xref: /openbmc/qemu/hw/vfio/pci.c (revision e4370165)
1 /*
2  * vfio based device assignment support
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Alex Williamson <alex.williamson@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
12  * Based on qemu-kvm device-assignment:
13  *  Adapted for KVM by Qumranet.
14  *  Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com)
15  *  Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com)
16  *  Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com)
17  *  Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com)
18  *  Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com)
19  */
20 
21 #include "qemu/osdep.h"
22 #include <linux/vfio.h>
23 #include <sys/ioctl.h>
24 
25 #include "hw/pci/msi.h"
26 #include "hw/pci/msix.h"
27 #include "hw/pci/pci_bridge.h"
28 #include "qemu/error-report.h"
29 #include "qemu/range.h"
30 #include "sysemu/kvm.h"
31 #include "sysemu/sysemu.h"
32 #include "pci.h"
33 #include "trace.h"
34 #include "qapi/error.h"
35 
36 #define MSIX_CAP_LENGTH 12
37 
38 static void vfio_disable_interrupts(VFIOPCIDevice *vdev);
39 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled);
40 
41 /*
42  * Disabling BAR mmaping can be slow, but toggling it around INTx can
43  * also be a huge overhead.  We try to get the best of both worlds by
44  * waiting until an interrupt to disable mmaps (subsequent transitions
45  * to the same state are effectively no overhead).  If the interrupt has
46  * been serviced and the time gap is long enough, we re-enable mmaps for
47  * performance.  This works well for things like graphics cards, which
48  * may not use their interrupt at all and are penalized to an unusable
49  * level by read/write BAR traps.  Other devices, like NICs, have more
50  * regular interrupts and see much better latency by staying in non-mmap
51  * mode.  We therefore set the default mmap_timeout such that a ping
52  * is just enough to keep the mmap disabled.  Users can experiment with
53  * other options with the x-intx-mmap-timeout-ms parameter (a value of
54  * zero disables the timer).
55  */
56 static void vfio_intx_mmap_enable(void *opaque)
57 {
58     VFIOPCIDevice *vdev = opaque;
59 
60     if (vdev->intx.pending) {
61         timer_mod(vdev->intx.mmap_timer,
62                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
63         return;
64     }
65 
66     vfio_mmap_set_enabled(vdev, true);
67 }
68 
69 static void vfio_intx_interrupt(void *opaque)
70 {
71     VFIOPCIDevice *vdev = opaque;
72 
73     if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
74         return;
75     }
76 
77     trace_vfio_intx_interrupt(vdev->vbasedev.name, 'A' + vdev->intx.pin);
78 
79     vdev->intx.pending = true;
80     pci_irq_assert(&vdev->pdev);
81     vfio_mmap_set_enabled(vdev, false);
82     if (vdev->intx.mmap_timeout) {
83         timer_mod(vdev->intx.mmap_timer,
84                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
85     }
86 }
87 
88 static void vfio_intx_eoi(VFIODevice *vbasedev)
89 {
90     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
91 
92     if (!vdev->intx.pending) {
93         return;
94     }
95 
96     trace_vfio_intx_eoi(vbasedev->name);
97 
98     vdev->intx.pending = false;
99     pci_irq_deassert(&vdev->pdev);
100     vfio_unmask_single_irqindex(vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
101 }
102 
103 static void vfio_intx_enable_kvm(VFIOPCIDevice *vdev, Error **errp)
104 {
105 #ifdef CONFIG_KVM
106     struct kvm_irqfd irqfd = {
107         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
108         .gsi = vdev->intx.route.irq,
109         .flags = KVM_IRQFD_FLAG_RESAMPLE,
110     };
111     struct vfio_irq_set *irq_set;
112     int ret, argsz;
113     int32_t *pfd;
114 
115     if (vdev->no_kvm_intx || !kvm_irqfds_enabled() ||
116         vdev->intx.route.mode != PCI_INTX_ENABLED ||
117         !kvm_resamplefds_enabled()) {
118         return;
119     }
120 
121     /* Get to a known interrupt state */
122     qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
123     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
124     vdev->intx.pending = false;
125     pci_irq_deassert(&vdev->pdev);
126 
127     /* Get an eventfd for resample/unmask */
128     if (event_notifier_init(&vdev->intx.unmask, 0)) {
129         error_setg(errp, "event_notifier_init failed eoi");
130         goto fail;
131     }
132 
133     /* KVM triggers it, VFIO listens for it */
134     irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
135 
136     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
137         error_setg_errno(errp, errno, "failed to setup resample irqfd");
138         goto fail_irqfd;
139     }
140 
141     argsz = sizeof(*irq_set) + sizeof(*pfd);
142 
143     irq_set = g_malloc0(argsz);
144     irq_set->argsz = argsz;
145     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
146     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
147     irq_set->start = 0;
148     irq_set->count = 1;
149     pfd = (int32_t *)&irq_set->data;
150 
151     *pfd = irqfd.resamplefd;
152 
153     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
154     g_free(irq_set);
155     if (ret) {
156         error_setg_errno(errp, -ret, "failed to setup INTx unmask fd");
157         goto fail_vfio;
158     }
159 
160     /* Let'em rip */
161     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
162 
163     vdev->intx.kvm_accel = true;
164 
165     trace_vfio_intx_enable_kvm(vdev->vbasedev.name);
166 
167     return;
168 
169 fail_vfio:
170     irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
171     kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
172 fail_irqfd:
173     event_notifier_cleanup(&vdev->intx.unmask);
174 fail:
175     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
176     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
177 #endif
178 }
179 
180 static void vfio_intx_disable_kvm(VFIOPCIDevice *vdev)
181 {
182 #ifdef CONFIG_KVM
183     struct kvm_irqfd irqfd = {
184         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
185         .gsi = vdev->intx.route.irq,
186         .flags = KVM_IRQFD_FLAG_DEASSIGN,
187     };
188 
189     if (!vdev->intx.kvm_accel) {
190         return;
191     }
192 
193     /*
194      * Get to a known state, hardware masked, QEMU ready to accept new
195      * interrupts, QEMU IRQ de-asserted.
196      */
197     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
198     vdev->intx.pending = false;
199     pci_irq_deassert(&vdev->pdev);
200 
201     /* Tell KVM to stop listening for an INTx irqfd */
202     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
203         error_report("vfio: Error: Failed to disable INTx irqfd: %m");
204     }
205 
206     /* We only need to close the eventfd for VFIO to cleanup the kernel side */
207     event_notifier_cleanup(&vdev->intx.unmask);
208 
209     /* QEMU starts listening for interrupt events. */
210     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
211 
212     vdev->intx.kvm_accel = false;
213 
214     /* If we've missed an event, let it re-fire through QEMU */
215     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
216 
217     trace_vfio_intx_disable_kvm(vdev->vbasedev.name);
218 #endif
219 }
220 
221 static void vfio_intx_update(PCIDevice *pdev)
222 {
223     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
224     PCIINTxRoute route;
225     Error *err = NULL;
226 
227     if (vdev->interrupt != VFIO_INT_INTx) {
228         return;
229     }
230 
231     route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
232 
233     if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
234         return; /* Nothing changed */
235     }
236 
237     trace_vfio_intx_update(vdev->vbasedev.name,
238                            vdev->intx.route.irq, route.irq);
239 
240     vfio_intx_disable_kvm(vdev);
241 
242     vdev->intx.route = route;
243 
244     if (route.mode != PCI_INTX_ENABLED) {
245         return;
246     }
247 
248     vfio_intx_enable_kvm(vdev, &err);
249     if (err) {
250         error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name);
251     }
252 
253     /* Re-enable the interrupt in cased we missed an EOI */
254     vfio_intx_eoi(&vdev->vbasedev);
255 }
256 
257 static int vfio_intx_enable(VFIOPCIDevice *vdev, Error **errp)
258 {
259     uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
260     int ret, argsz;
261     struct vfio_irq_set *irq_set;
262     int32_t *pfd;
263     Error *err = NULL;
264 
265     if (!pin) {
266         return 0;
267     }
268 
269     vfio_disable_interrupts(vdev);
270 
271     vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
272     pci_config_set_interrupt_pin(vdev->pdev.config, pin);
273 
274 #ifdef CONFIG_KVM
275     /*
276      * Only conditional to avoid generating error messages on platforms
277      * where we won't actually use the result anyway.
278      */
279     if (kvm_irqfds_enabled() && kvm_resamplefds_enabled()) {
280         vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
281                                                         vdev->intx.pin);
282     }
283 #endif
284 
285     ret = event_notifier_init(&vdev->intx.interrupt, 0);
286     if (ret) {
287         error_setg_errno(errp, -ret, "event_notifier_init failed");
288         return ret;
289     }
290 
291     argsz = sizeof(*irq_set) + sizeof(*pfd);
292 
293     irq_set = g_malloc0(argsz);
294     irq_set->argsz = argsz;
295     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
296     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
297     irq_set->start = 0;
298     irq_set->count = 1;
299     pfd = (int32_t *)&irq_set->data;
300 
301     *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
302     qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
303 
304     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
305     g_free(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         return -errno;
311     }
312 
313     vfio_intx_enable_kvm(vdev, &err);
314     if (err) {
315         error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name);
316     }
317 
318     vdev->interrupt = VFIO_INT_INTx;
319 
320     trace_vfio_intx_enable(vdev->vbasedev.name);
321 
322     return 0;
323 }
324 
325 static void vfio_intx_disable(VFIOPCIDevice *vdev)
326 {
327     int fd;
328 
329     timer_del(vdev->intx.mmap_timer);
330     vfio_intx_disable_kvm(vdev);
331     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
332     vdev->intx.pending = false;
333     pci_irq_deassert(&vdev->pdev);
334     vfio_mmap_set_enabled(vdev, true);
335 
336     fd = event_notifier_get_fd(&vdev->intx.interrupt);
337     qemu_set_fd_handler(fd, NULL, NULL, vdev);
338     event_notifier_cleanup(&vdev->intx.interrupt);
339 
340     vdev->interrupt = VFIO_INT_NONE;
341 
342     trace_vfio_intx_disable(vdev->vbasedev.name);
343 }
344 
345 /*
346  * MSI/X
347  */
348 static void vfio_msi_interrupt(void *opaque)
349 {
350     VFIOMSIVector *vector = opaque;
351     VFIOPCIDevice *vdev = vector->vdev;
352     MSIMessage (*get_msg)(PCIDevice *dev, unsigned vector);
353     void (*notify)(PCIDevice *dev, unsigned vector);
354     MSIMessage msg;
355     int nr = vector - vdev->msi_vectors;
356 
357     if (!event_notifier_test_and_clear(&vector->interrupt)) {
358         return;
359     }
360 
361     if (vdev->interrupt == VFIO_INT_MSIX) {
362         get_msg = msix_get_message;
363         notify = msix_notify;
364 
365         /* A masked vector firing needs to use the PBA, enable it */
366         if (msix_is_masked(&vdev->pdev, nr)) {
367             set_bit(nr, vdev->msix->pending);
368             memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, true);
369             trace_vfio_msix_pba_enable(vdev->vbasedev.name);
370         }
371     } else if (vdev->interrupt == VFIO_INT_MSI) {
372         get_msg = msi_get_message;
373         notify = msi_notify;
374     } else {
375         abort();
376     }
377 
378     msg = get_msg(&vdev->pdev, nr);
379     trace_vfio_msi_interrupt(vdev->vbasedev.name, nr, msg.address, msg.data);
380     notify(&vdev->pdev, nr);
381 }
382 
383 static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool msix)
384 {
385     struct vfio_irq_set *irq_set;
386     int ret = 0, i, argsz;
387     int32_t *fds;
388 
389     argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
390 
391     irq_set = g_malloc0(argsz);
392     irq_set->argsz = argsz;
393     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
394     irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
395     irq_set->start = 0;
396     irq_set->count = vdev->nr_vectors;
397     fds = (int32_t *)&irq_set->data;
398 
399     for (i = 0; i < vdev->nr_vectors; i++) {
400         int fd = -1;
401 
402         /*
403          * MSI vs MSI-X - The guest has direct access to MSI mask and pending
404          * bits, therefore we always use the KVM signaling path when setup.
405          * MSI-X mask and pending bits are emulated, so we want to use the
406          * KVM signaling path only when configured and unmasked.
