xref: /openbmc/qemu/hw/vfio/pci.c (revision c11b0583)
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 <linux/vfio.h>
22 #include <sys/ioctl.h>
23 #include <sys/mman.h>
24 #include <sys/stat.h>
25 #include <sys/types.h>
26 #include <unistd.h>
27 
28 #include "config.h"
29 #include "exec/address-spaces.h"
30 #include "exec/memory.h"
31 #include "hw/pci/msi.h"
32 #include "hw/pci/msix.h"
33 #include "hw/pci/pci.h"
34 #include "qemu-common.h"
35 #include "qemu/error-report.h"
36 #include "qemu/event_notifier.h"
37 #include "qemu/queue.h"
38 #include "qemu/range.h"
39 #include "sysemu/kvm.h"
40 #include "sysemu/sysemu.h"
41 #include "trace.h"
42 #include "hw/vfio/vfio.h"
43 #include "hw/vfio/vfio-common.h"
44 
45 struct VFIOPCIDevice;
46 
47 typedef struct VFIOQuirk {
48     MemoryRegion mem;
49     struct VFIOPCIDevice *vdev;
50     QLIST_ENTRY(VFIOQuirk) next;
51     struct {
52         uint32_t base_offset:TARGET_PAGE_BITS;
53         uint32_t address_offset:TARGET_PAGE_BITS;
54         uint32_t address_size:3;
55         uint32_t bar:3;
56 
57         uint32_t address_match;
58         uint32_t address_mask;
59 
60         uint32_t address_val:TARGET_PAGE_BITS;
61         uint32_t data_offset:TARGET_PAGE_BITS;
62         uint32_t data_size:3;
63 
64         uint8_t flags;
65         uint8_t read_flags;
66         uint8_t write_flags;
67     } data;
68 } VFIOQuirk;
69 
70 typedef struct VFIOBAR {
71     VFIORegion region;
72     bool ioport;
73     bool mem64;
74     QLIST_HEAD(, VFIOQuirk) quirks;
75 } VFIOBAR;
76 
77 typedef struct VFIOVGARegion {
78     MemoryRegion mem;
79     off_t offset;
80     int nr;
81     QLIST_HEAD(, VFIOQuirk) quirks;
82 } VFIOVGARegion;
83 
84 typedef struct VFIOVGA {
85     off_t fd_offset;
86     int fd;
87     VFIOVGARegion region[QEMU_PCI_VGA_NUM_REGIONS];
88 } VFIOVGA;
89 
90 typedef struct VFIOINTx {
91     bool pending; /* interrupt pending */
92     bool kvm_accel; /* set when QEMU bypass through KVM enabled */
93     uint8_t pin; /* which pin to pull for qemu_set_irq */
94     EventNotifier interrupt; /* eventfd triggered on interrupt */
95     EventNotifier unmask; /* eventfd for unmask on QEMU bypass */
96     PCIINTxRoute route; /* routing info for QEMU bypass */
97     uint32_t mmap_timeout; /* delay to re-enable mmaps after interrupt */
98     QEMUTimer *mmap_timer; /* enable mmaps after periods w/o interrupts */
99 } VFIOINTx;
100 
101 typedef struct VFIOMSIVector {
102     /*
103      * Two interrupt paths are configured per vector.  The first, is only used
104      * for interrupts injected via QEMU.  This is typically the non-accel path,
105      * but may also be used when we want QEMU to handle masking and pending
106      * bits.  The KVM path bypasses QEMU and is therefore higher performance,
107      * but requires masking at the device.  virq is used to track the MSI route
108      * through KVM, thus kvm_interrupt is only available when virq is set to a
109      * valid (>= 0) value.
110      */
111     EventNotifier interrupt;
112     EventNotifier kvm_interrupt;
113     struct VFIOPCIDevice *vdev; /* back pointer to device */
114     int virq;
115     bool use;
116 } VFIOMSIVector;
117 
118 enum {
119     VFIO_INT_NONE = 0,
120     VFIO_INT_INTx = 1,
121     VFIO_INT_MSI  = 2,
122     VFIO_INT_MSIX = 3,
123 };
124 
125 /* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
126 typedef struct VFIOMSIXInfo {
127     uint8_t table_bar;
128     uint8_t pba_bar;
129     uint16_t entries;
130     uint32_t table_offset;
131     uint32_t pba_offset;
132     MemoryRegion mmap_mem;
133     void *mmap;
134 } VFIOMSIXInfo;
135 
136 typedef struct VFIOPCIDevice {
137     PCIDevice pdev;
138     VFIODevice vbasedev;
139     VFIOINTx intx;
140     unsigned int config_size;
141     uint8_t *emulated_config_bits; /* QEMU emulated bits, little-endian */
142     off_t config_offset; /* Offset of config space region within device fd */
143     unsigned int rom_size;
144     off_t rom_offset; /* Offset of ROM region within device fd */
145     void *rom;
146     int msi_cap_size;
147     VFIOMSIVector *msi_vectors;
148     VFIOMSIXInfo *msix;
149     int nr_vectors; /* Number of MSI/MSIX vectors currently in use */
150     int interrupt; /* Current interrupt type */
151     VFIOBAR bars[PCI_NUM_REGIONS - 1]; /* No ROM */
152     VFIOVGA vga; /* 0xa0000, 0x3b0, 0x3c0 */
153     PCIHostDeviceAddress host;
154     EventNotifier err_notifier;
155     EventNotifier req_notifier;
156     int (*resetfn)(struct VFIOPCIDevice *);
157     uint32_t features;
158 #define VFIO_FEATURE_ENABLE_VGA_BIT 0
159 #define VFIO_FEATURE_ENABLE_VGA (1 << VFIO_FEATURE_ENABLE_VGA_BIT)
160 #define VFIO_FEATURE_ENABLE_REQ_BIT 1
161 #define VFIO_FEATURE_ENABLE_REQ (1 << VFIO_FEATURE_ENABLE_REQ_BIT)
162     int32_t bootindex;
163     uint8_t pm_cap;
164     bool has_vga;
165     bool pci_aer;
166     bool req_enabled;
167     bool has_flr;
168     bool has_pm_reset;
169     bool rom_read_failed;
170 } VFIOPCIDevice;
171 
172 typedef struct VFIORomBlacklistEntry {
173     uint16_t vendor_id;
174     uint16_t device_id;
175 } VFIORomBlacklistEntry;
176 
177 /*
178  * List of device ids/vendor ids for which to disable
179  * option rom loading. This avoids the guest hangs during rom
180  * execution as noticed with the BCM 57810 card for lack of a
181  * more better way to handle such issues.
182  * The  user can still override by specifying a romfile or
183  * rombar=1.
184  * Please see https://bugs.launchpad.net/qemu/+bug/1284874
185  * for an analysis of the 57810 card hang. When adding
186  * a new vendor id/device id combination below, please also add
187  * your card/environment details and information that could
188  * help in debugging to the bug tracking this issue
189  */
190 static const VFIORomBlacklistEntry romblacklist[] = {
191     /* Broadcom BCM 57810 */
192     { 0x14e4, 0x168e }
193 };
194 
195 #define MSIX_CAP_LENGTH 12
196 
197 static void vfio_disable_interrupts(VFIOPCIDevice *vdev);
198 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len);
199 static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
200                                   uint32_t val, int len);
201 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled);
202 
203 /*
204  * Disabling BAR mmaping can be slow, but toggling it around INTx can
205  * also be a huge overhead.  We try to get the best of both worlds by
206  * waiting until an interrupt to disable mmaps (subsequent transitions
207  * to the same state are effectively no overhead).  If the interrupt has
208  * been serviced and the time gap is long enough, we re-enable mmaps for
209  * performance.  This works well for things like graphics cards, which
210  * may not use their interrupt at all and are penalized to an unusable
211  * level by read/write BAR traps.  Other devices, like NICs, have more
212  * regular interrupts and see much better latency by staying in non-mmap
213  * mode.  We therefore set the default mmap_timeout such that a ping
214  * is just enough to keep the mmap disabled.  Users can experiment with
215  * other options with the x-intx-mmap-timeout-ms parameter (a value of
216  * zero disables the timer).
217  */
218 static void vfio_intx_mmap_enable(void *opaque)
219 {
220     VFIOPCIDevice *vdev = opaque;
221 
222     if (vdev->intx.pending) {
223         timer_mod(vdev->intx.mmap_timer,
224                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
225         return;
226     }
227 
228     vfio_mmap_set_enabled(vdev, true);
229 }
230 
231 static void vfio_intx_interrupt(void *opaque)
232 {
233     VFIOPCIDevice *vdev = opaque;
234 
235     if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
236         return;
237     }
238 
239     trace_vfio_intx_interrupt(vdev->vbasedev.name, 'A' + vdev->intx.pin);
240 
241     vdev->intx.pending = true;
242     pci_irq_assert(&vdev->pdev);
243     vfio_mmap_set_enabled(vdev, false);
244     if (vdev->intx.mmap_timeout) {
245         timer_mod(vdev->intx.mmap_timer,
246                        qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
247     }
248 }
249 
250 static void vfio_eoi(VFIODevice *vbasedev)
251 {
252     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
253 
254     if (!vdev->intx.pending) {
255         return;
256     }
257 
258     trace_vfio_eoi(vbasedev->name);
259 
260     vdev->intx.pending = false;
261     pci_irq_deassert(&vdev->pdev);
262     vfio_unmask_single_irqindex(vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
263 }
264 
265 static void vfio_enable_intx_kvm(VFIOPCIDevice *vdev)
266 {
267 #ifdef CONFIG_KVM
268     struct kvm_irqfd irqfd = {
269         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
270         .gsi = vdev->intx.route.irq,
271         .flags = KVM_IRQFD_FLAG_RESAMPLE,
272     };
273     struct vfio_irq_set *irq_set;
274     int ret, argsz;
275     int32_t *pfd;
276 
277     if (!VFIO_ALLOW_KVM_INTX || !kvm_irqfds_enabled() ||
278         vdev->intx.route.mode != PCI_INTX_ENABLED ||
279         !kvm_resamplefds_enabled()) {
280         return;
281     }
282 
283     /* Get to a known interrupt state */
284     qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
285     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
286     vdev->intx.pending = false;
287     pci_irq_deassert(&vdev->pdev);
288 
289     /* Get an eventfd for resample/unmask */
290     if (event_notifier_init(&vdev->intx.unmask, 0)) {
291         error_report("vfio: Error: event_notifier_init failed eoi");
292         goto fail;
293     }
294 
295     /* KVM triggers it, VFIO listens for it */
296     irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
297 
298     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
299         error_report("vfio: Error: Failed to setup resample irqfd: %m");
300         goto fail_irqfd;
301     }
302 
303     argsz = sizeof(*irq_set) + sizeof(*pfd);
304 
305     irq_set = g_malloc0(argsz);
306     irq_set->argsz = argsz;
307     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
308     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
309     irq_set->start = 0;
310     irq_set->count = 1;
311     pfd = (int32_t *)&irq_set->data;
312 
313     *pfd = irqfd.resamplefd;
314 
315     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
316     g_free(irq_set);
317     if (ret) {
318         error_report("vfio: Error: Failed to setup INTx unmask fd: %m");
319         goto fail_vfio;
320     }
321 
322     /* Let'em rip */
323     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
324 
325     vdev->intx.kvm_accel = true;
326 
327     trace_vfio_enable_intx_kvm(vdev->vbasedev.name);
328 
329     return;
330 
331 fail_vfio:
332     irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
333     kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
334 fail_irqfd:
335     event_notifier_cleanup(&vdev->intx.unmask);
336 fail:
337     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
338     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
339 #endif
340 }
341 
342 static void vfio_disable_intx_kvm(VFIOPCIDevice *vdev)
343 {
344 #ifdef CONFIG_KVM
345     struct kvm_irqfd irqfd = {
346         .fd = event_notifier_get_fd(&vdev->intx.interrupt),
347         .gsi = vdev->intx.route.irq,
348         .flags = KVM_IRQFD_FLAG_DEASSIGN,
349     };
350 
351     if (!vdev->intx.kvm_accel) {
352         return;
353     }
354 
355     /*
356      * Get to a known state, hardware masked, QEMU ready to accept new
357      * interrupts, QEMU IRQ de-asserted.
358      */
359     vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
360     vdev->intx.pending = false;
361     pci_irq_deassert(&vdev->pdev);
362 
363     /* Tell KVM to stop listening for an INTx irqfd */
364     if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
365         error_report("vfio: Error: Failed to disable INTx irqfd: %m");
366     }
367 
368     /* We only need to close the eventfd for VFIO to cleanup the kernel side */
369     event_notifier_cleanup(&vdev->intx.unmask);
370 
371     /* QEMU starts listening for interrupt events. */
372     qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
373 
374     vdev->intx.kvm_accel = false;
375 
376     /* If we've missed an event, let it re-fire through QEMU */
377     vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
378 
379     trace_vfio_disable_intx_kvm(vdev->vbasedev.name);
380 #endif
381 }
382 
383 static void vfio_update_irq(PCIDevice *pdev)
384 {
385     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
386     PCIINTxRoute route;
387 
388     if (vdev->interrupt != VFIO_INT_INTx) {
389         return;
390     }
391 
392     route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
393 
394     if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
395         return; /* Nothing changed */
396     }
397 
398     trace_vfio_update_irq(vdev->vbasedev.name,
399                           vdev->intx.route.irq, route.irq);
400 
401     vfio_disable_intx_kvm(vdev);
402 
403     vdev->intx.route = route;
404 
405     if (route.mode != PCI_INTX_ENABLED) {
406         return;
407     }
408 
409     vfio_enable_intx_kvm(vdev);
410 
411     /* Re-enable the interrupt in cased we missed an EOI */
412     vfio_eoi(&vdev->vbasedev);
413 }
414 
415 static int vfio_enable_intx(VFIOPCIDevice *vdev)
416 {
417     uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
418     int ret, argsz;
419     struct vfio_irq_set *irq_set;
420     int32_t *pfd;
421 
422     if (!pin) {
423         return 0;
424     }
425 
426     vfio_disable_interrupts(vdev);
427 
428     vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
429     pci_config_set_interrupt_pin(vdev->pdev.config, pin);
430 
431 #ifdef CONFIG_KVM
432     /*
433      * Only conditional to avoid generating error messages on platforms
434      * where we won't actually use the result anyway.
435      */
436     if (kvm_irqfds_enabled() && kvm_resamplefds_enabled()) {
437         vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
438                                                         vdev->intx.pin);
439     }
440 #endif
441 
442     ret = event_notifier_init(&vdev->intx.interrupt, 0);
443     if (ret) {
444         error_report("vfio: Error: event_notifier_init failed");
445         return ret;
446     }
447 
448     argsz = sizeof(*irq_set) + sizeof(*pfd);
449 
450     irq_set = g_malloc0(argsz);
451     irq_set->argsz = argsz;
452     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
453     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
454     irq_set->start = 0;
455     irq_set->count = 1;
456     pfd = (int32_t *)&irq_set->data;
457 
458     *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
459     qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
460 
461     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
462     g_free(irq_set);
463     if (ret) {
464         error_report("vfio: Error: Failed to setup INTx fd: %m");
465         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
466         event_notifier_cleanup(&vdev->intx.interrupt);
467         return -errno;
468     }
469 
470     vfio_enable_intx_kvm(vdev);
471 
472     vdev->interrupt = VFIO_INT_INTx;
473 
474     trace_vfio_enable_intx(vdev->vbasedev.name);
475 
476     return 0;
477 }
478 
479 static void vfio_disable_intx(VFIOPCIDevice *vdev)
480 {
481     int fd;
482 
483     timer_del(vdev->intx.mmap_timer);
484     vfio_disable_intx_kvm(vdev);
485     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX);
486     vdev->intx.pending = false;
487     pci_irq_deassert(&vdev->pdev);
488     vfio_mmap_set_enabled(vdev, true);
489 
490     fd = event_notifier_get_fd(&vdev->intx.interrupt);
491     qemu_set_fd_handler(fd, NULL, NULL, vdev);
492     event_notifier_cleanup(&vdev->intx.interrupt);
493 
494     vdev->interrupt = VFIO_INT_NONE;
495 
496     trace_vfio_disable_intx(vdev->vbasedev.name);
497 }
498 
499 /*
500  * MSI/X
501  */
502 static void vfio_msi_interrupt(void *opaque)
503 {
504     VFIOMSIVector *vector = opaque;
505     VFIOPCIDevice *vdev = vector->vdev;
506     int nr = vector - vdev->msi_vectors;
507 
508     if (!event_notifier_test_and_clear(&vector->interrupt)) {
509         return;
510     }
511 
512 #ifdef DEBUG_VFIO
513     MSIMessage msg;
514 
515     if (vdev->interrupt == VFIO_INT_MSIX) {
516         msg = msix_get_message(&vdev->pdev, nr);
517     } else if (vdev->interrupt == VFIO_INT_MSI) {
518         msg = msi_get_message(&vdev->pdev, nr);
519     } else {
520         abort();
521     }
522 
523     trace_vfio_msi_interrupt(vdev->vbasedev.name, nr, msg.address, msg.data);
524 #endif
525 
526     if (vdev->interrupt == VFIO_INT_MSIX) {
527         msix_notify(&vdev->pdev, nr);
528     } else if (vdev->interrupt == VFIO_INT_MSI) {
529         msi_notify(&vdev->pdev, nr);
530     } else {
531         error_report("vfio: MSI interrupt receieved, but not enabled?");
532     }
533 }
534 
535 static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool msix)
536 {
537     struct vfio_irq_set *irq_set;
538     int ret = 0, i, argsz;
539     int32_t *fds;
540 
541     argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
542 
543     irq_set = g_malloc0(argsz);
544     irq_set->argsz = argsz;
545     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
546     irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
547     irq_set->start = 0;
548     irq_set->count = vdev->nr_vectors;
549     fds = (int32_t *)&irq_set->data;
550 
551     for (i = 0; i < vdev->nr_vectors; i++) {
552         int fd = -1;
553 
554         /*
555          * MSI vs MSI-X - The guest has direct access to MSI mask and pending
556          * bits, therefore we always use the KVM signaling path when setup.
