]>
Commit | Line | Data |
---|---|---|
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> | |
22 | #include <unistd.h> | |
23 | #include <sys/ioctl.h> | |
24 | #include <sys/mman.h> | |
25 | #include <sys/stat.h> | |
26 | #include <sys/types.h> | |
27 | #include <linux/vfio.h> | |
28 | ||
29 | #include "config.h" | |
30 | #include "event_notifier.h" | |
31 | #include "exec-memory.h" | |
32 | #include "kvm.h" | |
33 | #include "memory.h" | |
34 | #include "msi.h" | |
35 | #include "msix.h" | |
36 | #include "qemu-error.h" | |
37 | #include "range.h" | |
38 | #include "vfio_pci_int.h" | |
39 | ||
40 | /* #define DEBUG_VFIO */ | |
41 | #ifdef DEBUG_VFIO | |
42 | #define DPRINTF(fmt, ...) \ | |
43 | do { fprintf(stderr, "vfio: " fmt, ## __VA_ARGS__); } while (0) | |
44 | #else | |
45 | #define DPRINTF(fmt, ...) \ | |
46 | do { } while (0) | |
47 | #endif | |
48 | ||
49 | #define MSIX_CAP_LENGTH 12 | |
50 | ||
51 | static QLIST_HEAD(, VFIOContainer) | |
52 | container_list = QLIST_HEAD_INITIALIZER(container_list); | |
53 | ||
54 | static QLIST_HEAD(, VFIOGroup) | |
55 | group_list = QLIST_HEAD_INITIALIZER(group_list); | |
56 | ||
57 | static void vfio_disable_interrupts(VFIODevice *vdev); | |
58 | static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len); | |
59 | static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled); | |
60 | ||
61 | /* | |
62 | * Common VFIO interrupt disable | |
63 | */ | |
64 | static void vfio_disable_irqindex(VFIODevice *vdev, int index) | |
65 | { | |
66 | struct vfio_irq_set irq_set = { | |
67 | .argsz = sizeof(irq_set), | |
68 | .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER, | |
69 | .index = index, | |
70 | .start = 0, | |
71 | .count = 0, | |
72 | }; | |
73 | ||
74 | ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set); | |
75 | ||
76 | vdev->interrupt = VFIO_INT_NONE; | |
77 | } | |
78 | ||
79 | /* | |
80 | * INTx | |
81 | */ | |
82 | static void vfio_unmask_intx(VFIODevice *vdev) | |
83 | { | |
84 | struct vfio_irq_set irq_set = { | |
85 | .argsz = sizeof(irq_set), | |
86 | .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK, | |
87 | .index = VFIO_PCI_INTX_IRQ_INDEX, | |
88 | .start = 0, | |
89 | .count = 1, | |
90 | }; | |
91 | ||
92 | ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set); | |
93 | } | |
94 | ||
ea486926 AW |
95 | /* |
96 | * Disabling BAR mmaping can be slow, but toggling it around INTx can | |
97 | * also be a huge overhead. We try to get the best of both worlds by | |
98 | * waiting until an interrupt to disable mmaps (subsequent transitions | |
99 | * to the same state are effectively no overhead). If the interrupt has | |
100 | * been serviced and the time gap is long enough, we re-enable mmaps for | |
101 | * performance. This works well for things like graphics cards, which | |
102 | * may not use their interrupt at all and are penalized to an unusable | |
103 | * level by read/write BAR traps. Other devices, like NICs, have more | |
104 | * regular interrupts and see much better latency by staying in non-mmap | |
105 | * mode. We therefore set the default mmap_timeout such that a ping | |
106 | * is just enough to keep the mmap disabled. Users can experiment with | |
107 | * other options with the x-intx-mmap-timeout-ms parameter (a value of | |
108 | * zero disables the timer). | |
109 | */ | |
110 | static void vfio_intx_mmap_enable(void *opaque) | |
111 | { | |
112 | VFIODevice *vdev = opaque; | |
113 | ||
114 | if (vdev->intx.pending) { | |
115 | qemu_mod_timer(vdev->intx.mmap_timer, | |
116 | qemu_get_clock_ms(vm_clock) + vdev->intx.mmap_timeout); | |
117 | return; | |
118 | } | |
119 | ||
120 | vfio_mmap_set_enabled(vdev, true); | |
121 | } | |
122 | ||
65501a74 AW |
123 | static void vfio_intx_interrupt(void *opaque) |
124 | { | |
125 | VFIODevice *vdev = opaque; | |
126 | ||
127 | if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) { | |
128 | return; | |
129 | } | |
130 | ||
131 | DPRINTF("%s(%04x:%02x:%02x.%x) Pin %c\n", __func__, vdev->host.domain, | |
132 | vdev->host.bus, vdev->host.slot, vdev->host.function, | |
133 | 'A' + vdev->intx.pin); | |
134 | ||
135 | vdev->intx.pending = true; | |
136 | qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 1); | |
ea486926 AW |
137 | vfio_mmap_set_enabled(vdev, false); |
138 | if (vdev->intx.mmap_timeout) { | |
139 | qemu_mod_timer(vdev->intx.mmap_timer, | |
140 | qemu_get_clock_ms(vm_clock) + vdev->intx.mmap_timeout); | |
141 | } | |
65501a74 AW |
142 | } |
143 | ||
144 | static void vfio_eoi(VFIODevice *vdev) | |
145 | { | |
146 | if (!vdev->intx.pending) { | |
147 | return; | |
148 | } | |
149 | ||
150 | DPRINTF("%s(%04x:%02x:%02x.%x) EOI\n", __func__, vdev->host.domain, | |
151 | vdev->host.bus, vdev->host.slot, vdev->host.function); | |
152 | ||
153 | vdev->intx.pending = false; | |
154 | qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0); | |
155 | vfio_unmask_intx(vdev); | |
156 | } | |
157 | ||
158 | typedef struct QEMU_PACKED VFIOIRQSetFD { | |
159 | struct vfio_irq_set irq_set; | |
160 | int32_t fd; | |
161 | } VFIOIRQSetFD; | |
162 | ||
163 | static int vfio_enable_intx(VFIODevice *vdev) | |
164 | { | |
165 | VFIOIRQSetFD irq_set_fd = { | |
166 | .irq_set = { | |
167 | .argsz = sizeof(irq_set_fd), | |
168 | .flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER, | |
169 | .index = VFIO_PCI_INTX_IRQ_INDEX, | |
170 | .start = 0, | |
171 | .count = 1, | |
172 | }, | |
173 | }; | |
174 | uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1); | |
175 | int ret; | |
176 | ||
ea486926 | 177 | if (!pin) { |
65501a74 AW |
178 | return 0; |
179 | } | |
180 | ||
181 | vfio_disable_interrupts(vdev); | |
182 | ||
183 | vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */ | |
184 | ret = event_notifier_init(&vdev->intx.interrupt, 0); | |
185 | if (ret) { | |
186 | error_report("vfio: Error: event_notifier_init failed\n"); | |
187 | return ret; | |
188 | } | |
189 | ||
190 | irq_set_fd.fd = event_notifier_get_fd(&vdev->intx.interrupt); | |
191 | qemu_set_fd_handler(irq_set_fd.fd, vfio_intx_interrupt, NULL, vdev); | |
192 | ||
193 | if (ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set_fd)) { | |
194 | error_report("vfio: Error: Failed to setup INTx fd: %m\n"); | |
195 | return -errno; | |
196 | } | |
197 | ||
65501a74 AW |
198 | vdev->interrupt = VFIO_INT_INTx; |
199 | ||
200 | DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain, | |
201 | vdev->host.bus, vdev->host.slot, vdev->host.function); | |
202 | ||
203 | return 0; | |
204 | } | |
205 | ||
206 | static void vfio_disable_intx(VFIODevice *vdev) | |
207 | { | |
208 | int fd; | |
209 | ||
ea486926 | 210 | qemu_del_timer(vdev->intx.mmap_timer); |
65501a74 AW |
211 | vfio_disable_irqindex(vdev, VFIO_PCI_INTX_IRQ_INDEX); |
212 | vdev->intx.pending = false; | |
213 | qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0); | |
214 | vfio_mmap_set_enabled(vdev, true); | |
215 | ||
216 | fd = event_notifier_get_fd(&vdev->intx.interrupt); | |
217 | qemu_set_fd_handler(fd, NULL, NULL, vdev); | |
218 | event_notifier_cleanup(&vdev->intx.interrupt); | |
219 | ||
220 | vdev->interrupt = VFIO_INT_NONE; | |
221 | ||
222 | DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain, | |
223 | vdev->host.bus, vdev->host.slot, vdev->host.function); | |
224 | } | |
225 | ||
226 | /* | |
227 | * MSI/X | |
228 | */ | |
229 | static void vfio_msi_interrupt(void *opaque) | |
230 | { | |
231 | VFIOMSIVector *vector = opaque; | |
232 | VFIODevice *vdev = vector->vdev; | |
233 | int nr = vector - vdev->msi_vectors; | |
234 | ||
235 | if (!event_notifier_test_and_clear(&vector->interrupt)) { | |
236 | return; | |
237 | } | |
238 | ||
239 | DPRINTF("%s(%04x:%02x:%02x.%x) vector %d\n", __func__, | |
240 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
241 | vdev->host.function, nr); | |
242 | ||
243 | if (vdev->interrupt == VFIO_INT_MSIX) { | |
244 | msix_notify(&vdev->pdev, nr); | |
245 | } else if (vdev->interrupt == VFIO_INT_MSI) { | |
246 | msi_notify(&vdev->pdev, nr); | |
247 | } else { | |
248 | error_report("vfio: MSI interrupt receieved, but not enabled?\n"); | |
249 | } | |
250 | } | |
