]> git.proxmox.com Git - qemu.git/blob - hw/misc/vfio.c
projects / qemu.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
aio / timers: Switch entire codebase to the new timer API
[qemu.git] / hw / misc / vfio.c
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 <linux/vfio.h>
23 #include <sys/ioctl.h>
24 #include <sys/mman.h>
25 #include <sys/stat.h>
26 #include <sys/types.h>
27 #include <unistd.h>
28
29 #include "config.h"
30 #include "exec/address-spaces.h"
31 #include "exec/memory.h"
32 #include "hw/pci/msi.h"
33 #include "hw/pci/msix.h"
34 #include "hw/pci/pci.h"
35 #include "qemu-common.h"
36 #include "qemu/error-report.h"
37 #include "qemu/event_notifier.h"
38 #include "qemu/queue.h"
39 #include "qemu/range.h"
40 #include "sysemu/kvm.h"
41 #include "sysemu/sysemu.h"
42
43 /* #define DEBUG_VFIO */
44 #ifdef DEBUG_VFIO
45 #define DPRINTF(fmt, ...) \
46 do { fprintf(stderr, "vfio: " fmt, ## __VA_ARGS__); } while (0)
47 #else
48 #define DPRINTF(fmt, ...) \
49 do { } while (0)
50 #endif
51
52 /* Extra debugging, trap acceleration paths for more logging */
53 #define VFIO_ALLOW_MMAP 1
54 #define VFIO_ALLOW_KVM_INTX 1
55
56 struct VFIODevice;
57
58 typedef struct VFIOQuirk {
59 MemoryRegion mem;
60 struct VFIODevice *vdev;
61 QLIST_ENTRY(VFIOQuirk) next;
62 struct {
63 uint32_t base_offset:TARGET_PAGE_BITS;
64 uint32_t address_offset:TARGET_PAGE_BITS;
65 uint32_t address_size:3;
66 uint32_t bar:3;
67
68 uint32_t address_match;
69 uint32_t address_mask;
70
71 uint32_t address_val:TARGET_PAGE_BITS;
72 uint32_t data_offset:TARGET_PAGE_BITS;
73 uint32_t data_size:3;
74
75 uint8_t flags;
76 uint8_t read_flags;
77 uint8_t write_flags;
78 } data;
79 } VFIOQuirk;
80
81 typedef struct VFIOBAR {
82 off_t fd_offset; /* offset of BAR within device fd */
83 int fd; /* device fd, allows us to pass VFIOBAR as opaque data */
84 MemoryRegion mem; /* slow, read/write access */
85 MemoryRegion mmap_mem; /* direct mapped access */
86 void *mmap;
87 size_t size;
88 uint32_t flags; /* VFIO region flags (rd/wr/mmap) */
89 uint8_t nr; /* cache the BAR number for debug */
90 bool ioport;
91 bool mem64;
92 QLIST_HEAD(, VFIOQuirk) quirks;
93 } VFIOBAR;
94
95 typedef struct VFIOVGARegion {
96 MemoryRegion mem;
97 off_t offset;
98 int nr;
99 QLIST_HEAD(, VFIOQuirk) quirks;
100 } VFIOVGARegion;
101
102 typedef struct VFIOVGA {
103 off_t fd_offset;
104 int fd;
105 VFIOVGARegion region[QEMU_PCI_VGA_NUM_REGIONS];
106 } VFIOVGA;
107
108 typedef struct VFIOINTx {
109 bool pending; /* interrupt pending */
110 bool kvm_accel; /* set when QEMU bypass through KVM enabled */
111 uint8_t pin; /* which pin to pull for qemu_set_irq */
112 EventNotifier interrupt; /* eventfd triggered on interrupt */
113 EventNotifier unmask; /* eventfd for unmask on QEMU bypass */
114 PCIINTxRoute route; /* routing info for QEMU bypass */
115 uint32_t mmap_timeout; /* delay to re-enable mmaps after interrupt */
116 QEMUTimer *mmap_timer; /* enable mmaps after periods w/o interrupts */
117 } VFIOINTx;
118
119 typedef struct VFIOMSIVector {
120 EventNotifier interrupt; /* eventfd triggered on interrupt */
121 struct VFIODevice *vdev; /* back pointer to device */
122 int virq; /* KVM irqchip route for QEMU bypass */
123 bool use;
124 } VFIOMSIVector;
125
126 enum {
127 VFIO_INT_NONE = 0,
128 VFIO_INT_INTx = 1,
129 VFIO_INT_MSI = 2,
130 VFIO_INT_MSIX = 3,
131 };
132
133 struct VFIOGroup;
134
135 typedef struct VFIOContainer {
136 int fd; /* /dev/vfio/vfio, empowered by the attached groups */
137 struct {
138 /* enable abstraction to support various iommu backends */
139 union {
140 MemoryListener listener; /* Used by type1 iommu */
141 };
142 void (*release)(struct VFIOContainer *);
143 } iommu_data;
144 QLIST_HEAD(, VFIOGroup) group_list;
145 QLIST_ENTRY(VFIOContainer) next;
146 } VFIOContainer;
147
148 /* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
149 typedef struct VFIOMSIXInfo {
150 uint8_t table_bar;
151 uint8_t pba_bar;
152 uint16_t entries;
153 uint32_t table_offset;
154 uint32_t pba_offset;
155 MemoryRegion mmap_mem;
156 void *mmap;
157 } VFIOMSIXInfo;
158
159 typedef struct VFIODevice {
160 PCIDevice pdev;
161 int fd;
162 VFIOINTx intx;
163 unsigned int config_size;
164 uint8_t *emulated_config_bits; /* QEMU emulated bits, little-endian */
165 off_t config_offset; /* Offset of config space region within device fd */
166 unsigned int rom_size;
167 off_t rom_offset; /* Offset of ROM region within device fd */
168 int msi_cap_size;
169 VFIOMSIVector *msi_vectors;
170 VFIOMSIXInfo *msix;
171 int nr_vectors; /* Number of MSI/MSIX vectors currently in use */
172 int interrupt; /* Current interrupt type */
173 VFIOBAR bars[PCI_NUM_REGIONS - 1]; /* No ROM */
174 VFIOVGA vga; /* 0xa0000, 0x3b0, 0x3c0 */
175 PCIHostDeviceAddress host;
176 QLIST_ENTRY(VFIODevice) next;
177 struct VFIOGroup *group;
178 EventNotifier err_notifier;
179 uint32_t features;
180 #define VFIO_FEATURE_ENABLE_VGA_BIT 0
181 #define VFIO_FEATURE_ENABLE_VGA (1 << VFIO_FEATURE_ENABLE_VGA_BIT)
182 int32_t bootindex;
183 uint8_t pm_cap;
184 bool reset_works;
185 bool has_vga;
186 bool pci_aer;
187 } VFIODevice;
188
189 typedef struct VFIOGroup {
190 int fd;
191 int groupid;
192 VFIOContainer *container;
193 QLIST_HEAD(, VFIODevice) device_list;
194 QLIST_ENTRY(VFIOGroup) next;
195 QLIST_ENTRY(VFIOGroup) container_next;
196 } VFIOGroup;
197
198 #define MSIX_CAP_LENGTH 12
199
200 static QLIST_HEAD(, VFIOContainer)
201 container_list = QLIST_HEAD_INITIALIZER(container_list);
202
203 static QLIST_HEAD(, VFIOGroup)
204 group_list = QLIST_HEAD_INITIALIZER(group_list);
205
206 static void vfio_disable_interrupts(VFIODevice *vdev);
207 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len);
208 static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
209 uint32_t val, int len);
210 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled);
211
212 /*
213 * Common VFIO interrupt disable
214 */
215 static void vfio_disable_irqindex(VFIODevice *vdev, int index)
216 {
217 struct vfio_irq_set irq_set = {
218 .argsz = sizeof(irq_set),
219 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
220 .index = index,
221 .start = 0,
222 .count = 0,
223 };
224
225 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
226 }
227
228 /*
229 * INTx
230 */
231 static void vfio_unmask_intx(VFIODevice *vdev)
232 {
233 struct vfio_irq_set irq_set = {
234 .argsz = sizeof(irq_set),
235 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
236 .index = VFIO_PCI_INTX_IRQ_INDEX,
237 .start = 0,
238 .count = 1,
239 };
240
241 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
242 }
243
244 #ifdef CONFIG_KVM /* Unused outside of CONFIG_KVM code */
245 static void vfio_mask_intx(VFIODevice *vdev)
246 {
247 struct vfio_irq_set irq_set = {
248 .argsz = sizeof(irq_set),
249 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
250 .index = VFIO_PCI_INTX_IRQ_INDEX,
251 .start = 0,
252 .count = 1,
253 };
254
255 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
256 }
257 #endif
258
259 /*
260 * Disabling BAR mmaping can be slow, but toggling it around INTx can
261 * also be a huge overhead. We try to get the best of both worlds by
262 * waiting until an interrupt to disable mmaps (subsequent transitions
263 * to the same state are effectively no overhead). If the interrupt has
264 * been serviced and the time gap is long enough, we re-enable mmaps for
265 * performance. This works well for things like graphics cards, which
266 * may not use their interrupt at all and are penalized to an unusable
267 * level by read/write BAR traps. Other devices, like NICs, have more
268 * regular interrupts and see much better latency by staying in non-mmap
269 * mode. We therefore set the default mmap_timeout such that a ping
270 * is just enough to keep the mmap disabled. Users can experiment with
271 * other options with the x-intx-mmap-timeout-ms parameter (a value of
272 * zero disables the timer).
273 */
274 static void vfio_intx_mmap_enable(void *opaque)
275 {
276 VFIODevice *vdev = opaque;
277
278 if (vdev->intx.pending) {
279 timer_mod(vdev->intx.mmap_timer,
280 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
281 return;
282 }
283
284 vfio_mmap_set_enabled(vdev, true);
285 }
286
287 static void vfio_intx_interrupt(void *opaque)
288 {
289 VFIODevice *vdev = opaque;
290
291 if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
292 return;
293 }
294
295 DPRINTF("%s(%04x:%02x:%02x.%x) Pin %c\n", __func__, vdev->host.domain,
296 vdev->host.bus, vdev->host.slot, vdev->host.function,
297 'A' + vdev->intx.pin);
298
299 vdev->intx.pending = true;
300 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 1);
301 vfio_mmap_set_enabled(vdev, false);
302 if (vdev->intx.mmap_timeout) {
303 timer_mod(vdev->intx.mmap_timer,
304 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout);
305 }
306 }
307
308 static void vfio_eoi(VFIODevice *vdev)
309 {
310 if (!vdev->intx.pending) {
311 return;
312 }
313
314 DPRINTF("%s(%04x:%02x:%02x.%x) EOI\n", __func__, vdev->host.domain,
315 vdev->host.bus, vdev->host.slot, vdev->host.function);
316
317 vdev->intx.pending = false;
318 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
319 vfio_unmask_intx(vdev);
320 }
321
322 static void vfio_enable_intx_kvm(VFIODevice *vdev)
323 {
324 #ifdef CONFIG_KVM
325 struct kvm_irqfd irqfd = {
326 .fd = event_notifier_get_fd(&vdev->intx.interrupt),
327 .gsi = vdev->intx.route.irq,
328 .flags = KVM_IRQFD_FLAG_RESAMPLE,
329 };
330 struct vfio_irq_set *irq_set;
331 int ret, argsz;
332 int32_t *pfd;
333
334 if (!VFIO_ALLOW_KVM_INTX || !kvm_irqfds_enabled() ||
335 vdev->intx.route.mode != PCI_INTX_ENABLED ||
336 !kvm_check_extension(kvm_state, KVM_CAP_IRQFD_RESAMPLE)) {
337 return;
338 }
339
340 /* Get to a known interrupt state */
341 qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
342 vfio_mask_intx(vdev);
343 vdev->intx.pending = false;
344 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
345
346 /* Get an eventfd for resample/unmask */
347 if (event_notifier_init(&vdev->intx.unmask, 0)) {
348 error_report("vfio: Error: event_notifier_init failed eoi");
349 goto fail;
350 }
351
352 /* KVM triggers it, VFIO listens for it */
353 irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
354
355 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
356 error_report("vfio: Error: Failed to setup resample irqfd: %m");
357 goto fail_irqfd;
358 }
359
360 argsz = sizeof(*irq_set) + sizeof(*pfd);
361
362 irq_set = g_malloc0(argsz);
363 irq_set->argsz = argsz;
364 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
365 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
366 irq_set->start = 0;
367 irq_set->count = 1;
368 pfd = (int32_t *)&irq_set->data;
369
370 *pfd = irqfd.resamplefd;
371
372 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
373 g_free(irq_set);
374 if (ret) {
375 error_report("vfio: Error: Failed to setup INTx unmask fd: %m");
376 goto fail_vfio;
377 }
378
379 /* Let'em rip */
380 vfio_unmask_intx(vdev);
381
382 vdev->intx.kvm_accel = true;
383
384 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel enabled\n",
385 __func__, vdev->host.domain, vdev->host.bus,
386 vdev->host.slot, vdev->host.function);
387
388 return;
389
390 fail_vfio:
391 irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
392 kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
393 fail_irqfd:
394 event_notifier_cleanup(&vdev->intx.unmask);
395 fail:
396 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
397 vfio_unmask_intx(vdev);
398 #endif
399 }
400
401 static void vfio_disable_intx_kvm(VFIODevice *vdev)
402 {
403 #ifdef CONFIG_KVM
404 struct kvm_irqfd irqfd = {
405 .fd = event_notifier_get_fd(&vdev->intx.interrupt),
406 .gsi = vdev->intx.route.irq,
407 .flags = KVM_IRQFD_FLAG_DEASSIGN,
408 };
409
410 if (!vdev->intx.kvm_accel) {
411 return;
412 }
413
414 /*
415 * Get to a known state, hardware masked, QEMU ready to accept new
416 * interrupts, QEMU IRQ de-asserted.