557          * MSI-X mask and pending bits are emulated, so we want to use the
558          * KVM signaling path only when configured and unmasked.
559          */
560         if (vdev->msi_vectors[i].use) {
561             if (vdev->msi_vectors[i].virq < 0 ||
562                 (msix && msix_is_masked(&vdev->pdev, i))) {
563                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
564             } else {
565                 fd = event_notifier_get_fd(&vdev->msi_vectors[i].kvm_interrupt);
566             }
567         }
568 
569         fds[i] = fd;
570     }
571 
572     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
573 
574     g_free(irq_set);
575 
576     return ret;
577 }
578 
579 static void vfio_add_kvm_msi_virq(VFIOMSIVector *vector, MSIMessage *msg,
580                                   bool msix)
581 {
582     int virq;
583 
584     if ((msix && !VFIO_ALLOW_KVM_MSIX) ||
585         (!msix && !VFIO_ALLOW_KVM_MSI) || !msg) {
586         return;
587     }
588 
589     if (event_notifier_init(&vector->kvm_interrupt, 0)) {
590         return;
591     }
592 
593     virq = kvm_irqchip_add_msi_route(kvm_state, *msg);
594     if (virq < 0) {
595         event_notifier_cleanup(&vector->kvm_interrupt);
596         return;
597     }
598 
599     if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
600                                        NULL, virq) < 0) {
601         kvm_irqchip_release_virq(kvm_state, virq);
602         event_notifier_cleanup(&vector->kvm_interrupt);
603         return;
604     }
605 
606     vector->virq = virq;
607 }
608 
609 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector)
610 {
611     kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt,
612                                           vector->virq);
613     kvm_irqchip_release_virq(kvm_state, vector->virq);
614     vector->virq = -1;
615     event_notifier_cleanup(&vector->kvm_interrupt);
616 }
617 
618 static void vfio_update_kvm_msi_virq(VFIOMSIVector *vector, MSIMessage msg)
619 {
620     kvm_irqchip_update_msi_route(kvm_state, vector->virq, msg);
621 }
622 
623 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
624                                    MSIMessage *msg, IOHandler *handler)
625 {
626     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
627     VFIOMSIVector *vector;
628     int ret;
629 
630     trace_vfio_msix_vector_do_use(vdev->vbasedev.name, nr);
631 
632     vector = &vdev->msi_vectors[nr];
633 
634     if (!vector->use) {
635         vector->vdev = vdev;
636         vector->virq = -1;
637         if (event_notifier_init(&vector->interrupt, 0)) {
638             error_report("vfio: Error: event_notifier_init failed");
639         }
640         vector->use = true;
641         msix_vector_use(pdev, nr);
642     }
643 
644     qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
645                         handler, NULL, vector);
646 
647     /*
648      * Attempt to enable route through KVM irqchip,
649      * default to userspace handling if unavailable.
650      */
651     if (vector->virq >= 0) {
652         if (!msg) {
653             vfio_remove_kvm_msi_virq(vector);
654         } else {
655             vfio_update_kvm_msi_virq(vector, *msg);
656         }
657     } else {
658         vfio_add_kvm_msi_virq(vector, msg, true);
659     }
660 
661     /*
662      * We don't want to have the host allocate all possible MSI vectors
663      * for a device if they're not in use, so we shutdown and incrementally
664      * increase them as needed.
665      */
666     if (vdev->nr_vectors < nr + 1) {
667         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
668         vdev->nr_vectors = nr + 1;
669         ret = vfio_enable_vectors(vdev, true);
670         if (ret) {
671             error_report("vfio: failed to enable vectors, %d", ret);
672         }
673     } else {
674         int argsz;
675         struct vfio_irq_set *irq_set;
676         int32_t *pfd;
677 
678         argsz = sizeof(*irq_set) + sizeof(*pfd);
679 
680         irq_set = g_malloc0(argsz);
681         irq_set->argsz = argsz;
682         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
683                          VFIO_IRQ_SET_ACTION_TRIGGER;
684         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
685         irq_set->start = nr;
686         irq_set->count = 1;
687         pfd = (int32_t *)&irq_set->data;
688 
689         if (vector->virq >= 0) {
690             *pfd = event_notifier_get_fd(&vector->kvm_interrupt);
691         } else {
692             *pfd = event_notifier_get_fd(&vector->interrupt);
693         }
694 
695         ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
696         g_free(irq_set);
697         if (ret) {
698             error_report("vfio: failed to modify vector, %d", ret);
699         }
700     }
701 
702     return 0;
703 }
704 
705 static int vfio_msix_vector_use(PCIDevice *pdev,
706                                 unsigned int nr, MSIMessage msg)
707 {
708     return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
709 }
710 
711 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
712 {
713     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
714     VFIOMSIVector *vector = &vdev->msi_vectors[nr];
715 
716     trace_vfio_msix_vector_release(vdev->vbasedev.name, nr);
717 
718     /*
719      * There are still old guests that mask and unmask vectors on every
720      * interrupt.  If we're using QEMU bypass with a KVM irqfd, leave all of
721      * the KVM setup in place, simply switch VFIO to use the non-bypass
722      * eventfd.  We'll then fire the interrupt through QEMU and the MSI-X
723      * core will mask the interrupt and set pending bits, allowing it to
724      * be re-asserted on unmask.  Nothing to do if already using QEMU mode.
725      */
726     if (vector->virq >= 0) {
727         int argsz;
728         struct vfio_irq_set *irq_set;
729         int32_t *pfd;
730 
731         argsz = sizeof(*irq_set) + sizeof(*pfd);
732 
733         irq_set = g_malloc0(argsz);
734         irq_set->argsz = argsz;
735         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
736                          VFIO_IRQ_SET_ACTION_TRIGGER;
737         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
738         irq_set->start = nr;
739         irq_set->count = 1;
740         pfd = (int32_t *)&irq_set->data;
741 
742         *pfd = event_notifier_get_fd(&vector->interrupt);
743 
744         ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
745 
746         g_free(irq_set);
747     }
748 }
749 
750 static void vfio_enable_msix(VFIOPCIDevice *vdev)
751 {
752     vfio_disable_interrupts(vdev);
753 
754     vdev->msi_vectors = g_malloc0(vdev->msix->entries * sizeof(VFIOMSIVector));
755 
756     vdev->interrupt = VFIO_INT_MSIX;
757 
758     /*
759      * Some communication channels between VF & PF or PF & fw rely on the
760      * physical state of the device and expect that enabling MSI-X from the
761      * guest enables the same on the host.  When our guest is Linux, the
762      * guest driver call to pci_enable_msix() sets the enabling bit in the
763      * MSI-X capability, but leaves the vector table masked.  We therefore
764      * can't rely on a vector_use callback (from request_irq() in the guest)
765      * to switch the physical device into MSI-X mode because that may come a
766      * long time after pci_enable_msix().  This code enables vector 0 with
767      * triggering to userspace, then immediately release the vector, leaving
768      * the physical device with no vectors enabled, but MSI-X enabled, just
769      * like the guest view.
770      */
771     vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
772     vfio_msix_vector_release(&vdev->pdev, 0);
773 
774     if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
775                                   vfio_msix_vector_release, NULL)) {
776         error_report("vfio: msix_set_vector_notifiers failed");
777     }
778 
779     trace_vfio_enable_msix(vdev->vbasedev.name);
780 }
781 
782 static void vfio_enable_msi(VFIOPCIDevice *vdev)
783 {
784     int ret, i;
785 
786     vfio_disable_interrupts(vdev);
787 
788     vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
789 retry:
790     vdev->msi_vectors = g_malloc0(vdev->nr_vectors * sizeof(VFIOMSIVector));
791 
792     for (i = 0; i < vdev->nr_vectors; i++) {
793         VFIOMSIVector *vector = &vdev->msi_vectors[i];
794         MSIMessage msg = msi_get_message(&vdev->pdev, i);
795 
796         vector->vdev = vdev;
797         vector->virq = -1;
798         vector->use = true;
799 
800         if (event_notifier_init(&vector->interrupt, 0)) {
801             error_report("vfio: Error: event_notifier_init failed");
802         }
803 
804         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
805                             vfio_msi_interrupt, NULL, vector);
806 
807         /*
808          * Attempt to enable route through KVM irqchip,
809          * default to userspace handling if unavailable.
810          */
811         vfio_add_kvm_msi_virq(vector, &msg, false);
812     }
813 
814     /* Set interrupt type prior to possible interrupts */
815     vdev->interrupt = VFIO_INT_MSI;
816 
817     ret = vfio_enable_vectors(vdev, false);
818     if (ret) {
819         if (ret < 0) {
820             error_report("vfio: Error: Failed to setup MSI fds: %m");
821         } else if (ret != vdev->nr_vectors) {
822             error_report("vfio: Error: Failed to enable %d "
823                          "MSI vectors, retry with %d", vdev->nr_vectors, ret);
824         }
825 
826         for (i = 0; i < vdev->nr_vectors; i++) {
827             VFIOMSIVector *vector = &vdev->msi_vectors[i];
828             if (vector->virq >= 0) {
829                 vfio_remove_kvm_msi_virq(vector);
830             }
831             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
832                                 NULL, NULL, NULL);
833             event_notifier_cleanup(&vector->interrupt);
834         }
835 
836         g_free(vdev->msi_vectors);
837 
838         if (ret > 0 && ret != vdev->nr_vectors) {
839             vdev->nr_vectors = ret;
840             goto retry;
841         }
842         vdev->nr_vectors = 0;
843 
844         /*
845          * Failing to setup MSI doesn't really fall within any specification.
846          * Let's try leaving interrupts disabled and hope the guest figures
847          * out to fall back to INTx for this device.
848          */
849         error_report("vfio: Error: Failed to enable MSI");
850         vdev->interrupt = VFIO_INT_NONE;
851 
852         return;
853     }
854 
855     trace_vfio_enable_msi(vdev->vbasedev.name, vdev->nr_vectors);
856 }
857 
858 static void vfio_disable_msi_common(VFIOPCIDevice *vdev)
859 {
860     int i;
861 
862     for (i = 0; i < vdev->nr_vectors; i++) {
863         VFIOMSIVector *vector = &vdev->msi_vectors[i];
864         if (vdev->msi_vectors[i].use) {
865             if (vector->virq >= 0) {
866                 vfio_remove_kvm_msi_virq(vector);
867             }
868             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
869                                 NULL, NULL, NULL);
870             event_notifier_cleanup(&vector->interrupt);
871         }
872     }
873 
874     g_free(vdev->msi_vectors);
875     vdev->msi_vectors = NULL;
876     vdev->nr_vectors = 0;
877     vdev->interrupt = VFIO_INT_NONE;
878 
879     vfio_enable_intx(vdev);
880 }
881 
882 static void vfio_disable_msix(VFIOPCIDevice *vdev)
883 {
884     int i;
885 
886     msix_unset_vector_notifiers(&vdev->pdev);
887 
888     /*
889      * MSI-X will only release vectors if MSI-X is still enabled on the
890      * device, check through the rest and release it ourselves if necessary.
891      */
892     for (i = 0; i < vdev->nr_vectors; i++) {
893         if (vdev->msi_vectors[i].use) {
894             vfio_msix_vector_release(&vdev->pdev, i);
895             msix_vector_unuse(&vdev->pdev, i);
896         }
897     }
898 
899     if (vdev->nr_vectors) {
900         vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX);
901     }
902 
903     vfio_disable_msi_common(vdev);
904 
905     trace_vfio_disable_msix(vdev->vbasedev.name);
906 }
907 
908 static void vfio_disable_msi(VFIOPCIDevice *vdev)
909 {
910     vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSI_IRQ_INDEX);
911     vfio_disable_msi_common(vdev);
912 
913     trace_vfio_disable_msi(vdev->vbasedev.name);
914 }
915 
916 static void vfio_update_msi(VFIOPCIDevice *vdev)
917 {
918     int i;
919 
920     for (i = 0; i < vdev->nr_vectors; i++) {
921         VFIOMSIVector *vector = &vdev->msi_vectors[i];
922         MSIMessage msg;
923 
924         if (!vector->use || vector->virq < 0) {
925             continue;
926         }
927 
928         msg = msi_get_message(&vdev->pdev, i);
929         vfio_update_kvm_msi_virq(vector, msg);
930     }
931 }
932 
933 static void vfio_pci_load_rom(VFIOPCIDevice *vdev)
934 {
935     struct vfio_region_info reg_info = {
936         .argsz = sizeof(reg_info),
937         .index = VFIO_PCI_ROM_REGION_INDEX
938     };
939     uint64_t size;
940     off_t off = 0;
941     ssize_t bytes;
942 
943     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info)) {
944         error_report("vfio: Error getting ROM info: %m");
945         return;
946     }
947 
948     trace_vfio_pci_load_rom(vdev->vbasedev.name, (unsigned long)reg_info.size,
949                             (unsigned long)reg_info.offset,
950                             (unsigned long)reg_info.flags);
951 
952     vdev->rom_size = size = reg_info.size;
953     vdev->rom_offset = reg_info.offset;
954 
955     if (!vdev->rom_size) {
956         vdev->rom_read_failed = true;
957         error_report("vfio-pci: Cannot read device rom at "
958                     "%s", vdev->vbasedev.name);
959         error_printf("Device option ROM contents are probably invalid "
960                     "(check dmesg).\nSkip option ROM probe with rombar=0, "
961                     "or load from file with romfile=\n");
962         return;
963     }
964 
965     vdev->rom = g_malloc(size);
966     memset(vdev->rom, 0xff, size);
967 
968     while (size) {
969         bytes = pread(vdev->vbasedev.fd, vdev->rom + off,
970                       size, vdev->rom_offset + off);
971         if (bytes == 0) {
972             break;
973         } else if (bytes > 0) {
974             off += bytes;
975             size -= bytes;
976         } else {
977             if (errno == EINTR || errno == EAGAIN) {
978                 continue;
979             }
980             error_report("vfio: Error reading device ROM: %m");
981             break;
982         }
983     }
984 }
985 
986 static uint64_t vfio_rom_read(void *opaque, hwaddr addr, unsigned size)
987 {
988     VFIOPCIDevice *vdev = opaque;
989     union {
990         uint8_t byte;
991         uint16_t word;
992         uint32_t dword;
993         uint64_t qword;
994     } val;
995     uint64_t data = 0;
996 
997     /* Load the ROM lazily when the guest tries to read it */
998     if (unlikely(!vdev->rom && !vdev->rom_read_failed)) {
999         vfio_pci_load_rom(vdev);
1000     }
1001 
1002     memcpy(&val, vdev->rom + addr,
1003            (addr < vdev->rom_size) ? MIN(size, vdev->rom_size - addr) : 0);
1004 
1005     switch (size) {
1006     case 1:
1007         data = val.byte;
1008         break;
1009     case 2:
1010         data = le16_to_cpu(val.word);
1011         break;
1012     case 4:
1013         data = le32_to_cpu(val.dword);
1014         break;
1015     default:
1016         hw_error("vfio: unsupported read size, %d bytes\n", size);
1017         break;
1018     }
1019 
1020     trace_vfio_rom_read(vdev->vbasedev.name, addr, size, data);
1021 
1022     return data;
1023 }
1024 
1025 static void vfio_rom_write(void *opaque, hwaddr addr,
1026                            uint64_t data, unsigned size)
1027 {
1028 }
1029 
1030 static const MemoryRegionOps vfio_rom_ops = {
1031     .read = vfio_rom_read,
1032     .write = vfio_rom_write,
1033     .endianness = DEVICE_LITTLE_ENDIAN,
1034 };
1035 
1036 static bool vfio_blacklist_opt_rom(VFIOPCIDevice *vdev)
1037 {
1038     PCIDevice *pdev = &vdev->pdev;
1039     uint16_t vendor_id, device_id;
1040     int count = 0;
1041 
1042     vendor_id = pci_get_word(pdev->config + PCI_VENDOR_ID);
1043     device_id = pci_get_word(pdev->config + PCI_DEVICE_ID);
1044 
1045     while (count < ARRAY_SIZE(romblacklist)) {
1046         if (romblacklist[count].vendor_id == vendor_id &&
1047             romblacklist[count].device_id == device_id) {
1048                 return true;
1049         }
1050         count++;
1051     }
1052 
1053     return false;
1054 }
1055 
1056 static void vfio_pci_size_rom(VFIOPCIDevice *vdev)
1057 {
1058     uint32_t orig, size = cpu_to_le32((uint32_t)PCI_ROM_ADDRESS_MASK);
1059     off_t offset = vdev->config_offset + PCI_ROM_ADDRESS;
1060     DeviceState *dev = DEVICE(vdev);
1061     char name[32];
1062     int fd = vdev->vbasedev.fd;
1063 
1064     if (vdev->pdev.romfile || !vdev->pdev.rom_bar) {
1065         /* Since pci handles romfile, just print a message and return */
1066         if (vfio_blacklist_opt_rom(vdev) && vdev->pdev.romfile) {
1067             error_printf("Warning : Device at %04x:%02x:%02x.%x "
1068                          "is known to cause system instability issues during "
1069                          "option rom execution. "
1070                          "Proceeding anyway since user specified romfile\n",
1071                          vdev->host.domain, vdev->host.bus, vdev->host.slot,
1072                          vdev->host.function);
1073         }
1074         return;
1075     }
1076 
1077     /*
1078      * Use the same size ROM BAR as the physical device.  The contents
1079      * will get filled in later when the guest tries to read it.
