xref: /qemu/hw/vfio/iommufd.c (revision 73b49878) 
1 /*
2  * iommufd container backend
3  *
4  * Copyright (C) 2023 Intel Corporation.
5  * Copyright Red Hat, Inc. 2023
6  *
7  * Authors: Yi Liu <yi.l.liu@intel.com>
8  *          Eric Auger <eric.auger@redhat.com>
9  *
10  * SPDX-License-Identifier: GPL-2.0-or-later
11  */
12 
13 #include "qemu/osdep.h"
14 #include <sys/ioctl.h>
15 #include <linux/vfio.h>
16 #include <linux/iommufd.h>
17 
18 #include "hw/vfio/vfio-common.h"
19 #include "qemu/error-report.h"
20 #include "trace.h"
21 #include "qapi/error.h"
22 #include "sysemu/iommufd.h"
23 #include "hw/qdev-core.h"
24 #include "sysemu/reset.h"
25 #include "qemu/cutils.h"
26 #include "qemu/chardev_open.h"
27 #include "pci.h"
28 
29 static int iommufd_cdev_map(const VFIOContainerBase *bcontainer, hwaddr iova,
30                             ram_addr_t size, void *vaddr, bool readonly)
31 {
32     const VFIOIOMMUFDContainer *container =
33         container_of(bcontainer, VFIOIOMMUFDContainer, bcontainer);
34 
35     return iommufd_backend_map_dma(container->be,
36                                    container->ioas_id,
37                                    iova, size, vaddr, readonly);
38 }
39 
40 static int iommufd_cdev_unmap(const VFIOContainerBase *bcontainer,
41                               hwaddr iova, ram_addr_t size,
42                               IOMMUTLBEntry *iotlb)
43 {
44     const VFIOIOMMUFDContainer *container =
45         container_of(bcontainer, VFIOIOMMUFDContainer, bcontainer);
46 
47     /* TODO: Handle dma_unmap_bitmap with iotlb args (migration) */
48     return iommufd_backend_unmap_dma(container->be,
49                                      container->ioas_id, iova, size);
50 }
51 
52 static int iommufd_cdev_kvm_device_add(VFIODevice *vbasedev, Error **errp)
53 {
54     return vfio_kvm_device_add_fd(vbasedev->fd, errp);
55 }
56 
57 static void iommufd_cdev_kvm_device_del(VFIODevice *vbasedev)
58 {
59     Error *err = NULL;
60 
61     if (vfio_kvm_device_del_fd(vbasedev->fd, &err)) {
62         error_report_err(err);
63     }
64 }
65 
66 static int iommufd_cdev_connect_and_bind(VFIODevice *vbasedev, Error **errp)
67 {
68     IOMMUFDBackend *iommufd = vbasedev->iommufd;
69     struct vfio_device_bind_iommufd bind = {
70         .argsz = sizeof(bind),
71         .flags = 0,
72     };
73     int ret;
74 
75     ret = iommufd_backend_connect(iommufd, errp);
76     if (ret) {
77         return ret;
78     }
79 
80     /*
81      * Add device to kvm-vfio to be prepared for the tracking
82      * in KVM. Especially for some emulated devices, it requires
83      * to have kvm information in the device open.