417 */
418 vfio_mask_intx(vdev);
419 vdev->intx.pending = false;
420 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
421
422 /* Tell KVM to stop listening for an INTx irqfd */
423 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
424 error_report("vfio: Error: Failed to disable INTx irqfd: %m");
425 }
426
427 /* We only need to close the eventfd for VFIO to cleanup the kernel side */
428 event_notifier_cleanup(&vdev->intx.unmask);
429
430 /* QEMU starts listening for interrupt events. */
431 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
432
433 vdev->intx.kvm_accel = false;
434
435 /* If we've missed an event, let it re-fire through QEMU */
436 vfio_unmask_intx(vdev);
437
438 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel disabled\n",
439 __func__, vdev->host.domain, vdev->host.bus,
440 vdev->host.slot, vdev->host.function);
441 #endif
442 }
443
444 static void vfio_update_irq(PCIDevice *pdev)
445 {
446 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
447 PCIINTxRoute route;
448
449 if (vdev->interrupt != VFIO_INT_INTx) {
450 return;
451 }
452
453 route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
454
455 if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
456 return; /* Nothing changed */
457 }
458
459 DPRINTF("%s(%04x:%02x:%02x.%x) IRQ moved %d -> %d\n", __func__,
460 vdev->host.domain, vdev->host.bus, vdev->host.slot,
461 vdev->host.function, vdev->intx.route.irq, route.irq);
462
463 vfio_disable_intx_kvm(vdev);
464
465 vdev->intx.route = route;
466
467 if (route.mode != PCI_INTX_ENABLED) {
468 return;
469 }
470
471 vfio_enable_intx_kvm(vdev);
472
473 /* Re-enable the interrupt in cased we missed an EOI */
474 vfio_eoi(vdev);
475 }
476
477 static int vfio_enable_intx(VFIODevice *vdev)
478 {
479 uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
480 int ret, argsz;
481 struct vfio_irq_set *irq_set;
482 int32_t *pfd;
483
484 if (!pin) {
485 return 0;
486 }
487
488 vfio_disable_interrupts(vdev);
489
490 vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
491
492 #ifdef CONFIG_KVM
493 /*
494 * Only conditional to avoid generating error messages on platforms
495 * where we won't actually use the result anyway.
496 */
497 if (kvm_irqfds_enabled() &&
498 kvm_check_extension(kvm_state, KVM_CAP_IRQFD_RESAMPLE)) {
499 vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
500 vdev->intx.pin);
501 }
502 #endif
503
504 ret = event_notifier_init(&vdev->intx.interrupt, 0);
505 if (ret) {
506 error_report("vfio: Error: event_notifier_init failed");
507 return ret;
508 }
509
510 argsz = sizeof(*irq_set) + sizeof(*pfd);
511
512 irq_set = g_malloc0(argsz);
513 irq_set->argsz = argsz;
514 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
515 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
516 irq_set->start = 0;
517 irq_set->count = 1;
518 pfd = (int32_t *)&irq_set->data;
519
520 *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
521 qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
522
523 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
524 g_free(irq_set);
525 if (ret) {
526 error_report("vfio: Error: Failed to setup INTx fd: %m");
527 qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
528 event_notifier_cleanup(&vdev->intx.interrupt);
529 return -errno;
530 }
531
532 vfio_enable_intx_kvm(vdev);
533
534 vdev->interrupt = VFIO_INT_INTx;
535
536 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
537 vdev->host.bus, vdev->host.slot, vdev->host.function);
538
539 return 0;
540 }
541
542 static void vfio_disable_intx(VFIODevice *vdev)
543 {
544 int fd;
545
546 timer_del(vdev->intx.mmap_timer);
547 vfio_disable_intx_kvm(vdev);
548 vfio_disable_irqindex(vdev, VFIO_PCI_INTX_IRQ_INDEX);
549 vdev->intx.pending = false;
550 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
551 vfio_mmap_set_enabled(vdev, true);
552
553 fd = event_notifier_get_fd(&vdev->intx.interrupt);
554 qemu_set_fd_handler(fd, NULL, NULL, vdev);
555 event_notifier_cleanup(&vdev->intx.interrupt);
556
557 vdev->interrupt = VFIO_INT_NONE;
558
559 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
560 vdev->host.bus, vdev->host.slot, vdev->host.function);
561 }
562
563 /*
564 * MSI/X
565 */
566 static void vfio_msi_interrupt(void *opaque)
567 {
568 VFIOMSIVector *vector = opaque;
569 VFIODevice *vdev = vector->vdev;
570 int nr = vector - vdev->msi_vectors;
571
572 if (!event_notifier_test_and_clear(&vector->interrupt)) {
573 return;
574 }
575
576 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d\n", __func__,
577 vdev->host.domain, vdev->host.bus, vdev->host.slot,
578 vdev->host.function, nr);
579
580 if (vdev->interrupt == VFIO_INT_MSIX) {
581 msix_notify(&vdev->pdev, nr);
582 } else if (vdev->interrupt == VFIO_INT_MSI) {
583 msi_notify(&vdev->pdev, nr);
584 } else {
585 error_report("vfio: MSI interrupt receieved, but not enabled?");
586 }
587 }
588
589 static int vfio_enable_vectors(VFIODevice *vdev, bool msix)
590 {
591 struct vfio_irq_set *irq_set;
592 int ret = 0, i, argsz;
593 int32_t *fds;
594
595 argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
596
597 irq_set = g_malloc0(argsz);
598 irq_set->argsz = argsz;
599 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
600 irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
601 irq_set->start = 0;
602 irq_set->count = vdev->nr_vectors;
603 fds = (int32_t *)&irq_set->data;
604
605 for (i = 0; i < vdev->nr_vectors; i++) {
606 if (!vdev->msi_vectors[i].use) {
607 fds[i] = -1;
608 continue;
609 }
610
611 fds[i] = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
612 }
613
614 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
615
616 g_free(irq_set);
617
618 return ret;
619 }
620
621 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
622 MSIMessage *msg, IOHandler *handler)
623 {
624 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
625 VFIOMSIVector *vector;
626 int ret;
627
628 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d used\n", __func__,
629 vdev->host.domain, vdev->host.bus, vdev->host.slot,
630 vdev->host.function, nr);
631
632 vector = &vdev->msi_vectors[nr];
633 vector->vdev = vdev;
634 vector->use = true;
635
636 msix_vector_use(pdev, nr);
637
638 if (event_notifier_init(&vector->interrupt, 0)) {
639 error_report("vfio: Error: event_notifier_init failed");
640 }
641
642 /*
643 * Attempt to enable route through KVM irqchip,
644 * default to userspace handling if unavailable.
645 */
646 vector->virq = msg ? kvm_irqchip_add_msi_route(kvm_state, *msg) : -1;
647 if (vector->virq < 0 ||
648 kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
649 vector->virq) < 0) {
650 if (vector->virq >= 0) {
651 kvm_irqchip_release_virq(kvm_state, vector->virq);
652 vector->virq = -1;
653 }
654 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
655 handler, NULL, vector);
656 }
657
658 /*
659 * We don't want to have the host allocate all possible MSI vectors
660 * for a device if they're not in use, so we shutdown and incrementally
661 * increase them as needed.
662 */
663 if (vdev->nr_vectors < nr + 1) {
664 vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
665 vdev->nr_vectors = nr + 1;
666 ret = vfio_enable_vectors(vdev, true);
667 if (ret) {
668 error_report("vfio: failed to enable vectors, %d", ret);
669 }
670 } else {
671 int argsz;
672 struct vfio_irq_set *irq_set;
673 int32_t *pfd;
674
675 argsz = sizeof(*irq_set) + sizeof(*pfd);
676
677 irq_set = g_malloc0(argsz);
678 irq_set->argsz = argsz;
679 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
680 VFIO_IRQ_SET_ACTION_TRIGGER;
681 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
682 irq_set->start = nr;
683 irq_set->count = 1;
684 pfd = (int32_t *)&irq_set->data;
685
686 *pfd = event_notifier_get_fd(&vector->interrupt);
687
688 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
689 g_free(irq_set);
690 if (ret) {
691 error_report("vfio: failed to modify vector, %d", ret);
692 }
693 }
694
695 return 0;
696 }
697
698 static int vfio_msix_vector_use(PCIDevice *pdev,
699 unsigned int nr, MSIMessage msg)
700 {
701 return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
702 }
703
704 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
705 {
706 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
707 VFIOMSIVector *vector = &vdev->msi_vectors[nr];
708 int argsz;
709 struct vfio_irq_set *irq_set;
710 int32_t *pfd;
711
712 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d released\n", __func__,
713 vdev->host.domain, vdev->host.bus, vdev->host.slot,
714 vdev->host.function, nr);
715
716 /*
717 * XXX What's the right thing to do here? This turns off the interrupt
718 * completely, but do we really just want to switch the interrupt to
719 * bouncing through userspace and let msix.c drop it? Not sure.
720 */
721 msix_vector_unuse(pdev, nr);
722
723 argsz = sizeof(*irq_set) + sizeof(*pfd);
724
725 irq_set = g_malloc0(argsz);
726 irq_set->argsz = argsz;
727 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
728 VFIO_IRQ_SET_ACTION_TRIGGER;
729 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
730 irq_set->start = nr;
731 irq_set->count = 1;
732 pfd = (int32_t *)&irq_set->data;
733
734 *pfd = -1;
735
736 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
737
738 g_free(irq_set);
739
740 if (vector->virq < 0) {
741 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
742 NULL, NULL, NULL);
743 } else {
744 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
745 vector->virq);
746 kvm_irqchip_release_virq(kvm_state, vector->virq);
747 vector->virq = -1;
748 }
749
750 event_notifier_cleanup(&vector->interrupt);
751 vector->use = false;
752 }
753
754 static void vfio_enable_msix(VFIODevice *vdev)
755 {
756 vfio_disable_interrupts(vdev);
757
758 vdev->msi_vectors = g_malloc0(vdev->msix->entries * sizeof(VFIOMSIVector));
759
760 vdev->interrupt = VFIO_INT_MSIX;
761
762 /*
763 * Some communication channels between VF & PF or PF & fw rely on the
764 * physical state of the device and expect that enabling MSI-X from the
765 * guest enables the same on the host. When our guest is Linux, the
766 * guest driver call to pci_enable_msix() sets the enabling bit in the
767 * MSI-X capability, but leaves the vector table masked. We therefore
768 * can't rely on a vector_use callback (from request_irq() in the guest)
769 * to switch the physical device into MSI-X mode because that may come a
770 * long time after pci_enable_msix(). This code enables vector 0 with
771 * triggering to userspace, then immediately release the vector, leaving
772 * the physical device with no vectors enabled, but MSI-X enabled, just
773 * like the guest view.
774 */
775 vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
776 vfio_msix_vector_release(&vdev->pdev, 0);
777
778 if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
779 vfio_msix_vector_release, NULL)) {
780 error_report("vfio: msix_set_vector_notifiers failed");
781 }
782
783 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
784 vdev->host.bus, vdev->host.slot, vdev->host.function);
785 }
786
787 static void vfio_enable_msi(VFIODevice *vdev)
788 {
789 int ret, i;
790
791 vfio_disable_interrupts(vdev);
792
793 vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
794 retry:
795 vdev->msi_vectors = g_malloc0(vdev->nr_vectors * sizeof(VFIOMSIVector));
796
797 for (i = 0; i < vdev->nr_vectors; i++) {
798 MSIMessage msg;
799 VFIOMSIVector *vector = &vdev->msi_vectors[i];
800
801 vector->vdev = vdev;
802 vector->use = true;
803
804 if (event_notifier_init(&vector->interrupt, 0)) {
805 error_report("vfio: Error: event_notifier_init failed");
806 }
807
808 msg = msi_get_message(&vdev->pdev, i);
809
810 /*
811 * Attempt to enable route through KVM irqchip,
812 * default to userspace handling if unavailable.
813 */
814 vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg);
815 if (vector->virq < 0 ||
816 kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
817 vector->virq) < 0) {
818 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
819 vfio_msi_interrupt, NULL, vector);
820 }
821 }
822
823 ret = vfio_enable_vectors(vdev, false);
824 if (ret) {
825 if (ret < 0) {
826 error_report("vfio: Error: Failed to setup MSI fds: %m");
827 } else if (ret != vdev->nr_vectors) {
828 error_report("vfio: Error: Failed to enable %d "
829 "MSI vectors, retry with %d", vdev->nr_vectors, ret);
830 }
831
832 for (i = 0; i < vdev->nr_vectors; i++) {
833 VFIOMSIVector *vector = &vdev->msi_vectors[i];
834 if (vector->virq >= 0) {
835 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
836 vector->virq);
837 kvm_irqchip_release_virq(kvm_state, vector->virq);
838 vector->virq = -1;
839 } else {
840 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
841 NULL, NULL, NULL);
842 }
843 event_notifier_cleanup(&vector->interrupt);
844 }
845
846 g_free(vdev->msi_vectors);
847
848 if (ret > 0 && ret != vdev->nr_vectors) {
849 vdev->nr_vectors = ret;
850 goto retry;
851 }
852 vdev->nr_vectors = 0;
853
854 return;
855 }
856
857 vdev->interrupt = VFIO_INT_MSI;
858
859 DPRINTF("%s(%04x:%02x:%02x.%x) Enabled %d MSI vectors\n", __func__,
860 vdev->host.domain, vdev->host.bus, vdev->host.slot,
861 vdev->host.function, vdev->nr_vectors);
862 }
863
864 static void vfio_disable_msi_common(VFIODevice *vdev)
865 {
866 g_free(vdev->msi_vectors);
867 vdev->msi_vectors = NULL;
868 vdev->nr_vectors = 0;
869 vdev->interrupt = VFIO_INT_NONE;
870
871 vfio_enable_intx(vdev);
872 }
873
874 static void vfio_disable_msix(VFIODevice *vdev)
875 {
876 msix_unset_vector_notifiers(&vdev->pdev);
877
878 if (vdev->nr_vectors) {
879 vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
880 }
881
882 vfio_disable_msi_common(vdev);
883
884 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
885 vdev->host.bus, vdev->host.slot, vdev->host.function);
886 }
887
888 static void vfio_disable_msi(VFIODevice *vdev)
889 {
890 int i;
891
892 vfio_disable_irqindex(vdev, VFIO_PCI_MSI_IRQ_INDEX);
893
894 for (i = 0; i < vdev->nr_vectors; i++) {
895 VFIOMSIVector *vector = &vdev->msi_vectors[i];
896
897 if (!vector->use) {
898 continue;
899 }
900
901 if (vector->virq >= 0) {
902 kvm_irqchip_remove_irqfd_notifier(kvm_state,
903 &vector->interrupt, vector->virq);
904 kvm_irqchip_release_virq(kvm_state, vector->virq);
905 vector->virq = -1;
906 } else {
907 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
908 NULL, NULL, NULL);
909 }
910
911 event_notifier_cleanup(&vector->interrupt);
912 }
913
914 vfio_disable_msi_common(vdev);
915
916 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
917 vdev->host.bus, vdev->host.slot, vdev->host.function);
918 }
919
920 /*
921 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
922 */
923 static void vfio_bar_write(void *opaque, hwaddr addr,
924 uint64_t data, unsigned size)
925 {
926 VFIOBAR *bar = opaque;
927 union {
928 uint8_t byte;
929 uint16_t word;
930 uint32_t dword;
931 uint64_t qword;
932 } buf;
933
934 switch (size) {
935 case 1:
936 buf.byte = data;
937 break;
938 case 2:
939 buf.word = cpu_to_le16(data);
940 break;
941 case 4:
942 buf.dword = cpu_to_le32(data);
943 break;
944 default:
945 hw_error("vfio: unsupported write size, %d bytes\n", size);
946 break;
947 }
948
949 if (pwrite(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
950 error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
951 __func__, addr, data, size);
952 }
953
954 #ifdef DEBUG_VFIO
955 {
956 VFIODevice *vdev = container_of(bar, VFIODevice, bars[bar->nr]);
957
958 DPRINTF("%s(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx", 0x%"PRIx64
959 ", %d)\n", __func__, vdev->host.domain, vdev->host.bus,
960 vdev->host.slot, vdev->host.function, bar->nr, addr,
961 data, size);
962 }
963 #endif
964
965 /*
966 * A read or write to a BAR always signals an INTx EOI. This will
967 * do nothing if not pending (including not in INTx mode). We assume
968 * that a BAR access is in response to an interrupt and that BAR
969 * accesses will service the interrupt. Unfortunately, we don't know
970 * which access will service the interrupt, so we're potentially
971 * getting quite a few host interrupts per guest interrupt.