251 | ||
252 | static int vfio_enable_vectors(VFIODevice *vdev, bool msix) | |
253 | { | |
254 | struct vfio_irq_set *irq_set; | |
255 | int ret = 0, i, argsz; | |
256 | int32_t *fds; | |
257 | ||
258 | argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds)); | |
259 | ||
260 | irq_set = g_malloc0(argsz); | |
261 | irq_set->argsz = argsz; | |
262 | irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER; | |
263 | irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX; | |
264 | irq_set->start = 0; | |
265 | irq_set->count = vdev->nr_vectors; | |
266 | fds = (int32_t *)&irq_set->data; | |
267 | ||
268 | for (i = 0; i < vdev->nr_vectors; i++) { | |
269 | if (!vdev->msi_vectors[i].use) { | |
270 | fds[i] = -1; | |
271 | continue; | |
272 | } | |
273 | ||
274 | fds[i] = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt); | |
275 | } | |
276 | ||
277 | ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set); | |
278 | ||
279 | g_free(irq_set); | |
280 | ||
281 | if (!ret) { | |
282 | vdev->interrupt = msix ? VFIO_INT_MSIX : VFIO_INT_MSI; | |
283 | } | |
284 | ||
285 | return ret; | |
286 | } | |
287 | ||
288 | static int vfio_msix_vector_use(PCIDevice *pdev, | |
289 | unsigned int nr, MSIMessage msg) | |
290 | { | |
291 | VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev); | |
292 | VFIOMSIVector *vector; | |
293 | int ret; | |
294 | ||
295 | DPRINTF("%s(%04x:%02x:%02x.%x) vector %d used\n", __func__, | |
296 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
297 | vdev->host.function, nr); | |
298 | ||
299 | if (vdev->interrupt != VFIO_INT_MSIX) { | |
300 | vfio_disable_interrupts(vdev); | |
301 | } | |
302 | ||
303 | if (!vdev->msi_vectors) { | |
304 | vdev->msi_vectors = g_malloc0(vdev->msix->entries * | |
305 | sizeof(VFIOMSIVector)); | |
306 | } | |
307 | ||
308 | vector = &vdev->msi_vectors[nr]; | |
309 | vector->vdev = vdev; | |
310 | vector->use = true; | |
311 | ||
312 | msix_vector_use(pdev, nr); | |
313 | ||
314 | if (event_notifier_init(&vector->interrupt, 0)) { | |
315 | error_report("vfio: Error: event_notifier_init failed\n"); | |
316 | } | |
317 | ||
318 | /* | |
319 | * Attempt to enable route through KVM irqchip, | |
320 | * default to userspace handling if unavailable. | |
321 | */ | |
322 | vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg); | |
323 | if (vector->virq < 0 || | |
324 | kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt, | |
325 | vector->virq) < 0) { | |
326 | if (vector->virq >= 0) { | |
327 | kvm_irqchip_release_virq(kvm_state, vector->virq); | |
328 | vector->virq = -1; | |
329 | } | |
330 | qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt), | |
331 | vfio_msi_interrupt, NULL, vector); | |
332 | } | |
333 | ||
334 | /* | |
335 | * We don't want to have the host allocate all possible MSI vectors | |
336 | * for a device if they're not in use, so we shutdown and incrementally | |
337 | * increase them as needed. | |
338 | */ | |
339 | if (vdev->nr_vectors < nr + 1) { | |
340 | int i; | |
341 | ||
342 | vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX); | |
343 | vdev->nr_vectors = nr + 1; | |
344 | ret = vfio_enable_vectors(vdev, true); | |
345 | if (ret) { | |
346 | error_report("vfio: failed to enable vectors, %d\n", ret); | |
347 | } | |
348 | ||
349 | /* We don't know if we've missed interrupts in the interim... */ | |
350 | for (i = 0; i < vdev->msix->entries; i++) { | |
351 | if (vdev->msi_vectors[i].use) { | |
352 | msix_notify(&vdev->pdev, i); | |
353 | } | |
354 | } | |
355 | } else { | |
356 | VFIOIRQSetFD irq_set_fd = { | |
357 | .irq_set = { | |
358 | .argsz = sizeof(irq_set_fd), | |
359 | .flags = VFIO_IRQ_SET_DATA_EVENTFD | | |
360 | VFIO_IRQ_SET_ACTION_TRIGGER, | |
361 | .index = VFIO_PCI_MSIX_IRQ_INDEX, | |
362 | .start = nr, | |
363 | .count = 1, | |
364 | }, | |
365 | .fd = event_notifier_get_fd(&vector->interrupt), | |
366 | }; | |
367 | ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set_fd); | |
368 | if (ret) { | |
369 | error_report("vfio: failed to modify vector, %d\n", ret); | |
370 | } | |
371 | ||
372 | /* | |
373 | * If we were connected to the hardware PBA we could skip this, | |
374 | * until then, a spurious interrupt is better than starvation. | |
375 | */ | |
376 | msix_notify(&vdev->pdev, nr); | |
377 | } | |
378 | ||
379 | return 0; | |
380 | } | |
381 | ||
382 | static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr) | |
383 | { | |
384 | VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev); | |
385 | VFIOMSIVector *vector = &vdev->msi_vectors[nr]; | |
386 | VFIOIRQSetFD irq_set_fd = { | |
387 | .irq_set = { | |
388 | .argsz = sizeof(irq_set_fd), | |
389 | .flags = VFIO_IRQ_SET_DATA_EVENTFD | | |
390 | VFIO_IRQ_SET_ACTION_TRIGGER, | |
391 | .index = VFIO_PCI_MSIX_IRQ_INDEX, | |
392 | .start = nr, | |
393 | .count = 1, | |
394 | }, | |
395 | .fd = -1, | |
396 | }; | |
397 | ||
398 | DPRINTF("%s(%04x:%02x:%02x.%x) vector %d released\n", __func__, | |
399 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
400 | vdev->host.function, nr); | |
401 | ||
402 | /* | |
403 | * XXX What's the right thing to do here? This turns off the interrupt | |
404 | * completely, but do we really just want to switch the interrupt to | |
405 | * bouncing through userspace and let msix.c drop it? Not sure. | |
406 | */ | |
407 | msix_vector_unuse(pdev, nr); | |
408 | ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set_fd); | |
409 | ||
410 | if (vector->virq < 0) { | |
411 | qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt), | |
412 | NULL, NULL, NULL); | |
413 | } else { | |
414 | kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt, | |
415 | vector->virq); | |
416 | kvm_irqchip_release_virq(kvm_state, vector->virq); | |
417 | vector->virq = -1; | |
418 | } | |
419 | ||
420 | event_notifier_cleanup(&vector->interrupt); | |
421 | vector->use = false; | |
422 | } | |
423 | ||
424 | /* TODO This should move to msi.c */ | |
425 | static MSIMessage msi_get_msg(PCIDevice *pdev, unsigned int vector) | |
426 | { | |
427 | uint16_t flags = pci_get_word(pdev->config + pdev->msi_cap + PCI_MSI_FLAGS); | |
428 | bool msi64bit = flags & PCI_MSI_FLAGS_64BIT; | |
429 | MSIMessage msg; | |
430 | ||
431 | if (msi64bit) { | |
432 | msg.address = pci_get_quad(pdev->config + | |
433 | pdev->msi_cap + PCI_MSI_ADDRESS_LO); | |
434 | } else { | |
435 | msg.address = pci_get_long(pdev->config + | |
436 | pdev->msi_cap + PCI_MSI_ADDRESS_LO); | |
437 | } | |
438 | ||
439 | msg.data = pci_get_word(pdev->config + pdev->msi_cap + | |
440 | (msi64bit ? PCI_MSI_DATA_64 : PCI_MSI_DATA_32)); | |
441 | msg.data += vector; | |
442 | ||
443 | return msg; | |
444 | } | |
445 | ||
446 | /* So should this */ | |
447 | static void msi_set_qsize(PCIDevice *pdev, uint8_t size) | |
448 | { | |
449 | uint8_t *config = pdev->config + pdev->msi_cap; | |
450 | uint16_t flags; | |
451 | ||
452 | flags = pci_get_word(config + PCI_MSI_FLAGS); | |
453 | flags = le16_to_cpu(flags); | |
454 | flags &= ~PCI_MSI_FLAGS_QSIZE; | |
455 | flags |= (size & 0x7) << 4; | |
456 | flags = cpu_to_le16(flags); | |
457 | pci_set_word(config + PCI_MSI_FLAGS, flags); | |
458 | } | |
459 | ||
460 | static void vfio_enable_msi(VFIODevice *vdev) | |
461 | { | |
462 | int ret, i; | |
463 | ||
464 | vfio_disable_interrupts(vdev); | |
465 | ||
466 | vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev); | |
467 | retry: | |
468 | vdev->msi_vectors = g_malloc0(vdev->nr_vectors * sizeof(VFIOMSIVector)); | |
469 | ||
470 | for (i = 0; i < vdev->nr_vectors; i++) { | |
471 | MSIMessage msg; | |
472 | VFIOMSIVector *vector = &vdev->msi_vectors[i]; | |
473 | ||
474 | vector->vdev = vdev; | |
475 | vector->use = true; | |
476 | ||
477 | if (event_notifier_init(&vector->interrupt, 0)) { | |
478 | error_report("vfio: Error: event_notifier_init failed\n"); | |
479 | } | |
480 | ||
481 | msg = msi_get_msg(&vdev->pdev, i); | |
482 | ||
483 | /* | |
484 | * Attempt to enable route through KVM irqchip, | |
485 | * default to userspace handling if unavailable. | |
486 | */ | |
487 | vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg); | |
