xref: /openbmc/qemu/hw/vfio/pci.c (revision 77d361b1)
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 "sysemu/kvm.h"
32 #include "sysemu/sysemu.h"
33 #include "pci.h"
34 #include "trace.h"
35 #include "qapi/error.h"
36 
37 #define MSIX_CAP_LENGTH 12
38 
39 static void vfio_disable_interrupts(VFIOPCIDevice *vdev);
40 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled);
41 
42 /*
43  * Disabling BAR mmaping can be slow, but toggling it around INTx can
44  * also be a huge overhead.  We try to get the best of both worlds by
45  * waiting until an interrupt to disable mmaps (subsequent transitions
46  * to the same state are effectively no overhead).  If the interrupt has
47  * been serviced and the time gap is long enough, we re-enable mmaps for
48  * performance.  This works well for things like graphics cards, which
49  * may not use their interrupt at all and are penalized to an unusable
50  * level by read/write BAR traps.  Other devices, like NICs, have more
51  * regular interrupts and see much better latency by staying in non-mmap
52  * mode.  We therefore set the default mmap_timeout such that a ping
53  * is just enough to keep the mmap disabled.  Users can experiment with
54  * other options with the x-intx-mmap-timeout-ms parameter (a value of
55  * zero disables the timer).
56  */
57 static void vfio_intx_mmap_enable(void *opaque)
58 {
59     VFIOPCIDevice *vdev = opaque;
60 
61     if (vdev->intx.pending) {
62         timer_mod(vdev->intx.mmap_timer,
63                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
64         return;
65     }
66 
67     vfio_mmap_set_enabled(vdev, true);
68 }
69 
70 static void vfio_intx_interrupt(void *opaque)
71 {
72     VFIOPCIDevice *vdev = opaque;
73 
74     if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
75         return;
76     }
77 
78     trace_vfio_intx_interrupt(vdev->vbasedev.name, 'A' + vdev->intx.pin);
79 
80     vdev->intx.pending = true;
81     pci_irq_assert(&vdev->pdev);
82     vfio_mmap_set_enabled(vdev, false);
83     if (vdev->intx.mmap_timeout) {
84         timer_mod(vdev->intx.mmap_timer,
85                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
86     }
87 }
88 
89 static void vfio_intx_eoi(VFIODevice *vbasedev)
90 {
91     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
92 
93     if (!vdev->intx.pending) {
94         return;
95     }
96 
97     trace_vfio_intx_eoi(vbasedev->name);
98 
99     vdev->intx.pending = false;
100     pci_irq_deassert(&vdev->pdev);
101     vfio_unmask_single_irqindex(vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
102 }
103 
104 static void vfio_intx_enable_kvm(VFIOPCIDevice *vdev, Error **errp)
105 {
106 #ifdef CONFIG_KVM
107     struct kvm_irqfd irqfd = {
108         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
109         .gsi = vdev->intx.route.irq,
110         .flags = KVM_IRQFD_FLAG_RESAMPLE,
111     };
112     struct vfio_irq_set *irq_set;
113     int ret, argsz;
114     int32_t *pfd;
115 
116     if (vdev->no_kvm_intx || !kvm_irqfds_enabled() ||
117         vdev->intx.route.mode != PCI_INTX_ENABLED ||
118         !kvm_resamplefds_enabled()) {
119         return;
120     }
121 
122     /* Get to a known interrupt state */
123     qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
124     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
125     vdev->intx.pending = false;
126     pci_irq_deassert(&vdev->pdev);
127 
128     /* Get an eventfd for resample/unmask */
129     if (event_notifier_init(&vdev->intx.unmask, 0)) {
130         error_setg(errp, "event_notifier_init failed eoi");
131         goto fail;
132     }
133 
134     /* KVM triggers it, VFIO listens for it */
135     irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
136 
137     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
138         error_setg_errno(errp, errno, "failed to setup resample irqfd");
139         goto fail_irqfd;
140     }
141 
142     argsz = sizeof(*irq_set) + sizeof(*pfd);
143 
144     irq_set = g_malloc0(argsz);
145     irq_set->argsz = argsz;
146     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
147     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
148     irq_set->start = 0;
149     irq_set->count = 1;
150     pfd = (int32_t *)&irq_set->data;
151 
152     *pfd = irqfd.resamplefd;
153 
154     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
155     g_free(irq_set);
156     if (ret) {
157         error_setg_errno(errp, -ret, "failed to setup INTx unmask fd");
158         goto fail_vfio;
159     }
160 
161     /* Let'em rip */
162     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
163 
164     vdev->intx.kvm_accel = true;
165 
166     trace_vfio_intx_enable_kvm(vdev->vbasedev.name);
167 
168     return;
169 
170 fail_vfio:
171     irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
172     kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
173 fail_irqfd:
174     event_notifier_cleanup(&vdev->intx.unmask);
175 fail:
176     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
177     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
178 #endif
179 }
180 
181 static void vfio_intx_disable_kvm(VFIOPCIDevice *vdev)
182 {
183 #ifdef CONFIG_KVM
184     struct kvm_irqfd irqfd = {
185         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
186         .gsi = vdev->intx.route.irq,
187         .flags = KVM_IRQFD_FLAG_DEASSIGN,
188     };
189 
190     if (!vdev->intx.kvm_accel) {
191         return;
192     }
193 
194     /*
195      * Get to a known state, hardware masked, QEMU ready to accept new
196      * interrupts, QEMU IRQ de-asserted.
197      */
198     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
199     vdev->intx.pending = false;
200     pci_irq_deassert(&vdev->pdev);
201 
202     /* Tell KVM to stop listening for an INTx irqfd */
203     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
204         error_report("vfio: Error: Failed to disable INTx irqfd: %m");
205     }
206 
207     /* We only need to close the eventfd for VFIO to cleanup the kernel side */
208     event_notifier_cleanup(&vdev->intx.unmask);
209 
210     /* QEMU starts listening for interrupt events. */
211     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
212 
213     vdev->intx.kvm_accel = false;
214 
215     /* If we've missed an event, let it re-fire through QEMU */
216     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
217 
218     trace_vfio_intx_disable_kvm(vdev->vbasedev.name);
219 #endif
220 }
221 
222 static void vfio_intx_update(PCIDevice *pdev)
223 {
224     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
225     PCIINTxRoute route;
226     Error *err = NULL;
227 
228     if (vdev->interrupt != VFIO_INT_INTx) {
229         return;
230     }
231 
232     route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
233 
234     if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
235         return; /* Nothing changed */
236     }
237 
238     trace_vfio_intx_update(vdev->vbasedev.name,
239                            vdev->intx.route.irq, route.irq);
240 
241     vfio_intx_disable_kvm(vdev);
242 
243     vdev->intx.route = route;
244 
245     if (route.mode != PCI_INTX_ENABLED) {
246         return;
247     }
248 
249     vfio_intx_enable_kvm(vdev, &err);
250     if (err) {
251         error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name);
252     }
253 
254     /* Re-enable the interrupt in cased we missed an EOI */
255     vfio_intx_eoi(&vdev->vbasedev);
256 }
257 
258 static int vfio_intx_enable(VFIOPCIDevice *vdev, Error **errp)
259 {
260     uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
261     int ret, argsz, retval = 0;
262     struct vfio_irq_set *irq_set;
263     int32_t *pfd;
264     Error *err = NULL;
265 
266     if (!pin) {
267         return 0;
268     }
269 
270     vfio_disable_interrupts(vdev);
271 
272     vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
273     pci_config_set_interrupt_pin(vdev->pdev.config, pin);
274 
275 #ifdef CONFIG_KVM
276     /*
277      * Only conditional to avoid generating error messages on platforms
278      * where we won't actually use the result anyway.
279      */
280     if (kvm_irqfds_enabled() && kvm_resamplefds_enabled()) {
281         vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
282                                                         vdev->intx.pin);
283     }
284 #endif
285 
286     ret = event_notifier_init(&vdev->intx.interrupt, 0);
287     if (ret) {
288         error_setg_errno(errp, -ret, "event_notifier_init failed");
289         return ret;
290     }
291 
292     argsz = sizeof(*irq_set) + sizeof(*pfd);
293 
294     irq_set = g_malloc0(argsz);
295     irq_set->argsz = argsz;
296     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
297     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
298     irq_set->start = 0;
299     irq_set->count = 1;
300     pfd = (int32_t *)&irq_set->data;
301 
302     *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
303     qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
304 
305     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
306     if (ret) {
307         error_setg_errno(errp, -ret, "failed to setup INTx fd");
308         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
309         event_notifier_cleanup(&vdev->intx.interrupt);
310         retval = -errno;
311         goto cleanup;
312     }
313 
314     vfio_intx_enable_kvm(vdev, &err);
315     if (err) {
316         error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name);
317     }
318 
319     vdev->interrupt = VFIO_INT_INTx;
320 
321     trace_vfio_intx_enable(vdev->vbasedev.name);
322 
323 cleanup:
324     g_free(irq_set);
325 
326     return retval;
327 }
328 
329 static void vfio_intx_disable(VFIOPCIDevice *vdev)
330 {
331     int fd;
332 
333     timer_del(vdev->intx.mmap_timer);
334     vfio_intx_disable_kvm(vdev);
335     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
336     vdev->intx.pending = false;
337     pci_irq_deassert(&vdev->pdev);
338     vfio_mmap_set_enabled(vdev, true);
339 
340     fd = event_notifier_get_fd(&vdev->intx.interrupt);
341     qemu_set_fd_handler(fd, NULL, NULL, vdev);
342     event_notifier_cleanup(&vdev->intx.interrupt);
343 
344     vdev->interrupt = VFIO_INT_NONE;
345 
346     trace_vfio_intx_disable(vdev->vbasedev.name);
347 }
348 
349 /*
350  * MSI/X
351  */
352 static void vfio_msi_interrupt(void *opaque)
353 {
354     VFIOMSIVector *vector = opaque;
355     VFIOPCIDevice *vdev = vector->vdev;
356     MSIMessage (*get_msg)(PCIDevice *dev, unsigned vector);
357     void (*notify)(PCIDevice *dev, unsigned vector);
358     MSIMessage msg;
359     int nr = vector - vdev->msi_vectors;
360 
361     if (!event_notifier_test_and_clear(&vector->interrupt)) {
362         return;
363     }
364 
365     if (vdev->interrupt == VFIO_INT_MSIX) {
366         get_msg = msix_get_message;
367         notify = msix_notify;
368 
369         /* A masked vector firing needs to use the PBA, enable it */
370         if (msix_is_masked(&vdev->pdev, nr)) {
371             set_bit(nr, vdev->msix->pending);
372             memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, true);
373             trace_vfio_msix_pba_enable(vdev->vbasedev.name);
374         }
375     } else if (vdev->interrupt == VFIO_INT_MSI) {
376         get_msg = msi_get_message;
377         notify = msi_notify;
378     } else {
379         abort();
380     }
381 
382     msg = get_msg(&vdev->pdev, nr);
383     trace_vfio_msi_interrupt(vdev->vbasedev.name, nr, msg.address, msg.data);
384     notify(&vdev->pdev, nr);
385 }
386 
387 static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool msix)
388 {
389     struct vfio_irq_set *irq_set;
390     int ret = 0, i, argsz;
391     int32_t *fds;
392 
393     argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
394 
395     irq_set = g_malloc0(argsz);
396     irq_set->argsz = argsz;
397     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
398     irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
399     irq_set->start = 0;
400     irq_set->count = vdev->nr_vectors;
401     fds = (int32_t *)&irq_set->data;
402 
403     for (i = 0; i < vdev->nr_vectors; i++) {
404         int fd = -1;
405 
406         /*
407          * MSI vs MSI-X - The guest has direct access to MSI mask and pending
408          * bits, therefore we always use the KVM signaling path when setup.
409          * MSI-X mask and pending bits are emulated, so we want to use the
410          * KVM signaling path only when configured and unmasked.
411          */
412         if (vdev->msi_vectors[i].use) {
413             if (vdev->msi_vectors[i].virq < 0 ||
414                 (msix && msix_is_masked(&vdev->pdev, i))) {
415                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
416             } else {
417                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].kvm_interrupt);
418             }
419         }
420 
421         fds[i] = fd;
422     }
423 
424     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
425 
426     g_free(irq_set);
427 
428     return ret;
429 }
430 
431 static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
432                                   int vector_n, bool msix)
433 {
434     int virq;
435 
436     if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) {
437         return;
438     }
439 
440     if (event_notifier_init(&vector->kvm_interrupt, 0)) {
441         return;
442     }
443 
444     virq = kvm_irqchip_add_msi_route(kvm_state, vector_n, &vdev->pdev);
445     if (virq < 0) {
446         event_notifier_cleanup(&vector->kvm_interrupt);
447         return;
448     }
449 
450     if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
451                                        NULL, virq) < 0) {
452         kvm_irqchip_release_virq(kvm_state, virq);
453         event_notifier_cleanup(&vector->kvm_interrupt);
454         return;
455     }
456 
457     vector->virq = virq;
458 }
459 
460 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
461 {
462     kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
463                                           vector->virq);
464     kvm_irqchip_release_virq(kvm_state, vector->virq);
465     vector->virq = -1;
466     event_notifier_cleanup(&vector->kvm_interrupt);
467 }
468 
469 static void vfio_update_kvm_msi_virq(VFIOMSIVector *vector, MSIMessage msg,
470                                      PCIDevice *pdev)
471 {
472     kvm_irqchip_update_msi_route(kvm_state, vector->virq, msg, pdev);
473     kvm_irqchip_commit_routes(kvm_state);
474 }
475 
476 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
477                                    MSIMessage *msg, IOHandler *handler)
478 {
479     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
480     VFIOMSIVector *vector;
481     int ret;
482 
483     trace_vfio_msix_vector_do_use(vdev->vbasedev.name, nr);
484 
485     vector = &vdev->msi_vectors[nr];
486 
487     if (!vector->use) {
488         vector->vdev = vdev;
489         vector->virq = -1;
490         if (event_notifier_init(&vector->interrupt, 0)) {
491             error_report("vfio: Error: event_notifier_init failed");
492         }
493         vector->use = true;
494         msix_vector_use(pdev, nr);
495     }
496 
497     qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
498                         handler, NULL, vector);
499 
500     /*
501      * Attempt to enable route through KVM irqchip,
502      * default to userspace handling if unavailable.
503      */
504     if (vector->virq >= 0) {
505         if (!msg) {
506             vfio_remove_kvm_msi_virq(vector);
507         } else {
508             vfio_update_kvm_msi_virq(vector, *msg, pdev);
509         }
510     } else {
511         if (msg) {
512             vfio_add_kvm_msi_virq(vdev, vector, nr, true);
513         }
514     }
515 
516     /*
517      * We don't want to have the host allocate all possible MSI vectors
518      * for a device if they're not in use, so we shutdown and incrementally
519      * increase them as needed.