1080      */
1081     if (pread(fd, &orig, 4, offset) != 4 ||
1082         pwrite(fd, &size, 4, offset) != 4 ||
1083         pread(fd, &size, 4, offset) != 4 ||
1084         pwrite(fd, &orig, 4, offset) != 4) {
1085         error_report("%s(%04x:%02x:%02x.%x) failed: %m",
1086                      __func__, vdev->host.domain, vdev->host.bus,
1087                      vdev->host.slot, vdev->host.function);
1088         return;
1089     }
1090 
1091     size = ~(le32_to_cpu(size) & PCI_ROM_ADDRESS_MASK) + 1;
1092 
1093     if (!size) {
1094         return;
1095     }
1096 
1097     if (vfio_blacklist_opt_rom(vdev)) {
1098         if (dev->opts && qemu_opt_get(dev->opts, "rombar")) {
1099             error_printf("Warning : Device at %04x:%02x:%02x.%x "
1100                          "is known to cause system instability issues during "
1101                          "option rom execution. "
1102                          "Proceeding anyway since user specified non zero value for "
1103                          "rombar\n",
1104                          vdev->host.domain, vdev->host.bus, vdev->host.slot,
1105                          vdev->host.function);
1106         } else {
1107             error_printf("Warning : Rom loading for device at "
1108                          "%04x:%02x:%02x.%x has been disabled due to "
1109                          "system instability issues. "
1110                          "Specify rombar=1 or romfile to force\n",
1111                          vdev->host.domain, vdev->host.bus, vdev->host.slot,
1112                          vdev->host.function);
1113             return;
1114         }
1115     }
1116 
1117     trace_vfio_pci_size_rom(vdev->vbasedev.name, size);
1118 
1119     snprintf(name, sizeof(name), "vfio[%04x:%02x:%02x.%x].rom",
1120              vdev->host.domain, vdev->host.bus, vdev->host.slot,
1121              vdev->host.function);
1122 
1123     memory_region_init_io(&vdev->pdev.rom, OBJECT(vdev),
1124                           &vfio_rom_ops, vdev, name, size);
1125 
1126     pci_register_bar(&vdev->pdev, PCI_ROM_SLOT,
1127                      PCI_BASE_ADDRESS_SPACE_MEMORY, &vdev->pdev.rom);
1128 
1129     vdev->pdev.has_rom = true;
1130     vdev->rom_read_failed = false;
1131 }
1132 
1133 static void vfio_vga_write(void *opaque, hwaddr addr,
1134                            uint64_t data, unsigned size)
1135 {
1136     VFIOVGARegion *region = opaque;
1137     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1138     union {
1139         uint8_t byte;
1140         uint16_t word;
1141         uint32_t dword;
1142         uint64_t qword;
1143     } buf;
1144     off_t offset = vga->fd_offset + region->offset + addr;
1145 
1146     switch (size) {
1147     case 1:
1148         buf.byte = data;
1149         break;
1150     case 2:
1151         buf.word = cpu_to_le16(data);
1152         break;
1153     case 4:
1154         buf.dword = cpu_to_le32(data);
1155         break;
1156     default:
1157         hw_error("vfio: unsupported write size, %d bytes", size);
1158         break;
1159     }
1160 
1161     if (pwrite(vga->fd, &buf, size, offset) != size) {
1162         error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
1163                      __func__, region->offset + addr, data, size);
1164     }
1165 
1166     trace_vfio_vga_write(region->offset + addr, data, size);
1167 }
1168 
1169 static uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size)
1170 {
1171     VFIOVGARegion *region = opaque;
1172     VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1173     union {
1174         uint8_t byte;
1175         uint16_t word;
1176         uint32_t dword;
1177         uint64_t qword;
1178     } buf;
1179     uint64_t data = 0;
1180     off_t offset = vga->fd_offset + region->offset + addr;
1181 
1182     if (pread(vga->fd, &buf, size, offset) != size) {
1183         error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
1184                      __func__, region->offset + addr, size);
1185         return (uint64_t)-1;
1186     }
1187 
1188     switch (size) {
1189     case 1:
1190         data = buf.byte;
1191         break;
1192     case 2:
1193         data = le16_to_cpu(buf.word);
1194         break;
1195     case 4:
1196         data = le32_to_cpu(buf.dword);
1197         break;
1198     default:
1199         hw_error("vfio: unsupported read size, %d bytes", size);
1200         break;
1201     }
1202 
1203     trace_vfio_vga_read(region->offset + addr, size, data);
1204 
1205     return data;
1206 }
1207 
1208 static const MemoryRegionOps vfio_vga_ops = {
1209     .read = vfio_vga_read,
1210     .write = vfio_vga_write,
1211     .endianness = DEVICE_LITTLE_ENDIAN,
1212 };
1213 
1214 /*
1215  * Device specific quirks
1216  */
1217 
1218 /* Is range1 fully contained within range2?  */
1219 static bool vfio_range_contained(uint64_t first1, uint64_t len1,
1220                                  uint64_t first2, uint64_t len2) {
1221     return (first1 >= first2 && first1 + len1 <= first2 + len2);
1222 }
1223 
1224 static bool vfio_flags_enabled(uint8_t flags, uint8_t mask)
1225 {
1226     return (mask && (flags & mask) == mask);
1227 }
1228 
1229 static uint64_t vfio_generic_window_quirk_read(void *opaque,
1230                                                hwaddr addr, unsigned size)
1231 {
1232     VFIOQuirk *quirk = opaque;
1233     VFIOPCIDevice *vdev = quirk->vdev;
1234     uint64_t data;
1235 
1236     if (vfio_flags_enabled(quirk->data.flags, quirk->data.read_flags) &&
1237         ranges_overlap(addr, size,
1238                        quirk->data.data_offset, quirk->data.data_size)) {
1239         hwaddr offset = addr - quirk->data.data_offset;
1240 
1241         if (!vfio_range_contained(addr, size, quirk->data.data_offset,
1242                                   quirk->data.data_size)) {
1243             hw_error("%s: window data read not fully contained: %s",
1244                      __func__, memory_region_name(&quirk->mem));
1245         }
1246 
1247         data = vfio_pci_read_config(&vdev->pdev,
1248                                     quirk->data.address_val + offset, size);
1249 
1250         trace_vfio_generic_window_quirk_read(memory_region_name(&quirk->mem),
1251                                              vdev->vbasedev.name,
1252                                              quirk->data.bar,
1253                                              addr, size, data);
1254     } else {
1255         data = vfio_region_read(&vdev->bars[quirk->data.bar].region,
1256                                 addr + quirk->data.base_offset, size);
1257     }
1258 
1259     return data;
1260 }
1261 
1262 static void vfio_generic_window_quirk_write(void *opaque, hwaddr addr,
1263                                             uint64_t data, unsigned size)
1264 {
1265     VFIOQuirk *quirk = opaque;
1266     VFIOPCIDevice *vdev = quirk->vdev;
1267 
1268     if (ranges_overlap(addr, size,
1269                        quirk->data.address_offset, quirk->data.address_size)) {
1270 
1271         if (addr != quirk->data.address_offset) {
1272             hw_error("%s: offset write into address window: %s",
1273                      __func__, memory_region_name(&quirk->mem));
1274         }
1275 
1276         if ((data & ~quirk->data.address_mask) == quirk->data.address_match) {
1277             quirk->data.flags |= quirk->data.write_flags |
1278                                  quirk->data.read_flags;
1279             quirk->data.address_val = data & quirk->data.address_mask;
1280         } else {
1281             quirk->data.flags &= ~(quirk->data.write_flags |
1282                                    quirk->data.read_flags);
1283         }
1284     }
1285 
1286     if (vfio_flags_enabled(quirk->data.flags, quirk->data.write_flags) &&
1287         ranges_overlap(addr, size,
1288                        quirk->data.data_offset, quirk->data.data_size)) {
1289         hwaddr offset = addr - quirk->data.data_offset;
1290 
1291         if (!vfio_range_contained(addr, size, quirk->data.data_offset,
1292                                   quirk->data.data_size)) {
1293             hw_error("%s: window data write not fully contained: %s",
1294                      __func__, memory_region_name(&quirk->mem));
1295         }
1296 
1297         vfio_pci_write_config(&vdev->pdev,
1298                               quirk->data.address_val + offset, data, size);
1299         trace_vfio_generic_window_quirk_write(memory_region_name(&quirk->mem),
1300                                               vdev->vbasedev.name,
1301                                               quirk->data.bar,
1302                                               addr, data, size);
1303         return;
1304     }
1305 
1306     vfio_region_write(&vdev->bars[quirk->data.bar].region,
1307                    addr + quirk->data.base_offset, data, size);
1308 }
1309 
1310 static const MemoryRegionOps vfio_generic_window_quirk = {
1311     .read = vfio_generic_window_quirk_read,
1312     .write = vfio_generic_window_quirk_write,
1313     .endianness = DEVICE_LITTLE_ENDIAN,
1314 };
1315 
1316 static uint64_t vfio_generic_quirk_read(void *opaque,
1317                                         hwaddr addr, unsigned size)
1318 {
1319     VFIOQuirk *quirk = opaque;
1320     VFIOPCIDevice *vdev = quirk->vdev;
1321     hwaddr base = quirk->data.address_match & TARGET_PAGE_MASK;
1322     hwaddr offset = quirk->data.address_match & ~TARGET_PAGE_MASK;
1323     uint64_t data;
1324 
1325     if (vfio_flags_enabled(quirk->data.flags, quirk->data.read_flags) &&
1326         ranges_overlap(addr, size, offset, quirk->data.address_mask + 1)) {
1327         if (!vfio_range_contained(addr, size, offset,
1328                                   quirk->data.address_mask + 1)) {
1329             hw_error("%s: read not fully contained: %s",
1330                      __func__, memory_region_name(&quirk->mem));
1331         }
1332 
1333         data = vfio_pci_read_config(&vdev->pdev, addr - offset, size);
1334 
1335         trace_vfio_generic_quirk_read(memory_region_name(&quirk->mem),
1336                                       vdev->vbasedev.name, quirk->data.bar,
1337                                       addr + base, size, data);
1338     } else {
1339         data = vfio_region_read(&vdev->bars[quirk->data.bar].region,
1340                                 addr + base, size);
1341     }
1342 
1343     return data;
1344 }
1345 
1346 static void vfio_generic_quirk_write(void *opaque, hwaddr addr,
1347                                      uint64_t data, unsigned size)
1348 {
1349     VFIOQuirk *quirk = opaque;
1350     VFIOPCIDevice *vdev = quirk->vdev;
1351     hwaddr base = quirk->data.address_match & TARGET_PAGE_MASK;
1352     hwaddr offset = quirk->data.address_match & ~TARGET_PAGE_MASK;
1353 
1354     if (vfio_flags_enabled(quirk->data.flags, quirk->data.write_flags) &&
1355         ranges_overlap(addr, size, offset, quirk->data.address_mask + 1)) {
1356         if (!vfio_range_contained(addr, size, offset,
1357                                   quirk->data.address_mask + 1)) {
1358             hw_error("%s: write not fully contained: %s",
1359                      __func__, memory_region_name(&quirk->mem));
1360         }
1361 
1362         vfio_pci_write_config(&vdev->pdev, addr - offset, data, size);
1363 
1364         trace_vfio_generic_quirk_write(memory_region_name(&quirk->mem),
1365                                        vdev->vbasedev.name, quirk->data.bar,
1366                                        addr + base, data, size);
1367     } else {
1368         vfio_region_write(&vdev->bars[quirk->data.bar].region,
1369                           addr + base, data, size);
1370     }
1371 }
1372 
1373 static const MemoryRegionOps vfio_generic_quirk = {
1374     .read = vfio_generic_quirk_read,
1375     .write = vfio_generic_quirk_write,
1376     .endianness = DEVICE_LITTLE_ENDIAN,
1377 };
1378 
1379 #define PCI_VENDOR_ID_ATI               0x1002
1380 
1381 /*
1382  * Radeon HD cards (HD5450 & HD7850) report the upper byte of the I/O port BAR
1383  * through VGA register 0x3c3.  On newer cards, the I/O port BAR is always
1384  * BAR4 (older cards like the X550 used BAR1, but we don't care to support
1385  * those).  Note that on bare metal, a read of 0x3c3 doesn't always return the
1386  * I/O port BAR address.  Originally this was coded to return the virtual BAR
1387  * address only if the physical register read returns the actual BAR address,
1388  * but users have reported greater success if we return the virtual address
1389  * unconditionally.
1390  */
1391 static uint64_t vfio_ati_3c3_quirk_read(void *opaque,
1392                                         hwaddr addr, unsigned size)
1393 {
1394     VFIOQuirk *quirk = opaque;
1395     VFIOPCIDevice *vdev = quirk->vdev;
1396     uint64_t data = vfio_pci_read_config(&vdev->pdev,
1397                                          PCI_BASE_ADDRESS_0 + (4 * 4) + 1,
1398                                          size);
1399     trace_vfio_ati_3c3_quirk_read(data);
1400 
1401     return data;
1402 }
1403 
1404 static const MemoryRegionOps vfio_ati_3c3_quirk = {
1405     .read = vfio_ati_3c3_quirk_read,
1406     .endianness = DEVICE_LITTLE_ENDIAN,
1407 };
1408 
1409 static void vfio_vga_probe_ati_3c3_quirk(VFIOPCIDevice *vdev)
1410 {
1411     PCIDevice *pdev = &vdev->pdev;
1412     VFIOQuirk *quirk;
1413 
1414     if (pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1415         return;
1416     }
1417 
1418     /*
1419      * As long as the BAR is >= 256 bytes it will be aligned such that the
1420      * lower byte is always zero.  Filter out anything else, if it exists.
1421      */
1422     if (!vdev->bars[4].ioport || vdev->bars[4].region.size < 256) {
1423         return;
1424     }
1425 
1426     quirk = g_malloc0(sizeof(*quirk));
1427     quirk->vdev = vdev;
1428 
1429     memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_ati_3c3_quirk, quirk,
1430                           "vfio-ati-3c3-quirk", 1);
1431     memory_region_add_subregion(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem,
1432                                 3 /* offset 3 bytes from 0x3c0 */, &quirk->mem);
1433 
1434     QLIST_INSERT_HEAD(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks,
1435                       quirk, next);
1436 
1437     trace_vfio_vga_probe_ati_3c3_quirk(vdev->vbasedev.name);
1438 }
1439 
1440 /*
1441  * Newer ATI/AMD devices, including HD5450 and HD7850, have a window to PCI
1442  * config space through MMIO BAR2 at offset 0x4000.  Nothing seems to access
1443  * the MMIO space directly, but a window to this space is provided through
1444  * I/O port BAR4.  Offset 0x0 is the address register and offset 0x4 is the
1445  * data register.  When the address is programmed to a range of 0x4000-0x4fff
1446  * PCI configuration space is available.  Experimentation seems to indicate
1447  * that only read-only access is provided, but we drop writes when the window
1448  * is enabled to config space nonetheless.
1449  */
1450 static void vfio_probe_ati_bar4_window_quirk(VFIOPCIDevice *vdev, int nr)
1451 {
1452     PCIDevice *pdev = &vdev->pdev;
1453     VFIOQuirk *quirk;
1454 
1455     if (!vdev->has_vga || nr != 4 ||
1456         pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1457         return;
1458     }
1459 
1460     quirk = g_malloc0(sizeof(*quirk));
1461     quirk->vdev = vdev;
1462     quirk->data.address_size = 4;
1463     quirk->data.data_offset = 4;
1464     quirk->data.data_size = 4;
1465     quirk->data.address_match = 0x4000;
1466     quirk->data.address_mask = PCIE_CONFIG_SPACE_SIZE - 1;
1467     quirk->data.bar = nr;
1468     quirk->data.read_flags = quirk->data.write_flags = 1;
1469 
1470     memory_region_init_io(&quirk->mem, OBJECT(vdev),
1471                           &vfio_generic_window_quirk, quirk,
1472                           "vfio-ati-bar4-window-quirk", 8);
1473     memory_region_add_subregion_overlap(&vdev->bars[nr].region.mem,
1474                           quirk->data.base_offset, &quirk->mem, 1);
1475 
1476     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1477 
1478     trace_vfio_probe_ati_bar4_window_quirk(vdev->vbasedev.name);
1479 }
1480 
1481 #define PCI_VENDOR_ID_REALTEK 0x10ec
1482 
1483 /*
1484  * RTL8168 devices have a backdoor that can access the MSI-X table.  At BAR2
1485  * offset 0x70 there is a dword data register, offset 0x74 is a dword address
1486  * register.  According to the Linux r8169 driver, the MSI-X table is addressed
1487  * when the "type" portion of the address register is set to 0x1.  This appears
1488  * to be bits 16:30.  Bit 31 is both a write indicator and some sort of
1489  * "address latched" indicator.  Bits 12:15 are a mask field, which we can
1490  * ignore because the MSI-X table should always be accessed as a dword (full
1491  * mask).  Bits 0:11 is offset within the type.