84      */
85     ret = iommufd_cdev_kvm_device_add(vbasedev, errp);
86     if (ret) {
87         goto err_kvm_device_add;
88     }
89 
90     /* Bind device to iommufd */
91     bind.iommufd = iommufd->fd;
92     ret = ioctl(vbasedev->fd, VFIO_DEVICE_BIND_IOMMUFD, &bind);
93     if (ret) {
94         error_setg_errno(errp, errno, "error bind device fd=%d to iommufd=%d",
95                          vbasedev->fd, bind.iommufd);
96         goto err_bind;
97     }
98 
99     vbasedev->devid = bind.out_devid;
100     trace_iommufd_cdev_connect_and_bind(bind.iommufd, vbasedev->name,
101                                         vbasedev->fd, vbasedev->devid);
102     return ret;
103 err_bind:
104     iommufd_cdev_kvm_device_del(vbasedev);
105 err_kvm_device_add:
106     iommufd_backend_disconnect(iommufd);
107     return ret;
108 }
109 
110 static void iommufd_cdev_unbind_and_disconnect(VFIODevice *vbasedev)
111 {
112     /* Unbind is automatically conducted when device fd is closed */
113     iommufd_cdev_kvm_device_del(vbasedev);
114     iommufd_backend_disconnect(vbasedev->iommufd);
115 }
116 
117 static int iommufd_cdev_getfd(const char *sysfs_path, Error **errp)
118 {
119     long int ret = -ENOTTY;
120     char *path, *vfio_dev_path = NULL, *vfio_path = NULL;
121     DIR *dir = NULL;
122     struct dirent *dent;
123     gchar *contents;
124     gsize length;
125     int major, minor;
126     dev_t vfio_devt;
127 
128     path = g_strdup_printf("%s/vfio-dev", sysfs_path);
129     dir = opendir(path);
130     if (!dir) {
131         error_setg_errno(errp, errno, "couldn't open directory %s", path);
132         goto out_free_path;
133     }
134 
135     while ((dent = readdir(dir))) {
136         if (!strncmp(dent->d_name, "vfio", 4)) {
137             vfio_dev_path = g_strdup_printf("%s/%s/dev", path, dent->d_name);
138             break;
139         }
140     }
141 
142     if (!vfio_dev_path) {
143         error_setg(errp, "failed to find vfio-dev/vfioX/dev");
144         goto out_close_dir;
145     }
146 
147     if (!g_file_get_contents(vfio_dev_path, &contents, &length, NULL)) {
148         error_setg(errp, "failed to load \"%s\"", vfio_dev_path);
149         goto out_free_dev_path;
150     }
151 
152     if (sscanf(contents, "%d:%d", &major, &minor) != 2) {
153         error_setg(errp, "failed to get major:minor for \"%s\"", vfio_dev_path);
154         goto out_free_dev_path;
155     }
156     g_free(contents);
157     vfio_devt = makedev(major, minor);
158 
159     vfio_path = g_strdup_printf("/dev/vfio/devices/%s", dent->d_name);
160     ret = open_cdev(vfio_path, vfio_devt);
161     if (ret < 0) {
162         error_setg(errp, "Failed to open %s", vfio_path);
163     }
164 
165     trace_iommufd_cdev_getfd(vfio_path, ret);
166     g_free(vfio_path);
167 
168 out_free_dev_path:
169     g_free(vfio_dev_path);
170 out_close_dir:
171     closedir(dir);
172 out_free_path:
173     if (*errp) {
174         error_prepend(errp, VFIO_MSG_PREFIX, path);
175     }
176     g_free(path);
177 
178     return ret;
179 }
180 
181 static int iommufd_cdev_attach_ioas_hwpt(VFIODevice *vbasedev, uint32_t id,
182                                          Error **errp)
183 {
184     int ret, iommufd = vbasedev->iommufd->fd;
185     struct vfio_device_attach_iommufd_pt attach_data = {
186         .argsz = sizeof(attach_data),
187         .flags = 0,
188         .pt_id = id,
189     };
190 
191     /* Attach device to an IOAS or hwpt within iommufd */
192     ret = ioctl(vbasedev->fd, VFIO_DEVICE_ATTACH_IOMMUFD_PT, &attach_data);
193     if (ret) {
194         error_setg_errno(errp, errno,
195                          "[iommufd=%d] error attach %s (%d) to id=%d",
196                          iommufd, vbasedev->name, vbasedev->fd, id);
197     } else {
