]> git.proxmox.com Git - qemu.git/blame - hw/misc/vfio.c
ne2000: pass device to ne2000_setup_io, use it as owner
[qemu.git] / hw / misc / vfio.c
CommitLineData
65501a74
AW
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>
6dcfdbad 22#include <linux/vfio.h>
65501a74
AW
23#include <sys/ioctl.h>
24#include <sys/mman.h>
25#include <sys/stat.h>
26#include <sys/types.h>
6dcfdbad 27#include <unistd.h>
65501a74
AW
28
29#include "config.h"
022c62cb 30#include "exec/address-spaces.h"
022c62cb 31#include "exec/memory.h"
83c9f4ca
PB
32#include "hw/pci/msi.h"
33#include "hw/pci/msix.h"
34#include "hw/pci/pci.h"
5c97e5eb 35#include "qemu-common.h"
1de7afc9 36#include "qemu/error-report.h"
6dcfdbad 37#include "qemu/event_notifier.h"
1de7afc9
PB
38#include "qemu/queue.h"
39#include "qemu/range.h"
6dcfdbad
AW
40#include "sysemu/kvm.h"
41#include "sysemu/sysemu.h"
65501a74
AW
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
82ca8912
AW
52/* Extra debugging, trap acceleration paths for more logging */
53#define VFIO_ALLOW_MMAP 1
54#define VFIO_ALLOW_KVM_INTX 1
55
7076eabc
AW
56struct VFIODevice;
57
58typedef struct VFIOQuirk {
59 MemoryRegion mem;
60 struct VFIODevice *vdev;
61 QLIST_ENTRY(VFIOQuirk) next;
62 uint32_t data;
63 uint32_t data2;
64} VFIOQuirk;
65
5c97e5eb
AW
66typedef struct VFIOBAR {
67 off_t fd_offset; /* offset of BAR within device fd */
68 int fd; /* device fd, allows us to pass VFIOBAR as opaque data */
69 MemoryRegion mem; /* slow, read/write access */
70 MemoryRegion mmap_mem; /* direct mapped access */
71 void *mmap;
72 size_t size;
73 uint32_t flags; /* VFIO region flags (rd/wr/mmap) */
74 uint8_t nr; /* cache the BAR number for debug */
7076eabc 75 QLIST_HEAD(, VFIOQuirk) quirks;
5c97e5eb
AW
76} VFIOBAR;
77
f15689c7
AW
78typedef struct VFIOVGARegion {
79 MemoryRegion mem;
80 off_t offset;
81 int nr;
7076eabc 82 QLIST_HEAD(, VFIOQuirk) quirks;
f15689c7
AW
83} VFIOVGARegion;
84
85typedef struct VFIOVGA {
86 off_t fd_offset;
87 int fd;
88 VFIOVGARegion region[QEMU_PCI_VGA_NUM_REGIONS];
89} VFIOVGA;
90
5c97e5eb
AW
91typedef struct VFIOINTx {
92 bool pending; /* interrupt pending */
93 bool kvm_accel; /* set when QEMU bypass through KVM enabled */
94 uint8_t pin; /* which pin to pull for qemu_set_irq */
95 EventNotifier interrupt; /* eventfd triggered on interrupt */
96 EventNotifier unmask; /* eventfd for unmask on QEMU bypass */
97 PCIINTxRoute route; /* routing info for QEMU bypass */
98 uint32_t mmap_timeout; /* delay to re-enable mmaps after interrupt */
99 QEMUTimer *mmap_timer; /* enable mmaps after periods w/o interrupts */
100} VFIOINTx;
101
5c97e5eb
AW
102typedef struct VFIOMSIVector {
103 EventNotifier interrupt; /* eventfd triggered on interrupt */
104 struct VFIODevice *vdev; /* back pointer to device */
105 int virq; /* KVM irqchip route for QEMU bypass */
106 bool use;
107} VFIOMSIVector;
108
109enum {
110 VFIO_INT_NONE = 0,
111 VFIO_INT_INTx = 1,
112 VFIO_INT_MSI = 2,
113 VFIO_INT_MSIX = 3,
114};
115
116struct VFIOGroup;
117
118typedef struct VFIOContainer {
119 int fd; /* /dev/vfio/vfio, empowered by the attached groups */
120 struct {
121 /* enable abstraction to support various iommu backends */
122 union {
123 MemoryListener listener; /* Used by type1 iommu */
124 };
125 void (*release)(struct VFIOContainer *);
126 } iommu_data;
127 QLIST_HEAD(, VFIOGroup) group_list;
128 QLIST_ENTRY(VFIOContainer) next;
129} VFIOContainer;
130
131/* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
132typedef struct VFIOMSIXInfo {
133 uint8_t table_bar;
134 uint8_t pba_bar;
135 uint16_t entries;
136 uint32_t table_offset;
137 uint32_t pba_offset;
138 MemoryRegion mmap_mem;
139 void *mmap;
140} VFIOMSIXInfo;
141
142typedef struct VFIODevice {
143 PCIDevice pdev;
144 int fd;
145 VFIOINTx intx;
146 unsigned int config_size;
4b5d5e87 147 uint8_t *emulated_config_bits; /* QEMU emulated bits, little-endian */
5c97e5eb
AW
148 off_t config_offset; /* Offset of config space region within device fd */
149 unsigned int rom_size;
150 off_t rom_offset; /* Offset of ROM region within device fd */
151 int msi_cap_size;
152 VFIOMSIVector *msi_vectors;
153 VFIOMSIXInfo *msix;
154 int nr_vectors; /* Number of MSI/MSIX vectors currently in use */
155 int interrupt; /* Current interrupt type */
156 VFIOBAR bars[PCI_NUM_REGIONS - 1]; /* No ROM */
f15689c7 157 VFIOVGA vga; /* 0xa0000, 0x3b0, 0x3c0 */
5c97e5eb
AW
158 PCIHostDeviceAddress host;
159 QLIST_ENTRY(VFIODevice) next;
160 struct VFIOGroup *group;
f15689c7
AW
161 uint32_t features;
162#define VFIO_FEATURE_ENABLE_VGA_BIT 0
163#define VFIO_FEATURE_ENABLE_VGA (1 << VFIO_FEATURE_ENABLE_VGA_BIT)
c29029dd 164 int32_t bootindex;
ba661818 165 uint8_t pm_cap;
5c97e5eb 166 bool reset_works;
f15689c7 167 bool has_vga;
5c97e5eb
AW
168} VFIODevice;
169
170typedef struct VFIOGroup {
171 int fd;
172 int groupid;
173 VFIOContainer *container;
174 QLIST_HEAD(, VFIODevice) device_list;
175 QLIST_ENTRY(VFIOGroup) next;
176 QLIST_ENTRY(VFIOGroup) container_next;
177} VFIOGroup;
178
65501a74
AW
179#define MSIX_CAP_LENGTH 12
180
181static QLIST_HEAD(, VFIOContainer)
182 container_list = QLIST_HEAD_INITIALIZER(container_list);
183
184static QLIST_HEAD(, VFIOGroup)
185 group_list = QLIST_HEAD_INITIALIZER(group_list);
186
187static void vfio_disable_interrupts(VFIODevice *vdev);
188static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len);
7076eabc
AW
189static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
190 uint32_t val, int len);
65501a74
AW
191static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled);
192
193/*
194 * Common VFIO interrupt disable
195 */
196static void vfio_disable_irqindex(VFIODevice *vdev, int index)
197{
198 struct vfio_irq_set irq_set = {
199 .argsz = sizeof(irq_set),
200 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
201 .index = index,
202 .start = 0,
203 .count = 0,
204 };
205
206 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
65501a74
AW
207}
208
209/*
210 * INTx
211 */
212static void vfio_unmask_intx(VFIODevice *vdev)
213{
214 struct vfio_irq_set irq_set = {
215 .argsz = sizeof(irq_set),
216 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
217 .index = VFIO_PCI_INTX_IRQ_INDEX,
218 .start = 0,
219 .count = 1,
220 };
221
222 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
223}
224
e1d1e586
AW
225#ifdef CONFIG_KVM /* Unused outside of CONFIG_KVM code */
226static void vfio_mask_intx(VFIODevice *vdev)
227{
228 struct vfio_irq_set irq_set = {
229 .argsz = sizeof(irq_set),
230 .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
231 .index = VFIO_PCI_INTX_IRQ_INDEX,
232 .start = 0,
233 .count = 1,
234 };
235
236 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
237}
238#endif
239
ea486926
AW
240/*
241 * Disabling BAR mmaping can be slow, but toggling it around INTx can
242 * also be a huge overhead. We try to get the best of both worlds by
243 * waiting until an interrupt to disable mmaps (subsequent transitions
244 * to the same state are effectively no overhead). If the interrupt has
245 * been serviced and the time gap is long enough, we re-enable mmaps for
246 * performance. This works well for things like graphics cards, which
247 * may not use their interrupt at all and are penalized to an unusable
248 * level by read/write BAR traps. Other devices, like NICs, have more
249 * regular interrupts and see much better latency by staying in non-mmap
250 * mode. We therefore set the default mmap_timeout such that a ping
251 * is just enough to keep the mmap disabled. Users can experiment with
252 * other options with the x-intx-mmap-timeout-ms parameter (a value of
253 * zero disables the timer).
254 */
255static void vfio_intx_mmap_enable(void *opaque)
256{
257 VFIODevice *vdev = opaque;
258
259 if (vdev->intx.pending) {
260 qemu_mod_timer(vdev->intx.mmap_timer,
261 qemu_get_clock_ms(vm_clock) + vdev->intx.mmap_timeout);
262 return;
263 }
264
265 vfio_mmap_set_enabled(vdev, true);
266}
267
65501a74
AW
268static void vfio_intx_interrupt(void *opaque)
269{
270 VFIODevice *vdev = opaque;
271
272 if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
273 return;
274 }
275
276 DPRINTF("%s(%04x:%02x:%02x.%x) Pin %c\n", __func__, vdev->host.domain,
277 vdev->host.bus, vdev->host.slot, vdev->host.function,
278 'A' + vdev->intx.pin);
279
280 vdev->intx.pending = true;
281 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 1);
ea486926
AW
282 vfio_mmap_set_enabled(vdev, false);
283 if (vdev->intx.mmap_timeout) {
284 qemu_mod_timer(vdev->intx.mmap_timer,
285 qemu_get_clock_ms(vm_clock) + vdev->intx.mmap_timeout);
286 }
65501a74
AW
287}
288
289static void vfio_eoi(VFIODevice *vdev)
290{
291 if (!vdev->intx.pending) {
292 return;
293 }
294
295 DPRINTF("%s(%04x:%02x:%02x.%x) EOI\n", __func__, vdev->host.domain,
296 vdev->host.bus, vdev->host.slot, vdev->host.function);
297
298 vdev->intx.pending = false;
299 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
300 vfio_unmask_intx(vdev);
301}
302
e1d1e586
AW
303static void vfio_enable_intx_kvm(VFIODevice *vdev)
304{
305#ifdef CONFIG_KVM
306 struct kvm_irqfd irqfd = {
307 .fd = event_notifier_get_fd(&vdev->intx.interrupt),
308 .gsi = vdev->intx.route.irq,
309 .flags = KVM_IRQFD_FLAG_RESAMPLE,
310 };
311 struct vfio_irq_set *irq_set;
312 int ret, argsz;
313 int32_t *pfd;
314
82ca8912 315 if (!VFIO_ALLOW_KVM_INTX || !kvm_irqfds_enabled() ||
e1d1e586
AW
316 vdev->intx.route.mode != PCI_INTX_ENABLED ||
317 !kvm_check_extension(kvm_state, KVM_CAP_IRQFD_RESAMPLE)) {
318 return;
319 }
320
321 /* Get to a known interrupt state */
322 qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev);
323 vfio_mask_intx(vdev);
324 vdev->intx.pending = false;
325 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
326
327 /* Get an eventfd for resample/unmask */
328 if (event_notifier_init(&vdev->intx.unmask, 0)) {
312fd5f2 329 error_report("vfio: Error: event_notifier_init failed eoi");
e1d1e586
AW
330 goto fail;
331 }
332
333 /* KVM triggers it, VFIO listens for it */
334 irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask);
335
336 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
312fd5f2 337 error_report("vfio: Error: Failed to setup resample irqfd: %m");
e1d1e586
AW
338 goto fail_irqfd;
339 }
340
341 argsz = sizeof(*irq_set) + sizeof(*pfd);
342
343 irq_set = g_malloc0(argsz);
344 irq_set->argsz = argsz;
345 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK;
346 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
347 irq_set->start = 0;
348 irq_set->count = 1;
349 pfd = (int32_t *)&irq_set->data;
350
351 *pfd = irqfd.resamplefd;
352
353 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
354 g_free(irq_set);
355 if (ret) {
312fd5f2 356 error_report("vfio: Error: Failed to setup INTx unmask fd: %m");
e1d1e586
AW
357 goto fail_vfio;
358 }
359
360 /* Let'em rip */
361 vfio_unmask_intx(vdev);
362
363 vdev->intx.kvm_accel = true;
364
365 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel enabled\n",
366 __func__, vdev->host.domain, vdev->host.bus,
367 vdev->host.slot, vdev->host.function);
368
369 return;
370
371fail_vfio:
372 irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN;
373 kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
374fail_irqfd:
375 event_notifier_cleanup(&vdev->intx.unmask);
376fail:
377 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
378 vfio_unmask_intx(vdev);
379#endif
380}
381
382static void vfio_disable_intx_kvm(VFIODevice *vdev)
383{
384#ifdef CONFIG_KVM
385 struct kvm_irqfd irqfd = {
386 .fd = event_notifier_get_fd(&vdev->intx.interrupt),
387 .gsi = vdev->intx.route.irq,
388 .flags = KVM_IRQFD_FLAG_DEASSIGN,
389 };
390
391 if (!vdev->intx.kvm_accel) {
392 return;
393 }
394
395 /*
396 * Get to a known state, hardware masked, QEMU ready to accept new
397 * interrupts, QEMU IRQ de-asserted.