407          */
408         if (vdev->msi_vectors[i].use) {
409             if (vdev->msi_vectors[i].virq < 0 ||
410                 (msix && msix_is_masked(&vdev->pdev, i))) {
411                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
412             } else {
413                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].kvm_interrupt);
414             }
415         }
416 
417         fds[i] = fd;
418     }
419 
420     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
421 
422     g_free(irq_set);
423 
424     return ret;
425 }
426 
427 static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector,
428                                   int vector_n, bool msix)
429 {
430     int virq;
431 
432     if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) {
433         return;
434     }
435 
436     if (event_notifier_init(&vector->kvm_interrupt, 0)) {
437         return;
438     }
439 
440     virq = kvm_irqchip_add_msi_route(kvm_state, vector_n, &vdev->pdev);
441     if (virq < 0) {
442         event_notifier_cleanup(&vector->kvm_interrupt);
443         return;
444     }
445 
446     if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
447                                        NULL, virq) < 0) {
448         kvm_irqchip_release_virq(kvm_state, virq);
449         event_notifier_cleanup(&vector->kvm_interrupt);
450         return;
451     }
452 
453     vector->virq = virq;
454 }
455 
456 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
457 {
458     kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
459                                           vector->virq);
460     kvm_irqchip_release_virq(kvm_state, vector->virq);
461     vector->virq = -1;
462     event_notifier_cleanup(&vector->kvm_interrupt);
463 }
464 
465 static void vfio_update_kvm_msi_virq(VFIOMSIVector *vector, MSIMessage msg,
466                                      PCIDevice *pdev)
467 {
468     kvm_irqchip_update_msi_route(kvm_state, vector->virq, msg, pdev);
469     kvm_irqchip_commit_routes(kvm_state);
470 }
471 
472 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
473                                    MSIMessage *msg, IOHandler *handler)
474 {
475     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
476     VFIOMSIVector *vector;
477     int ret;
478 
479     trace_vfio_msix_vector_do_use(vdev->vbasedev.name, nr);
480 
481     vector = &vdev->msi_vectors[nr];
482 
483     if (!vector->use) {
484         vector->vdev = vdev;
485         vector->virq = -1;
486         if (event_notifier_init(&vector->interrupt, 0)) {
487             error_report("vfio: Error: event_notifier_init failed");
488         }
489         vector->use = true;
490         msix_vector_use(pdev, nr);
491     }
492 
493     qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
494                         handler, NULL, vector);
495 
496     /*
497      * Attempt to enable route through KVM irqchip,
498      * default to userspace handling if unavailable.
499      */
500     if (vector->virq >= 0) {
501         if (!msg) {
502             vfio_remove_kvm_msi_virq(vector);
503         } else {
504             vfio_update_kvm_msi_virq(vector, *msg, pdev);
505         }
506     } else {
507         if (msg) {
508             vfio_add_kvm_msi_virq(vdev, vector, nr, true);
509         }
510     }
511 
512     /*
513      * We don't want to have the host allocate all possible MSI vectors
514      * for a device if they're not in use, so we shutdown and incrementally
515      * increase them as needed.
516      */
517     if (vdev->nr_vectors < nr + 1) {
518         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
519         vdev->nr_vectors = nr + 1;
520         ret = vfio_enable_vectors(vdev, true);
521         if (ret) {
522             error_report("vfio: failed to enable vectors, %d", ret);
523         }
524     } else {
525         int argsz;
526         struct vfio_irq_set *irq_set;
527         int32_t *pfd;
528 
529         argsz = sizeof(*irq_set) + sizeof(*pfd);
530 
531         irq_set = g_malloc0(argsz);
532         irq_set->argsz = argsz;
533         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
534                          VFIO_IRQ_SET_ACTION_TRIGGER;
535         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
536         irq_set->start = nr;
537         irq_set->count = 1;
538         pfd = (int32_t *)&irq_set->data;
539 
540         if (vector->virq >= 0) {
541             *pfd = event_notifier_get_fd(&vector->kvm_interrupt);
542         } else {
543             *pfd = event_notifier_get_fd(&vector->interrupt);
544         }
545 
546         ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
547         g_free(irq_set);
548         if (ret) {
549             error_report("vfio: failed to modify vector, %d", ret);
550         }
551     }
552 
553     /* Disable PBA emulation when nothing more is pending. */
554     clear_bit(nr, vdev->msix->pending);
555     if (find_first_bit(vdev->msix->pending,
556                        vdev->nr_vectors) == vdev->nr_vectors) {
557         memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
558         trace_vfio_msix_pba_disable(vdev->vbasedev.name);
559     }
560 
561     return 0;
562 }
563 
564 static int vfio_msix_vector_use(PCIDevice *pdev,
565                                 unsigned int nr, MSIMessage msg)
566 {
567     return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
568 }
569 
570 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
571 {
572     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
573     VFIOMSIVector *vector = &vdev->msi_vectors[nr];
574 
575     trace_vfio_msix_vector_release(vdev->vbasedev.name, nr);
576 
577     /*
578      * There are still old guests that mask and unmask vectors on every
579      * interrupt.  If we're using QEMU bypass with a KVM irqfd, leave all of
580      * the KVM setup in place, simply switch VFIO to use the non-bypass
581      * eventfd.  We'll then fire the interrupt through QEMU and the MSI-X
582      * core will mask the interrupt and set pending bits, allowing it to
583      * be re-asserted on unmask.  Nothing to do if already using QEMU mode.
584      */
585     if (vector->virq >= 0) {
586         int argsz;
587         struct vfio_irq_set *irq_set;
588         int32_t *pfd;
589 
590         argsz = sizeof(*irq_set) + sizeof(*pfd);
591 
592         irq_set = g_malloc0(argsz);
593         irq_set->argsz = argsz;
594         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
595                          VFIO_IRQ_SET_ACTION_TRIGGER;
596         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
597         irq_set->start = nr;
598         irq_set->count = 1;
599         pfd = (int32_t *)&irq_set->data;
600 
601         *pfd = event_notifier_get_fd(&vector->interrupt);
602 
603         ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
604 
605         g_free(irq_set);
606     }
607 }
608 
609 static void vfio_msix_enable(VFIOPCIDevice *vdev)
610 {
611     vfio_disable_interrupts(vdev);
612 
613     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->msix->entries);
614 
615     vdev->interrupt = VFIO_INT_MSIX;
616 
617     /*
618      * Some communication channels between VF & PF or PF & fw rely on the
619      * physical state of the device and expect that enabling MSI-X from the
620      * guest enables the same on the host.  When our guest is Linux, the
621      * guest driver call to pci_enable_msix() sets the enabling bit in the
622      * MSI-X capability, but leaves the vector table masked.  We therefore
623      * can't rely on a vector_use callback (from request_irq() in the guest)
624      * to switch the physical device into MSI-X mode because that may come a
625      * long time after pci_enable_msix().  This code enables vector 0 with
626      * triggering to userspace, then immediately release the vector, leaving
627      * the physical device with no vectors enabled, but MSI-X enabled, just
628      * like the guest view.
629      */
630     vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
631     vfio_msix_vector_release(&vdev->pdev, 0);
632 
633     if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
634                                   vfio_msix_vector_release, NULL)) {
635         error_report("vfio: msix_set_vector_notifiers failed");
636     }
637 
638     trace_vfio_msix_enable(vdev->vbasedev.name);
639 }
640 
641 static void vfio_msi_enable(VFIOPCIDevice *vdev)
642 {
643     int ret, i;
644 
645     vfio_disable_interrupts(vdev);
646 
647     vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
648 retry:
649     vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->nr_vectors);
650 
651     for (i = 0; i < vdev->nr_vectors; i++) {
652         VFIOMSIVector *vector = &vdev->msi_vectors[i];
653 
654         vector->vdev = vdev;
655         vector->virq = -1;
656         vector->use = true;
657 
658         if (event_notifier_init(&vector->interrupt, 0)) {
659             error_report("vfio: Error: event_notifier_init failed");
660         }
661 
662         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
663                             vfio_msi_interrupt, NULL, vector);
664 
665         /*
666          * Attempt to enable route through KVM irqchip,
667          * default to userspace handling if unavailable.
668          */
669         vfio_add_kvm_msi_virq(vdev, vector, i, false);
670     }
671 
672     /* Set interrupt type prior to possible interrupts */
673     vdev->interrupt = VFIO_INT_MSI;
674 
675     ret = vfio_enable_vectors(vdev, false);
676     if (ret) {
677         if (ret < 0) {
678             error_report("vfio: Error: Failed to setup MSI fds: %m");
679         } else if (ret != vdev->nr_vectors) {
680             error_report("vfio: Error: Failed to enable %d "
681                          "MSI vectors, retry with %d", vdev->nr_vectors, ret);
682         }
683 
684         for (i = 0; i < vdev->nr_vectors; i++) {
685             VFIOMSIVector *vector = &vdev->msi_vectors[i];
686             if (vector->virq >= 0) {
687                 vfio_remove_kvm_msi_virq(vector);
688             }
689             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
690                                 NULL, NULL, NULL);
691             event_notifier_cleanup(&vector->interrupt);
692         }
693 
694         g_free(vdev->msi_vectors);
695 
696         if (ret > 0 && ret != vdev->nr_vectors) {
697             vdev->nr_vectors = ret;
698             goto retry;
699         }
700         vdev->nr_vectors = 0;
701 
702         /*
703          * Failing to setup MSI doesn't really fall within any specification.
704          * Let's try leaving interrupts disabled and hope the guest figures
705          * out to fall back to INTx for this device.
706          */
707         error_report("vfio: Error: Failed to enable MSI");
708         vdev->interrupt = VFIO_INT_NONE;
709 
710         return;
711     }
712 
713     trace_vfio_msi_enable(vdev->vbasedev.name, vdev->nr_vectors);
714 }
715 
716 static void vfio_msi_disable_common(VFIOPCIDevice *vdev)
717 {
718     Error *err = NULL;
719     int i;
720 
721     for (i = 0; i < vdev->nr_vectors; i++) {
722         VFIOMSIVector *vector = &vdev->msi_vectors[i];
723         if (vdev->msi_vectors[i].use) {
724             if (vector->virq >= 0) {
725                 vfio_remove_kvm_msi_virq(vector);
726             }
727             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
728                                 NULL, NULL, NULL);
729             event_notifier_cleanup(&vector->interrupt);
730         }
731     }
732 
733     g_free(vdev->msi_vectors);
734     vdev->msi_vectors = NULL;
735     vdev->nr_vectors = 0;
736     vdev->interrupt = VFIO_INT_NONE;
737 
738     vfio_intx_enable(vdev, &err);
739     if (err) {
740         error_reportf_err(err, ERR_PREFIX, vdev->vbasedev.name);
741     }
742 }
743 
744 static void vfio_msix_disable(VFIOPCIDevice *vdev)
745 {
746     int i;
747 
748     msix_unset_vector_notifiers(&vdev->pdev);
749 
750     /*
751      * MSI-X will only release vectors if MSI-X is still enabled on the
752      * device, check through the rest and release it ourselves if necessary.
753      */
754     for (i = 0; i < vdev->nr_vectors; i++) {
755         if (vdev->msi_vectors[i].use) {
756             vfio_msix_vector_release(&vdev->pdev, i);
757             msix_vector_unuse(&vdev->pdev, i);
758         }
759     }
760 
761     if (vdev->nr_vectors) {
762         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
763     }
764 
765     vfio_msi_disable_common(vdev);
766 
767     memset(vdev->msix->pending, 0,
768            BITS_TO_LONGS(vdev->msix->entries) * sizeof(unsigned long));
769 
770     trace_vfio_msix_disable(vdev->vbasedev.name);
771 }
772 
773 static void vfio_msi_disable(VFIOPCIDevice *vdev)
774 {
775     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSI_IRQ_INDEX);
776     vfio_msi_disable_common(vdev);
777 
778     trace_vfio_msi_disable(vdev->vbasedev.name);
779 }
780 
781 static void vfio_update_msi(VFIOPCIDevice *vdev)
782 {
783     int i;
784 
785     for (i = 0; i < vdev->nr_vectors; i++) {
786         VFIOMSIVector *vector = &vdev->msi_vectors[i];
787         MSIMessage msg;
788 
789         if (!vector->use || vector->virq < 0) {
790             continue;
791         }
792 
793         msg = msi_get_message(&vdev->pdev, i);
794         vfio_update_kvm_msi_virq(vector, msg, &vdev->pdev);
795     }
796 }
797 
798 static void vfio_pci_load_rom(VFIOPCIDevice *vdev)
799 {
800     struct vfio_region_info *reg_info;
801     uint64_t size;
802     off_t off = 0;
803     ssize_t bytes;
804 
805     if (vfio_get_region_info(&vdev->vbasedev,
806                              VFIO_PCI_ROM_REGION_INDEX, &reg_info)) {
807         error_report("vfio: Error getting ROM info: %m");
808         return;
809     }
810 
811     trace_vfio_pci_load_rom(vdev->vbasedev.name, (unsigned long)reg_info->size,
812                             (unsigned long)reg_info->offset,
813                             (unsigned long)reg_info->flags);
814 
815     vdev->rom_size = size = reg_info->size;
816     vdev->rom_offset = reg_info->offset;
817 
818     g_free(reg_info);
819 
820     if (!vdev->rom_size) {
821         vdev->rom_read_failed = true;
822         error_report("vfio-pci: Cannot read device rom at "
823                     "%s", vdev->vbasedev.name);
824         error_printf("Device option ROM contents are probably invalid "
825                     "(check dmesg).\nSkip option ROM probe with rombar=0, "
826                     "or load from file with romfile=\n");
827         return;
828     }
829 
830     vdev->rom = g_malloc(size);
831     memset(vdev->rom, 0xff, size);
832 
833     while (size) {
834         bytes = pread(vdev->vbasedev.fd, vdev->rom + off,
835                       size, vdev->rom_offset + off);
836         if (bytes == 0) {
837             break;
838         } else if (bytes > 0) {
839             off += bytes;
840             size -= bytes;
841         } else {
842             if (errno == EINTR || errno == EAGAIN) {
843                 continue;
844             }
845             error_report("vfio: Error reading device ROM: %m");
846             break;
847         }
848     }
849 
850     /*
851      * Test the ROM signature against our device, if the vendor is correct
852      * but the device ID doesn't match, store the correct device ID and
853      * recompute the checksum.  Intel IGD devices need this and are known
854      * to have bogus checksums so we can't simply adjust the checksum.