972 */
973 vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
974 }
975
976 static uint64_t vfio_bar_read(void *opaque,
977 hwaddr addr, unsigned size)
978 {
979 VFIOBAR *bar = opaque;
980 union {
981 uint8_t byte;
982 uint16_t word;
983 uint32_t dword;
984 uint64_t qword;
985 } buf;
986 uint64_t data = 0;
987
988 if (pread(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
989 error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
990 __func__, addr, size);
991 return (uint64_t)-1;
992 }
993
994 switch (size) {
995 case 1:
996 data = buf.byte;
997 break;
998 case 2:
999 data = le16_to_cpu(buf.word);
1000 break;
1001 case 4:
1002 data = le32_to_cpu(buf.dword);
1003 break;
1004 default:
1005 hw_error("vfio: unsupported read size, %d bytes\n", size);
1006 break;
1007 }
1008
1009 #ifdef DEBUG_VFIO
1010 {
1011 VFIODevice *vdev = container_of(bar, VFIODevice, bars[bar->nr]);
1012
1013 DPRINTF("%s(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx
1014 ", %d) = 0x%"PRIx64"\n", __func__, vdev->host.domain,
1015 vdev->host.bus, vdev->host.slot, vdev->host.function,
1016 bar->nr, addr, size, data);
1017 }
1018 #endif
1019
1020 /* Same as write above */
1021 vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
1022
1023 return data;
1024 }
1025
1026 static const MemoryRegionOps vfio_bar_ops = {
1027 .read = vfio_bar_read,
1028 .write = vfio_bar_write,
1029 .endianness = DEVICE_LITTLE_ENDIAN,
1030 };
1031
1032 static void vfio_vga_write(void *opaque, hwaddr addr,
1033 uint64_t data, unsigned size)
1034 {
1035 VFIOVGARegion *region = opaque;
1036 VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1037 union {
1038 uint8_t byte;
1039 uint16_t word;
1040 uint32_t dword;
1041 uint64_t qword;
1042 } buf;
1043 off_t offset = vga->fd_offset + region->offset + addr;
1044
1045 switch (size) {
1046 case 1:
1047 buf.byte = data;
1048 break;
1049 case 2:
1050 buf.word = cpu_to_le16(data);
1051 break;
1052 case 4:
1053 buf.dword = cpu_to_le32(data);
1054 break;
1055 default:
1056 hw_error("vfio: unsupported write size, %d bytes\n", size);
1057 break;
1058 }
1059
1060 if (pwrite(vga->fd, &buf, size, offset) != size) {
1061 error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
1062 __func__, region->offset + addr, data, size);
1063 }
1064
1065 DPRINTF("%s(0x%"HWADDR_PRIx", 0x%"PRIx64", %d)\n",
1066 __func__, region->offset + addr, data, size);
1067 }
1068
1069 static uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size)
1070 {
1071 VFIOVGARegion *region = opaque;
1072 VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1073 union {
1074 uint8_t byte;
1075 uint16_t word;
1076 uint32_t dword;
1077 uint64_t qword;
1078 } buf;
1079 uint64_t data = 0;
1080 off_t offset = vga->fd_offset + region->offset + addr;
1081
1082 if (pread(vga->fd, &buf, size, offset) != size) {
1083 error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
1084 __func__, region->offset + addr, size);
1085 return (uint64_t)-1;
1086 }
1087
1088 switch (size) {
1089 case 1:
1090 data = buf.byte;
1091 break;
1092 case 2:
1093 data = le16_to_cpu(buf.word);
1094 break;
1095 case 4:
1096 data = le32_to_cpu(buf.dword);
1097 break;
1098 default:
1099 hw_error("vfio: unsupported read size, %d bytes\n", size);
1100 break;
1101 }
1102
1103 DPRINTF("%s(0x%"HWADDR_PRIx", %d) = 0x%"PRIx64"\n",
1104 __func__, region->offset + addr, size, data);
1105
1106 return data;
1107 }
1108
1109 static const MemoryRegionOps vfio_vga_ops = {
1110 .read = vfio_vga_read,
1111 .write = vfio_vga_write,
1112 .endianness = DEVICE_LITTLE_ENDIAN,
1113 };
1114
1115 /*
1116 * Device specific quirks
1117 */
1118
1119 /* Is range1 fully contained within range2? */
1120 static bool vfio_range_contained(uint64_t first1, uint64_t len1,
1121 uint64_t first2, uint64_t len2) {
1122 return (first1 >= first2 && first1 + len1 <= first2 + len2);
1123 }
1124
1125 static bool vfio_flags_enabled(uint8_t flags, uint8_t mask)
1126 {
1127 return (mask && (flags & mask) == mask);
1128 }
1129
1130 static uint64_t vfio_generic_window_quirk_read(void *opaque,
1131 hwaddr addr, unsigned size)
1132 {
1133 VFIOQuirk *quirk = opaque;
1134 VFIODevice *vdev = quirk->vdev;
1135 uint64_t data;
1136
1137 if (vfio_flags_enabled(quirk->data.flags, quirk->data.read_flags) &&
1138 ranges_overlap(addr, size,
1139 quirk->data.data_offset, quirk->data.data_size)) {
1140 hwaddr offset = addr - quirk->data.data_offset;
1141
1142 if (!vfio_range_contained(addr, size, quirk->data.data_offset,
1143 quirk->data.data_size)) {
1144 hw_error("%s: window data read not fully contained: %s\n",
1145 __func__, memory_region_name(&quirk->mem));
1146 }
1147
1148 data = vfio_pci_read_config(&vdev->pdev,
1149 quirk->data.address_val + offset, size);
1150
1151 DPRINTF("%s read(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx", %d) = 0x%"
1152 PRIx64"\n", memory_region_name(&quirk->mem), vdev->host.domain,
1153 vdev->host.bus, vdev->host.slot, vdev->host.function,
1154 quirk->data.bar, addr, size, data);
1155 } else {
1156 data = vfio_bar_read(&vdev->bars[quirk->data.bar],
1157 addr + quirk->data.base_offset, size);
1158 }
1159
1160 return data;
1161 }
1162
1163 static void vfio_generic_window_quirk_write(void *opaque, hwaddr addr,
1164 uint64_t data, unsigned size)
1165 {
1166 VFIOQuirk *quirk = opaque;
1167 VFIODevice *vdev = quirk->vdev;
1168
1169 if (ranges_overlap(addr, size,
1170 quirk->data.address_offset, quirk->data.address_size)) {
1171
1172 if (addr != quirk->data.address_offset) {
1173 hw_error("%s: offset write into address window: %s\n",
1174 __func__, memory_region_name(&quirk->mem));
1175 }
1176
1177 if ((data & ~quirk->data.address_mask) == quirk->data.address_match) {
1178 quirk->data.flags |= quirk->data.write_flags |
1179 quirk->data.read_flags;
1180 quirk->data.address_val = data & quirk->data.address_mask;
1181 } else {
1182 quirk->data.flags &= ~(quirk->data.write_flags |
1183 quirk->data.read_flags);
1184 }
1185 }
1186
1187 if (vfio_flags_enabled(quirk->data.flags, quirk->data.write_flags) &&
1188 ranges_overlap(addr, size,
1189 quirk->data.data_offset, quirk->data.data_size)) {
1190 hwaddr offset = addr - quirk->data.data_offset;
1191
1192 if (!vfio_range_contained(addr, size, quirk->data.data_offset,
1193 quirk->data.data_size)) {
1194 hw_error("%s: window data write not fully contained: %s\n",
1195 __func__, memory_region_name(&quirk->mem));
1196 }
1197
1198 vfio_pci_write_config(&vdev->pdev,
1199 quirk->data.address_val + offset, data, size);
1200 DPRINTF("%s write(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx", 0x%"
1201 PRIx64", %d)\n", memory_region_name(&quirk->mem),
1202 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1203 vdev->host.function, quirk->data.bar, addr, data, size);
1204 return;
1205 }
1206
1207 vfio_bar_write(&vdev->bars[quirk->data.bar],
1208 addr + quirk->data.base_offset, data, size);
1209 }
1210
1211 static const MemoryRegionOps vfio_generic_window_quirk = {
1212 .read = vfio_generic_window_quirk_read,
1213 .write = vfio_generic_window_quirk_write,
1214 .endianness = DEVICE_LITTLE_ENDIAN,
1215 };
1216
1217 static uint64_t vfio_generic_quirk_read(void *opaque,
1218 hwaddr addr, unsigned size)
1219 {
1220 VFIOQuirk *quirk = opaque;
1221 VFIODevice *vdev = quirk->vdev;
1222 hwaddr base = quirk->data.address_match & TARGET_PAGE_MASK;
1223 hwaddr offset = quirk->data.address_match & ~TARGET_PAGE_MASK;
1224 uint64_t data;
1225
1226 if (vfio_flags_enabled(quirk->data.flags, quirk->data.read_flags) &&
1227 ranges_overlap(addr, size, offset, quirk->data.address_mask + 1)) {
1228 if (!vfio_range_contained(addr, size, offset,
1229 quirk->data.address_mask + 1)) {
1230 hw_error("%s: read not fully contained: %s\n",
1231 __func__, memory_region_name(&quirk->mem));
1232 }
1233
1234 data = vfio_pci_read_config(&vdev->pdev, addr - offset, size);
1235
1236 DPRINTF("%s read(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx", %d) = 0x%"
1237 PRIx64"\n", memory_region_name(&quirk->mem), vdev->host.domain,
1238 vdev->host.bus, vdev->host.slot, vdev->host.function,
1239 quirk->data.bar, addr + base, size, data);
1240 } else {
1241 data = vfio_bar_read(&vdev->bars[quirk->data.bar], addr + base, size);
1242 }
1243
1244 return data;
1245 }
1246
1247 static void vfio_generic_quirk_write(void *opaque, hwaddr addr,
1248 uint64_t data, unsigned size)
1249 {
1250 VFIOQuirk *quirk = opaque;
1251 VFIODevice *vdev = quirk->vdev;
1252 hwaddr base = quirk->data.address_match & TARGET_PAGE_MASK;
1253 hwaddr offset = quirk->data.address_match & ~TARGET_PAGE_MASK;
1254
1255 if (vfio_flags_enabled(quirk->data.flags, quirk->data.write_flags) &&
1256 ranges_overlap(addr, size, offset, quirk->data.address_mask + 1)) {
1257 if (!vfio_range_contained(addr, size, offset,
1258 quirk->data.address_mask + 1)) {
1259 hw_error("%s: write not fully contained: %s\n",
1260 __func__, memory_region_name(&quirk->mem));
1261 }
1262
1263 vfio_pci_write_config(&vdev->pdev, addr - offset, data, size);
1264
1265 DPRINTF("%s write(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx", 0x%"
1266 PRIx64", %d)\n", memory_region_name(&quirk->mem),
1267 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1268 vdev->host.function, quirk->data.bar, addr + base, data, size);
1269 } else {
1270 vfio_bar_write(&vdev->bars[quirk->data.bar], addr + base, data, size);
1271 }
1272 }
1273
1274 static const MemoryRegionOps vfio_generic_quirk = {
1275 .read = vfio_generic_quirk_read,
1276 .write = vfio_generic_quirk_write,
1277 .endianness = DEVICE_LITTLE_ENDIAN,
1278 };
1279
1280 #define PCI_VENDOR_ID_ATI 0x1002
1281
1282 /*
1283 * Radeon HD cards (HD5450 & HD7850) report the upper byte of the I/O port BAR
1284 * through VGA register 0x3c3. On newer cards, the I/O port BAR is always
1285 * BAR4 (older cards like the X550 used BAR1, but we don't care to support
1286 * those). Note that on bare metal, a read of 0x3c3 doesn't always return the
1287 * I/O port BAR address. Originally this was coded to return the virtual BAR
1288 * address only if the physical register read returns the actual BAR address,
1289 * but users have reported greater success if we return the virtual address
1290 * unconditionally.
1291 */
1292 static uint64_t vfio_ati_3c3_quirk_read(void *opaque,
1293 hwaddr addr, unsigned size)
1294 {
1295 VFIOQuirk *quirk = opaque;
1296 VFIODevice *vdev = quirk->vdev;
1297 uint64_t data = vfio_pci_read_config(&vdev->pdev,
1298 PCI_BASE_ADDRESS_0 + (4 * 4) + 1,
1299 size);
1300 DPRINTF("%s(0x3c3, 1) = 0x%"PRIx64"\n", __func__, data);
1301
1302 return data;
1303 }
1304
1305 static const MemoryRegionOps vfio_ati_3c3_quirk = {
1306 .read = vfio_ati_3c3_quirk_read,
1307 .endianness = DEVICE_LITTLE_ENDIAN,
1308 };
1309
1310 static void vfio_vga_probe_ati_3c3_quirk(VFIODevice *vdev)
1311 {
1312 PCIDevice *pdev = &vdev->pdev;
1313 VFIOQuirk *quirk;
1314
1315 if (pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1316 return;
1317 }
1318
1319 /*
1320 * As long as the BAR is >= 256 bytes it will be aligned such that the
1321 * lower byte is always zero. Filter out anything else, if it exists.