1492  *
1493  * Example trace:
1494  *
1495  * Read from MSI-X table offset 0
1496  * vfio: vfio_bar_write(0000:05:00.0:BAR2+0x74, 0x1f000, 4) // store read addr
1497  * vfio: vfio_bar_read(0000:05:00.0:BAR2+0x74, 4) = 0x8001f000 // latch
1498  * vfio: vfio_bar_read(0000:05:00.0:BAR2+0x70, 4) = 0xfee00398 // read data
1499  *
1500  * Write 0xfee00000 to MSI-X table offset 0
1501  * vfio: vfio_bar_write(0000:05:00.0:BAR2+0x70, 0xfee00000, 4) // write data
1502  * vfio: vfio_bar_write(0000:05:00.0:BAR2+0x74, 0x8001f000, 4) // do write
1503  * vfio: vfio_bar_read(0000:05:00.0:BAR2+0x74, 4) = 0x1f000 // complete
1504  */
1505 
1506 static uint64_t vfio_rtl8168_window_quirk_read(void *opaque,
1507                                                hwaddr addr, unsigned size)
1508 {
1509     VFIOQuirk *quirk = opaque;
1510     VFIOPCIDevice *vdev = quirk->vdev;
1511 
1512     switch (addr) {
1513     case 4: /* address */
1514         if (quirk->data.flags) {
1515             trace_vfio_rtl8168_window_quirk_read_fake(
1516                     memory_region_name(&quirk->mem),
1517                     vdev->vbasedev.name);
1518 
1519             return quirk->data.address_match ^ 0x80000000U;
1520         }
1521         break;
1522     case 0: /* data */
1523         if (quirk->data.flags) {
1524             uint64_t val;
1525 
1526             trace_vfio_rtl8168_window_quirk_read_table(
1527                     memory_region_name(&quirk->mem),
1528                     vdev->vbasedev.name);
1529 
1530             if (!(vdev->pdev.cap_present & QEMU_PCI_CAP_MSIX)) {
1531                 return 0;
1532             }
1533 
1534             memory_region_dispatch_read(&vdev->pdev.msix_table_mmio,
1535                                         (hwaddr)(quirk->data.address_match
1536                                                  & 0xfff),
1537                                         &val,
1538                                         size,
1539                                         MEMTXATTRS_UNSPECIFIED);
1540             return val;
1541         }
1542     }
1543 
1544     trace_vfio_rtl8168_window_quirk_read_direct(memory_region_name(&quirk->mem),
1545                                                 vdev->vbasedev.name);
1546 
1547     return vfio_region_read(&vdev->bars[quirk->data.bar].region,
1548                             addr + 0x70, size);
1549 }
1550 
1551 static void vfio_rtl8168_window_quirk_write(void *opaque, hwaddr addr,
1552                                             uint64_t data, unsigned size)
1553 {
1554     VFIOQuirk *quirk = opaque;
1555     VFIOPCIDevice *vdev = quirk->vdev;
1556 
1557     switch (addr) {
1558     case 4: /* address */
1559         if ((data & 0x7fff0000) == 0x10000) {
1560             if (data & 0x80000000U &&
1561                 vdev->pdev.cap_present & QEMU_PCI_CAP_MSIX) {
1562 
1563                 trace_vfio_rtl8168_window_quirk_write_table(
1564                         memory_region_name(&quirk->mem),
1565                         vdev->vbasedev.name);
1566 
1567                 memory_region_dispatch_write(&vdev->pdev.msix_table_mmio,
1568                                              (hwaddr)(data & 0xfff),
1569                                              (uint64_t)quirk->data.address_mask,
1570                                              size, MEMTXATTRS_UNSPECIFIED);
1571             }
1572 
1573             quirk->data.flags = 1;
1574             quirk->data.address_match = data;
1575 
1576             return;
1577         }
1578         quirk->data.flags = 0;
1579         break;
1580     case 0: /* data */
1581         quirk->data.address_mask = data;
1582         break;
1583     }
1584 
1585     trace_vfio_rtl8168_window_quirk_write_direct(
1586             memory_region_name(&quirk->mem),
1587             vdev->vbasedev.name);
1588 
1589     vfio_region_write(&vdev->bars[quirk->data.bar].region,
1590                       addr + 0x70, data, size);
1591 }
1592 
1593 static const MemoryRegionOps vfio_rtl8168_window_quirk = {
1594     .read = vfio_rtl8168_window_quirk_read,
1595     .write = vfio_rtl8168_window_quirk_write,
1596     .valid = {
1597         .min_access_size = 4,
1598         .max_access_size = 4,
1599         .unaligned = false,
1600     },
1601     .endianness = DEVICE_LITTLE_ENDIAN,
1602 };
1603 
1604 static void vfio_probe_rtl8168_bar2_window_quirk(VFIOPCIDevice *vdev, int nr)
1605 {
1606     PCIDevice *pdev = &vdev->pdev;
1607     VFIOQuirk *quirk;
1608 
1609     if (pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_REALTEK ||
1610         pci_get_word(pdev->config + PCI_DEVICE_ID) != 0x8168 || nr != 2) {
1611         return;
1612     }
1613 
1614     quirk = g_malloc0(sizeof(*quirk));
1615     quirk->vdev = vdev;
1616     quirk->data.bar = nr;
1617 
1618     memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_rtl8168_window_quirk,
1619                           quirk, "vfio-rtl8168-window-quirk", 8);
1620     memory_region_add_subregion_overlap(&vdev->bars[nr].region.mem,
1621                                         0x70, &quirk->mem, 1);
1622 
1623     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1624 
1625     trace_vfio_probe_rtl8168_bar2_window_quirk(vdev->vbasedev.name);
1626 }
1627 /*
1628  * Trap the BAR2 MMIO window to config space as well.
1629  */
1630 static void vfio_probe_ati_bar2_4000_quirk(VFIOPCIDevice *vdev, int nr)
1631 {
1632     PCIDevice *pdev = &vdev->pdev;
1633     VFIOQuirk *quirk;
1634 
1635     /* Only enable on newer devices where BAR2 is 64bit */
1636     if (!vdev->has_vga || nr != 2 || !vdev->bars[2].mem64 ||
1637         pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1638         return;
1639     }
1640 
1641     quirk = g_malloc0(sizeof(*quirk));
1642     quirk->vdev = vdev;
1643     quirk->data.flags = quirk->data.read_flags = quirk->data.write_flags = 1;
1644     quirk->data.address_match = 0x4000;
1645     quirk->data.address_mask = PCIE_CONFIG_SPACE_SIZE - 1;
1646     quirk->data.bar = nr;
1647 
1648     memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_generic_quirk, quirk,
1649                           "vfio-ati-bar2-4000-quirk",
1650                           TARGET_PAGE_ALIGN(quirk->data.address_mask + 1));
1651     memory_region_add_subregion_overlap(&vdev->bars[nr].region.mem,
1652                           quirk->data.address_match & TARGET_PAGE_MASK,
1653                           &quirk->mem, 1);
1654 
1655     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1656 
1657     trace_vfio_probe_ati_bar2_4000_quirk(vdev->vbasedev.name);
1658 }
1659 
1660 /*
1661  * Older ATI/AMD cards like the X550 have a similar window to that above.
1662  * I/O port BAR1 provides a window to a mirror of PCI config space located
1663  * in BAR2 at offset 0xf00.  We don't care to support such older cards, but
1664  * note it for future reference.
1665  */
1666 
1667 #define PCI_VENDOR_ID_NVIDIA                    0x10de
1668 
1669 /*
1670  * Nvidia has several different methods to get to config space, the
1671  * nouveu project has several of these documented here:
1672  * https://github.com/pathscale/envytools/tree/master/hwdocs
1673  *
1674  * The first quirk is actually not documented in envytools and is found
1675  * on 10de:01d1 (NVIDIA Corporation G72 [GeForce 7300 LE]).  This is an
1676  * NV46 chipset.  The backdoor uses the legacy VGA I/O ports to access
1677  * the mirror of PCI config space found at BAR0 offset 0x1800.  The access
1678  * sequence first writes 0x338 to I/O port 0x3d4.  The target offset is
1679  * then written to 0x3d0.  Finally 0x538 is written for a read and 0x738
1680  * is written for a write to 0x3d4.  The BAR0 offset is then accessible
1681  * through 0x3d0.  This quirk doesn't seem to be necessary on newer cards
1682  * that use the I/O port BAR5 window but it doesn't hurt to leave it.
1683  */
1684 enum {
1685     NV_3D0_NONE = 0,
1686     NV_3D0_SELECT,
1687     NV_3D0_WINDOW,
1688     NV_3D0_READ,
1689     NV_3D0_WRITE,
1690 };
1691 
1692 static uint64_t vfio_nvidia_3d0_quirk_read(void *opaque,
1693                                            hwaddr addr, unsigned size)
1694 {
1695     VFIOQuirk *quirk = opaque;
1696     VFIOPCIDevice *vdev = quirk->vdev;
1697     PCIDevice *pdev = &vdev->pdev;
1698     uint64_t data = vfio_vga_read(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1699                                   addr + quirk->data.base_offset, size);
1700 
1701     if (quirk->data.flags == NV_3D0_READ && addr == quirk->data.data_offset) {
1702         data = vfio_pci_read_config(pdev, quirk->data.address_val, size);
1703         trace_vfio_nvidia_3d0_quirk_read(size, data);
1704     }
1705 
1706     quirk->data.flags = NV_3D0_NONE;
1707 
1708     return data;
1709 }
1710 
1711 static void vfio_nvidia_3d0_quirk_write(void *opaque, hwaddr addr,
1712                                         uint64_t data, unsigned size)
1713 {
1714     VFIOQuirk *quirk = opaque;
1715     VFIOPCIDevice *vdev = quirk->vdev;
1716     PCIDevice *pdev = &vdev->pdev;
1717 
1718     switch (quirk->data.flags) {
1719     case NV_3D0_NONE:
1720         if (addr == quirk->data.address_offset && data == 0x338) {
1721             quirk->data.flags = NV_3D0_SELECT;
1722         }
1723         break;
1724     case NV_3D0_SELECT:
1725         quirk->data.flags = NV_3D0_NONE;
1726         if (addr == quirk->data.data_offset &&
1727             (data & ~quirk->data.address_mask) == quirk->data.address_match) {
1728             quirk->data.flags = NV_3D0_WINDOW;
1729             quirk->data.address_val = data & quirk->data.address_mask;
1730         }
1731         break;
1732     case NV_3D0_WINDOW:
1733         quirk->data.flags = NV_3D0_NONE;
1734         if (addr == quirk->data.address_offset) {
1735             if (data == 0x538) {
1736                 quirk->data.flags = NV_3D0_READ;
1737             } else if (data == 0x738) {
1738                 quirk->data.flags = NV_3D0_WRITE;
1739             }
1740         }
1741         break;
1742     case NV_3D0_WRITE:
1743         quirk->data.flags = NV_3D0_NONE;
1744         if (addr == quirk->data.data_offset) {
1745             vfio_pci_write_config(pdev, quirk->data.address_val, data, size);
1746             trace_vfio_nvidia_3d0_quirk_write(data, size);
1747             return;
1748         }
1749         break;
1750     }
1751 
1752     vfio_vga_write(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1753                    addr + quirk->data.base_offset, data, size);
1754 }
1755 
1756 static const MemoryRegionOps vfio_nvidia_3d0_quirk = {
1757     .read = vfio_nvidia_3d0_quirk_read,
1758     .write = vfio_nvidia_3d0_quirk_write,
1759     .endianness = DEVICE_LITTLE_ENDIAN,
1760 };
1761 
1762 static void vfio_vga_probe_nvidia_3d0_quirk(VFIOPCIDevice *vdev)
1763 {
1764     PCIDevice *pdev = &vdev->pdev;
1765     VFIOQuirk *quirk;
1766 
1767     if (pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA ||
1768         !vdev->bars[1].region.size) {
1769         return;
1770     }
1771 
1772     quirk = g_malloc0(sizeof(*quirk));
1773     quirk->vdev = vdev;
1774     quirk->data.base_offset = 0x10;
1775     quirk->data.address_offset = 4;
1776     quirk->data.address_size = 2;
1777     quirk->data.address_match = 0x1800;
1778     quirk->data.address_mask = PCI_CONFIG_SPACE_SIZE - 1;
1779     quirk->data.data_offset = 0;
1780     quirk->data.data_size = 4;
1781 
1782     memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_nvidia_3d0_quirk,
1783                           quirk, "vfio-nvidia-3d0-quirk", 6);
1784     memory_region_add_subregion(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem,
1785                                 quirk->data.base_offset, &quirk->mem);
1786 
1787     QLIST_INSERT_HEAD(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks,
1788                       quirk, next);
1789 
1790     trace_vfio_vga_probe_nvidia_3d0_quirk(vdev->vbasedev.name);
1791 }
1792 
1793 /*
1794  * The second quirk is documented in envytools.  The I/O port BAR5 is just
1795  * a set of address/data ports to the MMIO BARs.  The BAR we care about is
1796  * again BAR0.  This backdoor is apparently a bit newer than the one above
1797  * so we need to not only trap 256 bytes @0x1800, but all of PCI config
1798  * space, including extended space is available at the 4k @0x88000.
1799  */
1800 enum {
1801     NV_BAR5_ADDRESS = 0x1,
1802     NV_BAR5_ENABLE = 0x2,
1803     NV_BAR5_MASTER = 0x4,
1804     NV_BAR5_VALID = 0x7,
1805 };
1806 
1807 static void vfio_nvidia_bar5_window_quirk_write(void *opaque, hwaddr addr,
1808                                                 uint64_t data, unsigned size)
1809 {
1810     VFIOQuirk *quirk = opaque;
1811 
1812     switch (addr) {
1813     case 0x0:
1814         if (data & 0x1) {
1815             quirk->data.flags |= NV_BAR5_MASTER;
1816         } else {
1817             quirk->data.flags &= ~NV_BAR5_MASTER;
1818         }
1819         break;
1820     case 0x4:
1821         if (data & 0x1) {
1822             quirk->data.flags |= NV_BAR5_ENABLE;
1823         } else {
1824             quirk->data.flags &= ~NV_BAR5_ENABLE;
1825         }
1826         break;
1827     case 0x8:
1828         if (quirk->data.flags & NV_BAR5_MASTER) {
1829             if ((data & ~0xfff) == 0x88000) {
1830                 quirk->data.flags |= NV_BAR5_ADDRESS;
1831                 quirk->data.address_val = data & 0xfff;
1832             } else if ((data & ~0xff) == 0x1800) {
1833                 quirk->data.flags |= NV_BAR5_ADDRESS;
1834                 quirk->data.address_val = data & 0xff;
1835             } else {
1836                 quirk->data.flags &= ~NV_BAR5_ADDRESS;
1837             }
1838         }
1839         break;
1840     }
1841 
1842     vfio_generic_window_quirk_write(opaque, addr, data, size);
1843 }
1844 
1845 static const MemoryRegionOps vfio_nvidia_bar5_window_quirk = {
1846     .read = vfio_generic_window_quirk_read,
1847     .write = vfio_nvidia_bar5_window_quirk_write,
1848     .valid.min_access_size = 4,
1849     .endianness = DEVICE_LITTLE_ENDIAN,
1850 };
1851 
1852 static void vfio_probe_nvidia_bar5_window_quirk(VFIOPCIDevice *vdev, int nr)
1853 {
1854     PCIDevice *pdev = &vdev->pdev;
1855     VFIOQuirk *quirk;
1856 
1857     if (!vdev->has_vga || nr != 5 ||
1858         pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1859         return;
1860     }
1861 
1862     quirk = g_malloc0(sizeof(*quirk));
1863     quirk->vdev = vdev;
1864     quirk->data.read_flags = quirk->data.write_flags = NV_BAR5_VALID;
1865     quirk->data.address_offset = 0x8;
1866     quirk->data.address_size = 0; /* actually 4, but avoids generic code */
1867     quirk->data.data_offset = 0xc;
1868     quirk->data.data_size = 4;
1869     quirk->data.bar = nr;
1870 
1871     memory_region_init_io(&quirk->mem, OBJECT(vdev),
1872                           &vfio_nvidia_bar5_window_quirk, quirk,
1873                           "vfio-nvidia-bar5-window-quirk", 16);
1874     memory_region_add_subregion_overlap(&vdev->bars[nr].region.mem,
1875                                         0, &quirk->mem, 1);
1876 
1877     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1878 
1879     trace_vfio_probe_nvidia_bar5_window_quirk(vdev->vbasedev.name);
1880 }
1881 
1882 static void vfio_nvidia_88000_quirk_write(void *opaque, hwaddr addr,
1883                                           uint64_t data, unsigned size)
1884 {
1885     VFIOQuirk *quirk = opaque;
1886     VFIOPCIDevice *vdev = quirk->vdev;
1887     PCIDevice *pdev = &vdev->pdev;
1888     hwaddr base = quirk->data.address_match & TARGET_PAGE_MASK;
1889 
1890     vfio_generic_quirk_write(opaque, addr, data, size);
1891 
1892     /*
1893      * Nvidia seems to acknowledge MSI interrupts by writing 0xff to the
1894      * MSI capability ID register.  Both the ID and next register are
1895      * read-only, so we allow writes covering either of those to real hw.
1896      * NB - only fixed for the 0x88000 MMIO window.
1897      */
1898     if ((pdev->cap_present & QEMU_PCI_CAP_MSI) &&
1899         vfio_range_contained(addr, size, pdev->msi_cap, PCI_MSI_FLAGS)) {
1900         vfio_region_write(&vdev->bars[quirk->data.bar].region,
1901                           addr + base, data, size);
1902     }
1903 }
1904 
1905 static const MemoryRegionOps vfio_nvidia_88000_quirk = {
1906     .read = vfio_generic_quirk_read,
1907     .write = vfio_nvidia_88000_quirk_write,
1908     .endianness = DEVICE_LITTLE_ENDIAN,
1909 };
1910 
1911 /*
1912  * Finally, BAR0 itself.  We want to redirect any accesses to either
1913  * 0x1800 or 0x88000 through the PCI config space access functions.
