xref: /openbmc/qemu/hw/vfio/migration.c (revision e50a24ea)
1 /*
2  * Migration support for VFIO devices
3  *
4  * Copyright NVIDIA, Inc. 2020
5  *
6  * This work is licensed under the terms of the GNU GPL, version 2. See
7  * the COPYING file in the top-level directory.
8  */
9 
10 #include "qemu/osdep.h"
11 #include "qemu/main-loop.h"
12 #include "qemu/cutils.h"
13 #include "qemu/units.h"
14 #include "qemu/error-report.h"
15 #include <linux/vfio.h>
16 #include <sys/ioctl.h>
17 
18 #include "sysemu/runstate.h"
19 #include "hw/vfio/vfio-common.h"
20 #include "migration/misc.h"
21 #include "migration/savevm.h"
22 #include "migration/vmstate.h"
23 #include "migration/qemu-file.h"
24 #include "migration/register.h"
25 #include "migration/blocker.h"
26 #include "qapi/error.h"
27 #include "qapi/qapi-events-vfio.h"
28 #include "exec/ramlist.h"
29 #include "exec/ram_addr.h"
30 #include "pci.h"
31 #include "trace.h"
32 #include "hw/hw.h"
33 
34 /*
35  * Flags to be used as unique delimiters for VFIO devices in the migration
36  * stream. These flags are composed as:
37  * 0xffffffff => MSB 32-bit all 1s
38  * 0xef10     => Magic ID, represents emulated (virtual) function IO
39  * 0x0000     => 16-bits reserved for flags
40  *
41  * The beginning of state information is marked by _DEV_CONFIG_STATE,
42  * _DEV_SETUP_STATE, or _DEV_DATA_STATE, respectively. The end of a
43  * certain state information is marked by _END_OF_STATE.
44  */
45 #define VFIO_MIG_FLAG_END_OF_STATE      (0xffffffffef100001ULL)
46 #define VFIO_MIG_FLAG_DEV_CONFIG_STATE  (0xffffffffef100002ULL)
47 #define VFIO_MIG_FLAG_DEV_SETUP_STATE   (0xffffffffef100003ULL)
48 #define VFIO_MIG_FLAG_DEV_DATA_STATE    (0xffffffffef100004ULL)
49 #define VFIO_MIG_FLAG_DEV_INIT_DATA_SENT (0xffffffffef100005ULL)
50 
51 /*
52  * This is an arbitrary size based on migration of mlx5 devices, where typically
53  * total device migration size is on the order of 100s of MB. Testing with
54  * larger values, e.g. 128MB and 1GB, did not show a performance improvement.
55  */
56 #define VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE (1 * MiB)
57 
58 static int64_t bytes_transferred;
59 
60 static const char *mig_state_to_str(enum vfio_device_mig_state state)
61 {
62     switch (state) {
63     case VFIO_DEVICE_STATE_ERROR:
64         return "ERROR";
65     case VFIO_DEVICE_STATE_STOP:
66         return "STOP";
67     case VFIO_DEVICE_STATE_RUNNING:
68         return "RUNNING";
69     case VFIO_DEVICE_STATE_STOP_COPY:
70         return "STOP_COPY";
71     case VFIO_DEVICE_STATE_RESUMING:
72         return "RESUMING";
73     case VFIO_DEVICE_STATE_RUNNING_P2P:
74         return "RUNNING_P2P";
75     case VFIO_DEVICE_STATE_PRE_COPY:
76         return "PRE_COPY";
77     case VFIO_DEVICE_STATE_PRE_COPY_P2P:
78         return "PRE_COPY_P2P";
79     default:
80         return "UNKNOWN STATE";
81     }
82 }
83 
84 static QapiVfioMigrationState
85 mig_state_to_qapi_state(enum vfio_device_mig_state state)
86 {
87     switch (state) {
88     case VFIO_DEVICE_STATE_STOP:
89         return QAPI_VFIO_MIGRATION_STATE_STOP;
90     case VFIO_DEVICE_STATE_RUNNING:
91         return QAPI_VFIO_MIGRATION_STATE_RUNNING;
92     case VFIO_DEVICE_STATE_STOP_COPY:
93         return QAPI_VFIO_MIGRATION_STATE_STOP_COPY;
94     case VFIO_DEVICE_STATE_RESUMING:
95         return QAPI_VFIO_MIGRATION_STATE_RESUMING;
96     case VFIO_DEVICE_STATE_RUNNING_P2P:
97         return QAPI_VFIO_MIGRATION_STATE_RUNNING_P2P;
98     case VFIO_DEVICE_STATE_PRE_COPY:
99         return QAPI_VFIO_MIGRATION_STATE_PRE_COPY;
100     case VFIO_DEVICE_STATE_PRE_COPY_P2P:
101         return QAPI_VFIO_MIGRATION_STATE_PRE_COPY_P2P;
102     default:
103         g_assert_not_reached();
104     }
105 }
106 
107 static void vfio_migration_send_event(VFIODevice *vbasedev)
108 {
109     VFIOMigration *migration = vbasedev->migration;
110     DeviceState *dev = vbasedev->dev;
111     g_autofree char *qom_path = NULL;
112     Object *obj;
113 
114     if (!vbasedev->migration_events) {
115         return;
116     }
117 
118     g_assert(vbasedev->ops->vfio_get_object);
119     obj = vbasedev->ops->vfio_get_object(vbasedev);
120     g_assert(obj);
121     qom_path = object_get_canonical_path(obj);
122 
123     qapi_event_send_vfio_migration(
124         dev->id, qom_path, mig_state_to_qapi_state(migration->device_state));
125 }
126 
127 static void vfio_migration_set_device_state(VFIODevice *vbasedev,
128                                             enum vfio_device_mig_state state)
129 {
130     VFIOMigration *migration = vbasedev->migration;
131 
132     trace_vfio_migration_set_device_state(vbasedev->name,
133                                           mig_state_to_str(state));
134 
135     migration->device_state = state;
136     vfio_migration_send_event(vbasedev);
137 }
138 
139 static int vfio_migration_set_state(VFIODevice *vbasedev,
140                                     enum vfio_device_mig_state new_state,
141                                     enum vfio_device_mig_state recover_state,