855      */
856     if (pci_get_word(vdev->rom) == 0xaa55 &&
857         pci_get_word(vdev->rom + 0x18) + 8 < vdev->rom_size &&
858         !memcmp(vdev->rom + pci_get_word(vdev->rom + 0x18), "PCIR", 4)) {
859         uint16_t vid, did;
860 
861         vid = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 4);
862         did = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6);
863 
864         if (vid == vdev->vendor_id && did != vdev->device_id) {
865             int i;
866             uint8_t csum, *data = vdev->rom;
867 
868             pci_set_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6,
869                          vdev->device_id);
870             data[6] = 0;
871 
872             for (csum = 0, i = 0; i < vdev->rom_size; i++) {
873                 csum += data[i];
874             }
875 
876             data[6] = -csum;
877         }
878     }
879 }
880 
881 static uint64_t vfio_rom_read(void *opaque, hwaddr addr, unsigned size)
882 {
883     VFIOPCIDevice *vdev = opaque;
884     union {
885         uint8_t byte;
886         uint16_t word;
887         uint32_t dword;
888         uint64_t qword;
889     } val;
890     uint64_t data = 0;
891 
892     /* Load the ROM lazily when the guest tries to read it */
893     if (unlikely(!vdev->rom && !vdev->rom_read_failed)) {
894         vfio_pci_load_rom(vdev);
895     }
896 
897     memcpy(&val, vdev->rom + addr,
898            (addr < vdev->rom_size) ? MIN(size, vdev->rom_size - addr) : 0);
899 
900     switch (size) {
901     case 1:
902         data = val.byte;
903         break;
904     case 2:
905         data = le16_to_cpu(val.word);
906         break;
907     case 4:
908         data = le32_to_cpu(val.dword);
909         break;
910     default:
911         hw_error("vfio: unsupported read size, %d bytes\n", size);
912         break;
913     }
914 
915     trace_vfio_rom_read(vdev->vbasedev.name, addr, size, data);
916 
917     return data;
918 }
919 
920 static void vfio_rom_write(void *opaque, hwaddr addr,
921                            uint64_t data, unsigned size)
922 {
923 }
924 
925 static const MemoryRegionOps vfio_rom_ops = {
926     .read = vfio_rom_read,
927     .write = vfio_rom_write,
928     .endianness = DEVICE_LITTLE_ENDIAN,
929 };
930 
931 static void vfio_pci_size_rom(VFIOPCIDevice *vdev)
932 {
933     uint32_t orig, size = cpu_to_le32((uint32_t)PCI_ROM_ADDRESS_MASK);
934     off_t offset = vdev->config_offset + PCI_ROM_ADDRESS;
935     DeviceState *dev = DEVICE(vdev);
936     char *name;
937     int fd = vdev->vbasedev.fd;
938 
939     if (vdev->pdev.romfile || !vdev->pdev.rom_bar) {
940         /* Since pci handles romfile, just print a message and return */
941         if (vfio_blacklist_opt_rom(vdev) && vdev->pdev.romfile) {
942             error_printf("Warning : Device at %s is known to cause system instability issues during option rom execution. Proceeding anyway since user specified romfile\n",
943                          vdev->vbasedev.name);
944         }
945         return;
946     }
947 
948     /*
949      * Use the same size ROM BAR as the physical device.  The contents
950      * will get filled in later when the guest tries to read it.
951      */
952     if (pread(fd, &orig, 4, offset) != 4 ||
953         pwrite(fd, &size, 4, offset) != 4 ||
954         pread(fd, &size, 4, offset) != 4 ||
955         pwrite(fd, &orig, 4, offset) != 4) {
956         error_report("%s(%s) failed: %m", __func__, vdev->vbasedev.name);
957         return;
958     }
959 
960     size = ~(le32_to_cpu(size) & PCI_ROM_ADDRESS_MASK) + 1;
961 
962     if (!size) {
963         return;
964     }
965 
966     if (vfio_blacklist_opt_rom(vdev)) {
967         if (dev->opts && qemu_opt_get(dev->opts, "rombar")) {
968             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",
969                          vdev->vbasedev.name);
970         } else {
971             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",
972                          vdev->vbasedev.name);
973             return;
974         }
975     }
976 
977     trace_vfio_pci_size_rom(vdev->vbasedev.name, size);
978 
979     name = g_strdup_printf("vfio[%s].rom", vdev->vbasedev.name);
980 
981     memory_region_init_io(&vdev->pdev.rom, OBJECT(vdev),
982                           &vfio_rom_ops, vdev, name, size);
983     g_free(name);
984 
985     pci_register_bar(&vdev->pdev, PCI_ROM_SLOT,
986                      PCI_BASE_ADDRESS_SPACE_MEMORY, &vdev->pdev.rom);
987 
988     vdev->pdev.has_rom = true;
989     vdev->rom_read_failed = false;
990 }
991 
992 void vfio_vga_write(void *opaque, hwaddr addr,
993                            uint64_t data, unsigned size)
994 {
995     VFIOVGARegion *region = opaque;
996     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
997     union {
998         uint8_t byte;
999         uint16_t word;
1000         uint32_t dword;
1001         uint64_t qword;
1002     } buf;
1003     off_t offset = vga->fd_offset + region->offset + addr;
1004 
1005     switch (size) {
1006     case 1:
1007         buf.byte = data;
1008         break;
1009     case 2:
1010         buf.word = cpu_to_le16(data);
1011         break;
1012     case 4:
1013         buf.dword = cpu_to_le32(data);
1014         break;
1015     default:
1016         hw_error("vfio: unsupported write size, %d bytes", size);
1017         break;
1018     }
1019 
1020     if (pwrite(vga->fd, &buf, size, offset) != size) {
1021         error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
1022                      __func__, region->offset + addr, data, size);
1023     }
1024 
1025     trace_vfio_vga_write(region->offset + addr, data, size);
1026 }
1027 
1028 uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size)
1029 {
1030     VFIOVGARegion *region = opaque;
1031     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1032     union {
1033         uint8_t byte;
1034         uint16_t word;
1035         uint32_t dword;
1036         uint64_t qword;
1037     } buf;
1038     uint64_t data = 0;
1039     off_t offset = vga->fd_offset + region->offset + addr;
1040 
1041     if (pread(vga->fd, &buf, size, offset) != size) {
1042         error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
1043                      __func__, region->offset + addr, size);
1044         return (uint64_t)-1;
1045     }
1046 
1047     switch (size) {
1048     case 1:
1049         data = buf.byte;
1050         break;
1051     case 2:
1052         data = le16_to_cpu(buf.word);
1053         break;
1054     case 4:
1055         data = le32_to_cpu(buf.dword);
1056         break;
1057     default:
1058         hw_error("vfio: unsupported read size, %d bytes", size);
1059         break;
1060     }
1061 
1062     trace_vfio_vga_read(region->offset + addr, size, data);
1063 
1064     return data;
1065 }
1066 
1067 static const MemoryRegionOps vfio_vga_ops = {
1068     .read = vfio_vga_read,
1069     .write = vfio_vga_write,
1070     .endianness = DEVICE_LITTLE_ENDIAN,
1071 };
1072 
1073 /*
1074  * PCI config space
1075  */
1076 uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
1077 {
1078     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
1079     uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val;
1080 
1081     memcpy(&emu_bits, vdev->emulated_config_bits + addr, len);
1082     emu_bits = le32_to_cpu(emu_bits);
1083 
1084     if (emu_bits) {
1085         emu_val = pci_default_read_config(pdev, addr, len);
1086     }
1087 
1088     if (~emu_bits & (0xffffffffU >> (32 - len * 8))) {
1089         ssize_t ret;
1090 
1091         ret = pread(vdev->vbasedev.fd, &phys_val, len,
1092                     vdev->config_offset + addr);
1093         if (ret != len) {
1094             error_report("%s(%s, 0x%x, 0x%x) failed: %m",
1095                          __func__, vdev->vbasedev.name, addr, len);
1096             return -errno;
1097         }
1098         phys_val = le32_to_cpu(phys_val);
1099     }
1100 
1101     val = (emu_val & emu_bits) | (phys_val & ~emu_bits);
1102 
1103     trace_vfio_pci_read_config(vdev->vbasedev.name, addr, len, val);
1104 
1105     return val;
1106 }
1107 
1108 void vfio_pci_write_config(PCIDevice *pdev,
1109                            uint32_t addr, uint32_t val, int len)
1110 {
1111     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
1112     uint32_t val_le = cpu_to_le32(val);
1113 
1114     trace_vfio_pci_write_config(vdev->vbasedev.name, addr, val, len);
1115 
1116     /* Write everything to VFIO, let it filter out what we can't write */
1117     if (pwrite(vdev->vbasedev.fd, &val_le, len, vdev->config_offset + addr)
1118                 != len) {
1119         error_report("%s(%s, 0x%x, 0x%x, 0x%x) failed: %m",
1120                      __func__, vdev->vbasedev.name, addr, val, len);
1121     }
1122 
1123     /* MSI/MSI-X Enabling/Disabling */
1124     if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
1125         ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
1126         int is_enabled, was_enabled = msi_enabled(pdev);
1127 
1128         pci_default_write_config(pdev, addr, val, len);
1129 
1130         is_enabled = msi_enabled(pdev);
1131 
1132         if (!was_enabled) {
1133             if (is_enabled) {
1134                 vfio_msi_enable(vdev);
1135             }
1136         } else {
1137             if (!is_enabled) {
1138                 vfio_msi_disable(vdev);
1139             } else {
1140                 vfio_update_msi(vdev);
1141             }
1142         }
1143     } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
1144         ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
1145         int is_enabled, was_enabled = msix_enabled(pdev);
1146 
1147         pci_default_write_config(pdev, addr, val, len);
1148 
1149         is_enabled = msix_enabled(pdev);
1150 
1151         if (!was_enabled && is_enabled) {
1152             vfio_msix_enable(vdev);
1153         } else if (was_enabled && !is_enabled) {
1154             vfio_msix_disable(vdev);
1155         }
1156     } else {
1157         /* Write everything to QEMU to keep emulated bits correct */
1158         pci_default_write_config(pdev, addr, val, len);
1159     }
1160 }
1161 
1162 /*
1163  * Interrupt setup
1164  */
1165 static void vfio_disable_interrupts(VFIOPCIDevice *vdev)
1166 {
1167     /*
1168      * More complicated than it looks.  Disabling MSI/X transitions the
1169      * device to INTx mode (if supported).  Therefore we need to first
1170      * disable MSI/X and then cleanup by disabling INTx.
1171      */
1172     if (vdev->interrupt == VFIO_INT_MSIX) {
1173         vfio_msix_disable(vdev);
1174     } else if (vdev->interrupt == VFIO_INT_MSI) {
1175         vfio_msi_disable(vdev);
1176     }
1177 
1178     if (vdev->interrupt == VFIO_INT_INTx) {
1179         vfio_intx_disable(vdev);
1180     }
1181 }
1182 
1183 static int vfio_msi_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1184 {
1185     uint16_t ctrl;
1186     bool msi_64bit, msi_maskbit;
1187     int ret, entries;
1188     Error *err = NULL;
1189 
1190     if (pread(vdev->vbasedev.fd, &ctrl, sizeof(ctrl),
1191               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1192         error_setg_errno(errp, errno, "failed reading MSI PCI_CAP_FLAGS");
1193         return -errno;
1194     }
1195     ctrl = le16_to_cpu(ctrl);
1196 
1197     msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
1198     msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
1199     entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
1200 
1201     trace_vfio_msi_setup(vdev->vbasedev.name, pos);
1202 
1203     ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit, &err);
1204     if (ret < 0) {
1205         if (ret == -ENOTSUP) {
1206             return 0;
1207         }
1208         error_prepend(&err, "msi_init failed: ");
1209         error_propagate(errp, err);
1210         return ret;
1211     }
1212     vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
1213 
1214     return 0;
1215 }
1216 
1217 static void vfio_pci_fixup_msix_region(VFIOPCIDevice *vdev)
1218 {
1219     off_t start, end;
1220     VFIORegion *region = &vdev->bars[vdev->msix->table_bar].region;
1221 
1222     /*
1223      * We expect to find a single mmap covering the whole BAR, anything else
1224      * means it's either unsupported or already setup.