398 */
399 vfio_mask_intx(vdev);
400 vdev->intx.pending = false;
401 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
402
403 /* Tell KVM to stop listening for an INTx irqfd */
404 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) {
312fd5f2 405 error_report("vfio: Error: Failed to disable INTx irqfd: %m");
e1d1e586
AW
406 }
407
408 /* We only need to close the eventfd for VFIO to cleanup the kernel side */
409 event_notifier_cleanup(&vdev->intx.unmask);
410
411 /* QEMU starts listening for interrupt events. */
412 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev);
413
414 vdev->intx.kvm_accel = false;
415
416 /* If we've missed an event, let it re-fire through QEMU */
417 vfio_unmask_intx(vdev);
418
419 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel disabled\n",
420 __func__, vdev->host.domain, vdev->host.bus,
421 vdev->host.slot, vdev->host.function);
422#endif
423}
424
425static void vfio_update_irq(PCIDevice *pdev)
426{
427 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
428 PCIINTxRoute route;
429
430 if (vdev->interrupt != VFIO_INT_INTx) {
431 return;
432 }
433
434 route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin);
435
436 if (!pci_intx_route_changed(&vdev->intx.route, &route)) {
437 return; /* Nothing changed */
438 }
439
440 DPRINTF("%s(%04x:%02x:%02x.%x) IRQ moved %d -> %d\n", __func__,
441 vdev->host.domain, vdev->host.bus, vdev->host.slot,
442 vdev->host.function, vdev->intx.route.irq, route.irq);
443
444 vfio_disable_intx_kvm(vdev);
445
446 vdev->intx.route = route;
447
448 if (route.mode != PCI_INTX_ENABLED) {
449 return;
450 }
451
452 vfio_enable_intx_kvm(vdev);
453
454 /* Re-enable the interrupt in cased we missed an EOI */
455 vfio_eoi(vdev);
456}
457
65501a74
AW
458static int vfio_enable_intx(VFIODevice *vdev)
459{
65501a74 460 uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
1a403133
AW
461 int ret, argsz;
462 struct vfio_irq_set *irq_set;
463 int32_t *pfd;
65501a74 464
ea486926 465 if (!pin) {
65501a74
AW
466 return 0;
467 }
468
469 vfio_disable_interrupts(vdev);
470
471 vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
e1d1e586
AW
472
473#ifdef CONFIG_KVM
474 /*
475 * Only conditional to avoid generating error messages on platforms
476 * where we won't actually use the result anyway.
477 */
d281084d
AW
478 if (kvm_irqfds_enabled() &&
479 kvm_check_extension(kvm_state, KVM_CAP_IRQFD_RESAMPLE)) {
e1d1e586
AW
480 vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev,
481 vdev->intx.pin);
482 }
483#endif
484
65501a74
AW
485 ret = event_notifier_init(&vdev->intx.interrupt, 0);
486 if (ret) {
312fd5f2 487 error_report("vfio: Error: event_notifier_init failed");
65501a74
AW
488 return ret;
489 }
490
1a403133
AW
491 argsz = sizeof(*irq_set) + sizeof(*pfd);
492
493 irq_set = g_malloc0(argsz);
494 irq_set->argsz = argsz;
495 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
496 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
497 irq_set->start = 0;
498 irq_set->count = 1;
499 pfd = (int32_t *)&irq_set->data;
500
501 *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
502 qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
65501a74 503
1a403133
AW
504 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
505 g_free(irq_set);
506 if (ret) {
312fd5f2 507 error_report("vfio: Error: Failed to setup INTx fd: %m");
1a403133 508 qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
ce59af2d 509 event_notifier_cleanup(&vdev->intx.interrupt);
65501a74
AW
510 return -errno;
511 }
512
e1d1e586
AW
513 vfio_enable_intx_kvm(vdev);
514
65501a74
AW
515 vdev->interrupt = VFIO_INT_INTx;
516
517 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
518 vdev->host.bus, vdev->host.slot, vdev->host.function);
519
520 return 0;
521}
522
523static void vfio_disable_intx(VFIODevice *vdev)
524{
525 int fd;
526
ea486926 527 qemu_del_timer(vdev->intx.mmap_timer);
e1d1e586 528 vfio_disable_intx_kvm(vdev);
65501a74
AW
529 vfio_disable_irqindex(vdev, VFIO_PCI_INTX_IRQ_INDEX);
530 vdev->intx.pending = false;
531 qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
532 vfio_mmap_set_enabled(vdev, true);
533
534 fd = event_notifier_get_fd(&vdev->intx.interrupt);
535 qemu_set_fd_handler(fd, NULL, NULL, vdev);
536 event_notifier_cleanup(&vdev->intx.interrupt);
537
538 vdev->interrupt = VFIO_INT_NONE;
539
540 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
541 vdev->host.bus, vdev->host.slot, vdev->host.function);
542}
543
544/*
545 * MSI/X
546 */
547static void vfio_msi_interrupt(void *opaque)
548{
549 VFIOMSIVector *vector = opaque;
550 VFIODevice *vdev = vector->vdev;
551 int nr = vector - vdev->msi_vectors;
552
553 if (!event_notifier_test_and_clear(&vector->interrupt)) {
554 return;
555 }
556
557 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d\n", __func__,
558 vdev->host.domain, vdev->host.bus, vdev->host.slot,
559 vdev->host.function, nr);
560
561 if (vdev->interrupt == VFIO_INT_MSIX) {
562 msix_notify(&vdev->pdev, nr);
563 } else if (vdev->interrupt == VFIO_INT_MSI) {
564 msi_notify(&vdev->pdev, nr);
565 } else {
312fd5f2 566 error_report("vfio: MSI interrupt receieved, but not enabled?");
65501a74
AW
567 }
568}
569
570static int vfio_enable_vectors(VFIODevice *vdev, bool msix)
571{
572 struct vfio_irq_set *irq_set;
573 int ret = 0, i, argsz;
574 int32_t *fds;
575
576 argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
577
578 irq_set = g_malloc0(argsz);
579 irq_set->argsz = argsz;
580 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
581 irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
582 irq_set->start = 0;
583 irq_set->count = vdev->nr_vectors;
584 fds = (int32_t *)&irq_set->data;
585
586 for (i = 0; i < vdev->nr_vectors; i++) {
587 if (!vdev->msi_vectors[i].use) {
588 fds[i] = -1;
589 continue;
590 }
591
592 fds[i] = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
593 }
594
595 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
596
597 g_free(irq_set);
598
65501a74
AW
599 return ret;
600}
601
b0223e29
AW
602static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr,
603 MSIMessage *msg, IOHandler *handler)
65501a74
AW
604{
605 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
606 VFIOMSIVector *vector;
607 int ret;
608
609 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d used\n", __func__,
610 vdev->host.domain, vdev->host.bus, vdev->host.slot,
611 vdev->host.function, nr);
612
65501a74
AW
613 vector = &vdev->msi_vectors[nr];
614 vector->vdev = vdev;
615 vector->use = true;
616
617 msix_vector_use(pdev, nr);
618
619 if (event_notifier_init(&vector->interrupt, 0)) {
312fd5f2 620 error_report("vfio: Error: event_notifier_init failed");
65501a74
AW
621 }
622
623 /*
624 * Attempt to enable route through KVM irqchip,
625 * default to userspace handling if unavailable.
626 */
b0223e29 627 vector->virq = msg ? kvm_irqchip_add_msi_route(kvm_state, *msg) : -1;
65501a74
AW
628 if (vector->virq < 0 ||
629 kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
630 vector->virq) < 0) {
631 if (vector->virq >= 0) {
632 kvm_irqchip_release_virq(kvm_state, vector->virq);
633 vector->virq = -1;
634 }
635 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
b0223e29 636 handler, NULL, vector);
65501a74
AW
637 }
638
639 /*
640 * We don't want to have the host allocate all possible MSI vectors
641 * for a device if they're not in use, so we shutdown and incrementally
642 * increase them as needed.
643 */
644 if (vdev->nr_vectors < nr + 1) {
65501a74
AW
645 vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
646 vdev->nr_vectors = nr + 1;
647 ret = vfio_enable_vectors(vdev, true);
648 if (ret) {
312fd5f2 649 error_report("vfio: failed to enable vectors, %d", ret);
65501a74 650 }
65501a74 651 } else {
1a403133
AW
652 int argsz;
653 struct vfio_irq_set *irq_set;
654 int32_t *pfd;
655
656 argsz = sizeof(*irq_set) + sizeof(*pfd);
657
658 irq_set = g_malloc0(argsz);
659 irq_set->argsz = argsz;
660 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
661 VFIO_IRQ_SET_ACTION_TRIGGER;
662 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
663 irq_set->start = nr;
664 irq_set->count = 1;
665 pfd = (int32_t *)&irq_set->data;
666
667 *pfd = event_notifier_get_fd(&vector->interrupt);
668
669 ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
670 g_free(irq_set);
65501a74 671 if (ret) {
312fd5f2 672 error_report("vfio: failed to modify vector, %d", ret);
65501a74 673 }
65501a74
AW
674 }
675
676 return 0;
677}
678
b0223e29
AW
679static int vfio_msix_vector_use(PCIDevice *pdev,
680 unsigned int nr, MSIMessage msg)
681{
682 return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt);
683}
684
65501a74
AW
685static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
686{
687 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
688 VFIOMSIVector *vector = &vdev->msi_vectors[nr];
1a403133
AW
689 int argsz;
690 struct vfio_irq_set *irq_set;
691 int32_t *pfd;
65501a74
AW
692
693 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d released\n", __func__,
694 vdev->host.domain, vdev->host.bus, vdev->host.slot,
695 vdev->host.function, nr);
696
697 /*
698 * XXX What's the right thing to do here? This turns off the interrupt
699 * completely, but do we really just want to switch the interrupt to
700 * bouncing through userspace and let msix.c drop it? Not sure.
701 */
702 msix_vector_unuse(pdev, nr);
1a403133
AW
703
704 argsz = sizeof(*irq_set) + sizeof(*pfd);
705
706 irq_set = g_malloc0(argsz);
707 irq_set->argsz = argsz;
708 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
709 VFIO_IRQ_SET_ACTION_TRIGGER;
710 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
711 irq_set->start = nr;
712 irq_set->count = 1;
713 pfd = (int32_t *)&irq_set->data;
714
715 *pfd = -1;
716
717 ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
718
719 g_free(irq_set);
65501a74
AW
720
721 if (vector->virq < 0) {
722 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
723 NULL, NULL, NULL);
724 } else {
725 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
726 vector->virq);
727 kvm_irqchip_release_virq(kvm_state, vector->virq);
728 vector->virq = -1;
729 }
730
731 event_notifier_cleanup(&vector->interrupt);
732 vector->use = false;
733}
734
fd704adc
AW
735static void vfio_enable_msix(VFIODevice *vdev)
736{
737 vfio_disable_interrupts(vdev);
738
739 vdev->msi_vectors = g_malloc0(vdev->msix->entries * sizeof(VFIOMSIVector));
740
741 vdev->interrupt = VFIO_INT_MSIX;
742
b0223e29
AW
743 /*
744 * Some communication channels between VF & PF or PF & fw rely on the
745 * physical state of the device and expect that enabling MSI-X from the
746 * guest enables the same on the host. When our guest is Linux, the
747 * guest driver call to pci_enable_msix() sets the enabling bit in the
748 * MSI-X capability, but leaves the vector table masked. We therefore
749 * can't rely on a vector_use callback (from request_irq() in the guest)
750 * to switch the physical device into MSI-X mode because that may come a
751 * long time after pci_enable_msix(). This code enables vector 0 with
752 * triggering to userspace, then immediately release the vector, leaving
753 * the physical device with no vectors enabled, but MSI-X enabled, just
754 * like the guest view.
755 */
756 vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL);
757 vfio_msix_vector_release(&vdev->pdev, 0);
758
fd704adc 759 if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
bbef882c 760 vfio_msix_vector_release, NULL)) {
312fd5f2 761 error_report("vfio: msix_set_vector_notifiers failed");
fd704adc
AW
762 }
763
764 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
765 vdev->host.bus, vdev->host.slot, vdev->host.function);
766}
767
65501a74
AW
768static void vfio_enable_msi(VFIODevice *vdev)
769{
770 int ret, i;
771
772 vfio_disable_interrupts(vdev);
773
774 vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
775retry:
776 vdev->msi_vectors = g_malloc0(vdev->nr_vectors * sizeof(VFIOMSIVector));
777
778 for (i = 0; i < vdev->nr_vectors; i++) {
779 MSIMessage msg;
780 VFIOMSIVector *vector = &vdev->msi_vectors[i];
781
782 vector->vdev = vdev;
783 vector->use = true;
784
785 if (event_notifier_init(&vector->interrupt, 0)) {
312fd5f2 786 error_report("vfio: Error: event_notifier_init failed");
65501a74
AW
787 }
788
a771c517 789 msg = msi_get_message(&vdev->pdev, i);
65501a74
AW
790
791 /*
792 * Attempt to enable route through KVM irqchip,
793 * default to userspace handling if unavailable.
794 */
795 vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg);
796 if (vector->virq < 0 ||
797 kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
798 vector->virq) < 0) {
799 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
800 vfio_msi_interrupt, NULL, vector);
801 }
802 }
803
804 ret = vfio_enable_vectors(vdev, false);
805 if (ret) {
806 if (ret < 0) {
312fd5f2 807 error_report("vfio: Error: Failed to setup MSI fds: %m");
65501a74
AW
808 } else if (ret != vdev->nr_vectors) {
809 error_report("vfio: Error: Failed to enable %d "
312fd5f2 810 "MSI vectors, retry with %d", vdev->nr_vectors, ret);
65501a74
AW
811 }
812
813 for (i = 0; i < vdev->nr_vectors; i++) {
814 VFIOMSIVector *vector = &vdev->msi_vectors[i];
815 if (vector->virq >= 0) {
816 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
817 vector->virq);
818 kvm_irqchip_release_virq(kvm_state, vector->virq);
819 vector->virq = -1;
820 } else {
821 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
822 NULL, NULL, NULL);
823 }
824 event_notifier_cleanup(&vector->interrupt);
825 }
826
827 g_free(vdev->msi_vectors);
828
829 if (ret > 0 && ret != vdev->nr_vectors) {
830 vdev->nr_vectors = ret;
831 goto retry;
832 }
833 vdev->nr_vectors = 0;
834
835 return;
836 }
837
fd704adc
AW
838 vdev->interrupt = VFIO_INT_MSI;
839
65501a74
AW
840 DPRINTF("%s(%04x:%02x:%02x.%x) Enabled %d MSI vectors\n", __func__,
841 vdev->host.domain, vdev->host.bus, vdev->host.slot,
842 vdev->host.function, vdev->nr_vectors);
843}
844
fd704adc
AW
845static void vfio_disable_msi_common(VFIODevice *vdev)
846{
847 g_free(vdev->msi_vectors);
848 vdev->msi_vectors = NULL;
849 vdev->nr_vectors = 0;
850 vdev->interrupt = VFIO_INT_NONE;
851
852 vfio_enable_intx(vdev);
853}
854
855static void vfio_disable_msix(VFIODevice *vdev)
856{
857 msix_unset_vector_notifiers(&vdev->pdev);
858
859 if (vdev->nr_vectors) {
860 vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
861 }
862
863 vfio_disable_msi_common(vdev);
864
a011b10e
AW
865 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
866 vdev->host.bus, vdev->host.slot, vdev->host.function);
fd704adc
AW
867}
868
869static void vfio_disable_msi(VFIODevice *vdev)
65501a74
AW
870{
871 int i;
872
fd704adc 873 vfio_disable_irqindex(vdev, VFIO_PCI_MSI_IRQ_INDEX);
65501a74
AW
874
875 for (i = 0; i < vdev->nr_vectors; i++) {
876 VFIOMSIVector *vector = &vdev->msi_vectors[i];
877
878 if (!vector->use) {
879 continue;
880 }
881
882 if (vector->virq >= 0) {
883 kvm_irqchip_remove_irqfd_notifier(kvm_state,
884 &vector->interrupt, vector->virq);
885 kvm_irqchip_release_virq(kvm_state, vector->virq);
886 vector->virq = -1;
887 } else {
888 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
889 NULL, NULL, NULL);
890 }
891
65501a74
AW
892 event_notifier_cleanup(&vector->interrupt);
893 }
894
fd704adc 895 vfio_disable_msi_common(vdev);
65501a74 896
fd704adc
AW
897 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
898 vdev->host.bus, vdev->host.slot, vdev->host.function);
65501a74
AW
899}
900
901/*
902 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
903 */
a8170e5e 904static void vfio_bar_write(void *opaque, hwaddr addr,
65501a74
AW
905 uint64_t data, unsigned size)
906{
907 VFIOBAR *bar = opaque;
908 union {
909 uint8_t byte;
910 uint16_t word;
911 uint32_t dword;
912 uint64_t qword;
913 } buf;
914
915 switch (size) {
916 case 1:
917 buf.byte = data;
918 break;
919 case 2:
920 buf.word = cpu_to_le16(data);
921 break;
922 case 4:
923 buf.dword = cpu_to_le32(data);
924 break;
925 default:
926 hw_error("vfio: unsupported write size, %d bytes\n", size);
927 break;
928 }
929
930 if (pwrite(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
312fd5f2 931 error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
65501a74
AW
932 __func__, addr, data, size);
933 }
934
82ca8912
AW
935#ifdef DEBUG_VFIO
936 {
937 VFIODevice *vdev = container_of(bar, VFIODevice, bars[bar->nr]);
938
939 DPRINTF("%s(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx", 0x%"PRIx64
940 ", %d)\n", __func__, vdev->host.domain, vdev->host.bus,
941 vdev->host.slot, vdev->host.function, bar->nr, addr,
942 data, size);
943 }
944#endif
65501a74
AW
945
946 /*
947 * A read or write to a BAR always signals an INTx EOI. This will
948 * do nothing if not pending (including not in INTx mode). We assume
949 * that a BAR access is in response to an interrupt and that BAR
950 * accesses will service the interrupt. Unfortunately, we don't know
951 * which access will service the interrupt, so we're potentially
952 * getting quite a few host interrupts per guest interrupt.