488 | if (vector->virq < 0 || | |
489 | kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt, | |
490 | vector->virq) < 0) { | |
491 | qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt), | |
492 | vfio_msi_interrupt, NULL, vector); | |
493 | } | |
494 | } | |
495 | ||
496 | ret = vfio_enable_vectors(vdev, false); | |
497 | if (ret) { | |
498 | if (ret < 0) { | |
499 | error_report("vfio: Error: Failed to setup MSI fds: %m\n"); | |
500 | } else if (ret != vdev->nr_vectors) { | |
501 | error_report("vfio: Error: Failed to enable %d " | |
502 | "MSI vectors, retry with %d\n", vdev->nr_vectors, ret); | |
503 | } | |
504 | ||
505 | for (i = 0; i < vdev->nr_vectors; i++) { | |
506 | VFIOMSIVector *vector = &vdev->msi_vectors[i]; | |
507 | if (vector->virq >= 0) { | |
508 | kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt, | |
509 | vector->virq); | |
510 | kvm_irqchip_release_virq(kvm_state, vector->virq); | |
511 | vector->virq = -1; | |
512 | } else { | |
513 | qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt), | |
514 | NULL, NULL, NULL); | |
515 | } | |
516 | event_notifier_cleanup(&vector->interrupt); | |
517 | } | |
518 | ||
519 | g_free(vdev->msi_vectors); | |
520 | ||
521 | if (ret > 0 && ret != vdev->nr_vectors) { | |
522 | vdev->nr_vectors = ret; | |
523 | goto retry; | |
524 | } | |
525 | vdev->nr_vectors = 0; | |
526 | ||
527 | return; | |
528 | } | |
529 | ||
530 | msi_set_qsize(&vdev->pdev, vdev->nr_vectors); | |
531 | ||
532 | DPRINTF("%s(%04x:%02x:%02x.%x) Enabled %d MSI vectors\n", __func__, | |
533 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
534 | vdev->host.function, vdev->nr_vectors); | |
535 | } | |
536 | ||
537 | static void vfio_disable_msi_x(VFIODevice *vdev, bool msix) | |
538 | { | |
539 | int i; | |
540 | ||
541 | vfio_disable_irqindex(vdev, msix ? VFIO_PCI_MSIX_IRQ_INDEX : | |
542 | VFIO_PCI_MSI_IRQ_INDEX); | |
543 | ||
544 | for (i = 0; i < vdev->nr_vectors; i++) { | |
545 | VFIOMSIVector *vector = &vdev->msi_vectors[i]; | |
546 | ||
547 | if (!vector->use) { | |
548 | continue; | |
549 | } | |
550 | ||
551 | if (vector->virq >= 0) { | |
552 | kvm_irqchip_remove_irqfd_notifier(kvm_state, | |
553 | &vector->interrupt, vector->virq); | |
554 | kvm_irqchip_release_virq(kvm_state, vector->virq); | |
555 | vector->virq = -1; | |
556 | } else { | |
557 | qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt), | |
558 | NULL, NULL, NULL); | |
559 | } | |
560 | ||
561 | if (msix) { | |
562 | msix_vector_unuse(&vdev->pdev, i); | |
563 | } | |
564 | ||
565 | event_notifier_cleanup(&vector->interrupt); | |
566 | } | |
567 | ||
568 | g_free(vdev->msi_vectors); | |
569 | vdev->msi_vectors = NULL; | |
570 | vdev->nr_vectors = 0; | |
571 | ||
572 | if (!msix) { | |
573 | msi_set_qsize(&vdev->pdev, 0); /* Actually still means 1 vector */ | |
574 | } | |
575 | ||
576 | DPRINTF("%s(%04x:%02x:%02x.%x, msi%s)\n", __func__, | |
577 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
578 | vdev->host.function, msix ? "x" : ""); | |
579 | ||
580 | vfio_enable_intx(vdev); | |
581 | } | |
582 | ||
583 | /* | |
584 | * IO Port/MMIO - Beware of the endians, VFIO is always little endian | |
585 | */ | |
586 | static void vfio_bar_write(void *opaque, target_phys_addr_t addr, | |
587 | uint64_t data, unsigned size) | |
588 | { | |
589 | VFIOBAR *bar = opaque; | |
590 | union { | |
591 | uint8_t byte; | |
592 | uint16_t word; | |
593 | uint32_t dword; | |
594 | uint64_t qword; | |
595 | } buf; | |
596 | ||
597 | switch (size) { | |
598 | case 1: | |
599 | buf.byte = data; | |
600 | break; | |
601 | case 2: | |
602 | buf.word = cpu_to_le16(data); | |
603 | break; | |
604 | case 4: | |
605 | buf.dword = cpu_to_le32(data); | |
606 | break; | |
607 | default: | |
608 | hw_error("vfio: unsupported write size, %d bytes\n", size); | |
609 | break; | |
610 | } | |
611 | ||
612 | if (pwrite(bar->fd, &buf, size, bar->fd_offset + addr) != size) { | |
613 | error_report("%s(,0x%"TARGET_PRIxPHYS", 0x%"PRIx64", %d) failed: %m\n", | |
614 | __func__, addr, data, size); | |
615 | } | |
616 | ||
617 | DPRINTF("%s(BAR%d+0x%"TARGET_PRIxPHYS", 0x%"PRIx64", %d)\n", | |
618 | __func__, bar->nr, addr, data, size); | |
619 | ||
620 | /* | |
621 | * A read or write to a BAR always signals an INTx EOI. This will | |
622 | * do nothing if not pending (including not in INTx mode). We assume | |
623 | * that a BAR access is in response to an interrupt and that BAR | |
624 | * accesses will service the interrupt. Unfortunately, we don't know | |
625 | * which access will service the interrupt, so we're potentially | |
626 | * getting quite a few host interrupts per guest interrupt. | |
627 | */ | |
628 | vfio_eoi(DO_UPCAST(VFIODevice, bars[bar->nr], bar)); | |
629 | } | |
630 | ||
631 | static uint64_t vfio_bar_read(void *opaque, | |
632 | target_phys_addr_t addr, unsigned size) | |
633 | { | |
634 | VFIOBAR *bar = opaque; | |
635 | union { | |
636 | uint8_t byte; | |
637 | uint16_t word; | |
638 | uint32_t dword; | |
639 | uint64_t qword; | |
640 | } buf; | |
641 | uint64_t data = 0; | |
642 | ||
643 | if (pread(bar->fd, &buf, size, bar->fd_offset + addr) != size) { | |
644 | error_report("%s(,0x%"TARGET_PRIxPHYS", %d) failed: %m\n", | |
645 | __func__, addr, size); | |
646 | return (uint64_t)-1; | |
647 | } | |
648 | ||
649 | switch (size) { | |
650 | case 1: | |
651 | data = buf.byte; | |
652 | break; | |
653 | case 2: | |
654 | data = le16_to_cpu(buf.word); | |
655 | break; | |
656 | case 4: | |
657 | data = le32_to_cpu(buf.dword); | |
658 | break; | |
659 | default: | |
660 | hw_error("vfio: unsupported read size, %d bytes\n", size); | |
661 | break; | |
662 | } | |
663 | ||
664 | DPRINTF("%s(BAR%d+0x%"TARGET_PRIxPHYS", %d) = 0x%"PRIx64"\n", | |
665 | __func__, bar->nr, addr, size, data); | |
666 | ||
667 | /* Same as write above */ | |
668 | vfio_eoi(DO_UPCAST(VFIODevice, bars[bar->nr], bar)); | |
669 | ||
670 | return data; | |
671 | } | |
672 | ||
673 | static const MemoryRegionOps vfio_bar_ops = { | |
674 | .read = vfio_bar_read, | |
675 | .write = vfio_bar_write, | |
676 | .endianness = DEVICE_LITTLE_ENDIAN, | |
677 | }; | |
678 | ||
679 | /* | |
680 | * PCI config space | |
681 | */ | |
682 | static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len) | |
683 | { | |
684 | VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev); | |
685 | uint32_t val = 0; | |
686 | ||
687 | /* | |
688 | * We only need QEMU PCI config support for the ROM BAR, the MSI and MSIX | |
689 | * capabilities, and the multifunction bit below. We let VFIO handle | |
690 | * virtualizing everything else. Performance is not a concern here. | |
691 | */ | |
692 | if (ranges_overlap(addr, len, PCI_ROM_ADDRESS, 4) || | |
693 | (pdev->cap_present & QEMU_PCI_CAP_MSIX && | |
694 | ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) || | |
695 | (pdev->cap_present & QEMU_PCI_CAP_MSI && | |
696 | ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size))) { | |
697 | ||
698 | val = pci_default_read_config(pdev, addr, len); | |
699 | } else { | |
700 | if (pread(vdev->fd, &val, len, vdev->config_offset + addr) != len) { | |
701 | error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m\n", | |
702 | __func__, vdev->host.domain, vdev->host.bus, | |
703 | vdev->host.slot, vdev->host.function, addr, len); | |
704 | return -errno; | |
705 | } | |
706 | val = le32_to_cpu(val); | |
707 | } | |
708 | ||
709 | /* Multifunction bit is virualized in QEMU */ | |
710 | if (unlikely(ranges_overlap(addr, len, PCI_HEADER_TYPE, 1))) { | |
711 | uint32_t mask = PCI_HEADER_TYPE_MULTI_FUNCTION; | |
712 | ||
713 | if (len == 4) { | |
714 | mask <<= 16; | |
715 | } | |
716 | ||
717 | if (pdev->cap_present & QEMU_PCI_CAP_MULTIFUNCTION) { | |
718 | val |= mask; | |
719 | } else { | |
720 | val &= ~mask; | |
721 | } | |
722 | } | |
723 | ||
724 | DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, len=0x%x) %x\n", __func__, | |
725 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
726 | vdev->host.function, addr, len, val); | |
727 | ||
728 | return val; | |
729 | } | |
730 | ||
731 | static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr, | |