520      */
521     if (vdev->nr_vectors < nr + 1) {
522         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
523         vdev->nr_vectors = nr + 1;
524         ret = vfio_enable_vectors(vdev, true);
525         if (ret) {
526             error_report("vfio: failed to enable vectors, %d", ret);
527         }
528     } else {
529         int argsz;
530         struct vfio_irq_set *irq_set;
531         int32_t *pfd;
532 
533         argsz = sizeof(*irq_set) + sizeof(*pfd);
534 
535         irq_set = g_malloc0(argsz);
536         irq_set->argsz = argsz;
537         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
538                          VFIO_IRQ_SET_ACTION_TRIGGER;
539         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
540         irq_set->start = nr;
541         irq_set->count = 1;
542         pfd = (int32_t *)&irq_set->data;
543 
544         if (vector->virq >= 0) {
545             *pfd = event_notifier_get_fd(&vector->kvm_interrupt);
546         } else {
547             *pfd = event_notifier_get_fd(&vector->interrupt);
548         }
549 
550         ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
551         g_free(irq_set);
552         if (ret) {
553             error_report("vfio: failed to modify vector, %d", ret);
554         }
555     }
556 
557     /* Disable PBA emulation when nothing more is pending. */
558     clear_bit(nr, vdev->msix->pending);
559     if (find_first_bit(vdev->msix->pending,
560                        vdev->nr_vectors) == vdev->nr_vectors) {
561         memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
562         trace_vfio_msix_pba_disable(vdev->vbasedev.name);
563     }
564 
565     return 0;
566 }
567 
568 static int vfio_msix_vector_use(PCIDevice *pdev,
569                                 unsigned int nr, MSIMessage msg)
570 {
571     return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
572 }
573 
574 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
575 {
576     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
577     VFIOMSIVector *vector = &vdev->msi_vectors[nr];
578 
579     trace_vfio_msix_vector_release(vdev->vbasedev.name, nr);
580 
581     /*
582      * There are still old guests that mask and unmask vectors on every
583      * interrupt.  If we're using QEMU bypass with a KVM irqfd, leave all of
584      * the KVM setup in place, simply switch VFIO to use the non-bypass
585      * eventfd.  We'll then fire the interrupt through QEMU and the MSI-X
586      * core will mask the interrupt and set pending bits, allowing it to
587      * be re-asserted on unmask.  Nothing to do if already using QEMU mode.
588      */
589     if (vector->virq >= 0) {
590         int argsz;
591         struct vfio_irq_set *irq_set;
592         int32_t *pfd;
593 
594         argsz = sizeof(*irq_set) + sizeof(*pfd);
595 
596         irq_set = g_malloc0(argsz);
597         irq_set->argsz = argsz;
598         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
599                          VFIO_IRQ_SET_ACTION_TRIGGER;
600         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
601         irq_set->start = nr;
602         irq_set->count = 1;
603         pfd = (int32_t *)&irq_set->data;
604 
605         *pfd = event_notifier_get_fd(&vector->interrupt);
606 
607         ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
608 
609         g_free(irq_set);
610     }
611 }
612 
613 static void vfio_msix_enable(VFIOPCIDevice *vdev)
614 {
615     vfio_disable_interrupts(vdev);
616 
617     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->msix->entries);
618 
619     vdev->interrupt = VFIO_INT_MSIX;
620 
621     /*
622      * Some communication channels between VF & PF or PF & fw rely on the
623      * physical state of the device and expect that enabling MSI-X from the
624      * guest enables the same on the host.  When our guest is Linux, the
625      * guest driver call to pci_enable_msix() sets the enabling bit in the
626      * MSI-X capability, but leaves the vector table masked.  We therefore
627      * can't rely on a vector_use callback (from request_irq() in the guest)
628      * to switch the physical device into MSI-X mode because that may come a
629      * long time after pci_enable_msix().  This code enables vector 0 with
630      * triggering to userspace, then immediately release the vector, leaving
631      * the physical device with no vectors enabled, but MSI-X enabled, just
632      * like the guest view.
633      */
634     vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
635     vfio_msix_vector_release(&vdev->pdev, 0);
636 
637     if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
638                                   vfio_msix_vector_release, NULL)) {
639         error_report("vfio: msix_set_vector_notifiers failed");
640     }
641 
642     trace_vfio_msix_enable(vdev->vbasedev.name);
643 }
644 
645 static void vfio_msi_enable(VFIOPCIDevice *vdev)
646 {
647     int ret, i;
648 
649     vfio_disable_interrupts(vdev);
650 
651     vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
652 retry:
653     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->nr_vectors);
654 
655     for (i = 0; i < vdev->nr_vectors; i++) {
656         VFIOMSIVector *vector = &vdev->msi_vectors[i];
657 
658         vector->vdev = vdev;
659         vector->virq = -1;
660         vector->use = true;
661 
662         if (event_notifier_init(&vector->interrupt, 0)) {
663             error_report("vfio: Error: event_notifier_init failed");
664         }
665 
666         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
667                             vfio_msi_interrupt, NULL, vector);
668 
669         /*
670          * Attempt to enable route through KVM irqchip,
671          * default to userspace handling if unavailable.
672          */
673         vfio_add_kvm_msi_virq(vdev, vector, i, false);
674     }
675 
676     /* Set interrupt type prior to possible interrupts */
677     vdev->interrupt = VFIO_INT_MSI;
678 
679     ret = vfio_enable_vectors(vdev, false);
680     if (ret) {
681         if (ret < 0) {
682             error_report("vfio: Error: Failed to setup MSI fds: %m");
683         } else if (ret != vdev->nr_vectors) {
684             error_report("vfio: Error: Failed to enable %d "
685                          "MSI vectors, retry with %d", vdev->nr_vectors, ret);
686         }
687 
688         for (i = 0; i < vdev->nr_vectors; i++) {
689             VFIOMSIVector *vector = &vdev->msi_vectors[i];
690             if (vector->virq >= 0) {
691                 vfio_remove_kvm_msi_virq(vector);
692             }
693             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
694                                 NULL, NULL, NULL);
695             event_notifier_cleanup(&vector->interrupt);
696         }
697 
698         g_free(vdev->msi_vectors);
699 
700         if (ret > 0 && ret != vdev->nr_vectors) {
701             vdev->nr_vectors = ret;
702             goto retry;
703         }
704         vdev->nr_vectors = 0;
705 
706         /*
707          * Failing to setup MSI doesn't really fall within any specification.
708          * Let's try leaving interrupts disabled and hope the guest figures
709          * out to fall back to INTx for this device.
710          */
711         error_report("vfio: Error: Failed to enable MSI");
712         vdev->interrupt = VFIO_INT_NONE;
713 
714         return;
715     }
716 
717     trace_vfio_msi_enable(vdev->vbasedev.name, vdev->nr_vectors);
718 }
719 
720 static void vfio_msi_disable_common(VFIOPCIDevice *vdev)
721 {
722     Error *err = NULL;
723     int i;
724 
725     for (i = 0; i < vdev->nr_vectors; i++) {
726         VFIOMSIVector *vector = &vdev->msi_vectors[i];
727         if (vdev->msi_vectors[i].use) {
728             if (vector->virq >= 0) {
729                 vfio_remove_kvm_msi_virq(vector);
730             }
731             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
732                                 NULL, NULL, NULL);
733             event_notifier_cleanup(&vector->interrupt);
734         }
735     }
736 
737     g_free(vdev->msi_vectors);
738     vdev->msi_vectors = NULL;
739     vdev->nr_vectors = 0;
740     vdev->interrupt = VFIO_INT_NONE;
741 
742     vfio_intx_enable(vdev, &err);
743     if (err) {
744         error_reportf_err(err, ERR_PREFIX, vdev->vbasedev.name);
745     }
746 }
747 
748 static void vfio_msix_disable(VFIOPCIDevice *vdev)
749 {
750     int i;
751 
752     msix_unset_vector_notifiers(&vdev->pdev);
753 
754     /*
755      * MSI-X will only release vectors if MSI-X is still enabled on the
756      * device, check through the rest and release it ourselves if necessary.
757      */
758     for (i = 0; i < vdev->nr_vectors; i++) {
759         if (vdev->msi_vectors[i].use) {
760             vfio_msix_vector_release(&vdev->pdev, i);
761             msix_vector_unuse(&vdev->pdev, i);
762         }
763     }
764 
765     if (vdev->nr_vectors) {
766         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
767     }
768 
769     vfio_msi_disable_common(vdev);
770 
771     memset(vdev->msix->pending, 0,
772            BITS_TO_LONGS(vdev->msix->entries) * sizeof(unsigned long));
773 
774     trace_vfio_msix_disable(vdev->vbasedev.name);
775 }
776 
777 static void vfio_msi_disable(VFIOPCIDevice *vdev)
778 {
779     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSI_IRQ_INDEX);
780     vfio_msi_disable_common(vdev);
781 
782     trace_vfio_msi_disable(vdev->vbasedev.name);
783 }
784 
785 static void vfio_update_msi(VFIOPCIDevice *vdev)
786 {
787     int i;
788 
789     for (i = 0; i < vdev->nr_vectors; i++) {
790         VFIOMSIVector *vector = &vdev->msi_vectors[i];
791         MSIMessage msg;
792 
793         if (!vector->use || vector->virq < 0) {
794             continue;
795         }
796 
797         msg = msi_get_message(&vdev->pdev, i);
798         vfio_update_kvm_msi_virq(vector, msg, &vdev->pdev);
799     }
800 }
801 
802 static void vfio_pci_load_rom(VFIOPCIDevice *vdev)
803 {
804     struct vfio_region_info *reg_info;
805     uint64_t size;
806     off_t off = 0;
807     ssize_t bytes;
808 
809     if (vfio_get_region_info(&vdev->vbasedev,
810                              VFIO_PCI_ROM_REGION_INDEX, &reg_info)) {
811         error_report("vfio: Error getting ROM info: %m");
812         return;
813     }
814 
815     trace_vfio_pci_load_rom(vdev->vbasedev.name, (unsigned long)reg_info->size,
816                             (unsigned long)reg_info->offset,
817                             (unsigned long)reg_info->flags);
818 
819     vdev->rom_size = size = reg_info->size;
820     vdev->rom_offset = reg_info->offset;
821 
822     g_free(reg_info);
823 
824     if (!vdev->rom_size) {
825         vdev->rom_read_failed = true;
826         error_report("vfio-pci: Cannot read device rom at "
827                     "%s", vdev->vbasedev.name);
828         error_printf("Device option ROM contents are probably invalid "
829                     "(check dmesg).\nSkip option ROM probe with rombar=0, "
830                     "or load from file with romfile=\n");
831         return;
832     }
833 
834     vdev->rom = g_malloc(size);
835     memset(vdev->rom, 0xff, size);
836 
837     while (size) {
838         bytes = pread(vdev->vbasedev.fd, vdev->rom + off,
839                       size, vdev->rom_offset + off);
840         if (bytes == 0) {
841             break;
842         } else if (bytes > 0) {
843             off += bytes;
844             size -= bytes;
845         } else {
846             if (errno == EINTR || errno == EAGAIN) {
847                 continue;
848             }
849             error_report("vfio: Error reading device ROM: %m");
850             break;
851         }
852     }
853 
854     /*
855      * Test the ROM signature against our device, if the vendor is correct
856      * but the device ID doesn't match, store the correct device ID and
857      * recompute the checksum.  Intel IGD devices need this and are known
858      * to have bogus checksums so we can't simply adjust the checksum.
859      */
860     if (pci_get_word(vdev->rom) == 0xaa55 &&
861         pci_get_word(vdev->rom + 0x18) + 8 < vdev->rom_size &&
862         !memcmp(vdev->rom + pci_get_word(vdev->rom + 0x18), "PCIR", 4)) {
863         uint16_t vid, did;
864 
865         vid = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 4);
866         did = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6);
867 
868         if (vid == vdev->vendor_id && did != vdev->device_id) {
869             int i;
870             uint8_t csum, *data = vdev->rom;
871 
872             pci_set_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6,
873                          vdev->device_id);
874             data[6] = 0;
875 
876             for (csum = 0, i = 0; i < vdev->rom_size; i++) {
877                 csum += data[i];
878             }
879 
880             data[6] = -csum;
881         }
882     }
883 }
884 
885 static uint64_t vfio_rom_read(void *opaque, hwaddr addr, unsigned size)
886 {
887     VFIOPCIDevice *vdev = opaque;
888     union {
889         uint8_t byte;
890         uint16_t word;
891         uint32_t dword;
892         uint64_t qword;
893     } val;
894     uint64_t data = 0;
895 
896     /* Load the ROM lazily when the guest tries to read it */
897     if (unlikely(!vdev->rom && !vdev->rom_read_failed)) {
898         vfio_pci_load_rom(vdev);
899     }
900 
901     memcpy(&val, vdev->rom + addr,
902            (addr < vdev->rom_size) ? MIN(size, vdev->rom_size - addr) : 0);
903 
904     switch (size) {
905     case 1:
906         data = val.byte;
907         break;
908     case 2:
909         data = le16_to_cpu(val.word);
910         break;
911     case 4:
912         data = le32_to_cpu(val.dword);
913         break;
914     default:
915         hw_error("vfio: unsupported read size, %d bytes\n", size);
916         break;
917     }
918 
919     trace_vfio_rom_read(vdev->vbasedev.name, addr, size, data);
920 
921     return data;
922 }
923 
924 static void vfio_rom_write(void *opaque, hwaddr addr,
925                            uint64_t data, unsigned size)
926 {
927 }
928 
929 static const MemoryRegionOps vfio_rom_ops = {
930     .read = vfio_rom_read,
931     .write = vfio_rom_write,
932     .endianness = DEVICE_LITTLE_ENDIAN,
933 };
934 
935 static void vfio_pci_size_rom(VFIOPCIDevice *vdev)
936 {
937     uint32_t orig, size = cpu_to_le32((uint32_t)PCI_ROM_ADDRESS_MASK);
938     off_t offset = vdev->config_offset + PCI_ROM_ADDRESS;
939     DeviceState *dev = DEVICE(vdev);
940     char *name;
941     int fd = vdev->vbasedev.fd;
942 
943     if (vdev->pdev.romfile || !vdev->pdev.rom_bar) {
944         /* Since pci handles romfile, just print a message and return */
945         if (vfio_blacklist_opt_rom(vdev) && vdev->pdev.romfile) {
946             error_printf("Warning : Device at %s is known to cause system instability issues during option rom execution. Proceeding anyway since user specified romfile\n",
947                          vdev->vbasedev.name);
948         }
949         return;
950     }
951 
952     /*
953      * Use the same size ROM BAR as the physical device.  The contents
954      * will get filled in later when the guest tries to read it.