1322 */
1323 if (!vdev->bars[4].ioport || vdev->bars[4].size < 256) {
1324 return;
1325 }
1326
1327 quirk = g_malloc0(sizeof(*quirk));
1328 quirk->vdev = vdev;
1329
1330 memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_ati_3c3_quirk, quirk,
1331 "vfio-ati-3c3-quirk", 1);
1332 memory_region_add_subregion(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem,
1333 3 /* offset 3 bytes from 0x3c0 */, &quirk->mem);
1334
1335 QLIST_INSERT_HEAD(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks,
1336 quirk, next);
1337
1338 DPRINTF("Enabled ATI/AMD quirk 0x3c3 BAR4for device %04x:%02x:%02x.%x\n",
1339 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1340 vdev->host.function);
1341 }
1342
1343 /*
1344 * Newer ATI/AMD devices, including HD5450 and HD7850, have a window to PCI
1345 * config space through MMIO BAR2 at offset 0x4000. Nothing seems to access
1346 * the MMIO space directly, but a window to this space is provided through
1347 * I/O port BAR4. Offset 0x0 is the address register and offset 0x4 is the
1348 * data register. When the address is programmed to a range of 0x4000-0x4fff
1349 * PCI configuration space is available. Experimentation seems to indicate
1350 * that only read-only access is provided, but we drop writes when the window
1351 * is enabled to config space nonetheless.
1352 */
1353 static void vfio_probe_ati_bar4_window_quirk(VFIODevice *vdev, int nr)
1354 {
1355 PCIDevice *pdev = &vdev->pdev;
1356 VFIOQuirk *quirk;
1357
1358 if (!vdev->has_vga || nr != 4 ||
1359 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1360 return;
1361 }
1362
1363 quirk = g_malloc0(sizeof(*quirk));
1364 quirk->vdev = vdev;
1365 quirk->data.address_size = 4;
1366 quirk->data.data_offset = 4;
1367 quirk->data.data_size = 4;
1368 quirk->data.address_match = 0x4000;
1369 quirk->data.address_mask = PCIE_CONFIG_SPACE_SIZE - 1;
1370 quirk->data.bar = nr;
1371 quirk->data.read_flags = quirk->data.write_flags = 1;
1372
1373 memory_region_init_io(&quirk->mem, OBJECT(vdev),
1374 &vfio_generic_window_quirk, quirk,
1375 "vfio-ati-bar4-window-quirk", 8);
1376 memory_region_add_subregion_overlap(&vdev->bars[nr].mem,
1377 quirk->data.base_offset, &quirk->mem, 1);
1378
1379 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1380
1381 DPRINTF("Enabled ATI/AMD BAR4 window quirk for device %04x:%02x:%02x.%x\n",
1382 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1383 vdev->host.function);
1384 }
1385
1386 /*
1387 * Trap the BAR2 MMIO window to config space as well.
1388 */
1389 static void vfio_probe_ati_bar2_4000_quirk(VFIODevice *vdev, int nr)
1390 {
1391 PCIDevice *pdev = &vdev->pdev;
1392 VFIOQuirk *quirk;
1393
1394 /* Only enable on newer devices where BAR2 is 64bit */
1395 if (!vdev->has_vga || nr != 2 || !vdev->bars[2].mem64 ||
1396 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1397 return;
1398 }
1399
1400 quirk = g_malloc0(sizeof(*quirk));
1401 quirk->vdev = vdev;
1402 quirk->data.flags = quirk->data.read_flags = quirk->data.write_flags = 1;
1403 quirk->data.address_match = 0x4000;
1404 quirk->data.address_mask = PCIE_CONFIG_SPACE_SIZE - 1;
1405 quirk->data.bar = nr;
1406
1407 memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_generic_quirk, quirk,
1408 "vfio-ati-bar2-4000-quirk",
1409 TARGET_PAGE_ALIGN(quirk->data.address_mask + 1));
1410 memory_region_add_subregion_overlap(&vdev->bars[nr].mem,
1411 quirk->data.address_match & TARGET_PAGE_MASK,
1412 &quirk->mem, 1);
1413
1414 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1415
1416 DPRINTF("Enabled ATI/AMD BAR2 0x4000 quirk for device %04x:%02x:%02x.%x\n",
1417 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1418 vdev->host.function);
1419 }
1420
1421 /*
1422 * Older ATI/AMD cards like the X550 have a similar window to that above.
1423 * I/O port BAR1 provides a window to a mirror of PCI config space located
1424 * in BAR2 at offset 0xf00. We don't care to support such older cards, but
1425 * note it for future reference.
1426 */
1427
1428 #define PCI_VENDOR_ID_NVIDIA 0x10de
1429
1430 /*
1431 * Nvidia has several different methods to get to config space, the
1432 * nouveu project has several of these documented here:
1433 * https://github.com/pathscale/envytools/tree/master/hwdocs
1434 *
1435 * The first quirk is actually not documented in envytools and is found
1436 * on 10de:01d1 (NVIDIA Corporation G72 [GeForce 7300 LE]). This is an
1437 * NV46 chipset. The backdoor uses the legacy VGA I/O ports to access
1438 * the mirror of PCI config space found at BAR0 offset 0x1800. The access
1439 * sequence first writes 0x338 to I/O port 0x3d4. The target offset is
1440 * then written to 0x3d0. Finally 0x538 is written for a read and 0x738
1441 * is written for a write to 0x3d4. The BAR0 offset is then accessible
1442 * through 0x3d0. This quirk doesn't seem to be necessary on newer cards
1443 * that use the I/O port BAR5 window but it doesn't hurt to leave it.
1444 */
1445 enum {
1446 NV_3D0_NONE = 0,
1447 NV_3D0_SELECT,
1448 NV_3D0_WINDOW,
1449 NV_3D0_READ,
1450 NV_3D0_WRITE,
1451 };
1452
1453 static uint64_t vfio_nvidia_3d0_quirk_read(void *opaque,
1454 hwaddr addr, unsigned size)
1455 {
1456 VFIOQuirk *quirk = opaque;
1457 VFIODevice *vdev = quirk->vdev;
1458 PCIDevice *pdev = &vdev->pdev;
1459 uint64_t data = vfio_vga_read(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1460 addr + quirk->data.base_offset, size);
1461
1462 if (quirk->data.flags == NV_3D0_READ && addr == quirk->data.data_offset) {
1463 data = vfio_pci_read_config(pdev, quirk->data.address_val, size);
1464 DPRINTF("%s(0x3d0, %d) = 0x%"PRIx64"\n", __func__, size, data);
1465 }
1466
1467 quirk->data.flags = NV_3D0_NONE;
1468
1469 return data;
1470 }
1471
1472 static void vfio_nvidia_3d0_quirk_write(void *opaque, hwaddr addr,
1473 uint64_t data, unsigned size)
1474 {
1475 VFIOQuirk *quirk = opaque;
1476 VFIODevice *vdev = quirk->vdev;
1477 PCIDevice *pdev = &vdev->pdev;
1478
1479 switch (quirk->data.flags) {
1480 case NV_3D0_NONE:
1481 if (addr == quirk->data.address_offset && data == 0x338) {
1482 quirk->data.flags = NV_3D0_SELECT;
1483 }
1484 break;
1485 case NV_3D0_SELECT:
1486 quirk->data.flags = NV_3D0_NONE;
1487 if (addr == quirk->data.data_offset &&
1488 (data & ~quirk->data.address_mask) == quirk->data.address_match) {
1489 quirk->data.flags = NV_3D0_WINDOW;
1490 quirk->data.address_val = data & quirk->data.address_mask;
1491 }
1492 break;
1493 case NV_3D0_WINDOW:
1494 quirk->data.flags = NV_3D0_NONE;
1495 if (addr == quirk->data.address_offset) {
1496 if (data == 0x538) {
1497 quirk->data.flags = NV_3D0_READ;
1498 } else if (data == 0x738) {
1499 quirk->data.flags = NV_3D0_WRITE;
1500 }
1501 }
1502 break;
1503 case NV_3D0_WRITE:
1504 quirk->data.flags = NV_3D0_NONE;
1505 if (addr == quirk->data.data_offset) {
1506 vfio_pci_write_config(pdev, quirk->data.address_val, data, size);
1507 DPRINTF("%s(0x3d0, 0x%"PRIx64", %d)\n", __func__, data, size);
1508 return;
1509 }
1510 break;
1511 }
1512
1513 vfio_vga_write(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1514 addr + quirk->data.base_offset, data, size);
1515 }
1516
1517 static const MemoryRegionOps vfio_nvidia_3d0_quirk = {
1518 .read = vfio_nvidia_3d0_quirk_read,
1519 .write = vfio_nvidia_3d0_quirk_write,
1520 .endianness = DEVICE_LITTLE_ENDIAN,
1521 };
1522
1523 static void vfio_vga_probe_nvidia_3d0_quirk(VFIODevice *vdev)
1524 {
1525 PCIDevice *pdev = &vdev->pdev;
1526 VFIOQuirk *quirk;
1527
1528 if (pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA ||
1529 !vdev->bars[1].size) {
1530 return;
1531 }
1532
1533 quirk = g_malloc0(sizeof(*quirk));
1534 quirk->vdev = vdev;
1535 quirk->data.base_offset = 0x10;
1536 quirk->data.address_offset = 4;
1537 quirk->data.address_size = 2;
1538 quirk->data.address_match = 0x1800;
1539 quirk->data.address_mask = PCI_CONFIG_SPACE_SIZE - 1;
1540 quirk->data.data_offset = 0;
1541 quirk->data.data_size = 4;
1542
1543 memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_nvidia_3d0_quirk,
1544 quirk, "vfio-nvidia-3d0-quirk", 6);
1545 memory_region_add_subregion(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem,
1546 quirk->data.base_offset, &quirk->mem);
1547
1548 QLIST_INSERT_HEAD(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks,
1549 quirk, next);
1550
1551 DPRINTF("Enabled NVIDIA VGA 0x3d0 quirk for device %04x:%02x:%02x.%x\n",
1552 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1553 vdev->host.function);
1554 }
1555
1556 /*
1557 * The second quirk is documented in envytools. The I/O port BAR5 is just
1558 * a set of address/data ports to the MMIO BARs. The BAR we care about is
1559 * again BAR0. This backdoor is apparently a bit newer than the one above
1560 * so we need to not only trap 256 bytes @0x1800, but all of PCI config
1561 * space, including extended space is available at the 4k @0x88000.
1562 */
1563 enum {
1564 NV_BAR5_ADDRESS = 0x1,
1565 NV_BAR5_ENABLE = 0x2,
1566 NV_BAR5_MASTER = 0x4,
1567 NV_BAR5_VALID = 0x7,
1568 };
1569
1570 static void vfio_nvidia_bar5_window_quirk_write(void *opaque, hwaddr addr,
1571 uint64_t data, unsigned size)
1572 {
1573 VFIOQuirk *quirk = opaque;
1574
1575 switch (addr) {
1576 case 0x0:
1577 if (data & 0x1) {
1578 quirk->data.flags |= NV_BAR5_MASTER;
1579 } else {
1580 quirk->data.flags &= ~NV_BAR5_MASTER;
1581 }
1582 break;
1583 case 0x4:
1584 if (data & 0x1) {
1585 quirk->data.flags |= NV_BAR5_ENABLE;
1586 } else {
1587 quirk->data.flags &= ~NV_BAR5_ENABLE;
1588 }
1589 break;
1590 case 0x8:
1591 if (quirk->data.flags & NV_BAR5_MASTER) {
1592 if ((data & ~0xfff) == 0x88000) {
1593 quirk->data.flags |= NV_BAR5_ADDRESS;
1594 quirk->data.address_val = data & 0xfff;
1595 } else if ((data & ~0xff) == 0x1800) {
1596 quirk->data.flags |= NV_BAR5_ADDRESS;
1597 quirk->data.address_val = data & 0xff;
1598 } else {
1599 quirk->data.flags &= ~NV_BAR5_ADDRESS;
1600 }
1601 }
1602 break;
1603 }
1604
1605 vfio_generic_window_quirk_write(opaque, addr, data, size);
1606 }
1607
1608 static const MemoryRegionOps vfio_nvidia_bar5_window_quirk = {
1609 .read = vfio_generic_window_quirk_read,
1610 .write = vfio_nvidia_bar5_window_quirk_write,
1611 .valid.min_access_size = 4,
1612 .endianness = DEVICE_LITTLE_ENDIAN,
1613 };
1614
1615 static void vfio_probe_nvidia_bar5_window_quirk(VFIODevice *vdev, int nr)
1616 {
1617 PCIDevice *pdev = &vdev->pdev;
1618 VFIOQuirk *quirk;
1619
1620 if (!vdev->has_vga || nr != 5 ||
1621 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1622 return;
1623 }
1624
1625 quirk = g_malloc0(sizeof(*quirk));
1626 quirk->vdev = vdev;
1627 quirk->data.read_flags = quirk->data.write_flags = NV_BAR5_VALID;
1628 quirk->data.address_offset = 0x8;
1629 quirk->data.address_size = 0; /* actually 4, but avoids generic code */
1630 quirk->data.data_offset = 0xc;
1631 quirk->data.data_size = 4;
1632 quirk->data.bar = nr;
1633
1634 memory_region_init_io(&quirk->mem, OBJECT(vdev),
1635 &vfio_nvidia_bar5_window_quirk, quirk,
1636 "vfio-nvidia-bar5-window-quirk", 16);
1637 memory_region_add_subregion_overlap(&vdev->bars[nr].mem, 0, &quirk->mem, 1);
1638
1639 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1640
1641 DPRINTF("Enabled NVIDIA BAR5 window quirk for device %04x:%02x:%02x.%x\n",
1642 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1643 vdev->host.function);
1644 }
1645
1646 /*
1647 * Finally, BAR0 itself. We want to redirect any accesses to either
1648 * 0x1800 or 0x88000 through the PCI config space access functions.