732 | uint32_t val, int len) | |
733 | { | |
734 | VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev); | |
735 | uint32_t val_le = cpu_to_le32(val); | |
736 | ||
737 | DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, 0x%x, len=0x%x)\n", __func__, | |
738 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
739 | vdev->host.function, addr, val, len); | |
740 | ||
741 | /* Write everything to VFIO, let it filter out what we can't write */ | |
742 | if (pwrite(vdev->fd, &val_le, len, vdev->config_offset + addr) != len) { | |
743 | error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m\n", | |
744 | __func__, vdev->host.domain, vdev->host.bus, | |
745 | vdev->host.slot, vdev->host.function, addr, val, len); | |
746 | } | |
747 | ||
748 | /* Write standard header bits to emulation */ | |
749 | if (addr < PCI_CONFIG_HEADER_SIZE) { | |
750 | pci_default_write_config(pdev, addr, val, len); | |
751 | return; | |
752 | } | |
753 | ||
754 | /* MSI/MSI-X Enabling/Disabling */ | |
755 | if (pdev->cap_present & QEMU_PCI_CAP_MSI && | |
756 | ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) { | |
757 | int is_enabled, was_enabled = msi_enabled(pdev); | |
758 | ||
759 | pci_default_write_config(pdev, addr, val, len); | |
760 | ||
761 | is_enabled = msi_enabled(pdev); | |
762 | ||
763 | if (!was_enabled && is_enabled) { | |
764 | vfio_enable_msi(vdev); | |
765 | } else if (was_enabled && !is_enabled) { | |
766 | vfio_disable_msi_x(vdev, false); | |
767 | } | |
768 | } | |
769 | ||
770 | if (pdev->cap_present & QEMU_PCI_CAP_MSIX && | |
771 | ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) { | |
772 | int is_enabled, was_enabled = msix_enabled(pdev); | |
773 | ||
774 | pci_default_write_config(pdev, addr, val, len); | |
775 | ||
776 | is_enabled = msix_enabled(pdev); | |
777 | ||
778 | if (!was_enabled && is_enabled) { | |
779 | /* vfio_msix_vector_use handles this automatically */ | |
780 | } else if (was_enabled && !is_enabled) { | |
781 | vfio_disable_msi_x(vdev, true); | |
782 | } | |
783 | } | |
784 | } | |
785 | ||
786 | /* | |
787 | * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86 | |
788 | */ | |
af6bc27e AW |
789 | static int vfio_dma_unmap(VFIOContainer *container, |
790 | target_phys_addr_t iova, ram_addr_t size) | |
791 | { | |
792 | struct vfio_iommu_type1_dma_unmap unmap = { | |
793 | .argsz = sizeof(unmap), | |
794 | .flags = 0, | |
795 | .iova = iova, | |
796 | .size = size, | |
797 | }; | |
798 | ||
799 | if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) { | |
800 | DPRINTF("VFIO_UNMAP_DMA: %d\n", -errno); | |
801 | return -errno; | |
802 | } | |
803 | ||
804 | return 0; | |
805 | } | |
806 | ||
65501a74 AW |
807 | static int vfio_dma_map(VFIOContainer *container, target_phys_addr_t iova, |
808 | ram_addr_t size, void *vaddr, bool readonly) | |
809 | { | |
810 | struct vfio_iommu_type1_dma_map map = { | |
811 | .argsz = sizeof(map), | |
812 | .flags = VFIO_DMA_MAP_FLAG_READ, | |
0f41dc18 | 813 | .vaddr = (__u64)(intptr_t)vaddr, |
65501a74 AW |
814 | .iova = iova, |
815 | .size = size, | |
816 | }; | |
817 | ||
818 | if (!readonly) { | |
819 | map.flags |= VFIO_DMA_MAP_FLAG_WRITE; | |
820 | } | |
821 | ||
822 | if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map)) { | |
823 | DPRINTF("VFIO_MAP_DMA: %d\n", -errno); | |
824 | return -errno; | |
825 | } | |
826 | ||
827 | return 0; | |
828 | } | |
829 | ||
65501a74 AW |
830 | static void vfio_listener_dummy1(MemoryListener *listener) |
831 | { | |
832 | /* We don't do batching (begin/commit) or care about logging */ | |
833 | } | |
834 | ||
835 | static void vfio_listener_dummy2(MemoryListener *listener, | |
836 | MemoryRegionSection *section) | |
837 | { | |
838 | /* We don't do logging or care about nops */ | |
839 | } | |
840 | ||
841 | static void vfio_listener_dummy3(MemoryListener *listener, | |
842 | MemoryRegionSection *section, | |
843 | bool match_data, uint64_t data, | |
844 | EventNotifier *e) | |
845 | { | |
846 | /* We don't care about eventfds */ | |
847 | } | |
848 | ||
849 | static bool vfio_listener_skipped_section(MemoryRegionSection *section) | |
850 | { | |
851 | return !memory_region_is_ram(section->mr); | |
852 | } | |
853 | ||
854 | static void vfio_listener_region_add(MemoryListener *listener, | |
855 | MemoryRegionSection *section) | |
856 | { | |
857 | VFIOContainer *container = container_of(listener, VFIOContainer, | |
858 | iommu_data.listener); | |
859 | target_phys_addr_t iova, end; | |
860 | void *vaddr; | |
861 | int ret; | |
862 | ||
863 | if (vfio_listener_skipped_section(section)) { | |
864 | DPRINTF("vfio: SKIPPING region_add %"TARGET_PRIxPHYS" - %"PRIx64"\n", | |
865 | section->offset_within_address_space, | |
866 | section->offset_within_address_space + section->size - 1); | |
867 | return; | |
868 | } | |
869 | ||
870 | if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) != | |
871 | (section->offset_within_region & ~TARGET_PAGE_MASK))) { | |
872 | error_report("%s received unaligned region\n", __func__); | |
873 | return; | |
874 | } | |
875 | ||
876 | iova = TARGET_PAGE_ALIGN(section->offset_within_address_space); | |
877 | end = (section->offset_within_address_space + section->size) & | |
878 | TARGET_PAGE_MASK; | |
879 | ||
880 | if (iova >= end) { | |
881 | return; | |
882 | } | |
883 | ||
884 | vaddr = memory_region_get_ram_ptr(section->mr) + | |
885 | section->offset_within_region + | |
886 | (iova - section->offset_within_address_space); | |
887 | ||
888 | DPRINTF("vfio: region_add %"TARGET_PRIxPHYS" - %"TARGET_PRIxPHYS" [%p]\n", | |
889 | iova, end - 1, vaddr); | |
890 | ||
891 | ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly); | |
892 | if (ret) { | |
893 | error_report("vfio_dma_map(%p, 0x%"TARGET_PRIxPHYS", " | |
894 | "0x%"TARGET_PRIxPHYS", %p) = %d (%m)\n", | |
895 | container, iova, end - iova, vaddr, ret); | |
896 | } | |
897 | } | |
898 | ||
899 | static void vfio_listener_region_del(MemoryListener *listener, | |
900 | MemoryRegionSection *section) | |
901 | { | |
902 | VFIOContainer *container = container_of(listener, VFIOContainer, | |
903 | iommu_data.listener); | |
904 | target_phys_addr_t iova, end; | |
905 | int ret; | |
906 | ||
907 | if (vfio_listener_skipped_section(section)) { | |
908 | DPRINTF("vfio: SKIPPING region_del %"TARGET_PRIxPHYS" - %"PRIx64"\n", | |
909 | section->offset_within_address_space, | |
910 | section->offset_within_address_space + section->size - 1); | |
911 | return; | |
912 | } | |
913 | ||
914 | if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) != | |
915 | (section->offset_within_region & ~TARGET_PAGE_MASK))) { | |
916 | error_report("%s received unaligned region\n", __func__); | |
917 | return; | |
918 | } | |
919 | ||
920 | iova = TARGET_PAGE_ALIGN(section->offset_within_address_space); | |
921 | end = (section->offset_within_address_space + section->size) & | |
922 | TARGET_PAGE_MASK; | |
923 | ||
924 | if (iova >= end) { | |
925 | return; | |
926 | } | |
927 | ||
928 | DPRINTF("vfio: region_del %"TARGET_PRIxPHYS" - %"TARGET_PRIxPHYS"\n", | |
929 | iova, end - 1); | |
930 | ||
931 | ret = vfio_dma_unmap(container, iova, end - iova); | |
932 | if (ret) { | |
933 | error_report("vfio_dma_unmap(%p, 0x%"TARGET_PRIxPHYS", " | |
934 | "0x%"TARGET_PRIxPHYS") = %d (%m)\n", | |
935 | container, iova, end - iova, ret); | |
936 | } | |
937 | } | |
938 | ||
939 | static MemoryListener vfio_memory_listener = { | |
940 | .begin = vfio_listener_dummy1, | |
941 | .commit = vfio_listener_dummy1, | |
942 | .region_add = vfio_listener_region_add, | |
943 | .region_del = vfio_listener_region_del, | |
944 | .region_nop = vfio_listener_dummy2, | |
945 | .log_start = vfio_listener_dummy2, | |
946 | .log_stop = vfio_listener_dummy2, | |
947 | .log_sync = vfio_listener_dummy2, | |
948 | .log_global_start = vfio_listener_dummy1, | |
949 | .log_global_stop = vfio_listener_dummy1, | |
950 | .eventfd_add = vfio_listener_dummy3, | |
951 | .eventfd_del = vfio_listener_dummy3, | |
952 | }; | |
953 | ||
954 | static void vfio_listener_release(VFIOContainer *container) | |
955 | { | |