198         trace_iommufd_cdev_attach_ioas_hwpt(iommufd, vbasedev->name,
199                                             vbasedev->fd, id);
200     }
201     return ret;
202 }
203 
204 static int iommufd_cdev_detach_ioas_hwpt(VFIODevice *vbasedev, Error **errp)
205 {
206     int ret, iommufd = vbasedev->iommufd->fd;
207     struct vfio_device_detach_iommufd_pt detach_data = {
208         .argsz = sizeof(detach_data),
209         .flags = 0,
210     };
211 
212     ret = ioctl(vbasedev->fd, VFIO_DEVICE_DETACH_IOMMUFD_PT, &detach_data);
213     if (ret) {
214         error_setg_errno(errp, errno, "detach %s failed", vbasedev->name);
215     } else {
216         trace_iommufd_cdev_detach_ioas_hwpt(iommufd, vbasedev->name);
217     }
218     return ret;
219 }
220 
221 static int iommufd_cdev_attach_container(VFIODevice *vbasedev,
222                                          VFIOIOMMUFDContainer *container,
223                                          Error **errp)
224 {
225     return iommufd_cdev_attach_ioas_hwpt(vbasedev, container->ioas_id, errp);
226 }
227 
228 static void iommufd_cdev_detach_container(VFIODevice *vbasedev,
229                                           VFIOIOMMUFDContainer *container)
230 {
231     Error *err = NULL;
232 
233     if (iommufd_cdev_detach_ioas_hwpt(vbasedev, &err)) {
234         error_report_err(err);
235     }
236 }
237 
238 static void iommufd_cdev_container_destroy(VFIOIOMMUFDContainer *container)
239 {
240     VFIOContainerBase *bcontainer = &container->bcontainer;
241 
242     if (!QLIST_EMPTY(&bcontainer->device_list)) {
243         return;
244     }
245     memory_listener_unregister(&bcontainer->listener);
246     vfio_container_destroy(bcontainer);
247     iommufd_backend_free_id(container->be, container->ioas_id);
248     g_free(container);
249 }
250 
251 static int iommufd_cdev_ram_block_discard_disable(bool state)
252 {
253     /*
254      * We support coordinated discarding of RAM via the RamDiscardManager.
255      */
256     return ram_block_uncoordinated_discard_disable(state);
257 }
258 
259 static int iommufd_cdev_get_info_iova_range(VFIOIOMMUFDContainer *container,
260                                             uint32_t ioas_id, Error **errp)
261 {
262     VFIOContainerBase *bcontainer = &container->bcontainer;
263     struct iommu_ioas_iova_ranges *info;
264     struct iommu_iova_range *iova_ranges;
265     int ret, sz, fd = container->be->fd;
266 
267     info = g_malloc0(sizeof(*info));
268     info->size = sizeof(*info);
269     info->ioas_id = ioas_id;
270 
271     ret = ioctl(fd, IOMMU_IOAS_IOVA_RANGES, info);
272     if (ret && errno != EMSGSIZE) {
273         goto error;
274     }
275 
276     sz = info->num_iovas * sizeof(struct iommu_iova_range);
277     info = g_realloc(info, sizeof(*info) + sz);
278     info->allowed_iovas = (uintptr_t)(info + 1);
279 
280     ret = ioctl(fd, IOMMU_IOAS_IOVA_RANGES, info);
281     if (ret) {
282         goto error;
283     }
284 
285     iova_ranges = (struct iommu_iova_range *)(uintptr_t)info->allowed_iovas;
286 
287     for (int i = 0; i < info->num_iovas; i++) {
288         Range *range = g_new(Range, 1);
289 
290         range_set_bounds(range, iova_ranges[i].start, iova_ranges[i].last);
291         bcontainer->iova_ranges =
292             range_list_insert(bcontainer->iova_ranges, range);
293     }
294     bcontainer->pgsizes = info->out_iova_alignment;
295 
296     g_free(info);
297     return 0;
298 
299 error:
300     ret = -errno;
301     g_free(info);
302     error_setg_errno(errp, errno, "Cannot get IOVA ranges");
303     return ret;
304 }
305 
306 static int iommufd_cdev_attach(const char *name, VFIODevice *vbasedev,
307                                AddressSpace *as, Error **errp)
308 {
309     VFIOContainerBase *bcontainer;
