?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Replace all setjmp()/longjmp() with sigsetjmp()/siglongjmp()
[qemu/pbrook.git] / hw / vfio_pci.c
blobad9ae360b2f6222c37f32592494657d8d2d7401f
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 <dirent.h>
22 #include <unistd.h>
23 #include <sys/ioctl.h>
24 #include <sys/mman.h>
25 #include <sys/stat.h>
26 #include <sys/types.h>
27 #include <linux/vfio.h>
28
29 #include "config.h"
30 #include "qemu/event_notifier.h"
31 #include "exec/address-spaces.h"
32 #include "sysemu/kvm.h"
33 #include "exec/memory.h"
34 #include "pci/msi.h"
35 #include "pci/msix.h"
36 #include "pci/pci.h"
37 #include "qemu-common.h"
38 #include "qemu/error-report.h"
39 #include "qemu/queue.h"
40 #include "qemu/range.h"
41
42 /* #define DEBUG_VFIO */
43 #ifdef DEBUG_VFIO
44 #define DPRINTF(fmt, ...) \
45 do { fprintf(stderr, "vfio: " fmt, ## __VA_ARGS__); } while (0)
46 #else
47 #define DPRINTF(fmt, ...) \
48 do { } while (0)
49 #endif
50
51 typedef struct VFIOBAR {
52 off_t fd_offset; /* offset of BAR within device fd */
53 int fd; /* device fd, allows us to pass VFIOBAR as opaque data */
54 MemoryRegion mem; /* slow, read/write access */
55 MemoryRegion mmap_mem; /* direct mapped access */
56 void *mmap;
57 size_t size;
58 uint32_t flags; /* VFIO region flags (rd/wr/mmap) */
59 uint8_t nr; /* cache the BAR number for debug */
60 } VFIOBAR;
61
62 typedef struct VFIOINTx {
63 bool pending; /* interrupt pending */
64 bool kvm_accel; /* set when QEMU bypass through KVM enabled */
65 uint8_t pin; /* which pin to pull for qemu_set_irq */
66 EventNotifier interrupt; /* eventfd triggered on interrupt */
67 EventNotifier unmask; /* eventfd for unmask on QEMU bypass */
68 PCIINTxRoute route; /* routing info for QEMU bypass */
69 uint32_t mmap_timeout; /* delay to re-enable mmaps after interrupt */
70 QEMUTimer *mmap_timer; /* enable mmaps after periods w/o interrupts */
71 } VFIOINTx;
72
73 struct VFIODevice;
74
75 typedef struct VFIOMSIVector {
76 EventNotifier interrupt; /* eventfd triggered on interrupt */
77 struct VFIODevice *vdev; /* back pointer to device */
78 int virq; /* KVM irqchip route for QEMU bypass */
79 bool use;
80 } VFIOMSIVector;
81
82 enum {
83 VFIO_INT_NONE = 0,
84 VFIO_INT_INTx = 1,
85 VFIO_INT_MSI = 2,
86 VFIO_INT_MSIX = 3,
87 };
88
89 struct VFIOGroup;
90
91 typedef struct VFIOContainer {
92 int fd; /* /dev/vfio/vfio, empowered by the attached groups */
93 struct {
94 /* enable abstraction to support various iommu backends */
95 union {
96 MemoryListener listener; /* Used by type1 iommu */
97 };
98 void (*release)(struct VFIOContainer *);
99 } iommu_data;
100 QLIST_HEAD(, VFIOGroup) group_list;
101 QLIST_ENTRY(VFIOContainer) next;
102 } VFIOContainer;
103
104 /* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
105 typedef struct VFIOMSIXInfo {
106 uint8_t table_bar;
107 uint8_t pba_bar;
108 uint16_t entries;
109 uint32_t table_offset;
110 uint32_t pba_offset;
111 MemoryRegion mmap_mem;
112 void *mmap;
113 } VFIOMSIXInfo;
114
115 typedef struct VFIODevice {
116 PCIDevice pdev;
117 int fd;
118 VFIOINTx intx;
119 unsigned int config_size;
120 off_t config_offset; /* Offset of config space region within device fd */
121 unsigned int rom_size;
122 off_t rom_offset; /* Offset of ROM region within device fd */
123 int msi_cap_size;
124 VFIOMSIVector *msi_vectors;
125 VFIOMSIXInfo *msix;
126 int nr_vectors; /* Number of MSI/MSIX vectors currently in use */
127 int interrupt; /* Current interrupt type */
128 VFIOBAR bars[PCI_NUM_REGIONS - 1]; /* No ROM */
129 PCIHostDeviceAddress host;
130 QLIST_ENTRY(VFIODevice) next;
131 struct VFIOGroup *group;
132 bool reset_works;
133 } VFIODevice;
134
135 typedef struct VFIOGroup {
136 int fd;
137 int groupid;
138 VFIOContainer *container;
139 QLIST_HEAD(, VFIODevice) device_list;
140 QLIST_ENTRY(VFIOGroup) next;
141 QLIST_ENTRY(VFIOGroup) container_next;
142 } VFIOGroup;
143
144 #define MSIX_CAP_LENGTH 12
145
146 static QLIST_HEAD(, VFIOContainer)
147 container_list = QLIST_HEAD_INITIALIZER(container_list);
148
149 static QLIST_HEAD(, VFIOGroup)
150 group_list = QLIST_HEAD_INITIALIZER(group_list);
151
152 static void vfio_disable_interrupts(VFIODevice *vdev);
153 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len);
154 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled);
155
156 /*
157 * Common VFIO interrupt disable
158 */
159 static void vfio_disable_irqindex(VFIODevice *vdev, int index)
160 {
161 struct vfio_irq_set irq_set = {
162 .argsz = sizeof(irq_set),
163 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
164 .index = index,
165 .start = 0,
166 .count = 0,
167 };
168
169 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
170 }
171
172 /*
173 * INTx
174 */
175 static void vfio_unmask_intx(VFIODevice *vdev)
176 {
177 struct vfio_irq_set irq_set = {
178 .argsz = sizeof(irq_set),
179 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
180 .index = VFIO_PCI_INTX_IRQ_INDEX,
181 .start = 0,
182 .count = 1,
183 };
184
185 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
186 }
187
188 #ifdef CONFIG_KVM /* Unused outside of CONFIG_KVM code */
189 static void vfio_mask_intx(VFIODevice *vdev)
190 {
191 struct vfio_irq_set irq_set = {
192 .argsz = sizeof(irq_set),
193 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
194 .index = VFIO_PCI_INTX_IRQ_INDEX,
195 .start = 0,
196 .count = 1,
197 };
198
199 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
200 }
201 #endif
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 VFIODevice *vdev = opaque;
221
222 if (vdev->intx.pending) {
223 qemu_mod_timer(vdev->intx.mmap_timer,
224 qemu_get_clock_ms(vm_clock) + 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 VFIODevice *vdev = opaque;
234
235 if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
236 return;
237 }
238
239 DPRINTF("%s(%04x:%02x:%02x.%x) Pin %c\n", __func__, vdev->host.domain,
240 vdev->host.bus, vdev->host.slot, vdev->host.function,
241 'A' + vdev->intx.pin);
242
243 vdev->intx.pending = true;
244 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 1);
245 vfio_mmap_set_enabled(vdev, false);
246 if (vdev->intx.mmap_timeout) {
247 qemu_mod_timer(vdev->intx.mmap_timer,
248 qemu_get_clock_ms(vm_clock) + vdev->intx.mmap_timeout);
249 }
250 }
251
252 static void vfio_eoi(VFIODevice *vdev)
253 {
254 if (!vdev->intx.pending) {
255 return;
256 }
257
258 DPRINTF("%s(%04x:%02x:%02x.%x) EOI\n", __func__, vdev->host.domain,
259 vdev->host.bus, vdev->host.slot, vdev->host.function);
260
261 vdev->intx.pending = false;
262 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
263 vfio_unmask_intx(vdev);
264 }
265
266 static void vfio_enable_intx_kvm(VFIODevice *vdev)
267 {
268 #ifdef CONFIG_KVM
269 struct kvm_irqfd irqfd = {
270 .fd = event_notifier_get_fd(&vdev->intx.interrupt),
271 .gsi = vdev->intx.route.irq,
272 .flags = KVM_IRQFD_FLAG_RESAMPLE,
273 };
274 struct vfio_irq_set *irq_set;
275 int ret, argsz;
276 int32_t *pfd;
277
278 if (!kvm_irqfds_enabled() ||
279 vdev->intx.route.mode != PCI_INTX_ENABLED ||
280 !kvm_check_extension(kvm_state, KVM_CAP_IRQFD_RESAMPLE)) {
281 return;
282 }
283
284 /* Get to a known interrupt state */
285 qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
286 vfio_mask_intx(vdev);
287 vdev->intx.pending = false;
288 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
289
290 /* Get an eventfd for resample/unmask */
291 if (event_notifier_init(&vdev->intx.unmask, 0)) {
292 error_report("vfio: Error: event_notifier_init failed eoi");
293 goto fail;
294 }
295
296 /* KVM triggers it, VFIO listens for it */
297 irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
298
299 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
300 error_report("vfio: Error: Failed to setup resample irqfd: %m");
301 goto fail_irqfd;
302 }
303
304 argsz = sizeof(*irq_set) + sizeof(*pfd);
305
306 irq_set = g_malloc0(argsz);