956 | memory_listener_unregister(&container->iommu_data.listener); | |
957 | } | |
958 | ||
959 | /* | |
960 | * Interrupt setup | |
961 | */ | |
962 | static void vfio_disable_interrupts(VFIODevice *vdev) | |
963 | { | |
964 | switch (vdev->interrupt) { | |
965 | case VFIO_INT_INTx: | |
966 | vfio_disable_intx(vdev); | |
967 | break; | |
968 | case VFIO_INT_MSI: | |
969 | vfio_disable_msi_x(vdev, false); | |
970 | break; | |
971 | case VFIO_INT_MSIX: | |
972 | vfio_disable_msi_x(vdev, true); | |
973 | break; | |
974 | } | |
975 | } | |
976 | ||
977 | static int vfio_setup_msi(VFIODevice *vdev, int pos) | |
978 | { | |
979 | uint16_t ctrl; | |
980 | bool msi_64bit, msi_maskbit; | |
981 | int ret, entries; | |
982 | ||
983 | /* | |
984 | * TODO: don't peek into msi_supported, let msi_init fail and | |
985 | * check for ENOTSUP | |
986 | */ | |
987 | if (!msi_supported) { | |
988 | return 0; | |
989 | } | |
990 | ||
991 | if (pread(vdev->fd, &ctrl, sizeof(ctrl), | |
992 | vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) { | |
993 | return -errno; | |
994 | } | |
995 | ctrl = le16_to_cpu(ctrl); | |
996 | ||
997 | msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT); | |
998 | msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT); | |
999 | entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1); | |
1000 | ||
1001 | DPRINTF("%04x:%02x:%02x.%x PCI MSI CAP @0x%x\n", vdev->host.domain, | |
1002 | vdev->host.bus, vdev->host.slot, vdev->host.function, pos); | |
1003 | ||
1004 | ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit); | |
1005 | if (ret < 0) { | |
1006 | error_report("vfio: msi_init failed\n"); | |
1007 | return ret; | |
1008 | } | |
1009 | vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0); | |
1010 | ||
1011 | return 0; | |
1012 | } | |
1013 | ||
1014 | /* | |
1015 | * We don't have any control over how pci_add_capability() inserts | |
1016 | * capabilities into the chain. In order to setup MSI-X we need a | |
1017 | * MemoryRegion for the BAR. In order to setup the BAR and not | |
1018 | * attempt to mmap the MSI-X table area, which VFIO won't allow, we | |
1019 | * need to first look for where the MSI-X table lives. So we | |
1020 | * unfortunately split MSI-X setup across two functions. | |
1021 | */ | |
1022 | static int vfio_early_setup_msix(VFIODevice *vdev) | |
1023 | { | |
1024 | uint8_t pos; | |
1025 | uint16_t ctrl; | |
1026 | uint32_t table, pba; | |
1027 | ||
1028 | pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX); | |
1029 | if (!pos) { | |
1030 | return 0; | |
1031 | } | |
1032 | ||
1033 | if (pread(vdev->fd, &ctrl, sizeof(ctrl), | |
1034 | vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) { | |
1035 | return -errno; | |
1036 | } | |
1037 | ||
1038 | if (pread(vdev->fd, &table, sizeof(table), | |
1039 | vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) { | |
1040 | return -errno; | |
1041 | } | |
1042 | ||
1043 | if (pread(vdev->fd, &pba, sizeof(pba), | |
1044 | vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) { | |
1045 | return -errno; | |
1046 | } | |
1047 | ||
1048 | ctrl = le16_to_cpu(ctrl); | |
1049 | table = le32_to_cpu(table); | |
1050 | pba = le32_to_cpu(pba); | |
1051 | ||
1052 | vdev->msix = g_malloc0(sizeof(*(vdev->msix))); | |
1053 | vdev->msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK; | |
1054 | vdev->msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK; | |
1055 | vdev->msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK; | |
1056 | vdev->msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK; | |
1057 | vdev->msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1; | |
1058 | ||
1059 | DPRINTF("%04x:%02x:%02x.%x " | |
1060 | "PCI MSI-X CAP @0x%x, BAR %d, offset 0x%x, entries %d\n", | |
1061 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
1062 | vdev->host.function, pos, vdev->msix->table_bar, | |
1063 | vdev->msix->table_offset, vdev->msix->entries); | |
1064 | ||
1065 | return 0; | |
1066 | } | |
1067 | ||
1068 | static int vfio_setup_msix(VFIODevice *vdev, int pos) | |
1069 | { | |
1070 | int ret; | |
1071 | ||
1072 | /* | |
1073 | * TODO: don't peek into msi_supported, let msix_init fail and | |
1074 | * check for ENOTSUP | |
1075 | */ | |
1076 | if (!msi_supported) { | |
1077 | return 0; | |
1078 | } | |
1079 | ||
1080 | ret = msix_init(&vdev->pdev, vdev->msix->entries, | |
1081 | &vdev->bars[vdev->msix->table_bar].mem, | |
1082 | vdev->msix->table_bar, vdev->msix->table_offset, | |
1083 | &vdev->bars[vdev->msix->pba_bar].mem, | |
1084 | vdev->msix->pba_bar, vdev->msix->pba_offset, pos); | |
1085 | if (ret < 0) { | |
1086 | error_report("vfio: msix_init failed\n"); | |
1087 | return ret; | |
1088 | } | |
1089 | ||
1090 | ret = msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use, | |
1091 | vfio_msix_vector_release); | |
1092 | if (ret) { | |
1093 | error_report("vfio: msix_set_vector_notifiers failed %d\n", ret); | |
1094 | msix_uninit(&vdev->pdev, &vdev->bars[vdev->msix->table_bar].mem, | |
1095 | &vdev->bars[vdev->msix->pba_bar].mem); | |
1096 | return ret; | |
1097 | } | |
1098 | ||
1099 | return 0; | |
1100 | } | |
1101 | ||
1102 | static void vfio_teardown_msi(VFIODevice *vdev) | |
1103 | { | |
1104 | msi_uninit(&vdev->pdev); | |
1105 | ||
1106 | if (vdev->msix) { | |
1107 | /* FIXME: Why can't unset just silently do nothing?? */ | |
1108 | if (vdev->pdev.msix_vector_use_notifier && | |
1109 | vdev->pdev.msix_vector_release_notifier) { | |
1110 | msix_unset_vector_notifiers(&vdev->pdev); | |
1111 | } | |
1112 | ||
1113 | msix_uninit(&vdev->pdev, &vdev->bars[vdev->msix->table_bar].mem, | |
1114 | &vdev->bars[vdev->msix->pba_bar].mem); | |
1115 | } | |
1116 | } | |
1117 | ||
1118 | /* | |
1119 | * Resource setup | |
1120 | */ | |
1121 | static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled) | |
1122 | { | |
1123 | int i; | |
1124 | ||
1125 | for (i = 0; i < PCI_ROM_SLOT; i++) { | |
1126 | VFIOBAR *bar = &vdev->bars[i]; | |
1127 | ||
1128 | if (!bar->size) { | |
1129 | continue; | |
1130 | } | |
1131 | ||
1132 | memory_region_set_enabled(&bar->mmap_mem, enabled); | |
1133 | if (vdev->msix && vdev->msix->table_bar == i) { | |
1134 | memory_region_set_enabled(&vdev->msix->mmap_mem, enabled); | |
1135 | } | |
1136 | } | |
1137 | } | |
1138 | ||
1139 | static void vfio_unmap_bar(VFIODevice *vdev, int nr) | |
1140 | { | |
1141 | VFIOBAR *bar = &vdev->bars[nr]; | |
1142 | ||
1143 | if (!bar->size) { | |
1144 | return; | |
1145 | } | |
1146 | ||
1147 | memory_region_del_subregion(&bar->mem, &bar->mmap_mem); | |
1148 | munmap(bar->mmap, memory_region_size(&bar->mmap_mem)); | |
1149 | ||
1150 | if (vdev->msix && vdev->msix->table_bar == nr) { | |
1151 | memory_region_del_subregion(&bar->mem, &vdev->msix->mmap_mem); | |
1152 | munmap(vdev->msix->mmap, memory_region_size(&vdev->msix->mmap_mem)); | |
1153 | } | |
1154 | ||
1155 | memory_region_destroy(&bar->mem); | |
1156 | } | |
1157 | ||
1158 | static int vfio_mmap_bar(VFIOBAR *bar, MemoryRegion *mem, MemoryRegion *submem, | |
1159 | void **map, size_t size, off_t offset, | |
1160 | const char *name) | |
1161 | { | |
1162 | int ret = 0; | |
1163 | ||
1164 | if (size && bar->flags & VFIO_REGION_INFO_FLAG_MMAP) { | |
1165 | int prot = 0; | |
1166 | ||
1167 | if (bar->flags & VFIO_REGION_INFO_FLAG_READ) { | |
1168 | prot |= PROT_READ; | |
1169 | } | |
1170 | ||
1171 | if (bar->flags & VFIO_REGION_INFO_FLAG_WRITE) { | |
1172 | prot |= PROT_WRITE; | |
1173 | } | |
1174 | ||
1175 | *map = mmap(NULL, size, prot, MAP_SHARED, | |
1176 | bar->fd, bar->fd_offset + offset); | |
1177 | if (*map == MAP_FAILED) { | |
1178 | *map = NULL; | |
1179 | ret = -errno; | |
1180 | goto empty_region; | |
1181 | } | |
1182 | ||
1183 | memory_region_init_ram_ptr(submem, name, size, *map); | |
1184 | } else { | |
1185 | empty_region: | |
1186 | /* Create a zero sized sub-region to make cleanup easy. */ | |
1187 | memory_region_init(submem, name, 0); | |
1188 | } | |
1189 | ||
1190 | memory_region_add_subregion(mem, offset, submem); | |
1191 | ||
1192 | return ret; | |
1193 | } | |
1194 | ||
1195 | static void vfio_map_bar(VFIODevice *vdev, int nr) | |
1196 | { | |