310     VFIOIOMMUFDContainer *container;
311     VFIOAddressSpace *space;
312     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
313     int ret, devfd;
314     uint32_t ioas_id;
315     Error *err = NULL;
316     const VFIOIOMMUClass *iommufd_vioc =
317         VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_IOMMUFD));
318 
319     if (vbasedev->fd < 0) {
320         devfd = iommufd_cdev_getfd(vbasedev->sysfsdev, errp);
321         if (devfd < 0) {
322             return devfd;
323         }
324         vbasedev->fd = devfd;
325     } else {
326         devfd = vbasedev->fd;
327     }
328 
329     ret = iommufd_cdev_connect_and_bind(vbasedev, errp);
330     if (ret) {
331         goto err_connect_bind;
332     }
333 
334     space = vfio_get_address_space(as);
335 
336     /* try to attach to an existing container in this space */
337     QLIST_FOREACH(bcontainer, &space->containers, next) {
338         container = container_of(bcontainer, VFIOIOMMUFDContainer, bcontainer);
339         if (bcontainer->ops != iommufd_vioc ||
340             vbasedev->iommufd != container->be) {
341             continue;
342         }
343         if (iommufd_cdev_attach_container(vbasedev, container, &err)) {
344             const char *msg = error_get_pretty(err);
345 
346             trace_iommufd_cdev_fail_attach_existing_container(msg);
347             error_free(err);
348             err = NULL;
349         } else {
350             ret = iommufd_cdev_ram_block_discard_disable(true);
351             if (ret) {
352                 error_setg(errp,
353                               "Cannot set discarding of RAM broken (%d)", ret);
354                 goto err_discard_disable;
355             }
356             goto found_container;
357         }
358     }
359 
360     /* Need to allocate a new dedicated container */
361     ret = iommufd_backend_alloc_ioas(vbasedev->iommufd, &ioas_id, errp);
362     if (ret < 0) {
363         goto err_alloc_ioas;
364     }
365 
366     trace_iommufd_cdev_alloc_ioas(vbasedev->iommufd->fd, ioas_id);
367 
368     container = g_malloc0(sizeof(*container));
369     container->be = vbasedev->iommufd;
370     container->ioas_id = ioas_id;
371 
372     bcontainer = &container->bcontainer;
373     vfio_container_init(bcontainer, space, iommufd_vioc);
374     QLIST_INSERT_HEAD(&space->containers, bcontainer, next);
375 
376     ret = iommufd_cdev_attach_container(vbasedev, container, errp);
377     if (ret) {
378         goto err_attach_container;
379     }
380 
381     ret = iommufd_cdev_ram_block_discard_disable(true);
382     if (ret) {
383         goto err_discard_disable;
384     }
385 
386     ret = iommufd_cdev_get_info_iova_range(container, ioas_id, &err);
387     if (ret) {
388         error_append_hint(&err,
389                    "Fallback to default 64bit IOVA range and 4K page size\n");
390         warn_report_err(err);
391         err = NULL;
392         bcontainer->pgsizes = qemu_real_host_page_size();
393     }
394 
395     bcontainer->listener = vfio_memory_listener;
396     memory_listener_register(&bcontainer->listener, bcontainer->space->as);
397 
398     if (bcontainer->error) {
399         ret = -1;
400         error_propagate_prepend(errp, bcontainer->error,
401                                 "memory listener initialization failed: ");
402         goto err_listener_register;
403     }
404 
405     bcontainer->initialized = true;
406 
407 found_container:
408     ret = ioctl(devfd, VFIO_DEVICE_GET_INFO, &dev_info);
409     if (ret) {
410         error_setg_errno(errp, errno, "error getting device info");
411         goto err_listener_register;
412     }
413 
414     /*
415      * TODO: examine RAM_BLOCK_DISCARD stuff, should we do group level
416      * for discarding incompatibility check as well?