955      */
956     if (pread(fd, &orig, 4, offset) != 4 ||
957         pwrite(fd, &size, 4, offset) != 4 ||
958         pread(fd, &size, 4, offset) != 4 ||
959         pwrite(fd, &orig, 4, offset) != 4) {
960         error_report("%s(%s) failed: %m", __func__, vdev->vbasedev.name);
961         return;
962     }
963 
964     size = ~(le32_to_cpu(size) & PCI_ROM_ADDRESS_MASK) + 1;
965 
966     if (!size) {
967         return;
968     }
969 
970     if (vfio_blacklist_opt_rom(vdev)) {
971         if (dev->opts && qemu_opt_get(dev->opts, "rombar")) {
972             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",
973                          vdev->vbasedev.name);
974         } else {
975             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",
976                          vdev->vbasedev.name);
977             return;
978         }
979     }
980 
981     trace_vfio_pci_size_rom(vdev->vbasedev.name, size);
982 
983     name = g_strdup_printf("vfio[%s].rom", vdev->vbasedev.name);
984 
985     memory_region_init_io(&vdev->pdev.rom, OBJECT(vdev),
986                           &vfio_rom_ops, vdev, name, size);
987     g_free(name);
988 
989     pci_register_bar(&vdev->pdev, PCI_ROM_SLOT,
990                      PCI_BASE_ADDRESS_SPACE_MEMORY, &vdev->pdev.rom);
991 
992     vdev->pdev.has_rom = true;
993     vdev->rom_read_failed = false;
994 }
995 
996 void vfio_vga_write(void *opaque, hwaddr addr,
997                            uint64_t data, unsigned size)
998 {
999     VFIOVGARegion *region = opaque;
1000     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1001     union {
1002         uint8_t byte;
1003         uint16_t word;
1004         uint32_t dword;
1005         uint64_t qword;
1006     } buf;
1007     off_t offset = vga->fd_offset + region->offset + addr;
1008 
1009     switch (size) {
1010     case 1:
1011         buf.byte = data;
1012         break;
1013     case 2:
1014         buf.word = cpu_to_le16(data);
1015         break;
1016     case 4:
1017         buf.dword = cpu_to_le32(data);
1018         break;
1019     default:
1020         hw_error("vfio: unsupported write size, %d bytes", size);
1021         break;
1022     }
1023 
1024     if (pwrite(vga->fd, &buf, size, offset) != size) {
1025         error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
1026                      __func__, region->offset + addr, data, size);
1027     }
1028 
1029     trace_vfio_vga_write(region->offset + addr, data, size);
1030 }
1031 
1032 uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size)
1033 {
1034     VFIOVGARegion *region = opaque;
1035     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1036     union {
1037         uint8_t byte;
1038         uint16_t word;
1039         uint32_t dword;
1040         uint64_t qword;
1041     } buf;
1042     uint64_t data = 0;
1043     off_t offset = vga->fd_offset + region->offset + addr;
1044 
1045     if (pread(vga->fd, &buf, size, offset) != size) {
1046         error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
1047                      __func__, region->offset + addr, size);
1048         return (uint64_t)-1;
1049     }
1050 
1051     switch (size) {
1052     case 1:
1053         data = buf.byte;
1054         break;
1055     case 2:
1056         data = le16_to_cpu(buf.word);
1057         break;
1058     case 4:
1059         data = le32_to_cpu(buf.dword);
1060         break;
1061     default:
1062         hw_error("vfio: unsupported read size, %d bytes", size);
1063         break;
1064     }
1065 
1066     trace_vfio_vga_read(region->offset + addr, size, data);
1067 
1068     return data;
1069 }
1070 
1071 static const MemoryRegionOps vfio_vga_ops = {
1072     .read = vfio_vga_read,
1073     .write = vfio_vga_write,
1074     .endianness = DEVICE_LITTLE_ENDIAN,
1075 };
1076 
1077 /*
1078  * Expand memory region of sub-page(size < PAGE_SIZE) MMIO BAR to page
1079  * size if the BAR is in an exclusive page in host so that we could map
1080  * this BAR to guest. But this sub-page BAR may not occupy an exclusive
1081  * page in guest. So we should set the priority of the expanded memory
1082  * region to zero in case of overlap with BARs which share the same page
1083  * with the sub-page BAR in guest. Besides, we should also recover the
1084  * size of this sub-page BAR when its base address is changed in guest
1085  * and not page aligned any more.
1086  */
1087 static void vfio_sub_page_bar_update_mapping(PCIDevice *pdev, int bar)
1088 {
1089     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
1090     VFIORegion *region = &vdev->bars[bar].region;
1091     MemoryRegion *mmap_mr, *region_mr, *base_mr;
1092     PCIIORegion *r;
1093     pcibus_t bar_addr;
1094     uint64_t size = region->size;
1095 
1096     /* Make sure that the whole region is allowed to be mmapped */
1097     if (region->nr_mmaps != 1 || !region->mmaps[0].mmap ||
1098         region->mmaps[0].size != region->size) {
1099         return;
1100     }
1101 
1102     r = &pdev->io_regions[bar];
1103     bar_addr = r->addr;
1104     base_mr = vdev->bars[bar].mr;
1105     region_mr = region->mem;
1106     mmap_mr = &region->mmaps[0].mem;
1107 
1108     /* If BAR is mapped and page aligned, update to fill PAGE_SIZE */
1109     if (bar_addr != PCI_BAR_UNMAPPED &&
1110         !(bar_addr & ~qemu_real_host_page_mask)) {
1111         size = qemu_real_host_page_size;
1112     }
1113 
1114     memory_region_transaction_begin();
1115 
1116     if (vdev->bars[bar].size < size) {
1117         memory_region_set_size(base_mr, size);
1118     }
1119     memory_region_set_size(region_mr, size);
1120     memory_region_set_size(mmap_mr, size);
1121     if (size != vdev->bars[bar].size && memory_region_is_mapped(base_mr)) {
1122         memory_region_del_subregion(r->address_space, base_mr);
1123         memory_region_add_subregion_overlap(r->address_space,
1124                                             bar_addr, base_mr, 0);
1125     }
1126 
1127     memory_region_transaction_commit();
1128 }
1129 
1130 /*
1131  * PCI config space
1132  */
1133 uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
1134 {
1135     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
1136     uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val;
1137 
1138     memcpy(&emu_bits, vdev->emulated_config_bits + addr, len);
1139     emu_bits = le32_to_cpu(emu_bits);
1140 
1141     if (emu_bits) {
1142         emu_val = pci_default_read_config(pdev, addr, len);
1143     }
1144 
1145     if (~emu_bits & (0xffffffffU >> (32 - len * 8))) {
1146         ssize_t ret;
1147 
1148         ret = pread(vdev->vbasedev.fd, &phys_val, len,
1149                     vdev->config_offset + addr);
1150         if (ret != len) {
1151             error_report("%s(%s, 0x%x, 0x%x) failed: %m",
1152                          __func__, vdev->vbasedev.name, addr, len);
1153             return -errno;
1154         }
1155         phys_val = le32_to_cpu(phys_val);
1156     }
1157 
1158     val = (emu_val & emu_bits) | (phys_val & ~emu_bits);
1159 
1160     trace_vfio_pci_read_config(vdev->vbasedev.name, addr, len, val);
1161 
1162     return val;
1163 }
1164 
1165 void vfio_pci_write_config(PCIDevice *pdev,
1166                            uint32_t addr, uint32_t val, int len)
1167 {
1168     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
1169     uint32_t val_le = cpu_to_le32(val);
1170 
1171     trace_vfio_pci_write_config(vdev->vbasedev.name, addr, val, len);
1172 
1173     /* Write everything to VFIO, let it filter out what we can't write */
1174     if (pwrite(vdev->vbasedev.fd, &val_le, len, vdev->config_offset + addr)
1175                 != len) {
1176         error_report("%s(%s, 0x%x, 0x%x, 0x%x) failed: %m",
1177                      __func__, vdev->vbasedev.name, addr, val, len);
1178     }
1179 
1180     /* MSI/MSI-X Enabling/Disabling */
1181     if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
1182         ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
1183         int is_enabled, was_enabled = msi_enabled(pdev);
1184 
1185         pci_default_write_config(pdev, addr, val, len);
1186 
1187         is_enabled = msi_enabled(pdev);
1188 
1189         if (!was_enabled) {
1190             if (is_enabled) {
1191                 vfio_msi_enable(vdev);
1192             }
1193         } else {
1194             if (!is_enabled) {
1195                 vfio_msi_disable(vdev);
1196             } else {
1197                 vfio_update_msi(vdev);
1198             }
1199         }
1200     } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
1201         ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
1202         int is_enabled, was_enabled = msix_enabled(pdev);
1203 
1204         pci_default_write_config(pdev, addr, val, len);
1205 
1206         is_enabled = msix_enabled(pdev);
1207 
1208         if (!was_enabled && is_enabled) {
1209             vfio_msix_enable(vdev);
1210         } else if (was_enabled && !is_enabled) {
1211             vfio_msix_disable(vdev);
1212         }
1213     } else if (ranges_overlap(addr, len, PCI_BASE_ADDRESS_0, 24) ||
1214         range_covers_byte(addr, len, PCI_COMMAND)) {
1215         pcibus_t old_addr[PCI_NUM_REGIONS - 1];
1216         int bar;
1217 
1218         for (bar = 0; bar < PCI_ROM_SLOT; bar++) {
1219             old_addr[bar] = pdev->io_regions[bar].addr;
1220         }
1221 
1222         pci_default_write_config(pdev, addr, val, len);
1223 
1224         for (bar = 0; bar < PCI_ROM_SLOT; bar++) {
1225             if (old_addr[bar] != pdev->io_regions[bar].addr &&
1226                 vdev->bars[bar].region.size > 0 &&
1227                 vdev->bars[bar].region.size < qemu_real_host_page_size) {
1228                 vfio_sub_page_bar_update_mapping(pdev, bar);
1229             }
1230         }
1231     } else {
1232         /* Write everything to QEMU to keep emulated bits correct */
1233         pci_default_write_config(pdev, addr, val, len);
1234     }
1235 }
1236 
1237 /*
1238  * Interrupt setup
1239  */
1240 static void vfio_disable_interrupts(VFIOPCIDevice *vdev)
1241 {
1242     /*
1243      * More complicated than it looks.  Disabling MSI/X transitions the
1244      * device to INTx mode (if supported).  Therefore we need to first
1245      * disable MSI/X and then cleanup by disabling INTx.
1246      */
1247     if (vdev->interrupt == VFIO_INT_MSIX) {
1248         vfio_msix_disable(vdev);
1249     } else if (vdev->interrupt == VFIO_INT_MSI) {
1250         vfio_msi_disable(vdev);
1251     }
1252 
1253     if (vdev->interrupt == VFIO_INT_INTx) {
1254         vfio_intx_disable(vdev);
1255     }
1256 }
1257 
1258 static int vfio_msi_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1259 {
1260     uint16_t ctrl;
1261     bool msi_64bit, msi_maskbit;
1262     int ret, entries;
1263     Error *err = NULL;
1264 
1265     if (pread(vdev->vbasedev.fd, &ctrl, sizeof(ctrl),
1266               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1267         error_setg_errno(errp, errno, "failed reading MSI PCI_CAP_FLAGS");
1268         return -errno;
1269     }
1270     ctrl = le16_to_cpu(ctrl);
1271 
1272     msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
1273     msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
1274     entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
1275 
1276     trace_vfio_msi_setup(vdev->vbasedev.name, pos);
1277 
1278     ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit, &err);
1279     if (ret < 0) {
1280         if (ret == -ENOTSUP) {
1281             return 0;
1282         }
1283         error_prepend(&err, "msi_init failed: ");
1284         error_propagate(errp, err);
1285         return ret;
1286     }
1287     vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
1288 
1289     return 0;
1290 }
1291 
1292 static void vfio_pci_fixup_msix_region(VFIOPCIDevice *vdev)
1293 {
1294     off_t start, end;
1295     VFIORegion *region = &vdev->bars[vdev->msix->table_bar].region;
1296 
1297     /*
1298      * If the host driver allows mapping of a MSIX data, we are going to
1299      * do map the entire BAR and emulate MSIX table on top of that.
1300      */
1301     if (vfio_has_region_cap(&vdev->vbasedev, region->nr,
1302                             VFIO_REGION_INFO_CAP_MSIX_MAPPABLE)) {
1303         return;
1304     }
1305 
1306     /*
1307      * We expect to find a single mmap covering the whole BAR, anything else
1308      * means it's either unsupported or already setup.
1309      */
1310     if (region->nr_mmaps != 1 || region->mmaps[0].offset ||
1311         region->size != region->mmaps[0].size) {
1312         return;
1313     }
1314 
1315     /* MSI-X table start and end aligned to host page size */
1316     start = vdev->msix->table_offset & qemu_real_host_page_mask;
1317     end = REAL_HOST_PAGE_ALIGN((uint64_t)vdev->msix->table_offset +
1318                                (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
1319 
1320     /*
1321      * Does the MSI-X table cover the beginning of the BAR?  The whole BAR?
1322      * NB - Host page size is necessarily a power of two and so is the PCI
1323      * BAR (not counting EA yet), therefore if we have host page aligned
1324      * @start and @end, then any remainder of the BAR before or after those
1325      * must be at least host page sized and therefore mmap'able.