1197 | VFIOBAR *bar = &vdev->bars[nr]; | |
1198 | unsigned size = bar->size; | |
1199 | char name[64]; | |
1200 | uint32_t pci_bar; | |
1201 | uint8_t type; | |
1202 | int ret; | |
1203 | ||
1204 | /* Skip both unimplemented BARs and the upper half of 64bit BARS. */ | |
1205 | if (!size) { | |
1206 | return; | |
1207 | } | |
1208 | ||
1209 | snprintf(name, sizeof(name), "VFIO %04x:%02x:%02x.%x BAR %d", | |
1210 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
1211 | vdev->host.function, nr); | |
1212 | ||
1213 | /* Determine what type of BAR this is for registration */ | |
1214 | ret = pread(vdev->fd, &pci_bar, sizeof(pci_bar), | |
1215 | vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr)); | |
1216 | if (ret != sizeof(pci_bar)) { | |
1217 | error_report("vfio: Failed to read BAR %d (%m)\n", nr); | |
1218 | return; | |
1219 | } | |
1220 | ||
1221 | pci_bar = le32_to_cpu(pci_bar); | |
1222 | type = pci_bar & (pci_bar & PCI_BASE_ADDRESS_SPACE_IO ? | |
1223 | ~PCI_BASE_ADDRESS_IO_MASK : ~PCI_BASE_ADDRESS_MEM_MASK); | |
1224 | ||
1225 | /* A "slow" read/write mapping underlies all BARs */ | |
1226 | memory_region_init_io(&bar->mem, &vfio_bar_ops, bar, name, size); | |
1227 | pci_register_bar(&vdev->pdev, nr, type, &bar->mem); | |
1228 | ||
1229 | /* | |
1230 | * We can't mmap areas overlapping the MSIX vector table, so we | |
1231 | * potentially insert a direct-mapped subregion before and after it. | |
1232 | */ | |
1233 | if (vdev->msix && vdev->msix->table_bar == nr) { | |
1234 | size = vdev->msix->table_offset & TARGET_PAGE_MASK; | |
1235 | } | |
1236 | ||
1237 | strncat(name, " mmap", sizeof(name) - strlen(name) - 1); | |
1238 | if (vfio_mmap_bar(bar, &bar->mem, | |
1239 | &bar->mmap_mem, &bar->mmap, size, 0, name)) { | |
1240 | error_report("%s unsupported. Performance may be slow\n", name); | |
1241 | } | |
1242 | ||
1243 | if (vdev->msix && vdev->msix->table_bar == nr) { | |
1244 | unsigned start; | |
1245 | ||
1246 | start = TARGET_PAGE_ALIGN(vdev->msix->table_offset + | |
1247 | (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE)); | |
1248 | ||
1249 | size = start < bar->size ? bar->size - start : 0; | |
1250 | strncat(name, " msix-hi", sizeof(name) - strlen(name) - 1); | |
1251 | /* VFIOMSIXInfo contains another MemoryRegion for this mapping */ | |
1252 | if (vfio_mmap_bar(bar, &bar->mem, &vdev->msix->mmap_mem, | |
1253 | &vdev->msix->mmap, size, start, name)) { | |
1254 | error_report("%s unsupported. Performance may be slow\n", name); | |
1255 | } | |
1256 | } | |
1257 | } | |
1258 | ||
1259 | static void vfio_map_bars(VFIODevice *vdev) | |
1260 | { | |
1261 | int i; | |
1262 | ||
1263 | for (i = 0; i < PCI_ROM_SLOT; i++) { | |
1264 | vfio_map_bar(vdev, i); | |
1265 | } | |
1266 | } | |
1267 | ||
1268 | static void vfio_unmap_bars(VFIODevice *vdev) | |
1269 | { | |
1270 | int i; | |
1271 | ||
1272 | for (i = 0; i < PCI_ROM_SLOT; i++) { | |
1273 | vfio_unmap_bar(vdev, i); | |
1274 | } | |
1275 | } | |
1276 | ||
1277 | /* | |
1278 | * General setup | |
1279 | */ | |
1280 | static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos) | |
1281 | { | |
1282 | uint8_t tmp, next = 0xff; | |
1283 | ||
1284 | for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp; | |
1285 | tmp = pdev->config[tmp + 1]) { | |
1286 | if (tmp > pos && tmp < next) { | |
1287 | next = tmp; | |
1288 | } | |
1289 | } | |
1290 | ||
1291 | return next - pos; | |
1292 | } | |
1293 | ||
1294 | static int vfio_add_std_cap(VFIODevice *vdev, uint8_t pos) | |
1295 | { | |
1296 | PCIDevice *pdev = &vdev->pdev; | |
1297 | uint8_t cap_id, next, size; | |
1298 | int ret; | |
1299 | ||
1300 | cap_id = pdev->config[pos]; | |
1301 | next = pdev->config[pos + 1]; | |
1302 | ||
1303 | /* | |
1304 | * If it becomes important to configure capabilities to their actual | |
1305 | * size, use this as the default when it's something we don't recognize. | |
1306 | * Since QEMU doesn't actually handle many of the config accesses, | |
1307 | * exact size doesn't seem worthwhile. | |
1308 | */ | |
1309 | size = vfio_std_cap_max_size(pdev, pos); | |
1310 | ||
1311 | /* | |
1312 | * pci_add_capability always inserts the new capability at the head | |
1313 | * of the chain. Therefore to end up with a chain that matches the | |
1314 | * physical device, we insert from the end by making this recursive. | |
1315 | * This is also why we pre-caclulate size above as cached config space | |
1316 | * will be changed as we unwind the stack. | |
1317 | */ | |
1318 | if (next) { | |
1319 | ret = vfio_add_std_cap(vdev, next); | |
1320 | if (ret) { | |
1321 | return ret; | |
1322 | } | |
1323 | } else { | |
1324 | pdev->config[PCI_CAPABILITY_LIST] = 0; /* Begin the rebuild */ | |
1325 | } | |
1326 | ||
1327 | switch (cap_id) { | |
1328 | case PCI_CAP_ID_MSI: | |
1329 | ret = vfio_setup_msi(vdev, pos); | |
1330 | break; | |
1331 | case PCI_CAP_ID_MSIX: | |
1332 | ret = vfio_setup_msix(vdev, pos); | |
1333 | break; | |
1334 | default: | |
1335 | ret = pci_add_capability(pdev, cap_id, pos, size); | |
1336 | break; | |
1337 | } | |
1338 | ||
1339 | if (ret < 0) { | |
1340 | error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability " | |
1341 | "0x%x[0x%x]@0x%x: %d\n", vdev->host.domain, | |
1342 | vdev->host.bus, vdev->host.slot, vdev->host.function, | |
1343 | cap_id, size, pos, ret); | |
1344 | return ret; | |
1345 | } | |
1346 | ||
1347 | return 0; | |
1348 | } | |
1349 | ||
1350 | static int vfio_add_capabilities(VFIODevice *vdev) | |
1351 | { | |
1352 | PCIDevice *pdev = &vdev->pdev; | |
1353 | ||
1354 | if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) || | |
1355 | !pdev->config[PCI_CAPABILITY_LIST]) { | |
1356 | return 0; /* Nothing to add */ | |
1357 | } | |
1358 | ||
1359 | return vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST]); | |
1360 | } | |
1361 | ||
1362 | static int vfio_load_rom(VFIODevice *vdev) | |
1363 | { | |
1364 | uint64_t size = vdev->rom_size; | |
1365 | char name[32]; | |
1366 | off_t off = 0, voff = vdev->rom_offset; | |
1367 | ssize_t bytes; | |
1368 | void *ptr; | |
1369 | ||
1370 | /* If loading ROM from file, pci handles it */ | |
1371 | if (vdev->pdev.romfile || !vdev->pdev.rom_bar || !size) { | |
1372 | return 0; | |
1373 | } | |
1374 | ||
1375 | DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain, | |
1376 | vdev->host.bus, vdev->host.slot, vdev->host.function); | |
1377 | ||
1378 | snprintf(name, sizeof(name), "vfio[%04x:%02x:%02x.%x].rom", | |
1379 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
1380 | vdev->host.function); | |
1381 | memory_region_init_ram(&vdev->pdev.rom, name, size); | |
1382 | ptr = memory_region_get_ram_ptr(&vdev->pdev.rom); | |
1383 | memset(ptr, 0xff, size); | |
1384 | ||
1385 | while (size) { | |
1386 | bytes = pread(vdev->fd, ptr + off, size, voff + off); | |
1387 | if (bytes == 0) { | |
1388 | break; /* expect that we could get back less than the ROM BAR */ | |
1389 | } else if (bytes > 0) { | |
1390 | off += bytes; | |
1391 | size -= bytes; | |
1392 | } else { | |
1393 | if (errno == EINTR || errno == EAGAIN) { | |
1394 | continue; | |
1395 | } | |
1396 | error_report("vfio: Error reading device ROM: %m\n"); | |
1397 | memory_region_destroy(&vdev->pdev.rom); | |
1398 | return -errno; | |
1399 | } | |
1400 | } | |
1401 | ||
1402 | pci_register_bar(&vdev->pdev, PCI_ROM_SLOT, 0, &vdev->pdev.rom); | |
1403 | vdev->pdev.has_rom = true; | |
1404 | return 0; | |
1405 | } | |
1406 | ||
1407 | static int vfio_connect_container(VFIOGroup *group) | |
1408 | { | |
1409 | VFIOContainer *container; | |
1410 | int ret, fd; | |
1411 | ||
1412 | if (group->container) { | |
1413 | return 0; | |
1414 | } | |
1415 | ||
1416 | QLIST_FOREACH(container, &container_list, next) { | |
1417 | if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) { | |
1418 | group->container = container; | |
1419 | QLIST_INSERT_HEAD(&container->group_list, group, container_next); | |
1420 | return 0; | |
1421 | } | |
1422 | } | |
1423 | ||
1424 | fd = qemu_open("/dev/vfio/vfio", O_RDWR); | |
1425 | if (fd < 0) { | |
1426 | error_report("vfio: failed to open /dev/vfio/vfio: %m\n"); | |
1427 | return -errno; | |