1914  *
1915  * NB - quirk at a page granularity or else they don't seem to work when
1916  *      BARs are mmap'd
1917  *
1918  * Here's offset 0x88000...
1919  */
1920 static void vfio_probe_nvidia_bar0_88000_quirk(VFIOPCIDevice *vdev, int nr)
1921 {
1922     PCIDevice *pdev = &vdev->pdev;
1923     VFIOQuirk *quirk;
1924     uint16_t vendor, class;
1925 
1926     vendor = pci_get_word(pdev->config + PCI_VENDOR_ID);
1927     class = pci_get_word(pdev->config + PCI_CLASS_DEVICE);
1928 
1929     if (nr != 0 || vendor != PCI_VENDOR_ID_NVIDIA ||
1930         class != PCI_CLASS_DISPLAY_VGA) {
1931         return;
1932     }
1933 
1934     quirk = g_malloc0(sizeof(*quirk));
1935     quirk->vdev = vdev;
1936     quirk->data.flags = quirk->data.read_flags = quirk->data.write_flags = 1;
1937     quirk->data.address_match = 0x88000;
1938     quirk->data.address_mask = PCIE_CONFIG_SPACE_SIZE - 1;
1939     quirk->data.bar = nr;
1940 
1941     memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_nvidia_88000_quirk,
1942                           quirk, "vfio-nvidia-bar0-88000-quirk",
1943                           TARGET_PAGE_ALIGN(quirk->data.address_mask + 1));
1944     memory_region_add_subregion_overlap(&vdev->bars[nr].region.mem,
1945                           quirk->data.address_match & TARGET_PAGE_MASK,
1946                           &quirk->mem, 1);
1947 
1948     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1949 
1950     trace_vfio_probe_nvidia_bar0_88000_quirk(vdev->vbasedev.name);
1951 }
1952 
1953 /*
1954  * And here's the same for BAR0 offset 0x1800...
1955  */
1956 static void vfio_probe_nvidia_bar0_1800_quirk(VFIOPCIDevice *vdev, int nr)
1957 {
1958     PCIDevice *pdev = &vdev->pdev;
1959     VFIOQuirk *quirk;
1960 
1961     if (!vdev->has_vga || nr != 0 ||
1962         pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1963         return;
1964     }
1965 
1966     /* Log the chipset ID */
1967     trace_vfio_probe_nvidia_bar0_1800_quirk_id(
1968             (unsigned int)(vfio_region_read(&vdev->bars[0].region, 0, 4) >> 20)
1969             & 0xff);
1970 
1971     quirk = g_malloc0(sizeof(*quirk));
1972     quirk->vdev = vdev;
1973     quirk->data.flags = quirk->data.read_flags = quirk->data.write_flags = 1;
1974     quirk->data.address_match = 0x1800;
1975     quirk->data.address_mask = PCI_CONFIG_SPACE_SIZE - 1;
1976     quirk->data.bar = nr;
1977 
1978     memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_generic_quirk, quirk,
1979                           "vfio-nvidia-bar0-1800-quirk",
1980                           TARGET_PAGE_ALIGN(quirk->data.address_mask + 1));
1981     memory_region_add_subregion_overlap(&vdev->bars[nr].region.mem,
1982                           quirk->data.address_match & TARGET_PAGE_MASK,
1983                           &quirk->mem, 1);
1984 
1985     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1986 
1987     trace_vfio_probe_nvidia_bar0_1800_quirk(vdev->vbasedev.name);
1988 }
1989 
1990 /*
1991  * TODO - Some Nvidia devices provide config access to their companion HDA
1992  * device and even to their parent bridge via these config space mirrors.
1993  * Add quirks for those regions.
1994  */
1995 
1996 /*
1997  * Common quirk probe entry points.
1998  */
1999 static void vfio_vga_quirk_setup(VFIOPCIDevice *vdev)
2000 {
2001     vfio_vga_probe_ati_3c3_quirk(vdev);
2002     vfio_vga_probe_nvidia_3d0_quirk(vdev);
2003 }
2004 
2005 static void vfio_vga_quirk_teardown(VFIOPCIDevice *vdev)
2006 {
2007     VFIOQuirk *quirk;
2008     int i;
2009 
2010     for (i = 0; i < ARRAY_SIZE(vdev->vga.region); i++) {
2011         QLIST_FOREACH(quirk, &vdev->vga.region[i].quirks, next) {
2012             memory_region_del_subregion(&vdev->vga.region[i].mem, &quirk->mem);
2013         }
2014     }
2015 }
2016 
2017 static void vfio_vga_quirk_free(VFIOPCIDevice *vdev)
2018 {
2019     int i;
2020 
2021     for (i = 0; i < ARRAY_SIZE(vdev->vga.region); i++) {
2022         while (!QLIST_EMPTY(&vdev->vga.region[i].quirks)) {
2023             VFIOQuirk *quirk = QLIST_FIRST(&vdev->vga.region[i].quirks);
2024             object_unparent(OBJECT(&quirk->mem));
2025             QLIST_REMOVE(quirk, next);
2026             g_free(quirk);
2027         }
2028     }
2029 }
2030 
2031 static void vfio_bar_quirk_setup(VFIOPCIDevice *vdev, int nr)
2032 {
2033     vfio_probe_ati_bar4_window_quirk(vdev, nr);
2034     vfio_probe_ati_bar2_4000_quirk(vdev, nr);
2035     vfio_probe_nvidia_bar5_window_quirk(vdev, nr);
2036     vfio_probe_nvidia_bar0_88000_quirk(vdev, nr);
2037     vfio_probe_nvidia_bar0_1800_quirk(vdev, nr);
2038     vfio_probe_rtl8168_bar2_window_quirk(vdev, nr);
2039 }
2040 
2041 static void vfio_bar_quirk_teardown(VFIOPCIDevice *vdev, int nr)
2042 {
2043     VFIOBAR *bar = &vdev->bars[nr];
2044     VFIOQuirk *quirk;
2045 
2046     QLIST_FOREACH(quirk, &bar->quirks, next) {
2047         memory_region_del_subregion(&bar->region.mem, &quirk->mem);
2048     }
2049 }
2050 
2051 static void vfio_bar_quirk_free(VFIOPCIDevice *vdev, int nr)
2052 {
2053     VFIOBAR *bar = &vdev->bars[nr];
2054 
2055     while (!QLIST_EMPTY(&bar->quirks)) {
2056         VFIOQuirk *quirk = QLIST_FIRST(&bar->quirks);
2057         object_unparent(OBJECT(&quirk->mem));
2058         QLIST_REMOVE(quirk, next);
2059         g_free(quirk);
2060     }
2061 }
2062 
2063 /*
2064  * PCI config space
2065  */
2066 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
2067 {
2068     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
2069     uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val;
2070 
2071     memcpy(&emu_bits, vdev->emulated_config_bits + addr, len);
2072     emu_bits = le32_to_cpu(emu_bits);
2073 
2074     if (emu_bits) {
2075         emu_val = pci_default_read_config(pdev, addr, len);
2076     }
2077 
2078     if (~emu_bits & (0xffffffffU >> (32 - len * 8))) {
2079         ssize_t ret;
2080 
2081         ret = pread(vdev->vbasedev.fd, &phys_val, len,
2082                     vdev->config_offset + addr);
2083         if (ret != len) {
2084             error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m",
2085                          __func__, vdev->host.domain, vdev->host.bus,
2086                          vdev->host.slot, vdev->host.function, addr, len);
2087             return -errno;
2088         }
2089         phys_val = le32_to_cpu(phys_val);
2090     }
2091 
2092     val = (emu_val & emu_bits) | (phys_val & ~emu_bits);
2093 
2094     trace_vfio_pci_read_config(vdev->vbasedev.name, addr, len, val);
2095 
2096     return val;
2097 }
2098 
2099 static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
2100                                   uint32_t val, int len)
2101 {
2102     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
2103     uint32_t val_le = cpu_to_le32(val);
2104 
2105     trace_vfio_pci_write_config(vdev->vbasedev.name, addr, val, len);
2106 
2107     /* Write everything to VFIO, let it filter out what we can't write */
2108     if (pwrite(vdev->vbasedev.fd, &val_le, len, vdev->config_offset + addr)
2109                 != len) {
2110         error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m",
2111                      __func__, vdev->host.domain, vdev->host.bus,
2112                      vdev->host.slot, vdev->host.function, addr, val, len);
2113     }
2114 
2115     /* MSI/MSI-X Enabling/Disabling */
2116     if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
2117         ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
2118         int is_enabled, was_enabled = msi_enabled(pdev);
2119 
2120         pci_default_write_config(pdev, addr, val, len);
2121 
2122         is_enabled = msi_enabled(pdev);
2123 
2124         if (!was_enabled) {
2125             if (is_enabled) {
2126                 vfio_enable_msi(vdev);
2127             }
2128         } else {
2129             if (!is_enabled) {
2130                 vfio_disable_msi(vdev);
2131             } else {
2132                 vfio_update_msi(vdev);
2133             }
2134         }
2135     } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
2136         ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
2137         int is_enabled, was_enabled = msix_enabled(pdev);
2138 
2139         pci_default_write_config(pdev, addr, val, len);
2140 
2141         is_enabled = msix_enabled(pdev);
2142 
2143         if (!was_enabled && is_enabled) {
2144             vfio_enable_msix(vdev);
2145         } else if (was_enabled && !is_enabled) {
2146             vfio_disable_msix(vdev);
2147         }
2148     } else {
2149         /* Write everything to QEMU to keep emulated bits correct */
2150         pci_default_write_config(pdev, addr, val, len);
2151     }
2152 }
2153 
2154 /*
2155  * Interrupt setup
2156  */
2157 static void vfio_disable_interrupts(VFIOPCIDevice *vdev)
2158 {
2159     /*
2160      * More complicated than it looks.  Disabling MSI/X transitions the
2161      * device to INTx mode (if supported).  Therefore we need to first
2162      * disable MSI/X and then cleanup by disabling INTx.
2163      */
2164     if (vdev->interrupt == VFIO_INT_MSIX) {
2165         vfio_disable_msix(vdev);
2166     } else if (vdev->interrupt == VFIO_INT_MSI) {
2167         vfio_disable_msi(vdev);
2168     }
2169 
2170     if (vdev->interrupt == VFIO_INT_INTx) {
2171         vfio_disable_intx(vdev);
2172     }
2173 }
2174 
2175 static int vfio_setup_msi(VFIOPCIDevice *vdev, int pos)
2176 {
2177     uint16_t ctrl;
2178     bool msi_64bit, msi_maskbit;
2179     int ret, entries;
2180 
2181     if (pread(vdev->vbasedev.fd, &ctrl, sizeof(ctrl),
2182               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
2183         return -errno;
2184     }
2185     ctrl = le16_to_cpu(ctrl);
2186 
2187     msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
2188     msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
2189     entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
2190 
2191     trace_vfio_setup_msi(vdev->vbasedev.name, pos);
2192 
2193     ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit);
2194     if (ret < 0) {
2195         if (ret == -ENOTSUP) {
2196             return 0;
2197         }
2198         error_report("vfio: msi_init failed");
2199         return ret;
2200     }
2201     vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
2202 
2203     return 0;
2204 }
2205 
2206 /*
2207  * We don't have any control over how pci_add_capability() inserts
2208  * capabilities into the chain.  In order to setup MSI-X we need a
2209  * MemoryRegion for the BAR.  In order to setup the BAR and not
2210  * attempt to mmap the MSI-X table area, which VFIO won't allow, we
2211  * need to first look for where the MSI-X table lives.  So we
2212  * unfortunately split MSI-X setup across two functions.
2213  */
2214 static int vfio_early_setup_msix(VFIOPCIDevice *vdev)
2215 {
2216     uint8_t pos;
2217     uint16_t ctrl;
2218     uint32_t table, pba;
2219     int fd = vdev->vbasedev.fd;
2220 
2221     pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
2222     if (!pos) {
2223         return 0;
2224     }
2225 
2226     if (pread(fd, &ctrl, sizeof(ctrl),
2227               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
2228         return -errno;
2229     }
2230 
2231     if (pread(fd, &table, sizeof(table),
2232               vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
2233         return -errno;
2234     }
2235 
2236     if (pread(fd, &pba, sizeof(pba),
2237               vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
2238         return -errno;
2239     }
2240 
2241     ctrl = le16_to_cpu(ctrl);
2242     table = le32_to_cpu(table);
2243     pba = le32_to_cpu(pba);
2244 
2245     vdev->msix = g_malloc0(sizeof(*(vdev->msix)));
2246     vdev->msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
2247     vdev->msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
2248     vdev->msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
2249     vdev->msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
2250     vdev->msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
2251 
2252     /*
2253      * Test the size of the pba_offset variable and catch if it extends outside
2254      * of the specified BAR. If it is the case, we need to apply a hardware
2255      * specific quirk if the device is known or we have a broken configuration.
2256      */
2257     if (vdev->msix->pba_offset >=
2258         vdev->bars[vdev->msix->pba_bar].region.size) {
2259 
2260         PCIDevice *pdev = &vdev->pdev;
2261         uint16_t vendor = pci_get_word(pdev->config + PCI_VENDOR_ID);
2262         uint16_t device = pci_get_word(pdev->config + PCI_DEVICE_ID);
2263 
2264         /*
2265          * Chelsio T5 Virtual Function devices are encoded as 0x58xx for T5
2266          * adapters. The T5 hardware returns an incorrect value of 0x8000 for
2267          * the VF PBA offset while the BAR itself is only 8k. The correct value
2268          * is 0x1000, so we hard code that here.
2269          */
2270         if (vendor == PCI_VENDOR_ID_CHELSIO && (device & 0xff00) == 0x5800) {
2271             vdev->msix->pba_offset = 0x1000;
2272         } else {
2273             error_report("vfio: Hardware reports invalid configuration, "
2274                          "MSIX PBA outside of specified BAR");
2275             return -EINVAL;
2276         }
2277     }
2278 
2279     trace_vfio_early_setup_msix(vdev->vbasedev.name, pos,
2280                                 vdev->msix->table_bar,
2281                                 vdev->msix->table_offset,
2282                                 vdev->msix->entries);
2283 
2284     return 0;
2285 }
2286 
2287 static int vfio_setup_msix(VFIOPCIDevice *vdev, int pos)
2288 {
2289     int ret;
2290 
2291     ret = msix_init(&vdev->pdev, vdev->msix->entries,
2292                     &vdev->bars[vdev->msix->table_bar].region.mem,
2293                     vdev->msix->table_bar, vdev->msix->table_offset,
2294                     &vdev->bars[vdev->msix->pba_bar].region.mem,
2295                     vdev->msix->pba_bar, vdev->msix->pba_offset, pos);
2296     if (ret < 0) {
2297         if (ret == -ENOTSUP) {
2298             return 0;
2299         }
2300         error_report("vfio: msix_init failed");
2301         return ret;
2302     }
2303 
2304     return 0;
2305 }
2306 
2307 static void vfio_teardown_msi(VFIOPCIDevice *vdev)
2308 {
2309     msi_uninit(&vdev->pdev);
2310 
2311     if (vdev->msix) {
2312         msix_uninit(&vdev->pdev,
2313                     &vdev->bars[vdev->msix->table_bar].region.mem,
2314                     &vdev->bars[vdev->msix->pba_bar].region.mem);
2315     }
2316 }
2317 
2318 /*
2319  * Resource setup
2320  */
2321 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled)
2322 {
2323     int i;
2324 
2325     for (i = 0; i < PCI_ROM_SLOT; i++) {
2326         VFIOBAR *bar = &vdev->bars[i];
2327 
2328         if (!bar->region.size) {
2329             continue;
2330         }
2331 
2332         memory_region_set_enabled(&bar->region.mmap_mem, enabled);
2333         if (vdev->msix && vdev->msix->table_bar == i) {
2334             memory_region_set_enabled(&vdev->msix->mmap_mem, enabled);
2335         }
2336     }
2337 }
2338 
2339 static void vfio_unregister_bar(VFIOPCIDevice *vdev, int nr)
2340 {
2341     VFIOBAR *bar = &vdev->bars[nr];
2342 
2343     if (!bar->region.size) {
2344         return;
2345     }
2346 
2347     vfio_bar_quirk_teardown(vdev, nr);
2348 
2349     memory_region_del_subregion(&bar->region.mem, &bar->region.mmap_mem);
2350 
2351     if (vdev->msix && vdev->msix->table_bar == nr) {
2352         memory_region_del_subregion(&bar->region.mem, &vdev->msix->mmap_mem);
2353     }
2354 }
2355 
2356 static void vfio_unmap_bar(VFIOPCIDevice *vdev, int nr)
2357 {
2358     VFIOBAR *bar = &vdev->bars[nr];
2359 
2360     if (!bar->region.size) {
2361         return;
2362     }
2363 
2364     vfio_bar_quirk_free(vdev, nr);
2365 
2366     munmap(bar->region.mmap, memory_region_size(&bar->region.mmap_mem));
2367 
2368     if (vdev->msix && vdev->msix->table_bar == nr) {
2369         munmap(vdev->msix->mmap, memory_region_size(&vdev->msix->mmap_mem));
2370     }
2371 }
2372 
2373 static void vfio_map_bar(VFIOPCIDevice *vdev, int nr)
2374 {
2375     VFIOBAR *bar = &vdev->bars[nr];
2376     uint64_t size = bar->region.size;
2377     char name[64];
2378     uint32_t pci_bar;
2379     uint8_t type;
2380     int ret;
2381 
2382     /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
2383     if (!size) {
2384         return;
2385     }
2386 
2387     snprintf(name, sizeof(name), "VFIO %04x:%02x:%02x.%x BAR %d",
2388              vdev->host.domain, vdev->host.bus, vdev->host.slot,
2389              vdev->host.function, nr);
2390 
2391     /* Determine what type of BAR this is for registration */
2392     ret = pread(vdev->vbasedev.fd, &pci_bar, sizeof(pci_bar),
2393                 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
2394     if (ret != sizeof(pci_bar)) {
2395         error_report("vfio: Failed to read BAR %d (%m)", nr);
2396         return;
2397     }
2398 
2399     pci_bar = le32_to_cpu(pci_bar);
2400     bar->ioport = (pci_bar & PCI_BASE_ADDRESS_SPACE_IO);
2401     bar->mem64 = bar->ioport ? 0 : (pci_bar & PCI_BASE_ADDRESS_MEM_TYPE_64);
2402     type = pci_bar & (bar->ioport ? ~PCI_BASE_ADDRESS_IO_MASK :
2403                                     ~PCI_BASE_ADDRESS_MEM_MASK);
2404 
2405     /* A "slow" read/write mapping underlies all BARs */
2406     memory_region_init_io(&bar->region.mem, OBJECT(vdev), &vfio_region_ops,
2407                           bar, name, size);
2408     pci_register_bar(&vdev->pdev, nr, type, &bar->region.mem);
2409 
2410     /*
2411      * We can't mmap areas overlapping the MSIX vector table, so we
2412      * potentially insert a direct-mapped subregion before and after it.