1649 *
1650 * NB - quirk at a page granularity or else they don't seem to work when
1651 * BARs are mmap'd
1652 *
1653 * Here's offset 0x88000...
1654 */
1655 static void vfio_probe_nvidia_bar0_88000_quirk(VFIODevice *vdev, int nr)
1656 {
1657 PCIDevice *pdev = &vdev->pdev;
1658 VFIOQuirk *quirk;
1659
1660 if (!vdev->has_vga || nr != 0 ||
1661 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1662 return;
1663 }
1664
1665 quirk = g_malloc0(sizeof(*quirk));
1666 quirk->vdev = vdev;
1667 quirk->data.flags = quirk->data.read_flags = quirk->data.write_flags = 1;
1668 quirk->data.address_match = 0x88000;
1669 quirk->data.address_mask = PCIE_CONFIG_SPACE_SIZE - 1;
1670 quirk->data.bar = nr;
1671
1672 memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_generic_quirk,
1673 quirk, "vfio-nvidia-bar0-88000-quirk",
1674 TARGET_PAGE_ALIGN(quirk->data.address_mask + 1));
1675 memory_region_add_subregion_overlap(&vdev->bars[nr].mem,
1676 quirk->data.address_match & TARGET_PAGE_MASK,
1677 &quirk->mem, 1);
1678
1679 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1680
1681 DPRINTF("Enabled NVIDIA BAR0 0x88000 quirk for device %04x:%02x:%02x.%x\n",
1682 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1683 vdev->host.function);
1684 }
1685
1686 /*
1687 * And here's the same for BAR0 offset 0x1800...
1688 */
1689 static void vfio_probe_nvidia_bar0_1800_quirk(VFIODevice *vdev, int nr)
1690 {
1691 PCIDevice *pdev = &vdev->pdev;
1692 VFIOQuirk *quirk;
1693
1694 if (!vdev->has_vga || nr != 0 ||
1695 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1696 return;
1697 }
1698
1699 /* Log the chipset ID */
1700 DPRINTF("Nvidia NV%02x\n",
1701 (unsigned int)(vfio_bar_read(&vdev->bars[0], 0, 4) >> 20) & 0xff);
1702
1703 quirk = g_malloc0(sizeof(*quirk));
1704 quirk->vdev = vdev;
1705 quirk->data.flags = quirk->data.read_flags = quirk->data.write_flags = 1;
1706 quirk->data.address_match = 0x1800;
1707 quirk->data.address_mask = PCI_CONFIG_SPACE_SIZE - 1;
1708 quirk->data.bar = nr;
1709
1710 memory_region_init_io(&quirk->mem, OBJECT(vdev), &vfio_generic_quirk, quirk,
1711 "vfio-nvidia-bar0-1800-quirk",
1712 TARGET_PAGE_ALIGN(quirk->data.address_mask + 1));
1713 memory_region_add_subregion_overlap(&vdev->bars[nr].mem,
1714 quirk->data.address_match & TARGET_PAGE_MASK,
1715 &quirk->mem, 1);
1716
1717 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1718
1719 DPRINTF("Enabled NVIDIA BAR0 0x1800 quirk for device %04x:%02x:%02x.%x\n",
1720 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1721 vdev->host.function);
1722 }
1723
1724 /*
1725 * TODO - Some Nvidia devices provide config access to their companion HDA
1726 * device and even to their parent bridge via these config space mirrors.
1727 * Add quirks for those regions.
1728 */
1729
1730 /*
1731 * Common quirk probe entry points.
1732 */
1733 static void vfio_vga_quirk_setup(VFIODevice *vdev)
1734 {
1735 vfio_vga_probe_ati_3c3_quirk(vdev);
1736 vfio_vga_probe_nvidia_3d0_quirk(vdev);
1737 }
1738
1739 static void vfio_vga_quirk_teardown(VFIODevice *vdev)
1740 {
1741 int i;
1742
1743 for (i = 0; i < ARRAY_SIZE(vdev->vga.region); i++) {
1744 while (!QLIST_EMPTY(&vdev->vga.region[i].quirks)) {
1745 VFIOQuirk *quirk = QLIST_FIRST(&vdev->vga.region[i].quirks);
1746 memory_region_del_subregion(&vdev->vga.region[i].mem, &quirk->mem);
1747 QLIST_REMOVE(quirk, next);
1748 g_free(quirk);
1749 }
1750 }
1751 }
1752
1753 static void vfio_bar_quirk_setup(VFIODevice *vdev, int nr)
1754 {
1755 vfio_probe_ati_bar4_window_quirk(vdev, nr);
1756 vfio_probe_ati_bar2_4000_quirk(vdev, nr);
1757 vfio_probe_nvidia_bar5_window_quirk(vdev, nr);
1758 vfio_probe_nvidia_bar0_88000_quirk(vdev, nr);
1759 vfio_probe_nvidia_bar0_1800_quirk(vdev, nr);
1760 }
1761
1762 static void vfio_bar_quirk_teardown(VFIODevice *vdev, int nr)
1763 {
1764 VFIOBAR *bar = &vdev->bars[nr];
1765
1766 while (!QLIST_EMPTY(&bar->quirks)) {
1767 VFIOQuirk *quirk = QLIST_FIRST(&bar->quirks);
1768 memory_region_del_subregion(&bar->mem, &quirk->mem);
1769 QLIST_REMOVE(quirk, next);
1770 g_free(quirk);
1771 }
1772 }
1773
1774 /*
1775 * PCI config space
1776 */
1777 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
1778 {
1779 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1780 uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val;
1781
1782 memcpy(&emu_bits, vdev->emulated_config_bits + addr, len);
1783 emu_bits = le32_to_cpu(emu_bits);
1784
1785 if (emu_bits) {
1786 emu_val = pci_default_read_config(pdev, addr, len);
1787 }
1788
1789 if (~emu_bits & (0xffffffffU >> (32 - len * 8))) {
1790 ssize_t ret;
1791
1792 ret = pread(vdev->fd, &phys_val, len, vdev->config_offset + addr);
1793 if (ret != len) {
1794 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m",
1795 __func__, vdev->host.domain, vdev->host.bus,
1796 vdev->host.slot, vdev->host.function, addr, len);
1797 return -errno;
1798 }
1799 phys_val = le32_to_cpu(phys_val);
1800 }
1801
1802 val = (emu_val & emu_bits) | (phys_val & ~emu_bits);
1803
1804 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, len=0x%x) %x\n", __func__,
1805 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1806 vdev->host.function, addr, len, val);
1807
1808 return val;
1809 }
1810
1811 static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
1812 uint32_t val, int len)
1813 {
1814 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1815 uint32_t val_le = cpu_to_le32(val);
1816
1817 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, 0x%x, len=0x%x)\n", __func__,
1818 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1819 vdev->host.function, addr, val, len);
1820
1821 /* Write everything to VFIO, let it filter out what we can't write */
1822 if (pwrite(vdev->fd, &val_le, len, vdev->config_offset + addr) != len) {
1823 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m",
1824 __func__, vdev->host.domain, vdev->host.bus,
1825 vdev->host.slot, vdev->host.function, addr, val, len);
1826 }
1827
1828 /* MSI/MSI-X Enabling/Disabling */
1829 if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
1830 ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
1831 int is_enabled, was_enabled = msi_enabled(pdev);
1832
1833 pci_default_write_config(pdev, addr, val, len);
1834
1835 is_enabled = msi_enabled(pdev);
1836
1837 if (!was_enabled && is_enabled) {
1838 vfio_enable_msi(vdev);
1839 } else if (was_enabled && !is_enabled) {
1840 vfio_disable_msi(vdev);
1841 }
1842 } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
1843 ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
1844 int is_enabled, was_enabled = msix_enabled(pdev);
1845
1846 pci_default_write_config(pdev, addr, val, len);
1847
1848 is_enabled = msix_enabled(pdev);
1849
1850 if (!was_enabled && is_enabled) {
1851 vfio_enable_msix(vdev);
1852 } else if (was_enabled && !is_enabled) {
1853 vfio_disable_msix(vdev);
1854 }
1855 } else {
1856 /* Write everything to QEMU to keep emulated bits correct */
1857 pci_default_write_config(pdev, addr, val, len);
1858 }
1859 }
1860
1861 /*
1862 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
1863 */
1864 static int vfio_dma_unmap(VFIOContainer *container,
1865 hwaddr iova, ram_addr_t size)
1866 {
1867 struct vfio_iommu_type1_dma_unmap unmap = {
1868 .argsz = sizeof(unmap),
1869 .flags = 0,
1870 .iova = iova,
1871 .size = size,
1872 };
1873
1874 if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
1875 DPRINTF("VFIO_UNMAP_DMA: %d\n", -errno);
1876 return -errno;
1877 }
1878
1879 return 0;
1880 }
1881
1882 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
1883 ram_addr_t size, void *vaddr, bool readonly)
1884 {
1885 struct vfio_iommu_type1_dma_map map = {
1886 .argsz = sizeof(map),
1887 .flags = VFIO_DMA_MAP_FLAG_READ,
1888 .vaddr = (__u64)(uintptr_t)vaddr,
1889 .iova = iova,
1890 .size = size,
1891 };
1892
1893 if (!readonly) {
1894 map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
1895 }
1896
1897 /*
1898 * Try the mapping, if it fails with EBUSY, unmap the region and try
1899 * again. This shouldn't be necessary, but we sometimes see it in
1900 * the the VGA ROM space.
1901 */
1902 if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
1903 (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
1904 ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
1905 return 0;
1906 }
1907
1908 DPRINTF("VFIO_MAP_DMA: %d\n", -errno);
1909 return -errno;
1910 }
1911
1912 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
1913 {
1914 return !memory_region_is_ram(section->mr);
1915 }
1916
1917 static void vfio_listener_region_add(MemoryListener *listener,
1918 MemoryRegionSection *section)
1919 {
1920 VFIOContainer *container = container_of(listener, VFIOContainer,
1921 iommu_data.listener);
1922 hwaddr iova, end;
1923 void *vaddr;
1924 int ret;
1925
1926 assert(!memory_region_is_iommu(section->mr));
1927
1928 if (vfio_listener_skipped_section(section)) {
1929 DPRINTF("SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
1930 section->offset_within_address_space,
1931 section->offset_within_address_space + section->size - 1);
1932 return;
1933 }
1934
1935 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
1936 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
1937 error_report("%s received unaligned region", __func__);
1938 return;
1939 }
1940
1941 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
1942 end = (section->offset_within_address_space + int128_get64(section->size)) &
1943 TARGET_PAGE_MASK;
1944
1945 if (iova >= end) {
1946 return;
1947 }
1948
1949 vaddr = memory_region_get_ram_ptr(section->mr) +
1950 section->offset_within_region +
1951 (iova - section->offset_within_address_space);
1952
1953 DPRINTF("region_add %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
1954 iova, end - 1, vaddr);
1955
1956 memory_region_ref(section->mr);
1957 ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
1958 if (ret) {
1959 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
1960 "0x%"HWADDR_PRIx", %p) = %d (%m)",
1961 container, iova, end - iova, vaddr, ret);
1962 }
1963 }
1964
1965 static void vfio_listener_region_del(MemoryListener *listener,
1966 MemoryRegionSection *section)
1967 {
1968 VFIOContainer *container = container_of(listener, VFIOContainer,
1969 iommu_data.listener);
1970 hwaddr iova, end;
1971 int ret;
1972
1973 if (vfio_listener_skipped_section(section)) {
1974 DPRINTF("SKIPPING region_del %"HWADDR_PRIx" - %"PRIx64"\n",
1975 section->offset_within_address_space,
1976 section->offset_within_address_space + section->size - 1);
1977 return;
1978 }
1979
1980 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
1981 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
1982 error_report("%s received unaligned region", __func__);
1983 return;
1984 }
1985
1986 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
1987 end = (section->offset_within_address_space + int128_get64(section->size)) &
1988 TARGET_PAGE_MASK;
1989
1990 if (iova >= end) {
1991 return;
1992 }
1993
1994 DPRINTF("region_del %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
1995 iova, end - 1);
1996
1997 ret = vfio_dma_unmap(container, iova, end - iova);
1998 memory_region_unref(section->mr);
1999 if (ret) {
2000 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
2001 "0x%"HWADDR_PRIx") = %d (%m)",
2002 container, iova, end - iova, ret);
2003 }
2004 }
2005
2006 static MemoryListener vfio_memory_listener = {
2007 .region_add = vfio_listener_region_add,
2008 .region_del = vfio_listener_region_del,
2009 };
2010
2011 static void vfio_listener_release(VFIOContainer *container)
2012 {
2013 memory_listener_unregister(&container->iommu_data.listener);
2014 }
2015
2016 /*
2017 * Interrupt setup
2018 */
2019 static void vfio_disable_interrupts(VFIODevice *vdev)
2020 {
2021 switch (vdev->interrupt) {
2022 case VFIO_INT_INTx:
2023 vfio_disable_intx(vdev);
2024 break;
2025 case VFIO_INT_MSI:
2026 vfio_disable_msi(vdev);
2027 break;
2028 case VFIO_INT_MSIX:
2029 vfio_disable_msix(vdev);
2030 break;
2031 }
2032 }
2033
2034 static int vfio_setup_msi(VFIODevice *vdev, int pos)
2035 {
2036 uint16_t ctrl;
2037 bool msi_64bit, msi_maskbit;
2038 int ret, entries;
2039
2040 if (pread(vdev->fd, &ctrl, sizeof(ctrl),
2041 vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
2042 return -errno;
2043 }
2044 ctrl = le16_to_cpu(ctrl);
2045
2046 msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
2047 msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
2048 entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
2049
2050 DPRINTF("%04x:%02x:%02x.%x PCI MSI CAP @0x%x\n", vdev->host.domain,
2051 vdev->host.bus, vdev->host.slot, vdev->host.function, pos);
2052
2053 ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit);
2054 if (ret < 0) {
2055 if (ret == -ENOTSUP) {
2056 return 0;
2057 }
2058 error_report("vfio: msi_init failed");
2059 return ret;
2060 }
2061 vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
2062
2063 return 0;
2064 }
2065
2066 /*
2067 * We don't have any control over how pci_add_capability() inserts
2068 * capabilities into the chain. In order to setup MSI-X we need a
2069 * MemoryRegion for the BAR. In order to setup the BAR and not
2070 * attempt to mmap the MSI-X table area, which VFIO won't allow, we
2071 * need to first look for where the MSI-X table lives. So we
2072 * unfortunately split MSI-X setup across two functions.