307 irq_set->argsz = argsz;
308 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
309 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
310 irq_set->start = 0;
311 irq_set->count = 1;
312 pfd = (int32_t *)&irq_set->data;
313
314 *pfd = irqfd.resamplefd;
315
316 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
317 g_free(irq_set);
318 if (ret) {
319 error_report("vfio: Error: Failed to setup INTx unmask fd: %m");
320 goto fail_vfio;
321 }
322
323 /* Let'em rip */
324 vfio_unmask_intx(vdev);
325
326 vdev->intx.kvm_accel = true;
327
328 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel enabled\n",
329 __func__, vdev->host.domain, vdev->host.bus,
330 vdev->host.slot, vdev->host.function);
331
332 return;
333
334 fail_vfio:
335 irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
336 kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
337 fail_irqfd:
338 event_notifier_cleanup(&vdev->intx.unmask);
339 fail:
340 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
341 vfio_unmask_intx(vdev);
342 #endif
343 }
344
345 static void vfio_disable_intx_kvm(VFIODevice *vdev)
346 {
347 #ifdef CONFIG_KVM
348 struct kvm_irqfd irqfd = {
349 .fd = event_notifier_get_fd(&vdev->intx.interrupt),
350 .gsi = vdev->intx.route.irq,
351 .flags = KVM_IRQFD_FLAG_DEASSIGN,
352 };
353
354 if (!vdev->intx.kvm_accel) {
355 return;
356 }
357
358 /*
359 * Get to a known state, hardware masked, QEMU ready to accept new
360 * interrupts, QEMU IRQ de-asserted.
361 */
362 vfio_mask_intx(vdev);
363 vdev->intx.pending = false;
364 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
365
366 /* Tell KVM to stop listening for an INTx irqfd */
367 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
368 error_report("vfio: Error: Failed to disable INTx irqfd: %m");
369 }
370
371 /* We only need to close the eventfd for VFIO to cleanup the kernel side */
372 event_notifier_cleanup(&vdev->intx.unmask);
373
374 /* QEMU starts listening for interrupt events. */
375 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
376
377 vdev->intx.kvm_accel = false;
378
379 /* If we've missed an event, let it re-fire through QEMU */
380 vfio_unmask_intx(vdev);
381
382 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel disabled\n",
383 __func__, vdev->host.domain, vdev->host.bus,
384 vdev->host.slot, vdev->host.function);
385 #endif
386 }
387
388 static void vfio_update_irq(PCIDevice *pdev)
389 {
390 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
391 PCIINTxRoute route;
392
393 if (vdev->interrupt != VFIO_INT_INTx) {
394 return;
395 }
396
397 route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
398
399 if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
400 return; /* Nothing changed */
401 }
402
403 DPRINTF("%s(%04x:%02x:%02x.%x) IRQ moved %d -> %d\n", __func__,
404 vdev->host.domain, vdev->host.bus, vdev->host.slot,
405 vdev->host.function, vdev->intx.route.irq, route.irq);
406
407 vfio_disable_intx_kvm(vdev);
408
409 vdev->intx.route = route;
410
411 if (route.mode != PCI_INTX_ENABLED) {
412 return;
413 }
414
415 vfio_enable_intx_kvm(vdev);
416
417 /* Re-enable the interrupt in cased we missed an EOI */
418 vfio_eoi(vdev);
419 }
420
421 static int vfio_enable_intx(VFIODevice *vdev)
422 {
423 uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
424 int ret, argsz;
425 struct vfio_irq_set *irq_set;
426 int32_t *pfd;
427
428 if (!pin) {
429 return 0;
430 }
431
432 vfio_disable_interrupts(vdev);
433
434 vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
435
436 #ifdef CONFIG_KVM
437 /*
438 * Only conditional to avoid generating error messages on platforms
439 * where we won't actually use the result anyway.
440 */
441 if (kvm_irqfds_enabled() &&
442 kvm_check_extension(kvm_state, KVM_CAP_IRQFD_RESAMPLE)) {
443 vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
444 vdev->intx.pin);
445 }
446 #endif
447
448 ret = event_notifier_init(&vdev->intx.interrupt, 0);
449 if (ret) {
450 error_report("vfio: Error: event_notifier_init failed");
451 return ret;
452 }
453
454 argsz = sizeof(*irq_set) + sizeof(*pfd);
455
456 irq_set = g_malloc0(argsz);
457 irq_set->argsz = argsz;
458 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
459 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
460 irq_set->start = 0;
461 irq_set->count = 1;
462 pfd = (int32_t *)&irq_set->data;
463
464 *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
465 qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
466
467 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
468 g_free(irq_set);
469 if (ret) {
470 error_report("vfio: Error: Failed to setup INTx fd: %m");
471 qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
472 event_notifier_cleanup(&vdev->intx.interrupt);
473 return -errno;
474 }
475
476 vfio_enable_intx_kvm(vdev);
477
478 vdev->interrupt = VFIO_INT_INTx;
479
480 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
481 vdev->host.bus, vdev->host.slot, vdev->host.function);
482
483 return 0;
484 }
485
486 static void vfio_disable_intx(VFIODevice *vdev)
487 {
488 int fd;
489
490 qemu_del_timer(vdev->intx.mmap_timer);
491 vfio_disable_intx_kvm(vdev);
492 vfio_disable_irqindex(vdev, VFIO_PCI_INTX_IRQ_INDEX);
493 vdev->intx.pending = false;
494 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
495 vfio_mmap_set_enabled(vdev, true);
496
497 fd = event_notifier_get_fd(&vdev->intx.interrupt);
498 qemu_set_fd_handler(fd, NULL, NULL, vdev);
499 event_notifier_cleanup(&vdev->intx.interrupt);
500
501 vdev->interrupt = VFIO_INT_NONE;
502
503 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
504 vdev->host.bus, vdev->host.slot, vdev->host.function);
505 }
506
507 /*
508 * MSI/X
509 */
510 static void vfio_msi_interrupt(void *opaque)
511 {
512 VFIOMSIVector *vector = opaque;
513 VFIODevice *vdev = vector->vdev;
514 int nr = vector - vdev->msi_vectors;
515
516 if (!event_notifier_test_and_clear(&vector->interrupt)) {
517 return;
518 }
519
520 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d\n", __func__,
521 vdev->host.domain, vdev->host.bus, vdev->host.slot,
522 vdev->host.function, nr);
523
524 if (vdev->interrupt == VFIO_INT_MSIX) {
525 msix_notify(&vdev->pdev, nr);
526 } else if (vdev->interrupt == VFIO_INT_MSI) {
527 msi_notify(&vdev->pdev, nr);
528 } else {
529 error_report("vfio: MSI interrupt receieved, but not enabled?");
530 }
531 }
532
533 static int vfio_enable_vectors(VFIODevice *vdev, bool msix)
534 {
535 struct vfio_irq_set *irq_set;
536 int ret = 0, i, argsz;
537 int32_t *fds;
538
539 argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
540
541 irq_set = g_malloc0(argsz);
542 irq_set->argsz = argsz;
543 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
544 irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
545 irq_set->start = 0;
546 irq_set->count = vdev->nr_vectors;
547 fds = (int32_t *)&irq_set->data;
548
549 for (i = 0; i < vdev->nr_vectors; i++) {
550 if (!vdev->msi_vectors[i].use) {
551 fds[i] = -1;
552 continue;
553 }
554
555 fds[i] = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
556 }
557
558 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
559
560 g_free(irq_set);
561
562 return ret;
563 }
564
565 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
566 MSIMessage *msg, IOHandler *handler)
567 {
568 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
569 VFIOMSIVector *vector;
570 int ret;
571
572 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d used\n", __func__,
573 vdev->host.domain, vdev->host.bus, vdev->host.slot,
574 vdev->host.function, nr);
575
576 vector = &vdev->msi_vectors[nr];
577 vector->vdev = vdev;
578 vector->use = true;
579
580 msix_vector_use(pdev, nr);
581
582 if (event_notifier_init(&vector->interrupt, 0)) {
583 error_report("vfio: Error: event_notifier_init failed");
584 }
585
586 /*
587 * Attempt to enable route through KVM irqchip,
588 * default to userspace handling if unavailable.