953 */
3a4f2816 954 vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
65501a74
AW
955}
956
957static uint64_t vfio_bar_read(void *opaque,
a8170e5e 958 hwaddr addr, unsigned size)
65501a74
AW
959{
960 VFIOBAR *bar = opaque;
961 union {
962 uint8_t byte;
963 uint16_t word;
964 uint32_t dword;
965 uint64_t qword;
966 } buf;
967 uint64_t data = 0;
968
969 if (pread(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
312fd5f2 970 error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
65501a74
AW
971 __func__, addr, size);
972 return (uint64_t)-1;
973 }
974
975 switch (size) {
976 case 1:
977 data = buf.byte;
978 break;
979 case 2:
980 data = le16_to_cpu(buf.word);
981 break;
982 case 4:
983 data = le32_to_cpu(buf.dword);
984 break;
985 default:
986 hw_error("vfio: unsupported read size, %d bytes\n", size);
987 break;
988 }
989
82ca8912
AW
990#ifdef DEBUG_VFIO
991 {
992 VFIODevice *vdev = container_of(bar, VFIODevice, bars[bar->nr]);
993
994 DPRINTF("%s(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx
995 ", %d) = 0x%"PRIx64"\n", __func__, vdev->host.domain,
996 vdev->host.bus, vdev->host.slot, vdev->host.function,
997 bar->nr, addr, size, data);
998 }
999#endif
65501a74
AW
1000
1001 /* Same as write above */
3a4f2816 1002 vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
65501a74
AW
1003
1004 return data;
1005}
1006
1007static const MemoryRegionOps vfio_bar_ops = {
1008 .read = vfio_bar_read,
1009 .write = vfio_bar_write,
1010 .endianness = DEVICE_LITTLE_ENDIAN,
1011};
1012
f15689c7
AW
1013static void vfio_vga_write(void *opaque, hwaddr addr,
1014 uint64_t data, unsigned size)
1015{
1016 VFIOVGARegion *region = opaque;
1017 VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1018 union {
1019 uint8_t byte;
1020 uint16_t word;
1021 uint32_t dword;
1022 uint64_t qword;
1023 } buf;
1024 off_t offset = vga->fd_offset + region->offset + addr;
1025
1026 switch (size) {
1027 case 1:
1028 buf.byte = data;
1029 break;
1030 case 2:
1031 buf.word = cpu_to_le16(data);
1032 break;
1033 case 4:
1034 buf.dword = cpu_to_le32(data);
1035 break;
1036 default:
1037 hw_error("vfio: unsupported write size, %d bytes\n", size);
1038 break;
1039 }
1040
1041 if (pwrite(vga->fd, &buf, size, offset) != size) {
1042 error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m",
1043 __func__, region->offset + addr, data, size);
1044 }
1045
1046 DPRINTF("%s(0x%"HWADDR_PRIx", 0x%"PRIx64", %d)\n",
1047 __func__, region->offset + addr, data, size);
1048}
1049
1050static uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size)
1051{
1052 VFIOVGARegion *region = opaque;
1053 VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]);
1054 union {
1055 uint8_t byte;
1056 uint16_t word;
1057 uint32_t dword;
1058 uint64_t qword;
1059 } buf;
1060 uint64_t data = 0;
1061 off_t offset = vga->fd_offset + region->offset + addr;
1062
1063 if (pread(vga->fd, &buf, size, offset) != size) {
1064 error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m",
1065 __func__, region->offset + addr, size);
1066 return (uint64_t)-1;
1067 }
1068
1069 switch (size) {
1070 case 1:
1071 data = buf.byte;
1072 break;
1073 case 2:
1074 data = le16_to_cpu(buf.word);
1075 break;
1076 case 4:
1077 data = le32_to_cpu(buf.dword);
1078 break;
1079 default:
1080 hw_error("vfio: unsupported read size, %d bytes\n", size);
1081 break;
1082 }
1083
1084 DPRINTF("%s(0x%"HWADDR_PRIx", %d) = 0x%"PRIx64"\n",
1085 __func__, region->offset + addr, size, data);
1086
1087 return data;
1088}
1089
1090static const MemoryRegionOps vfio_vga_ops = {
1091 .read = vfio_vga_read,
1092 .write = vfio_vga_write,
1093 .endianness = DEVICE_LITTLE_ENDIAN,
1094};
1095
7076eabc
AW
1096/*
1097 * Device specific quirks
1098 */
1099
1100#define PCI_VENDOR_ID_ATI 0x1002
1101
1102/*
1103 * Device 1002:68f9 (Advanced Micro Devices [AMD] nee ATI Cedar PRO [Radeon
1104 * HD 5450/6350]) reports the upper byte of the physical address of the
1105 * I/O port BAR4 through VGA register 0x3c3. The BAR is 256 bytes, so the
1106 * lower byte is known to be zero. Probing for this quirk reads 0xff from
1107 * port 0x3c3 on some devices so we store the physical address and replace
1108 * reads with the virtual address any time it matches. XXX Research when
1109 * to enable quirk.
1110 */
1111static uint64_t vfio_ati_3c3_quirk_read(void *opaque,
1112 hwaddr addr, unsigned size)
1113{
1114 VFIOQuirk *quirk = opaque;
1115 VFIODevice *vdev = quirk->vdev;
1116 PCIDevice *pdev = &vdev->pdev;
1117 uint64_t data = vfio_vga_read(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1118 addr + 0x3, size);
1119
1120 if (data == quirk->data) {
1121 data = pci_get_byte(pdev->config + PCI_BASE_ADDRESS_4 + 1);
1122 DPRINTF("%s(0x3c3, 1) = 0x%"PRIx64"\n", __func__, data);
1123 }
1124
1125 return data;
1126}
1127
1128static const MemoryRegionOps vfio_ati_3c3_quirk = {
1129 .read = vfio_ati_3c3_quirk_read,
1130 .endianness = DEVICE_LITTLE_ENDIAN,
1131};
1132
1133static void vfio_vga_probe_ati_3c3_quirk(VFIODevice *vdev)
1134{
1135 PCIDevice *pdev = &vdev->pdev;
1136 off_t physoffset = vdev->config_offset + PCI_BASE_ADDRESS_4;
1137 uint32_t physbar;
1138 VFIOQuirk *quirk;
1139
1140 if (pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI ||
1141 vdev->bars[4].size < 256) {
1142 return;
1143 }
1144
1145 /* Get I/O port BAR physical address */
1146 if (pread(vdev->fd, &physbar, 4, physoffset) != 4) {
1147 error_report("vfio: probe failed for ATI/AMD 0x3c3 quirk on device "
1148 "%04x:%02x:%02x.%x", vdev->host.domain,
1149 vdev->host.bus, vdev->host.slot, vdev->host.function);
1150 return;
1151 }
1152
1153 quirk = g_malloc0(sizeof(*quirk));
1154 quirk->vdev = vdev;
1155 quirk->data = (physbar >> 8) & 0xff;
1156
2c9b15ca 1157 memory_region_init_io(&quirk->mem, NULL, &vfio_ati_3c3_quirk, quirk,
7076eabc
AW
1158 "vfio-ati-3c3-quirk", 1);
1159 memory_region_add_subregion(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem, 3,
1160 &quirk->mem);
1161
1162 QLIST_INSERT_HEAD(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks,
1163 quirk, next);
1164
1165 DPRINTF("Enabled ATI/AMD quirk 0x3c3 for device %04x:%02x:%02x.%x\n",
1166 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1167 vdev->host.function);
1168}
1169
1170/*
1171 * Device 1002:68f9 (Advanced Micro Devices [AMD] nee ATI Cedar PRO [Radeon
1172 * HD 5450/6350]) reports the physical address of MMIO BAR0 through a
1173 * write/read operation on I/O port BAR4. When uint32_t 0x4010 is written
1174 * to offset 0x0, the subsequent read from offset 0x4 returns the contents
1175 * of BAR0. Test for this quirk on all ATI/AMD devices. XXX - Note that
1176 * 0x10 is the offset of BAR0 in config sapce, is this a window to all of
1177 * config space?
1178 */
1179static uint64_t vfio_ati_4010_quirk_read(void *opaque,
1180 hwaddr addr, unsigned size)
1181{
1182 VFIOQuirk *quirk = opaque;
1183 VFIODevice *vdev = quirk->vdev;
1184 PCIDevice *pdev = &vdev->pdev;
1185 uint64_t data = vfio_bar_read(&vdev->bars[4], addr, size);
1186
1187 if (addr == 4 && size == 4 && quirk->data) {
1188 data = pci_get_long(pdev->config + PCI_BASE_ADDRESS_0);
1189 DPRINTF("%s(BAR4+0x4) = 0x%"PRIx64"\n", __func__, data);
1190 }
1191
1192 quirk->data = 0;
1193
1194 return data;
1195}
1196
1197static void vfio_ati_4010_quirk_write(void *opaque, hwaddr addr,
1198 uint64_t data, unsigned size)
1199{
1200 VFIOQuirk *quirk = opaque;
1201 VFIODevice *vdev = quirk->vdev;
1202
1203 vfio_bar_write(&vdev->bars[4], addr, data, size);
1204
1205 quirk->data = (addr == 0 && size == 4 && data == 0x4010) ? 1 : 0;
1206}
1207
1208static const MemoryRegionOps vfio_ati_4010_quirk = {
1209 .read = vfio_ati_4010_quirk_read,
1210 .write = vfio_ati_4010_quirk_write,
1211 .endianness = DEVICE_LITTLE_ENDIAN,
1212};
1213
1214static void vfio_probe_ati_4010_quirk(VFIODevice *vdev, int nr)
1215{
1216 PCIDevice *pdev = &vdev->pdev;
1217 off_t physoffset = vdev->config_offset + PCI_BASE_ADDRESS_0;
1218 uint32_t physbar0;
1219 uint64_t data;
1220 VFIOQuirk *quirk;
1221
1222 if (!vdev->has_vga || nr != 4 || !vdev->bars[0].size ||
1223 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1224 return;
1225 }
1226
1227 /* Get I/O port BAR physical address */
1228 if (pread(vdev->fd, &physbar0, 4, physoffset) != 4) {
1229 error_report("vfio: probe failed for ATI/AMD 0x4010 quirk on device "
1230 "%04x:%02x:%02x.%x", vdev->host.domain,
1231 vdev->host.bus, vdev->host.slot, vdev->host.function);
1232 return;
1233 }
1234
1235 /* Write 0x4010 to I/O port BAR offset 0 */
1236 vfio_bar_write(&vdev->bars[4], 0, 0x4010, 4);
1237 /* Read back result */
1238 data = vfio_bar_read(&vdev->bars[4], 4, 4);
1239
1240 /* If the register matches the physical address of BAR0, we need a quirk */
1241 if (data != physbar0) {
1242 return;
1243 }
1244
1245 quirk = g_malloc0(sizeof(*quirk));
1246 quirk->vdev = vdev;
1247
2c9b15ca 1248 memory_region_init_io(&quirk->mem, NULL, &vfio_ati_4010_quirk, quirk,
7076eabc
AW
1249 "vfio-ati-4010-quirk", 8);
1250 memory_region_add_subregion_overlap(&vdev->bars[nr].mem, 0, &quirk->mem, 1);
1251
1252 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1253
1254 DPRINTF("Enabled ATI/AMD quirk 0x4010 for device %04x:%02x:%02x.%x\n",
1255 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1256 vdev->host.function);
1257}
1258
1259/*
1260 * Device 1002:5b63 (Advanced Micro Devices [AMD] nee ATI RV370 [Radeon X550])
1261 * retrieves the upper half of the MMIO BAR0 physical address by writing
1262 * 0xf10 to I/O port BAR1 offset 0 and reading the result from offset 6.
1263 * XXX - 0x10 is the offset of BAR0 in PCI config space, this could provide
1264 * full access to config space. Config space is little endian, so the data
1265 * register probably starts at 0x4.