1326      */
1327     if (!start) {
1328         if (end >= region->size) {
1329             region->nr_mmaps = 0;
1330             g_free(region->mmaps);
1331             region->mmaps = NULL;
1332             trace_vfio_msix_fixup(vdev->vbasedev.name,
1333                                   vdev->msix->table_bar, 0, 0);
1334         } else {
1335             region->mmaps[0].offset = end;
1336             region->mmaps[0].size = region->size - end;
1337             trace_vfio_msix_fixup(vdev->vbasedev.name,
1338                               vdev->msix->table_bar, region->mmaps[0].offset,
1339                               region->mmaps[0].offset + region->mmaps[0].size);
1340         }
1341 
1342     /* Maybe it's aligned at the end of the BAR */
1343     } else if (end >= region->size) {
1344         region->mmaps[0].size = start;
1345         trace_vfio_msix_fixup(vdev->vbasedev.name,
1346                               vdev->msix->table_bar, region->mmaps[0].offset,
1347                               region->mmaps[0].offset + region->mmaps[0].size);
1348 
1349     /* Otherwise it must split the BAR */
1350     } else {
1351         region->nr_mmaps = 2;
1352         region->mmaps = g_renew(VFIOMmap, region->mmaps, 2);
1353 
1354         memcpy(&region->mmaps[1], &region->mmaps[0], sizeof(VFIOMmap));
1355 
1356         region->mmaps[0].size = start;
1357         trace_vfio_msix_fixup(vdev->vbasedev.name,
1358                               vdev->msix->table_bar, region->mmaps[0].offset,
1359                               region->mmaps[0].offset + region->mmaps[0].size);
1360 
1361         region->mmaps[1].offset = end;
1362         region->mmaps[1].size = region->size - end;
1363         trace_vfio_msix_fixup(vdev->vbasedev.name,
1364                               vdev->msix->table_bar, region->mmaps[1].offset,
1365                               region->mmaps[1].offset + region->mmaps[1].size);
1366     }
1367 }
1368 
1369 static void vfio_pci_relocate_msix(VFIOPCIDevice *vdev, Error **errp)
1370 {
1371     int target_bar = -1;
1372     size_t msix_sz;
1373 
1374     if (!vdev->msix || vdev->msix_relo == OFF_AUTOPCIBAR_OFF) {
1375         return;
1376     }
1377 
1378     /* The actual minimum size of MSI-X structures */
1379     msix_sz = (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE) +
1380               (QEMU_ALIGN_UP(vdev->msix->entries, 64) / 8);
1381     /* Round up to host pages, we don't want to share a page */
1382     msix_sz = REAL_HOST_PAGE_ALIGN(msix_sz);
1383     /* PCI BARs must be a power of 2 */
1384     msix_sz = pow2ceil(msix_sz);
1385 
1386     if (vdev->msix_relo == OFF_AUTOPCIBAR_AUTO) {
1387         /*
1388          * TODO: Lookup table for known devices.
1389          *
1390          * Logically we might use an algorithm here to select the BAR adding
1391          * the least additional MMIO space, but we cannot programatically
1392          * predict the driver dependency on BAR ordering or sizing, therefore
1393          * 'auto' becomes a lookup for combinations reported to work.
1394          */
1395         if (target_bar < 0) {
1396             error_setg(errp, "No automatic MSI-X relocation available for "
1397                        "device %04x:%04x", vdev->vendor_id, vdev->device_id);
1398             return;
1399         }
1400     } else {
1401         target_bar = (int)(vdev->msix_relo - OFF_AUTOPCIBAR_BAR0);
1402     }
1403 
1404     /* I/O port BARs cannot host MSI-X structures */
1405     if (vdev->bars[target_bar].ioport) {
1406         error_setg(errp, "Invalid MSI-X relocation BAR %d, "
1407                    "I/O port BAR", target_bar);
1408         return;
1409     }
1410 
1411     /* Cannot use a BAR in the "shadow" of a 64-bit BAR */
1412     if (!vdev->bars[target_bar].size &&
1413          target_bar > 0 && vdev->bars[target_bar - 1].mem64) {
1414         error_setg(errp, "Invalid MSI-X relocation BAR %d, "
1415                    "consumed by 64-bit BAR %d", target_bar, target_bar - 1);
1416         return;
1417     }
1418 
1419     /* 2GB max size for 32-bit BARs, cannot double if already > 1G */
1420     if (vdev->bars[target_bar].size > (1 * 1024 * 1024 * 1024) &&
1421         !vdev->bars[target_bar].mem64) {
1422         error_setg(errp, "Invalid MSI-X relocation BAR %d, "
1423                    "no space to extend 32-bit BAR", target_bar);
1424         return;
1425     }
1426 
1427     /*
1428      * If adding a new BAR, test if we can make it 64bit.  We make it
1429      * prefetchable since QEMU MSI-X emulation has no read side effects
1430      * and doing so makes mapping more flexible.
1431      */
1432     if (!vdev->bars[target_bar].size) {
1433         if (target_bar < (PCI_ROM_SLOT - 1) &&
1434             !vdev->bars[target_bar + 1].size) {
1435             vdev->bars[target_bar].mem64 = true;
1436             vdev->bars[target_bar].type = PCI_BASE_ADDRESS_MEM_TYPE_64;
1437         }
1438         vdev->bars[target_bar].type |= PCI_BASE_ADDRESS_MEM_PREFETCH;
1439         vdev->bars[target_bar].size = msix_sz;
1440         vdev->msix->table_offset = 0;
1441     } else {
1442         vdev->bars[target_bar].size = MAX(vdev->bars[target_bar].size * 2,
1443                                           msix_sz * 2);
1444         /*
1445          * Due to above size calc, MSI-X always starts halfway into the BAR,
1446          * which will always be a separate host page.
1447          */
1448         vdev->msix->table_offset = vdev->bars[target_bar].size / 2;
1449     }
1450 
1451     vdev->msix->table_bar = target_bar;
1452     vdev->msix->pba_bar = target_bar;
1453     /* Requires 8-byte alignment, but PCI_MSIX_ENTRY_SIZE guarantees that */
1454     vdev->msix->pba_offset = vdev->msix->table_offset +
1455                                   (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE);
1456 
1457     trace_vfio_msix_relo(vdev->vbasedev.name,
1458                          vdev->msix->table_bar, vdev->msix->table_offset);
1459 }
1460 
1461 /*
1462  * We don't have any control over how pci_add_capability() inserts
1463  * capabilities into the chain.  In order to setup MSI-X we need a
1464  * MemoryRegion for the BAR.  In order to setup the BAR and not
1465  * attempt to mmap the MSI-X table area, which VFIO won't allow, we
1466  * need to first look for where the MSI-X table lives.  So we
1467  * unfortunately split MSI-X setup across two functions.
1468  */
1469 static void vfio_msix_early_setup(VFIOPCIDevice *vdev, Error **errp)
1470 {
1471     uint8_t pos;
1472     uint16_t ctrl;
1473     uint32_t table, pba;
1474     int fd = vdev->vbasedev.fd;
1475     VFIOMSIXInfo *msix;
1476 
1477     pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
1478     if (!pos) {
1479         return;
1480     }
1481 
1482     if (pread(fd, &ctrl, sizeof(ctrl),
1483               vdev->config_offset + pos + PCI_MSIX_FLAGS) != sizeof(ctrl)) {
1484         error_setg_errno(errp, errno, "failed to read PCI MSIX FLAGS");
1485         return;
1486     }
1487 
1488     if (pread(fd, &table, sizeof(table),
1489               vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
1490         error_setg_errno(errp, errno, "failed to read PCI MSIX TABLE");
1491         return;
1492     }
1493 
1494     if (pread(fd, &pba, sizeof(pba),
1495               vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
1496         error_setg_errno(errp, errno, "failed to read PCI MSIX PBA");
1497         return;
1498     }
1499 
1500     ctrl = le16_to_cpu(ctrl);
1501     table = le32_to_cpu(table);
1502     pba = le32_to_cpu(pba);
1503 
1504     msix = g_malloc0(sizeof(*msix));
1505     msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
1506     msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
1507     msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
1508     msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
1509     msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
1510 
1511     /*
1512      * Test the size of the pba_offset variable and catch if it extends outside
1513      * of the specified BAR. If it is the case, we need to apply a hardware
1514      * specific quirk if the device is known or we have a broken configuration.
1515      */
1516     if (msix->pba_offset >= vdev->bars[msix->pba_bar].region.size) {
1517         /*
1518          * Chelsio T5 Virtual Function devices are encoded as 0x58xx for T5
1519          * adapters. The T5 hardware returns an incorrect value of 0x8000 for
1520          * the VF PBA offset while the BAR itself is only 8k. The correct value
1521          * is 0x1000, so we hard code that here.
1522          */
1523         if (vdev->vendor_id == PCI_VENDOR_ID_CHELSIO &&
1524             (vdev->device_id & 0xff00) == 0x5800) {
1525             msix->pba_offset = 0x1000;
1526         } else {
1527             error_setg(errp, "hardware reports invalid configuration, "
1528                        "MSIX PBA outside of specified BAR");
1529             g_free(msix);
1530             return;
1531         }
1532     }
1533 
1534     trace_vfio_msix_early_setup(vdev->vbasedev.name, pos, msix->table_bar,
1535                                 msix->table_offset, msix->entries);
1536     vdev->msix = msix;
1537 
1538     vfio_pci_fixup_msix_region(vdev);
1539 
1540     vfio_pci_relocate_msix(vdev, errp);
1541 }
1542 
1543 static int vfio_msix_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1544 {
1545     int ret;
1546     Error *err = NULL;
1547 
1548     vdev->msix->pending = g_malloc0(BITS_TO_LONGS(vdev->msix->entries) *
1549                                     sizeof(unsigned long));
1550     ret = msix_init(&vdev->pdev, vdev->msix->entries,
1551                     vdev->bars[vdev->msix->table_bar].mr,
1552                     vdev->msix->table_bar, vdev->msix->table_offset,
1553                     vdev->bars[vdev->msix->pba_bar].mr,
1554                     vdev->msix->pba_bar, vdev->msix->pba_offset, pos,
1555                     &err);
1556     if (ret < 0) {
1557         if (ret == -ENOTSUP) {
1558             error_report_err(err);
1559             return 0;
1560         }
1561 
1562         error_propagate(errp, err);
1563         return ret;
1564     }
1565 
1566     /*
1567      * The PCI spec suggests that devices provide additional alignment for
1568      * MSI-X structures and avoid overlapping non-MSI-X related registers.
1569      * For an assigned device, this hopefully means that emulation of MSI-X
1570      * structures does not affect the performance of the device.  If devices
1571      * fail to provide that alignment, a significant performance penalty may
1572      * result, for instance Mellanox MT27500 VFs:
1573      * http://www.spinics.net/lists/kvm/msg125881.html
1574      *
1575      * The PBA is simply not that important for such a serious regression and
1576      * most drivers do not appear to look at it.  The solution for this is to
1577      * disable the PBA MemoryRegion unless it's being used.  We disable it
1578      * here and only enable it if a masked vector fires through QEMU.  As the
1579      * vector-use notifier is called, which occurs on unmask, we test whether
1580      * PBA emulation is needed and again disable if not.
1581      */
1582     memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
1583 
1584     /*
1585      * The emulated machine may provide a paravirt interface for MSIX setup
1586      * so it is not strictly necessary to emulate MSIX here. This becomes
1587      * helpful when frequently accessed MMIO registers are located in
1588      * subpages adjacent to the MSIX table but the MSIX data containing page
1589      * cannot be mapped because of a host page size bigger than the MSIX table
1590      * alignment.
1591      */
1592     if (object_property_get_bool(OBJECT(qdev_get_machine()),
1593                                  "vfio-no-msix-emulation", NULL)) {
1594         memory_region_set_enabled(&vdev->pdev.msix_table_mmio, false);
1595     }
1596 
1597     return 0;
1598 }
1599 
1600 static void vfio_teardown_msi(VFIOPCIDevice *vdev)
1601 {
1602     msi_uninit(&vdev->pdev);
1603 
1604     if (vdev->msix) {
1605         msix_uninit(&vdev->pdev,
1606                     vdev->bars[vdev->msix->table_bar].mr,
1607                     vdev->bars[vdev->msix->pba_bar].mr);
1608         g_free(vdev->msix->pending);
1609     }
1610 }
1611 
1612 /*
1613  * Resource setup
1614  */
1615 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled)
1616 {
1617     int i;
1618 
1619     for (i = 0; i < PCI_ROM_SLOT; i++) {
1620         vfio_region_mmaps_set_enabled(&vdev->bars[i].region, enabled);
1621     }
1622 }
1623 
1624 static void vfio_bar_prepare(VFIOPCIDevice *vdev, int nr)
1625 {
1626     VFIOBAR *bar = &vdev->bars[nr];
1627 
1628     uint32_t pci_bar;
1629     int ret;
1630 
1631     /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1632     if (!bar->region.size) {
1633         return;
1634     }
1635 
1636     /* Determine what type of BAR this is for registration */
1637     ret = pread(vdev->vbasedev.fd, &pci_bar, sizeof(pci_bar),
1638                 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
1639     if (ret != sizeof(pci_bar)) {
1640         error_report("vfio: Failed to read BAR %d (%m)", nr);
1641         return;
1642     }
1643 
1644     pci_bar = le32_to_cpu(pci_bar);
1645     bar->ioport = (pci_bar & PCI_BASE_ADDRESS_SPACE_IO);
1646     bar->mem64 = bar->ioport ? 0 : (pci_bar & PCI_BASE_ADDRESS_MEM_TYPE_64);
1647     bar->type = pci_bar & (bar->ioport ? ~PCI_BASE_ADDRESS_IO_MASK :
1648                                          ~PCI_BASE_ADDRESS_MEM_MASK);
1649     bar->size = bar->region.size;
1650 }
1651 
1652 static void vfio_bars_prepare(VFIOPCIDevice *vdev)
1653 {
1654     int i;
1655 
1656     for (i = 0; i < PCI_ROM_SLOT; i++) {
1657         vfio_bar_prepare(vdev, i);
1658     }
1659 }
1660 
1661 static void vfio_bar_register(VFIOPCIDevice *vdev, int nr)
1662 {
1663     VFIOBAR *bar = &vdev->bars[nr];
1664     char *name;
1665 
1666     if (!bar->size) {
1667         return;
1668     }
1669 
1670     bar->mr = g_new0(MemoryRegion, 1);
1671     name = g_strdup_printf("%s base BAR %d", vdev->vbasedev.name, nr);
1672     memory_region_init_io(bar->mr, OBJECT(vdev), NULL, NULL, name, bar->size);
1673     g_free(name);
1674 
1675     if (bar->region.size) {
1676         memory_region_add_subregion(bar->mr, 0, bar->region.mem);
1677 
1678         if (vfio_region_mmap(&bar->region)) {
1679             error_report("Failed to mmap %s BAR %d. Performance may be slow",
1680                          vdev->vbasedev.name, nr);
1681         }
1682     }
1683 
1684     pci_register_bar(&vdev->pdev, nr, bar->type, bar->mr);
1685 }
1686 
1687 static void vfio_bars_register(VFIOPCIDevice *vdev)
1688 {
1689     int i;
1690 
1691     for (i = 0; i < PCI_ROM_SLOT; i++) {
1692         vfio_bar_register(vdev, i);
1693     }
1694 }
1695 
1696 static void vfio_bars_exit(VFIOPCIDevice *vdev)
1697 {
1698     int i;
1699 
1700     for (i = 0; i < PCI_ROM_SLOT; i++) {
1701         VFIOBAR *bar = &vdev->bars[i];
1702 
1703         vfio_bar_quirk_exit(vdev, i);
1704         vfio_region_exit(&bar->region);
1705         if (bar->region.size) {
1706             memory_region_del_subregion(bar->mr, bar->region.mem);
1707         }
1708     }
1709 
1710     if (vdev->vga) {
1711         pci_unregister_vga(&vdev->pdev);
1712         vfio_vga_quirk_exit(vdev);
1713     }
1714 }
1715 
1716 static void vfio_bars_finalize(VFIOPCIDevice *vdev)
1717 {
1718     int i;
1719 
1720     for (i = 0; i < PCI_ROM_SLOT; i++) {
1721         VFIOBAR *bar = &vdev->bars[i];
1722 
1723         vfio_bar_quirk_finalize(vdev, i);
1724         vfio_region_finalize(&bar->region);
1725         if (bar->size) {
1726             object_unparent(OBJECT(bar->mr));
1727             g_free(bar->mr);
1728         }
1729     }
1730 
1731     if (vdev->vga) {
1732         vfio_vga_quirk_finalize(vdev);
1733         for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) {
1734             object_unparent(OBJECT(&vdev->vga->region[i].mem));
1735         }
1736         g_free(vdev->vga);
1737     }
1738 }
1739 
1740 /*
1741  * General setup
1742  */
1743 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
1744 {
1745     uint8_t tmp;
1746     uint16_t next = PCI_CONFIG_SPACE_SIZE;
1747 
1748     for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
1749          tmp = pdev->config[tmp + PCI_CAP_LIST_NEXT]) {
1750         if (tmp > pos && tmp < next) {
1751             next = tmp;
1752         }
1753     }
1754 
1755     return next - pos;
1756 }
1757 
1758 
1759 static uint16_t vfio_ext_cap_max_size(const uint8_t *config, uint16_t pos)
1760 {
1761     uint16_t tmp, next = PCIE_CONFIG_SPACE_SIZE;
1762 
1763     for (tmp = PCI_CONFIG_SPACE_SIZE; tmp;
1764         tmp = PCI_EXT_CAP_NEXT(pci_get_long(config + tmp))) {
1765         if (tmp > pos && tmp < next) {
1766             next = tmp;
1767         }
1768     }
1769 
1770     return next - pos;
1771 }
1772 
1773 static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask)
1774 {
1775     pci_set_word(buf, (pci_get_word(buf) & ~mask) | val);
1776 }
1777 
1778 static void vfio_add_emulated_word(VFIOPCIDevice *vdev, int pos,
1779                                    uint16_t val, uint16_t mask)
1780 {
1781     vfio_set_word_bits(vdev->pdev.config + pos, val, mask);
1782     vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask);
1783     vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask);
1784 }
1785 
1786 static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask)
1787 {
1788     pci_set_long(buf, (pci_get_long(buf) & ~mask) | val);
1789 }
1790 
1791 static void vfio_add_emulated_long(VFIOPCIDevice *vdev, int pos,
1792                                    uint32_t val, uint32_t mask)
1793 {
1794     vfio_set_long_bits(vdev->pdev.config + pos, val, mask);
1795     vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask);
1796     vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask);
1797 }
1798 
1799 static int vfio_setup_pcie_cap(VFIOPCIDevice *vdev, int pos, uint8_t size,
1800                                Error **errp)
1801 {
1802     uint16_t flags;
1803     uint8_t type;
1804 
1805     flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS);
1806     type = (flags & PCI_EXP_FLAGS_TYPE) >> 4;
1807 
1808     if (type != PCI_EXP_TYPE_ENDPOINT &&
1809         type != PCI_EXP_TYPE_LEG_END &&
1810         type != PCI_EXP_TYPE_RC_END) {
1811 
1812         error_setg(errp, "assignment of PCIe type 0x%x "
1813                    "devices is not currently supported", type);
1814         return -EINVAL;
1815     }
1816 
1817     if (!pci_bus_is_express(pci_get_bus(&vdev->pdev))) {
1818         PCIBus *bus = pci_get_bus(&vdev->pdev);
1819         PCIDevice *bridge;
1820 
1821         /*
1822          * Traditionally PCI device assignment exposes the PCIe capability
1823          * as-is on non-express buses.  The reason being that some drivers
1824          * simply assume that it's there, for example tg3.  However when
1825          * we're running on a native PCIe machine type, like Q35, we need
1826          * to hide the PCIe capability.  The reason for this is twofold;
1827          * first Windows guests get a Code 10 error when the PCIe capability
1828          * is exposed in this configuration.  Therefore express devices won't
1829          * work at all unless they're attached to express buses in the VM.