2073 */
2074 static int vfio_early_setup_msix(VFIODevice *vdev)
2075 {
2076 uint8_t pos;
2077 uint16_t ctrl;
2078 uint32_t table, pba;
2079
2080 pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
2081 if (!pos) {
2082 return 0;
2083 }
2084
2085 if (pread(vdev->fd, &ctrl, sizeof(ctrl),
2086 vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
2087 return -errno;
2088 }
2089
2090 if (pread(vdev->fd, &table, sizeof(table),
2091 vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
2092 return -errno;
2093 }
2094
2095 if (pread(vdev->fd, &pba, sizeof(pba),
2096 vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
2097 return -errno;
2098 }
2099
2100 ctrl = le16_to_cpu(ctrl);
2101 table = le32_to_cpu(table);
2102 pba = le32_to_cpu(pba);
2103
2104 vdev->msix = g_malloc0(sizeof(*(vdev->msix)));
2105 vdev->msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
2106 vdev->msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
2107 vdev->msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
2108 vdev->msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
2109 vdev->msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
2110
2111 DPRINTF("%04x:%02x:%02x.%x "
2112 "PCI MSI-X CAP @0x%x, BAR %d, offset 0x%x, entries %d\n",
2113 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2114 vdev->host.function, pos, vdev->msix->table_bar,
2115 vdev->msix->table_offset, vdev->msix->entries);
2116
2117 return 0;
2118 }
2119
2120 static int vfio_setup_msix(VFIODevice *vdev, int pos)
2121 {
2122 int ret;
2123
2124 ret = msix_init(&vdev->pdev, vdev->msix->entries,
2125 &vdev->bars[vdev->msix->table_bar].mem,
2126 vdev->msix->table_bar, vdev->msix->table_offset,
2127 &vdev->bars[vdev->msix->pba_bar].mem,
2128 vdev->msix->pba_bar, vdev->msix->pba_offset, pos);
2129 if (ret < 0) {
2130 if (ret == -ENOTSUP) {
2131 return 0;
2132 }
2133 error_report("vfio: msix_init failed");
2134 return ret;
2135 }
2136
2137 return 0;
2138 }
2139
2140 static void vfio_teardown_msi(VFIODevice *vdev)
2141 {
2142 msi_uninit(&vdev->pdev);
2143
2144 if (vdev->msix) {
2145 msix_uninit(&vdev->pdev, &vdev->bars[vdev->msix->table_bar].mem,
2146 &vdev->bars[vdev->msix->pba_bar].mem);
2147 }
2148 }
2149
2150 /*
2151 * Resource setup
2152 */
2153 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled)
2154 {
2155 int i;
2156
2157 for (i = 0; i < PCI_ROM_SLOT; i++) {
2158 VFIOBAR *bar = &vdev->bars[i];
2159
2160 if (!bar->size) {
2161 continue;
2162 }
2163
2164 memory_region_set_enabled(&bar->mmap_mem, enabled);
2165 if (vdev->msix && vdev->msix->table_bar == i) {
2166 memory_region_set_enabled(&vdev->msix->mmap_mem, enabled);
2167 }
2168 }
2169 }
2170
2171 static void vfio_unmap_bar(VFIODevice *vdev, int nr)
2172 {
2173 VFIOBAR *bar = &vdev->bars[nr];
2174
2175 if (!bar->size) {
2176 return;
2177 }
2178
2179 vfio_bar_quirk_teardown(vdev, nr);
2180
2181 memory_region_del_subregion(&bar->mem, &bar->mmap_mem);
2182 munmap(bar->mmap, memory_region_size(&bar->mmap_mem));
2183
2184 if (vdev->msix && vdev->msix->table_bar == nr) {
2185 memory_region_del_subregion(&bar->mem, &vdev->msix->mmap_mem);
2186 munmap(vdev->msix->mmap, memory_region_size(&vdev->msix->mmap_mem));
2187 }
2188
2189 memory_region_destroy(&bar->mem);
2190 }
2191
2192 static int vfio_mmap_bar(VFIODevice *vdev, VFIOBAR *bar,
2193 MemoryRegion *mem, MemoryRegion *submem,
2194 void **map, size_t size, off_t offset,
2195 const char *name)
2196 {
2197 int ret = 0;
2198
2199 if (VFIO_ALLOW_MMAP && size && bar->flags & VFIO_REGION_INFO_FLAG_MMAP) {
2200 int prot = 0;
2201
2202 if (bar->flags & VFIO_REGION_INFO_FLAG_READ) {
2203 prot |= PROT_READ;
2204 }
2205
2206 if (bar->flags & VFIO_REGION_INFO_FLAG_WRITE) {
2207 prot |= PROT_WRITE;
2208 }
2209
2210 *map = mmap(NULL, size, prot, MAP_SHARED,
2211 bar->fd, bar->fd_offset + offset);
2212 if (*map == MAP_FAILED) {
2213 *map = NULL;
2214 ret = -errno;
2215 goto empty_region;
2216 }
2217
2218 memory_region_init_ram_ptr(submem, OBJECT(vdev), name, size, *map);
2219 } else {
2220 empty_region:
2221 /* Create a zero sized sub-region to make cleanup easy. */
2222 memory_region_init(submem, OBJECT(vdev), name, 0);
2223 }
2224
2225 memory_region_add_subregion(mem, offset, submem);
2226
2227 return ret;
2228 }
2229
2230 static void vfio_map_bar(VFIODevice *vdev, int nr)
2231 {
2232 VFIOBAR *bar = &vdev->bars[nr];
2233 unsigned size = bar->size;
2234 char name[64];
2235 uint32_t pci_bar;
2236 uint8_t type;
2237 int ret;
2238
2239 /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
2240 if (!size) {
2241 return;
2242 }
2243
2244 snprintf(name, sizeof(name), "VFIO %04x:%02x:%02x.%x BAR %d",
2245 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2246 vdev->host.function, nr);
2247
2248 /* Determine what type of BAR this is for registration */
2249 ret = pread(vdev->fd, &pci_bar, sizeof(pci_bar),
2250 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
2251 if (ret != sizeof(pci_bar)) {
2252 error_report("vfio: Failed to read BAR %d (%m)", nr);
2253 return;
2254 }
2255
2256 pci_bar = le32_to_cpu(pci_bar);
2257 bar->ioport = (pci_bar & PCI_BASE_ADDRESS_SPACE_IO);
2258 bar->mem64 = bar->ioport ? 0 : (pci_bar & PCI_BASE_ADDRESS_MEM_TYPE_64);
2259 type = pci_bar & (bar->ioport ? ~PCI_BASE_ADDRESS_IO_MASK :
2260 ~PCI_BASE_ADDRESS_MEM_MASK);
2261
2262 /* A "slow" read/write mapping underlies all BARs */
2263 memory_region_init_io(&bar->mem, OBJECT(vdev), &vfio_bar_ops,
2264 bar, name, size);
2265 pci_register_bar(&vdev->pdev, nr, type, &bar->mem);
2266
2267 /*
2268 * We can't mmap areas overlapping the MSIX vector table, so we
2269 * potentially insert a direct-mapped subregion before and after it.
2270 */
2271 if (vdev->msix && vdev->msix->table_bar == nr) {
2272 size = vdev->msix->table_offset & TARGET_PAGE_MASK;
2273 }
2274
2275 strncat(name, " mmap", sizeof(name) - strlen(name) - 1);
2276 if (vfio_mmap_bar(vdev, bar, &bar->mem,
2277 &bar->mmap_mem, &bar->mmap, size, 0, name)) {
2278 error_report("%s unsupported. Performance may be slow", name);
2279 }
2280
2281 if (vdev->msix && vdev->msix->table_bar == nr) {
2282 unsigned start;
2283
2284 start = TARGET_PAGE_ALIGN(vdev->msix->table_offset +
2285 (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
2286
2287 size = start < bar->size ? bar->size - start : 0;
2288 strncat(name, " msix-hi", sizeof(name) - strlen(name) - 1);
2289 /* VFIOMSIXInfo contains another MemoryRegion for this mapping */
2290 if (vfio_mmap_bar(vdev, bar, &bar->mem, &vdev->msix->mmap_mem,
2291 &vdev->msix->mmap, size, start, name)) {
2292 error_report("%s unsupported. Performance may be slow", name);
2293 }
2294 }
2295
2296 vfio_bar_quirk_setup(vdev, nr);
2297 }
2298
2299 static void vfio_map_bars(VFIODevice *vdev)
2300 {
2301 int i;
2302
2303 for (i = 0; i < PCI_ROM_SLOT; i++) {
2304 vfio_map_bar(vdev, i);
2305 }
2306
2307 if (vdev->has_vga) {
2308 memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_MEM].mem,
2309 OBJECT(vdev), &vfio_vga_ops,
2310 &vdev->vga.region[QEMU_PCI_VGA_MEM],
2311 "vfio-vga-mmio@0xa0000",
2312 QEMU_PCI_VGA_MEM_SIZE);
2313 memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem,
2314 OBJECT(vdev), &vfio_vga_ops,
2315 &vdev->vga.region[QEMU_PCI_VGA_IO_LO],
2316 "vfio-vga-io@0x3b0",
2317 QEMU_PCI_VGA_IO_LO_SIZE);
2318 memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem,
2319 OBJECT(vdev), &vfio_vga_ops,
2320 &vdev->vga.region[QEMU_PCI_VGA_IO_HI],
2321 "vfio-vga-io@0x3c0",
2322 QEMU_PCI_VGA_IO_HI_SIZE);
2323
2324 pci_register_vga(&vdev->pdev, &vdev->vga.region[QEMU_PCI_VGA_MEM].mem,
2325 &vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem,
2326 &vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem);
2327 vfio_vga_quirk_setup(vdev);
2328 }
2329 }
2330
2331 static void vfio_unmap_bars(VFIODevice *vdev)
2332 {
2333 int i;
2334
2335 for (i = 0; i < PCI_ROM_SLOT; i++) {
2336 vfio_unmap_bar(vdev, i);
2337 }
2338
2339 if (vdev->has_vga) {
2340 vfio_vga_quirk_teardown(vdev);
2341 pci_unregister_vga(&vdev->pdev);
2342 memory_region_destroy(&vdev->vga.region[QEMU_PCI_VGA_MEM].mem);
2343 memory_region_destroy(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem);
2344 memory_region_destroy(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem);
2345 }
2346 }
2347
2348 /*
2349 * General setup
2350 */
2351 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
2352 {
2353 uint8_t tmp, next = 0xff;
2354
2355 for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
2356 tmp = pdev->config[tmp + 1]) {
2357 if (tmp > pos && tmp < next) {
2358 next = tmp;
2359 }
2360 }
2361
2362 return next - pos;
2363 }
2364
2365 static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask)
2366 {
2367 pci_set_word(buf, (pci_get_word(buf) & ~mask) | val);
2368 }
2369
2370 static void vfio_add_emulated_word(VFIODevice *vdev, int pos,
2371 uint16_t val, uint16_t mask)
2372 {
2373 vfio_set_word_bits(vdev->pdev.config + pos, val, mask);
2374 vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask);
2375 vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask);
2376 }
2377
2378 static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask)
2379 {
2380 pci_set_long(buf, (pci_get_long(buf) & ~mask) | val);
2381 }
2382
2383 static void vfio_add_emulated_long(VFIODevice *vdev, int pos,
2384 uint32_t val, uint32_t mask)
2385 {
2386 vfio_set_long_bits(vdev->pdev.config + pos, val, mask);
2387 vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask);
2388 vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask);
2389 }
2390
2391 static int vfio_setup_pcie_cap(VFIODevice *vdev, int pos, uint8_t size)
2392 {
2393 uint16_t flags;
2394 uint8_t type;
2395
2396 flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS);
2397 type = (flags & PCI_EXP_FLAGS_TYPE) >> 4;
2398
2399 if (type != PCI_EXP_TYPE_ENDPOINT &&
2400 type != PCI_EXP_TYPE_LEG_END &&
2401 type != PCI_EXP_TYPE_RC_END) {
2402
2403 error_report("vfio: Assignment of PCIe type 0x%x "
2404 "devices is not currently supported", type);
2405 return -EINVAL;
2406 }
2407
2408 if (!pci_bus_is_express(vdev->pdev.bus)) {
2409 /*
2410 * Use express capability as-is on PCI bus. It doesn't make much
2411 * sense to even expose, but some drivers (ex. tg3) depend on it
2412 * and guests don't seem to be particular about it. We'll need
2413 * to revist this or force express devices to express buses if we
2414 * ever expose an IOMMU to the guest.
2415 */
2416 } else if (pci_bus_is_root(vdev->pdev.bus)) {
2417 /*
2418 * On a Root Complex bus Endpoints become Root Complex Integrated
2419 * Endpoints, which changes the type and clears the LNK & LNK2 fields.
2420 */
2421 if (type == PCI_EXP_TYPE_ENDPOINT) {
2422 vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
2423 PCI_EXP_TYPE_RC_END << 4,
2424 PCI_EXP_FLAGS_TYPE);
2425
2426 /* Link Capabilities, Status, and Control goes away */
2427 if (size > PCI_EXP_LNKCTL) {
2428 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0);
2429 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
2430 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0);
2431
2432 #ifndef PCI_EXP_LNKCAP2
2433 #define PCI_EXP_LNKCAP2 44
2434 #endif
2435 #ifndef PCI_EXP_LNKSTA2
2436 #define PCI_EXP_LNKSTA2 50
2437 #endif
2438 /* Link 2 Capabilities, Status, and Control goes away */
2439 if (size > PCI_EXP_LNKCAP2) {
2440 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0);
2441 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0);
2442 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0);
2443 }
2444 }
2445
2446 } else if (type == PCI_EXP_TYPE_LEG_END) {
2447 /*
2448 * Legacy endpoints don't belong on the root complex. Windows
2449 * seems to be happier with devices if we skip the capability.
2450 */
2451 return 0;
2452 }
2453
2454 } else {
2455 /*
2456 * Convert Root Complex Integrated Endpoints to regular endpoints.
2457 * These devices don't support LNK/LNK2 capabilities, so make them up.