1266 */
1267static uint64_t vfio_ati_f10_quirk_read(void *opaque,
1268 hwaddr addr, unsigned size)
1269{
1270 VFIOQuirk *quirk = opaque;
1271 VFIODevice *vdev = quirk->vdev;
1272 PCIDevice *pdev = &vdev->pdev;
1273 uint64_t data = vfio_bar_read(&vdev->bars[1], addr, size);
1274
1275 if (addr == 6 && size == 2 && quirk->data) {
1276 data = pci_get_word(pdev->config + PCI_BASE_ADDRESS_0 + 2);
1277 DPRINTF("%s(BAR1+0x6) = 0x%"PRIx64"\n", __func__, data);
1278 }
1279
1280 quirk->data = 0;
1281
1282 return data;
1283}
1284
1285static void vfio_ati_f10_quirk_write(void *opaque, hwaddr addr,
1286 uint64_t data, unsigned size)
1287{
1288 VFIOQuirk *quirk = opaque;
1289 VFIODevice *vdev = quirk->vdev;
1290
1291 vfio_bar_write(&vdev->bars[1], addr, data, size);
1292
1293 quirk->data = (addr == 0 && size == 4 && data == 0xf10) ? 1 : 0;
1294}
1295
1296static const MemoryRegionOps vfio_ati_f10_quirk = {
1297 .read = vfio_ati_f10_quirk_read,
1298 .write = vfio_ati_f10_quirk_write,
1299 .endianness = DEVICE_LITTLE_ENDIAN,
1300};
1301
1302static void vfio_probe_ati_f10_quirk(VFIODevice *vdev, int nr)
1303{
1304 PCIDevice *pdev = &vdev->pdev;
1305 off_t physoffset = vdev->config_offset + PCI_BASE_ADDRESS_0;
1306 uint32_t physbar0;
1307 uint64_t data;
1308 VFIOQuirk *quirk;
1309
1310 if (!vdev->has_vga || nr != 1 || !vdev->bars[0].size ||
1311 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_ATI) {
1312 return;
1313 }
1314
1315 /* Get I/O port BAR physical address */
1316 if (pread(vdev->fd, &physbar0, 4, physoffset) != 4) {
1317 error_report("vfio: probe failed for ATI/AMD 0xf10 quirk on device "
1318 "%04x:%02x:%02x.%x", vdev->host.domain,
1319 vdev->host.bus, vdev->host.slot, vdev->host.function);
1320 return;
1321 }
1322
1323 vfio_bar_write(&vdev->bars[1], 0, 0xf10, 4);
1324 data = vfio_bar_read(&vdev->bars[1], 0x6, 2);
1325
1326 /* If the register matches the physical address of BAR0, we need a quirk */
1327 if (data != (le32_to_cpu(physbar0) >> 16)) {
1328 return;
1329 }
1330
1331 quirk = g_malloc0(sizeof(*quirk));
1332 quirk->vdev = vdev;
1333
2c9b15ca 1334 memory_region_init_io(&quirk->mem, NULL, &vfio_ati_f10_quirk, quirk,
7076eabc
AW
1335 "vfio-ati-f10-quirk", 8);
1336 memory_region_add_subregion_overlap(&vdev->bars[nr].mem, 0, &quirk->mem, 1);
1337
1338 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1339
1340 DPRINTF("Enabled ATI/AMD quirk 0xf10 for device %04x:%02x:%02x.%x\n",
1341 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1342 vdev->host.function);
1343}
1344
1345#define PCI_VENDOR_ID_NVIDIA 0x10de
1346
1347/*
1348 * Nvidia has several different methods to get to config space, the
1349 * nouveu project has several of these documented here:
1350 * https://github.com/pathscale/envytools/tree/master/hwdocs
1351 *
1352 * The first quirk is actually not documented in envytools and is found
1353 * on 10de:01d1 (NVIDIA Corporation G72 [GeForce 7300 LE]). This is an
1354 * NV46 chipset. The backdoor uses the legacy VGA I/O ports to access
1355 * the mirror of PCI config space found at BAR0 offset 0x1800. The access
1356 * sequence first writes 0x338 to I/O port 0x3d4. The target offset is
1357 * then written to 0x3d0. Finally 0x538 is written for a read and 0x738
1358 * is written for a write to 0x3d4. The BAR0 offset is then accessible
1359 * through 0x3d0. This quirk doesn't seem to be necessary on newer cards
1360 * that use the I/O port BAR5 window but it doesn't hurt to leave it.
1361 */
1362enum {
1363 NV_3D0_NONE,
1364 NV_3D0_SELECT,
1365 NV_3D0_WINDOW,
1366 NV_3D0_READ,
1367 NV_3D0_WRITE,
1368};
1369
1370static uint64_t vfio_nvidia_3d0_quirk_read(void *opaque,
1371 hwaddr addr, unsigned size)
1372{
1373 VFIOQuirk *quirk = opaque;
1374 VFIODevice *vdev = quirk->vdev;
1375 PCIDevice *pdev = &vdev->pdev;
1376 uint64_t data = vfio_vga_read(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1377 addr + 0x10, size);
1378
1379 if (quirk->data == NV_3D0_READ && addr == 0) {
1380 data = vfio_pci_read_config(pdev, quirk->data2, size);
1381 DPRINTF("%s(0x3d0, %d) = 0x%"PRIx64"\n", __func__, size, data);
1382 }
1383
1384 quirk->data = NV_3D0_NONE;
1385
1386 return data;
1387}
1388
1389static void vfio_nvidia_3d0_quirk_write(void *opaque, hwaddr addr,
1390 uint64_t data, unsigned size)
1391{
1392 VFIOQuirk *quirk = opaque;
1393 VFIODevice *vdev = quirk->vdev;
1394 PCIDevice *pdev = &vdev->pdev;
1395
1396 switch (quirk->data) {
1397 case NV_3D0_NONE:
1398 if (addr == 4 && data == 0x338) {
1399 quirk->data = NV_3D0_SELECT;
1400 }
1401 break;
1402 case NV_3D0_SELECT:
1403 quirk->data = NV_3D0_NONE;
1404 if (addr == 0 && (data & ~0xff) == 0x1800) {
1405 quirk->data = NV_3D0_WINDOW;
1406 quirk->data2 = data & 0xff;
1407 }
1408 break;
1409 case NV_3D0_WINDOW:
1410 quirk->data = NV_3D0_NONE;
1411 if (addr == 4) {
1412 if (data == 0x538) {
1413 quirk->data = NV_3D0_READ;
1414 } else if (data == 0x738) {
1415 quirk->data = NV_3D0_WRITE;
1416 }
1417 }
1418 break;
1419 case NV_3D0_WRITE:
1420 quirk->data = NV_3D0_NONE;
1421 if (addr == 0) {
1422 vfio_pci_write_config(pdev, quirk->data2, data, size);
1423 DPRINTF("%s(0x3d0, 0x%"PRIx64", %d)\n", __func__, data, size);
1424 return;
1425 }
1426 break;
1427 default:
1428 quirk->data = NV_3D0_NONE;
1429 }
1430
1431 vfio_vga_write(&vdev->vga.region[QEMU_PCI_VGA_IO_HI],
1432 addr + 0x10, data, size);
1433}
1434
1435static const MemoryRegionOps vfio_nvidia_3d0_quirk = {
1436 .read = vfio_nvidia_3d0_quirk_read,
1437 .write = vfio_nvidia_3d0_quirk_write,
1438 .endianness = DEVICE_LITTLE_ENDIAN,
1439};
1440
1441static void vfio_vga_probe_nvidia_3d0_quirk(VFIODevice *vdev)
1442{
1443 PCIDevice *pdev = &vdev->pdev;
1444 VFIOQuirk *quirk;
1445
1446 if (pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA ||
1447 !vdev->bars[1].size) {
1448 return;
1449 }
1450
1451 quirk = g_malloc0(sizeof(*quirk));
1452 quirk->vdev = vdev;
1453
2c9b15ca 1454 memory_region_init_io(&quirk->mem, NULL, &vfio_nvidia_3d0_quirk, quirk,
7076eabc
AW
1455 "vfio-nvidia-3d0-quirk", 6);
1456 memory_region_add_subregion(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem,
1457 0x10, &quirk->mem);
1458
1459 QLIST_INSERT_HEAD(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks,
1460 quirk, next);
1461
1462 DPRINTF("Enabled NVIDIA VGA 0x3d0 quirk for device %04x:%02x:%02x.%x\n",
1463 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1464 vdev->host.function);
1465}
1466
1467/*
1468 * The second quirk is documented in envytools. The I/O port BAR5 is just
1469 * a set of address/data ports to the MMIO BARs. The BAR we care about is
1470 * again BAR0. This backdoor is apparently a bit newer than the one above
1471 * so we need to not only trap 256 bytes @0x1800, but all of PCI config
1472 * space, including extended space is available at the 4k @0x88000.
1473 */
1474enum {
1475 NV_BAR5_ADDRESS = 0x1,
1476 NV_BAR5_ENABLE = 0x2,
1477 NV_BAR5_MASTER = 0x4,
1478 NV_BAR5_VALID = 0x7,
1479};
1480
1481static uint64_t vfio_nvidia_bar5_window_quirk_read(void *opaque,
1482 hwaddr addr, unsigned size)
1483{
1484 VFIOQuirk *quirk = opaque;
1485 VFIODevice *vdev = quirk->vdev;
1486 uint64_t data = vfio_bar_read(&vdev->bars[5], addr, size);
1487
1488 if (addr == 0xc && quirk->data == NV_BAR5_VALID) {
1489 data = vfio_pci_read_config(&vdev->pdev, quirk->data2, size);
1490 DPRINTF("%s(%04x:%02x:%02x.%x:BAR5+0x%"HWADDR_PRIx", %d) = 0x%"
1491 PRIx64"\n", __func__, vdev->host.domain, vdev->host.bus,
1492 vdev->host.slot, vdev->host.function, addr, size, data);
1493 }
1494
1495 return data;
1496}
1497
1498static void vfio_nvidia_bar5_window_quirk_write(void *opaque, hwaddr addr,
1499 uint64_t data, unsigned size)
1500{
1501 VFIOQuirk *quirk = opaque;
1502 VFIODevice *vdev = quirk->vdev;
1503
1504 /*
1505 * Use quirk->data to track enables and quirk->data2 for the offset
1506 */
1507 switch (addr) {
1508 case 0x0:
1509 if (data & 0x1) {
1510 quirk->data |= NV_BAR5_MASTER;
1511 } else {
1512 quirk->data &= ~NV_BAR5_MASTER;
1513 }
1514 break;
1515 case 0x4:
1516 if (data & 0x1) {
1517 quirk->data |= NV_BAR5_ENABLE;
1518 } else {
1519 quirk->data &= ~NV_BAR5_ENABLE;
1520 }
1521 break;
1522 case 0x8:
1523 if (quirk->data & NV_BAR5_MASTER) {
1524 if ((data & ~0xfff) == 0x88000) {
1525 quirk->data |= NV_BAR5_ADDRESS;
1526 quirk->data2 = data & 0xfff;
1527 } else if ((data & ~0xff) == 0x1800) {
1528 quirk->data |= NV_BAR5_ADDRESS;
1529 quirk->data2 = data & 0xff;
1530 } else {
1531 quirk->data &= ~NV_BAR5_ADDRESS;
1532 }
1533 }
1534 break;
1535 case 0xc:
1536 if (quirk->data == NV_BAR5_VALID) {
1537 vfio_pci_write_config(&vdev->pdev, quirk->data2, data, size);
1538 DPRINTF("%s(%04x:%02x:%02x.%x:BAR5+0x%"HWADDR_PRIx", 0x%"
1539 PRIx64", %d)\n", __func__, vdev->host.domain,
1540 vdev->host.bus, vdev->host.slot, vdev->host.function,
1541 addr, data, size);
1542 return;
1543 }
1544 }
1545
1546 vfio_bar_write(&vdev->bars[5], addr, data, size);
1547}
1548
1549static const MemoryRegionOps vfio_nvidia_bar5_window_quirk = {
1550 .read = vfio_nvidia_bar5_window_quirk_read,
1551 .write = vfio_nvidia_bar5_window_quirk_write,
1552 .valid.min_access_size = 4,
1553 .endianness = DEVICE_LITTLE_ENDIAN,
1554};
1555
1556static void vfio_probe_nvidia_bar5_window_quirk(VFIODevice *vdev, int nr)
1557{
1558 PCIDevice *pdev = &vdev->pdev;
1559 VFIOQuirk *quirk;
1560
1561 if (!vdev->has_vga || nr != 5 ||
1562 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1563 return;
1564 }
1565
1566 quirk = g_malloc0(sizeof(*quirk));
1567 quirk->vdev = vdev;
1568
2c9b15ca 1569 memory_region_init_io(&quirk->mem, NULL, &vfio_nvidia_bar5_window_quirk, quirk,
7076eabc
AW
1570 "vfio-nvidia-bar5-window-quirk", 16);
1571 memory_region_add_subregion_overlap(&vdev->bars[nr].mem, 0, &quirk->mem, 1);
1572
1573 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1574
1575 DPRINTF("Enabled NVIDIA BAR5 window quirk for device %04x:%02x:%02x.%x\n",
1576 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1577 vdev->host.function);
1578}
1579
1580/*
1581 * Finally, BAR0 itself. We want to redirect any accesses to either
1582 * 0x1800 or 0x88000 through the PCI config space access functions.
1583 *
1584 * NB - quirk at a page granularity or else they don't seem to work when
1585 * BARs are mmap'd
1586 *
1587 * Here's offset 0x88000...
1588 */
1589static uint64_t vfio_nvidia_bar0_88000_quirk_read(void *opaque,
1590 hwaddr addr, unsigned size)
1591{
1592 VFIOQuirk *quirk = opaque;
1593 VFIODevice *vdev = quirk->vdev;
1594 hwaddr base = 0x88000 & TARGET_PAGE_MASK;
1595 hwaddr offset = 0x88000 & ~TARGET_PAGE_MASK;
1596 uint64_t data = vfio_bar_read(&vdev->bars[0], addr + base, size);
1597
1598 if (ranges_overlap(addr, size, offset, PCI_CONFIG_SPACE_SIZE)) {
1599 data = vfio_pci_read_config(&vdev->pdev, addr - offset, size);
1600
1601 DPRINTF("%s(%04x:%02x:%02x.%x:BAR0+0x%"HWADDR_PRIx", %d) = 0x%"
1602 PRIx64"\n", __func__, vdev->host.domain, vdev->host.bus,
1603 vdev->host.slot, vdev->host.function, addr + base, size, data);
1604 }
1605
1606 return data;
1607}
1608
1609static void vfio_nvidia_bar0_88000_quirk_write(void *opaque, hwaddr addr,
1610 uint64_t data, unsigned size)
1611{
1612 VFIOQuirk *quirk = opaque;
1613 VFIODevice *vdev = quirk->vdev;
1614 hwaddr base = 0x88000 & TARGET_PAGE_MASK;
1615 hwaddr offset = 0x88000 & ~TARGET_PAGE_MASK;
1616
1617 if (ranges_overlap(addr, size, offset, PCI_CONFIG_SPACE_SIZE)) {
1618 vfio_pci_write_config(&vdev->pdev, addr - offset, data, size);
1619
1620 DPRINTF("%s(%04x:%02x:%02x.%x:BAR0+0x%"HWADDR_PRIx", 0x%"
1621 PRIx64", %d)\n", __func__, vdev->host.domain, vdev->host.bus,
1622 vdev->host.slot, vdev->host.function, addr + base, data, size);
1623 } else {
1624 vfio_bar_write(&vdev->bars[0], addr + base, data, size);
1625 }
1626}
1627
1628static const MemoryRegionOps vfio_nvidia_bar0_88000_quirk = {
1629 .read = vfio_nvidia_bar0_88000_quirk_read,
1630 .write = vfio_nvidia_bar0_88000_quirk_write,
1631 .endianness = DEVICE_LITTLE_ENDIAN,
1632};
1633
1634static void vfio_probe_nvidia_bar0_88000_quirk(VFIODevice *vdev, int nr)
1635{
1636 PCIDevice *pdev = &vdev->pdev;
1637 VFIOQuirk *quirk;
1638
1639 if (!vdev->has_vga || nr != 0 ||
1640 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1641 return;
1642 }
1643
1644 quirk = g_malloc0(sizeof(*quirk));
1645 quirk->vdev = vdev;
1646
2c9b15ca 1647 memory_region_init_io(&quirk->mem, NULL, &vfio_nvidia_bar0_88000_quirk, quirk,
7076eabc
AW
1648 "vfio-nvidia-bar0-88000-quirk",
1649 TARGET_PAGE_ALIGN(PCIE_CONFIG_SPACE_SIZE));
1650 memory_region_add_subregion_overlap(&vdev->bars[nr].mem,
1651 0x88000 & TARGET_PAGE_MASK,
1652 &quirk->mem, 1);
1653
1654 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1655
1656 DPRINTF("Enabled NVIDIA BAR0 0x88000 quirk for device %04x:%02x:%02x.%x\n",
1657 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1658 vdev->host.function);
1659}
1660
1661/*
1662 * And here's the same for BAR0 offset 0x1800...