1830          * Second, a native PCIe machine introduces the possibility of fine
1831          * granularity IOMMUs supporting both translation and isolation.
1832          * Guest code to discover the IOMMU visibility of a device, such as
1833          * IOMMU grouping code on Linux, is very aware of device types and
1834          * valid transitions between bus types.  An express device on a non-
1835          * express bus is not a valid combination on bare metal systems.
1836          *
1837          * Drivers that require a PCIe capability to make the device
1838          * functional are simply going to need to have their devices placed
1839          * on a PCIe bus in the VM.
1840          */
1841         while (!pci_bus_is_root(bus)) {
1842             bridge = pci_bridge_get_device(bus);
1843             bus = pci_get_bus(bridge);
1844         }
1845 
1846         if (pci_bus_is_express(bus)) {
1847             return 0;
1848         }
1849 
1850     } else if (pci_bus_is_root(pci_get_bus(&vdev->pdev))) {
1851         /*
1852          * On a Root Complex bus Endpoints become Root Complex Integrated
1853          * Endpoints, which changes the type and clears the LNK & LNK2 fields.
1854          */
1855         if (type == PCI_EXP_TYPE_ENDPOINT) {
1856             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1857                                    PCI_EXP_TYPE_RC_END << 4,
1858                                    PCI_EXP_FLAGS_TYPE);
1859 
1860             /* Link Capabilities, Status, and Control goes away */
1861             if (size > PCI_EXP_LNKCTL) {
1862                 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0);
1863                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1864                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0);
1865 
1866 #ifndef PCI_EXP_LNKCAP2
1867 #define PCI_EXP_LNKCAP2 44
1868 #endif
1869 #ifndef PCI_EXP_LNKSTA2
1870 #define PCI_EXP_LNKSTA2 50
1871 #endif
1872                 /* Link 2 Capabilities, Status, and Control goes away */
1873                 if (size > PCI_EXP_LNKCAP2) {
1874                     vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0);
1875                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0);
1876                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0);
1877                 }
1878             }
1879 
1880         } else if (type == PCI_EXP_TYPE_LEG_END) {
1881             /*
1882              * Legacy endpoints don't belong on the root complex.  Windows
1883              * seems to be happier with devices if we skip the capability.
1884              */
1885             return 0;
1886         }
1887 
1888     } else {
1889         /*
1890          * Convert Root Complex Integrated Endpoints to regular endpoints.
1891          * These devices don't support LNK/LNK2 capabilities, so make them up.
1892          */
1893         if (type == PCI_EXP_TYPE_RC_END) {
1894             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1895                                    PCI_EXP_TYPE_ENDPOINT << 4,
1896                                    PCI_EXP_FLAGS_TYPE);
1897             vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP,
1898                                    PCI_EXP_LNK_MLW_1 | PCI_EXP_LNK_LS_25, ~0);
1899             vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1900         }
1901 
1902         /* Mark the Link Status bits as emulated to allow virtual negotiation */
1903         vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA,
1904                                pci_get_word(vdev->pdev.config + pos +
1905                                             PCI_EXP_LNKSTA),
1906                                PCI_EXP_LNKCAP_MLW | PCI_EXP_LNKCAP_SLS);
1907     }
1908 
1909     /*
1910      * Intel 82599 SR-IOV VFs report an invalid PCIe capability version 0
1911      * (Niantic errate #35) causing Windows to error with a Code 10 for the
1912      * device on Q35.  Fixup any such devices to report version 1.  If we
1913      * were to remove the capability entirely the guest would lose extended
1914      * config space.
1915      */
1916     if ((flags & PCI_EXP_FLAGS_VERS) == 0) {
1917         vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1918                                1, PCI_EXP_FLAGS_VERS);
1919     }
1920 
1921     pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size,
1922                              errp);
1923     if (pos < 0) {
1924         return pos;
1925     }
1926 
1927     vdev->pdev.exp.exp_cap = pos;
1928 
1929     return pos;
1930 }
1931 
1932 static void vfio_check_pcie_flr(VFIOPCIDevice *vdev, uint8_t pos)
1933 {
1934     uint32_t cap = pci_get_long(vdev->pdev.config + pos + PCI_EXP_DEVCAP);
1935 
1936     if (cap & PCI_EXP_DEVCAP_FLR) {
1937         trace_vfio_check_pcie_flr(vdev->vbasedev.name);
1938         vdev->has_flr = true;
1939     }
1940 }
1941 
1942 static void vfio_check_pm_reset(VFIOPCIDevice *vdev, uint8_t pos)
1943 {
1944     uint16_t csr = pci_get_word(vdev->pdev.config + pos + PCI_PM_CTRL);
1945 
1946     if (!(csr & PCI_PM_CTRL_NO_SOFT_RESET)) {
1947         trace_vfio_check_pm_reset(vdev->vbasedev.name);
1948         vdev->has_pm_reset = true;
1949     }
1950 }
1951 
1952 static void vfio_check_af_flr(VFIOPCIDevice *vdev, uint8_t pos)
1953 {
1954     uint8_t cap = pci_get_byte(vdev->pdev.config + pos + PCI_AF_CAP);
1955 
1956     if ((cap & PCI_AF_CAP_TP) && (cap & PCI_AF_CAP_FLR)) {
1957         trace_vfio_check_af_flr(vdev->vbasedev.name);
1958         vdev->has_flr = true;
1959     }
1960 }
1961 
1962 static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos, Error **errp)
1963 {
1964     PCIDevice *pdev = &vdev->pdev;
1965     uint8_t cap_id, next, size;
1966     int ret;
1967 
1968     cap_id = pdev->config[pos];
1969     next = pdev->config[pos + PCI_CAP_LIST_NEXT];
1970 
1971     /*
1972      * If it becomes important to configure capabilities to their actual
1973      * size, use this as the default when it's something we don't recognize.
1974      * Since QEMU doesn't actually handle many of the config accesses,
1975      * exact size doesn't seem worthwhile.
1976      */
1977     size = vfio_std_cap_max_size(pdev, pos);
1978 
1979     /*
1980      * pci_add_capability always inserts the new capability at the head
1981      * of the chain.  Therefore to end up with a chain that matches the
1982      * physical device, we insert from the end by making this recursive.
1983      * This is also why we pre-calculate size above as cached config space
1984      * will be changed as we unwind the stack.
1985      */
1986     if (next) {
1987         ret = vfio_add_std_cap(vdev, next, errp);
1988         if (ret) {
1989             return ret;
1990         }
1991     } else {
1992         /* Begin the rebuild, use QEMU emulated list bits */
1993         pdev->config[PCI_CAPABILITY_LIST] = 0;
1994         vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
1995         vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
1996 
1997         ret = vfio_add_virt_caps(vdev, errp);
1998         if (ret) {
1999             return ret;
2000         }
2001     }
2002 
2003     /* Scale down size, esp in case virt caps were added above */
2004     size = MIN(size, vfio_std_cap_max_size(pdev, pos));
2005 
2006     /* Use emulated next pointer to allow dropping caps */
2007     pci_set_byte(vdev->emulated_config_bits + pos + PCI_CAP_LIST_NEXT, 0xff);
2008 
2009     switch (cap_id) {
2010     case PCI_CAP_ID_MSI:
2011         ret = vfio_msi_setup(vdev, pos, errp);
2012         break;
2013     case PCI_CAP_ID_EXP:
2014         vfio_check_pcie_flr(vdev, pos);
2015         ret = vfio_setup_pcie_cap(vdev, pos, size, errp);
2016         break;
2017     case PCI_CAP_ID_MSIX:
2018         ret = vfio_msix_setup(vdev, pos, errp);
2019         break;
2020     case PCI_CAP_ID_PM:
2021         vfio_check_pm_reset(vdev, pos);
2022         vdev->pm_cap = pos;
2023         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
2024         break;
2025     case PCI_CAP_ID_AF:
2026         vfio_check_af_flr(vdev, pos);
2027         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
2028         break;
2029     default:
2030         ret = pci_add_capability(pdev, cap_id, pos, size, errp);
2031         break;
2032     }
2033 
2034     if (ret < 0) {
2035         error_prepend(errp,
2036                       "failed to add PCI capability 0x%x[0x%x]@0x%x: ",
2037                       cap_id, size, pos);
2038         return ret;
2039     }
2040 
2041     return 0;
2042 }
2043 
2044 static void vfio_add_ext_cap(VFIOPCIDevice *vdev)
2045 {
2046     PCIDevice *pdev = &vdev->pdev;
2047     uint32_t header;
2048     uint16_t cap_id, next, size;
2049     uint8_t cap_ver;
2050     uint8_t *config;
2051 
2052     /* Only add extended caps if we have them and the guest can see them */
2053     if (!pci_is_express(pdev) || !pci_bus_is_express(pci_get_bus(pdev)) ||
2054         !pci_get_long(pdev->config + PCI_CONFIG_SPACE_SIZE)) {
2055         return;
2056     }
2057 
2058     /*
2059      * pcie_add_capability always inserts the new capability at the tail
2060      * of the chain.  Therefore to end up with a chain that matches the
2061      * physical device, we cache the config space to avoid overwriting
2062      * the original config space when we parse the extended capabilities.
2063      */
2064     config = g_memdup(pdev->config, vdev->config_size);
2065 
2066     /*
2067      * Extended capabilities are chained with each pointing to the next, so we
2068      * can drop anything other than the head of the chain simply by modifying
2069      * the previous next pointer.  Seed the head of the chain here such that
2070      * we can simply skip any capabilities we want to drop below, regardless
2071      * of their position in the chain.  If this stub capability still exists
2072      * after we add the capabilities we want to expose, update the capability
2073      * ID to zero.  Note that we cannot seed with the capability header being
2074      * zero as this conflicts with definition of an absent capability chain
2075      * and prevents capabilities beyond the head of the list from being added.
2076      * By replacing the dummy capability ID with zero after walking the device
2077      * chain, we also transparently mark extended capabilities as absent if
2078      * no capabilities were added.  Note that the PCIe spec defines an absence
2079      * of extended capabilities to be determined by a value of zero for the
2080      * capability ID, version, AND next pointer.  A non-zero next pointer
2081      * should be sufficient to indicate additional capabilities are present,
2082      * which will occur if we call pcie_add_capability() below.  The entire
2083      * first dword is emulated to support this.
2084      *
2085      * NB. The kernel side does similar masking, so be prepared that our
2086      * view of the device may also contain a capability ID zero in the head
2087      * of the chain.  Skip it for the same reason that we cannot seed the
2088      * chain with a zero capability.
