) < 0 || irq_info.count < 1) {
2738         return;
2739     }
2740 
2741     if (event_notifier_init(&vdev->req_notifier, 0)) {
2742         error_report("vfio: Unable to init event notifier for device request");
2743         return;
2744     }
2745 
2746     argsz = sizeof(*irq_set) + sizeof(*pfd);
2747 
2748     irq_set = g_malloc0(argsz);
2749     irq_set->argsz = argsz;
2750     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2751                      VFIO_IRQ_SET_ACTION_TRIGGER;
2752     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
2753     irq_set->start = 0;
2754     irq_set->count = 1;
2755     pfd = (int32_t *)&irq_set->data;
2756 
2757     *pfd = event_notifier_get_fd(&vdev->req_notifier);
2758     qemu_set_fd_handler(*pfd, vfio_req_notifier_handler, NULL, vdev);
2759 
2760     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
2761         error_report("vfio: Failed to set up device request notification");
2762         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
2763         event_notifier_cleanup(&vdev->req_notifier);
2764     } else {
2765         vdev->req_enabled = true;
2766     }
2767 
2768     g_free(irq_set);
2769 }
2770 
2771 static void vfio_unregister_req_notifier(VFIOPCIDevice *vdev)
2772 {
2773     int argsz;
2774     struct vfio_irq_set *irq_set;
2775     int32_t *pfd;
2776 
2777     if (!vdev->req_enabled) {
2778         return;
2779     }
2780 
2781     argsz = sizeof(*irq_set) + sizeof(*pfd);
2782 
2783     irq_set = g_malloc0(argsz);
2784     irq_set->argsz = argsz;
2785     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2786                      VFIO_IRQ_SET_ACTION_TRIGGER;
2787     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
2788     irq_set->start = 0;
2789     irq_set->count = 1;
2790     pfd = (int32_t *)&irq_set->data;
2791     *pfd = -1;
2792 
2793     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
2794         error_report("vfio: Failed to de-assign device request fd: %m");
2795     }
2796     g_free(irq_set);
2797     qemu_set_fd_handler(event_notifier_get_fd(&vdev->req_notifier),
2798                         NULL, NULL, vdev);
2799     event_notifier_cleanup(&vdev->req_notifier);
2800 
2801     vdev->req_enabled = false;
2802 }
2803 
2804 static void vfio_realize(PCIDevice *pdev, Error **errp)
2805 {
2806     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
2807     VFIODevice *vbasedev_iter;
2808     VFIOGroup *group;
2809     char *tmp, *subsys, group_path[PATH_MAX], *group_name;
2810     Error *err = NULL;
2811     ssize_t len;
2812     struct stat st;
2813     int groupid;
2814     int i, ret;
2815     bool is_mdev;
2816 
2817     if (!vdev->vbasedev.sysfsdev) {
2818         if (!(~vdev->host.domain || ~vdev->host.bus ||
2819               ~vdev->host.slot || ~vdev->host.function)) {
2820             error_setg(errp, "No provided host device");
2821             error_append_hint(errp, "Use -device vfio-pci,host=DDDD:BB:DD.F "
2822                               "or -device vfio-pci,sysfsdev=PATH_TO_DEVICE\n");
2823             return;
2824         }
2825         vdev->vbasedev.sysfsdev =
2826             g_strdup_printf("/sys/bus/pci/devices/%04x:%02x:%02x.%01x",
2827                             vdev->host.domain, vdev->host.bus,
2828                             vdev->host.slot, vdev->host.function);
2829     }
2830 
2831     if (stat(vdev->vbasedev.sysfsdev, &st) < 0) {
2832         error_setg_errno(errp, errno, "no such host device");
2833         error_prepend(errp, VFIO_MSG_PREFIX, vdev->vbasedev.sysfsdev);
2834         return;
2835     }
2836 
2837     vdev->vbasedev.name = g_path_get_basename(vdev->vbasedev.sysfsdev);