589 */
590 vector->virq = msg ? kvm_irqchip_add_msi_route(kvm_state, *msg) : -1;
591 if (vector->virq < 0 ||
592 kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
593 vector->virq) < 0) {
594 if (vector->virq >= 0) {
595 kvm_irqchip_release_virq(kvm_state, vector->virq);
596 vector->virq = -1;
597 }
598 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
599 handler, NULL, vector);
600 }
601
602 /*
603 * We don't want to have the host allocate all possible MSI vectors
604 * for a device if they're not in use, so we shutdown and incrementally
605 * increase them as needed.
606 */
607 if (vdev->nr_vectors < nr + 1) {
608 vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
609 vdev->nr_vectors = nr + 1;
610 ret = vfio_enable_vectors(vdev, true);
611 if (ret) {
612 error_report("vfio: failed to enable vectors, %d", ret);
613 }
614 } else {
615 int argsz;
616 struct vfio_irq_set *irq_set;
617 int32_t *pfd;
618
619 argsz = sizeof(*irq_set) + sizeof(*pfd);
620
621 irq_set = g_malloc0(argsz);
622 irq_set->argsz = argsz;
623 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
624 VFIO_IRQ_SET_ACTION_TRIGGER;
625 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
626 irq_set->start = nr;
627 irq_set->count = 1;
628 pfd = (int32_t *)&irq_set->data;
629
630 *pfd = event_notifier_get_fd(&vector->interrupt);
631
632 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
633 g_free(irq_set);
634 if (ret) {
635 error_report("vfio: failed to modify vector, %d", ret);
636 }
637 }
638
639 return 0;
640 }
641
642 static int vfio_msix_vector_use(PCIDevice *pdev,
643 unsigned int nr, MSIMessage msg)
644 {
645 return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
646 }
647
648 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
649 {
650 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
651 VFIOMSIVector *vector = &vdev->msi_vectors[nr];
652 int argsz;
653 struct vfio_irq_set *irq_set;
654 int32_t *pfd;
655
656 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d released\n", __func__,
657 vdev->host.domain, vdev->host.bus, vdev->host.slot,
658 vdev->host.function, nr);
659
660 /*
661 * XXX What's the right thing to do here? This turns off the interrupt
662 * completely, but do we really just want to switch the interrupt to
663 * bouncing through userspace and let msix.c drop it? Not sure.
664 */
665 msix_vector_unuse(pdev, nr);
666
667 argsz = sizeof(*irq_set) + sizeof(*pfd);
668
669 irq_set = g_malloc0(argsz);
670 irq_set->argsz = argsz;
671 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
672 VFIO_IRQ_SET_ACTION_TRIGGER;
673 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
674 irq_set->start = nr;
675 irq_set->count = 1;
676 pfd = (int32_t *)&irq_set->data;
677
678 *pfd = -1;
679
680 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
681
682 g_free(irq_set);
683
684 if (vector->virq < 0) {
685 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
686 NULL, NULL, NULL);
687 } else {
688 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
689 vector->virq);
690 kvm_irqchip_release_virq(kvm_state, vector->virq);
691 vector->virq = -1;
692 }
693
694 event_notifier_cleanup(&vector->interrupt);
695 vector->use = false;
696 }
697
698 static void vfio_enable_msix(VFIODevice *vdev)
699 {
700 vfio_disable_interrupts(vdev);
701
702 vdev->msi_vectors = g_malloc0(vdev->msix->entries * sizeof(VFIOMSIVector));
703
704 vdev->interrupt = VFIO_INT_MSIX;
705
706 /*
707 * Some communication channels between VF & PF or PF & fw rely on the
708 * physical state of the device and expect that enabling MSI-X from the
709 * guest enables the same on the host. When our guest is Linux, the
710 * guest driver call to pci_enable_msix() sets the enabling bit in the
711 * MSI-X capability, but leaves the vector table masked. We therefore
712 * can't rely on a vector_use callback (from request_irq() in the guest)
713 * to switch the physical device into MSI-X mode because that may come a
714 * long time after pci_enable_msix(). This code enables vector 0 with
715 * triggering to userspace, then immediately release the vector, leaving
716 * the physical device with no vectors enabled, but MSI-X enabled, just
717 * like the guest view.
718 */
719 vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
720 vfio_msix_vector_release(&vdev->pdev, 0);
721
722 if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
723 vfio_msix_vector_release, NULL)) {
724 error_report("vfio: msix_set_vector_notifiers failed");
725 }
726
727 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
728 vdev->host.bus, vdev->host.slot, vdev->host.function);
729 }
730
731 static void vfio_enable_msi(VFIODevice *vdev)
732 {
733 int ret, i;
734
735 vfio_disable_interrupts(vdev);
736
737 vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
738 retry:
739 vdev->msi_vectors = g_malloc0(vdev->nr_vectors * sizeof(VFIOMSIVector));
740
741 for (i = 0; i < vdev->nr_vectors; i++) {
742 MSIMessage msg;
743 VFIOMSIVector *vector = &vdev->msi_vectors[i];
744
745 vector->vdev = vdev;
746 vector->use = true;
747
748 if (event_notifier_init(&vector->interrupt, 0)) {
749 error_report("vfio: Error: event_notifier_init failed");
750 }
751
752 msg = msi_get_message(&vdev->pdev, i);
753
754 /*
755 * Attempt to enable route through KVM irqchip,
756 * default to userspace handling if unavailable.