2089      */
2090     pci_set_long(pdev->config + PCI_CONFIG_SPACE_SIZE,
2091                  PCI_EXT_CAP(0xFFFF, 0, 0));
2092     pci_set_long(pdev->wmask + PCI_CONFIG_SPACE_SIZE, 0);
2093     pci_set_long(vdev->emulated_config_bits + PCI_CONFIG_SPACE_SIZE, ~0);
2094 
2095     for (next = PCI_CONFIG_SPACE_SIZE; next;
2096          next = PCI_EXT_CAP_NEXT(pci_get_long(config + next))) {
2097         header = pci_get_long(config + next);
2098         cap_id = PCI_EXT_CAP_ID(header);
2099         cap_ver = PCI_EXT_CAP_VER(header);
2100 
2101         /*
2102          * If it becomes important to configure extended capabilities to their
2103          * actual size, use this as the default when it's something we don't
2104          * recognize. Since QEMU doesn't actually handle many of the config
2105          * accesses, exact size doesn't seem worthwhile.
2106          */
2107         size = vfio_ext_cap_max_size(config, next);
2108 
2109         /* Use emulated next pointer to allow dropping extended caps */
2110         pci_long_test_and_set_mask(vdev->emulated_config_bits + next,
2111                                    PCI_EXT_CAP_NEXT_MASK);
2112 
2113         switch (cap_id) {
2114         case 0: /* kernel masked capability */
2115         case PCI_EXT_CAP_ID_SRIOV: /* Read-only VF BARs confuse OVMF */
2116         case PCI_EXT_CAP_ID_ARI: /* XXX Needs next function virtualization */
2117             trace_vfio_add_ext_cap_dropped(vdev->vbasedev.name, cap_id, next);
2118             break;
2119         default:
2120             pcie_add_capability(pdev, cap_id, cap_ver, next, size);
2121         }
2122 
2123     }
2124 
2125     /* Cleanup chain head ID if necessary */
2126     if (pci_get_word(pdev->config + PCI_CONFIG_SPACE_SIZE) == 0xFFFF) {
2127         pci_set_word(pdev->config + PCI_CONFIG_SPACE_SIZE, 0);
2128     }
2129 
2130     g_free(config);
2131     return;
2132 }
2133 
2134 static int vfio_add_capabilities(VFIOPCIDevice *vdev, Error **errp)
2135 {
2136     PCIDevice *pdev = &vdev->pdev;
2137     int ret;
2138 
2139     if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
2140         !pdev->config[PCI_CAPABILITY_LIST]) {
2141         return 0; /* Nothing to add */
2142     }
2143 
2144     ret = vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST], errp);
2145     if (ret) {
2146         return ret;
2147     }
2148 
2149     vfio_add_ext_cap(vdev);
2150     return 0;
2151 }
2152 
2153 static void vfio_pci_pre_reset(VFIOPCIDevice *vdev)
2154 {
2155     PCIDevice *pdev = &vdev->pdev;
2156     uint16_t cmd;
2157 
2158     vfio_disable_interrupts(vdev);
2159 
2160     /* Make sure the device is in D0 */
2161     if (vdev->pm_cap) {
2162         uint16_t pmcsr;
2163         uint8_t state;
2164 
2165         pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2166         state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2167         if (state) {
2168             pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
2169             vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2);
2170             /* vfio handles the necessary delay here */
2171             pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2172             state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2173             if (state) {
2174                 error_report("vfio: Unable to power on device, stuck in D%d",
2175                              state);
2176             }
2177         }
2178     }
2179 
2180     /*
2181      * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
2182      * Also put INTx Disable in known state.
2183      */
2184     cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
2185     cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
2186              PCI_COMMAND_INTX_DISABLE);
2187     vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
2188 }
2189 
2190 static void vfio_pci_post_reset(VFIOPCIDevice *vdev)
2191 {
2192     Error *err = NULL;
2193     int nr;
2194 
2195     vfio_intx_enable(vdev, &err);
2196     if (err) {
2197         error_reportf_err(err, ERR_PREFIX, vdev->vbasedev.name);
2198     }
2199 
2200     for (nr = 0; nr < PCI_NUM_REGIONS - 1; ++nr) {
2201         off_t addr = vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr);
2202         uint32_t val = 0;
2203         uint32_t len = sizeof(val);
2204 
2205         if (pwrite(vdev->vbasedev.fd, &val, len, addr) != len) {
2206             error_report("%s(%s) reset bar %d failed: %m", __func__,
2207                          vdev->vbasedev.name, nr);
2208         }
2209     }
2210 
2211     vfio_quirk_reset(vdev);
2212 }
2213 
2214 static bool vfio_pci_host_match(PCIHostDeviceAddress *addr, const char *name)
2215 {
2216     char tmp[13];
2217 
2218     sprintf(tmp, "%04x:%02x:%02x.%1x", addr->domain,
2219             addr->bus, addr->slot, addr->function);
2220 
2221     return (strcmp(tmp, name) == 0);
2222 }
2223 
2224 static int vfio_pci_hot_reset(VFIOPCIDevice *vdev, bool single)
2225 {
2226     VFIOGroup *group;
2227     struct vfio_pci_hot_reset_info *info;
2228     struct vfio_pci_dependent_device *devices;
2229     struct vfio_pci_hot_reset *reset;
2230     int32_t *fds;
2231     int ret, i, count;
2232     bool multi = false;
2233 
2234     trace_vfio_pci_hot_reset(vdev->vbasedev.name, single ? "one" : "multi");
2235 
2236     if (!single) {
2237         vfio_pci_pre_reset(vdev);
2238     }
2239     vdev->vbasedev.needs_reset = false;
2240 
2241     info = g_malloc0(sizeof(*info));
2242     info->argsz = sizeof(*info);
2243 
2244     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2245     if (ret && errno != ENOSPC) {
2246         ret = -errno;
2247         if (!vdev->has_pm_reset) {
2248             error_report("vfio: Cannot reset device %s, "
2249                          "no available reset mechanism.", vdev->vbasedev.name);
2250         }
2251         goto out_single;
2252     }
2253 
2254     count = info->count;
2255     info = g_realloc(info, sizeof(*info) + (count * sizeof(*devices)));
2256     info->argsz = sizeof(*info) + (count * sizeof(*devices));
2257     devices = &info->devices[0];
2258 
2259     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2260     if (ret) {
2261         ret = -errno;
2262         error_report("vfio: hot reset info failed: %m");
2263         goto out_single;
2264     }
2265 
2266     trace_vfio_pci_hot_reset_has_dep_devices(vdev->vbasedev.name);
2267 
2268     /* Verify that we have all the groups required */
2269     for (i = 0; i < info->count; i++) {
2270         PCIHostDeviceAddress host;
2271         VFIOPCIDevice *tmp;
2272         VFIODevice *vbasedev_iter;
2273 
2274         host.domain = devices[i].segment;
2275         host.bus = devices[i].bus;
2276         host.slot = PCI_SLOT(devices[i].devfn);
2277         host.function = PCI_FUNC(devices[i].devfn);
2278 
2279         trace_vfio_pci_hot_reset_dep_devices(host.domain,
2280                 host.bus, host.slot, host.function, devices[i].group_id);
2281 
2282         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2283             continue;
2284         }
2285 
2286         QLIST_FOREACH(group, &vfio_group_list, next) {
2287             if (group->groupid == devices[i].group_id) {
2288                 break;
2289             }
2290         }
2291 
2292         if (!group) {
2293             if (!vdev->has_pm_reset) {
2294                 error_report("vfio: Cannot reset device %s, "
2295                              "depends on group %d which is not owned.",
2296                              vdev->vbasedev.name, devices[i].group_id);
2297             }
2298             ret = -EPERM;
2299             goto out;
2300         }
2301 
2302         /* Prep dependent devices for reset and clear our marker. */
2303         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2304             if (!vbasedev_iter->dev->realized ||
2305                 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2306                 continue;
2307             }
2308             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2309             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2310                 if (single) {
2311                     ret = -EINVAL;
2312                     goto out_single;
2313                 }
2314                 vfio_pci_pre_reset(tmp);
2315                 tmp->vbasedev.needs_reset = false;
2316                 multi = true;
2317                 break;
2318             }
2319         }
2320     }
2321 
2322     if (!single && !multi) {
2323         ret = -EINVAL;
2324         goto out_single;
2325     }
2326 
2327     /* Determine how many group fds need to be passed */
2328     count = 0;
2329     QLIST_FOREACH(group, &vfio_group_list, next) {
2330         for (i = 0; i < info->count; i++) {
2331             if (group->groupid == devices[i].group_id) {
2332                 count++;
2333                 break;
2334             }
2335         }
2336     }
2337 
2338     reset = g_malloc0(sizeof(*reset) + (count * sizeof(*fds)));
2339     reset->argsz = sizeof(*reset) + (count * sizeof(*fds));
2340     fds = &reset->group_fds[0];
2341 
2342     /* Fill in group fds */
2343     QLIST_FOREACH(group, &vfio_group_list, next) {
2344         for (i = 0; i < info->count; i++) {
2345             if (group->groupid == devices[i].group_id) {
2346                 fds[reset->count++] = group->fd;
2347                 break;
2348             }
2349         }
2350     }
2351 
2352     /* Bus reset! */
2353     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_PCI_HOT_RESET, reset);
2354     g_free(reset);
2355 
2356     trace_vfio_pci_hot_reset_result(vdev->vbasedev.name,
2357                                     ret ? "%m" : "Success");
2358 
2359 out:
2360     /* Re-enable INTx on affected devices */
2361     for (i = 0; i < info->count; i++) {
2362         PCIHostDeviceAddress host;
2363         VFIOPCIDevice *tmp;
2364         VFIODevice *vbasedev_iter;
2365 
2366         host.domain = devices[i].segment;
2367         host.bus = devices[i].bus;
2368         host.slot = PCI_SLOT(devices[i].devfn);
2369         host.function = PCI_FUNC(devices[i].devfn);
2370 
2371         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2372             continue;
2373         }
2374 
2375         QLIST_FOREACH(group, &vfio_group_list, next) {
2376             if (group->groupid == devices[i].group_id) {
2377                 break;
2378             }
2379         }
2380 
2381         if (!group) {
2382             break;
2383         }
2384 
2385         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2386             if (!vbasedev_iter->dev->realized ||
2387                 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2388                 continue;
2389             }
2390             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2391             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2392                 vfio_pci_post_reset(tmp);
2393                 break;
2394             }
2395         }
2396     }
2397 out_single:
2398     if (!single) {
2399         vfio_pci_post_reset(vdev);
2400     }
2401     g_free(info);
2402 
2403     return ret;
2404 }
2405 
2406 /*
2407  * We want to differentiate hot reset of mulitple in-use devices vs hot reset
2408  * of a single in-use device.  VFIO_DEVICE_RESET will already handle the case
2409  * of doing hot resets when there is only a single device per bus.  The in-use
2410  * here refers to how many VFIODevices are affected.  A hot reset that affects
2411  * multiple devices, but only a single in-use device, means that we can call
2412  * it from our bus ->reset() callback since the extent is effectively a single
2413  * device.  This allows us to make use of it in the hotplug path.  When there
2414  * are multiple in-use devices, we can only trigger the hot reset during a
2415  * system reset and thus from our reset handler.  We separate _one vs _multi
2416  * here so that we don't overlap and do a double reset on the system reset
2417  * path where both our reset handler and ->reset() callback are used.  Calling
2418  * _one() will only do a hot reset for the one in-use devices case, calling
2419  * _multi() will do nothing if a _one() would have been sufficient.