2413      */
2414     if (vdev->msix && vdev->msix->table_bar == nr) {
2415         size = vdev->msix->table_offset & qemu_real_host_page_mask;
2416     }
2417 
2418     strncat(name, " mmap", sizeof(name) - strlen(name) - 1);
2419     if (vfio_mmap_region(OBJECT(vdev), &bar->region, &bar->region.mem,
2420                       &bar->region.mmap_mem, &bar->region.mmap,
2421                       size, 0, name)) {
2422         error_report("%s unsupported. Performance may be slow", name);
2423     }
2424 
2425     if (vdev->msix && vdev->msix->table_bar == nr) {
2426         uint64_t start;
2427 
2428         start = REAL_HOST_PAGE_ALIGN((uint64_t)vdev->msix->table_offset +
2429                                      (vdev->msix->entries *
2430                                       PCI_MSIX_ENTRY_SIZE));
2431 
2432         size = start < bar->region.size ? bar->region.size - start : 0;
2433         strncat(name, " msix-hi", sizeof(name) - strlen(name) - 1);
2434         /* VFIOMSIXInfo contains another MemoryRegion for this mapping */
2435         if (vfio_mmap_region(OBJECT(vdev), &bar->region, &bar->region.mem,
2436                           &vdev->msix->mmap_mem,
2437                           &vdev->msix->mmap, size, start, name)) {
2438             error_report("%s unsupported. Performance may be slow", name);
2439         }
2440     }
2441 
2442     vfio_bar_quirk_setup(vdev, nr);
2443 }
2444 
2445 static void vfio_map_bars(VFIOPCIDevice *vdev)
2446 {
2447     int i;
2448 
2449     for (i = 0; i < PCI_ROM_SLOT; i++) {
2450         vfio_map_bar(vdev, i);
2451     }
2452 
2453     if (vdev->has_vga) {
2454         memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_MEM].mem,
2455                               OBJECT(vdev), &vfio_vga_ops,
2456                               &vdev->vga.region[QEMU_PCI_VGA_MEM],
2457                               "vfio-vga-mmio@0xa0000",
2458                               QEMU_PCI_VGA_MEM_SIZE);
2459         memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem,
2460                               OBJECT(vdev), &vfio_vga_ops,
2461                               &vdev->vga.region[QEMU_PCI_VGA_IO_LO],
2462                               "vfio-vga-io@0x3b0",
2463                               QEMU_PCI_VGA_IO_LO_SIZE);
2464         memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem,
2465                               OBJECT(vdev), &vfio_vga_ops,
2466                               &vdev->vga.region[QEMU_PCI_VGA_IO_HI],
2467                               "vfio-vga-io@0x3c0",
2468                               QEMU_PCI_VGA_IO_HI_SIZE);
2469 
2470         pci_register_vga(&vdev->pdev, &vdev->vga.region[QEMU_PCI_VGA_MEM].mem,
2471                          &vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem,
2472                          &vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem);
2473         vfio_vga_quirk_setup(vdev);
2474     }
2475 }
2476 
2477 static void vfio_unregister_bars(VFIOPCIDevice *vdev)
2478 {
2479     int i;
2480 
2481     for (i = 0; i < PCI_ROM_SLOT; i++) {
2482         vfio_unregister_bar(vdev, i);
2483     }
2484 
2485     if (vdev->has_vga) {
2486         vfio_vga_quirk_teardown(vdev);
2487         pci_unregister_vga(&vdev->pdev);
2488     }
2489 }
2490 
2491 static void vfio_unmap_bars(VFIOPCIDevice *vdev)
2492 {
2493     int i;
2494 
2495     for (i = 0; i < PCI_ROM_SLOT; i++) {
2496         vfio_unmap_bar(vdev, i);
2497     }
2498 
2499     if (vdev->has_vga) {
2500         vfio_vga_quirk_free(vdev);
2501     }
2502 }
2503 
2504 /*
2505  * General setup
2506  */
2507 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
2508 {
2509     uint8_t tmp, next = 0xff;
2510 
2511     for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
2512          tmp = pdev->config[tmp + 1]) {
2513         if (tmp > pos && tmp < next) {
2514             next = tmp;
2515         }
2516     }
2517 
2518     return next - pos;
2519 }
2520 
2521 static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask)
2522 {
2523     pci_set_word(buf, (pci_get_word(buf) & ~mask) | val);
2524 }
2525 
2526 static void vfio_add_emulated_word(VFIOPCIDevice *vdev, int pos,
2527                                    uint16_t val, uint16_t mask)
2528 {
2529     vfio_set_word_bits(vdev->pdev.config + pos, val, mask);
2530     vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask);
2531     vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask);
2532 }
2533 
2534 static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask)
2535 {
2536     pci_set_long(buf, (pci_get_long(buf) & ~mask) | val);
2537 }
2538 
2539 static void vfio_add_emulated_long(VFIOPCIDevice *vdev, int pos,
2540                                    uint32_t val, uint32_t mask)
2541 {
2542     vfio_set_long_bits(vdev->pdev.config + pos, val, mask);
2543     vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask);
2544     vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask);
2545 }
2546 
2547 static int vfio_setup_pcie_cap(VFIOPCIDevice *vdev, int pos, uint8_t size)
2548 {
2549     uint16_t flags;
2550     uint8_t type;
2551 
2552     flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS);
2553     type = (flags & PCI_EXP_FLAGS_TYPE) >> 4;
2554 
2555     if (type != PCI_EXP_TYPE_ENDPOINT &&
2556         type != PCI_EXP_TYPE_LEG_END &&
2557         type != PCI_EXP_TYPE_RC_END) {
2558 
2559         error_report("vfio: Assignment of PCIe type 0x%x "
2560                      "devices is not currently supported", type);
2561         return -EINVAL;
2562     }
2563 
2564     if (!pci_bus_is_express(vdev->pdev.bus)) {
2565         /*
2566          * Use express capability as-is on PCI bus.  It doesn't make much
2567          * sense to even expose, but some drivers (ex. tg3) depend on it
2568          * and guests don't seem to be particular about it.  We'll need
2569          * to revist this or force express devices to express buses if we
2570          * ever expose an IOMMU to the guest.
2571          */
2572     } else if (pci_bus_is_root(vdev->pdev.bus)) {
2573         /*
2574          * On a Root Complex bus Endpoints become Root Complex Integrated
2575          * Endpoints, which changes the type and clears the LNK & LNK2 fields.
2576          */
2577         if (type == PCI_EXP_TYPE_ENDPOINT) {
2578             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
2579                                    PCI_EXP_TYPE_RC_END << 4,
2580                                    PCI_EXP_FLAGS_TYPE);
2581 
2582             /* Link Capabilities, Status, and Control goes away */
2583             if (size > PCI_EXP_LNKCTL) {
2584                 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0);
2585                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
2586                 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0);
2587 
2588 #ifndef PCI_EXP_LNKCAP2
2589 #define PCI_EXP_LNKCAP2 44
2590 #endif
2591 #ifndef PCI_EXP_LNKSTA2
2592 #define PCI_EXP_LNKSTA2 50
2593 #endif
2594                 /* Link 2 Capabilities, Status, and Control goes away */
2595                 if (size > PCI_EXP_LNKCAP2) {
2596                     vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0);
2597                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0);
2598                     vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0);
2599                 }
2600             }
2601 
2602         } else if (type == PCI_EXP_TYPE_LEG_END) {
2603             /*
2604              * Legacy endpoints don't belong on the root complex.  Windows
2605              * seems to be happier with devices if we skip the capability.
2606              */
2607             return 0;
2608         }
2609 
2610     } else {
2611         /*
2612          * Convert Root Complex Integrated Endpoints to regular endpoints.
2613          * These devices don't support LNK/LNK2 capabilities, so make them up.
2614          */
2615         if (type == PCI_EXP_TYPE_RC_END) {
2616             vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
2617                                    PCI_EXP_TYPE_ENDPOINT << 4,
2618                                    PCI_EXP_FLAGS_TYPE);
2619             vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP,
2620                                    PCI_EXP_LNK_MLW_1 | PCI_EXP_LNK_LS_25, ~0);
2621             vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
2622         }
2623 
2624         /* Mark the Link Status bits as emulated to allow virtual negotiation */
2625         vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA,
2626                                pci_get_word(vdev->pdev.config + pos +
2627                                             PCI_EXP_LNKSTA),
2628                                PCI_EXP_LNKCAP_MLW | PCI_EXP_LNKCAP_SLS);
2629     }
2630 
2631     pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size);
2632     if (pos >= 0) {
2633         vdev->pdev.exp.exp_cap = pos;
2634     }
2635 
2636     return pos;
2637 }
2638 
2639 static void vfio_check_pcie_flr(VFIOPCIDevice *vdev, uint8_t pos)
2640 {
2641     uint32_t cap = pci_get_long(vdev->pdev.config + pos + PCI_EXP_DEVCAP);
2642 
2643     if (cap & PCI_EXP_DEVCAP_FLR) {
2644         trace_vfio_check_pcie_flr(vdev->vbasedev.name);
2645         vdev->has_flr = true;
2646     }
2647 }
2648 
2649 static void vfio_check_pm_reset(VFIOPCIDevice *vdev, uint8_t pos)
2650 {
2651     uint16_t csr = pci_get_word(vdev->pdev.config + pos + PCI_PM_CTRL);
2652 
2653     if (!(csr & PCI_PM_CTRL_NO_SOFT_RESET)) {
2654         trace_vfio_check_pm_reset(vdev->vbasedev.name);
2655         vdev->has_pm_reset = true;
2656     }
2657 }
2658 
2659 static void vfio_check_af_flr(VFIOPCIDevice *vdev, uint8_t pos)
2660 {
2661     uint8_t cap = pci_get_byte(vdev->pdev.config + pos + PCI_AF_CAP);
2662 
2663     if ((cap & PCI_AF_CAP_TP) && (cap & PCI_AF_CAP_FLR)) {
2664         trace_vfio_check_af_flr(vdev->vbasedev.name);
2665         vdev->has_flr = true;
2666     }
2667 }
2668 
2669 static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos)
2670 {
2671     PCIDevice *pdev = &vdev->pdev;
2672     uint8_t cap_id, next, size;
2673     int ret;
2674 
2675     cap_id = pdev->config[pos];
2676     next = pdev->config[pos + 1];
2677 
2678     /*
2679      * If it becomes important to configure capabilities to their actual
2680      * size, use this as the default when it's something we don't recognize.
2681      * Since QEMU doesn't actually handle many of the config accesses,
2682      * exact size doesn't seem worthwhile.
2683      */
2684     size = vfio_std_cap_max_size(pdev, pos);
2685 
2686     /*
2687      * pci_add_capability always inserts the new capability at the head
2688      * of the chain.  Therefore to end up with a chain that matches the
2689      * physical device, we insert from the end by making this recursive.
2690      * This is also why we pre-caclulate size above as cached config space
2691      * will be changed as we unwind the stack.
2692      */
2693     if (next) {
2694         ret = vfio_add_std_cap(vdev, next);
2695         if (ret) {
2696             return ret;
2697         }
2698     } else {
2699         /* Begin the rebuild, use QEMU emulated list bits */
2700         pdev->config[PCI_CAPABILITY_LIST] = 0;
2701         vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
2702         vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
2703     }
2704 
2705     /* Use emulated next pointer to allow dropping caps */
2706     pci_set_byte(vdev->emulated_config_bits + pos + 1, 0xff);
2707 
2708     switch (cap_id) {
2709     case PCI_CAP_ID_MSI:
2710         ret = vfio_setup_msi(vdev, pos);
2711         break;
2712     case PCI_CAP_ID_EXP:
2713         vfio_check_pcie_flr(vdev, pos);
2714         ret = vfio_setup_pcie_cap(vdev, pos, size);
2715         break;
2716     case PCI_CAP_ID_MSIX:
2717         ret = vfio_setup_msix(vdev, pos);
2718         break;
2719     case PCI_CAP_ID_PM:
2720         vfio_check_pm_reset(vdev, pos);
2721         vdev->pm_cap = pos;
2722         ret = pci_add_capability(pdev, cap_id, pos, size);
2723         break;
2724     case PCI_CAP_ID_AF:
2725         vfio_check_af_flr(vdev, pos);
2726         ret = pci_add_capability(pdev, cap_id, pos, size);
2727         break;
2728     default:
2729         ret = pci_add_capability(pdev, cap_id, pos, size);
2730         break;
2731     }
2732 
2733     if (ret < 0) {
2734         error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability "
2735                      "0x%x[0x%x]@0x%x: %d", vdev->host.domain,
2736                      vdev->host.bus, vdev->host.slot, vdev->host.function,
2737                      cap_id, size, pos, ret);
2738         return ret;
2739     }
2740 
2741     return 0;
2742 }
2743 
2744 static int vfio_add_capabilities(VFIOPCIDevice *vdev)
2745 {
2746     PCIDevice *pdev = &vdev->pdev;
2747 
2748     if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
2749         !pdev->config[PCI_CAPABILITY_LIST]) {
2750         return 0; /* Nothing to add */
2751     }
2752 
2753     return vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST]);
2754 }
2755 
2756 static void vfio_pci_pre_reset(VFIOPCIDevice *vdev)
2757 {
2758     PCIDevice *pdev = &vdev->pdev;
2759     uint16_t cmd;
2760 
2761     vfio_disable_interrupts(vdev);
2762 
2763     /* Make sure the device is in D0 */
2764     if (vdev->pm_cap) {
2765         uint16_t pmcsr;
2766         uint8_t state;
2767 
2768         pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2769         state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2770         if (state) {
2771             pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
2772             vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2);
2773             /* vfio handles the necessary delay here */
2774             pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
2775             state = pmcsr & PCI_PM_CTRL_STATE_MASK;
2776             if (state) {
2777                 error_report("vfio: Unable to power on device, stuck in D%d",
2778                              state);
2779             }
2780         }
2781     }
2782 
2783     /*
2784      * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
2785      * Also put INTx Disable in known state.