757 */
758 vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg);
759 if (vector->virq < 0 ||
760 kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
761 vector->virq) < 0) {
762 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
763 vfio_msi_interrupt, NULL, vector);
764 }
765 }
766
767 ret = vfio_enable_vectors(vdev, false);
768 if (ret) {
769 if (ret < 0) {
770 error_report("vfio: Error: Failed to setup MSI fds: %m");
771 } else if (ret != vdev->nr_vectors) {
772 error_report("vfio: Error: Failed to enable %d "
773 "MSI vectors, retry with %d", vdev->nr_vectors, ret);
774 }
775
776 for (i = 0; i < vdev->nr_vectors; i++) {
777 VFIOMSIVector *vector = &vdev->msi_vectors[i];
778 if (vector->virq >= 0) {
779 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
780 vector->virq);
781 kvm_irqchip_release_virq(kvm_state, vector->virq);
782 vector->virq = -1;
783 } else {
784 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
785 NULL, NULL, NULL);
786 }
787 event_notifier_cleanup(&vector->interrupt);
788 }
789
790 g_free(vdev->msi_vectors);
791
792 if (ret > 0 && ret != vdev->nr_vectors) {
793 vdev->nr_vectors = ret;
794 goto retry;
795 }
796 vdev->nr_vectors = 0;
797
798 return;
799 }
800
801 vdev->interrupt = VFIO_INT_MSI;
802
803 DPRINTF("%s(%04x:%02x:%02x.%x) Enabled %d MSI vectors\n", __func__,
804 vdev->host.domain, vdev->host.bus, vdev->host.slot,
805 vdev->host.function, vdev->nr_vectors);
806 }
807
808 static void vfio_disable_msi_common(VFIODevice *vdev)
809 {
810 g_free(vdev->msi_vectors);
811 vdev->msi_vectors = NULL;
812 vdev->nr_vectors = 0;
813 vdev->interrupt = VFIO_INT_NONE;
814
815 vfio_enable_intx(vdev);
816 }
817
818 static void vfio_disable_msix(VFIODevice *vdev)
819 {
820 msix_unset_vector_notifiers(&vdev->pdev);
821
822 if (vdev->nr_vectors) {
823 vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
824 }
825
826 vfio_disable_msi_common(vdev);
827
828 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
829 vdev->host.bus, vdev->host.slot, vdev->host.function);
830 }
831
832 static void vfio_disable_msi(VFIODevice *vdev)
833 {
834 int i;
835
836 vfio_disable_irqindex(vdev, VFIO_PCI_MSI_IRQ_INDEX);
837
838 for (i = 0; i < vdev->nr_vectors; i++) {
839 VFIOMSIVector *vector = &vdev->msi_vectors[i];
840
841 if (!vector->use) {
842 continue;
843 }
844
845 if (vector->virq >= 0) {
846 kvm_irqchip_remove_irqfd_notifier(kvm_state,
847 &vector->interrupt, vector->virq);
848 kvm_irqchip_release_virq(kvm_state, vector->virq);
849 vector->virq = -1;
850 } else {
851 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
852 NULL, NULL, NULL);
853 }
854
855 event_notifier_cleanup(&vector->interrupt);
856 }
857
858 vfio_disable_msi_common(vdev);
859
860 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
861 vdev->host.bus, vdev->host.slot, vdev->host.function);
862 }
863
864 /*
865 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
866 */
867 static void vfio_bar_write(void *opaque, hwaddr addr,
868 uint64_t data, unsigned size)
869 {
870 VFIOBAR *bar = opaque;
871 union {
872 uint8_t byte;
873 uint16_t word;
874 uint32_t dword;
875 uint64_t qword;
876 } buf;
877
878 switch (size) {
879 case 1:
880 buf.byte = data;
881 break;
882 case 2:
883 buf.word = cpu_to_le16(data);
884 break;
885 case 4:
886 buf.dword = cpu_to_le32(data);
887 break;
888 default:
889 hw_error("vfio: unsupported write size, %d bytes\n", size);
890 break;
891 }
892
893 if (pwrite(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
894 error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
895 __func__, addr, data, size);
896 }
897
898 DPRINTF("%s(BAR%d+0x%"HWADDR_PRIx", 0x%"PRIx64", %d)\n",
899 __func__, bar->nr, addr, data, size);
900
901 /*
902 * A read or write to a BAR always signals an INTx EOI. This will
903 * do nothing if not pending (including not in INTx mode). We assume
904 * that a BAR access is in response to an interrupt and that BAR
905 * accesses will service the interrupt. Unfortunately, we don't know
906 * which access will service the interrupt, so we're potentially
907 * getting quite a few host interrupts per guest interrupt.
908 */
909 vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
910 }
911
912 static uint64_t vfio_bar_read(void *opaque,
913 hwaddr addr, unsigned size)
914 {
915 VFIOBAR *bar = opaque;
916 union {
917 uint8_t byte;
918 uint16_t word;
919 uint32_t dword;
920 uint64_t qword;
921 } buf;
922 uint64_t data = 0;
923
924 if (pread(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
925 error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
926 __func__, addr, size);
927 return (uint64_t)-1;
928 }
929
930 switch (size) {
931 case 1:
932 data = buf.byte;
933 break;
934 case 2:
935 data = le16_to_cpu(buf.word);
936 break;
937 case 4:
938 data = le32_to_cpu(buf.dword);
939 break;
940 default:
941 hw_error("vfio: unsupported read size, %d bytes\n", size);
942 break;
943 }
944
945 DPRINTF("%s(BAR%d+0x%"HWADDR_PRIx", %d) = 0x%"PRIx64"\n",
946 __func__, bar->nr, addr, size, data);
947
948 /* Same as write above */
949 vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
950
951 return data;
952 }
953
954 static const MemoryRegionOps vfio_bar_ops = {
955 .read = vfio_bar_read,
956 .write = vfio_bar_write,
957 .endianness = DEVICE_LITTLE_ENDIAN,
958 };
959
960 /*
961 * PCI config space
962 */
963 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
964 {
965 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
966 uint32_t val = 0;
967
968 /*
969 * We only need QEMU PCI config support for the ROM BAR, the MSI and MSIX
970 * capabilities, and the multifunction bit below. We let VFIO handle
971 * virtualizing everything else. Performance is not a concern here.
972 */
973 if (ranges_overlap(addr, len, PCI_ROM_ADDRESS, 4) ||
974 (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
975 ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) ||
976 (pdev->cap_present & QEMU_PCI_CAP_MSI &&
977 ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size))) {
978
979 val = pci_default_read_config(pdev, addr, len);
980 } else {
981 if (pread(vdev->fd, &val, len, vdev->config_offset + addr) != len) {
982 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m",
983 __func__, vdev->host.domain, vdev->host.bus,
984 vdev->host.slot, vdev->host.function, addr, len);
985 return -errno;
986 }
987 val = le32_to_cpu(val);
988 }
989
990 /* Multifunction bit is virualized in QEMU */
991 if (unlikely(ranges_overlap(addr, len, PCI_HEADER_TYPE, 1))) {
992 uint32_t mask = PCI_HEADER_TYPE_MULTI_FUNCTION;
993
994 if (len == 4) {
995 mask <<= 16;
996 }
997
998 if (pdev->cap_present & QEMU_PCI_CAP_MULTIFUNCTION) {
999 val |= mask;
1000 } else {
1001 val &= ~mask;
1002 }
1003 }
1004
1005 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, len=0x%x) %x\n", __func__,
1006 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1007 vdev->host.function, addr, len, val);
1008
1009 return val;
1010 }
1011
1012 static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
1013 uint32_t val, int len)
1014 {
1015 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1016 uint32_t val_le = cpu_to_le32(val);
1017
1018 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, 0x%x, len=0x%x)\n", __func__,
1019 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1020 vdev->host.function, addr, val, len);
1021
1022 /* Write everything to VFIO, let it filter out what we can't write */
1023 if (pwrite(vdev->fd, &val_le, len, vdev->config_offset + addr) != len) {
1024 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m",
1025 __func__, vdev->host.domain, vdev->host.bus,
1026 vdev->host.slot, vdev->host.function, addr, val, len);
1027 }
1028
1029 /* Write standard header bits to emulation */
1030 if (addr < PCI_CONFIG_HEADER_SIZE) {
1031 pci_default_write_config(pdev, addr, val, len);
1032 return;
1033 }
1034
1035 /* MSI/MSI-X Enabling/Disabling */
1036 if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
1037 ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
1038 int is_enabled, was_enabled = msi_enabled(pdev);
1039
1040 pci_default_write_config(pdev, addr, val, len);
1041
1042 is_enabled = msi_enabled(pdev);
1043
1044 if (!was_enabled && is_enabled) {
1045 vfio_enable_msi(vdev);
1046 } else if (was_enabled && !is_enabled) {
1047 vfio_disable_msi(vdev);
1048 }
1049 }
1050
1051 if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
1052 ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
1053 int is_enabled, was_enabled = msix_enabled(pdev);
1054
1055 pci_default_write_config(pdev, addr, val, len);
1056
1057 is_enabled = msix_enabled(pdev);
1058
1059 if (!was_enabled && is_enabled) {
1060 vfio_enable_msix(vdev);
1061 } else if (was_enabled && !is_enabled) {
1062 vfio_disable_msix(vdev);
1063 }
1064 }
1065 }
1066
1067 /*
1068 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
1069 */
1070 static int vfio_dma_unmap(VFIOContainer *container,
1071 hwaddr iova, ram_addr_t size)
1072 {
1073 struct vfio_iommu_type1_dma_unmap unmap = {
1074 .argsz = sizeof(unmap),
1075 .flags = 0,
1076 .iova = iova,
1077 .size = size,
1078 };
1079
1080 if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
1081 DPRINTF("VFIO_UNMAP_DMA: %d\n", -errno);
1082 return -errno;
1083 }
1084
1085 return 0;
1086 }
1087
1088 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
1089 ram_addr_t size, void *vaddr, bool readonly)
1090 {
1091 struct vfio_iommu_type1_dma_map map = {
1092 .argsz = sizeof(map),
1093 .flags = VFIO_DMA_MAP_FLAG_READ,
1094 .vaddr = (__u64)(uintptr_t)vaddr,
1095 .iova = iova,
1096 .size = size,
1097 };
1098
1099 if (!readonly) {
1100 map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
1101 }
1102
1103 /*
1104 * Try the mapping, if it fails with EBUSY, unmap the region and try
1105 * again. This shouldn't be necessary, but we sometimes see it in
1106 * the the VGA ROM space.