2838     vdev->vbasedev.ops = &vfio_pci_ops;
2839     vdev->vbasedev.type = VFIO_DEVICE_TYPE_PCI;
2840     vdev->vbasedev.dev = &vdev->pdev.qdev;
2841 
2842     tmp = g_strdup_printf("%s/iommu_group", vdev->vbasedev.sysfsdev);
2843     len = readlink(tmp, group_path, sizeof(group_path));
2844     g_free(tmp);
2845 
2846     if (len <= 0 || len >= sizeof(group_path)) {
2847         error_setg_errno(errp, len < 0 ? errno : ENAMETOOLONG,
2848                          "no iommu_group found");
2849         goto error;
2850     }
2851 
2852     group_path[len] = 0;
2853 
2854     group_name = basename(group_path);
2855     if (sscanf(group_name, "%d", &groupid) != 1) {
2856         error_setg_errno(errp, errno, "failed to read %s", group_path);
2857         goto error;
2858     }
2859 
2860     trace_vfio_realize(vdev->vbasedev.name, groupid);
2861 
2862     group = vfio_get_group(groupid, pci_device_iommu_address_space(pdev), errp);
2863     if (!group) {
2864         goto error;
2865     }
2866 
2867     QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2868         if (strcmp(vbasedev_iter->name, vdev->vbasedev.name) == 0) {
2869             error_setg(errp, "device is already attached");
2870             vfio_put_group(group);
2871             goto error;
2872         }
2873     }
2874 
2875     /*
2876      * Mediated devices *might* operate compatibly with memory ballooning, but
2877      * we cannot know for certain, it depends on whether the mdev vendor driver
2878      * stays in sync with the active working set of the guest driver.  Prevent
2879      * the x-balloon-allowed option unless this is minimally an mdev device.
2880      */
2881     tmp = g_strdup_printf("%s/subsystem", vdev->vbasedev.sysfsdev);
2882     subsys = realpath(tmp, NULL);
2883     g_free(tmp);
2884     is_mdev = subsys && (strcmp(subsys, "/sys/bus/mdev") == 0);
2885     free(subsys);
2886 
2887     trace_vfio_mdev(vdev->vbasedev.name, is_mdev);
2888 
2889     if (vdev->vbasedev.balloon_allowed && !is_mdev) {
2890         error_setg(errp, "x-balloon-allowed only potentially compatible "
2891                    "with mdev devices");
2892         vfio_put_group(group);
2893         goto error;
2894     }
2895 
2896     ret = vfio_get_device(group, vdev->vbasedev.name, &vdev->vbasedev, errp);
2897     if (ret) {
2898         vfio_put_group(group);
2899         goto error;
2900     }
2901 
2902     vfio_populate_device(vdev, &err);
2903     if (err) {
2904         error_propagate(errp, err);
2905         goto error;
2906     }
2907 
2908     /* Get a copy of config space */
2909     ret = pread(vdev->vbasedev.fd, vdev->pdev.config,
2910                 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
2911                 vdev->config_offset);
2912     if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
2913         ret = ret < 0 ? -errno : -EFAULT;
2914         error_setg_errno(errp, -ret, "failed to read device config space");
2915         goto error;
2916     }
2917 
2918     /* vfio emulates a lot for us, but some bits need extra love */
2919     vdev->emulated_config_bits = g_malloc0(vdev->config_size);
2920 
2921     /* QEMU can choose to expose the ROM or not */
2922     memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4);
2923     /* QEMU can also add or extend BARs */
2924     memset(vdev->emulated_config_bits + PCI_BASE_ADDRESS_0, 0xff, 6 * 4);
2925 
2926     /*
2927      * The PCI spec reserves vendor ID 0xffff as an invalid value.  The
2928      * device ID is managed by the vendor and need only be a 16-bit value.
2929      * Allow any 16-bit value for subsystem so they can be hidden or changed.