417      */
418     if (vbasedev->ram_block_discard_allowed) {
419         iommufd_cdev_ram_block_discard_disable(false);
420     }
421 
422     vbasedev->group = 0;
423     vbasedev->num_irqs = dev_info.num_irqs;
424     vbasedev->num_regions = dev_info.num_regions;
425     vbasedev->flags = dev_info.flags;
426     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
427     vbasedev->bcontainer = bcontainer;
428     QLIST_INSERT_HEAD(&bcontainer->device_list, vbasedev, container_next);
429     QLIST_INSERT_HEAD(&vfio_device_list, vbasedev, global_next);
430 
431     trace_iommufd_cdev_device_info(vbasedev->name, devfd, vbasedev->num_irqs,
432                                    vbasedev->num_regions, vbasedev->flags);
433     return 0;
434 
435 err_listener_register:
436     iommufd_cdev_ram_block_discard_disable(false);
437 err_discard_disable:
438     iommufd_cdev_detach_container(vbasedev, container);
439 err_attach_container:
440     iommufd_cdev_container_destroy(container);
441 err_alloc_ioas:
442     vfio_put_address_space(space);
443     iommufd_cdev_unbind_and_disconnect(vbasedev);
444 err_connect_bind:
445     close(vbasedev->fd);
446     return ret;
447 }
448 
449 static void iommufd_cdev_detach(VFIODevice *vbasedev)
450 {
451     VFIOContainerBase *bcontainer = vbasedev->bcontainer;
452     VFIOAddressSpace *space = bcontainer->space;
453     VFIOIOMMUFDContainer *container = container_of(bcontainer,
454                                                    VFIOIOMMUFDContainer,
455                                                    bcontainer);
456     QLIST_REMOVE(vbasedev, global_next);
457     QLIST_REMOVE(vbasedev, container_next);
458     vbasedev->bcontainer = NULL;
459 
460     if (!vbasedev->ram_block_discard_allowed) {
461         iommufd_cdev_ram_block_discard_disable(false);
462     }
463 
464     iommufd_cdev_detach_container(vbasedev, container);
465     iommufd_cdev_container_destroy(container);
466     vfio_put_address_space(space);
467 
468     iommufd_cdev_unbind_and_disconnect(vbasedev);
469     close(vbasedev->fd);
470 }
471 
472 static VFIODevice *iommufd_cdev_pci_find_by_devid(__u32 devid)
473 {
474     VFIODevice *vbasedev_iter;
475     const VFIOIOMMUClass *iommufd_vioc =
476         VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_IOMMUFD));
477 
478     QLIST_FOREACH(vbasedev_iter, &vfio_device_list, global_next) {
479         if (vbasedev_iter->bcontainer->ops != iommufd_vioc) {
480             continue;
481         }
482         if (devid == vbasedev_iter->devid) {
483             return vbasedev_iter;
484         }
485     }
486     return NULL;
487 }
488 
489 static VFIOPCIDevice *
490 iommufd_cdev_dep_get_realized_vpdev(struct vfio_pci_dependent_device *dep_dev,
491                                     VFIODevice *reset_dev)
492 {
493     VFIODevice *vbasedev_tmp;
494 
495     if (dep_dev->devid == reset_dev->devid ||
496         dep_dev->devid == VFIO_PCI_DEVID_OWNED) {
497         return NULL;
498     }
499 
500     vbasedev_tmp = iommufd_cdev_pci_find_by_devid(dep_dev->devid);
501     if (!vbasedev_tmp || !vbasedev_tmp->dev->realized ||
502         vbasedev_tmp->type != VFIO_DEVICE_TYPE_PCI) {
503         return NULL;
504     }
505 
506     return container_of(vbasedev_tmp, VFIOPCIDevice, vbasedev);
507 }
508 
509 static int iommufd_cdev_pci_hot_reset(VFIODevice *vbasedev, bool single)
510 {
511     VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
512     struct vfio_pci_hot_reset_info *info = NULL;
513     struct vfio_pci_dependent_device *devices;
514     struct vfio_pci_hot_reset *reset;
515     int ret, i;
516     bool multi = false;
517 
518     trace_vfio_pci_hot_reset(vdev->vbasedev.name, single ? "one" : "multi");
519 
520     if (!single) {
521         vfio_pci_pre_reset(vdev);
522     }
523     vdev->vbasedev.needs_reset = false;
524 
525     ret = vfio_pci_get_pci_hot_reset_info(vdev, &info);