1107 */
1108 if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
1109 (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
1110 ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
1111 return 0;
1112 }
1113
1114 DPRINTF("VFIO_MAP_DMA: %d\n", -errno);
1115 return -errno;
1116 }
1117
1118 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
1119 {
1120 return !memory_region_is_ram(section->mr);
1121 }
1122
1123 static void vfio_listener_region_add(MemoryListener *listener,
1124 MemoryRegionSection *section)
1125 {
1126 VFIOContainer *container = container_of(listener, VFIOContainer,
1127 iommu_data.listener);
1128 hwaddr iova, end;
1129 void *vaddr;
1130 int ret;
1131
1132 if (vfio_listener_skipped_section(section)) {
1133 DPRINTF("vfio: SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
1134 section->offset_within_address_space,
1135 section->offset_within_address_space + section->size - 1);
1136 return;
1137 }
1138
1139 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
1140 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
1141 error_report("%s received unaligned region", __func__);
1142 return;
1143 }
1144
1145 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
1146 end = (section->offset_within_address_space + section->size) &
1147 TARGET_PAGE_MASK;
1148
1149 if (iova >= end) {
1150 return;
1151 }
1152
1153 vaddr = memory_region_get_ram_ptr(section->mr) +
1154 section->offset_within_region +
1155 (iova - section->offset_within_address_space);
1156
1157 DPRINTF("vfio: region_add %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
1158 iova, end - 1, vaddr);
1159
1160 ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
1161 if (ret) {
1162 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
1163 "0x%"HWADDR_PRIx", %p) = %d (%m)",
1164 container, iova, end - iova, vaddr, ret);
1165 }
1166 }
1167
1168 static void vfio_listener_region_del(MemoryListener *listener,
1169 MemoryRegionSection *section)
1170 {
1171 VFIOContainer *container = container_of(listener, VFIOContainer,
1172 iommu_data.listener);
1173 hwaddr iova, end;
1174 int ret;
1175
1176 if (vfio_listener_skipped_section(section)) {
1177 DPRINTF("vfio: SKIPPING region_del %"HWADDR_PRIx" - %"PRIx64"\n",
1178 section->offset_within_address_space,
1179 section->offset_within_address_space + section->size - 1);
1180 return;
1181 }
1182
1183 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
1184 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
1185 error_report("%s received unaligned region", __func__);
1186 return;
1187 }
1188
1189 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
1190 end = (section->offset_within_address_space + section->size) &
1191 TARGET_PAGE_MASK;
1192
1193 if (iova >= end) {
1194 return;
1195 }
1196
1197 DPRINTF("vfio: region_del %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
1198 iova, end - 1);
1199
1200 ret = vfio_dma_unmap(container, iova, end - iova);
1201 if (ret) {
1202 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
1203 "0x%"HWADDR_PRIx") = %d (%m)",
1204 container, iova, end - iova, ret);
1205 }
1206 }
1207
1208 static MemoryListener vfio_memory_listener = {
1209 .region_add = vfio_listener_region_add,
1210 .region_del = vfio_listener_region_del,
1211 };
1212
1213 static void vfio_listener_release(VFIOContainer *container)
1214 {
1215 memory_listener_unregister(&container->iommu_data.listener);
1216 }
1217
1218 /*
1219 * Interrupt setup
1220 */
1221 static void vfio_disable_interrupts(VFIODevice *vdev)
1222 {
1223 switch (vdev->interrupt) {
1224 case VFIO_INT_INTx:
1225 vfio_disable_intx(vdev);
1226 break;
1227 case VFIO_INT_MSI:
1228 vfio_disable_msi(vdev);
1229 break;
1230 case VFIO_INT_MSIX:
1231 vfio_disable_msix(vdev);
1232 break;
1233 }
1234 }
1235
1236 static int vfio_setup_msi(VFIODevice *vdev, int pos)
1237 {
1238 uint16_t ctrl;
1239 bool msi_64bit, msi_maskbit;
1240 int ret, entries;
1241
1242 if (pread(vdev->fd, &ctrl, sizeof(ctrl),
1243 vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1244 return -errno;
1245 }
1246 ctrl = le16_to_cpu(ctrl);
1247
1248 msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
1249 msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
1250 entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
1251
1252 DPRINTF("%04x:%02x:%02x.%x PCI MSI CAP @0x%x\n", vdev->host.domain,
1253 vdev->host.bus, vdev->host.slot, vdev->host.function, pos);
1254
1255 ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit);
1256 if (ret < 0) {
1257 if (ret == -ENOTSUP) {
1258 return 0;
1259 }
1260 error_report("vfio: msi_init failed");
1261 return ret;
1262 }
1263 vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
1264
1265 return 0;
1266 }
1267
1268 /*
1269 * We don't have any control over how pci_add_capability() inserts
1270 * capabilities into the chain. In order to setup MSI-X we need a
1271 * MemoryRegion for the BAR. In order to setup the BAR and not
1272 * attempt to mmap the MSI-X table area, which VFIO won't allow, we
1273 * need to first look for where the MSI-X table lives. So we
1274 * unfortunately split MSI-X setup across two functions.
1275 */
1276 static int vfio_early_setup_msix(VFIODevice *vdev)
1277 {
1278 uint8_t pos;
1279 uint16_t ctrl;
1280 uint32_t table, pba;
1281
1282 pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
1283 if (!pos) {
1284 return 0;
1285 }
1286
1287 if (pread(vdev->fd, &ctrl, sizeof(ctrl),
1288 vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1289 return -errno;
1290 }
1291
1292 if (pread(vdev->fd, &table, sizeof(table),
1293 vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
1294 return -errno;
1295 }
1296
1297 if (pread(vdev->fd, &pba, sizeof(pba),
1298 vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
1299 return -errno;
1300 }
1301
1302 ctrl = le16_to_cpu(ctrl);
1303 table = le32_to_cpu(table);
1304 pba = le32_to_cpu(pba);
1305
1306 vdev->msix = g_malloc0(sizeof(*(vdev->msix)));
1307 vdev->msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
1308 vdev->msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
1309 vdev->msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
1310 vdev->msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
1311 vdev->msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
1312
1313 DPRINTF("%04x:%02x:%02x.%x "
1314 "PCI MSI-X CAP @0x%x, BAR %d, offset 0x%x, entries %d\n",
1315 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1316 vdev->host.function, pos, vdev->msix->table_bar,
1317 vdev->msix->table_offset, vdev->msix->entries);
1318