2930      */
2931     if (vdev->vendor_id != PCI_ANY_ID) {
2932         if (vdev->vendor_id >= 0xffff) {
2933             error_setg(errp, "invalid PCI vendor ID provided");
2934             goto error;
2935         }
2936         vfio_add_emulated_word(vdev, PCI_VENDOR_ID, vdev->vendor_id, ~0);
2937         trace_vfio_pci_emulated_vendor_id(vdev->vbasedev.name, vdev->vendor_id);
2938     } else {
2939         vdev->vendor_id = pci_get_word(pdev->config + PCI_VENDOR_ID);
2940     }
2941 
2942     if (vdev->device_id != PCI_ANY_ID) {
2943         if (vdev->device_id > 0xffff) {
2944             error_setg(errp, "invalid PCI device ID provided");
2945             goto error;
2946         }
2947         vfio_add_emulated_word(vdev, PCI_DEVICE_ID, vdev->device_id, ~0);
2948         trace_vfio_pci_emulated_device_id(vdev->vbasedev.name, vdev->device_id);
2949     } else {
2950         vdev->device_id = pci_get_word(pdev->config + PCI_DEVICE_ID);
2951     }
2952 
2953     if (vdev->sub_vendor_id != PCI_ANY_ID) {
2954         if (vdev->sub_vendor_id > 0xffff) {
2955             error_setg(errp, "invalid PCI subsystem vendor ID provided");
2956             goto error;
2957         }
2958         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_VENDOR_ID,
2959                                vdev->sub_vendor_id, ~0);
2960         trace_vfio_pci_emulated_sub_vendor_id(vdev->vbasedev.name,
2961                                               vdev->sub_vendor_id);
2962     }
2963 
2964     if (vdev->sub_device_id != PCI_ANY_ID) {
2965         if (vdev->sub_device_id > 0xffff) {
2966             error_setg(errp, "invalid PCI subsystem device ID provided");
2967             goto error;
2968         }
2969         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_ID, vdev->sub_device_id, ~0);
2970         trace_vfio_pci_emulated_sub_device_id(vdev->vbasedev.name,
2971                                               vdev->sub_device_id);
2972     }
2973 
2974     /* QEMU can change multi-function devices to single function, or reverse */
2975     vdev->emulated_config_bits[PCI_HEADER_TYPE] =
2976                                               PCI_HEADER_TYPE_MULTI_FUNCTION;
2977 
2978     /* Restore or clear multifunction, this is always controlled by QEMU */
2979     if (vdev->pdev.cap_present & QEMU_PCI_CAP_MULTIFUNCTION) {
2980         vdev->pdev.config[PCI_HEADER_TYPE] |= PCI_HEADER_TYPE_MULTI_FUNCTION;
2981     } else {
2982         vdev->pdev.config[PCI_HEADER_TYPE] &= ~PCI_HEADER_TYPE_MULTI_FUNCTION;
2983     }
2984 
2985     /*
2986      * Clear host resource mapping info.  If we choose not to register a
2987      * BAR, such as might be the case with the option ROM, we can get
2988      * confusing, unwritable, residual addresses from the host here.
2989      */
2990     memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
2991     memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
2992 
2993     vfio_pci_size_rom(vdev);
2994 
2995     vfio_bars_prepare(vdev);
2996 
2997     vfio_msix_early_setup(vdev, &err);
2998     if (err) {
2999         error_propagate(errp, err);
3000         goto error;
3001     }
3002 
3003     vfio_bars_register(vdev);
3004 
3005     ret = vfio_add_capabilities(vdev, errp);
3006     if (ret) {
3007         goto out_teardown;
3008     }
3009 
3010     if (vdev->vga) {
3011         vfio_vga_quirk_setup(vdev);
3012     }
3013 
3014     for (i = 0; i < PCI_ROM_SLOT; i++) {
3015         vfio_bar_quirk_setup(vdev, i);
3016     }
3017 
3018     if (!vdev->igd_opregion &&
3019         vdev->features & VFIO_FEATURE_ENABLE_IGD_OPREGION) {
3020         struct vfio_region_info *opregion;
3021 
3022         if (vdev->pdev.qdev.hotplugged) {
3023             error_setg(errp,
3024                        "cannot support IGD OpRegion feature on hotplugged "
3025                        "device");
3026             goto out_teardown;
3027         }
3028 