1663 */
1664static uint64_t vfio_nvidia_bar0_1800_quirk_read(void *opaque,
1665 hwaddr addr, unsigned size)
1666{
1667 VFIOQuirk *quirk = opaque;
1668 VFIODevice *vdev = quirk->vdev;
1669 hwaddr base = 0x1800 & TARGET_PAGE_MASK;
1670 hwaddr offset = 0x1800 & ~TARGET_PAGE_MASK;
1671 uint64_t data = vfio_bar_read(&vdev->bars[0], addr + base, size);
1672
1673 if (ranges_overlap(addr, size, offset, PCI_CONFIG_SPACE_SIZE)) {
1674 data = vfio_pci_read_config(&vdev->pdev, addr - offset, size);
1675
1676 DPRINTF("%s(%04x:%02x:%02x.%x:BAR0+0x%"HWADDR_PRIx", %d) = 0x%"
1677 PRIx64"\n", __func__, vdev->host.domain, vdev->host.bus,
1678 vdev->host.slot, vdev->host.function, addr + base, size, data);
1679 }
1680
1681 return data;
1682}
1683
1684static void vfio_nvidia_bar0_1800_quirk_write(void *opaque, hwaddr addr,
1685 uint64_t data, unsigned size)
1686{
1687 VFIOQuirk *quirk = opaque;
1688 VFIODevice *vdev = quirk->vdev;
1689 hwaddr base = 0x1800 & TARGET_PAGE_MASK;
1690 hwaddr offset = 0x1800 & ~TARGET_PAGE_MASK;
1691
1692 if (ranges_overlap(addr, size, offset, PCI_CONFIG_SPACE_SIZE)) {
1693 vfio_pci_write_config(&vdev->pdev, addr - offset, data, size);
1694
1695 DPRINTF("%s(%04x:%02x:%02x.%x:BAR0+0x%"HWADDR_PRIx", 0x%"
1696 PRIx64", %d)\n", __func__, vdev->host.domain, vdev->host.bus,
1697 vdev->host.slot, vdev->host.function, addr + base, data, size);
1698 } else {
1699 vfio_bar_write(&vdev->bars[0], addr + base, data, size);
1700 }
1701}
1702
1703static const MemoryRegionOps vfio_nvidia_bar0_1800_quirk = {
1704 .read = vfio_nvidia_bar0_1800_quirk_read,
1705 .write = vfio_nvidia_bar0_1800_quirk_write,
1706 .endianness = DEVICE_LITTLE_ENDIAN,
1707};
1708
1709static void vfio_probe_nvidia_bar0_1800_quirk(VFIODevice *vdev, int nr)
1710{
1711 PCIDevice *pdev = &vdev->pdev;
1712 VFIOQuirk *quirk;
1713
1714 if (!vdev->has_vga || nr != 0 ||
1715 pci_get_word(pdev->config + PCI_VENDOR_ID) != PCI_VENDOR_ID_NVIDIA) {
1716 return;
1717 }
1718
1719 /* Log the chipset ID */
1720 DPRINTF("Nvidia NV%02x\n",
1721 (unsigned int)(vfio_bar_read(&vdev->bars[0], 0, 4) >> 20) & 0xff);
1722
1723 quirk = g_malloc0(sizeof(*quirk));
1724 quirk->vdev = vdev;
1725
2c9b15ca 1726 memory_region_init_io(&quirk->mem, NULL, &vfio_nvidia_bar0_1800_quirk, quirk,
7076eabc
AW
1727 "vfio-nvidia-bar0-1800-quirk",
1728 TARGET_PAGE_ALIGN(PCI_CONFIG_SPACE_SIZE));
1729 memory_region_add_subregion_overlap(&vdev->bars[nr].mem,
1730 0x1800 & TARGET_PAGE_MASK,
1731 &quirk->mem, 1);
1732
1733 QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1734
1735 DPRINTF("Enabled NVIDIA BAR0 0x1800 quirk for device %04x:%02x:%02x.%x\n",
1736 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1737 vdev->host.function);
1738}
1739
1740/*
1741 * TODO - Some Nvidia devices provide config access to their companion HDA
1742 * device and even to their parent bridge via these config space mirrors.
1743 * Add quirks for those regions.
1744 */
1745
1746/*
1747 * Common quirk probe entry points.
1748 */
1749static void vfio_vga_quirk_setup(VFIODevice *vdev)
1750{
1751 vfio_vga_probe_ati_3c3_quirk(vdev);
1752 vfio_vga_probe_nvidia_3d0_quirk(vdev);
1753}
1754
1755static void vfio_vga_quirk_teardown(VFIODevice *vdev)
1756{
1757 int i;
1758
1759 for (i = 0; i < ARRAY_SIZE(vdev->vga.region); i++) {
1760 while (!QLIST_EMPTY(&vdev->vga.region[i].quirks)) {
1761 VFIOQuirk *quirk = QLIST_FIRST(&vdev->vga.region[i].quirks);
1762 memory_region_del_subregion(&vdev->vga.region[i].mem, &quirk->mem);
1763 QLIST_REMOVE(quirk, next);
1764 g_free(quirk);
1765 }
1766 }
1767}
1768
1769static void vfio_bar_quirk_setup(VFIODevice *vdev, int nr)
1770{
1771 vfio_probe_ati_4010_quirk(vdev, nr);
1772 vfio_probe_ati_f10_quirk(vdev, nr);
1773 vfio_probe_nvidia_bar5_window_quirk(vdev, nr);
1774 vfio_probe_nvidia_bar0_88000_quirk(vdev, nr);
1775 vfio_probe_nvidia_bar0_1800_quirk(vdev, nr);
1776}
1777
1778static void vfio_bar_quirk_teardown(VFIODevice *vdev, int nr)
1779{
1780 VFIOBAR *bar = &vdev->bars[nr];
1781
1782 while (!QLIST_EMPTY(&bar->quirks)) {
1783 VFIOQuirk *quirk = QLIST_FIRST(&bar->quirks);
1784 memory_region_del_subregion(&bar->mem, &quirk->mem);
1785 QLIST_REMOVE(quirk, next);
1786 g_free(quirk);
1787 }
1788}
1789
65501a74
AW
1790/*
1791 * PCI config space
1792 */
1793static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
1794{
1795 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
4b5d5e87 1796 uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val;
65501a74 1797
4b5d5e87
AW
1798 memcpy(&emu_bits, vdev->emulated_config_bits + addr, len);
1799 emu_bits = le32_to_cpu(emu_bits);
65501a74 1800
4b5d5e87
AW
1801 if (emu_bits) {
1802 emu_val = pci_default_read_config(pdev, addr, len);
1803 }
1804
1805 if (~emu_bits & (0xffffffffU >> (32 - len * 8))) {
1806 ssize_t ret;
1807
1808 ret = pread(vdev->fd, &phys_val, len, vdev->config_offset + addr);
1809 if (ret != len) {
312fd5f2 1810 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m",
65501a74
AW
1811 __func__, vdev->host.domain, vdev->host.bus,
1812 vdev->host.slot, vdev->host.function, addr, len);
1813 return -errno;
1814 }
4b5d5e87 1815 phys_val = le32_to_cpu(phys_val);
65501a74
AW
1816 }
1817
4b5d5e87 1818 val = (emu_val & emu_bits) | (phys_val & ~emu_bits);
65501a74
AW
1819
1820 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, len=0x%x) %x\n", __func__,
1821 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1822 vdev->host.function, addr, len, val);
1823
1824 return val;
1825}
1826
1827static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
1828 uint32_t val, int len)
1829{
1830 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1831 uint32_t val_le = cpu_to_le32(val);
1832
1833 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, 0x%x, len=0x%x)\n", __func__,
1834 vdev->host.domain, vdev->host.bus, vdev->host.slot,
1835 vdev->host.function, addr, val, len);
1836
1837 /* Write everything to VFIO, let it filter out what we can't write */
1838 if (pwrite(vdev->fd, &val_le, len, vdev->config_offset + addr) != len) {
312fd5f2 1839 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m",
65501a74
AW
1840 __func__, vdev->host.domain, vdev->host.bus,
1841 vdev->host.slot, vdev->host.function, addr, val, len);
1842 }
1843
65501a74
AW
1844 /* MSI/MSI-X Enabling/Disabling */
1845 if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
1846 ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
1847 int is_enabled, was_enabled = msi_enabled(pdev);
1848
1849 pci_default_write_config(pdev, addr, val, len);
1850
1851 is_enabled = msi_enabled(pdev);
1852
1853 if (!was_enabled && is_enabled) {
1854 vfio_enable_msi(vdev);
1855 } else if (was_enabled && !is_enabled) {
fd704adc 1856 vfio_disable_msi(vdev);
65501a74 1857 }
4b5d5e87 1858 } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
65501a74
AW
1859 ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
1860 int is_enabled, was_enabled = msix_enabled(pdev);
1861
1862 pci_default_write_config(pdev, addr, val, len);
1863
1864 is_enabled = msix_enabled(pdev);
1865
1866 if (!was_enabled && is_enabled) {
fd704adc 1867 vfio_enable_msix(vdev);
65501a74 1868 } else if (was_enabled && !is_enabled) {
fd704adc 1869 vfio_disable_msix(vdev);
65501a74 1870 }
4b5d5e87
AW
1871 } else {
1872 /* Write everything to QEMU to keep emulated bits correct */
1873 pci_default_write_config(pdev, addr, val, len);
65501a74
AW
1874 }
1875}
1876
1877/*
1878 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
1879 */
af6bc27e 1880static int vfio_dma_unmap(VFIOContainer *container,
a8170e5e 1881 hwaddr iova, ram_addr_t size)
af6bc27e
AW
1882{
1883 struct vfio_iommu_type1_dma_unmap unmap = {
1884 .argsz = sizeof(unmap),
1885 .flags = 0,
1886 .iova = iova,
1887 .size = size,
1888 };
1889
1890 if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
1891 DPRINTF("VFIO_UNMAP_DMA: %d\n", -errno);
1892 return -errno;
1893 }
1894
1895 return 0;
1896}
1897
a8170e5e 1898static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
65501a74
AW
1899 ram_addr_t size, void *vaddr, bool readonly)
1900{
1901 struct vfio_iommu_type1_dma_map map = {
1902 .argsz = sizeof(map),
1903 .flags = VFIO_DMA_MAP_FLAG_READ,
5976cdd5 1904 .vaddr = (__u64)(uintptr_t)vaddr,
65501a74
AW
1905 .iova = iova,
1906 .size = size,
1907 };
1908
1909 if (!readonly) {
1910 map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
1911 }
1912
12af1344
AW
1913 /*
1914 * Try the mapping, if it fails with EBUSY, unmap the region and try
1915 * again. This shouldn't be necessary, but we sometimes see it in
1916 * the the VGA ROM space.