1319 return 0;
1320 }
1321
1322 static int vfio_setup_msix(VFIODevice *vdev, int pos)
1323 {
1324 int ret;
1325
1326 ret = msix_init(&vdev->pdev, vdev->msix->entries,
1327 &vdev->bars[vdev->msix->table_bar].mem,
1328 vdev->msix->table_bar, vdev->msix->table_offset,
1329 &vdev->bars[vdev->msix->pba_bar].mem,
1330 vdev->msix->pba_bar, vdev->msix->pba_offset, pos);
1331 if (ret < 0) {
1332 if (ret == -ENOTSUP) {
1333 return 0;
1334 }
1335 error_report("vfio: msix_init failed");
1336 return ret;
1337 }
1338
1339 return 0;
1340 }
1341
1342 static void vfio_teardown_msi(VFIODevice *vdev)
1343 {
1344 msi_uninit(&vdev->pdev);
1345
1346 if (vdev->msix) {
1347 msix_uninit(&vdev->pdev, &vdev->bars[vdev->msix->table_bar].mem,
1348 &vdev->bars[vdev->msix->pba_bar].mem);
1349 }
1350 }
1351
1352 /*
1353 * Resource setup
1354 */
1355 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled)
1356 {
1357 int i;
1358
1359 for (i = 0; i < PCI_ROM_SLOT; i++) {
1360 VFIOBAR *bar = &vdev->bars[i];
1361
1362 if (!bar->size) {
1363 continue;
1364 }
1365
1366 memory_region_set_enabled(&bar->mmap_mem, enabled);
1367 if (vdev->msix && vdev->msix->table_bar == i) {
1368 memory_region_set_enabled(&vdev->msix->mmap_mem, enabled);
1369 }
1370 }
1371 }
1372
1373 static void vfio_unmap_bar(VFIODevice *vdev, int nr)
1374 {
1375 VFIOBAR *bar = &vdev->bars[nr];
1376
1377 if (!bar->size) {
1378 return;
1379 }
1380
1381 memory_region_del_subregion(&bar->mem, &bar->mmap_mem);
1382 munmap(bar->mmap, memory_region_size(&bar->mmap_mem));
1383
1384 if (vdev->msix && vdev->msix->table_bar == nr) {
1385 memory_region_del_subregion(&bar->mem, &vdev->msix->mmap_mem);
1386 munmap(vdev->msix->mmap, memory_region_size(&vdev->msix->mmap_mem));
1387 }
1388
1389 memory_region_destroy(&bar->mem);
1390 }
1391
1392 static int vfio_mmap_bar(VFIOBAR *bar, MemoryRegion *mem, MemoryRegion *submem,
1393 void **map, size_t size, off_t offset,
1394 const char *name)
1395 {
1396 int ret = 0;
1397
1398 if (size && bar->flags & VFIO_REGION_INFO_FLAG_MMAP) {
1399 int prot = 0;
1400
1401 if (bar->flags & VFIO_REGION_INFO_FLAG_READ) {
1402 prot |= PROT_READ;
1403 }
1404
1405 if (bar->flags & VFIO_REGION_INFO_FLAG_WRITE) {
1406 prot |= PROT_WRITE;
1407 }
1408
1409 *map = mmap(NULL, size, prot, MAP_SHARED,
1410 bar->fd, bar->fd_offset + offset);
1411 if (*map == MAP_FAILED) {
1412 *map = NULL;
1413 ret = -errno;
1414 goto empty_region;
1415 }
1416
1417 memory_region_init_ram_ptr(submem, name, size, *map);
1418 } else {
1419 empty_region:
1420 /* Create a zero sized sub-region to make cleanup easy. */
1421 memory_region_init(submem, name, 0);
1422 }
1423
1424 memory_region_add_subregion(mem, offset, submem);
1425
1426 return ret;
1427 }
1428
1429 static void vfio_map_bar(VFIODevice *vdev, int nr)
1430 {
1431 VFIOBAR *bar = &vdev->bars[nr];
1432 unsigned size = bar->size;
1433 char name[64];
1434 uint32_t pci_bar;
1435 uint8_t type;
1436 int ret;
1437
1438 /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1439 if (!size) {
1440 return;
1441 }
1442
1443 snprintf(name, sizeof(name), "VFIO %04x:%02x:%02x.%x BAR %d",
1444 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1445 vdev->host.function, nr);
1446
1447 /* Determine what type of BAR this is for registration */
1448 ret = pread(vdev->fd, &pci_bar, sizeof(pci_bar),
1449 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
1450 if (ret != sizeof(pci_bar)) {
1451 error_report("vfio: Failed to read BAR %d (%m)", nr);
1452 return;
1453 }
1454
1455 pci_bar = le32_to_cpu(pci_bar);
1456 type = pci_bar & (pci_bar & PCI_BASE_ADDRESS_SPACE_IO ?
1457 ~PCI_BASE_ADDRESS_IO_MASK : ~PCI_BASE_ADDRESS_MEM_MASK);
1458
1459 /* A "slow" read/write mapping underlies all BARs */
1460 memory_region_init_io(&bar->mem, &vfio_bar_ops, bar, name, size);
1461 pci_register_bar(&vdev->pdev, nr, type, &bar->mem);
1462
1463 /*
1464 * We can't mmap areas overlapping the MSIX vector table, so we
1465 * potentially insert a direct-mapped subregion before and after it.
1466 */
1467 if (vdev->msix && vdev->msix->table_bar == nr) {
1468 size = vdev->msix->table_offset & TARGET_PAGE_MASK;
1469 }
1470
1471 strncat(name, " mmap", sizeof(name) - strlen(name) - 1);
1472 if (vfio_mmap_bar(bar, &bar->mem,
1473 &bar->mmap_mem, &bar->mmap, size, 0, name)) {
1474 error_report("%s unsupported. Performance may be slow", name);
1475 }
1476
1477 if (vdev->msix && vdev->msix->table_bar == nr) {
1478 unsigned start;
1479
1480 start = TARGET_PAGE_ALIGN(vdev->msix->table_offset +
1481 (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
1482
1483 size = start < bar->size ? bar->size - start : 0;
1484 strncat(name, " msix-hi", sizeof(name) - strlen(name) - 1);
1485 /* VFIOMSIXInfo contains another MemoryRegion for this mapping */
1486 if (vfio_mmap_bar(bar, &bar->mem, &vdev->msix->mmap_mem,
1487 &vdev->msix->mmap, size, start, name)) {
1488 error_report("%s unsupported. Performance may be slow", name);
1489 }
1490 }
1491 }
1492
1493 static void vfio_map_bars(VFIODevice *vdev)
1494 {
1495 int i;
1496
1497 for (i = 0; i < PCI_ROM_SLOT; i++) {
1498 vfio_map_bar(vdev, i);
1499 }
1500 }
1501
1502 static void vfio_unmap_bars(VFIODevice *vdev)
1503 {
1504 int i;
1505
1506 for (i = 0; i < PCI_ROM_SLOT; i++) {
1507 vfio_unmap_bar(vdev, i);
1508 }
1509 }
1510
1511 /*
1512 * General setup
1513 */
1514 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
1515 {
1516 uint8_t tmp, next = 0xff;
1517
1518 for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
1519 tmp = pdev->config[tmp + 1]) {
1520 if (tmp > pos && tmp < next) {
1521 next = tmp;
1522 }
1523 }
1524
1525 return next - pos;
1526 }
1527
1528 static int vfio_add_std_cap(VFIODevice *vdev, uint8_t pos)
1529 {
1530 PCIDevice *pdev = &vdev->pdev;
1531 uint8_t cap_id, next, size;
1532 int ret;
1533
1534 cap_id = pdev->config[pos];
1535 next = pdev->config[pos + 1];
1536
1537 /*
1538 * If it becomes important to configure capabilities to their actual
1539 * size, use this as the default when it's something we don't recognize.
1540 * Since QEMU doesn't actually handle many of the config accesses,
1541 * exact size doesn't seem worthwhile.
1542 */
1543 size = vfio_std_cap_max_size(pdev, pos);
1544
1545 /*
1546 * pci_add_capability always inserts the new capability at the head
1547 * of the chain. Therefore to end up with a chain that matches the
1548 * physical device, we insert from the end by making this recursive.
1549 * This is also why we pre-caclulate size above as cached config space
1550 * will be changed as we unwind the stack.