3029         ret = vfio_get_dev_region_info(&vdev->vbasedev,
3030                         VFIO_REGION_TYPE_PCI_VENDOR_TYPE | PCI_VENDOR_ID_INTEL,
3031                         VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION, &opregion);
3032         if (ret) {
3033             error_setg_errno(errp, -ret,
3034                              "does not support requested IGD OpRegion feature");
3035             goto out_teardown;
3036         }
3037 
3038         ret = vfio_pci_igd_opregion_init(vdev, opregion, errp);
3039         g_free(opregion);
3040         if (ret) {
3041             goto out_teardown;
3042         }
3043     }
3044 
3045     /* QEMU emulates all of MSI & MSIX */
3046     if (pdev->cap_present & QEMU_PCI_CAP_MSIX) {
3047         memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff,
3048                MSIX_CAP_LENGTH);
3049     }
3050 
3051     if (pdev->cap_present & QEMU_PCI_CAP_MSI) {
3052         memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff,
3053                vdev->msi_cap_size);
3054     }
3055 
3056     if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
3057         vdev->intx.mmap_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL,
3058                                                   vfio_intx_mmap_enable, vdev);
3059         pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_intx_update);
3060         ret = vfio_intx_enable(vdev, errp);
3061         if (ret) {
3062             goto out_teardown;
3063         }
3064     }
3065 
3066     if (vdev->display != ON_OFF_AUTO_OFF) {
3067         ret = vfio_display_probe(vdev, errp);
3068         if (ret) {
3069             goto out_teardown;
3070         }
3071     }
3072     if (vdev->enable_ramfb && vdev->dpy == NULL) {
3073         error_setg(errp, "ramfb=on requires display=on");
3074         goto out_teardown;
3075     }
3076 
3077     vfio_register_err_notifier(vdev);
3078     vfio_register_req_notifier(vdev);
3079     vfio_setup_resetfn_quirk(vdev);
3080 
3081     return;
3082 
3083 out_teardown:
3084     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3085     vfio_teardown_msi(vdev);
3086     vfio_bars_exit(vdev);
3087 error:
3088     error_prepend(errp, VFIO_MSG_PREFIX, vdev->vbasedev.name);
3089 }
3090 
3091 static void vfio_instance_finalize(Object *obj)
3092 {
3093     VFIOPCIDevice *vdev = PCI_VFIO(obj);
3094     VFIOGroup *group = vdev->vbasedev.group;
3095 
3096     vfio_display_finalize(vdev);
3097     vfio_bars_finalize(vdev);
3098     g_free(vdev->emulated_config_bits);
3099     g_free(vdev->rom);
3100     /*
3101      * XXX Leaking igd_opregion is not an oversight, we can't remove the
3102      * fw_cfg entry therefore leaking this allocation seems like the safest
3103      * option.
3104      *
3105      * g_free(vdev->igd_opregion);
3106      */
3107     vfio_put_device(vdev);
3108     vfio_put_group(group);
3109 }
3110 
3111 static void vfio_exitfn(PCIDevice *pdev)
3112 {
3113     VFIOPCIDevice *vdev = PCI_VFIO(pdev);
3114 
3115     vfio_unregister_req_notifier(vdev);
3116     vfio_unregister_err_notifier(vdev);
3117     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3118     vfio_disable_interrupts(vdev);
3119     if (vdev->intx.mmap_timer) {
3120         timer_free(vdev->intx.mmap_timer);
3121     }
3122     vfio_teardown_msi(vdev);
3123     vfio_bars_exit(vdev);
3124 }
3125 
3126 static void vfio_pci_reset(DeviceState *dev)
3127 {
3128     VFIOPCIDevice *vdev = PCI_VFIO(dev);
3129 
3130     trace_vfio_pci_reset(vdev->vbasedev.name);
3131 
3132     vfio_pci_pre_reset(vdev);
3133 
3134     if (vdev->display != ON_OFF_AUTO_OFF) {
3135         vfio_display_reset(vdev);
3136     }
3137 
3138     if (vdev->resetfn && !vdev->resetfn(vdev)) {
3139         goto post_reset;
3140     }
3141 
3142     if (vdev->vbasedev.reset_works &&
3143         (vdev->has_flr || !vdev->has_pm_reset) &&
3144         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