2420  */
2421 static int vfio_pci_hot_reset_one(VFIOPCIDevice *vdev)
2422 {
2423     return vfio_pci_hot_reset(vdev, true);
2424 }
2425 
2426 static int vfio_pci_hot_reset_multi(VFIODevice *vbasedev)
2427 {
2428     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2429     return vfio_pci_hot_reset(vdev, false);
2430 }
2431 
2432 static void vfio_pci_compute_needs_reset(VFIODevice *vbasedev)
2433 {
2434     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2435     if (!vbasedev->reset_works || (!vdev->has_flr && vdev->has_pm_reset)) {
2436         vbasedev->needs_reset = true;
2437     }
2438 }
2439 
2440 static VFIODeviceOps vfio_pci_ops = {
2441     .vfio_compute_needs_reset = vfio_pci_compute_needs_reset,
2442     .vfio_hot_reset_multi = vfio_pci_hot_reset_multi,
2443     .vfio_eoi = vfio_intx_eoi,
2444 };
2445 
2446 int vfio_populate_vga(VFIOPCIDevice *vdev, Error **errp)
2447 {
2448     VFIODevice *vbasedev = &vdev->vbasedev;
2449     struct vfio_region_info *reg_info;
2450     int ret;
2451 
2452     ret = vfio_get_region_info(vbasedev, VFIO_PCI_VGA_REGION_INDEX, &reg_info);
2453     if (ret) {
2454         error_setg_errno(errp, -ret,
2455                          "failed getting region info for VGA region index %d",
2456                          VFIO_PCI_VGA_REGION_INDEX);
2457         return ret;
2458     }
2459 
2460     if (!(reg_info->flags & VFIO_REGION_INFO_FLAG_READ) ||
2461         !(reg_info->flags & VFIO_REGION_INFO_FLAG_WRITE) ||
2462         reg_info->size < 0xbffff + 1) {
2463         error_setg(errp, "unexpected VGA info, flags 0x%lx, size 0x%lx",
2464                    (unsigned long)reg_info->flags,
2465                    (unsigned long)reg_info->size);
2466         g_free(reg_info);
2467         return -EINVAL;
2468     }
2469 
2470     vdev->vga = g_new0(VFIOVGA, 1);
2471 
2472     vdev->vga->fd_offset = reg_info->offset;
2473     vdev->vga->fd = vdev->vbasedev.fd;
2474 
2475     g_free(reg_info);
2476 
2477     vdev->vga->region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE;
2478     vdev->vga->region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM;
2479     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_MEM].quirks);
2480 
2481     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2482                           OBJECT(vdev), &vfio_vga_ops,
2483                           &vdev->vga->region[QEMU_PCI_VGA_MEM],
2484                           "vfio-vga-mmio@0xa0000",
2485                           QEMU_PCI_VGA_MEM_SIZE);
2486 
2487     vdev->vga->region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE;
2488     vdev->vga->region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO;
2489     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].quirks);
2490 
2491     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2492                           OBJECT(vdev), &vfio_vga_ops,
2493                           &vdev->vga->region[QEMU_PCI_VGA_IO_LO],
2494                           "vfio-vga-io@0x3b0",
2495                           QEMU_PCI_VGA_IO_LO_SIZE);
2496 
2497     vdev->vga->region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE;
2498     vdev->vga->region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI;
2499     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks);
2500 
2501     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
2502                           OBJECT(vdev), &vfio_vga_ops,
2503                           &vdev->vga->region[QEMU_PCI_VGA_IO_HI],
2504                           "vfio-vga-io@0x3c0",
2505                           QEMU_PCI_VGA_IO_HI_SIZE);
2506 
2507     pci_register_vga(&vdev->pdev, &vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2508                      &vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2509                      &vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem);
2510 
2511     return 0;
2512 }
2513 
2514 static void vfio_populate_device(VFIOPCIDevice *vdev, Error **errp)
2515 {
2516     VFIODevice *vbasedev = &vdev->vbasedev;
2517     struct vfio_region_info *reg_info;
2518     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) };
2519     int i, ret = -1;
2520 
2521     /* Sanity check device */
2522     if (!(vbasedev->flags & VFIO_DEVICE_FLAGS_PCI)) {
2523         error_setg(errp, "this isn't a PCI device");
2524         return;
2525     }
2526 
2527     if (vbasedev->num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
2528         error_setg(errp, "unexpected number of io regions %u",
2529                    vbasedev->num_regions);
2530         return;
2531     }
2532 
2533     if (vbasedev->num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
2534         error_setg(errp, "unexpected number of irqs %u", vbasedev->num_irqs);
2535         return;
2536     }
2537 
2538     for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
2539         char *name = g_strdup_printf("%s BAR %d", vbasedev->name, i);
2540 
2541         ret = vfio_region_setup(OBJECT(vdev), vbasedev,
2542                                 &vdev->bars[i].region, i, name);
2543         g_free(name);
2544 
2545         if (ret) {
2546             error_setg_errno(errp, -ret, "failed to get region %d info", i);
2547             return;
2548         }
2549 
2550         QLIST_INIT(&vdev->bars[i].quirks);
2551     }
2552 
2553     ret = vfio_get_region_info(vbasedev,
2554                                VFIO_PCI_CONFIG_REGION_INDEX, &reg_info);
2555     if (ret) {
2556         error_setg_errno(errp, -ret, "failed to get config info");
2557         return;
2558     }
2559 
2560     trace_vfio_populate_device_config(vdev->vbasedev.name,
2561                                       (unsigned long)reg_info->size,
2562                                       (unsigned long)reg_info->offset,
2563                                       (unsigned long)reg_info->flags);
2564 
2565     vdev->config_size = reg_info->size;
2566     if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
2567         vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
2568     }
2569     vdev->config_offset = reg_info->offset;
2570 
2571     g_free(reg_info);
2572 
2573     if (vdev->features & VFIO_FEATURE_ENABLE_VGA) {
2574         ret = vfio_populate_vga(vdev, errp);
2575         if (ret) {
2576             error_append_hint(errp, "device does not support "
2577                               "requested feature x-vga\n");
2578             return;
2579         }
2580     }
2581 
2582     irq_info.index = VFIO_PCI_ERR_IRQ_INDEX;
2583 
2584     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info);
2585     if (ret) {
2586         /* This can fail for an old kernel or legacy PCI dev */
2587         trace_vfio_populate_device_get_irq_info_failure();
2588     } else if (irq_info.count == 1) {
2589         vdev->pci_aer = true;
2590     } else {
2591         error_report(WARN_PREFIX
2592                      "Could not enable error recovery for the device",
2593                      vbasedev->name);
2594     }
2595 }
2596 
2597 static void vfio_put_device(VFIOPCIDevice *vdev)
2598 {
2599     g_free(vdev->vbasedev.name);
2600     g_free(vdev->msix);
2601 
2602     vfio_put_base_device(&vdev->vbasedev);
2603 }
2604 
2605 static void vfio_err_notifier_handler(void *opaque)
2606 {
2607     VFIOPCIDevice *vdev = opaque;
2608 
2609     if (!event_notifier_test_and_clear(&vdev->err_notifier)) {
2610         return;
2611     }
2612 
2613     /*
2614      * TBD. Retrieve the error details and decide what action
2615      * needs to be taken. One of the actions could be to pass
2616      * the error to the guest and have the guest driver recover
2617      * from the error. This requires that PCIe capabilities be
2618      * exposed to the guest. For now, we just terminate the
2619      * guest to contain the error.
2620      */
2621 
2622     error_report("%s(%s) Unrecoverable error detected. Please collect any data possible and then kill the guest", __func__, vdev->vbasedev.name);
2623 
2624     vm_stop(RUN_STATE_INTERNAL_ERROR);
2625 }
2626 
2627 /*
2628  * Registers error notifier for devices supporting error recovery.
2629  * If we encounter a failure in this function, we report an error
2630  * and continue after disabling error recovery support for the
2631  * device.
2632  */
2633 static void vfio_register_err_notifier(VFIOPCIDevice *vdev)
2634 {
2635     int ret;
2636     int argsz;
2637     struct vfio_irq_set *irq_set;
2638     int32_t *pfd;
2639 
2640     if (!vdev->pci_aer) {
2641         return;
2642     }
2643 
2644     if (event_notifier_init(&vdev->err_notifier, 0)) {
2645         error_report("vfio: Unable to init event notifier for error detection");
2646         vdev->pci_aer = false;
2647         return;
2648     }
2649 
2650     argsz = sizeof(*irq_set) + sizeof(*pfd);
2651 
2652     irq_set = g_malloc0(argsz);
2653     irq_set->argsz = argsz;
2654     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2655                      VFIO_IRQ_SET_ACTION_TRIGGER;
2656     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
2657     irq_set->start = 0;
2658     irq_set->count = 1;
2659     pfd = (int32_t *)&irq_set->data;
2660 
2661     *pfd = event_notifier_get_fd(&vdev->err_notifier);
2662     qemu_set_fd_handler(*pfd, vfio_err_notifier_handler, NULL, vdev);
2663 
2664     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
2665     if (ret) {
2666         error_report("vfio: Failed to set up error notification");
2667         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
2668         event_notifier_cleanup(&vdev->err_notifier);
2669         vdev->pci_aer = false;
2670     }
2671     g_free(irq_set);
2672 }
2673 
2674 static void vfio_unregister_err_notifier(VFIOPCIDevice *vdev)
2675 {
2676     int argsz;
2677     struct vfio_irq_set *irq_set;
2678     int32_t *pfd;
2679     int ret;
2680 
2681     if (!vdev->pci_aer) {
2682         return;
2683     }
2684 
2685     argsz = sizeof(*irq_set) + sizeof(*pfd);
2686 
2687     irq_set = g_malloc0(argsz);
2688     irq_set->argsz = argsz;
2689     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2690                      VFIO_IRQ_SET_ACTION_TRIGGER;
2691     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
2692     irq_set->start = 0;
2693     irq_set->count = 1;
2694     pfd = (int32_t *)&irq_set->data;
2695     *pfd = -1;
2696 
2697     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
2698     if (ret) {
2699         error_report("vfio: Failed to de-assign error fd: %m");
2700     }
2701     g_free(irq_set);
2702     qemu_set_fd_handler(event_notifier_get_fd(&vdev->err_notifier),
2703                         NULL, NULL, vdev);
2704     event_notifier_cleanup(&vdev->err_notifier);
2705 }
2706 
2707 static void vfio_req_notifier_handler(void *opaque)
2708 {
2709     VFIOPCIDevice *vdev = opaque;
2710     Error *err = NULL;
2711 
2712     if (!event_notifier_test_and_clear(&vdev->req_notifier)) {
2713         return;
2714     }
2715 
2716     qdev_unplug(&vdev->pdev.qdev, &err);
2717     if (err) {
2718         error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name);
2719     }
2720 }
2721 
2722 static void vfio_register_req_notifier(VFIOPCIDevice *vdev)
2723 {
2724     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info),
2725                                       .index = VFIO_PCI_REQ_IRQ_INDEX };
2726     int argsz;
2727     struct vfio_irq_set *irq_set;
2728     int32_t *pfd;
2729 
2730     if (!(vdev->features & VFIO_FEATURE_ENABLE_REQ)) {
2731         return;
2732     }
2733 
2734     if (ioctl(vdev->vbasedev.fd,
2735               VFIO_DEVICE_GET_IRQ_INFO, &irq_info) < 0 || irq_info.count < 1) {
2736         return;
2737     }
2738 
2739     if (event_notifier_init(&vdev->req_notifier, 0)) {
2740         error_report("vfio: Unable to init event notifier for device request");
2741         return;
2742     }
2743 
2744     argsz = sizeof(*irq_set) + sizeof(*pfd);
2745 
2746     irq_set = g_malloc0(argsz);
2747     irq_set->argsz = argsz;
2748     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2749                      VFIO_IRQ_SET_ACTION_TRIGGER;
2750     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
2751     irq_set->start = 0;
2752     irq_set->count = 1;
2753     pfd = (int32_t *)&irq_set->data;
2754 
2755     *pfd = event_notifier_get_fd(&vdev->req_notifier);
2756     qemu_set_fd_handler(*pfd, vfio_req_notifier_handler, NULL, vdev);
2757 
2758     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
2759         error_report("vfio: Failed to set up device request notification");
2760         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
2761         event_notifier_cleanup(&vdev->req_notifier);
2762     } else {
2763         vdev->req_enabled = true;
2764     }
2765 
2766     g_free(irq_set);
2767 }
2768 
2769 static void vfio_unregister_req_notifier(VFIOPCIDevice *vdev)
2770 {
2771     int argsz;
2772     struct vfio_irq_set *irq_set;
2773     int32_t *pfd;
2774 
2775     if (!vdev->req_enabled) {
2776         return;
2777     }
2778 
2779     argsz = sizeof(*irq_set) + sizeof(*pfd);
2780 
2781     irq_set = g_malloc0(argsz);
2782     irq_set->argsz = argsz;
2783     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2784                      VFIO_IRQ_SET_ACTION_TRIGGER;
2785     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
2786     irq_set->start = 0;
2787     irq_set->count = 1;
2788     pfd = (int32_t *)&irq_set->data;
2789     *pfd = -1;
2790 
2791     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
2792         error_report("vfio: Failed to de-assign device request fd: %m");
2793     }
2794     g_free(irq_set);
2795     qemu_set_fd_handler(event_notifier_get_fd(&vdev->req_notifier),
2796                         NULL, NULL, vdev);
2797     event_notifier_cleanup(&vdev->req_notifier);
2798 
2799     vdev->req_enabled = false;
2800 }
2801 
2802 static void vfio_realize(PCIDevice *pdev, Error **errp)
2803 {
2804     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
2805     VFIODevice *vbasedev_iter;
2806     VFIOGroup *group;
2807     char *tmp, group_path[PATH_MAX], *group_name;
2808     Error *err = NULL;
2809     ssize_t len;
2810     struct stat st;
2811     int groupid;
2812     int i, ret;
2813 
2814     if (!vdev->vbasedev.sysfsdev) {
2815         if (!(~vdev->host.domain || ~vdev->host.bus ||
2816               ~vdev->host.slot || ~vdev->host.function)) {
2817             error_setg(errp, "No provided host device");
2818             error_append_hint(errp, "Use -device vfio-pci,host=DDDD:BB:DD.F "
2819                               "or -device vfio-pci,sysfsdev=PATH_TO_DEVICE\n");
2820             return;
2821         }
2822         vdev->vbasedev.sysfsdev =
2823             g_strdup_printf("/sys/bus/pci/devices/%04x:%02x:%02x.%01x",
2824                             vdev->host.domain, vdev->host.bus,
2825                             vdev->host.slot, vdev->host.function);
2826     }
2827 
2828     if (stat(vdev->vbasedev.sysfsdev, &st) < 0) {
2829         error_setg_errno(errp, errno, "no such host device");
2830         error_prepend(errp, ERR_PREFIX, vdev->vbasedev.sysfsdev);
2831         return;
2832     }
2833 
2834     vdev->vbasedev.name = g_path_get_basename(vdev->vbasedev.sysfsdev);
2835     vdev->vbasedev.ops = &vfio_pci_ops;
2836     vdev->vbasedev.type = VFIO_DEVICE_TYPE_PCI;
2837     vdev->vbasedev.dev = &vdev->pdev.qdev;
2838 
2839     tmp = g_strdup_printf("%s/iommu_group", vdev->vbasedev.sysfsdev);
2840     len = readlink(tmp, group_path, sizeof(group_path));
2841     g_free(tmp);
2842 
2843     if (len <= 0 || len >= sizeof(group_path)) {
2844         error_setg_errno(errp, len < 0 ? errno : ENAMETOOLONG,
2845                          "no iommu_group found");
2846         goto error;
2847     }
2848 
2849     group_path[len] = 0;
2850 
2851     group_name = basename(group_path);
2852     if (sscanf(group_name, "%d", &groupid) != 1) {
2853         error_setg_errno(errp, errno, "failed to read %s", group_path);
2854         goto error;
2855     }
2856 
2857     trace_vfio_realize(vdev->vbasedev.name, groupid);
2858 
2859     group = vfio_get_group(groupid, pci_device_iommu_address_space(pdev), errp);
2860     if (!group) {
2861         goto error;
2862     }
2863 
2864     QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2865         if (strcmp(vbasedev_iter->name, vdev->vbasedev.name) == 0) {
2866             error_setg(errp, "device is already attached");
2867             vfio_put_group(group);
2868             goto error;
2869         }
2870     }
2871 
2872     ret = vfio_get_device(group, vdev->vbasedev.name, &vdev->vbasedev, errp);
2873     if (ret) {
2874         vfio_put_group(group);
2875         goto error;
2876     }
2877 
2878     vfio_populate_device(vdev, &err);
2879     if (err) {
2880         error_propagate(errp, err);
2881         goto error;
2882     }
2883 
2884     /* Get a copy of config space */
2885     ret = pread(vdev->vbasedev.fd, vdev->pdev.config,
2886                 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
2887                 vdev->config_offset);
2888     if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
2889         ret = ret < 0 ? -errno : -EFAULT;
2890         error_setg_errno(errp, -ret, "failed to read device config space");
2891         goto error;
2892     }
2893 
2894     /* vfio emulates a lot for us, but some bits need extra love */
2895     vdev->emulated_config_bits = g_malloc0(vdev->config_size);
2896 
2897     /* QEMU can choose to expose the ROM or not */
2898     memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4);
2899     /* QEMU can also add or extend BARs */
2900     memset(vdev->emulated_config_bits + PCI_BASE_ADDRESS_0, 0xff, 6 * 4);
2901 
2902     /*
2903      * The PCI spec reserves vendor ID 0xffff as an invalid value.  The
2904      * device ID is managed by the vendor and need only be a 16-bit value.
2905      * Allow any 16-bit value for subsystem so they can be hidden or changed.