2786      */
2787     cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
2788     cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
2789              PCI_COMMAND_INTX_DISABLE);
2790     vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
2791 }
2792 
2793 static void vfio_pci_post_reset(VFIOPCIDevice *vdev)
2794 {
2795     vfio_enable_intx(vdev);
2796 }
2797 
2798 static bool vfio_pci_host_match(PCIHostDeviceAddress *host1,
2799                                 PCIHostDeviceAddress *host2)
2800 {
2801     return (host1->domain == host2->domain && host1->bus == host2->bus &&
2802             host1->slot == host2->slot && host1->function == host2->function);
2803 }
2804 
2805 static int vfio_pci_hot_reset(VFIOPCIDevice *vdev, bool single)
2806 {
2807     VFIOGroup *group;
2808     struct vfio_pci_hot_reset_info *info;
2809     struct vfio_pci_dependent_device *devices;
2810     struct vfio_pci_hot_reset *reset;
2811     int32_t *fds;
2812     int ret, i, count;
2813     bool multi = false;
2814 
2815     trace_vfio_pci_hot_reset(vdev->vbasedev.name, single ? "one" : "multi");
2816 
2817     vfio_pci_pre_reset(vdev);
2818     vdev->vbasedev.needs_reset = false;
2819 
2820     info = g_malloc0(sizeof(*info));
2821     info->argsz = sizeof(*info);
2822 
2823     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2824     if (ret && errno != ENOSPC) {
2825         ret = -errno;
2826         if (!vdev->has_pm_reset) {
2827             error_report("vfio: Cannot reset device %04x:%02x:%02x.%x, "
2828                          "no available reset mechanism.", vdev->host.domain,
2829                          vdev->host.bus, vdev->host.slot, vdev->host.function);
2830         }
2831         goto out_single;
2832     }
2833 
2834     count = info->count;
2835     info = g_realloc(info, sizeof(*info) + (count * sizeof(*devices)));
2836     info->argsz = sizeof(*info) + (count * sizeof(*devices));
2837     devices = &info->devices[0];
2838 
2839     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info);
2840     if (ret) {
2841         ret = -errno;
2842         error_report("vfio: hot reset info failed: %m");
2843         goto out_single;
2844     }
2845 
2846     trace_vfio_pci_hot_reset_has_dep_devices(vdev->vbasedev.name);
2847 
2848     /* Verify that we have all the groups required */
2849     for (i = 0; i < info->count; i++) {
2850         PCIHostDeviceAddress host;
2851         VFIOPCIDevice *tmp;
2852         VFIODevice *vbasedev_iter;
2853 
2854         host.domain = devices[i].segment;
2855         host.bus = devices[i].bus;
2856         host.slot = PCI_SLOT(devices[i].devfn);
2857         host.function = PCI_FUNC(devices[i].devfn);
2858 
2859         trace_vfio_pci_hot_reset_dep_devices(host.domain,
2860                 host.bus, host.slot, host.function, devices[i].group_id);
2861 
2862         if (vfio_pci_host_match(&host, &vdev->host)) {
2863             continue;
2864         }
2865 
2866         QLIST_FOREACH(group, &vfio_group_list, next) {
2867             if (group->groupid == devices[i].group_id) {
2868                 break;
2869             }
2870         }
2871 
2872         if (!group) {
2873             if (!vdev->has_pm_reset) {
2874                 error_report("vfio: Cannot reset device %s, "
2875                              "depends on group %d which is not owned.",
2876                              vdev->vbasedev.name, devices[i].group_id);
2877             }
2878             ret = -EPERM;
2879             goto out;
2880         }
2881 
2882         /* Prep dependent devices for reset and clear our marker. */
2883         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2884             if (vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2885                 continue;
2886             }
2887             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2888             if (vfio_pci_host_match(&host, &tmp->host)) {
2889                 if (single) {
2890                     ret = -EINVAL;
2891                     goto out_single;
2892                 }
2893                 vfio_pci_pre_reset(tmp);
2894                 tmp->vbasedev.needs_reset = false;
2895                 multi = true;
2896                 break;
2897             }
2898         }
2899     }
2900 
2901     if (!single && !multi) {
2902         ret = -EINVAL;
2903         goto out_single;
2904     }
2905 
2906     /* Determine how many group fds need to be passed */
2907     count = 0;
2908     QLIST_FOREACH(group, &vfio_group_list, next) {
2909         for (i = 0; i < info->count; i++) {
2910             if (group->groupid == devices[i].group_id) {
2911                 count++;
2912                 break;
2913             }
2914         }
2915     }
2916 
2917     reset = g_malloc0(sizeof(*reset) + (count * sizeof(*fds)));
2918     reset->argsz = sizeof(*reset) + (count * sizeof(*fds));
2919     fds = &reset->group_fds[0];
2920 
2921     /* Fill in group fds */
2922     QLIST_FOREACH(group, &vfio_group_list, next) {
2923         for (i = 0; i < info->count; i++) {
2924             if (group->groupid == devices[i].group_id) {
2925                 fds[reset->count++] = group->fd;
2926                 break;
2927             }
2928         }
2929     }
2930 
2931     /* Bus reset! */
2932     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_PCI_HOT_RESET, reset);
2933     g_free(reset);
2934 
2935     trace_vfio_pci_hot_reset_result(vdev->vbasedev.name,
2936                                     ret ? "%m" : "Success");
2937 
2938 out:
2939     /* Re-enable INTx on affected devices */
2940     for (i = 0; i < info->count; i++) {
2941         PCIHostDeviceAddress host;
2942         VFIOPCIDevice *tmp;
2943         VFIODevice *vbasedev_iter;
2944 
2945         host.domain = devices[i].segment;
2946         host.bus = devices[i].bus;
2947         host.slot = PCI_SLOT(devices[i].devfn);
2948         host.function = PCI_FUNC(devices[i].devfn);
2949 
2950         if (vfio_pci_host_match(&host, &vdev->host)) {
2951             continue;
2952         }
2953 
2954         QLIST_FOREACH(group, &vfio_group_list, next) {
2955             if (group->groupid == devices[i].group_id) {
2956                 break;
2957             }
2958         }
2959 
2960         if (!group) {
2961             break;
2962         }
2963 
2964         QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
2965             if (vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) {
2966                 continue;
2967             }
2968             tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev);
2969             if (vfio_pci_host_match(&host, &tmp->host)) {
2970                 vfio_pci_post_reset(tmp);
2971                 break;
2972             }
2973         }
2974     }
2975 out_single:
2976     vfio_pci_post_reset(vdev);
2977     g_free(info);
2978 
2979     return ret;
2980 }
2981 
2982 /*
2983  * We want to differentiate hot reset of mulitple in-use devices vs hot reset
2984  * of a single in-use device.  VFIO_DEVICE_RESET will already handle the case
2985  * of doing hot resets when there is only a single device per bus.  The in-use
2986  * here refers to how many VFIODevices are affected.  A hot reset that affects
2987  * multiple devices, but only a single in-use device, means that we can call
2988  * it from our bus ->reset() callback since the extent is effectively a single
2989  * device.  This allows us to make use of it in the hotplug path.  When there
2990  * are multiple in-use devices, we can only trigger the hot reset during a
2991  * system reset and thus from our reset handler.  We separate _one vs _multi
2992  * here so that we don't overlap and do a double reset on the system reset
2993  * path where both our reset handler and ->reset() callback are used.  Calling
2994  * _one() will only do a hot reset for the one in-use devices case, calling
2995  * _multi() will do nothing if a _one() would have been sufficient.
2996  */
2997 static int vfio_pci_hot_reset_one(VFIOPCIDevice *vdev)
2998 {
2999     return vfio_pci_hot_reset(vdev, true);
3000 }
3001 
3002 static int vfio_pci_hot_reset_multi(VFIODevice *vbasedev)
3003 {
3004     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
3005     return vfio_pci_hot_reset(vdev, false);
3006 }
3007 
3008 static void vfio_pci_compute_needs_reset(VFIODevice *vbasedev)
3009 {
3010     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
3011     if (!vbasedev->reset_works || (!vdev->has_flr && vdev->has_pm_reset)) {
3012         vbasedev->needs_reset = true;
3013     }
3014 }
3015 
3016 static VFIODeviceOps vfio_pci_ops = {
3017     .vfio_compute_needs_reset = vfio_pci_compute_needs_reset,
3018     .vfio_hot_reset_multi = vfio_pci_hot_reset_multi,
3019     .vfio_eoi = vfio_eoi,
3020 };
3021 
3022 static int vfio_populate_device(VFIOPCIDevice *vdev)
3023 {
3024     VFIODevice *vbasedev = &vdev->vbasedev;
3025     struct vfio_region_info reg_info = { .argsz = sizeof(reg_info) };
3026     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) };
3027     int i, ret = -1;
3028 
3029     /* Sanity check device */
3030     if (!(vbasedev->flags & VFIO_DEVICE_FLAGS_PCI)) {
3031         error_report("vfio: Um, this isn't a PCI device");
3032         goto error;
3033     }
3034 
3035     if (vbasedev->num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
3036         error_report("vfio: unexpected number of io regions %u",
3037                      vbasedev->num_regions);
3038         goto error;
3039     }
3040 
3041     if (vbasedev->num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
3042         error_report("vfio: unexpected number of irqs %u", vbasedev->num_irqs);
3043         goto error;
3044     }
3045 
3046     for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
3047         reg_info.index = i;
3048 
3049         ret = ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
3050         if (ret) {
3051             error_report("vfio: Error getting region %d info: %m", i);
3052             goto error;
3053         }
3054 
3055         trace_vfio_populate_device_region(vbasedev->name, i,
3056                                           (unsigned long)reg_info.size,
3057                                           (unsigned long)reg_info.offset,
3058                                           (unsigned long)reg_info.flags);
3059 
3060         vdev->bars[i].region.vbasedev = vbasedev;
3061         vdev->bars[i].region.flags = reg_info.flags;
3062         vdev->bars[i].region.size = reg_info.size;
3063         vdev->bars[i].region.fd_offset = reg_info.offset;
3064         vdev->bars[i].region.nr = i;
3065         QLIST_INIT(&vdev->bars[i].quirks);
3066     }
3067 
3068     reg_info.index = VFIO_PCI_CONFIG_REGION_INDEX;
3069 
3070     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
3071     if (ret) {
3072         error_report("vfio: Error getting config info: %m");
3073         goto error;
3074     }
3075 
3076     trace_vfio_populate_device_config(vdev->vbasedev.name,
3077                                       (unsigned long)reg_info.size,
3078                                       (unsigned long)reg_info.offset,
3079                                       (unsigned long)reg_info.flags);
3080 
3081     vdev->config_size = reg_info.size;
3082     if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
3083         vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
3084     }
3085     vdev->config_offset = reg_info.offset;
3086 
3087     if ((vdev->features & VFIO_FEATURE_ENABLE_VGA) &&
3088         vbasedev->num_regions > VFIO_PCI_VGA_REGION_INDEX) {
3089         struct vfio_region_info vga_info = {
3090             .argsz = sizeof(vga_info),
3091             .index = VFIO_PCI_VGA_REGION_INDEX,
3092          };
3093 
3094         ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_REGION_INFO, &vga_info);
3095         if (ret) {
3096             error_report(
3097                 "vfio: Device does not support requested feature x-vga");
3098             goto error;
3099         }
3100 
3101         if (!(vga_info.flags & VFIO_REGION_INFO_FLAG_READ) ||
3102             !(vga_info.flags & VFIO_REGION_INFO_FLAG_WRITE) ||
3103             vga_info.size < 0xbffff + 1) {
3104             error_report("vfio: Unexpected VGA info, flags 0x%lx, size 0x%lx",
3105                          (unsigned long)vga_info.flags,
3106                          (unsigned long)vga_info.size);
3107             goto error;
3108         }
3109 
3110         vdev->vga.fd_offset = vga_info.offset;
3111         vdev->vga.fd = vdev->vbasedev.fd;
3112 
3113         vdev->vga.region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE;
3114         vdev->vga.region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM;
3115         QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_MEM].quirks);
3116 
3117         vdev->vga.region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE;
3118         vdev->vga.region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO;
3119         QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].quirks);
3120 
3121         vdev->vga.region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE;
3122         vdev->vga.region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI;
3123         QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks);
3124 
3125         vdev->has_vga = true;
3126     }
3127 
3128     irq_info.index = VFIO_PCI_ERR_IRQ_INDEX;
3129 
3130     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info);
3131     if (ret) {
3132         /* This can fail for an old kernel or legacy PCI dev */
3133         trace_vfio_populate_device_get_irq_info_failure();
3134         ret = 0;
3135     } else if (irq_info.count == 1) {
3136         vdev->pci_aer = true;
3137     } else {
3138         error_report("vfio: %s "
3139                      "Could not enable error recovery for the device",
3140                      vbasedev->name);
3141     }
3142 
3143 error:
3144     return ret;
3145 }
3146 
3147 static void vfio_put_device(VFIOPCIDevice *vdev)
3148 {
3149     g_free(vdev->vbasedev.name);
3150     if (vdev->msix) {
3151         object_unparent(OBJECT(&vdev->msix->mmap_mem));
3152         g_free(vdev->msix);
3153         vdev->msix = NULL;
3154     }
3155     vfio_put_base_device(&vdev->vbasedev);
3156 }
3157 
3158 static void vfio_err_notifier_handler(void *opaque)
3159 {
3160     VFIOPCIDevice *vdev = opaque;
3161 
3162     if (!event_notifier_test_and_clear(&vdev->err_notifier)) {
3163         return;
3164     }
3165 
3166     /*
3167      * TBD. Retrieve the error details and decide what action
3168      * needs to be taken. One of the actions could be to pass
3169      * the error to the guest and have the guest driver recover
3170      * from the error. This requires that PCIe capabilities be
3171      * exposed to the guest. For now, we just terminate the
3172      * guest to contain the error.
3173      */
3174 
3175     error_report("%s(%04x:%02x:%02x.%x) Unrecoverable error detected.  "
3176                  "Please collect any data possible and then kill the guest",
3177                  __func__, vdev->host.domain, vdev->host.bus,
3178                  vdev->host.slot, vdev->host.function);
3179 
3180     vm_stop(RUN_STATE_INTERNAL_ERROR);
3181 }
3182 
3183 /*
3184  * Registers error notifier for devices supporting error recovery.
3185  * If we encounter a failure in this function, we report an error
3186  * and continue after disabling error recovery support for the
3187  * device.
3188  */
3189 static void vfio_register_err_notifier(VFIOPCIDevice *vdev)
3190 {
3191     int ret;
3192     int argsz;
3193     struct vfio_irq_set *irq_set;
3194     int32_t *pfd;
3195 
3196     if (!vdev->pci_aer) {
3197         return;
3198     }
3199 
3200     if (event_notifier_init(&vdev->err_notifier, 0)) {
3201         error_report("vfio: Unable to init event notifier for error detection");
3202         vdev->pci_aer = false;
3203         return;
3204     }
3205 
3206     argsz = sizeof(*irq_set) + sizeof(*pfd);
3207 
3208     irq_set = g_malloc0(argsz);
3209     irq_set->argsz = argsz;
3210     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
3211                      VFIO_IRQ_SET_ACTION_TRIGGER;
3212     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
3213     irq_set->start = 0;
3214     irq_set->count = 1;
3215     pfd = (int32_t *)&irq_set->data;
3216 
3217     *pfd = event_notifier_get_fd(&vdev->err_notifier);
3218     qemu_set_fd_handler(*pfd, vfio_err_notifier_handler, NULL, vdev);
3219 
3220     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
3221     if (ret) {
3222         error_report("vfio: Failed to set up error notification");
3223         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
3224         event_notifier_cleanup(&vdev->err_notifier);
3225         vdev->pci_aer = false;
3226     }
3227     g_free(irq_set);
3228 }
3229 
3230 static void vfio_unregister_err_notifier(VFIOPCIDevice *vdev)
3231 {
3232     int argsz;
3233     struct vfio_irq_set *irq_set;
3234     int32_t *pfd;
3235     int ret;
3236 
3237     if (!vdev->pci_aer) {
3238         return;
3239     }
3240 
3241     argsz = sizeof(*irq_set) + sizeof(*pfd);
3242 
3243     irq_set = g_malloc0(argsz);
3244     irq_set->argsz = argsz;
3245     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
3246                      VFIO_IRQ_SET_ACTION_TRIGGER;
3247     irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
3248     irq_set->start = 0;
3249     irq_set->count = 1;
3250     pfd = (int32_t *)&irq_set->data;
3251     *pfd = -1;
3252 
3253     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set);
3254     if (ret) {
3255         error_report("vfio: Failed to de-assign error fd: %m");
3256     }
3257     g_free(irq_set);
3258     qemu_set_fd_handler(event_notifier_get_fd(&vdev->err_notifier),
3259                         NULL, NULL, vdev);
3260     event_notifier_cleanup(&vdev->err_notifier);
3261 }
3262 
3263 static void vfio_req_notifier_handler(void *opaque)
3264 {
3265     VFIOPCIDevice *vdev = opaque;
3266 
3267     if (!event_notifier_test_and_clear(&vdev->req_notifier)) {
3268         return;
3269     }
3270 
3271     qdev_unplug(&vdev->pdev.qdev, NULL);
3272 }
3273 
3274 static void vfio_register_req_notifier(VFIOPCIDevice *vdev)
3275 {
3276     struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info),
3277                                       .index = VFIO_PCI_REQ_IRQ_INDEX };
3278     int argsz;
3279     struct vfio_irq_set *irq_set;
3280     int32_t *pfd;
3281 
3282     if (!(vdev->features & VFIO_FEATURE_ENABLE_REQ)) {
3283         return;
3284     }
3285 
3286     if (ioctl(vdev->vbasedev.fd,
3287               VFIO_DEVICE_GET_IRQ_INFO, &irq_info) < 0 || irq_info.count < 1) {
3288         return;
3289     }
3290 
3291     if (event_notifier_init(&vdev->req_notifier, 0)) {
3292         error_report("vfio: Unable to init event notifier for device request");
3293         return;
3294     }
3295 
3296     argsz = sizeof(*irq_set) + sizeof(*pfd);
3297 
3298     irq_set = g_malloc0(argsz);
3299     irq_set->argsz = argsz;
3300     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
3301                      VFIO_IRQ_SET_ACTION_TRIGGER;
3302     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
3303     irq_set->start = 0;
3304     irq_set->count = 1;
3305     pfd = (int32_t *)&irq_set->data;
3306 
3307     *pfd = event_notifier_get_fd(&vdev->req_notifier);
3308     qemu_set_fd_handler(*pfd, vfio_req_notifier_handler, NULL, vdev);
3309 
3310     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
3311         error_report("vfio: Failed to set up device request notification");
3312         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
3313         event_notifier_cleanup(&vdev->req_notifier);
3314     } else {
3315         vdev->req_enabled = true;
3316     }
3317 
3318     g_free(irq_set);
3319 }
3320 
3321 static void vfio_unregister_req_notifier(VFIOPCIDevice *vdev)
3322 {
3323     int argsz;
3324     struct vfio_irq_set *irq_set;
3325     int32_t *pfd;
3326 
3327     if (!vdev->req_enabled) {
3328         return;
3329     }
3330 
3331     argsz = sizeof(*irq_set) + sizeof(*pfd);
3332 
3333     irq_set = g_malloc0(argsz);
3334     irq_set->argsz = argsz;
3335     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
3336                      VFIO_IRQ_SET_ACTION_TRIGGER;
3337     irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
3338     irq_set->start = 0;
3339     irq_set->count = 1;
3340     pfd = (int32_t *)&irq_set->data;
3341     *pfd = -1;
3342 
3343     if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
3344         error_report("vfio: Failed to de-assign device request fd: %m");
3345     }
3346     g_free(irq_set);
3347     qemu_set_fd_handler(event_notifier_get_fd(&vdev->req_notifier),
3348                         NULL, NULL, vdev);
3349     event_notifier_cleanup(&vdev->req_notifier);
3350 
3351     vdev->req_enabled = false;
3352 }
3353 
3354 /*
3355  * AMD Radeon PCI config reset, based on Linux:
3356  *   drivers/gpu/drm/radeon/ci_smc.c:ci_is_smc_running()
3357  *   drivers/gpu/drm/radeon/radeon_device.c:radeon_pci_config_reset
3358  *   drivers/gpu/drm/radeon/ci_smc.c:ci_reset_smc()
3359  *   drivers/gpu/drm/radeon/ci_smc.c:ci_stop_smc_clock()
3360  * IDs: include/drm/drm_pciids.h
3361  * Registers: http://cgit.freedesktop.org/~agd5f/linux/commit/?id=4e2aa447f6f0
3362  *
3363  * Bonaire and Hawaii GPUs do not respond to a bus reset.  This is a bug in the
3364  * hardware that should be fixed on future ASICs.  The symptom of this is that
3365  * once the accerlated driver loads, Windows guests will bsod on subsequent
3366  * attmpts to load the driver, such as after VM reset or shutdown/restart.  To
3367  * work around this, we do an AMD specific PCI config reset, followed by an SMC
3368  * reset.  The PCI config reset only works if SMC firmware is running, so we
3369  * have a dependency on the state of the device as to whether this reset will
3370  * be effective.  There are still cases where we won't be able to kick the
3371  * device into working, but this greatly improves the usability overall.  The
3372  * config reset magic is relatively common on AMD GPUs, but the setup and SMC
3373  * poking is largely ASIC specific.