3145         trace_vfio_pci_reset_flr(vdev->vbasedev.name);
3146         goto post_reset;
3147     }
3148 
3149     /* See if we can do our own bus reset */
3150     if (!vfio_pci_hot_reset_one(vdev)) {
3151         goto post_reset;
3152     }
3153 
3154     /* If nothing else works and the device supports PM reset, use it */
3155     if (vdev->vbasedev.reset_works && vdev->has_pm_reset &&
3156         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
3157         trace_vfio_pci_reset_pm(vdev->vbasedev.name);
3158         goto post_reset;
3159     }
3160 
3161 post_reset:
3162     vfio_pci_post_reset(vdev);
3163 }
3164 
3165 static void vfio_instance_init(Object *obj)
3166 {
3167     PCIDevice *pci_dev = PCI_DEVICE(obj);
3168     VFIOPCIDevice *vdev = PCI_VFIO(obj);
3169 
3170     device_add_bootindex_property(obj, &vdev->bootindex,
3171                                   "bootindex", NULL,
3172                                   &pci_dev->qdev, NULL);
3173     vdev->host.domain = ~0U;
3174     vdev->host.bus = ~0U;
3175     vdev->host.slot = ~0U;
3176     vdev->host.function = ~0U;
3177 
3178     vdev->nv_gpudirect_clique = 0xFF;
3179 
3180     /* QEMU_PCI_CAP_EXPRESS initialization does not depend on QEMU command
3181      * line, therefore, no need to wait to realize like other devices */
3182     pci_dev->cap_present |= QEMU_PCI_CAP_EXPRESS;
3183 }
3184 
3185 static Property vfio_pci_dev_properties[] = {
3186     DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIOPCIDevice, host),
3187     DEFINE_PROP_STRING("sysfsdev", VFIOPCIDevice, vbasedev.sysfsdev),
3188     DEFINE_PROP_ON_OFF_AUTO("display", VFIOPCIDevice,
3189                             display, ON_OFF_AUTO_OFF),
3190     DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIOPCIDevice,
3191                        intx.mmap_timeout, 1100),
3192     DEFINE_PROP_BIT("x-vga", VFIOPCIDevice, features,
3193                     VFIO_FEATURE_ENABLE_VGA_BIT, false),
3194     DEFINE_PROP_BIT("x-req", VFIOPCIDevice, features,
3195                     VFIO_FEATURE_ENABLE_REQ_BIT, true),
3196     DEFINE_PROP_BIT("x-igd-opregion", VFIOPCIDevice, features,
3197                     VFIO_FEATURE_ENABLE_IGD_OPREGION_BIT, false),
3198     DEFINE_PROP_BOOL("x-no-mmap", VFIOPCIDevice, vbasedev.no_mmap, false),
3199     DEFINE_PROP_BOOL("x-balloon-allowed", VFIOPCIDevice,
3200                      vbasedev.balloon_allowed, false),
3201     DEFINE_PROP_BOOL("x-no-kvm-intx", VFIOPCIDevice, no_kvm_intx, false),
3202     DEFINE_PROP_BOOL("x-no-kvm-msi", VFIOPCIDevice, no_kvm_msi, false),
3203     DEFINE_PROP_BOOL("x-no-kvm-msix", VFIOPCIDevice, no_kvm_msix, false),
3204     DEFINE_PROP_BOOL("x-no-geforce-quirks", VFIOPCIDevice,
3205                      no_geforce_quirks, false),
3206     DEFINE_PROP_BOOL("x-no-kvm-ioeventfd", VFIOPCIDevice, no_kvm_ioeventfd,
3207                      false),
3208     DEFINE_PROP_BOOL("x-no-vfio-ioeventfd", VFIOPCIDevice, no_vfio_ioeventfd,
3209                      false),
3210     DEFINE_PROP_UINT32("x-pci-vendor-id", VFIOPCIDevice, vendor_id, PCI_ANY_ID),
3211     DEFINE_PROP_UINT32("x-pci-device-id", VFIOPCIDevice, device_id, PCI_ANY_ID),
3212     DEFINE_PROP_UINT32("x-pci-sub-vendor-id", VFIOPCIDevice,
3213                        sub_vendor_id, PCI_ANY_ID),
3214     DEFINE_PROP_UINT32("x-pci-sub-device-id", VFIOPCIDevice,
3215                        sub_device_id, PCI_ANY_ID),
3216     DEFINE_PROP_UINT32("x-igd-gms", VFIOPCIDevice, igd_gms, 0),
3217     DEFINE_PROP_UNSIGNED_NODEFAULT("x-nv-gpudirect-clique", VFIOPCIDevice,
3218                                    nv_gpudirect_clique,
3219                                    qdev_prop_nv_gpudirect_clique, uint8_t),
3220     DEFINE_PROP_OFF_AUTO_PCIBAR("x-msix-relocation", VFIOPCIDevice, msix_relo,
3221                                 OFF_AUTOPCIBAR_OFF),
3222     /*
3223      * TODO - support passed fds... is this necessary?