2906      */
2907     if (vdev->vendor_id != PCI_ANY_ID) {
2908         if (vdev->vendor_id >= 0xffff) {
2909             error_setg(errp, "invalid PCI vendor ID provided");
2910             goto error;
2911         }
2912         vfio_add_emulated_word(vdev, PCI_VENDOR_ID, vdev->vendor_id, ~0);
2913         trace_vfio_pci_emulated_vendor_id(vdev->vbasedev.name, vdev->vendor_id);
2914     } else {
2915         vdev->vendor_id = pci_get_word(pdev->config + PCI_VENDOR_ID);
2916     }
2917 
2918     if (vdev->device_id != PCI_ANY_ID) {
2919         if (vdev->device_id > 0xffff) {
2920             error_setg(errp, "invalid PCI device ID provided");
2921             goto error;
2922         }
2923         vfio_add_emulated_word(vdev, PCI_DEVICE_ID, vdev->device_id, ~0);
2924         trace_vfio_pci_emulated_device_id(vdev->vbasedev.name, vdev->device_id);
2925     } else {
2926         vdev->device_id = pci_get_word(pdev->config + PCI_DEVICE_ID);
2927     }
2928 
2929     if (vdev->sub_vendor_id != PCI_ANY_ID) {
2930         if (vdev->sub_vendor_id > 0xffff) {
2931             error_setg(errp, "invalid PCI subsystem vendor ID provided");
2932             goto error;
2933         }
2934         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_VENDOR_ID,
2935                                vdev->sub_vendor_id, ~0);
2936         trace_vfio_pci_emulated_sub_vendor_id(vdev->vbasedev.name,
2937                                               vdev->sub_vendor_id);
2938     }
2939 
2940     if (vdev->sub_device_id != PCI_ANY_ID) {
2941         if (vdev->sub_device_id > 0xffff) {
2942             error_setg(errp, "invalid PCI subsystem device ID provided");
2943             goto error;
2944         }
2945         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_ID, vdev->sub_device_id, ~0);
2946         trace_vfio_pci_emulated_sub_device_id(vdev->vbasedev.name,
2947                                               vdev->sub_device_id);
2948     }
2949 
2950     /* QEMU can change multi-function devices to single function, or reverse */
2951     vdev->emulated_config_bits[PCI_HEADER_TYPE] =
2952                                               PCI_HEADER_TYPE_MULTI_FUNCTION;
2953 
2954     /* Restore or clear multifunction, this is always controlled by QEMU */
2955     if (vdev->pdev.cap_present & QEMU_PCI_CAP_MULTIFUNCTION) {
2956         vdev->pdev.config[PCI_HEADER_TYPE] |= PCI_HEADER_TYPE_MULTI_FUNCTION;
2957     } else {
2958         vdev->pdev.config[PCI_HEADER_TYPE] &= ~PCI_HEADER_TYPE_MULTI_FUNCTION;
2959     }
2960 
2961     /*
2962      * Clear host resource mapping info.  If we choose not to register a
2963      * BAR, such as might be the case with the option ROM, we can get
2964      * confusing, unwritable, residual addresses from the host here.
2965      */
2966     memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
2967     memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
2968 
2969     vfio_pci_size_rom(vdev);
2970 
2971     vfio_bars_prepare(vdev);
2972 
2973     vfio_msix_early_setup(vdev, &err);
2974     if (err) {
2975         error_propagate(errp, err);
2976         goto error;
2977     }
2978 
2979     vfio_bars_register(vdev);
2980 
2981     ret = vfio_add_capabilities(vdev, errp);
2982     if (ret) {
2983         goto out_teardown;
2984     }
2985 
2986     if (vdev->vga) {
2987         vfio_vga_quirk_setup(vdev);
2988     }
2989 
2990     for (i = 0; i < PCI_ROM_SLOT; i++) {
2991         vfio_bar_quirk_setup(vdev, i);
2992     }
2993 
2994     if (!vdev->igd_opregion &&
2995         vdev->features & VFIO_FEATURE_ENABLE_IGD_OPREGION) {
2996         struct vfio_region_info *opregion;
2997 
2998         if (vdev->pdev.qdev.hotplugged) {
2999             error_setg(errp,
3000                        "cannot support IGD OpRegion feature on hotplugged "
3001                        "device");
3002             goto out_teardown;
3003         }
3004 
3005         ret = vfio_get_dev_region_info(&vdev->vbasedev,
3006                         VFIO_REGION_TYPE_PCI_VENDOR_TYPE | PCI_VENDOR_ID_INTEL,
3007                         VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION, &opregion);
3008         if (ret) {
3009             error_setg_errno(errp, -ret,
3010                              "does not support requested IGD OpRegion feature");
3011             goto out_teardown;
3012         }
3013 
3014         ret = vfio_pci_igd_opregion_init(vdev, opregion, errp);
3015         g_free(opregion);
3016         if (ret) {
3017             goto out_teardown;
3018         }
3019     }
3020 
3021     /* QEMU emulates all of MSI & MSIX */
3022     if (pdev->cap_present & QEMU_PCI_CAP_MSIX) {
3023         memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff,
3024                MSIX_CAP_LENGTH);
3025     }
3026 
3027     if (pdev->cap_present & QEMU_PCI_CAP_MSI) {
3028         memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff,
3029                vdev->msi_cap_size);
3030     }
3031 
3032     if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
3033         vdev->intx.mmap_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL,
3034                                                   vfio_intx_mmap_enable, vdev);
3035         pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_intx_update);
3036         ret = vfio_intx_enable(vdev, errp);
3037         if (ret) {
3038             goto out_teardown;
3039         }
3040     }
3041 
3042     if (vdev->display != ON_OFF_AUTO_OFF) {
3043         ret = vfio_display_probe(vdev, errp);
3044         if (ret) {
3045             goto out_teardown;
3046         }
3047     }
3048 
3049     vfio_register_err_notifier(vdev);
3050     vfio_register_req_notifier(vdev);
3051     vfio_setup_resetfn_quirk(vdev);
3052 
3053     return;
3054 
3055 out_teardown:
3056     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3057     vfio_teardown_msi(vdev);
3058     vfio_bars_exit(vdev);
3059 error:
3060     error_prepend(errp, ERR_PREFIX, vdev->vbasedev.name);
3061 }
3062 
3063 static void vfio_instance_finalize(Object *obj)
3064 {
3065     PCIDevice *pci_dev = PCI_DEVICE(obj);
3066     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pci_dev);
3067     VFIOGroup *group = vdev->vbasedev.group;
3068 
3069     vfio_display_finalize(vdev);
3070     vfio_bars_finalize(vdev);
3071     g_free(vdev->emulated_config_bits);
3072     g_free(vdev->rom);
3073     /*
3074      * XXX Leaking igd_opregion is not an oversight, we can't remove the
3075      * fw_cfg entry therefore leaking this allocation seems like the safest
3076      * option.
3077      *
3078      * g_free(vdev->igd_opregion);
3079      */
3080     vfio_put_device(vdev);
3081     vfio_put_group(group);
3082 }
3083 
3084 static void vfio_exitfn(PCIDevice *pdev)
3085 {
3086     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
3087 
3088     vfio_unregister_req_notifier(vdev);
3089     vfio_unregister_err_notifier(vdev);
3090     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3091     vfio_disable_interrupts(vdev);
3092     if (vdev->intx.mmap_timer) {
3093         timer_free(vdev->intx.mmap_timer);
3094     }
3095     vfio_teardown_msi(vdev);
3096     vfio_bars_exit(vdev);
3097 }
3098 
3099 static void vfio_pci_reset(DeviceState *dev)
3100 {
3101     PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
3102     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
3103 
3104     trace_vfio_pci_reset(vdev->vbasedev.name);
3105 
3106     vfio_pci_pre_reset(vdev);
3107 
3108     if (vdev->display != ON_OFF_AUTO_OFF) {
3109         vfio_display_reset(vdev);
3110     }
3111 
3112     if (vdev->resetfn && !vdev->resetfn(vdev)) {
3113         goto post_reset;
3114     }
3115 
3116     if (vdev->vbasedev.reset_works &&
3117         (vdev->has_flr || !vdev->has_pm_reset) &&
3118         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
3119         trace_vfio_pci_reset_flr(vdev->vbasedev.name);
3120         goto post_reset;
3121     }
3122 
3123     /* See if we can do our own bus reset */
3124     if (!vfio_pci_hot_reset_one(vdev)) {
3125         goto post_reset;
3126     }
3127 
3128     /* If nothing else works and the device supports PM reset, use it */
3129     if (vdev->vbasedev.reset_works && vdev->has_pm_reset &&
3130         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
3131         trace_vfio_pci_reset_pm(vdev->vbasedev.name);
3132         goto post_reset;
3133     }
3134 
3135 post_reset:
3136     vfio_pci_post_reset(vdev);
3137 }
3138 
3139 static void vfio_instance_init(Object *obj)
3140 {
3141     PCIDevice *pci_dev = PCI_DEVICE(obj);
3142     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, PCI_DEVICE(obj));
3143 
3144     device_add_bootindex_property(obj, &vdev->bootindex,
3145                                   "bootindex", NULL,
3146                                   &pci_dev->qdev, NULL);
3147     vdev->host.domain = ~0U;
3148     vdev->host.bus = ~0U;
3149     vdev->host.slot = ~0U;
3150     vdev->host.function = ~0U;
3151 
3152     vdev->nv_gpudirect_clique = 0xFF;
3153 
3154     /* QEMU_PCI_CAP_EXPRESS initialization does not depend on QEMU command
3155      * line, therefore, no need to wait to realize like other devices */
3156     pci_dev->cap_present |= QEMU_PCI_CAP_EXPRESS;
3157 }
3158 
3159 static Property vfio_pci_dev_properties[] = {
3160     DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIOPCIDevice, host),
3161     DEFINE_PROP_STRING("sysfsdev", VFIOPCIDevice, vbasedev.sysfsdev),
3162     DEFINE_PROP_ON_OFF_AUTO("display", VFIOPCIDevice,
3163                             display, ON_OFF_AUTO_OFF),
3164     DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIOPCIDevice,
3165                        intx.mmap_timeout, 1100),
3166     DEFINE_PROP_BIT("x-vga", VFIOPCIDevice, features,
3167                     VFIO_FEATURE_ENABLE_VGA_BIT, false),
3168     DEFINE_PROP_BIT("x-req", VFIOPCIDevice, features,
3169                     VFIO_FEATURE_ENABLE_REQ_BIT, true),
3170     DEFINE_PROP_BIT("x-igd-opregion", VFIOPCIDevice, features,
3171                     VFIO_FEATURE_ENABLE_IGD_OPREGION_BIT, false),
3172     DEFINE_PROP_BOOL("x-no-mmap", VFIOPCIDevice, vbasedev.no_mmap, false),
3173     DEFINE_PROP_BOOL("x-no-kvm-intx", VFIOPCIDevice, no_kvm_intx, false),
3174     DEFINE_PROP_BOOL("x-no-kvm-msi", VFIOPCIDevice, no_kvm_msi, false),
3175     DEFINE_PROP_BOOL("x-no-kvm-msix", VFIOPCIDevice, no_kvm_msix, false),
3176     DEFINE_PROP_BOOL("x-no-geforce-quirks", VFIOPCIDevice,
3177                      no_geforce_quirks, false),
3178     DEFINE_PROP_BOOL("x-no-kvm-ioeventfd", VFIOPCIDevice, no_kvm_ioeventfd,
3179                      false),
3180     DEFINE_PROP_BOOL("x-no-vfio-ioeventfd", VFIOPCIDevice, no_vfio_ioeventfd,
3181                      false),
3182     DEFINE_PROP_UINT32("x-pci-vendor-id", VFIOPCIDevice, vendor_id, PCI_ANY_ID),
3183     DEFINE_PROP_UINT32("x-pci-device-id", VFIOPCIDevice, device_id, PCI_ANY_ID),
3184     DEFINE_PROP_UINT32("x-pci-sub-vendor-id", VFIOPCIDevice,
3185                        sub_vendor_id, PCI_ANY_ID),
3186     DEFINE_PROP_UINT32("x-pci-sub-device-id", VFIOPCIDevice,
3187                        sub_device_id, PCI_ANY_ID),
3188     DEFINE_PROP_UINT32("x-igd-gms", VFIOPCIDevice, igd_gms, 0),
3189     DEFINE_PROP_UNSIGNED_NODEFAULT("x-nv-gpudirect-clique", VFIOPCIDevice,
3190                                    nv_gpudirect_clique,
3191                                    qdev_prop_nv_gpudirect_clique, uint8_t),
3192     DEFINE_PROP_OFF_AUTO_PCIBAR("x-msix-relocation", VFIOPCIDevice, msix_relo,
3193                                 OFF_AUTOPCIBAR_OFF),
3194     /*
3195      * TODO - support passed fds... is this necessary?
3196      * DEFINE_PROP_STRING("vfiofd", VFIOPCIDevice, vfiofd_name),
3197      * DEFINE_PROP_STRING("vfiogroupfd, VFIOPCIDevice, vfiogroupfd_name),
3198      */
3199     DEFINE_PROP_END_OF_LIST(),
3200 };
3201 
3202 static const VMStateDescription vfio_pci_vmstate = {
3203     .name = "vfio-pci",
3204     .unmigratable = 1,
3205 };
3206 
3207 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
3208 {
3209     DeviceClass *dc = DEVICE_CLASS(klass);
3210     PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
3211 
3212     dc->reset = vfio_pci_reset;
3213     dc->props = vfio_pci_dev_properties;
3214     dc->vmsd = &vfio_pci_vmstate;
3215     dc->desc = "VFIO-based PCI device assignment";
3216     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
3217     pdc->realize = vfio_realize;
3218     pdc->exit = vfio_exitfn;
3219     pdc->config_read = vfio_pci_read_config;
3220     pdc->config_write = vfio_pci_write_config;
3221 }
3222 
3223 static const TypeInfo vfio_pci_dev_info = {
3224     .name = "vfio-pci",
3225     .parent = TYPE_PCI_DEVICE,
3226     .instance_size = sizeof(VFIOPCIDevice),
3227     .class_init = vfio_pci_dev_class_init,
3228     .instance_init = vfio_instance_init,
3229     .instance_finalize = vfio_instance_finalize,
3230     .interfaces = (InterfaceInfo[]) {
3231         { INTERFACE_PCIE_DEVICE },
3232         { INTERFACE_CONVENTIONAL_PCI_DEVICE },
3233         { }
3234     },
3235 };
3236 
3237 static void register_vfio_pci_dev_type(void)
3238 {
3239     type_register_static(&vfio_pci_dev_info);
3240 }
3241 
3242 type_init(register_vfio_pci_dev_type)
3243