1428 | } | |
1429 | ||
1430 | ret = ioctl(fd, VFIO_GET_API_VERSION); | |
1431 | if (ret != VFIO_API_VERSION) { | |
1432 | error_report("vfio: supported vfio version: %d, " | |
1433 | "reported version: %d\n", VFIO_API_VERSION, ret); | |
1434 | close(fd); | |
1435 | return -EINVAL; | |
1436 | } | |
1437 | ||
1438 | container = g_malloc0(sizeof(*container)); | |
1439 | container->fd = fd; | |
1440 | ||
1441 | if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) { | |
1442 | ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd); | |
1443 | if (ret) { | |
1444 | error_report("vfio: failed to set group container: %m\n"); | |
1445 | g_free(container); | |
1446 | close(fd); | |
1447 | return -errno; | |
1448 | } | |
1449 | ||
1450 | ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU); | |
1451 | if (ret) { | |
1452 | error_report("vfio: failed to set iommu for container: %m\n"); | |
1453 | g_free(container); | |
1454 | close(fd); | |
1455 | return -errno; | |
1456 | } | |
1457 | ||
1458 | container->iommu_data.listener = vfio_memory_listener; | |
1459 | container->iommu_data.release = vfio_listener_release; | |
1460 | ||
1461 | memory_listener_register(&container->iommu_data.listener, | |
1462 | get_system_memory()); | |
1463 | } else { | |
1464 | error_report("vfio: No available IOMMU models\n"); | |
1465 | g_free(container); | |
1466 | close(fd); | |
1467 | return -EINVAL; | |
1468 | } | |
1469 | ||
1470 | QLIST_INIT(&container->group_list); | |
1471 | QLIST_INSERT_HEAD(&container_list, container, next); | |
1472 | ||
1473 | group->container = container; | |
1474 | QLIST_INSERT_HEAD(&container->group_list, group, container_next); | |
1475 | ||
1476 | return 0; | |
1477 | } | |
1478 | ||
1479 | static void vfio_disconnect_container(VFIOGroup *group) | |
1480 | { | |
1481 | VFIOContainer *container = group->container; | |
1482 | ||
1483 | if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) { | |
1484 | error_report("vfio: error disconnecting group %d from container\n", | |
1485 | group->groupid); | |
1486 | } | |
1487 | ||
1488 | QLIST_REMOVE(group, container_next); | |
1489 | group->container = NULL; | |
1490 | ||
1491 | if (QLIST_EMPTY(&container->group_list)) { | |
1492 | if (container->iommu_data.release) { | |
1493 | container->iommu_data.release(container); | |
1494 | } | |
1495 | QLIST_REMOVE(container, next); | |
1496 | DPRINTF("vfio_disconnect_container: close container->fd\n"); | |
1497 | close(container->fd); | |
1498 | g_free(container); | |
1499 | } | |
1500 | } | |
1501 | ||
1502 | static VFIOGroup *vfio_get_group(int groupid) | |
1503 | { | |
1504 | VFIOGroup *group; | |
1505 | char path[32]; | |
1506 | struct vfio_group_status status = { .argsz = sizeof(status) }; | |
1507 | ||
1508 | QLIST_FOREACH(group, &group_list, next) { | |
1509 | if (group->groupid == groupid) { | |
1510 | return group; | |
1511 | } | |
1512 | } | |
1513 | ||
1514 | group = g_malloc0(sizeof(*group)); | |
1515 | ||
1516 | snprintf(path, sizeof(path), "/dev/vfio/%d", groupid); | |
1517 | group->fd = qemu_open(path, O_RDWR); | |
1518 | if (group->fd < 0) { | |
1519 | error_report("vfio: error opening %s: %m\n", path); | |
1520 | g_free(group); | |
1521 | return NULL; | |
1522 | } | |
1523 | ||
1524 | if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) { | |
1525 | error_report("vfio: error getting group status: %m\n"); | |
1526 | close(group->fd); | |
1527 | g_free(group); | |
1528 | return NULL; | |
1529 | } | |
1530 | ||
1531 | if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) { | |
1532 | error_report("vfio: error, group %d is not viable, please ensure " | |
1533 | "all devices within the iommu_group are bound to their " | |
1534 | "vfio bus driver.\n", groupid); | |
1535 | close(group->fd); | |
1536 | g_free(group); | |
1537 | return NULL; | |
1538 | } | |
1539 | ||
1540 | group->groupid = groupid; | |
1541 | QLIST_INIT(&group->device_list); | |
1542 | ||
1543 | if (vfio_connect_container(group)) { | |
1544 | error_report("vfio: failed to setup container for group %d\n", groupid); | |
1545 | close(group->fd); | |
1546 | g_free(group); | |
1547 | return NULL; | |
1548 | } | |
1549 | ||
1550 | QLIST_INSERT_HEAD(&group_list, group, next); | |
1551 | ||
1552 | return group; | |
1553 | } | |
1554 | ||
1555 | static void vfio_put_group(VFIOGroup *group) | |
1556 | { | |
1557 | if (!QLIST_EMPTY(&group->device_list)) { | |
1558 | return; | |
1559 | } | |
1560 | ||
1561 | vfio_disconnect_container(group); | |
1562 | QLIST_REMOVE(group, next); | |
1563 | DPRINTF("vfio_put_group: close group->fd\n"); | |
1564 | close(group->fd); | |
1565 | g_free(group); | |
1566 | } | |
1567 | ||
1568 | static int vfio_get_device(VFIOGroup *group, const char *name, VFIODevice *vdev) | |
1569 | { | |
1570 | struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) }; | |
1571 | struct vfio_region_info reg_info = { .argsz = sizeof(reg_info) }; | |
1572 | int ret, i; | |
1573 | ||
1574 | ret = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name); | |
1575 | if (ret < 0) { | |
1576 | error_report("vfio: error getting device %s from group %d: %m\n", | |
1577 | name, group->groupid); | |
1578 | error_report("Verify all devices in group %d are bound to vfio-pci " | |
1579 | "or pci-stub and not already in use\n", group->groupid); | |
1580 | return ret; | |
1581 | } | |
1582 | ||
1583 | vdev->fd = ret; | |
1584 | vdev->group = group; | |
1585 | QLIST_INSERT_HEAD(&group->device_list, vdev, next); | |
1586 | ||
1587 | /* Sanity check device */ | |
1588 | ret = ioctl(vdev->fd, VFIO_DEVICE_GET_INFO, &dev_info); | |
1589 | if (ret) { | |
1590 | error_report("vfio: error getting device info: %m\n"); | |
1591 | goto error; | |
1592 | } | |
1593 | ||
1594 | DPRINTF("Device %s flags: %u, regions: %u, irgs: %u\n", name, | |
1595 | dev_info.flags, dev_info.num_regions, dev_info.num_irqs); | |
1596 | ||
1597 | if (!(dev_info.flags & VFIO_DEVICE_FLAGS_PCI)) { | |
1598 | error_report("vfio: Um, this isn't a PCI device\n"); | |
1599 | goto error; | |
1600 | } | |
1601 | ||
1602 | vdev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET); | |
1603 | if (!vdev->reset_works) { | |
1604 | error_report("Warning, device %s does not support reset\n", name); | |
1605 | } | |
1606 | ||
1607 | if (dev_info.num_regions != VFIO_PCI_NUM_REGIONS) { | |
1608 | error_report("vfio: unexpected number of io regions %u\n", | |
1609 | dev_info.num_regions); | |
1610 | goto error; | |
1611 | } | |
1612 | ||
1613 | if (dev_info.num_irqs != VFIO_PCI_NUM_IRQS) { | |
1614 | error_report("vfio: unexpected number of irqs %u\n", dev_info.num_irqs); | |
1615 | goto error; | |
1616 | } | |
1617 | ||
1618 | for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) { | |
1619 | reg_info.index = i; | |
1620 | ||
1621 | ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, ®_info); | |
1622 | if (ret) { | |
1623 | error_report("vfio: Error getting region %d info: %m\n", i); | |
1624 | goto error; | |
1625 | } | |
1626 | ||
1627 | DPRINTF("Device %s region %d:\n", name, i); | |
1628 | DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n", | |
1629 | (unsigned long)reg_info.size, (unsigned long)reg_info.offset, | |
1630 | (unsigned long)reg_info.flags); | |
1631 | ||
1632 | vdev->bars[i].flags = reg_info.flags; | |
1633 | vdev->bars[i].size = reg_info.size; | |
1634 | vdev->bars[i].fd_offset = reg_info.offset; | |
1635 | vdev->bars[i].fd = vdev->fd; | |
1636 | vdev->bars[i].nr = i; | |
1637 | } | |
1638 | ||
1639 | reg_info.index = VFIO_PCI_ROM_REGION_INDEX; | |
1640 | ||
1641 | ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, ®_info); | |
1642 | if (ret) { | |
1643 | error_report("vfio: Error getting ROM info: %m\n"); | |
1644 | goto error; | |
1645 | } | |
1646 | ||
1647 | DPRINTF("Device %s ROM:\n", name); | |
1648 | DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n", | |
1649 | (unsigned long)reg_info.size, (unsigned long)reg_info.offset, | |
1650 | (unsigned long)reg_info.flags); | |
1651 | ||
1652 | vdev->rom_size = reg_info.size; | |
1653 | vdev->rom_offset = reg_info.offset; | |
1654 | ||
1655 | reg_info.index = VFIO_PCI_CONFIG_REGION_INDEX; | |