1917 */
1918 if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
1919 (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
1920 ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
1921 return 0;
65501a74
AW
1922 }
1923
12af1344
AW
1924 DPRINTF("VFIO_MAP_DMA: %d\n", -errno);
1925 return -errno;
65501a74
AW
1926}
1927
65501a74
AW
1928static bool vfio_listener_skipped_section(MemoryRegionSection *section)
1929{
1930 return !memory_region_is_ram(section->mr);
1931}
1932
1933static void vfio_listener_region_add(MemoryListener *listener,
1934 MemoryRegionSection *section)
1935{
1936 VFIOContainer *container = container_of(listener, VFIOContainer,
1937 iommu_data.listener);
a8170e5e 1938 hwaddr iova, end;
65501a74
AW
1939 void *vaddr;
1940 int ret;
1941
06d985f5
AK
1942 assert(!memory_region_is_iommu(section->mr));
1943
65501a74 1944 if (vfio_listener_skipped_section(section)) {
82ca8912 1945 DPRINTF("SKIPPING region_add %"HWADDR_PRIx" - %"PRIx64"\n",
65501a74
AW
1946 section->offset_within_address_space,
1947 section->offset_within_address_space + section->size - 1);
1948 return;
1949 }
1950
1951 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
1952 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
312fd5f2 1953 error_report("%s received unaligned region", __func__);
65501a74
AW
1954 return;
1955 }
1956
1957 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
052e87b0 1958 end = (section->offset_within_address_space + int128_get64(section->size)) &
65501a74
AW
1959 TARGET_PAGE_MASK;
1960
1961 if (iova >= end) {
1962 return;
1963 }
1964
1965 vaddr = memory_region_get_ram_ptr(section->mr) +
1966 section->offset_within_region +
1967 (iova - section->offset_within_address_space);
1968
82ca8912 1969 DPRINTF("region_add %"HWADDR_PRIx" - %"HWADDR_PRIx" [%p]\n",
65501a74
AW
1970 iova, end - 1, vaddr);
1971
dfde4e6e 1972 memory_region_ref(section->mr);
65501a74
AW
1973 ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
1974 if (ret) {
a8170e5e 1975 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
312fd5f2 1976 "0x%"HWADDR_PRIx", %p) = %d (%m)",
65501a74
AW
1977 container, iova, end - iova, vaddr, ret);
1978 }
1979}
1980
1981static void vfio_listener_region_del(MemoryListener *listener,
1982 MemoryRegionSection *section)
1983{
1984 VFIOContainer *container = container_of(listener, VFIOContainer,
1985 iommu_data.listener);
a8170e5e 1986 hwaddr iova, end;
65501a74
AW
1987 int ret;
1988
1989 if (vfio_listener_skipped_section(section)) {
82ca8912 1990 DPRINTF("SKIPPING region_del %"HWADDR_PRIx" - %"PRIx64"\n",
65501a74
AW
1991 section->offset_within_address_space,
1992 section->offset_within_address_space + section->size - 1);
1993 return;
1994 }
1995
1996 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
1997 (section->offset_within_region & ~TARGET_PAGE_MASK))) {
312fd5f2 1998 error_report("%s received unaligned region", __func__);
65501a74
AW
1999 return;
2000 }
2001
2002 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
052e87b0 2003 end = (section->offset_within_address_space + int128_get64(section->size)) &
65501a74
AW
2004 TARGET_PAGE_MASK;
2005
2006 if (iova >= end) {
2007 return;
2008 }
2009
82ca8912 2010 DPRINTF("region_del %"HWADDR_PRIx" - %"HWADDR_PRIx"\n",
65501a74
AW
2011 iova, end - 1);
2012
2013 ret = vfio_dma_unmap(container, iova, end - iova);
dfde4e6e 2014 memory_region_unref(section->mr);
65501a74 2015 if (ret) {
a8170e5e 2016 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
312fd5f2 2017 "0x%"HWADDR_PRIx") = %d (%m)",
65501a74
AW
2018 container, iova, end - iova, ret);
2019 }
2020}
2021
2022static MemoryListener vfio_memory_listener = {
65501a74
AW
2023 .region_add = vfio_listener_region_add,
2024 .region_del = vfio_listener_region_del,
65501a74
AW
2025};
2026
2027static void vfio_listener_release(VFIOContainer *container)
2028{
2029 memory_listener_unregister(&container->iommu_data.listener);
2030}
2031
2032/*
2033 * Interrupt setup
2034 */
2035static void vfio_disable_interrupts(VFIODevice *vdev)
2036{
2037 switch (vdev->interrupt) {
2038 case VFIO_INT_INTx:
2039 vfio_disable_intx(vdev);
2040 break;
2041 case VFIO_INT_MSI:
fd704adc 2042 vfio_disable_msi(vdev);
65501a74
AW
2043 break;
2044 case VFIO_INT_MSIX:
fd704adc 2045 vfio_disable_msix(vdev);
65501a74
AW
2046 break;
2047 }
2048}
2049
2050static int vfio_setup_msi(VFIODevice *vdev, int pos)
2051{
2052 uint16_t ctrl;
2053 bool msi_64bit, msi_maskbit;
2054 int ret, entries;
2055
65501a74
AW
2056 if (pread(vdev->fd, &ctrl, sizeof(ctrl),
2057 vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
2058 return -errno;
2059 }
2060 ctrl = le16_to_cpu(ctrl);
2061
2062 msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
2063 msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
2064 entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
2065
2066 DPRINTF("%04x:%02x:%02x.%x PCI MSI CAP @0x%x\n", vdev->host.domain,
2067 vdev->host.bus, vdev->host.slot, vdev->host.function, pos);
2068
2069 ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit);
2070 if (ret < 0) {
e43b9a5a
AW
2071 if (ret == -ENOTSUP) {
2072 return 0;
2073 }
312fd5f2 2074 error_report("vfio: msi_init failed");
65501a74
AW
2075 return ret;
2076 }
2077 vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
2078
2079 return 0;
2080}
2081
2082/*
2083 * We don't have any control over how pci_add_capability() inserts
2084 * capabilities into the chain. In order to setup MSI-X we need a
2085 * MemoryRegion for the BAR. In order to setup the BAR and not
2086 * attempt to mmap the MSI-X table area, which VFIO won't allow, we
2087 * need to first look for where the MSI-X table lives. So we
2088 * unfortunately split MSI-X setup across two functions.
2089 */
2090static int vfio_early_setup_msix(VFIODevice *vdev)
2091{
2092 uint8_t pos;
2093 uint16_t ctrl;
2094 uint32_t table, pba;
2095
2096 pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
2097 if (!pos) {
2098 return 0;
2099 }
2100
2101 if (pread(vdev->fd, &ctrl, sizeof(ctrl),
2102 vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
2103 return -errno;
2104 }
2105
2106 if (pread(vdev->fd, &table, sizeof(table),
2107 vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
2108 return -errno;
2109 }
2110
2111 if (pread(vdev->fd, &pba, sizeof(pba),
2112 vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
2113 return -errno;
2114 }
2115
2116 ctrl = le16_to_cpu(ctrl);
2117 table = le32_to_cpu(table);
2118 pba = le32_to_cpu(pba);
2119
2120 vdev->msix = g_malloc0(sizeof(*(vdev->msix)));
2121 vdev->msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
2122 vdev->msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
2123 vdev->msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
2124 vdev->msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
2125 vdev->msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
2126
2127 DPRINTF("%04x:%02x:%02x.%x "
2128 "PCI MSI-X CAP @0x%x, BAR %d, offset 0x%x, entries %d\n",
2129 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2130 vdev->host.function, pos, vdev->msix->table_bar,
2131 vdev->msix->table_offset, vdev->msix->entries);
2132
2133 return 0;
2134}
2135
2136static int vfio_setup_msix(VFIODevice *vdev, int pos)
2137{
2138 int ret;
2139
65501a74
AW
2140 ret = msix_init(&vdev->pdev, vdev->msix->entries,
2141 &vdev->bars[vdev->msix->table_bar].mem,
2142 vdev->msix->table_bar, vdev->msix->table_offset,
2143 &vdev->bars[vdev->msix->pba_bar].mem,
2144 vdev->msix->pba_bar, vdev->msix->pba_offset, pos);
2145 if (ret < 0) {
e43b9a5a
AW
2146 if (ret == -ENOTSUP) {
2147 return 0;
2148 }
312fd5f2 2149 error_report("vfio: msix_init failed");
65501a74
AW
2150 return ret;
2151 }
2152
65501a74
AW
2153 return 0;
2154}
2155
2156static void vfio_teardown_msi(VFIODevice *vdev)
2157{
2158 msi_uninit(&vdev->pdev);
2159
2160 if (vdev->msix) {
65501a74
AW
2161 msix_uninit(&vdev->pdev, &vdev->bars[vdev->msix->table_bar].mem,
2162 &vdev->bars[vdev->msix->pba_bar].mem);
2163 }
2164}
2165
2166/*
2167 * Resource setup
2168 */
2169static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled)
2170{
2171 int i;
2172
2173 for (i = 0; i < PCI_ROM_SLOT; i++) {
2174 VFIOBAR *bar = &vdev->bars[i];
2175
2176 if (!bar->size) {
2177 continue;
2178 }
2179
2180 memory_region_set_enabled(&bar->mmap_mem, enabled);
2181 if (vdev->msix && vdev->msix->table_bar == i) {
2182 memory_region_set_enabled(&vdev->msix->mmap_mem, enabled);
2183 }
2184 }
2185}
2186
2187static void vfio_unmap_bar(VFIODevice *vdev, int nr)
2188{
2189 VFIOBAR *bar = &vdev->bars[nr];
2190
2191 if (!bar->size) {
2192 return;
2193 }
2194
7076eabc
AW
2195 vfio_bar_quirk_teardown(vdev, nr);
2196
65501a74
AW
2197 memory_region_del_subregion(&bar->mem, &bar->mmap_mem);
2198 munmap(bar->mmap, memory_region_size(&bar->mmap_mem));
2199
2200 if (vdev->msix && vdev->msix->table_bar == nr) {
2201 memory_region_del_subregion(&bar->mem, &vdev->msix->mmap_mem);
2202 munmap(vdev->msix->mmap, memory_region_size(&vdev->msix->mmap_mem));
2203 }
2204
2205 memory_region_destroy(&bar->mem);
2206}
2207
2208static int vfio_mmap_bar(VFIOBAR *bar, MemoryRegion *mem, MemoryRegion *submem,
2209 void **map, size_t size, off_t offset,
2210 const char *name)
2211{
2212 int ret = 0;
2213
82ca8912 2214 if (VFIO_ALLOW_MMAP && size && bar->flags & VFIO_REGION_INFO_FLAG_MMAP) {
65501a74
AW
2215 int prot = 0;
2216
2217 if (bar->flags & VFIO_REGION_INFO_FLAG_READ) {
2218 prot |= PROT_READ;
2219 }
2220
2221 if (bar->flags & VFIO_REGION_INFO_FLAG_WRITE) {
2222 prot |= PROT_WRITE;
2223 }
2224
2225 *map = mmap(NULL, size, prot, MAP_SHARED,
2226 bar->fd, bar->fd_offset + offset);
2227 if (*map == MAP_FAILED) {
2228 *map = NULL;
2229 ret = -errno;
2230 goto empty_region;
2231 }
2232
2c9b15ca 2233 memory_region_init_ram_ptr(submem, NULL, name, size, *map);
65501a74
AW
2234 } else {
2235empty_region:
2236 /* Create a zero sized sub-region to make cleanup easy. */
2c9b15ca 2237 memory_region_init(submem, NULL, name, 0);
65501a74
AW
2238 }
2239
2240 memory_region_add_subregion(mem, offset, submem);
2241
2242 return ret;
2243}
2244
2245static void vfio_map_bar(VFIODevice *vdev, int nr)
2246{
2247 VFIOBAR *bar = &vdev->bars[nr];
2248 unsigned size = bar->size;
2249 char name[64];
2250 uint32_t pci_bar;
2251 uint8_t type;
2252 int ret;
2253
2254 /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
2255 if (!size) {
2256 return;
2257 }
2258
2259 snprintf(name, sizeof(name), "VFIO %04x:%02x:%02x.%x BAR %d",
2260 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2261 vdev->host.function, nr);
2262
2263 /* Determine what type of BAR this is for registration */
2264 ret = pread(vdev->fd, &pci_bar, sizeof(pci_bar),
2265 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
2266 if (ret != sizeof(pci_bar)) {
312fd5f2 2267 error_report("vfio: Failed to read BAR %d (%m)", nr);
65501a74
AW
2268 return;
2269 }
2270
2271 pci_bar = le32_to_cpu(pci_bar);
2272 type = pci_bar & (pci_bar & PCI_BASE_ADDRESS_SPACE_IO ?
2273 ~PCI_BASE_ADDRESS_IO_MASK : ~PCI_BASE_ADDRESS_MEM_MASK);
2274
2275 /* A "slow" read/write mapping underlies all BARs */
2c9b15ca 2276 memory_region_init_io(&bar->mem, NULL, &vfio_bar_ops, bar, name, size);
65501a74
AW
2277 pci_register_bar(&vdev->pdev, nr, type, &bar->mem);
2278
2279 /*
2280 * We can't mmap areas overlapping the MSIX vector table, so we
2281 * potentially insert a direct-mapped subregion before and after it.
2282 */
2283 if (vdev->msix && vdev->msix->table_bar == nr) {
2284 size = vdev->msix->table_offset & TARGET_PAGE_MASK;
2285 }
2286
2287 strncat(name, " mmap", sizeof(name) - strlen(name) - 1);
2288 if (vfio_mmap_bar(bar, &bar->mem,
2289 &bar->mmap_mem, &bar->mmap, size, 0, name)) {
312fd5f2 2290 error_report("%s unsupported. Performance may be slow", name);
65501a74
AW
2291 }
2292
2293 if (vdev->msix && vdev->msix->table_bar == nr) {
2294 unsigned start;
2295
2296 start = TARGET_PAGE_ALIGN(vdev->msix->table_offset +
2297 (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
2298
2299 size = start < bar->size ? bar->size - start : 0;
2300 strncat(name, " msix-hi", sizeof(name) - strlen(name) - 1);
2301 /* VFIOMSIXInfo contains another MemoryRegion for this mapping */
2302 if (vfio_mmap_bar(bar, &bar->mem, &vdev->msix->mmap_mem,
2303 &vdev->msix->mmap, size, start, name)) {
312fd5f2 2304 error_report("%s unsupported. Performance may be slow", name);
65501a74
AW
2305 }
2306 }
7076eabc
AW
2307
2308 vfio_bar_quirk_setup(vdev, nr);
65501a74
AW
2309}
2310
2311static void vfio_map_bars(VFIODevice *vdev)
2312{
2313 int i;
2314
2315 for (i = 0; i < PCI_ROM_SLOT; i++) {
2316 vfio_map_bar(vdev, i);
2317 }
f15689c7
AW
2318
2319 if (vdev->has_vga) {
2c9b15ca 2320 memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_MEM].mem, NULL,
f15689c7
AW
2321 &vfio_vga_ops,
2322 &vdev->vga.region[QEMU_PCI_VGA_MEM],
2323 "vfio-vga-mmio@0xa0000",
2324 QEMU_PCI_VGA_MEM_SIZE);
2c9b15ca 2325 memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem, NULL,
f15689c7
AW
2326 &vfio_vga_ops,
2327 &vdev->vga.region[QEMU_PCI_VGA_IO_LO],
2328 "vfio-vga-io@0x3b0",
2329 QEMU_PCI_VGA_IO_LO_SIZE);
2c9b15ca 2330 memory_region_init_io(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem, NULL,
f15689c7
AW
2331 &vfio_vga_ops,
2332 &vdev->vga.region[QEMU_PCI_VGA_IO_HI],
2333 "vfio-vga-io@0x3c0",
2334 QEMU_PCI_VGA_IO_HI_SIZE);
2335
2336 pci_register_vga(&vdev->pdev, &vdev->vga.region[QEMU_PCI_VGA_MEM].mem,
2337 &vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem,
2338 &vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem);
7076eabc 2339 vfio_vga_quirk_setup(vdev);
f15689c7 2340 }
65501a74
AW
2341}
2342
2343static void vfio_unmap_bars(VFIODevice *vdev)
2344{
2345 int i;
2346
2347 for (i = 0; i < PCI_ROM_SLOT; i++) {
2348 vfio_unmap_bar(vdev, i);
2349 }
f15689c7
AW
2350
2351 if (vdev->has_vga) {
7076eabc 2352 vfio_vga_quirk_teardown(vdev);
f15689c7
AW
2353 pci_unregister_vga(&vdev->pdev);
2354 memory_region_destroy(&vdev->vga.region[QEMU_PCI_VGA_MEM].mem);
2355 memory_region_destroy(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].mem);
2356 memory_region_destroy(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].mem);
2357 }
65501a74
AW
2358}
2359
2360/*
2361 * General setup
2362 */
2363static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
2364{
2365 uint8_t tmp, next = 0xff;
2366
2367 for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
2368 tmp = pdev->config[tmp + 1]) {
2369 if (tmp > pos && tmp < next) {
2370 next = tmp;
2371 }
2372 }
2373
2374 return next - pos;
2375}
2376
96adc5c7
AW
2377static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask)
2378{
2379 pci_set_word(buf, (pci_get_word(buf) & ~mask) | val);
2380}
2381
2382static void vfio_add_emulated_word(VFIODevice *vdev, int pos,
2383 uint16_t val, uint16_t mask)
2384{
2385 vfio_set_word_bits(vdev->pdev.config + pos, val, mask);
2386 vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask);
2387 vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask);
2388}
2389
2390static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask)
2391{
2392 pci_set_long(buf, (pci_get_long(buf) & ~mask) | val);
2393}
2394
2395static void vfio_add_emulated_long(VFIODevice *vdev, int pos,
2396 uint32_t val, uint32_t mask)
2397{
2398 vfio_set_long_bits(vdev->pdev.config + pos, val, mask);
2399 vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask);
2400 vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask);
2401}
2402
2403static int vfio_setup_pcie_cap(VFIODevice *vdev, int pos, uint8_t size)
2404{
2405 uint16_t flags;
2406 uint8_t type;
2407
2408 flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS);
2409 type = (flags & PCI_EXP_FLAGS_TYPE) >> 4;
2410
2411 if (type != PCI_EXP_TYPE_ENDPOINT &&
2412 type != PCI_EXP_TYPE_LEG_END &&
2413 type != PCI_EXP_TYPE_RC_END) {
2414
2415 error_report("vfio: Assignment of PCIe type 0x%x "
2416 "devices is not currently supported", type);
2417 return -EINVAL;
2418 }
2419
2420 if (!pci_bus_is_express(vdev->pdev.bus)) {
2421 /*
2422 * Use express capability as-is on PCI bus. It doesn't make much
2423 * sense to even expose, but some drivers (ex. tg3) depend on it
2424 * and guests don't seem to be particular about it. We'll need
2425 * to revist this or force express devices to express buses if we
2426 * ever expose an IOMMU to the guest.