1225      */
1226     if (region->nr_mmaps != 1 || region->mmaps[0].offset ||
1227         region->size != region->mmaps[0].size) {
1228         return;
1229     }
1230 
1231     /* MSI-X table start and end aligned to host page size */
1232     start = vdev->msix->table_offset & qemu_real_host_page_mask;
1233     end = REAL_HOST_PAGE_ALIGN((uint64_t)vdev->msix->table_offset +
1234                                (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
1235 
1236     /*
1237      * Does the MSI-X table cover the beginning of the BAR?  The whole BAR?
1238      * NB - Host page size is necessarily a power of two and so is the PCI
1239      * BAR (not counting EA yet), therefore if we have host page aligned
1240      * @start and @end, then any remainder of the BAR before or after those
1241      * must be at least host page sized and therefore mmap'able.
1242      */
1243     if (!start) {
1244         if (end >= region->size) {
1245             region->nr_mmaps = 0;
1246             g_free(region->mmaps);
1247             region->mmaps = NULL;
1248             trace_vfio_msix_fixup(vdev->vbasedev.name,
1249                                   vdev->msix->table_bar, 0, 0);
1250         } else {
1251             region->mmaps[0].offset = end;
1252             region->mmaps[0].size = region->size - end;
1253             trace_vfio_msix_fixup(vdev->vbasedev.name,
1254                               vdev->msix->table_bar, region->mmaps[0].offset,
1255                               region->mmaps[0].offset + region->mmaps[0].size);
1256         }
1257 
1258     /* Maybe it's aligned at the end of the BAR */
1259     } else if (end >= region->size) {
1260         region->mmaps[0].size = start;
1261         trace_vfio_msix_fixup(vdev->vbasedev.name,
1262                               vdev->msix->table_bar, region->mmaps[0].offset,
1263                               region->mmaps[0].offset + region->mmaps[0].size);
1264 
1265     /* Otherwise it must split the BAR */
1266     } else {
1267         region->nr_mmaps = 2;
1268         region->mmaps = g_renew(VFIOMmap, region->mmaps, 2);
1269 
1270         memcpy(&region->mmaps[1], &region->mmaps[0], sizeof(VFIOMmap));
1271 
1272         region->mmaps[0].size = start;
1273         trace_vfio_msix_fixup(vdev->vbasedev.name,
1274                               vdev->msix->table_bar, region->mmaps[0].offset,
1275                               region->mmaps[0].offset + region->mmaps[0].size);
1276 
1277         region->mmaps[1].offset = end;
1278         region->mmaps[1].size = region->size - end;
1279         trace_vfio_msix_fixup(vdev->vbasedev.name,
1280                               vdev->msix->table_bar, region->mmaps[1].offset,
1281                               region->mmaps[1].offset + region->mmaps[1].size);
1282     }
1283 }
1284 
1285 /*
1286  * We don't have any control over how pci_add_capability() inserts
1287  * capabilities into the chain.  In order to setup MSI-X we need a
1288  * MemoryRegion for the BAR.  In order to setup the BAR and not
1289  * attempt to mmap the MSI-X table area, which VFIO won't allow, we
1290  * need to first look for where the MSI-X table lives.  So we
1291  * unfortunately split MSI-X setup across two functions.
1292  */
1293 static void vfio_msix_early_setup(VFIOPCIDevice *vdev, Error **errp)
1294 {
1295     uint8_t pos;
1296     uint16_t ctrl;
1297     uint32_t table, pba;
1298     int fd = vdev->vbasedev.fd;
1299     VFIOMSIXInfo *msix;
1300 
1301     pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
1302     if (!pos) {
1303         return;
1304     }
1305 
1306     if (pread(fd, &ctrl, sizeof(ctrl),
1307               vdev->config_offset + pos + PCI_MSIX_FLAGS) != sizeof(ctrl)) {
1308         error_setg_errno(errp, errno, "failed to read PCI MSIX FLAGS");
1309         return;
1310     }
1311 
1312     if (pread(fd, &table, sizeof(table),
1313               vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
1314         error_setg_errno(errp, errno, "failed to read PCI MSIX TABLE");
1315         return;
1316     }
1317 
1318     if (pread(fd, &pba, sizeof(pba),
1319               vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
1320         error_setg_errno(errp, errno, "failed to read PCI MSIX PBA");
1321         return;
1322     }
1323 
1324     ctrl = le16_to_cpu(ctrl);
1325     table = le32_to_cpu(table);
1326     pba = le32_to_cpu(pba);
1327 
1328     msix = g_malloc0(sizeof(*msix));
1329     msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
1330     msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
1331     msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
1332     msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
1333     msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
1334 
1335     /*
1336      * Test the size of the pba_offset variable and catch if it extends outside
1337      * of the specified BAR. If it is the case, we need to apply a hardware
1338      * specific quirk if the device is known or we have a broken configuration.
1339      */
1340     if (msix->pba_offset >= vdev->bars[msix->pba_bar].region.size) {
1341         /*
1342          * Chelsio T5 Virtual Function devices are encoded as 0x58xx for T5
1343          * adapters. The T5 hardware returns an incorrect value of 0x8000 for
1344          * the VF PBA offset while the BAR itself is only 8k. The correct value
1345          * is 0x1000, so we hard code that here.
1346          */
1347         if (vdev->vendor_id == PCI_VENDOR_ID_CHELSIO &&
1348             (vdev->device_id & 0xff00) == 0x5800) {
1349             msix->pba_offset = 0x1000;
1350         } else {
1351             error_setg(errp, "hardware reports invalid configuration, "
1352                        "MSIX PBA outside of specified BAR");
1353             g_free(msix);
1354             return;
1355         }
1356     }
1357 
1358     trace_vfio_msix_early_setup(vdev->vbasedev.name, pos, msix->table_bar,
1359                                 msix->table_offset, msix->entries);
1360     vdev->msix = msix;
1361 
1362     vfio_pci_fixup_msix_region(vdev);
1363 }
1364 
1365 static int vfio_msix_setup(VFIOPCIDevice *vdev, int pos, Error **errp)
1366 {
1367     int ret;
1368 
1369     vdev->msix->pending = g_malloc0(BITS_TO_LONGS(vdev->msix->entries) *
1370                                     sizeof(unsigned long));
1371     ret = msix_init(&vdev->pdev, vdev->msix->entries,
1372                     vdev->bars[vdev->msix->table_bar].region.mem,
1373                     vdev->msix->table_bar, vdev->msix->table_offset,
1374                     vdev->bars[vdev->msix->pba_bar].region.mem,
1375                     vdev->msix->pba_bar, vdev->msix->pba_offset, pos);
1376     if (ret < 0) {
1377         if (ret == -ENOTSUP) {
1378             return 0;
1379         }
1380         error_setg(errp, "msix_init failed");
1381         return ret;
1382     }
1383 
1384     /*
1385      * The PCI spec suggests that devices provide additional alignment for
1386      * MSI-X structures and avoid overlapping non-MSI-X related registers.
1387      * For an assigned device, this hopefully means that emulation of MSI-X
1388      * structures does not affect the performance of the device.  If devices
1389      * fail to provide that alignment, a significant performance penalty may
1390      * result, for instance Mellanox MT27500 VFs:
1391      * http://www.spinics.net/lists/kvm/msg125881.html
1392      *
1393      * The PBA is simply not that important for such a serious regression and
1394      * most drivers do not appear to look at it.  The solution for this is to
1395      * disable the PBA MemoryRegion unless it's being used.  We disable it
1396      * here and only enable it if a masked vector fires through QEMU.  As the
1397      * vector-use notifier is called, which occurs on unmask, we test whether
1398      * PBA emulation is needed and again disable if not.
1399      */
1400     memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false);
1401 
1402     return 0;
1403 }
1404 
1405 static void vfio_teardown_msi(VFIOPCIDevice *vdev)
1406 {
1407     msi_uninit(&vdev->pdev);
1408 
1409     if (vdev->msix) {
1410         msix_uninit(&vdev->pdev,
1411                     vdev->bars[vdev->msix->table_bar].region.mem,
1412                     vdev->bars[vdev->msix->pba_bar].region.mem);
1413         g_free(vdev->msix->pending);
1414     }
1415 }
1416 
1417 /*
1418  * Resource setup
1419  */
1420 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled)
1421 {
1422     int i;
1423 
1424     for (i = 0; i < PCI_ROM_SLOT; i++) {
1425         vfio_region_mmaps_set_enabled(&vdev->bars[i].region, enabled);
1426     }
1427 }
1428 
1429 static void vfio_bar_setup(VFIOPCIDevice *vdev, int nr)
1430 {
1431     VFIOBAR *bar = &vdev->bars[nr];
1432 
1433     uint32_t pci_bar;
1434     uint8_t type;
1435     int ret;
1436 
1437     /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1438     if (!bar->region.size) {
1439         return;
1440     }
1441 
1442     /* Determine what type of BAR this is for registration */
1443     ret = pread(vdev->vbasedev.fd, &pci_bar, sizeof(pci_bar),
1444                 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
1445     if (ret != sizeof(pci_bar)) {
1446         error_report("vfio: Failed to read BAR %d (%m)", nr);
1447         return;
1448     }
1449 
1450     pci_bar = le32_to_cpu(pci_bar);
1451     bar->ioport = (pci_bar & PCI_BASE_ADDRESS_SPACE_IO);
1452     bar->mem64 = bar->ioport ? 0 : (pci_bar & PCI_BASE_ADDRESS_MEM_TYPE_64);
1453     type = pci_bar & (bar->ioport ? ~PCI_BASE_ADDRESS_IO_MASK :
1454                                     ~PCI_BASE_ADDRESS_MEM_MASK);
1455 
1456     if (vfio_region_mmap(&bar->region)) {
1457         error_report("Failed to mmap %s BAR %d. Performance may be slow",
1458                      vdev->vbasedev.name, nr);
1459     }
1460 
1461     pci_register_bar(&vdev->pdev, nr, type, bar->region.mem);
1462 }
1463 
1464 static void vfio_bars_setup(VFIOPCIDevice *vdev)
1465 {
1466     int i;
1467 
1468     for (i = 0; i < PCI_ROM_SLOT; i++) {
1469         vfio_bar_setup(vdev, i);
1470     }
1471 }
1472 
1473 static void vfio_bars_exit(VFIOPCIDevice *vdev)
1474 {
1475     int i;
1476 
1477     for (i = 0; i < PCI_ROM_SLOT; i++) {
1478         vfio_bar_quirk_exit(vdev, i);
1479         vfio_region_exit(&vdev->bars[i].region);
1480     }
1481 
1482     if (vdev->vga) {
1483         pci_unregister_vga(&vdev->pdev);
1484         vfio_vga_quirk_exit(vdev);
1485     }
1486 }
1487 
1488 static void vfio_bars_finalize(VFIOPCIDevice *vdev)
1489 {
1490     int i;
1491 
1492     for (i = 0; i < PCI_ROM_SLOT; i++) {
1493         vfio_bar_quirk_finalize(vdev, i);
1494         vfio_region_finalize(&vdev->bars[i].region);
1495     }
1496 
1497     if (vdev->vga) {
1498         vfio_vga_quirk_finalize(vdev);
1499         for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) {
1500             object_unparent(OBJECT(&vdev->vga->region[i].mem));
1501         }
1502         g_free(vdev->vga);
1503     }
1504 }
1505 
1506 /*
1507  * General setup
1508  */
1509 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
1510 {
1511     uint8_t tmp;
1512     uint16_t next = PCI_CONFIG_SPACE_SIZE;
1513 
1514     for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
1515          tmp = pdev->config[tmp + PCI_CAP_LIST_NEXT]) {
1516         if (tmp > pos && tmp < next) {
1517             next = tmp;
1518         }
1519     }
1520 
1521     return next - pos;
1522 }
1523 
1524 
1525 static uint16_t vfio_ext_cap_max_size(const uint8_t *config, uint16_t pos)
1526 {
1527     uint16_t tmp, next = PCIE_CONFIG_SPACE_SIZE;
1528 
1529     for (tmp = PCI_CONFIG_SPACE_SIZE; tmp;
1530         tmp = PCI_EXT_CAP_NEXT(pci_get_long(config + tmp))) {
1531         if (tmp > pos && tmp < next) {
1532             next = tmp;
1533         }
1534     }
1535 
1536     return next - pos;
1537 }
1538 
1539 static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask)
1540 {
1541     pci_set_word(buf, (pci_get_word(buf) & ~mask) | val);
1542 }
1543 
1544 static void vfio_add_emulated_word(VFIOPCIDevice *vdev, int pos,
1545                                    uint16_t val, uint16_t mask)
1546 {
1547     vfio_set_word_bits(vdev->pdev.config + pos, val, mask);
1548     vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask);
1549     vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask);
1550 }
1551 
1552 static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask)
1553 {
1554     pci_set_long(buf, (pci_get_long(buf) & ~mask) | val);
1555 }
1556 
1557 static void vfio_add_emulated_long(VFIOPCIDevice *vdev, int pos,
1558                                    uint32_t val, uint32_t mask)
1559 {
1560     vfio_set_long_bits(vdev->pdev.config + pos, val, mask);
1561     vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask);
1562     vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask);
1563 }
1564 
1565 static int vfio_setup_pcie_cap(VFIOPCIDevice *vdev, int pos, uint8_t size,
1566                                Error **errp)
1567 {
1568     uint16_t flags;
1569     uint8_t type;
1570 
1571     flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS);
1572     type = (flags & PCI_EXP_FLAGS_TYPE) >> 4;
1573 
1574     if (type != PCI_EXP_TYPE_ENDPOINT &&
1575         type != PCI_EXP_TYPE_LEG_END &&
1576         type != PCI_EXP_TYPE_RC_END) {
1577 
1578         error_setg(errp, "assignment of PCIe type 0x%x "
1579                    "devices is not currently supported", type);
1580         return -EINVAL;
1581     }
1582 
1583     if (!pci_bus_is_express(vdev->pdev.bus)) {
1584         PCIBus *bus = vdev->pdev.bus;
1585         PCIDevice *bridge;
1586 
1587         /*
1588          * Traditionally PCI device assignment exposes the PCIe capability
1589          * as-is on non-express buses.  The reason being that some drivers
1590          * simply assume that it's there, for example tg3.  However when
1591          * we're running on a native PCIe machine type, like Q35, we need
1592          * to hide the PCIe capability.  The reason for this is twofold;
1593          * first Windows guests get a Code 10 error when the PCIe capability
1594          * is exposed in this configuration.  Therefore express devices won't
1595          * work at all unless they're attached to express buses in the VM.