3224      * DEFINE_PROP_STRING("vfiofd", VFIOPCIDevice, vfiofd_name),
3225      * DEFINE_PROP_STRING("vfiogroupfd, VFIOPCIDevice, vfiogroupfd_name),
3226      */
3227     DEFINE_PROP_END_OF_LIST(),
3228 };
3229 
3230 static const VMStateDescription vfio_pci_vmstate = {
3231     .name = "vfio-pci",
3232     .unmigratable = 1,
3233 };
3234 
3235 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
3236 {
3237     DeviceClass *dc = DEVICE_CLASS(klass);
3238     PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
3239 
3240     dc->reset = vfio_pci_reset;
3241     dc->props = vfio_pci_dev_properties;
3242     dc->vmsd = &vfio_pci_vmstate;
3243     dc->desc = "VFIO-based PCI device assignment";
3244     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
3245     pdc->realize = vfio_realize;
3246     pdc->exit = vfio_exitfn;
3247     pdc->config_read = vfio_pci_read_config;
3248     pdc->config_write = vfio_pci_write_config;
3249 }
3250 
3251 static const TypeInfo vfio_pci_dev_info = {
3252     .name = TYPE_VFIO_PCI,
3253     .parent = TYPE_PCI_DEVICE,
3254     .instance_size = sizeof(VFIOPCIDevice),
3255     .class_init = vfio_pci_dev_class_init,
3256     .instance_init = vfio_instance_init,
3257     .instance_finalize = vfio_instance_finalize,
3258     .interfaces = (InterfaceInfo[]) {
3259         { INTERFACE_PCIE_DEVICE },
3260         { INTERFACE_CONVENTIONAL_PCI_DEVICE },
3261         { }
3262     },
3263 };
3264 
3265 static Property vfio_pci_dev_nohotplug_properties[] = {
3266     DEFINE_PROP_BOOL("ramfb", VFIOPCIDevice, enable_ramfb, false),
3267     DEFINE_PROP_END_OF_LIST(),
3268 };
3269 
3270 static void vfio_pci_nohotplug_dev_class_init(ObjectClass *klass, void *data)
3271 {
3272     DeviceClass *dc = DEVICE_CLASS(klass);
3273 
3274     dc->props = vfio_pci_dev_nohotplug_properties;
3275     dc->hotpluggable = false;
3276 }
3277 
3278 static const TypeInfo vfio_pci_nohotplug_dev_info = {
3279     .name = "vfio-pci-nohotplug",
3280     .parent = "vfio-pci",
3281     .instance_size = sizeof(VFIOPCIDevice),
3282     .class_init = vfio_pci_nohotplug_dev_class_init,
3283 };
3284 
3285 static void register_vfio_pci_dev_type(void)
3286 {
3287     type_register_static(&vfio_pci_dev_info);
3288     type_register_static(&vfio_pci_nohotplug_dev_info);
3289 }
3290 
3291 type_init(register_vfio_pci_dev_type)
3292