2427 */
2428 } else if (pci_bus_is_root(vdev->pdev.bus)) {
2429 /*
2430 * On a Root Complex bus Endpoints become Root Complex Integrated
2431 * Endpoints, which changes the type and clears the LNK & LNK2 fields.
2432 */
2433 if (type == PCI_EXP_TYPE_ENDPOINT) {
2434 vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
2435 PCI_EXP_TYPE_RC_END << 4,
2436 PCI_EXP_FLAGS_TYPE);
2437
2438 /* Link Capabilities, Status, and Control goes away */
2439 if (size > PCI_EXP_LNKCTL) {
2440 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0);
2441 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
2442 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0);
2443
2444#ifndef PCI_EXP_LNKCAP2
2445#define PCI_EXP_LNKCAP2 44
2446#endif
2447#ifndef PCI_EXP_LNKSTA2
2448#define PCI_EXP_LNKSTA2 50
2449#endif
2450 /* Link 2 Capabilities, Status, and Control goes away */
2451 if (size > PCI_EXP_LNKCAP2) {
2452 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0);
2453 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0);
2454 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0);
2455 }
2456 }
2457
2458 } else if (type == PCI_EXP_TYPE_LEG_END) {
2459 /*
2460 * Legacy endpoints don't belong on the root complex. Windows
2461 * seems to be happier with devices if we skip the capability.
2462 */
2463 return 0;
2464 }
2465
2466 } else {
2467 /*
2468 * Convert Root Complex Integrated Endpoints to regular endpoints.
2469 * These devices don't support LNK/LNK2 capabilities, so make them up.
2470 */
2471 if (type == PCI_EXP_TYPE_RC_END) {
2472 vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS,
2473 PCI_EXP_TYPE_ENDPOINT << 4,
2474 PCI_EXP_FLAGS_TYPE);
2475 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP,
2476 PCI_EXP_LNK_MLW_1 | PCI_EXP_LNK_LS_25, ~0);
2477 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0);
2478 }
2479
2480 /* Mark the Link Status bits as emulated to allow virtual negotiation */
2481 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA,
2482 pci_get_word(vdev->pdev.config + pos +
2483 PCI_EXP_LNKSTA),
2484 PCI_EXP_LNKCAP_MLW | PCI_EXP_LNKCAP_SLS);
2485 }
2486
2487 pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size);
2488 if (pos >= 0) {
2489 vdev->pdev.exp.exp_cap = pos;
2490 }
2491
2492 return pos;
2493}
2494
65501a74
AW
2495static int vfio_add_std_cap(VFIODevice *vdev, uint8_t pos)
2496{
2497 PCIDevice *pdev = &vdev->pdev;
2498 uint8_t cap_id, next, size;
2499 int ret;
2500
2501 cap_id = pdev->config[pos];
2502 next = pdev->config[pos + 1];
2503
2504 /*
2505 * If it becomes important to configure capabilities to their actual
2506 * size, use this as the default when it's something we don't recognize.
2507 * Since QEMU doesn't actually handle many of the config accesses,
2508 * exact size doesn't seem worthwhile.
2509 */
2510 size = vfio_std_cap_max_size(pdev, pos);
2511
2512 /*
2513 * pci_add_capability always inserts the new capability at the head
2514 * of the chain. Therefore to end up with a chain that matches the
2515 * physical device, we insert from the end by making this recursive.
2516 * This is also why we pre-caclulate size above as cached config space
2517 * will be changed as we unwind the stack.
2518 */
2519 if (next) {
2520 ret = vfio_add_std_cap(vdev, next);
2521 if (ret) {
2522 return ret;
2523 }
2524 } else {
96adc5c7
AW
2525 /* Begin the rebuild, use QEMU emulated list bits */
2526 pdev->config[PCI_CAPABILITY_LIST] = 0;
2527 vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff;
2528 vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST;
65501a74
AW
2529 }
2530
96adc5c7
AW
2531 /* Use emulated next pointer to allow dropping caps */
2532 pci_set_byte(vdev->emulated_config_bits + pos + 1, 0xff);
2533
65501a74
AW
2534 switch (cap_id) {
2535 case PCI_CAP_ID_MSI:
2536 ret = vfio_setup_msi(vdev, pos);
2537 break;
96adc5c7
AW
2538 case PCI_CAP_ID_EXP:
2539 ret = vfio_setup_pcie_cap(vdev, pos, size);
2540 break;
65501a74
AW
2541 case PCI_CAP_ID_MSIX:
2542 ret = vfio_setup_msix(vdev, pos);
2543 break;
ba661818
AW
2544 case PCI_CAP_ID_PM:
2545 vdev->pm_cap = pos;
65501a74
AW
2546 default:
2547 ret = pci_add_capability(pdev, cap_id, pos, size);
2548 break;
2549 }
2550
2551 if (ret < 0) {
2552 error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability "
312fd5f2 2553 "0x%x[0x%x]@0x%x: %d", vdev->host.domain,
65501a74
AW
2554 vdev->host.bus, vdev->host.slot, vdev->host.function,
2555 cap_id, size, pos, ret);
2556 return ret;
2557 }
2558
2559 return 0;
2560}
2561
2562static int vfio_add_capabilities(VFIODevice *vdev)
2563{
2564 PCIDevice *pdev = &vdev->pdev;
2565
2566 if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
2567 !pdev->config[PCI_CAPABILITY_LIST]) {
2568 return 0; /* Nothing to add */
2569 }
2570
2571 return vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST]);
2572}
2573
2574static int vfio_load_rom(VFIODevice *vdev)
2575{
2576 uint64_t size = vdev->rom_size;
2577 char name[32];
2578 off_t off = 0, voff = vdev->rom_offset;
2579 ssize_t bytes;
2580 void *ptr;
2581
2582 /* If loading ROM from file, pci handles it */
2583 if (vdev->pdev.romfile || !vdev->pdev.rom_bar || !size) {
2584 return 0;
2585 }
2586
2587 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
2588 vdev->host.bus, vdev->host.slot, vdev->host.function);
2589
2590 snprintf(name, sizeof(name), "vfio[%04x:%02x:%02x.%x].rom",
2591 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2592 vdev->host.function);
2c9b15ca 2593 memory_region_init_ram(&vdev->pdev.rom, NULL, name, size);
65501a74
AW
2594 ptr = memory_region_get_ram_ptr(&vdev->pdev.rom);
2595 memset(ptr, 0xff, size);
2596
2597 while (size) {
2598 bytes = pread(vdev->fd, ptr + off, size, voff + off);
2599 if (bytes == 0) {
2600 break; /* expect that we could get back less than the ROM BAR */
2601 } else if (bytes > 0) {
2602 off += bytes;
2603 size -= bytes;
2604 } else {
2605 if (errno == EINTR || errno == EAGAIN) {
2606 continue;
2607 }
312fd5f2 2608 error_report("vfio: Error reading device ROM: %m");
65501a74
AW
2609 memory_region_destroy(&vdev->pdev.rom);
2610 return -errno;
2611 }
2612 }
2613
2614 pci_register_bar(&vdev->pdev, PCI_ROM_SLOT, 0, &vdev->pdev.rom);
2615 vdev->pdev.has_rom = true;
2616 return 0;
2617}
2618
2619static int vfio_connect_container(VFIOGroup *group)
2620{
2621 VFIOContainer *container;
2622 int ret, fd;
2623
2624 if (group->container) {
2625 return 0;
2626 }
2627
2628 QLIST_FOREACH(container, &container_list, next) {
2629 if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
2630 group->container = container;
2631 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2632 return 0;
2633 }
2634 }
2635
2636 fd = qemu_open("/dev/vfio/vfio", O_RDWR);
2637 if (fd < 0) {
312fd5f2 2638 error_report("vfio: failed to open /dev/vfio/vfio: %m");
65501a74
AW
2639 return -errno;
2640 }
2641
2642 ret = ioctl(fd, VFIO_GET_API_VERSION);
2643 if (ret != VFIO_API_VERSION) {
2644 error_report("vfio: supported vfio version: %d, "
312fd5f2 2645 "reported version: %d", VFIO_API_VERSION, ret);
65501a74
AW
2646 close(fd);
2647 return -EINVAL;
2648 }
2649
2650 container = g_malloc0(sizeof(*container));
2651 container->fd = fd;
2652
2653 if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) {
2654 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
2655 if (ret) {
312fd5f2 2656 error_report("vfio: failed to set group container: %m");
65501a74
AW
2657 g_free(container);
2658 close(fd);
2659 return -errno;
2660 }
2661
2662 ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU);
2663 if (ret) {
312fd5f2 2664 error_report("vfio: failed to set iommu for container: %m");
65501a74
AW
2665 g_free(container);
2666 close(fd);
2667 return -errno;
2668 }
2669
2670 container->iommu_data.listener = vfio_memory_listener;
2671 container->iommu_data.release = vfio_listener_release;
2672
f6790af6 2673 memory_listener_register(&container->iommu_data.listener, &address_space_memory);
65501a74 2674 } else {
312fd5f2 2675 error_report("vfio: No available IOMMU models");
65501a74
AW
2676 g_free(container);
2677 close(fd);
2678 return -EINVAL;
2679 }
2680
2681 QLIST_INIT(&container->group_list);
2682 QLIST_INSERT_HEAD(&container_list, container, next);
2683
2684 group->container = container;
2685 QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2686
2687 return 0;
2688}
2689
2690static void vfio_disconnect_container(VFIOGroup *group)
2691{
2692 VFIOContainer *container = group->container;
2693
2694 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
312fd5f2 2695 error_report("vfio: error disconnecting group %d from container",
65501a74
AW
2696 group->groupid);
2697 }
2698
2699 QLIST_REMOVE(group, container_next);
2700 group->container = NULL;
2701
2702 if (QLIST_EMPTY(&container->group_list)) {
2703 if (container->iommu_data.release) {
2704 container->iommu_data.release(container);
2705 }
2706 QLIST_REMOVE(container, next);
2707 DPRINTF("vfio_disconnect_container: close container->fd\n");
2708 close(container->fd);
2709 g_free(container);
2710 }
2711}
2712
2713static VFIOGroup *vfio_get_group(int groupid)
2714{
2715 VFIOGroup *group;
2716 char path[32];
2717 struct vfio_group_status status = { .argsz = sizeof(status) };
2718
2719 QLIST_FOREACH(group, &group_list, next) {
2720 if (group->groupid == groupid) {
2721 return group;
2722 }
2723 }
2724
2725 group = g_malloc0(sizeof(*group));
2726
2727 snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
2728 group->fd = qemu_open(path, O_RDWR);
2729 if (group->fd < 0) {
312fd5f2 2730 error_report("vfio: error opening %s: %m", path);
65501a74
AW
2731 g_free(group);
2732 return NULL;
2733 }
2734
2735 if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
312fd5f2 2736 error_report("vfio: error getting group status: %m");
65501a74
AW
2737 close(group->fd);
2738 g_free(group);
2739 return NULL;
2740 }
2741
2742 if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
2743 error_report("vfio: error, group %d is not viable, please ensure "
2744 "all devices within the iommu_group are bound to their "
312fd5f2 2745 "vfio bus driver.", groupid);
65501a74
AW
2746 close(group->fd);
2747 g_free(group);
2748 return NULL;
2749 }
2750
2751 group->groupid = groupid;
2752 QLIST_INIT(&group->device_list);
2753
2754 if (vfio_connect_container(group)) {
312fd5f2 2755 error_report("vfio: failed to setup container for group %d", groupid);
65501a74
AW
2756 close(group->fd);
2757 g_free(group);
2758 return NULL;
2759 }
2760
2761 QLIST_INSERT_HEAD(&group_list, group, next);
2762
2763 return group;
2764}
2765
2766static void vfio_put_group(VFIOGroup *group)
2767{
2768 if (!QLIST_EMPTY(&group->device_list)) {
2769 return;
2770 }
2771
2772 vfio_disconnect_container(group);
2773 QLIST_REMOVE(group, next);
2774 DPRINTF("vfio_put_group: close group->fd\n");
2775 close(group->fd);
2776 g_free(group);
2777}
2778
2779static int vfio_get_device(VFIOGroup *group, const char *name, VFIODevice *vdev)
2780{
2781 struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
2782 struct vfio_region_info reg_info = { .argsz = sizeof(reg_info) };
2783 int ret, i;
2784
2785 ret = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
2786 if (ret < 0) {
1a9522cc 2787 error_report("vfio: error getting device %s from group %d: %m",
65501a74 2788 name, group->groupid);
1a9522cc 2789 error_printf("Verify all devices in group %d are bound to vfio-pci "
65501a74
AW
2790 "or pci-stub and not already in use\n", group->groupid);
2791 return ret;
2792 }
2793
2794 vdev->fd = ret;
2795 vdev->group = group;
2796 QLIST_INSERT_HEAD(&group->device_list, vdev, next);
2797
2798 /* Sanity check device */
2799 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_INFO, &dev_info);
2800 if (ret) {
312fd5f2 2801 error_report("vfio: error getting device info: %m");
65501a74
AW
2802 goto error;
2803 }
2804
2805 DPRINTF("Device %s flags: %u, regions: %u, irgs: %u\n", name,
2806 dev_info.flags, dev_info.num_regions, dev_info.num_irqs);
2807
2808 if (!(dev_info.flags & VFIO_DEVICE_FLAGS_PCI)) {
312fd5f2 2809 error_report("vfio: Um, this isn't a PCI device");
65501a74
AW
2810 goto error;
2811 }
2812
2813 vdev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
2814 if (!vdev->reset_works) {
312fd5f2 2815 error_report("Warning, device %s does not support reset", name);
65501a74
AW
2816 }
2817
8fc94e5a 2818 if (dev_info.num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) {
312fd5f2 2819 error_report("vfio: unexpected number of io regions %u",
65501a74
AW
2820 dev_info.num_regions);
2821 goto error;
2822 }
2823
8fc94e5a 2824 if (dev_info.num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) {
312fd5f2 2825 error_report("vfio: unexpected number of irqs %u", dev_info.num_irqs);
65501a74
AW
2826 goto error;
2827 }
2828
2829 for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
2830 reg_info.index = i;
2831