1596          * Second, a native PCIe machine introduces the possibility of fine
1597          * granularity IOMMUs supporting both translation and isolation.
1598          * Guest code to discover the IOMMU visibility of a device, such as
1599          * IOMMU grouping code on Linux, is very aware of device types and
1600          * valid transitions between bus types.  An express device on a non-
1601          * express bus is not a valid combination on bare metal systems.
1602          *
1603          * Drivers that require a PCIe capability to make the device
1604          * functional are simply going to need to have their devices placed
1605          * on a PCIe bus in the VM.
1606          */
1607         while (!pci_bus_is_root(bus)) {
1608             bridge = pci_bridge_get_device(bus);
1609             bus = bridge->bus;
1610         }
1611 
1612         if (pci_bus_is_express(bus)) {
1613             return 0;
1614         }
1615 
1616     } else if (pci_bus_is_root(vdev->pdev.bus)) {
1617         /*
1618          * On a Root Complex bus Endpoints become Root Complex Integrated
1619          * Endpoints, which changes the type and clears the LNK & LNK2 fields.
1620          */
1621         if (type == PCI_EXP_TYPE_ENDPOINT) {
1622             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1623                                    PCI_EXP_TYPE_RC_END << 4,
1624                                    PCI_EXP_FLAGS_TYPE);
1625 
1626             /* Link Capabilities, Status, and Control goes away */
1627             if (size > PCI_EXP_LNKCTL) {
1628                 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0);
1629                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1630                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0);
1631 
1632 #ifndef PCI_EXP_LNKCAP2
1633 #define PCI_EXP_LNKCAP2 44
1634 #endif
1635 #ifndef PCI_EXP_LNKSTA2
1636 #define PCI_EXP_LNKSTA2 50
1637 #endif
1638                 /* Link 2 Capabilities, Status, and Control goes away */
1639                 if (size > PCI_EXP_LNKCAP2) {
1640                     vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0);
1641                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0);
1642                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0);
1643                 }
1644             }
1645 
1646         } else if (type == PCI_EXP_TYPE_LEG_END) {
1647             /*
1648              * Legacy endpoints don't belong on the root complex.  Windows
1649              * seems to be happier with devices if we skip the capability.
1650              */
1651             return 0;
1652         }
1653 
1654     } else {
1655         /*
1656          * Convert Root Complex Integrated Endpoints to regular endpoints.
1657          * These devices don't support LNK/LNK2 capabilities, so make them up.
1658          */
1659         if (type == PCI_EXP_TYPE_RC_END) {
1660             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
1661                                    PCI_EXP_TYPE_ENDPOINT << 4,
1662                                    PCI_EXP_FLAGS_TYPE);
1663             vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP,
1664                                    PCI_EXP_LNK_MLW_1 | PCI_EXP_LNK_LS_25, ~0);
1665             vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
1666         }
1667 
1668         /* Mark the Link Status bits as emulated to allow virtual negotiation */
1669         vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA,
1670                                pci_get_word(vdev->pdev.config + pos +
1671                                             PCI_EXP_LNKSTA),
1672                                PCI_EXP_LNKCAP_MLW | PCI_EXP_LNKCAP_SLS);
1673     }
1674 
1675     pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size);
1676     if (pos >= 0) {
1677         vdev->pdev.exp.exp_cap = pos;
1678     }
1679 
1680     return pos;
1681 }
1682 
1683 static void vfio_check_pcie_flr(VFIOPCIDevice *vdev, uint8_t pos)
1684 {
1685     uint32_t cap = pci_get_long(vdev->pdev.config + pos + PCI_EXP_DEVCAP);
1686 
1687     if (cap & PCI_EXP_DEVCAP_FLR) {
1688         trace_vfio_check_pcie_flr(vdev->vbasedev.name);
1689         vdev->has_flr = true;
1690     }
1691 }
1692 
1693 static void vfio_check_pm_reset(VFIOPCIDevice *vdev, uint8_t pos)
1694 {
1695     uint16_t csr = pci_get_word(vdev->pdev.config + pos + PCI_PM_CTRL);
1696 
1697     if (!(csr & PCI_PM_CTRL_NO_SOFT_RESET)) {
1698         trace_vfio_check_pm_reset(vdev->vbasedev.name);
1699         vdev->has_pm_reset = true;
1700     }
1701 }
1702 
1703 static void vfio_check_af_flr(VFIOPCIDevice *vdev, uint8_t pos)
1704 {
1705     uint8_t cap = pci_get_byte(vdev->pdev.config + pos + PCI_AF_CAP);
1706 
1707     if ((cap & PCI_AF_CAP_TP) && (cap & PCI_AF_CAP_FLR)) {
1708         trace_vfio_check_af_flr(vdev->vbasedev.name);
1709         vdev->has_flr = true;
1710     }
1711 }
1712 
1713 static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos, Error **errp)
1714 {
1715     PCIDevice *pdev = &vdev->pdev;
1716     uint8_t cap_id, next, size;
1717     int ret;
1718 
1719     cap_id = pdev->config[pos];
1720     next = pdev->config[pos + PCI_CAP_LIST_NEXT];
1721 
1722     /*
1723      * If it becomes important to configure capabilities to their actual
1724      * size, use this as the default when it's something we don't recognize.
1725      * Since QEMU doesn't actually handle many of the config accesses,
1726      * exact size doesn't seem worthwhile.
1727      */
1728     size = vfio_std_cap_max_size(pdev, pos);
1729 
1730     /*
1731      * pci_add_capability always inserts the new capability at the head
1732      * of the chain.  Therefore to end up with a chain that matches the
1733      * physical device, we insert from the end by making this recursive.
1734      * This is also why we pre-calculate size above as cached config space
1735      * will be changed as we unwind the stack.
1736      */
1737     if (next) {
1738         ret = vfio_add_std_cap(vdev, next, errp);
1739         if (ret) {
1740             goto out;
1741         }
1742     } else {
1743         /* Begin the rebuild, use QEMU emulated list bits */
1744         pdev->config[PCI_CAPABILITY_LIST] = 0;
1745         vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
1746         vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
1747     }
1748 
1749     /* Use emulated next pointer to allow dropping caps */
1750     pci_set_byte(vdev->emulated_config_bits + pos + PCI_CAP_LIST_NEXT, 0xff);
1751 
1752     switch (cap_id) {
1753     case PCI_CAP_ID_MSI:
1754         ret = vfio_msi_setup(vdev, pos, errp);
1755         break;
1756     case PCI_CAP_ID_EXP:
1757         vfio_check_pcie_flr(vdev, pos);
1758         ret = vfio_setup_pcie_cap(vdev, pos, size, errp);
1759         break;
1760     case PCI_CAP_ID_MSIX:
1761         ret = vfio_msix_setup(vdev, pos, errp);
1762         break;
1763     case PCI_CAP_ID_PM:
1764         vfio_check_pm_reset(vdev, pos);
1765         vdev->pm_cap = pos;
1766         ret = pci_add_capability2(pdev, cap_id, pos, size, errp);
1767         break;
1768     case PCI_CAP_ID_AF:
1769         vfio_check_af_flr(vdev, pos);
1770         ret = pci_add_capability2(pdev, cap_id, pos, size, errp);
1771         break;
1772     default:
1773         ret = pci_add_capability2(pdev, cap_id, pos, size, errp);
1774         break;
1775     }
1776 out:
1777     if (ret < 0) {
1778         error_prepend(errp,
1779                       "failed to add PCI capability 0x%x[0x%x]@0x%x: ",
1780                       cap_id, size, pos);
1781         return ret;
1782     }
1783 
1784     return 0;
1785 }
1786 
1787 static void vfio_add_ext_cap(VFIOPCIDevice *vdev)
1788 {
1789     PCIDevice *pdev = &vdev->pdev;
1790     uint32_t header;
1791     uint16_t cap_id, next, size;
1792     uint8_t cap_ver;
1793     uint8_t *config;
1794 
1795     /* Only add extended caps if we have them and the guest can see them */
1796     if (!pci_is_express(pdev) || !pci_bus_is_express(pdev->bus) ||
1797         !pci_get_long(pdev->config + PCI_CONFIG_SPACE_SIZE)) {
1798         return;
1799     }
1800 
1801     /*
1802      * pcie_add_capability always inserts the new capability at the tail
1803      * of the chain.  Therefore to end up with a chain that matches the
1804      * physical device, we cache the config space to avoid overwriting
1805      * the original config space when we parse the extended capabilities.
1806      */
1807     config = g_memdup(pdev->config, vdev->config_size);
1808 
1809     /*
1810      * Extended capabilities are chained with each pointing to the next, so we
1811      * can drop anything other than the head of the chain simply by modifying
1812      * the previous next pointer.  For the head of the chain, we can modify the
1813      * capability ID to something that cannot match a valid capability.  ID
1814      * 0 is reserved for this since absence of capabilities is indicated by
1815      * 0 for the ID, version, AND next pointer.  However, pcie_add_capability()
1816      * uses ID 0 as reserved for list management and will incorrectly match and
1817      * assert if we attempt to pre-load the head of the chain with with this
1818      * ID.  Use ID 0xFFFF temporarily since it is also seems to be reserved in
1819      * part for identifying absence of capabilities in a root complex register
1820      * block.  If the ID still exists after adding capabilities, switch back to
1821      * zero.  We'll mark this entire first dword as emulated for this purpose.
1822      */
1823     pci_set_long(pdev->config + PCI_CONFIG_SPACE_SIZE,
1824                  PCI_EXT_CAP(0xFFFF, 0, 0));
1825     pci_set_long(pdev->wmask + PCI_CONFIG_SPACE_SIZE, 0);
1826     pci_set_long(vdev->emulated_config_bits + PCI_CONFIG_SPACE_SIZE, ~0);
1827 
1828     for (next = PCI_CONFIG_SPACE_SIZE; next;
1829          next = PCI_EXT_CAP_NEXT(pci_get_long(config + next))) {
1830         header = pci_get_long(config + next);
1831         cap_id = PCI_EXT_CAP_ID(header);
1832         cap_ver = PCI_EXT_CAP_VER(header);
1833 
1834         /*
1835          * If it becomes important to configure extended capabilities to their
1836          * actual size, use this as the default when it's something we don't
1837          * recognize. Since QEMU doesn't actually handle many of the config
1838          * accesses, exact size doesn't seem worthwhile.
1839          */
1840         size = vfio_ext_cap_max_size(config, next);
1841 
1842         /* Use emulated next pointer to allow dropping extended caps */
1843         pci_long_test_and_set_mask(vdev->emulated_config_bits + next,
1844                                    PCI_EXT_CAP_NEXT_MASK);
1845 
1846         switch (cap_id) {
1847         case PCI_EXT_CAP_ID_SRIOV: /* Read-only VF BARs confuse OVMF */
1848         case PCI_EXT_CAP_ID_ARI: /* XXX Needs next function virtualization */
1849             trace_vfio_add_ext_cap_dropped(vdev->vbasedev.name, cap_id, next);
1850             break;
1851         default:
1852             pcie_add_capability(pdev, cap_id, cap_ver, next, size);
1853         }
1854 
1855     }
1856 
1857     /* Cleanup chain head ID if necessary */
1858     if (pci_get_word(pdev->config + PCI_CONFIG_SPACE_SIZE) == 0xFFFF) {
1859         pci_set_word(pdev->config + PCI_CONFIG_SPACE_SIZE, 0);
1860     }
1861 
1862     g_free(config);
1863     return;
1864 }
1865 
1866 static int vfio_add_capabilities(VFIOPCIDevice *vdev, Error **errp)
1867 {
1868     PCIDevice *pdev = &vdev->pdev;
1869     int ret;
1870 
1871     if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
1872         !pdev->config[PCI_CAPABILITY_LIST]) {
1873         return 0; /* Nothing to add */
1874     }
1875 
1876     ret = vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST], errp);
1877     if (ret) {
1878         return ret;
1879     }
1880 
1881     vfio_add_ext_cap(vdev);
1882     return 0;
1883 }
1884 
1885 static void vfio_pci_pre_reset(VFIOPCIDevice *vdev)
1886 {
1887     PCIDevice *pdev = &vdev->pdev;
1888     uint16_t cmd;
1889 
1890     vfio_disable_interrupts(vdev);
1891 
1892     /* Make sure the device is in D0 */
1893     if (vdev->pm_cap) {
1894         uint16_t pmcsr;
1895         uint8_t state;
1896 
1897         pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
1898         state = pmcsr & PCI_PM_CTRL_STATE_MASK;
1899         if (state) {
1900             pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
1901             vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2);
1902             /* vfio handles the necessary delay here */
1903             pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
1904             state = pmcsr & PCI_PM_CTRL_STATE_MASK;
1905             if (state) {
1906                 error_report("vfio: Unable to power on device, stuck in D%d",
1907                              state);
1908             }
1909         }
1910     }
1911 
1912     /*
1913      * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
1914      * Also put INTx Disable in known state.