2458 */
2459 if (type == PCI_EXP_TYPE_RC_END) {
2460 vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
2461 PCI_EXP_TYPE_ENDPOINT << 4,
2462 PCI_EXP_FLAGS_TYPE);
2463 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP,
2464 PCI_EXP_LNK_MLW_1 | PCI_EXP_LNK_LS_25, ~0);
2465 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
2466 }
2467
2468 /* Mark the Link Status bits as emulated to allow virtual negotiation */
2469 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA,
2470 pci_get_word(vdev->pdev.config + pos +
2471 PCI_EXP_LNKSTA),
2472 PCI_EXP_LNKCAP_MLW | PCI_EXP_LNKCAP_SLS);
2473 }
2474
2475 pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size);
2476 if (pos >= 0) {
2477 vdev->pdev.exp.exp_cap = pos;
2478 }
2479
2480 return pos;
2481 }
2482
2483 static int vfio_add_std_cap(VFIODevice *vdev, uint8_t pos)
2484 {
2485 PCIDevice *pdev = &vdev->pdev;
2486 uint8_t cap_id, next, size;
2487 int ret;
2488
2489 cap_id = pdev->config[pos];
2490 next = pdev->config[pos + 1];
2491
2492 /*
2493 * If it becomes important to configure capabilities to their actual
2494 * size, use this as the default when it's something we don't recognize.
2495 * Since QEMU doesn't actually handle many of the config accesses,
2496 * exact size doesn't seem worthwhile.
2497 */
2498 size = vfio_std_cap_max_size(pdev, pos);
2499
2500 /*
2501 * pci_add_capability always inserts the new capability at the head
2502 * of the chain. Therefore to end up with a chain that matches the
2503 * physical device, we insert from the end by making this recursive.
2504 * This is also why we pre-caclulate size above as cached config space
2505 * will be changed as we unwind the stack.
2506 */
2507 if (next) {
2508 ret = vfio_add_std_cap(vdev, next);
2509 if (ret) {
2510 return ret;
2511 }
2512 } else {
2513 /* Begin the rebuild, use QEMU emulated list bits */
2514 pdev->config[PCI_CAPABILITY_LIST] = 0;
2515 vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
2516 vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
2517 }
2518
2519 /* Use emulated next pointer to allow dropping caps */
2520 pci_set_byte(vdev->emulated_config_bits + pos + 1, 0xff);
2521
2522 switch (cap_id) {
2523 case PCI_CAP_ID_MSI:
2524 ret = vfio_setup_msi(vdev, pos);
2525 break;
2526 case PCI_CAP_ID_EXP:
2527 ret = vfio_setup_pcie_cap(vdev, pos, size);
2528 break;
2529 case PCI_CAP_ID_MSIX:
2530 ret = vfio_setup_msix(vdev, pos);
2531 break;
2532 case PCI_CAP_ID_PM:
2533 vdev->pm_cap = pos;
2534 default:
2535 ret = pci_add_capability(pdev, cap_id, pos, size);
2536 break;
2537 }
2538
2539 if (ret < 0) {
2540 error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability "
2541 "0x%x[0x%x]@0x%x: %d", vdev->host.domain,
2542 vdev->host.bus, vdev->host.slot, vdev->host.function,
2543 cap_id, size, pos, ret);
2544 return ret;
2545 }
2546
2547 return 0;
2548 }
2549
2550 static int vfio_add_capabilities(VFIODevice *vdev)
2551 {
2552 PCIDevice *pdev = &vdev->pdev;
2553
2554 if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
2555 !pdev->config[PCI_CAPABILITY_LIST]) {
2556 return 0; /* Nothing to add */
2557 }
2558
2559 return vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST]);
2560 }
2561
2562 static int vfio_load_rom(VFIODevice *vdev)
2563 {
2564 uint64_t size = vdev->rom_size;
2565 char name[32];
2566 off_t off = 0, voff = vdev->rom_offset;
2567 ssize_t bytes;
2568 void *ptr;
2569
2570 /* If loading ROM from file, pci handles it */
2571 if (vdev->pdev.romfile || !vdev->pdev.rom_bar || !size) {
2572 return 0;
2573 }
2574
2575 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
2576 vdev->host.bus, vdev->host.slot, vdev->host.function);
2577
2578 snprintf(name, sizeof(name), "vfio[%04x:%02x:%02x.%x].rom",
2579 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2580 vdev->host.function);
2581 memory_region_init_ram(&vdev->pdev.rom, OBJECT(vdev), name, size);
2582 ptr = memory_region_get_ram_ptr(&vdev->pdev.rom);
2583 memset(ptr, 0xff, size);
2584
2585 while (size) {
2586 bytes = pread(vdev->fd, ptr + off, size, voff + off);
2587 if (bytes == 0) {
2588 break; /* expect that we could get back less than the ROM BAR */
2589 } else if (bytes > 0) {
2590 off += bytes;
2591 size -= bytes;
2592 } else {
2593 if (errno == EINTR || errno == EAGAIN) {
2594 continue;
2595 }
2596 error_report("vfio: Error reading device ROM: %m");
2597 memory_region_destroy(&vdev->pdev.rom);
2598 return -errno;
2599 }
2600 }
2601
2602 pci_register_bar(&vdev->pdev, PCI_ROM_SLOT, 0, &vdev->pdev.rom);
2603 vdev->pdev.has_rom = true;
2604 return 0;
2605 }
2606
2607 static int vfio_connect_container(VFIOGroup *group)
2608 {
2609 VFIOContainer *container;
2610 int ret, fd;
2611
2612 if (group->container) {
2613 return 0;
2614 }
2615
2616 QLIST_FOREACH(container, &container_list, next) {
2617 if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
2618 group->container = container;
2619 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2620 return 0;
2621 }
2622 }
2623
2624 fd = qemu_open("/dev/vfio/vfio", O_RDWR);
2625 if (fd < 0) {
2626 error_report("vfio: failed to open /dev/vfio/vfio: %m");
2627 return -errno;
2628 }
2629
2630 ret = ioctl(fd, VFIO_GET_API_VERSION);
2631 if (ret != VFIO_API_VERSION) {
2632 error_report("vfio: supported vfio version: %d, "
2633 "reported version: %d", VFIO_API_VERSION, ret);
2634 close(fd);
2635 return -EINVAL;
2636 }
2637
2638 container = g_malloc0(sizeof(*container));
2639 container->fd = fd;
2640
2641 if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) {
2642 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
2643 if (ret) {
2644 error_report("vfio: failed to set group container: %m");
2645 g_free(container);
2646 close(fd);
2647 return -errno;
2648 }
2649
2650 ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU);
2651 if (ret) {
2652 error_report("vfio: failed to set iommu for container: %m");
2653 g_free(container);
2654 close(fd);
2655 return -errno;
2656 }
2657
2658 container->iommu_data.listener = vfio_memory_listener;
2659 container->iommu_data.release = vfio_listener_release;
2660
2661 memory_listener_register(&container->iommu_data.listener, &address_space_memory);
2662 } else {
2663 error_report("vfio: No available IOMMU models");
2664 g_free(container);
2665 close(fd);
2666 return -EINVAL;
2667 }
2668
2669 QLIST_INIT(&container->group_list);
2670 QLIST_INSERT_HEAD(&container_list, container, next);
2671
2672 group->container = container;
2673 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2674
2675 return 0;
2676 }
2677
2678 static void vfio_disconnect_container(VFIOGroup *group)
2679 {
2680 VFIOContainer *container = group->container;
2681
2682 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
2683 error_report("vfio: error disconnecting group %d from container",
2684 group->groupid);
2685 }
2686
2687 QLIST_REMOVE(group, container_next);
2688 group->container = NULL;
2689
2690 if (QLIST_EMPTY(&container->group_list)) {
2691 if (container->iommu_data.release) {
2692 container->iommu_data.release(container);
2693 }
2694 QLIST_REMOVE(container, next);
2695 DPRINTF("vfio_disconnect_container: close container->fd\n");
2696 close(container->fd);
2697 g_free(container);
2698 }
2699 }
2700
2701 static VFIOGroup *vfio_get_group(int groupid)
2702 {
2703 VFIOGroup *group;
2704 char path[32];
2705 struct vfio_group_status status = { .argsz = sizeof(status) };
2706
2707 QLIST_FOREACH(group, &group_list, next) {
2708 if (group->groupid == groupid) {
2709 return group;
2710 }
2711 }
2712
2713 group = g_malloc0(sizeof(*group));
2714
2715 snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
2716 group->fd = qemu_open(path, O_RDWR);
2717 if (group->fd < 0) {
2718 error_report("vfio: error opening %s: %m", path);
2719 g_free(group);
2720 return NULL;
2721 }
2722
2723 if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
2724 error_report("vfio: error getting group status: %m");
2725 close(group->fd);
2726 g_free(group);
2727 return NULL;
2728 }
2729
2730 if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
2731 error_report("vfio: error, group %d is not viable, please ensure "
2732 "all devices within the iommu_group are bound to their "
2733 "vfio bus driver.", groupid);
2734 close(group->fd);
2735 g_free(group);
2736 return NULL;
2737 }
2738
2739 group->groupid = groupid;
2740 QLIST_INIT(&group->device_list);
2741
2742 if (vfio_connect_container(group)) {
2743 error_report("vfio: failed to setup container for group %d", groupid);
2744 close(group->fd);
2745 g_free(group);
2746 return NULL;
2747 }
2748
2749 QLIST_INSERT_HEAD(&group_list, group, next);
2750
2751 return group;
2752 }
2753
2754 static void vfio_put_group(VFIOGroup *group)
2755 {
2756 if (!QLIST_EMPTY(&group->device_list)) {
2757 return;
2758 }
2759
2760 vfio_disconnect_container(group);
2761 QLIST_REMOVE(group, next);
2762 DPRINTF("vfio_put_group: close group->fd\n");
2763 close(group->fd);
2764 g_free(group);
2765 }
2766
2767 static int vfio_get_device(VFIOGroup *group, const char *name, VFIODevice *vdev)
2768 {
2769 struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
2770 struct vfio_region_info reg_info = { .argsz = sizeof(reg_info) };
2771 struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) };
2772 int ret, i;
2773
2774 ret = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
2775 if (ret < 0) {
2776 error_report("vfio: error getting device %s from group %d: %m",
2777 name, group->groupid);
2778 error_printf("Verify all devices in group %d are bound to vfio-pci "
2779 "or pci-stub and not already in use\n", group->groupid);
2780 return ret;
2781 }
2782
2783 vdev->fd = ret;
2784 vdev->group = group;
2785 QLIST_INSERT_HEAD(&group->device_list, vdev, next);
2786
2787 /* Sanity check device */
2788 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_INFO, &dev_info);
2789 if (ret) {
2790 error_report("vfio: error getting device info: %m");
2791 goto error;
2792 }
2793
2794 DPRINTF("Device %s flags: %u, regions: %u, irgs: %u\n", name,
2795 dev_info.flags, dev_info.num_regions, dev_info.num_irqs);
2796
2797 if (!(dev_info.flags & VFIO_DEVICE_FLAGS_PCI)) {
2798 error_report("vfio: Um, this isn't a PCI device");
2799 goto error;
2800 }
2801
2802 vdev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
2803 if (!vdev->reset_works) {
2804 error_report("Warning, device %s does not support reset", name);
2805 }
2806
2807 if (dev_info.num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
2808 error_report("vfio: unexpected number of io regions %u",
2809 dev_info.num_regions);
2810 goto error;
2811 }
2812
2813 if (dev_info.num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
2814 error_report("vfio: unexpected number of irqs %u", dev_info.num_irqs);
2815 goto error;
2816 }
2817
2818 for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
2819 reg_info.index = i;
2820
2821 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
2822 if (ret) {
2823 error_report("vfio: Error getting region %d info: %m", i);
2824 goto error;
2825 }
2826
2827 DPRINTF("Device %s region %d:\n", name, i);
2828 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
2829 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
2830 (unsigned long)reg_info.flags);
2831
2832 vdev->bars[i].flags = reg_info.flags;
2833 vdev->bars[i].size = reg_info.size;
2834 vdev->bars[i].fd_offset = reg_info.offset;
2835 vdev->bars[i].fd = vdev->fd;
2836 vdev->bars[i].nr = i;
2837 QLIST_INIT(&vdev->bars[i].quirks);
2838 }
2839
2840 reg_info.index = VFIO_PCI_ROM_REGION_INDEX;
2841
2842 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
2843 if (ret) {
2844 error_report("vfio: Error getting ROM info: %m");
2845 goto error;
2846 }
2847
2848 DPRINTF("Device %s ROM:\n", name);
2849 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
2850 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
2851 (unsigned long)reg_info.flags);
2852
2853 vdev->rom_size = reg_info.size;
2854 vdev->rom_offset = reg_info.offset;
2855
2856 reg_info.index = VFIO_PCI_CONFIG_REGION_INDEX;
2857
2858 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
2859 if (ret) {
2860 error_report("vfio: Error getting config info: %m");
2861 goto error;
2862 }
2863
2864 DPRINTF("Device %s config:\n", name);
2865 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
2866 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
2867 (unsigned long)reg_info.flags);
2868
2869 vdev->config_size = reg_info.size;
2870 if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
2871 vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
2872 }
2873 vdev->config_offset = reg_info.offset;
2874
2875 if ((vdev->features & VFIO_FEATURE_ENABLE_VGA) &&
2876 dev_info.num_regions > VFIO_PCI_VGA_REGION_INDEX) {
2877 struct vfio_region_info vga_info = {
2878 .argsz = sizeof(vga_info),
2879 .index = VFIO_PCI_VGA_REGION_INDEX,
2880 };
2881
2882 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &vga_info);
2883 if (ret) {
2884 error_report(
2885 "vfio: Device does not support requested feature x-vga");
2886 goto error;
2887 }
2888
2889 if (!(vga_info.flags & VFIO_REGION_INFO_FLAG_READ) ||
2890 !(vga_info.flags & VFIO_REGION_INFO_FLAG_WRITE) ||
2891 vga_info.size < 0xbffff + 1) {
2892 error_report("vfio: Unexpected VGA info, flags 0x%lx, size 0x%lx",
2893 (unsigned long)vga_info.flags,
2894 (unsigned long)vga_info.size);
2895 goto error;
2896 }
2897
2898 vdev->vga.fd_offset = vga_info.offset;
2899 vdev->vga.fd = vdev->fd;
2900
2901 vdev->vga.region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE;
2902 vdev->vga.region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM;
2903 QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_MEM].quirks);
2904
2905 vdev->vga.region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE;
2906 vdev->vga.region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO;
2907 QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].quirks);
2908
2909 vdev->vga.region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE;
2910 vdev->vga.region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI;
2911 QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks);
2912
2913 vdev->has_vga = true;
2914 }
2915 irq_info.index = VFIO_PCI_ERR_IRQ_INDEX;
2916
2917 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info);
2918 if (ret) {
2919 /* This can fail for an old kernel or legacy PCI dev */
2920 DPRINTF("VFIO_DEVICE_GET_IRQ_INFO failure ret=%d\n", ret);
2921 ret = 0;
2922 } else if (irq_info.count == 1) {
2923 vdev->pci_aer = true;
2924 } else {
2925 error_report("vfio: Warning: "
2926 "Could not enable error recovery for the device\n");
2927 }
2928
2929 error:
2930 if (ret) {
2931 QLIST_REMOVE(vdev, next);
2932 vdev->group = NULL;
2933 close(vdev->fd);
2934 }
2935 return ret;
2936 }
2937
2938 static void vfio_put_device(VFIODevice *vdev)
2939 {
2940 QLIST_REMOVE(vdev, next);
2941 vdev->group = NULL;
2942 DPRINTF("vfio_put_device: close vdev->fd\n");
2943 close(vdev->fd);
2944 if (vdev->msix) {
2945 g_free(vdev->msix);
2946 vdev->msix = NULL;
2947 }
2948 }
2949
2950 static void vfio_err_notifier_handler(void *opaque)
2951 {
2952 VFIODevice *vdev = opaque;
2953
2954 if (!event_notifier_test_and_clear(&vdev->err_notifier)) {
2955 return;
2956 }
2957
2958 /*
2959 * TBD. Retrieve the error details and decide what action
2960 * needs to be taken. One of the actions could be to pass
2961 * the error to the guest and have the guest driver recover
2962 * from the error. This requires that PCIe capabilities be
2963 * exposed to the guest. For now, we just terminate the
2964 * guest to contain the error.