3374  */
3375 static bool vfio_radeon_smc_is_running(VFIOPCIDevice *vdev)
3376 {
3377     uint32_t clk, pc_c;
3378 
3379     /*
3380      * Registers 200h and 204h are index and data registers for accessing
3381      * indirect configuration registers within the device.
3382      */
3383     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000004, 4);
3384     clk = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
3385     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000370, 4);
3386     pc_c = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
3387 
3388     return (!(clk & 1) && (0x20100 <= pc_c));
3389 }
3390 
3391 /*
3392  * The scope of a config reset is controlled by a mode bit in the misc register
3393  * and a fuse, exposed as a bit in another register.  The fuse is the default
3394  * (0 = GFX, 1 = whole GPU), the misc bit is a toggle, with the forumula
3395  * scope = !(misc ^ fuse), where the resulting scope is defined the same as
3396  * the fuse.  A truth table therefore tells us that if misc == fuse, we need
3397  * to flip the value of the bit in the misc register.
3398  */
3399 static void vfio_radeon_set_gfx_only_reset(VFIOPCIDevice *vdev)
3400 {
3401     uint32_t misc, fuse;
3402     bool a, b;
3403 
3404     vfio_region_write(&vdev->bars[5].region, 0x200, 0xc00c0000, 4);
3405     fuse = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
3406     b = fuse & 64;
3407 
3408     vfio_region_write(&vdev->bars[5].region, 0x200, 0xc0000010, 4);
3409     misc = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
3410     a = misc & 2;
3411 
3412     if (a == b) {
3413         vfio_region_write(&vdev->bars[5].region, 0x204, misc ^ 2, 4);
3414         vfio_region_read(&vdev->bars[5].region, 0x204, 4); /* flush */
3415     }
3416 }
3417 
3418 static int vfio_radeon_reset(VFIOPCIDevice *vdev)
3419 {
3420     PCIDevice *pdev = &vdev->pdev;
3421     int i, ret = 0;
3422     uint32_t data;
3423 
3424     /* Defer to a kernel implemented reset */
3425     if (vdev->vbasedev.reset_works) {
3426         return -ENODEV;
3427     }
3428 
3429     /* Enable only memory BAR access */
3430     vfio_pci_write_config(pdev, PCI_COMMAND, PCI_COMMAND_MEMORY, 2);
3431 
3432     /* Reset only works if SMC firmware is loaded and running */
3433     if (!vfio_radeon_smc_is_running(vdev)) {
3434         ret = -EINVAL;
3435         goto out;
3436     }
3437 
3438     /* Make sure only the GFX function is reset */
3439     vfio_radeon_set_gfx_only_reset(vdev);
3440 
3441     /* AMD PCI config reset */
3442     vfio_pci_write_config(pdev, 0x7c, 0x39d5e86b, 4);
3443     usleep(100);
3444 
3445     /* Read back the memory size to make sure we're out of reset */
3446     for (i = 0; i < 100000; i++) {
3447         if (vfio_region_read(&vdev->bars[5].region, 0x5428, 4) != 0xffffffff) {
3448             break;
3449         }
3450         usleep(1);
3451     }
3452 
3453     /* Reset SMC */
3454     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000000, 4);
3455     data = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
3456     data |= 1;
3457     vfio_region_write(&vdev->bars[5].region, 0x204, data, 4);
3458 
3459     /* Disable SMC clock */
3460     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000004, 4);
3461     data = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
3462     data |= 1;
3463     vfio_region_write(&vdev->bars[5].region, 0x204, data, 4);
3464 
3465 out:
3466     /* Restore PCI command register */
3467     vfio_pci_write_config(pdev, PCI_COMMAND, 0, 2);
3468 
3469     return ret;
3470 }
3471 
3472 static void vfio_setup_resetfn(VFIOPCIDevice *vdev)
3473 {
3474     PCIDevice *pdev = &vdev->pdev;
3475     uint16_t vendor, device;
3476 
3477     vendor = pci_get_word(pdev->config + PCI_VENDOR_ID);
3478     device = pci_get_word(pdev->config + PCI_DEVICE_ID);
3479 
3480     switch (vendor) {
3481     case 0x1002:
3482         switch (device) {
3483         /* Bonaire */
3484         case 0x6649: /* Bonaire [FirePro W5100] */
3485         case 0x6650:
3486         case 0x6651:
3487         case 0x6658: /* Bonaire XTX [Radeon R7 260X] */
3488         case 0x665c: /* Bonaire XT [Radeon HD 7790/8770 / R9 260 OEM] */
3489         case 0x665d: /* Bonaire [Radeon R7 200 Series] */
3490         /* Hawaii */
3491         case 0x67A0: /* Hawaii XT GL [FirePro W9100] */
3492         case 0x67A1: /* Hawaii PRO GL [FirePro W8100] */
3493         case 0x67A2:
3494         case 0x67A8:
3495         case 0x67A9:
3496         case 0x67AA:
3497         case 0x67B0: /* Hawaii XT [Radeon R9 290X] */
3498         case 0x67B1: /* Hawaii PRO [Radeon R9 290] */
3499         case 0x67B8:
3500         case 0x67B9:
3501         case 0x67BA:
3502         case 0x67BE:
3503             vdev->resetfn = vfio_radeon_reset;
3504             break;
3505         }
3506         break;
3507     }
3508 }
3509 
3510 static int vfio_initfn(PCIDevice *pdev)
3511 {
3512     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
3513     VFIODevice *vbasedev_iter;
3514     VFIOGroup *group;
3515     char path[PATH_MAX], iommu_group_path[PATH_MAX], *group_name;
3516     ssize_t len;
3517     struct stat st;
3518     int groupid;
3519     int ret;
3520 
3521     /* Check that the host device exists */
3522     snprintf(path, sizeof(path),
3523              "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/",
3524              vdev->host.domain, vdev->host.bus, vdev->host.slot,
3525              vdev->host.function);
3526     if (stat(path, &st) < 0) {
3527         error_report("vfio: error: no such host device: %s", path);
3528         return -errno;
3529     }
3530 
3531     vdev->vbasedev.ops = &vfio_pci_ops;
3532 
3533     vdev->vbasedev.type = VFIO_DEVICE_TYPE_PCI;
3534     vdev->vbasedev.name = g_strdup_printf("%04x:%02x:%02x.%01x",
3535                                           vdev->host.domain, vdev->host.bus,
3536                                           vdev->host.slot, vdev->host.function);
3537 
3538     strncat(path, "iommu_group", sizeof(path) - strlen(path) - 1);
3539 
3540     len = readlink(path, iommu_group_path, sizeof(path));
3541     if (len <= 0 || len >= sizeof(path)) {
3542         error_report("vfio: error no iommu_group for device");
3543         return len < 0 ? -errno : -ENAMETOOLONG;
3544     }
3545 
3546     iommu_group_path[len] = 0;
3547     group_name = basename(iommu_group_path);
3548 
3549     if (sscanf(group_name, "%d", &groupid) != 1) {
3550         error_report("vfio: error reading %s: %m", path);
3551         return -errno;
3552     }
3553 
3554     trace_vfio_initfn(vdev->vbasedev.name, groupid);
3555 
3556     group = vfio_get_group(groupid, pci_device_iommu_address_space(pdev));
3557     if (!group) {
3558         error_report("vfio: failed to get group %d", groupid);
3559         return -ENOENT;
3560     }
3561 
3562     snprintf(path, sizeof(path), "%04x:%02x:%02x.%01x",
3563             vdev->host.domain, vdev->host.bus, vdev->host.slot,
3564             vdev->host.function);
3565 
3566     QLIST_FOREACH(vbasedev_iter, &group->device_list, next) {
3567         if (strcmp(vbasedev_iter->name, vdev->vbasedev.name) == 0) {
3568             error_report("vfio: error: device %s is already attached", path);
3569             vfio_put_group(group);
3570             return -EBUSY;
3571         }
3572     }
3573 
3574     ret = vfio_get_device(group, path, &vdev->vbasedev);
3575     if (ret) {
3576         error_report("vfio: failed to get device %s", path);
3577         vfio_put_group(group);
3578         return ret;
3579     }
3580 
3581     ret = vfio_populate_device(vdev);
3582     if (ret) {
3583         return ret;
3584     }
3585 
3586     /* Get a copy of config space */
3587     ret = pread(vdev->vbasedev.fd, vdev->pdev.config,
3588                 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
3589                 vdev->config_offset);
3590     if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
3591         ret = ret < 0 ? -errno : -EFAULT;
3592         error_report("vfio: Failed to read device config space");
3593         return ret;
3594     }
3595 
3596     /* vfio emulates a lot for us, but some bits need extra love */
3597     vdev->emulated_config_bits = g_malloc0(vdev->config_size);
3598 
3599     /* QEMU can choose to expose the ROM or not */
3600     memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4);
3601 
3602     /* QEMU can change multi-function devices to single function, or reverse */
3603     vdev->emulated_config_bits[PCI_HEADER_TYPE] =
3604                                               PCI_HEADER_TYPE_MULTI_FUNCTION;
3605 
3606     /* Restore or clear multifunction, this is always controlled by QEMU */
3607     if (vdev->pdev.cap_present & QEMU_PCI_CAP_MULTIFUNCTION) {
3608         vdev->pdev.config[PCI_HEADER_TYPE] |= PCI_HEADER_TYPE_MULTI_FUNCTION;
3609     } else {
3610         vdev->pdev.config[PCI_HEADER_TYPE] &= ~PCI_HEADER_TYPE_MULTI_FUNCTION;
3611     }
3612 
3613     /*
3614      * Clear host resource mapping info.  If we choose not to register a
3615      * BAR, such as might be the case with the option ROM, we can get
3616      * confusing, unwritable, residual addresses from the host here.
3617      */
3618     memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
3619     memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
3620 
3621     vfio_pci_size_rom(vdev);
3622 
3623     ret = vfio_early_setup_msix(vdev);
3624     if (ret) {
3625         return ret;
3626     }
3627 
3628     vfio_map_bars(vdev);
3629 
3630     ret = vfio_add_capabilities(vdev);
3631     if (ret) {
3632         goto out_teardown;
3633     }
3634 
3635     /* QEMU emulates all of MSI & MSIX */
3636     if (pdev->cap_present & QEMU_PCI_CAP_MSIX) {
3637         memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff,
3638                MSIX_CAP_LENGTH);
3639     }
3640 
3641     if (pdev->cap_present & QEMU_PCI_CAP_MSI) {
3642         memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff,
3643                vdev->msi_cap_size);
3644     }
3645 
3646     if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
3647         vdev->intx.mmap_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL,
3648                                                   vfio_intx_mmap_enable, vdev);
3649         pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_update_irq);
3650         ret = vfio_enable_intx(vdev);
3651         if (ret) {
3652             goto out_teardown;
3653         }
3654     }
3655 
3656     vfio_register_err_notifier(vdev);
3657     vfio_register_req_notifier(vdev);
3658     vfio_setup_resetfn(vdev);
3659 
3660     return 0;
3661 
3662 out_teardown:
3663     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3664     vfio_teardown_msi(vdev);
3665     vfio_unregister_bars(vdev);
3666     return ret;
3667 }
3668 
3669 static void vfio_instance_finalize(Object *obj)
3670 {
3671     PCIDevice *pci_dev = PCI_DEVICE(obj);
3672     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pci_dev);
3673     VFIOGroup *group = vdev->vbasedev.group;
3674 
3675     vfio_unmap_bars(vdev);
3676     g_free(vdev->emulated_config_bits);
3677     g_free(vdev->rom);
3678     vfio_put_device(vdev);
3679     vfio_put_group(group);
3680 }
3681 
3682 static void vfio_exitfn(PCIDevice *pdev)
3683 {
3684     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
3685 
3686     vfio_unregister_req_notifier(vdev);
3687     vfio_unregister_err_notifier(vdev);
3688     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3689     vfio_disable_interrupts(vdev);
3690     if (vdev->intx.mmap_timer) {
3691         timer_free(vdev->intx.mmap_timer);
3692     }
3693     vfio_teardown_msi(vdev);
3694     vfio_unregister_bars(vdev);
3695 }
3696 
3697 static void vfio_pci_reset(DeviceState *dev)
3698 {
3699     PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
3700     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev);
3701 
3702     trace_vfio_pci_reset(vdev->vbasedev.name);
3703 
3704     vfio_pci_pre_reset(vdev);
3705 
3706     if (vdev->resetfn && !vdev->resetfn(vdev)) {
3707         goto post_reset;
3708     }
3709 
3710     if (vdev->vbasedev.reset_works &&
3711         (vdev->has_flr || !vdev->has_pm_reset) &&
3712         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
3713         trace_vfio_pci_reset_flr(vdev->vbasedev.name);
3714         goto post_reset;
3715     }
3716 
3717     /* See if we can do our own bus reset */
3718     if (!vfio_pci_hot_reset_one(vdev)) {
3719         goto post_reset;
3720     }
3721 
3722     /* If nothing else works and the device supports PM reset, use it */
3723     if (vdev->vbasedev.reset_works && vdev->has_pm_reset &&
3724         !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) {
3725         trace_vfio_pci_reset_pm(vdev->vbasedev.name);
3726         goto post_reset;
3727     }
3728 
3729 post_reset:
3730     vfio_pci_post_reset(vdev);
3731 }
3732 
3733 static void vfio_instance_init(Object *obj)
3734 {
3735     PCIDevice *pci_dev = PCI_DEVICE(obj);
3736     VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, PCI_DEVICE(obj));
3737 
3738     device_add_bootindex_property(obj, &vdev->bootindex,
3739                                   "bootindex", NULL,
3740                                   &pci_dev->qdev, NULL);
3741 }
3742 
3743 static Property vfio_pci_dev_properties[] = {
3744     DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIOPCIDevice, host),
3745     DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIOPCIDevice,
3746                        intx.mmap_timeout, 1100),
3747     DEFINE_PROP_BIT("x-vga", VFIOPCIDevice, features,
3748                     VFIO_FEATURE_ENABLE_VGA_BIT, false),
3749     DEFINE_PROP_BIT("x-req", VFIOPCIDevice, features,
3750                     VFIO_FEATURE_ENABLE_REQ_BIT, true),
3751     DEFINE_PROP_BOOL("x-mmap", VFIOPCIDevice, vbasedev.allow_mmap, true),
3752     /*
3753      * TODO - support passed fds... is this necessary?
3754      * DEFINE_PROP_STRING("vfiofd", VFIOPCIDevice, vfiofd_name),
3755      * DEFINE_PROP_STRING("vfiogroupfd, VFIOPCIDevice, vfiogroupfd_name),
3756      */
3757     DEFINE_PROP_END_OF_LIST(),
3758 };
3759 
3760 static const VMStateDescription vfio_pci_vmstate = {
3761     .name = "vfio-pci",
3762     .unmigratable = 1,
3763 };
3764 
3765 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
3766 {
3767     DeviceClass *dc = DEVICE_CLASS(klass);
3768     PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
3769 
3770     dc->reset = vfio_pci_reset;
3771     dc->props = vfio_pci_dev_properties;
3772     dc->vmsd = &vfio_pci_vmstate;
3773     dc->desc = "VFIO-based PCI device assignment";
3774     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
3775     pdc->init = vfio_initfn;
3776     pdc->exit = vfio_exitfn;
3777     pdc->config_read = vfio_pci_read_config;
3778     pdc->config_write = vfio_pci_write_config;
3779     pdc->is_express = 1; /* We might be */
3780 }
3781 
3782 static const TypeInfo vfio_pci_dev_info = {
3783     .name = "vfio-pci",
3784     .parent = TYPE_PCI_DEVICE,
3785     .instance_size = sizeof(VFIOPCIDevice),
3786     .class_init = vfio_pci_dev_class_init,
3787     .instance_init = vfio_instance_init,
3788     .instance_finalize = vfio_instance_finalize,
3789 };
3790 
3791 static void register_vfio_pci_dev_type(void)
3792 {
3793     type_register_static(&vfio_pci_dev_info);
3794 }
3795 
3796 type_init(register_vfio_pci_dev_type)
3797