526 
527     if (ret) {
528         goto out_single;
529     }
530 
531     assert(info->flags & VFIO_PCI_HOT_RESET_FLAG_DEV_ID);
532 
533     devices = &info->devices[0];
534 
535     if (!(info->flags & VFIO_PCI_HOT_RESET_FLAG_DEV_ID_OWNED)) {
536         if (!vdev->has_pm_reset) {
537             for (i = 0; i < info->count; i++) {
538                 if (devices[i].devid == VFIO_PCI_DEVID_NOT_OWNED) {
539                     error_report("vfio: Cannot reset device %s, "
540                                  "depends on device %04x:%02x:%02x.%x "
541                                  "which is not owned.",
542                                  vdev->vbasedev.name, devices[i].segment,
543                                  devices[i].bus, PCI_SLOT(devices[i].devfn),
544                                  PCI_FUNC(devices[i].devfn));
545                 }
546             }
547         }
548         ret = -EPERM;
549         goto out_single;
550     }
551 
552     trace_vfio_pci_hot_reset_has_dep_devices(vdev->vbasedev.name);
553 
554     for (i = 0; i < info->count; i++) {
555         VFIOPCIDevice *tmp;
556 
557         trace_iommufd_cdev_pci_hot_reset_dep_devices(devices[i].segment,
558                                                      devices[i].bus,
559                                                      PCI_SLOT(devices[i].devfn),
560                                                      PCI_FUNC(devices[i].devfn),
561                                                      devices[i].devid);
562 
563         /*
564          * If a VFIO cdev device is resettable, all the dependent devices
565          * are either bound to same iommufd or within same iommu_groups as
566          * one of the iommufd bound devices.
567          */
568         assert(devices[i].devid != VFIO_PCI_DEVID_NOT_OWNED);
569 
570         tmp = iommufd_cdev_dep_get_realized_vpdev(&devices[i], &vdev->vbasedev);
571         if (!tmp) {
572             continue;
573         }
574 
575         if (single) {
576             ret = -EINVAL;
577             goto out_single;
578         }
579         vfio_pci_pre_reset(tmp);
580         tmp->vbasedev.needs_reset = false;
581         multi = true;
582     }
583 
584     if (!single && !multi) {
585         ret = -EINVAL;
586         goto out_single;
587     }
588 
589     /* Use zero length array for hot reset with iommufd backend */
590     reset = g_malloc0(sizeof(*reset));
591     reset->argsz = sizeof(*reset);
592 
593      /* Bus reset! */
594     ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_PCI_HOT_RESET, reset);
595     g_free(reset);
596     if (ret) {
597         ret = -errno;
598     }
599 
600     trace_vfio_pci_hot_reset_result(vdev->vbasedev.name,
601                                     ret ? strerror(errno) : "Success");
602 
603     /* Re-enable INTx on affected devices */
604     for (i = 0; i < info->count; i++) {
605         VFIOPCIDevice *tmp;
606 
607         tmp = iommufd_cdev_dep_get_realized_vpdev(&devices[i], &vdev->vbasedev);
608         if (!tmp) {
609             continue;
610         }
611         vfio_pci_post_reset(tmp);
612     }
613 out_single:
614     if (!single) {
615         vfio_pci_post_reset(vdev);
616     }
617     g_free(info);
618 
619     return ret;
620 }
621 
622 static void vfio_iommu_iommufd_class_init(ObjectClass *klass, void *data)
623 {
624     VFIOIOMMUClass *vioc = VFIO_IOMMU_CLASS(klass);
625 
626     vioc->dma_map = iommufd_cdev_map;
627     vioc->dma_unmap = iommufd_cdev_unmap;
628     vioc->attach_device = iommufd_cdev_attach;
629     vioc->detach_device = iommufd_cdev_detach;
630     vioc->pci_hot_reset = iommufd_cdev_pci_hot_reset;
631 };
632 
633 static const TypeInfo types[] = {
634     {
635         .name = TYPE_VFIO_IOMMU_IOMMUFD,
636         .parent = TYPE_VFIO_IOMMU,
637         .class_init = vfio_iommu_iommufd_class_init,
638     },
639 };
640 
641 DEFINE_TYPES(types)
642