2965 */
2966
2967 error_report("%s (%04x:%02x:%02x.%x)"
2968 "Unrecoverable error detected...\n"
2969 "Please collect any data possible and then kill the guest",
2970 __func__, vdev->host.domain, vdev->host.bus,
2971 vdev->host.slot, vdev->host.function);
2972
2973 vm_stop(RUN_STATE_IO_ERROR);
2974 }
2975
2976 /*
2977 * Registers error notifier for devices supporting error recovery.
2978 * If we encounter a failure in this function, we report an error
2979 * and continue after disabling error recovery support for the
2980 * device.
2981 */
2982 static void vfio_register_err_notifier(VFIODevice *vdev)
2983 {
2984 int ret;
2985 int argsz;
2986 struct vfio_irq_set *irq_set;
2987 int32_t *pfd;
2988
2989 if (!vdev->pci_aer) {
2990 return;
2991 }
2992
2993 if (event_notifier_init(&vdev->err_notifier, 0)) {
2994 error_report("vfio: Warning: "
2995 "Unable to init event notifier for error detection\n");
2996 vdev->pci_aer = false;
2997 return;
2998 }
2999
3000 argsz = sizeof(*irq_set) + sizeof(*pfd);
3001
3002 irq_set = g_malloc0(argsz);
3003 irq_set->argsz = argsz;
3004 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
3005 VFIO_IRQ_SET_ACTION_TRIGGER;
3006 irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
3007 irq_set->start = 0;
3008 irq_set->count = 1;
3009 pfd = (int32_t *)&irq_set->data;
3010
3011 *pfd = event_notifier_get_fd(&vdev->err_notifier);
3012 qemu_set_fd_handler(*pfd, vfio_err_notifier_handler, NULL, vdev);
3013
3014 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
3015 if (ret) {
3016 error_report("vfio: Failed to set up error notification\n");
3017 qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
3018 event_notifier_cleanup(&vdev->err_notifier);
3019 vdev->pci_aer = false;
3020 }
3021 g_free(irq_set);
3022 }
3023
3024 static void vfio_unregister_err_notifier(VFIODevice *vdev)
3025 {
3026 int argsz;
3027 struct vfio_irq_set *irq_set;
3028 int32_t *pfd;
3029 int ret;
3030
3031 if (!vdev->pci_aer) {
3032 return;
3033 }
3034
3035 argsz = sizeof(*irq_set) + sizeof(*pfd);
3036
3037 irq_set = g_malloc0(argsz);
3038 irq_set->argsz = argsz;
3039 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
3040 VFIO_IRQ_SET_ACTION_TRIGGER;
3041 irq_set->index = VFIO_PCI_ERR_IRQ_INDEX;
3042 irq_set->start = 0;
3043 irq_set->count = 1;
3044 pfd = (int32_t *)&irq_set->data;
3045 *pfd = -1;
3046
3047 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
3048 if (ret) {
3049 error_report("vfio: Failed to de-assign error fd: %d\n", ret);
3050 }
3051 g_free(irq_set);
3052 qemu_set_fd_handler(event_notifier_get_fd(&vdev->err_notifier),
3053 NULL, NULL, vdev);
3054 event_notifier_cleanup(&vdev->err_notifier);
3055 }
3056
3057 static int vfio_initfn(PCIDevice *pdev)
3058 {
3059 VFIODevice *pvdev, *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
3060 VFIOGroup *group;
3061 char path[PATH_MAX], iommu_group_path[PATH_MAX], *group_name;
3062 ssize_t len;
3063 struct stat st;
3064 int groupid;
3065 int ret;
3066
3067 /* Check that the host device exists */
3068 snprintf(path, sizeof(path),
3069 "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/",
3070 vdev->host.domain, vdev->host.bus, vdev->host.slot,
3071 vdev->host.function);
3072 if (stat(path, &st) < 0) {
3073 error_report("vfio: error: no such host device: %s", path);
3074 return -errno;
3075 }
3076
3077 strncat(path, "iommu_group", sizeof(path) - strlen(path) - 1);
3078
3079 len = readlink(path, iommu_group_path, PATH_MAX);
3080 if (len <= 0) {
3081 error_report("vfio: error no iommu_group for device");
3082 return -errno;
3083 }
3084
3085 iommu_group_path[len] = 0;
3086 group_name = basename(iommu_group_path);
3087
3088 if (sscanf(group_name, "%d", &groupid) != 1) {
3089 error_report("vfio: error reading %s: %m", path);
3090 return -errno;
3091 }
3092
3093 DPRINTF("%s(%04x:%02x:%02x.%x) group %d\n", __func__, vdev->host.domain,
3094 vdev->host.bus, vdev->host.slot, vdev->host.function, groupid);
3095
3096 group = vfio_get_group(groupid);
3097 if (!group) {
3098 error_report("vfio: failed to get group %d", groupid);
3099 return -ENOENT;
3100 }
3101
3102 snprintf(path, sizeof(path), "%04x:%02x:%02x.%01x",
3103 vdev->host.domain, vdev->host.bus, vdev->host.slot,
3104 vdev->host.function);
3105
3106 QLIST_FOREACH(pvdev, &group->device_list, next) {
3107 if (pvdev->host.domain == vdev->host.domain &&
3108 pvdev->host.bus == vdev->host.bus &&
3109 pvdev->host.slot == vdev->host.slot &&
3110 pvdev->host.function == vdev->host.function) {
3111
3112 error_report("vfio: error: device %s is already attached", path);
3113 vfio_put_group(group);
3114 return -EBUSY;
3115 }
3116 }
3117
3118 ret = vfio_get_device(group, path, vdev);
3119 if (ret) {
3120 error_report("vfio: failed to get device %s", path);
3121 vfio_put_group(group);
3122 return ret;
3123 }
3124
3125 /* Get a copy of config space */
3126 ret = pread(vdev->fd, vdev->pdev.config,
3127 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
3128 vdev->config_offset);
3129 if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
3130 ret = ret < 0 ? -errno : -EFAULT;
3131 error_report("vfio: Failed to read device config space");
3132 goto out_put;
3133 }
3134
3135 /* vfio emulates a lot for us, but some bits need extra love */
3136 vdev->emulated_config_bits = g_malloc0(vdev->config_size);
3137
3138 /* QEMU can choose to expose the ROM or not */
3139 memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4);
3140
3141 /* QEMU can change multi-function devices to single function, or reverse */
3142 vdev->emulated_config_bits[PCI_HEADER_TYPE] =
3143 PCI_HEADER_TYPE_MULTI_FUNCTION;
3144
3145 /*
3146 * Clear host resource mapping info. If we choose not to register a
3147 * BAR, such as might be the case with the option ROM, we can get
3148 * confusing, unwritable, residual addresses from the host here.
3149 */
3150 memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
3151 memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
3152
3153 vfio_load_rom(vdev);
3154
3155 ret = vfio_early_setup_msix(vdev);
3156 if (ret) {
3157 goto out_put;
3158 }
3159
3160 vfio_map_bars(vdev);
3161
3162 ret = vfio_add_capabilities(vdev);
3163 if (ret) {
3164 goto out_teardown;
3165 }
3166
3167 /* QEMU emulates all of MSI & MSIX */
3168 if (pdev->cap_present & QEMU_PCI_CAP_MSIX) {
3169 memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff,
3170 MSIX_CAP_LENGTH);
3171 }
3172
3173 if (pdev->cap_present & QEMU_PCI_CAP_MSI) {
3174 memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff,
3175 vdev->msi_cap_size);
3176 }
3177
3178 if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
3179 vdev->intx.mmap_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL,
3180 vfio_intx_mmap_enable, vdev);
3181 pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_update_irq);
3182 ret = vfio_enable_intx(vdev);
3183 if (ret) {
3184 goto out_teardown;
3185 }
3186 }
3187
3188 add_boot_device_path(vdev->bootindex, &pdev->qdev, NULL);
3189 vfio_register_err_notifier(vdev);
3190
3191 return 0;
3192
3193 out_teardown:
3194 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3195 vfio_teardown_msi(vdev);
3196 vfio_unmap_bars(vdev);
3197 out_put:
3198 g_free(vdev->emulated_config_bits);
3199 vfio_put_device(vdev);
3200 vfio_put_group(group);
3201 return ret;
3202 }
3203
3204 static void vfio_exitfn(PCIDevice *pdev)
3205 {
3206 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
3207 VFIOGroup *group = vdev->group;
3208
3209 vfio_unregister_err_notifier(vdev);
3210 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3211 vfio_disable_interrupts(vdev);
3212 if (vdev->intx.mmap_timer) {
3213 timer_free(vdev->intx.mmap_timer);
3214 }
3215 vfio_teardown_msi(vdev);
3216 vfio_unmap_bars(vdev);
3217 g_free(vdev->emulated_config_bits);
3218 vfio_put_device(vdev);
3219 vfio_put_group(group);
3220 }
3221
3222 static void vfio_pci_reset(DeviceState *dev)
3223 {
3224 PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
3225 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
3226 uint16_t cmd;
3227
3228 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
3229 vdev->host.bus, vdev->host.slot, vdev->host.function);
3230
3231 vfio_disable_interrupts(vdev);
3232
3233 /* Make sure the device is in D0 */
3234 if (vdev->pm_cap) {
3235 uint16_t pmcsr;
3236 uint8_t state;
3237
3238 pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
3239 state = pmcsr & PCI_PM_CTRL_STATE_MASK;
3240 if (state) {
3241 pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
3242 vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2);
3243 /* vfio handles the necessary delay here */
3244 pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
3245 state = pmcsr & PCI_PM_CTRL_STATE_MASK;
3246 if (state) {
3247 error_report("vfio: Unable to power on device, stuck in D%d\n",
3248 state);
3249 }
3250 }
3251 }
3252
3253 /*
3254 * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
3255 * Also put INTx Disable in known state.
3256 */
3257 cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
3258 cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
3259 PCI_COMMAND_INTX_DISABLE);
3260 vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
3261
3262 if (vdev->reset_works) {
3263 if (ioctl(vdev->fd, VFIO_DEVICE_RESET)) {
3264 error_report("vfio: Error unable to reset physical device "
3265 "(%04x:%02x:%02x.%x): %m", vdev->host.domain,
3266 vdev->host.bus, vdev->host.slot, vdev->host.function);
3267 }
3268 }
3269
3270 vfio_enable_intx(vdev);
3271 }
3272
3273 static Property vfio_pci_dev_properties[] = {
3274 DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIODevice, host),
3275 DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIODevice,
3276 intx.mmap_timeout, 1100),
3277 DEFINE_PROP_BIT("x-vga", VFIODevice, features,
3278 VFIO_FEATURE_ENABLE_VGA_BIT, false),
3279 DEFINE_PROP_INT32("bootindex", VFIODevice, bootindex, -1),
3280 /*
3281 * TODO - support passed fds... is this necessary?
3282 * DEFINE_PROP_STRING("vfiofd", VFIODevice, vfiofd_name),
3283 * DEFINE_PROP_STRING("vfiogroupfd, VFIODevice, vfiogroupfd_name),
3284 */
3285 DEFINE_PROP_END_OF_LIST(),
3286 };
3287
3288 static const VMStateDescription vfio_pci_vmstate = {
3289 .name = "vfio-pci",
3290 .unmigratable = 1,
3291 };
3292
3293 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
3294 {
3295 DeviceClass *dc = DEVICE_CLASS(klass);
3296 PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
3297
3298 dc->reset = vfio_pci_reset;
3299 dc->props = vfio_pci_dev_properties;
3300 dc->vmsd = &vfio_pci_vmstate;
3301 dc->desc = "VFIO-based PCI device assignment";
3302 set_bit(DEVICE_CATEGORY_MISC, dc->categories);
3303 pdc->init = vfio_initfn;
3304 pdc->exit = vfio_exitfn;
3305 pdc->config_read = vfio_pci_read_config;
3306 pdc->config_write = vfio_pci_write_config;
3307 pdc->is_express = 1; /* We might be */
3308 }
3309
3310 static const TypeInfo vfio_pci_dev_info = {
3311 .name = "vfio-pci",
3312 .parent = TYPE_PCI_DEVICE,
3313 .instance_size = sizeof(VFIODevice),
3314 .class_init = vfio_pci_dev_class_init,
3315 };
3316
3317 static void register_vfio_pci_dev_type(void)
3318 {
3319 type_register_static(&vfio_pci_dev_info);
3320 }
3321
3322 type_init(register_vfio_pci_dev_type)
Qemu copy for testing