1915      */
1916     cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
1917     cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
1918              PCI_COMMAND_INTX_DISABLE);
1919     vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
1920 }
1921 
1922 static void vfio_pci_post_reset(VFIOPCIDevice *vdev)
1923 {
1924     Error *err = NULL;
1925 
1926     vfio_intx_enable(vdev, &err);
1927     if (err) {
1928         error_reportf_err(err, ERR_PREFIX, vdev->vbasedev.name);
1929     }
1930 }
1931 
1932 static bool vfio_pci_host_match(PCIHostDeviceAddress *addr, const char *name)
1933 {
1934     char tmp[13];
1935 
1936     sprintf(tmp, "%04x:%02x:%02x.%1x", addr->domain,
1937             addr->bus, addr->slot, addr->function);
1938 
1939     return (strcmp(tmp, name) == 0);
1940 }
1941 
1942 static int vfio_pci_hot_reset(VFIOPCIDevice *vdev, bool single)
1943 {
1944     VFIOGroup *group;
1945     struct vfio_pci_hot_reset_info *info;
1946     struct vfio_pci_dependent_device *devices;
1947     struct vfio_pci_hot_reset *reset;
1948     int32_t *fds;
1949     int ret, i, count;
1950     bool multi = false;
1951 
1952     trace_vfio_pci_hot_reset(vdev->vbasedev.name, single ? "one" : "multi");
1953 
1954     if (!single) {
1955         vfio_pci_pre_reset(vdev);
1956     }
1957     vdev->vbasedev.needs_reset = false;
1958 
1959     info = g_malloc0(sizeof(*info));
1960     info->argsz = sizeof(*info);
1961 
1962     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
1963     if (ret && errno != ENOSPC) {
1964         ret = -errno;
1965         if (!vdev->has_pm_reset) {
1966             error_report("vfio: Cannot reset device %s, "
1967                          "no available reset mechanism.", vdev->vbasedev.name);
1968         }
1969         goto out_single;
1970     }
1971 
1972     count = info->count;
1973     info = g_realloc(info, sizeof(*info) + (count * sizeof(*devices)));
1974     info->argsz = sizeof(*info) + (count * sizeof(*devices));
1975     devices = &info->devices[0];
1976 
1977     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
1978     if (ret) {
1979         ret = -errno;
1980         error_report("vfio: hot reset info failed: %m");
1981         goto out_single;
1982     }
1983 
1984     trace_vfio_pci_hot_reset_has_dep_devices(vdev->vbasedev.name);
1985 
1986     /* Verify that we have all the groups required */
1987     for (i = 0; i < info->count; i++) {
1988         PCIHostDeviceAddress host;
1989         VFIOPCIDevice *tmp;
1990         VFIODevice *vbasedev_iter;
1991 
1992         host.domain = devices[i].segment;
1993         host.bus = devices[i].bus;
1994         host.slot = PCI_SLOT(devices[i].devfn);
1995         host.function = PCI_FUNC(devices[i].devfn);
1996 
1997         trace_vfio_pci_hot_reset_dep_devices(host.domain,
1998                 host.bus, host.slot, host.function, devices[i].group_id);
1999 
2000         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2001             continue;
2002         }
2003 
2004         QLIST_FOREACH(group, &vfio_group_list, next) {
2005             if (group->groupid == devices[i].group_id) {
2006                 break;
2007             }
2008         }
2009 
2010         if (!group) {
2011             if (!vdev->has_pm_reset) {
2012                 error_report("vfio: Cannot reset device %s, "
2013                              "depends on group %d which is not owned.",
2014                              vdev->vbasedev.name, devices[i].group_id);
2015             }
2016             ret = -EPERM;
2017             goto out;
2018         }
2019 
2020         /* Prep dependent devices for reset and clear our marker. */
2021         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2022             if (vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2023                 continue;
2024             }
2025             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2026             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2027                 if (single) {
2028                     ret = -EINVAL;
2029                     goto out_single;
2030                 }
2031                 vfio_pci_pre_reset(tmp);
2032                 tmp->vbasedev.needs_reset = false;
2033                 multi = true;
2034                 break;
2035             }
2036         }
2037     }
2038 
2039     if (!single && !multi) {
2040         ret = -EINVAL;
2041         goto out_single;
2042     }
2043 
2044     /* Determine how many group fds need to be passed */
2045     count = 0;
2046     QLIST_FOREACH(group, &vfio_group_list, next) {
2047         for (i = 0; i < info->count; i++) {
2048             if (group->groupid == devices[i].group_id) {
2049                 count++;
2050                 break;
2051             }
2052         }
2053     }
2054 
2055     reset = g_malloc0(sizeof(*reset) + (count * sizeof(*fds)));
2056     reset->argsz = sizeof(*reset) + (count * sizeof(*fds));
2057     fds = &reset->group_fds[0];
2058 
2059     /* Fill in group fds */
2060     QLIST_FOREACH(group, &vfio_group_list, next) {
2061         for (i = 0; i < info->count; i++) {
2062             if (group->groupid == devices[i].group_id) {
2063                 fds[reset->count++] = group->fd;
2064                 break;
2065             }
2066         }
2067     }
2068 
2069     /* Bus reset! */
2070     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_PCI_HOT_RESET, reset);
2071     g_free(reset);
2072 
2073     trace_vfio_pci_hot_reset_result(vdev->vbasedev.name,
2074                                     ret ? "%m" : "Success");
2075 
2076 out:
2077     /* Re-enable INTx on affected devices */
2078     for (i = 0; i < info->count; i++) {
2079         PCIHostDeviceAddress host;
2080         VFIOPCIDevice *tmp;
2081         VFIODevice *vbasedev_iter;
2082 
2083         host.domain = devices[i].segment;
2084         host.bus = devices[i].bus;
2085         host.slot = PCI_SLOT(devices[i].devfn);
2086         host.function = PCI_FUNC(devices[i].devfn);
2087 
2088         if (vfio_pci_host_match(&host, vdev->vbasedev.name)) {
2089             continue;
2090         }
2091 
2092         QLIST_FOREACH(group, &vfio_group_list, next) {
2093             if (group->groupid == devices[i].group_id) {
2094                 break;
2095             }
2096         }
2097 
2098         if (!group) {
2099             break;
2100         }
2101 
2102         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2103             if (vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2104                 continue;
2105             }
2106             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2107             if (vfio_pci_host_match(&host, tmp->vbasedev.name)) {
2108                 vfio_pci_post_reset(tmp);
2109                 break;
2110             }
2111         }
2112     }
2113 out_single:
2114     if (!single) {
2115         vfio_pci_post_reset(vdev);
2116     }
2117     g_free(info);
2118 
2119     return ret;
2120 }
2121 
2122 /*
2123  * We want to differentiate hot reset of mulitple in-use devices vs hot reset
2124  * of a single in-use device.  VFIO_DEVICE_RESET will already handle the case
2125  * of doing hot resets when there is only a single device per bus.  The in-use
2126  * here refers to how many VFIODevices are affected.  A hot reset that affects
2127  * multiple devices, but only a single in-use device, means that we can call
2128  * it from our bus ->reset() callback since the extent is effectively a single
2129  * device.  This allows us to make use of it in the hotplug path.  When there
2130  * are multiple in-use devices, we can only trigger the hot reset during a
2131  * system reset and thus from our reset handler.  We separate _one vs _multi
2132  * here so that we don't overlap and do a double reset on the system reset
2133  * path where both our reset handler and ->reset() callback are used.  Calling
2134  * _one() will only do a hot reset for the one in-use devices case, calling
2135  * _multi() will do nothing if a _one() would have been sufficient.
2136  */
2137 static int vfio_pci_hot_reset_one(VFIOPCIDevice *vdev)
2138 {
2139     return vfio_pci_hot_reset(vdev, true);
2140 }
2141 
2142 static int vfio_pci_hot_reset_multi(VFIODevice *vbasedev)
2143 {
2144     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2145     return vfio_pci_hot_reset(vdev, false);
2146 }
2147 
2148 static void vfio_pci_compute_needs_reset(VFIODevice *vbasedev)
2149 {
2150     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
2151     if (!vbasedev->reset_works || (!vdev->has_flr && vdev->has_pm_reset)) {
2152         vbasedev->needs_reset = true;
2153     }
2154 }
2155 
2156 static VFIODeviceOps vfio_pci_ops = {
2157     .vfio_compute_needs_reset = vfio_pci_compute_needs_reset,
2158     .vfio_hot_reset_multi = vfio_pci_hot_reset_multi,
2159     .vfio_eoi = vfio_intx_eoi,
2160 };
2161 
2162 int vfio_populate_vga(VFIOPCIDevice *vdev, Error **errp)
2163 {
2164     VFIODevice *vbasedev = &vdev->vbasedev;
2165     struct vfio_region_info *reg_info;
2166     int ret;
2167 
2168     ret = vfio_get_region_info(vbasedev, VFIO_PCI_VGA_REGION_INDEX, &reg_info);
2169     if (ret) {
2170         error_setg_errno(errp, -ret,
2171                          "failed getting region info for VGA region index %d",
2172                          VFIO_PCI_VGA_REGION_INDEX);
2173         return ret;
2174     }
2175 
2176     if (!(reg_info->flags & VFIO_REGION_INFO_FLAG_READ) ||
2177         !(reg_info->flags & VFIO_REGION_INFO_FLAG_WRITE) ||
2178         reg_info->size < 0xbffff + 1) {
2179         error_setg(errp, "unexpected VGA info, flags 0x%lx, size 0x%lx",
2180                    (unsigned long)reg_info->flags,
2181                    (unsigned long)reg_info->size);
2182         g_free(reg_info);
2183         return -EINVAL;
2184     }
2185 
2186     vdev->vga = g_new0(VFIOVGA, 1);
2187 
2188     vdev->vga->fd_offset = reg_info->offset;
2189     vdev->vga->fd = vdev->vbasedev.fd;
2190 
2191     g_free(reg_info);
2192 
2193     vdev->vga->region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE;
2194     vdev->vga->region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM;
2195     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_MEM].quirks);
2196 
2197     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2198                           OBJECT(vdev), &vfio_vga_ops,
2199                           &vdev->vga->region[QEMU_PCI_VGA_MEM],
2200                           "vfio-vga-mmio@0xa0000",
2201                           QEMU_PCI_VGA_MEM_SIZE);
2202 
2203     vdev->vga->region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE;
2204     vdev->vga->region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO;
2205     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].quirks);
2206 
2207     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2208                           OBJECT(vdev), &vfio_vga_ops,
2209                           &vdev->vga->region[QEMU_PCI_VGA_IO_LO],
2210                           "vfio-vga-io@0x3b0",
2211                           QEMU_PCI_VGA_IO_LO_SIZE);
2212 
2213     vdev->vga->region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE;
2214     vdev->vga->region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI;
2215     QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks);
2216 
2217     memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
2218                           OBJECT(vdev), &vfio_vga_ops,
2219                           &vdev->vga->region[QEMU_PCI_VGA_IO_HI],
2220                           "vfio-vga-io@0x3c0",
2221                           QEMU_PCI_VGA_IO_HI_SIZE);
2222 
2223     pci_register_vga(&vdev->pdev, &vdev->vga->region[QEMU_PCI_VGA_MEM].mem,
2224                      &vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem,
2225                      &vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem);
2226 
2227     return 0;
2228 }
2229 
2230 static void vfio_populate_device(VFIOPCIDevice *vdev, Error **errp)
2231 {
2232     VFIODevice *vbasedev = &vdev->vbasedev;
2233     struct vfio_region_info *reg_info;
2234     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) };
2235     int i, ret = -1;
2236 
2237     /* Sanity check device */
2238     if (!(vbasedev->flags & VFIO_DEVICE_FLAGS_PCI)) {
2239         error_setg(errp, "this isn't a PCI device");
2240         return;
2241     }
2242 
2243     if (vbasedev->num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
2244         error_setg(errp, "unexpected number of io regions %u",
2245                    vbasedev->num_regions);
2246         return;
2247     }
2248 
2249     if (vbasedev->num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
2250         error_setg(errp, "unexpected number of irqs %u", vbasedev->num_irqs);
2251         return;
2252     }
2253 
2254     for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
2255         char *name = g_strdup_printf("%s BAR %d", vbasedev->name, i);
2256 
2257         ret = vfio_region_setup(OBJECT(vdev), vbasedev,
2258                                 &vdev->bars[i].region, i, name);
2259         g_free(name);
2260 
2261         if (ret) {
2262             error_setg_errno(errp, -ret, "failed to get region %d info", i);
2263             return;
2264         }
2265 
2266         QLIST_INIT(&vdev->bars[i].quirks);
2267     }
2268 
2269     ret = vfio_get_region_info(vbasedev,
2270                                VFIO_PCI_CONFIG_REGION_INDEX, &reg_info);
2271     if (ret) {
2272         error_setg_errno(errp, -ret, "failed to get config info");
2273         return;
2274     }
2275 
2276     trace_vfio_populate_device_config(vdev->vbasedev.name,
2277                                       (unsigned long)reg_info->size,
2278                                       (unsigned long)reg_info->offset,
2279                                       (unsigned long)reg_info->flags);
2280 
2281     vdev->config_size = reg_info->size;
2282     if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
2283         vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
2284     }
2285     vdev->config_offset = reg_info->offset;
2286 
2287     g_free(reg_info);
2288 
2289     if (vdev->features & VFIO_FEATURE_ENABLE_VGA) {
2290         ret = vfio_populate_vga(vdev, errp);
2291         if (ret) {
2292             error_append_hint(errp, "device does not support "
2293                               "requested feature x-vga\n");
2294             return;
2295         }
2296     }
2297 
2298     irq_info.index = VFIO_PCI_ERR_IRQ_INDEX;
2299 
2300     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info);
2301     if (ret) {
2302         /* This can fail for an old kernel or legacy PCI dev */
2303         trace_vfio_populate_device_get_irq_info_failure();
2304     } else if (irq_info.count == 1) {
2305         vdev->pci_aer = true;
2306     } else {
2307         error_report(WARN_PREFIX
2308                      "Could not enable error recovery for the device",
2309                      vbasedev->name);
2310     }
2311 }
2312 
2313 static void vfio_put_device(VFIOPCIDevice *vdev)
2314 {
2315     g_free(vdev->vbasedev.name);
2316     g_free(vdev->msix);
2317 
2318     vfio_put_base_device(&vdev->vbasedev);
2319 }
2320 
2321 static void vfio_err_notifier_handler(void *opaque)
2322 {
2323     VFIOPCIDevice *vdev = opaque;
2324 
2325     if (!event_notifier_test_and_clear(&vdev->err_notifier)) {
2326         return;
2327     }
2328 
2329     /*
2330      * TBD. Retrieve the error details and decide what action
2331      * needs to be taken. One of the actions could be to pass
2332      * the error to the guest and have the guest driver recover
2333      * from the error. This requires that PCIe capabilities be
2334      * exposed to the guest. For now, we just terminate the
2335      * guest to contain the error.