2832 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
2833 if (ret) {
312fd5f2 2834 error_report("vfio: Error getting region %d info: %m", i);
65501a74
AW
2835 goto error;
2836 }
2837
2838 DPRINTF("Device %s region %d:\n", name, i);
2839 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
2840 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
2841 (unsigned long)reg_info.flags);
2842
2843 vdev->bars[i].flags = reg_info.flags;
2844 vdev->bars[i].size = reg_info.size;
2845 vdev->bars[i].fd_offset = reg_info.offset;
2846 vdev->bars[i].fd = vdev->fd;
2847 vdev->bars[i].nr = i;
7076eabc 2848 QLIST_INIT(&vdev->bars[i].quirks);
65501a74
AW
2849 }
2850
2851 reg_info.index = VFIO_PCI_ROM_REGION_INDEX;
2852
2853 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
2854 if (ret) {
312fd5f2 2855 error_report("vfio: Error getting ROM info: %m");
65501a74
AW
2856 goto error;
2857 }
2858
2859 DPRINTF("Device %s ROM:\n", name);
2860 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
2861 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
2862 (unsigned long)reg_info.flags);
2863
2864 vdev->rom_size = reg_info.size;
2865 vdev->rom_offset = reg_info.offset;
2866
2867 reg_info.index = VFIO_PCI_CONFIG_REGION_INDEX;
2868
2869 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
2870 if (ret) {
312fd5f2 2871 error_report("vfio: Error getting config info: %m");
65501a74
AW
2872 goto error;
2873 }
2874
2875 DPRINTF("Device %s config:\n", name);
2876 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
2877 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
2878 (unsigned long)reg_info.flags);
2879
2880 vdev->config_size = reg_info.size;
6a659bbf
AW
2881 if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) {
2882 vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS;
2883 }
65501a74
AW
2884 vdev->config_offset = reg_info.offset;
2885
f15689c7
AW
2886 if ((vdev->features & VFIO_FEATURE_ENABLE_VGA) &&
2887 dev_info.num_regions > VFIO_PCI_VGA_REGION_INDEX) {
2888 struct vfio_region_info vga_info = {
2889 .argsz = sizeof(vga_info),
2890 .index = VFIO_PCI_VGA_REGION_INDEX,
2891 };
2892
2893 ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &vga_info);
2894 if (ret) {
2895 error_report(
2896 "vfio: Device does not support requested feature x-vga");
2897 goto error;
2898 }
2899
2900 if (!(vga_info.flags & VFIO_REGION_INFO_FLAG_READ) ||
2901 !(vga_info.flags & VFIO_REGION_INFO_FLAG_WRITE) ||
2902 vga_info.size < 0xbffff + 1) {
2903 error_report("vfio: Unexpected VGA info, flags 0x%lx, size 0x%lx",
2904 (unsigned long)vga_info.flags,
2905 (unsigned long)vga_info.size);
2906 goto error;
2907 }
2908
2909 vdev->vga.fd_offset = vga_info.offset;
2910 vdev->vga.fd = vdev->fd;
2911
2912 vdev->vga.region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE;
2913 vdev->vga.region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM;
7076eabc 2914 QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_MEM].quirks);
f15689c7
AW
2915
2916 vdev->vga.region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE;
2917 vdev->vga.region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO;
7076eabc 2918 QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_IO_LO].quirks);
f15689c7
AW
2919
2920 vdev->vga.region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE;
2921 vdev->vga.region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI;
7076eabc 2922 QLIST_INIT(&vdev->vga.region[QEMU_PCI_VGA_IO_HI].quirks);
f15689c7
AW
2923
2924 vdev->has_vga = true;
2925 }
2926
65501a74
AW
2927error:
2928 if (ret) {
2929 QLIST_REMOVE(vdev, next);
2930 vdev->group = NULL;
2931 close(vdev->fd);
2932 }
2933 return ret;
2934}
2935
2936static void vfio_put_device(VFIODevice *vdev)
2937{
2938 QLIST_REMOVE(vdev, next);
2939 vdev->group = NULL;
2940 DPRINTF("vfio_put_device: close vdev->fd\n");
2941 close(vdev->fd);
2942 if (vdev->msix) {
2943 g_free(vdev->msix);
2944 vdev->msix = NULL;
2945 }
2946}
2947
2948static int vfio_initfn(PCIDevice *pdev)
2949{
2950 VFIODevice *pvdev, *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
2951 VFIOGroup *group;
2952 char path[PATH_MAX], iommu_group_path[PATH_MAX], *group_name;
2953 ssize_t len;
2954 struct stat st;
2955 int groupid;
2956 int ret;
2957
2958 /* Check that the host device exists */
2959 snprintf(path, sizeof(path),
2960 "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/",
2961 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2962 vdev->host.function);
2963 if (stat(path, &st) < 0) {
312fd5f2 2964 error_report("vfio: error: no such host device: %s", path);
65501a74
AW
2965 return -errno;
2966 }
2967
2968 strncat(path, "iommu_group", sizeof(path) - strlen(path) - 1);
2969
2970 len = readlink(path, iommu_group_path, PATH_MAX);
2971 if (len <= 0) {
312fd5f2 2972 error_report("vfio: error no iommu_group for device");
65501a74
AW
2973 return -errno;
2974 }
2975
2976 iommu_group_path[len] = 0;
2977 group_name = basename(iommu_group_path);
2978
2979 if (sscanf(group_name, "%d", &groupid) != 1) {
312fd5f2 2980 error_report("vfio: error reading %s: %m", path);
65501a74
AW
2981 return -errno;
2982 }
2983
2984 DPRINTF("%s(%04x:%02x:%02x.%x) group %d\n", __func__, vdev->host.domain,
2985 vdev->host.bus, vdev->host.slot, vdev->host.function, groupid);
2986
2987 group = vfio_get_group(groupid);
2988 if (!group) {
312fd5f2 2989 error_report("vfio: failed to get group %d", groupid);
65501a74
AW
2990 return -ENOENT;
2991 }
2992
2993 snprintf(path, sizeof(path), "%04x:%02x:%02x.%01x",
2994 vdev->host.domain, vdev->host.bus, vdev->host.slot,
2995 vdev->host.function);
2996
2997 QLIST_FOREACH(pvdev, &group->device_list, next) {
2998 if (pvdev->host.domain == vdev->host.domain &&
2999 pvdev->host.bus == vdev->host.bus &&
3000 pvdev->host.slot == vdev->host.slot &&
3001 pvdev->host.function == vdev->host.function) {
3002
312fd5f2 3003 error_report("vfio: error: device %s is already attached", path);
65501a74
AW
3004 vfio_put_group(group);
3005 return -EBUSY;
3006 }
3007 }
3008
3009 ret = vfio_get_device(group, path, vdev);
3010 if (ret) {
312fd5f2 3011 error_report("vfio: failed to get device %s", path);
65501a74
AW
3012 vfio_put_group(group);
3013 return ret;
3014 }
3015
3016 /* Get a copy of config space */
3017 ret = pread(vdev->fd, vdev->pdev.config,
3018 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
3019 vdev->config_offset);
3020 if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
3021 ret = ret < 0 ? -errno : -EFAULT;
312fd5f2 3022 error_report("vfio: Failed to read device config space");
65501a74
AW
3023 goto out_put;
3024 }
3025
4b5d5e87
AW
3026 /* vfio emulates a lot for us, but some bits need extra love */
3027 vdev->emulated_config_bits = g_malloc0(vdev->config_size);
3028
3029 /* QEMU can choose to expose the ROM or not */
3030 memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4);
3031
3032 /* QEMU can change multi-function devices to single function, or reverse */
3033 vdev->emulated_config_bits[PCI_HEADER_TYPE] =
3034 PCI_HEADER_TYPE_MULTI_FUNCTION;
3035
65501a74
AW
3036 /*
3037 * Clear host resource mapping info. If we choose not to register a
3038 * BAR, such as might be the case with the option ROM, we can get
3039 * confusing, unwritable, residual addresses from the host here.
3040 */
3041 memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
3042 memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
3043
3044 vfio_load_rom(vdev);
3045
3046 ret = vfio_early_setup_msix(vdev);
3047 if (ret) {
3048 goto out_put;
3049 }
3050
3051 vfio_map_bars(vdev);
3052
3053 ret = vfio_add_capabilities(vdev);
3054 if (ret) {
3055 goto out_teardown;
3056 }
3057
4b5d5e87
AW
3058 /* QEMU emulates all of MSI & MSIX */
3059 if (pdev->cap_present & QEMU_PCI_CAP_MSIX) {
3060 memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff,
3061 MSIX_CAP_LENGTH);
3062 }
3063
3064 if (pdev->cap_present & QEMU_PCI_CAP_MSI) {
3065 memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff,
3066 vdev->msi_cap_size);
3067 }
3068
65501a74 3069 if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
ea486926
AW
3070 vdev->intx.mmap_timer = qemu_new_timer_ms(vm_clock,
3071 vfio_intx_mmap_enable, vdev);
e1d1e586 3072 pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_update_irq);
65501a74
AW
3073 ret = vfio_enable_intx(vdev);
3074 if (ret) {
3075 goto out_teardown;
3076 }
3077 }
3078
c29029dd
AW
3079 add_boot_device_path(vdev->bootindex, &pdev->qdev, NULL);
3080
65501a74
AW
3081 return 0;
3082
3083out_teardown:
3084 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3085 vfio_teardown_msi(vdev);
3086 vfio_unmap_bars(vdev);
3087out_put:
4b5d5e87 3088 g_free(vdev->emulated_config_bits);
65501a74
AW
3089 vfio_put_device(vdev);
3090 vfio_put_group(group);
3091 return ret;
3092}
3093
3094static void vfio_exitfn(PCIDevice *pdev)
3095{
3096 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
3097 VFIOGroup *group = vdev->group;
3098
3099 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
3100 vfio_disable_interrupts(vdev);
ea486926
AW
3101 if (vdev->intx.mmap_timer) {
3102 qemu_free_timer(vdev->intx.mmap_timer);
3103 }
65501a74
AW
3104 vfio_teardown_msi(vdev);
3105 vfio_unmap_bars(vdev);
4b5d5e87 3106 g_free(vdev->emulated_config_bits);
65501a74
AW
3107 vfio_put_device(vdev);
3108 vfio_put_group(group);
3109}
3110
3111static void vfio_pci_reset(DeviceState *dev)
3112{
3113 PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
3114 VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
5834a83f 3115 uint16_t cmd;
65501a74 3116
5834a83f
AW
3117 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
3118 vdev->host.bus, vdev->host.slot, vdev->host.function);
3119
3120 vfio_disable_interrupts(vdev);
65501a74 3121
ba661818
AW
3122 /* Make sure the device is in D0 */
3123 if (vdev->pm_cap) {
3124 uint16_t pmcsr;
3125 uint8_t state;
3126
3127 pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
3128 state = pmcsr & PCI_PM_CTRL_STATE_MASK;
3129 if (state) {
3130 pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
3131 vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2);
3132 /* vfio handles the necessary delay here */
3133 pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2);
3134 state = pmcsr & PCI_PM_CTRL_STATE_MASK;
3135 if (state) {
3136 error_report("vfio: Unable to power on device, stuck in D%d\n",
3137 state);
3138 }
3139 }
3140 }
3141
5834a83f
AW
3142 /*
3143 * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
3144 * Also put INTx Disable in known state.
3145 */
3146 cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
3147 cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
3148 PCI_COMMAND_INTX_DISABLE);
3149 vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
3150
3151 if (vdev->reset_works) {
3152 if (ioctl(vdev->fd, VFIO_DEVICE_RESET)) {
3153 error_report("vfio: Error unable to reset physical device "
312fd5f2 3154 "(%04x:%02x:%02x.%x): %m", vdev->host.domain,
5834a83f
AW
3155 vdev->host.bus, vdev->host.slot, vdev->host.function);
3156 }
65501a74 3157 }
5834a83f
AW
3158
3159 vfio_enable_intx(vdev);
65501a74
AW
3160}
3161
3162static Property vfio_pci_dev_properties[] = {
3163 DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIODevice, host),
ea486926
AW
3164 DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIODevice,
3165 intx.mmap_timeout, 1100),
f15689c7
AW
3166 DEFINE_PROP_BIT("x-vga", VFIODevice, features,
3167 VFIO_FEATURE_ENABLE_VGA_BIT, false),
c29029dd 3168 DEFINE_PROP_INT32("bootindex", VFIODevice, bootindex, -1),
65501a74
AW
3169 /*
3170 * TODO - support passed fds... is this necessary?
3171 * DEFINE_PROP_STRING("vfiofd", VFIODevice, vfiofd_name),
3172 * DEFINE_PROP_STRING("vfiogroupfd, VFIODevice, vfiogroupfd_name),
3173 */
3174 DEFINE_PROP_END_OF_LIST(),
3175};
3176
d9f0e638
AW
3177static const VMStateDescription vfio_pci_vmstate = {
3178 .name = "vfio-pci",
3179 .unmigratable = 1,
3180};
65501a74
AW
3181
3182static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
3183{
3184 DeviceClass *dc = DEVICE_CLASS(klass);
3185 PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
3186
3187 dc->reset = vfio_pci_reset;
3188 dc->props = vfio_pci_dev_properties;
d9f0e638
AW
3189 dc->vmsd = &vfio_pci_vmstate;
3190 dc->desc = "VFIO-based PCI device assignment";
65501a74
AW
3191 pdc->init = vfio_initfn;
3192 pdc->exit = vfio_exitfn;
3193 pdc->config_read = vfio_pci_read_config;
3194 pdc->config_write = vfio_pci_write_config;
6a659bbf 3195 pdc->is_express = 1; /* We might be */
65501a74
AW
3196}
3197
3198static const TypeInfo vfio_pci_dev_info = {
3199 .name = "vfio-pci",
3200 .parent = TYPE_PCI_DEVICE,
3201 .instance_size = sizeof(VFIODevice),
3202 .class_init = vfio_pci_dev_class_init,
3203};
3204
3205static void register_vfio_pci_dev_type(void)
3206{
3207 type_register_static(&vfio_pci_dev_info);
3208}
3209
3210type_init(register_vfio_pci_dev_type)