1656 | ||
1657 | ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, ®_info); | |
1658 | if (ret) { | |
1659 | error_report("vfio: Error getting config info: %m\n"); | |
1660 | goto error; | |
1661 | } | |
1662 | ||
1663 | DPRINTF("Device %s config:\n", name); | |
1664 | DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n", | |
1665 | (unsigned long)reg_info.size, (unsigned long)reg_info.offset, | |
1666 | (unsigned long)reg_info.flags); | |
1667 | ||
1668 | vdev->config_size = reg_info.size; | |
1669 | vdev->config_offset = reg_info.offset; | |
1670 | ||
1671 | error: | |
1672 | if (ret) { | |
1673 | QLIST_REMOVE(vdev, next); | |
1674 | vdev->group = NULL; | |
1675 | close(vdev->fd); | |
1676 | } | |
1677 | return ret; | |
1678 | } | |
1679 | ||
1680 | static void vfio_put_device(VFIODevice *vdev) | |
1681 | { | |
1682 | QLIST_REMOVE(vdev, next); | |
1683 | vdev->group = NULL; | |
1684 | DPRINTF("vfio_put_device: close vdev->fd\n"); | |
1685 | close(vdev->fd); | |
1686 | if (vdev->msix) { | |
1687 | g_free(vdev->msix); | |
1688 | vdev->msix = NULL; | |
1689 | } | |
1690 | } | |
1691 | ||
1692 | static int vfio_initfn(PCIDevice *pdev) | |
1693 | { | |
1694 | VFIODevice *pvdev, *vdev = DO_UPCAST(VFIODevice, pdev, pdev); | |
1695 | VFIOGroup *group; | |
1696 | char path[PATH_MAX], iommu_group_path[PATH_MAX], *group_name; | |
1697 | ssize_t len; | |
1698 | struct stat st; | |
1699 | int groupid; | |
1700 | int ret; | |
1701 | ||
1702 | /* Check that the host device exists */ | |
1703 | snprintf(path, sizeof(path), | |
1704 | "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/", | |
1705 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
1706 | vdev->host.function); | |
1707 | if (stat(path, &st) < 0) { | |
1708 | error_report("vfio: error: no such host device: %s\n", path); | |
1709 | return -errno; | |
1710 | } | |
1711 | ||
1712 | strncat(path, "iommu_group", sizeof(path) - strlen(path) - 1); | |
1713 | ||
1714 | len = readlink(path, iommu_group_path, PATH_MAX); | |
1715 | if (len <= 0) { | |
1716 | error_report("vfio: error no iommu_group for device\n"); | |
1717 | return -errno; | |
1718 | } | |
1719 | ||
1720 | iommu_group_path[len] = 0; | |
1721 | group_name = basename(iommu_group_path); | |
1722 | ||
1723 | if (sscanf(group_name, "%d", &groupid) != 1) { | |
1724 | error_report("vfio: error reading %s: %m\n", path); | |
1725 | return -errno; | |
1726 | } | |
1727 | ||
1728 | DPRINTF("%s(%04x:%02x:%02x.%x) group %d\n", __func__, vdev->host.domain, | |
1729 | vdev->host.bus, vdev->host.slot, vdev->host.function, groupid); | |
1730 | ||
1731 | group = vfio_get_group(groupid); | |
1732 | if (!group) { | |
1733 | error_report("vfio: failed to get group %d\n", groupid); | |
1734 | return -ENOENT; | |
1735 | } | |
1736 | ||
1737 | snprintf(path, sizeof(path), "%04x:%02x:%02x.%01x", | |
1738 | vdev->host.domain, vdev->host.bus, vdev->host.slot, | |
1739 | vdev->host.function); | |
1740 | ||
1741 | QLIST_FOREACH(pvdev, &group->device_list, next) { | |
1742 | if (pvdev->host.domain == vdev->host.domain && | |
1743 | pvdev->host.bus == vdev->host.bus && | |
1744 | pvdev->host.slot == vdev->host.slot && | |
1745 | pvdev->host.function == vdev->host.function) { | |
1746 | ||
1747 | error_report("vfio: error: device %s is already attached\n", path); | |
1748 | vfio_put_group(group); | |
1749 | return -EBUSY; | |
1750 | } | |
1751 | } | |
1752 | ||
1753 | ret = vfio_get_device(group, path, vdev); | |
1754 | if (ret) { | |
1755 | error_report("vfio: failed to get device %s\n", path); | |
1756 | vfio_put_group(group); | |
1757 | return ret; | |
1758 | } | |
1759 | ||
1760 | /* Get a copy of config space */ | |
1761 | ret = pread(vdev->fd, vdev->pdev.config, | |
1762 | MIN(pci_config_size(&vdev->pdev), vdev->config_size), | |
1763 | vdev->config_offset); | |
1764 | if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) { | |
1765 | ret = ret < 0 ? -errno : -EFAULT; | |
1766 | error_report("vfio: Failed to read device config space\n"); | |
1767 | goto out_put; | |
1768 | } | |
1769 | ||
1770 | /* | |
1771 | * Clear host resource mapping info. If we choose not to register a | |
1772 | * BAR, such as might be the case with the option ROM, we can get | |
1773 | * confusing, unwritable, residual addresses from the host here. | |
1774 | */ | |
1775 | memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24); | |
1776 | memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4); | |
1777 | ||
1778 | vfio_load_rom(vdev); | |
1779 | ||
1780 | ret = vfio_early_setup_msix(vdev); | |
1781 | if (ret) { | |
1782 | goto out_put; | |
1783 | } | |
1784 | ||
1785 | vfio_map_bars(vdev); | |
1786 | ||
1787 | ret = vfio_add_capabilities(vdev); | |
1788 | if (ret) { | |
1789 | goto out_teardown; | |
1790 | } | |
1791 | ||
1792 | if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) { | |
ea486926 AW |
1793 | vdev->intx.mmap_timer = qemu_new_timer_ms(vm_clock, |
1794 | vfio_intx_mmap_enable, vdev); | |
65501a74 AW |
1795 | ret = vfio_enable_intx(vdev); |
1796 | if (ret) { | |
1797 | goto out_teardown; | |
1798 | } | |
1799 | } | |
1800 | ||
1801 | return 0; | |
1802 | ||
1803 | out_teardown: | |
1804 | pci_device_set_intx_routing_notifier(&vdev->pdev, NULL); | |
1805 | vfio_teardown_msi(vdev); | |
1806 | vfio_unmap_bars(vdev); | |
1807 | out_put: | |
1808 | vfio_put_device(vdev); | |
1809 | vfio_put_group(group); | |
1810 | return ret; | |
1811 | } | |
1812 | ||
1813 | static void vfio_exitfn(PCIDevice *pdev) | |
1814 | { | |
1815 | VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev); | |
1816 | VFIOGroup *group = vdev->group; | |
1817 | ||
1818 | pci_device_set_intx_routing_notifier(&vdev->pdev, NULL); | |
1819 | vfio_disable_interrupts(vdev); | |
ea486926 AW |
1820 | if (vdev->intx.mmap_timer) { |
1821 | qemu_free_timer(vdev->intx.mmap_timer); | |
1822 | } | |
65501a74 AW |
1823 | vfio_teardown_msi(vdev); |
1824 | vfio_unmap_bars(vdev); | |
1825 | vfio_put_device(vdev); | |
1826 | vfio_put_group(group); | |
1827 | } | |
1828 | ||
1829 | static void vfio_pci_reset(DeviceState *dev) | |
1830 | { | |
1831 | PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev); | |
1832 | VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev); | |
1833 | ||
1834 | if (!vdev->reset_works) { | |
1835 | return; | |
1836 | } | |
1837 | ||
1838 | if (ioctl(vdev->fd, VFIO_DEVICE_RESET)) { | |
1839 | error_report("vfio: Error unable to reset physical device " | |
1840 | "(%04x:%02x:%02x.%x): %m\n", vdev->host.domain, | |
1841 | vdev->host.bus, vdev->host.slot, vdev->host.function); | |
1842 | } | |
1843 | } | |
1844 | ||
1845 | static Property vfio_pci_dev_properties[] = { | |
1846 | DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIODevice, host), | |
ea486926 AW |
1847 | DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIODevice, |
1848 | intx.mmap_timeout, 1100), | |
65501a74 AW |
1849 | /* |
1850 | * TODO - support passed fds... is this necessary? | |
1851 | * DEFINE_PROP_STRING("vfiofd", VFIODevice, vfiofd_name), | |
1852 | * DEFINE_PROP_STRING("vfiogroupfd, VFIODevice, vfiogroupfd_name), | |
1853 | */ | |
1854 | DEFINE_PROP_END_OF_LIST(), | |
1855 | }; | |
1856 | ||
1857 | ||
1858 | static void vfio_pci_dev_class_init(ObjectClass *klass, void *data) | |
1859 | { | |
1860 | DeviceClass *dc = DEVICE_CLASS(klass); | |
1861 | PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass); | |
1862 | ||
1863 | dc->reset = vfio_pci_reset; | |
1864 | dc->props = vfio_pci_dev_properties; | |
1865 | pdc->init = vfio_initfn; | |
1866 | pdc->exit = vfio_exitfn; | |
1867 | pdc->config_read = vfio_pci_read_config; | |
1868 | pdc->config_write = vfio_pci_write_config; | |
1869 | } | |
1870 | ||
1871 | static const TypeInfo vfio_pci_dev_info = { | |
1872 | .name = "vfio-pci", | |
1873 | .parent = TYPE_PCI_DEVICE, | |
1874 | .instance_size = sizeof(VFIODevice), | |
1875 | .class_init = vfio_pci_dev_class_init, | |
1876 | }; | |
1877 | ||
1878 | static void register_vfio_pci_dev_type(void) | |
1879 | { | |
1880 | type_register_static(&vfio_pci_dev_info); | |
1881 | } | |
1882 | ||
1883 | type_init(register_vfio_pci_dev_type) |