2336      */
2337 
2338     error_report("%s(%s) Unrecoverable error detected. Please collect any data possible and then kill the guest", __func__, vdev->vbasedev.name);
2339 
2340     vm_stop(RUN_STATE_INTERNAL_ERROR);
2341 }
2342 
2343 /*
2344  * Registers error notifier for devices supporting error recovery.
2345  * If we encounter a failure in this function, we report an error
2346  * and continue after disabling error recovery support for the
2347  * device.
2348  */
2349 static void vfio_register_err_notifier(VFIOPCIDevice *vdev)
2350 {
2351     int ret;
2352     int argsz;
2353     struct vfio_irq_set *irq_set;
2354     int32_t *pfd;
2355 
2356     if (!vdev->pci_aer) {
2357         return;
2358     }
2359 
2360     if (event_notifier_init(&vdev->err_notifier, 0)) {
2361         error_report("vfio: Unable to init event notifier for error detection");
2362         vdev->pci_aer = false;
2363         return;
2364     }
2365 
2366     argsz = sizeof(*irq_set) + sizeof(*pfd);
2367 
2368     irq_set = g_malloc0(argsz);
2369     irq_set->argsz = argsz;
2370     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2371                      VFIO_IRQ_SET_ACTION_TRIGGER;
2372     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
2373     irq_set->start = 0;
2374     irq_set->count = 1;
2375     pfd = (int32_t *)&irq_set->data;
2376 
2377     *pfd = event_notifier_get_fd(&vdev->err_notifier);
2378     qemu_set_fd_handler(*pfd, vfio_err_notifier_handler, NULL, vdev);
2379 
2380     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
2381     if (ret) {
2382         error_report("vfio: Failed to set up error notification");
2383         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
2384         event_notifier_cleanup(&vdev->err_notifier);
2385         vdev->pci_aer = false;
2386     }
2387     g_free(irq_set);
2388 }
2389 
2390 static void vfio_unregister_err_notifier(VFIOPCIDevice *vdev)
2391 {
2392     int argsz;
2393     struct vfio_irq_set *irq_set;
2394     int32_t *pfd;
2395     int ret;
2396 
2397     if (!vdev->pci_aer) {
2398         return;
2399     }
2400 
2401     argsz = sizeof(*irq_set) + sizeof(*pfd);
2402 
2403     irq_set = g_malloc0(argsz);
2404     irq_set->argsz = argsz;
2405     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2406                      VFIO_IRQ_SET_ACTION_TRIGGER;
2407     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
2408     irq_set->start = 0;
2409     irq_set->count = 1;
2410     pfd = (int32_t *)&irq_set->data;
2411     *pfd = -1;
2412 
2413     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
2414     if (ret) {
2415         error_report("vfio: Failed to de-assign error fd: %m");
2416     }
2417     g_free(irq_set);
2418     qemu_set_fd_handler(event_notifier_get_fd(&vdev->err_notifier),
2419                         NULL, NULL, vdev);
2420     event_notifier_cleanup(&vdev->err_notifier);
2421 }
2422 
2423 static void vfio_req_notifier_handler(void *opaque)
2424 {
2425     VFIOPCIDevice *vdev = opaque;
2426 
2427     if (!event_notifier_test_and_clear(&vdev->req_notifier)) {
2428         return;
2429     }
2430 
2431     qdev_unplug(&vdev->pdev.qdev, NULL);
2432 }
2433 
2434 static void vfio_register_req_notifier(VFIOPCIDevice *vdev)
2435 {
2436     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info),
2437                                       .index = VFIO_PCI_REQ_IRQ_INDEX };
2438     int argsz;
2439     struct vfio_irq_set *irq_set;
2440     int32_t *pfd;
2441 
2442     if (!(vdev->features & VFIO_FEATURE_ENABLE_REQ)) {
2443         return;
2444     }
2445 
2446     if (ioctl(vdev->vbasedev.fd,
2447               VFIO_DEVICE_GET_IRQ_INFO, &irq_info) < 0 || irq_info.count < 1) {
2448         return;
2449     }
2450 
2451     if (event_notifier_init(&vdev->req_notifier, 0)) {
2452         error_report("vfio: Unable to init event notifier for device request");
2453         return;
2454     }
2455 
2456     argsz = sizeof(*irq_set) + sizeof(*pfd);
2457 
2458     irq_set = g_malloc0(argsz);
2459     irq_set->argsz = argsz;
2460     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2461                      VFIO_IRQ_SET_ACTION_TRIGGER;
2462     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
2463     irq_set->start = 0;
2464     irq_set->count = 1;
2465     pfd = (int32_t *)&irq_set->data;
2466 
2467     *pfd = event_notifier_get_fd(&vdev->req_notifier);
2468     qemu_set_fd_handler(*pfd, vfio_req_notifier_handler, NULL, vdev);
2469 
2470     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
2471         error_report("vfio: Failed to set up device request notification");
2472         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
2473         event_notifier_cleanup(&vdev->req_notifier);
2474     } else {
2475         vdev->req_enabled = true;
2476     }
2477 
2478     g_free(irq_set);
2479 }
2480 
2481 static void vfio_unregister_req_notifier(VFIOPCIDevice *vdev)
2482 {
2483     int argsz;
2484     struct vfio_irq_set *irq_set;
2485     int32_t *pfd;
2486 
2487     if (!vdev->req_enabled) {
2488         return;
2489     }
2490 
2491     argsz = sizeof(*irq_set) + sizeof(*pfd);
2492 
2493     irq_set = g_malloc0(argsz);
2494     irq_set->argsz = argsz;
2495     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
2496                      VFIO_IRQ_SET_ACTION_TRIGGER;
2497     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
2498     irq_set->start = 0;
2499     irq_set->count = 1;
2500     pfd = (int32_t *)&irq_set->data;
2501     *pfd = -1;
2502 
2503     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
2504         error_report("vfio: Failed to de-assign device request fd: %m");
2505     }
2506     g_free(irq_set);
2507     qemu_set_fd_handler(event_notifier_get_fd(&vdev->req_notifier),
2508                         NULL, NULL, vdev);
2509     event_notifier_cleanup(&vdev->req_notifier);
2510 
2511     vdev->req_enabled = false;
2512 }
2513 
2514 static void vfio_realize(PCIDevice *pdev, Error **errp)
2515 {
2516     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
2517     VFIODevice *vbasedev_iter;
2518     VFIOGroup *group;
2519     char *tmp, group_path[PATH_MAX], *group_name;
2520     Error *err = NULL;
2521     ssize_t len;
2522     struct stat st;
2523     int groupid;
2524     int i, ret;
2525 
2526     if (!vdev->vbasedev.sysfsdev) {
2527         if (!(~vdev->host.domain || ~vdev->host.bus ||
2528               ~vdev->host.slot || ~vdev->host.function)) {
2529             error_setg(errp, "No provided host device");
2530             error_append_hint(errp, "Use -vfio-pci,host=DDDD:BB:DD.F "
2531                               "or -vfio-pci,sysfsdev=PATH_TO_DEVICE\n");
2532             return;
2533         }
2534         vdev->vbasedev.sysfsdev =
2535             g_strdup_printf("/sys/bus/pci/devices/%04x:%02x:%02x.%01x",
2536                             vdev->host.domain, vdev->host.bus,
2537                             vdev->host.slot, vdev->host.function);
2538     }
2539 
2540     if (stat(vdev->vbasedev.sysfsdev, &st) < 0) {
2541         error_setg_errno(errp, errno, "no such host device");
2542         error_prepend(errp, ERR_PREFIX, vdev->vbasedev.sysfsdev);
2543         return;
2544     }
2545 
2546     vdev->vbasedev.name = g_strdup(basename(vdev->vbasedev.sysfsdev));
2547     vdev->vbasedev.ops = &vfio_pci_ops;
2548     vdev->vbasedev.type = VFIO_DEVICE_TYPE_PCI;
2549 
2550     tmp = g_strdup_printf("%s/iommu_group", vdev->vbasedev.sysfsdev);
2551     len = readlink(tmp, group_path, sizeof(group_path));
2552     g_free(tmp);
2553 
2554     if (len <= 0 || len >= sizeof(group_path)) {
2555         error_setg_errno(errp, len < 0 ? errno : ENAMETOOLONG,
2556                          "no iommu_group found");
2557         goto error;
2558     }
2559 
2560     group_path[len] = 0;
2561 
2562     group_name = basename(group_path);
2563     if (sscanf(group_name, "%d", &groupid) != 1) {
2564         error_setg_errno(errp, errno, "failed to read %s", group_path);
2565         goto error;
2566     }
2567 
2568     trace_vfio_realize(vdev->vbasedev.name, groupid);
2569 
2570     group = vfio_get_group(groupid, pci_device_iommu_address_space(pdev), errp);
2571     if (!group) {
2572         goto error;
2573     }
2574 
2575     QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2576         if (strcmp(vbasedev_iter->name, vdev->vbasedev.name) == 0) {
2577             error_setg(errp, "device is already attached");
2578             vfio_put_group(group);
2579             goto error;
2580         }
2581     }
2582 
2583     ret = vfio_get_device(group, vdev->vbasedev.name, &vdev->vbasedev, errp);
2584     if (ret) {
2585         vfio_put_group(group);
2586         goto error;
2587     }
2588 
2589     vfio_populate_device(vdev, &err);
2590     if (err) {
2591         error_propagate(errp, err);
2592         goto error;
2593     }
2594 
2595     /* Get a copy of config space */
2596     ret = pread(vdev->vbasedev.fd, vdev->pdev.config,
2597                 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
2598                 vdev->config_offset);
2599     if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
2600         ret = ret < 0 ? -errno : -EFAULT;
2601         error_setg_errno(errp, -ret, "failed to read device config space");
2602         goto error;
2603     }
2604 
2605     /* vfio emulates a lot for us, but some bits need extra love */
2606     vdev->emulated_config_bits = g_malloc0(vdev->config_size);
2607 
2608     /* QEMU can choose to expose the ROM or not */
2609     memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4);
2610 
2611     /*
2612      * The PCI spec reserves vendor ID 0xffff as an invalid value.  The
2613      * device ID is managed by the vendor and need only be a 16-bit value.