1551 */
1552 if (next) {
1553 ret = vfio_add_std_cap(vdev, next);
1554 if (ret) {
1555 return ret;
1556 }
1557 } else {
1558 pdev->config[PCI_CAPABILITY_LIST] = 0; /* Begin the rebuild */
1559 }
1560
1561 switch (cap_id) {
1562 case PCI_CAP_ID_MSI:
1563 ret = vfio_setup_msi(vdev, pos);
1564 break;
1565 case PCI_CAP_ID_MSIX:
1566 ret = vfio_setup_msix(vdev, pos);
1567 break;
1568 default:
1569 ret = pci_add_capability(pdev, cap_id, pos, size);
1570 break;
1571 }
1572
1573 if (ret < 0) {
1574 error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability "
1575 "0x%x[0x%x]@0x%x: %d", vdev->host.domain,
1576 vdev->host.bus, vdev->host.slot, vdev->host.function,
1577 cap_id, size, pos, ret);
1578 return ret;
1579 }
1580
1581 return 0;
1582 }
1583
1584 static int vfio_add_capabilities(VFIODevice *vdev)
1585 {
1586 PCIDevice *pdev = &vdev->pdev;
1587
1588 if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
1589 !pdev->config[PCI_CAPABILITY_LIST]) {
1590 return 0; /* Nothing to add */
1591 }
1592
1593 return vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST]);
1594 }
1595
1596 static int vfio_load_rom(VFIODevice *vdev)
1597 {
1598 uint64_t size = vdev->rom_size;
1599 char name[32];
1600 off_t off = 0, voff = vdev->rom_offset;
1601 ssize_t bytes;
1602 void *ptr;
1603
1604 /* If loading ROM from file, pci handles it */
1605 if (vdev->pdev.romfile || !vdev->pdev.rom_bar || !size) {
1606 return 0;
1607 }
1608
1609 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
1610 vdev->host.bus, vdev->host.slot, vdev->host.function);
1611
1612 snprintf(name, sizeof(name), "vfio[%04x:%02x:%02x.%x].rom",
1613 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1614 vdev->host.function);
1615 memory_region_init_ram(&vdev->pdev.rom, name, size);
1616 ptr = memory_region_get_ram_ptr(&vdev->pdev.rom);
1617 memset(ptr, 0xff, size);
1618
1619 while (size) {
1620 bytes = pread(vdev->fd, ptr + off, size, voff + off);
1621 if (bytes == 0) {
1622 break; /* expect that we could get back less than the ROM BAR */
1623 } else if (bytes > 0) {
1624 off += bytes;
1625 size -= bytes;
1626 } else {
1627 if (errno == EINTR || errno == EAGAIN) {
1628 continue;
1629 }
1630 error_report("vfio: Error reading device ROM: %m");
1631 memory_region_destroy(&vdev->pdev.rom);
1632 return -errno;
1633 }
1634 }
1635
1636 pci_register_bar(&vdev->pdev, PCI_ROM_SLOT, 0, &vdev->pdev.rom);
1637 vdev->pdev.has_rom = true;
1638 return 0;
1639 }
1640
1641 static int vfio_connect_container(VFIOGroup *group)
1642 {
1643 VFIOContainer *container;
1644 int ret, fd;
1645
1646 if (group->container) {
1647 return 0;
1648 }
1649
1650 QLIST_FOREACH(container, &container_list, next) {
1651 if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
1652 group->container = container;
1653 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1654 return 0;
1655 }
1656 }
1657
1658 fd = qemu_open("/dev/vfio/vfio", O_RDWR);
1659 if (fd < 0) {
1660 error_report("vfio: failed to open /dev/vfio/vfio: %m");
1661 return -errno;
1662 }
1663
1664 ret = ioctl(fd, VFIO_GET_API_VERSION);
1665 if (ret != VFIO_API_VERSION) {
1666 error_report("vfio: supported vfio version: %d, "
1667 "reported version: %d", VFIO_API_VERSION, ret);
1668 close(fd);
1669 return -EINVAL;
1670 }
1671
1672 container = g_malloc0(sizeof(*container));
1673 container->fd = fd;
1674
1675 if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) {
1676 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1677 if (ret) {
1678 error_report("vfio: failed to set group container: %m");
1679 g_free(container);
1680 close(fd);
1681 return -errno;
1682 }
1683
1684 ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU);
1685 if (ret) {
1686 error_report("vfio: failed to set iommu for container: %m");
1687 g_free(container);
1688 close(fd);
1689 return -errno;
1690 }
1691
1692 container->iommu_data.listener = vfio_memory_listener;
1693 container->iommu_data.release = vfio_listener_release;
1694
1695 memory_listener_register(&container->iommu_data.listener, &address_space_memory);
1696 } else {
1697 error_report("vfio: No available IOMMU models");
1698 g_free(container);
1699 close(fd);
1700 return -EINVAL;
1701 }
1702
1703 QLIST_INIT(&container->group_list);
1704 QLIST_INSERT_HEAD(&container_list, container, next);
1705
1706 group->container = container;
1707 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1708
1709 return 0;
1710 }
1711
1712 static void vfio_disconnect_container(VFIOGroup *group)
1713 {
1714 VFIOContainer *container = group->container;
1715
1716 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1717 error_report("vfio: error disconnecting group %d from container",
1718 group->groupid);
1719 }
1720
1721 QLIST_REMOVE(group, container_next);
1722 group->container = NULL;
1723
1724 if (QLIST_EMPTY(&container->group_list)) {
1725 if (container->iommu_data.release) {
1726 container->iommu_data.release(container);
1727 }
1728 QLIST_REMOVE(container, next);
1729 DPRINTF("vfio_disconnect_container: close container->fd\n");
1730 close(container->fd);
1731 g_free(container);
1732 }
1733 }
1734
1735 static VFIOGroup *vfio_get_group(int groupid)
1736 {
1737 VFIOGroup *group;
1738 char path[32];
1739 struct vfio_group_status status = { .argsz = sizeof(status) };
1740
1741 QLIST_FOREACH(group, &group_list, next) {
1742 if (group->groupid == groupid) {
1743 return group;
1744 }
1745 }
1746
1747 group = g_malloc0(sizeof(*group));
1748
1749 snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1750 group->fd = qemu_open(path, O_RDWR);
1751 if (group->fd < 0) {
1752 error_report("vfio: error opening %s: %m", path);
1753 g_free(group);
1754 return NULL;
1755 }
1756
1757 if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1758 error_report("vfio: error getting group status: %m");
1759 close(group->fd);
1760 g_free(group);
1761 return NULL;
1762 }
1763
1764 if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1765 error_report("vfio: error, group %d is not viable, please ensure "
1766 "all devices within the iommu_group are bound to their "
1767 "vfio bus driver.", groupid);
1768 close(group->fd);
1769 g_free(group);
1770 return NULL;
1771 }
1772
1773 group->groupid = groupid;
1774 QLIST_INIT(&group->device_list);
1775
1776 if (vfio_connect_container(group)) {
1777 error_report("vfio: failed to setup container for group %d", groupid);
1778 close(group->fd);
1779 g_free(group);
1780 return NULL;
1781 }
1782
1783 QLIST_INSERT_HEAD(&group_list, group, next);
1784
1785 return group;
1786 }
1787
1788 static void vfio_put_group(VFIOGroup *group)
1789 {
1790 if (!QLIST_EMPTY(&group->device_list)) {
1791 return;
1792 }
1793
1794 vfio_disconnect_container(group);
1795 QLIST_REMOVE(group, next);
1796 DPRINTF("vfio_put_group: close group->fd\n");
1797 close(group->fd);
1798 g_free(group);
1799 }
1800
1801 static int vfio_get_device(VFIOGroup *group, const char *name, VFIODevice *vdev)
1802 {
1803 struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1804 struct vfio_region_info reg_info = { .argsz = sizeof(reg_info) };
1805 int ret, i;
1806
1807 ret = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1808 if (ret < 0) {
1809 error_report("vfio: error getting device %s from group %d: %m",
1810 name, group->groupid);
1811 error_printf("Verify all devices in group %d are bound to vfio-pci "
1812 "or pci-stub and not already in use\n", group->groupid);
1813 return ret;
1814 }
1815
1816 vdev->fd = ret;
1817 vdev->group = group;
1818 QLIST_INSERT_HEAD(&group->device_list, vdev, next);
1819
1820 /* Sanity check device */
1821 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_INFO, &dev_info);
1822 if (ret) {
1823 error_report("vfio: error getting device info: %m");
1824 goto error;
1825 }
1826
1827 DPRINTF("Device %s flags: %u, regions: %u, irgs: %u\n", name,
1828 dev_info.flags, dev_info.num_regions, dev_info.num_irqs);
1829
1830 if (!(dev_info.flags & VFIO_DEVICE_FLAGS_PCI)) {
1831 error_report("vfio: Um, this isn't a PCI device");
1832 goto error;
1833 }
1834
1835 vdev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1836 if (!vdev->reset_works) {
1837 error_report("Warning, device %s does not support reset", name);
1838 }
1839
1840 if (dev_info.num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
1841 error_report("vfio: unexpected number of io regions %u",
1842 dev_info.num_regions);
1843 goto error;
1844 }
1845
1846 if (dev_info.num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
1847 error_report("vfio: unexpected number of irqs %u", dev_info.num_irqs);
1848 goto error;
1849 }
1850
1851 for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
1852 reg_info.index = i;
1853
1854 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
1855 if (ret) {