2614      * Allow any 16-bit value for subsystem so they can be hidden or changed.
2615      */
2616     if (vdev->vendor_id != PCI_ANY_ID) {
2617         if (vdev->vendor_id >= 0xffff) {
2618             error_setg(errp, "invalid PCI vendor ID provided");
2619             goto error;
2620         }
2621         vfio_add_emulated_word(vdev, PCI_VENDOR_ID, vdev->vendor_id, ~0);
2622         trace_vfio_pci_emulated_vendor_id(vdev->vbasedev.name, vdev->vendor_id);
2623     } else {
2624         vdev->vendor_id = pci_get_word(pdev->config + PCI_VENDOR_ID);
2625     }
2626 
2627     if (vdev->device_id != PCI_ANY_ID) {
2628         if (vdev->device_id > 0xffff) {
2629             error_setg(errp, "invalid PCI device ID provided");
2630             goto error;
2631         }
2632         vfio_add_emulated_word(vdev, PCI_DEVICE_ID, vdev->device_id, ~0);
2633         trace_vfio_pci_emulated_device_id(vdev->vbasedev.name, vdev->device_id);
2634     } else {
2635         vdev->device_id = pci_get_word(pdev->config + PCI_DEVICE_ID);
2636     }
2637 
2638     if (vdev->sub_vendor_id != PCI_ANY_ID) {
2639         if (vdev->sub_vendor_id > 0xffff) {
2640             error_setg(errp, "invalid PCI subsystem vendor ID provided");
2641             goto error;
2642         }
2643         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_VENDOR_ID,
2644                                vdev->sub_vendor_id, ~0);
2645         trace_vfio_pci_emulated_sub_vendor_id(vdev->vbasedev.name,
2646                                               vdev->sub_vendor_id);
2647     }
2648 
2649     if (vdev->sub_device_id != PCI_ANY_ID) {
2650         if (vdev->sub_device_id > 0xffff) {
2651             error_setg(errp, "invalid PCI subsystem device ID provided");
2652             goto error;
2653         }
2654         vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_ID, vdev->sub_device_id, ~0);
2655         trace_vfio_pci_emulated_sub_device_id(vdev->vbasedev.name,
2656                                               vdev->sub_device_id);
2657     }
2658 
2659     /* QEMU can change multi-function devices to single function, or reverse */
2660     vdev->emulated_config_bits[PCI_HEADER_TYPE] =
2661                                               PCI_HEADER_TYPE_MULTI_FUNCTION;
2662 
2663     /* Restore or clear multifunction, this is always controlled by QEMU */
2664     if (vdev->pdev.cap_present & QEMU_PCI_CAP_MULTIFUNCTION) {
2665         vdev->pdev.config[PCI_HEADER_TYPE] |= PCI_HEADER_TYPE_MULTI_FUNCTION;
2666     } else {
2667         vdev->pdev.config[PCI_HEADER_TYPE] &= ~PCI_HEADER_TYPE_MULTI_FUNCTION;
2668     }
2669 
2670     /*
2671      * Clear host resource mapping info.  If we choose not to register a
2672      * BAR, such as might be the case with the option ROM, we can get
2673      * confusing, unwritable, residual addresses from the host here.
2674      */
2675     memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
2676     memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
2677 
2678     vfio_pci_size_rom(vdev);
2679 
2680     vfio_msix_early_setup(vdev, &err);
2681     if (err) {
2682         error_propagate(errp, err);
2683         goto error;
2684     }
2685 
2686     vfio_bars_setup(vdev);
2687 
2688     ret = vfio_add_capabilities(vdev, errp);
2689     if (ret) {
2690         goto out_teardown;
2691     }
2692 
2693     if (vdev->vga) {
2694         vfio_vga_quirk_setup(vdev);
2695     }
2696 
2697     for (i = 0; i < PCI_ROM_SLOT; i++) {
2698         vfio_bar_quirk_setup(vdev, i);
2699     }
2700 
2701     if (!vdev->igd_opregion &&
2702         vdev->features & VFIO_FEATURE_ENABLE_IGD_OPREGION) {
2703         struct vfio_region_info *opregion;
2704 
2705         if (vdev->pdev.qdev.hotplugged) {
2706             error_setg(errp,
2707                        "cannot support IGD OpRegion feature on hotplugged "
2708                        "device");
2709             goto out_teardown;
2710         }
2711 
2712         ret = vfio_get_dev_region_info(&vdev->vbasedev,
2713                         VFIO_REGION_TYPE_PCI_VENDOR_TYPE | PCI_VENDOR_ID_INTEL,
2714                         VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION, &opregion);
2715         if (ret) {
2716             error_setg_errno(errp, -ret,
2717                              "does not support requested IGD OpRegion feature");
2718             goto out_teardown;
2719         }
2720 
2721         ret = vfio_pci_igd_opregion_init(vdev, opregion, errp);
2722         g_free(opregion);
2723         if (ret) {
2724             goto out_teardown;
2725         }
2726     }
2727 
2728     /* QEMU emulates all of MSI & MSIX */
2729     if (pdev->cap_present & QEMU_PCI_CAP_MSIX) {
2730         memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff,
2731                MSIX_CAP_LENGTH);
2732     }
2733 
2734     if (pdev->cap_present & QEMU_PCI_CAP_MSI) {
2735         memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff,
2736                vdev->msi_cap_size);
2737     }
2738 
2739     if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
2740         vdev->intx.mmap_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL,
2741                                                   vfio_intx_mmap_enable, vdev);
2742         pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_intx_update);
2743         ret = vfio_intx_enable(vdev, errp);
2744         if (ret) {
2745             goto out_teardown;
2746         }
2747     }
2748 
2749     vfio_register_err_notifier(vdev);
2750     vfio_register_req_notifier(vdev);
2751     vfio_setup_resetfn_quirk(vdev);
2752 
2753     return;
2754 
2755 out_teardown:
2756     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
2757     vfio_teardown_msi(vdev);
2758     vfio_bars_exit(vdev);
2759 error:
2760     error_prepend(errp, ERR_PREFIX, vdev->vbasedev.name);
2761 }
2762 
2763 static void vfio_instance_finalize(Object *obj)
2764 {
2765     PCIDevice *pci_dev = PCI_DEVICE(obj);
2766     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pci_dev);
2767     VFIOGroup *group = vdev->vbasedev.group;
2768 
2769     vfio_bars_finalize(vdev);
2770     g_free(vdev->emulated_config_bits);
2771     g_free(vdev->rom);
2772     /*
2773      * XXX Leaking igd_opregion is not an oversight, we can't remove the
2774      * fw_cfg entry therefore leaking this allocation seems like the safest
2775      * option.
2776      *
2777      * g_free(vdev->igd_opregion);
2778      */
2779     vfio_put_device(vdev);
2780     vfio_put_group(group);
2781 }
2782 
2783 static void vfio_exitfn(PCIDevice *pdev)
2784 {
2785     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
2786 
2787     vfio_unregister_req_notifier(vdev);
2788     vfio_unregister_err_notifier(vdev);
2789     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
2790     vfio_disable_interrupts(vdev);
2791     if (vdev->intx.mmap_timer) {
2792         timer_free(vdev->intx.mmap_timer);
2793     }
2794     vfio_teardown_msi(vdev);
2795     vfio_bars_exit(vdev);
2796 }
2797 
2798 static void vfio_pci_reset(DeviceState *dev)
2799 {
2800     PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
2801     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
2802 
2803     trace_vfio_pci_reset(vdev->vbasedev.name);
2804 
2805     vfio_pci_pre_reset(vdev);
2806 
2807     if (vdev->resetfn && !vdev->resetfn(vdev)) {
2808         goto post_reset;
2809     }
2810 
2811     if (vdev->vbasedev.reset_works &&
2812         (vdev->has_flr || !vdev->has_pm_reset) &&
2813         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
2814         trace_vfio_pci_reset_flr(vdev->vbasedev.name);
2815         goto post_reset;
2816     }
2817 
2818     /* See if we can do our own bus reset */
2819     if (!vfio_pci_hot_reset_one(vdev)) {
2820         goto post_reset;
2821     }
2822 
2823     /* If nothing else works and the device supports PM reset, use it */
2824     if (vdev->vbasedev.reset_works && vdev->has_pm_reset &&
2825         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
2826         trace_vfio_pci_reset_pm(vdev->vbasedev.name);
2827         goto post_reset;
2828     }
2829 
2830 post_reset:
2831     vfio_pci_post_reset(vdev);
2832 }
2833 
2834 static void vfio_instance_init(Object *obj)
2835 {
2836     PCIDevice *pci_dev = PCI_DEVICE(obj);
2837     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, PCI_DEVICE(obj));
2838 
2839     device_add_bootindex_property(obj, &vdev->bootindex,
2840                                   "bootindex", NULL,
2841                                   &pci_dev->qdev, NULL);
2842     vdev->host.domain = ~0U;
2843     vdev->host.bus = ~0U;
2844     vdev->host.slot = ~0U;
2845     vdev->host.function = ~0U;
2846 }
2847 
2848 static Property vfio_pci_dev_properties[] = {
2849     DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIOPCIDevice, host),
2850     DEFINE_PROP_STRING("sysfsdev", VFIOPCIDevice, vbasedev.sysfsdev),
2851     DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIOPCIDevice,
2852                        intx.mmap_timeout, 1100),
2853     DEFINE_PROP_BIT("x-vga", VFIOPCIDevice, features,
2854                     VFIO_FEATURE_ENABLE_VGA_BIT, false),
2855     DEFINE_PROP_BIT("x-req", VFIOPCIDevice, features,
2856                     VFIO_FEATURE_ENABLE_REQ_BIT, true),
2857     DEFINE_PROP_BIT("x-igd-opregion", VFIOPCIDevice, features,
2858                     VFIO_FEATURE_ENABLE_IGD_OPREGION_BIT, false),
2859     DEFINE_PROP_BOOL("x-no-mmap", VFIOPCIDevice, vbasedev.no_mmap, false),
2860     DEFINE_PROP_BOOL("x-no-kvm-intx", VFIOPCIDevice, no_kvm_intx, false),
2861     DEFINE_PROP_BOOL("x-no-kvm-msi", VFIOPCIDevice, no_kvm_msi, false),
2862     DEFINE_PROP_BOOL("x-no-kvm-msix", VFIOPCIDevice, no_kvm_msix, false),
2863     DEFINE_PROP_UINT32("x-pci-vendor-id", VFIOPCIDevice, vendor_id, PCI_ANY_ID),
2864     DEFINE_PROP_UINT32("x-pci-device-id", VFIOPCIDevice, device_id, PCI_ANY_ID),
2865     DEFINE_PROP_UINT32("x-pci-sub-vendor-id", VFIOPCIDevice,
2866                        sub_vendor_id, PCI_ANY_ID),
2867     DEFINE_PROP_UINT32("x-pci-sub-device-id", VFIOPCIDevice,
2868                        sub_device_id, PCI_ANY_ID),
2869     DEFINE_PROP_UINT32("x-igd-gms", VFIOPCIDevice, igd_gms, 0),
2870     /*
2871      * TODO - support passed fds... is this necessary?
2872      * DEFINE_PROP_STRING("vfiofd", VFIOPCIDevice, vfiofd_name),
2873      * DEFINE_PROP_STRING("vfiogroupfd, VFIOPCIDevice, vfiogroupfd_name),
2874      */
2875     DEFINE_PROP_END_OF_LIST(),
2876 };
2877 
2878 static const VMStateDescription vfio_pci_vmstate = {
2879     .name = "vfio-pci",
2880     .unmigratable = 1,
2881 };
2882 
2883 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
2884 {
2885     DeviceClass *dc = DEVICE_CLASS(klass);
2886     PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
2887 
2888     dc->reset = vfio_pci_reset;
2889     dc->props = vfio_pci_dev_properties;
2890     dc->vmsd = &vfio_pci_vmstate;
2891     dc->desc = "VFIO-based PCI device assignment";
2892     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
2893     pdc->realize = vfio_realize;
2894     pdc->exit = vfio_exitfn;
2895     pdc->config_read = vfio_pci_read_config;
2896     pdc->config_write = vfio_pci_write_config;
2897     pdc->is_express = 1; /* We might be */
2898 }
2899 
2900 static const TypeInfo vfio_pci_dev_info = {
2901     .name = "vfio-pci",
2902     .parent = TYPE_PCI_DEVICE,
2903     .instance_size = sizeof(VFIOPCIDevice),
2904     .class_init = vfio_pci_dev_class_init,
2905     .instance_init = vfio_instance_init,
2906     .instance_finalize = vfio_instance_finalize,
2907 };
2908 
2909 static void register_vfio_pci_dev_type(void)
2910 {
2911     type_register_static(&vfio_pci_dev_info);
2912 }
2913 
2914 type_init(register_vfio_pci_dev_type)
2915