1856 error_report("vfio: Error getting region %d info: %m", i);
1857 goto error;
1858 }
1859
1860 DPRINTF("Device %s region %d:\n", name, i);
1861 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1862 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
1863 (unsigned long)reg_info.flags);
1864
1865 vdev->bars[i].flags = reg_info.flags;
1866 vdev->bars[i].size = reg_info.size;
1867 vdev->bars[i].fd_offset = reg_info.offset;
1868 vdev->bars[i].fd = vdev->fd;
1869 vdev->bars[i].nr = i;
1870 }
1871
1872 reg_info.index = VFIO_PCI_ROM_REGION_INDEX;
1873
1874 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
1875 if (ret) {
1876 error_report("vfio: Error getting ROM info: %m");
1877 goto error;
1878 }
1879
1880 DPRINTF("Device %s ROM:\n", name);
1881 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1882 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
1883 (unsigned long)reg_info.flags);
1884
1885 vdev->rom_size = reg_info.size;
1886 vdev->rom_offset = reg_info.offset;
1887
1888 reg_info.index = VFIO_PCI_CONFIG_REGION_INDEX;
1889
1890 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
1891 if (ret) {
1892 error_report("vfio: Error getting config info: %m");
1893 goto error;
1894 }
1895
1896 DPRINTF("Device %s config:\n", name);
1897 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1898 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
1899 (unsigned long)reg_info.flags);
1900
1901 vdev->config_size = reg_info.size;
1902 if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
1903 vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
1904 }
1905 vdev->config_offset = reg_info.offset;
1906
1907 error:
1908 if (ret) {
1909 QLIST_REMOVE(vdev, next);
1910 vdev->group = NULL;
1911 close(vdev->fd);
1912 }
1913 return ret;
1914 }
1915
1916 static void vfio_put_device(VFIODevice *vdev)
1917 {
1918 QLIST_REMOVE(vdev, next);
1919 vdev->group = NULL;
1920 DPRINTF("vfio_put_device: close vdev->fd\n");
1921 close(vdev->fd);
1922 if (vdev->msix) {
1923 g_free(vdev->msix);
1924 vdev->msix = NULL;
1925 }
1926 }
1927
1928 static int vfio_initfn(PCIDevice *pdev)
1929 {
1930 VFIODevice *pvdev, *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1931 VFIOGroup *group;
1932 char path[PATH_MAX], iommu_group_path[PATH_MAX], *group_name;
1933 ssize_t len;
1934 struct stat st;
1935 int groupid;
1936 int ret;
1937
1938 /* Check that the host device exists */
1939 snprintf(path, sizeof(path),
1940 "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/",
1941 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1942 vdev->host.function);
1943 if (stat(path, &st) < 0) {
1944 error_report("vfio: error: no such host device: %s", path);
1945 return -errno;
1946 }
1947
1948 strncat(path, "iommu_group", sizeof(path) - strlen(path) - 1);
1949
1950 len = readlink(path, iommu_group_path, PATH_MAX);
1951 if (len <= 0) {
1952 error_report("vfio: error no iommu_group for device");
1953 return -errno;
1954 }
1955
1956 iommu_group_path[len] = 0;
1957 group_name = basename(iommu_group_path);
1958
1959 if (sscanf(group_name, "%d", &groupid) != 1) {
1960 error_report("vfio: error reading %s: %m", path);
1961 return -errno;
1962 }
1963
1964 DPRINTF("%s(%04x:%02x:%02x.%x) group %d\n", __func__, vdev->host.domain,
1965 vdev->host.bus, vdev->host.slot, vdev->host.function, groupid);
1966
1967 group = vfio_get_group(groupid);
1968 if (!group) {
1969 error_report("vfio: failed to get group %d", groupid);
1970 return -ENOENT;
1971 }
1972
1973 snprintf(path, sizeof(path), "%04x:%02x:%02x.%01x",
1974 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1975 vdev->host.function);
1976
1977 QLIST_FOREACH(pvdev, &group->device_list, next) {
1978 if (pvdev->host.domain == vdev->host.domain &&
1979 pvdev->host.bus == vdev->host.bus &&
1980 pvdev->host.slot == vdev->host.slot &&
1981 pvdev->host.function == vdev->host.function) {
1982
1983 error_report("vfio: error: device %s is already attached", path);
1984 vfio_put_group(group);
1985 return -EBUSY;
1986 }
1987 }
1988
1989 ret = vfio_get_device(group, path, vdev);
1990 if (ret) {
1991 error_report("vfio: failed to get device %s", path);
1992 vfio_put_group(group);
1993 return ret;
1994 }
1995
1996 /* Get a copy of config space */
1997 ret = pread(vdev->fd, vdev->pdev.config,
1998 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
1999 vdev->config_offset);
2000 if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
2001 ret = ret < 0 ? -errno : -EFAULT;
2002 error_report("vfio: Failed to read device config space");
2003 goto out_put;
2004 }
2005
2006 /*
2007 * Clear host resource mapping info. If we choose not to register a
2008 * BAR, such as might be the case with the option ROM, we can get
2009 * confusing, unwritable, residual addresses from the host here.
2010 */
2011 memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
2012 memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
2013
2014 vfio_load_rom(vdev);
2015
2016 ret = vfio_early_setup_msix(vdev);
2017 if (ret) {
2018 goto out_put;
2019 }
2020
2021 vfio_map_bars(vdev);
2022
2023 ret = vfio_add_capabilities(vdev);
2024 if (ret) {
2025 goto out_teardown;
2026 }
2027
2028 if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
2029 vdev->intx.mmap_timer = qemu_new_timer_ms(vm_clock,
2030 vfio_intx_mmap_enable, vdev);
2031 pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_update_irq);
2032 ret = vfio_enable_intx(vdev);
2033 if (ret) {
2034 goto out_teardown;
2035 }
2036 }
2037
2038 return 0;
2039
2040 out_teardown:
2041 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
2042 vfio_teardown_msi(vdev);
2043 vfio_unmap_bars(vdev);
2044 out_put:
2045 vfio_put_device(vdev);
2046 vfio_put_group(group);
2047 return ret;
2048 }
2049
2050 static void vfio_exitfn(PCIDevice *pdev)
2051 {
2052 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
2053 VFIOGroup *group = vdev->group;
2054
2055 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
2056 vfio_disable_interrupts(vdev);
2057 if (vdev->intx.mmap_timer) {
2058 qemu_free_timer(vdev->intx.mmap_timer);
2059 }
2060 vfio_teardown_msi(vdev);
2061 vfio_unmap_bars(vdev);
2062 vfio_put_device(vdev);
2063 vfio_put_group(group);
2064 }
2065
2066 static void vfio_pci_reset(DeviceState *dev)
2067 {
2068 PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
2069 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
2070 uint16_t cmd;
2071
2072 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
2073 vdev->host.bus, vdev->host.slot, vdev->host.function);
2074
2075 vfio_disable_interrupts(vdev);
2076
2077 /*
2078 * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
2079 * Also put INTx Disable in known state.
2080 */
2081 cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
2082 cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
2083 PCI_COMMAND_INTX_DISABLE);
2084 vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
2085
2086 if (vdev->reset_works) {
2087 if (ioctl(vdev->fd, VFIO_DEVICE_RESET)) {
2088 error_report("vfio: Error unable to reset physical device "
2089 "(%04x:%02x:%02x.%x): %m", vdev->host.domain,
2090 vdev->host.bus, vdev->host.slot, vdev->host.function);
2091 }
2092 }
2093
2094 vfio_enable_intx(vdev);
2095 }
2096
2097 static Property vfio_pci_dev_properties[] = {
2098 DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIODevice, host),
2099 DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIODevice,
2100 intx.mmap_timeout, 1100),
2101 /*
2102 * TODO - support passed fds... is this necessary?
2103 * DEFINE_PROP_STRING("vfiofd", VFIODevice, vfiofd_name),
2104 * DEFINE_PROP_STRING("vfiogroupfd, VFIODevice, vfiogroupfd_name),
2105 */
2106 DEFINE_PROP_END_OF_LIST(),
2107 };
2108
2109 static const VMStateDescription vfio_pci_vmstate = {
2110 .name = "vfio-pci",
2111 .unmigratable = 1,
2112 };
2113
2114 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
2115 {
2116 DeviceClass *dc = DEVICE_CLASS(klass);
2117 PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
2118
2119 dc->reset = vfio_pci_reset;
2120 dc->props = vfio_pci_dev_properties;
2121 dc->vmsd = &vfio_pci_vmstate;
2122 dc->desc = "VFIO-based PCI device assignment";
2123 pdc->init = vfio_initfn;
2124 pdc->exit = vfio_exitfn;
2125 pdc->config_read = vfio_pci_read_config;
2126 pdc->config_write = vfio_pci_write_config;
2127 pdc->is_express = 1; /* We might be */
2128 }
2129
2130 static const TypeInfo vfio_pci_dev_info = {
2131 .name = "vfio-pci",
2132 .parent = TYPE_PCI_DEVICE,
2133 .instance_size = sizeof(VFIODevice),
2134 .class_init = vfio_pci_dev_class_init,
2135 };
2136
2137 static void register_vfio_pci_dev_type(void)
2138 {
2139 type_register_static(&vfio_pci_dev_info);
2140 }
2141
2142 type_init(register_vfio_pci_dev_type)
Various patches