142                                     Error **errp)
143 {
144     VFIOMigration *migration = vbasedev->migration;
145     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
146                               sizeof(struct vfio_device_feature_mig_state),
147                               sizeof(uint64_t))] = {};
148     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
149     struct vfio_device_feature_mig_state *mig_state =
150         (struct vfio_device_feature_mig_state *)feature->data;
151     int ret;
152     g_autofree char *error_prefix =
153         g_strdup_printf("%s: Failed setting device state to %s.",
154                         vbasedev->name, mig_state_to_str(new_state));
155 
156     trace_vfio_migration_set_state(vbasedev->name, mig_state_to_str(new_state),
157                                    mig_state_to_str(recover_state));
158 
159     if (new_state == migration->device_state) {
160         return 0;
161     }
162 
163     feature->argsz = sizeof(buf);
164     feature->flags =
165         VFIO_DEVICE_FEATURE_SET | VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE;
166     mig_state->device_state = new_state;
167     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
168         /* Try to set the device in some good state */
169         ret = -errno;
170 
171         if (recover_state == VFIO_DEVICE_STATE_ERROR) {
172             error_setg_errno(errp, errno,
173                              "%s Recover state is ERROR. Resetting device",
174                              error_prefix);
175 
176             goto reset_device;
177         }
178 
179         error_setg_errno(errp, errno,
180                          "%s Setting device in recover state %s",
181                          error_prefix, mig_state_to_str(recover_state));
182 
183         mig_state->device_state = recover_state;
184         if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
185             ret = -errno;
186             /*
187              * If setting the device in recover state fails, report
188              * the error here and propagate the first error.
189              */
190             error_report(
191                 "%s: Failed setting device in recover state, err: %s. Resetting device",
192                          vbasedev->name, strerror(errno));
193 
194             goto reset_device;
195         }
196 
197         vfio_migration_set_device_state(vbasedev, recover_state);
198 
199         return ret;
200     }
201 
202     vfio_migration_set_device_state(vbasedev, new_state);
203     if (mig_state->data_fd != -1) {
204         if (migration->data_fd != -1) {
205             /*
206              * This can happen if the device is asynchronously reset and
207              * terminates a data transfer.
208              */
209             error_setg(errp, "%s: data_fd out of sync", vbasedev->name);
210             close(mig_state->data_fd);
211 
212             return -EBADF;
213         }
214 
215         migration->data_fd = mig_state->data_fd;
216     }
217 
218     return 0;
219 
220 reset_device:
221     if (ioctl(vbasedev->fd, VFIO_DEVICE_RESET)) {
222         hw_error("%s: Failed resetting device, err: %s", vbasedev->name,
223                  strerror(errno));
224     }
225 
226     vfio_migration_set_device_state(vbasedev, VFIO_DEVICE_STATE_RUNNING);
227 
228     return ret;
229 }
230 
231 /*
232  * Some device state transitions require resetting the device if they fail.
233  * This function sets the device in new_state and resets the device if that
234  * fails. Reset is done by using ERROR as the recover state.
235  */
236 static int
237 vfio_migration_set_state_or_reset(VFIODevice *vbasedev,
238                                   enum vfio_device_mig_state new_state,
239                                   Error **errp)
240 {
241     return vfio_migration_set_state(vbasedev, new_state,
242                                     VFIO_DEVICE_STATE_ERROR, errp);
243 }
244 
245 static int vfio_load_buffer(QEMUFile *f, VFIODevice *vbasedev,
246                             uint64_t data_size)
247 {
248     VFIOMigration *migration = vbasedev->migration;
249     int ret;
250 
251     ret = qemu_file_get_to_fd(f, migration->data_fd, data_size);
252     trace_vfio_load_state_device_data(vbasedev->name, data_size, ret);
253 
254     return ret;
255 }
256 
257 static int vfio_save_device_config_state(QEMUFile *f, void *opaque,
258                                          Error **errp)
259 {
260     VFIODevice *vbasedev = opaque;
261     int ret;
262 
263     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_CONFIG_STATE);
264 
265     if (vbasedev->ops && vbasedev->ops->vfio_save_config) {
266         ret = vbasedev->ops->vfio_save_config(vbasedev, f, errp);
267         if (ret) {
268             return ret;
269         }
270     }
271 
272     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
273 
274     trace_vfio_save_device_config_state(vbasedev->name);
275 
276     ret = qemu_file_get_error(f);
277     if (ret < 0) {
278         error_setg_errno(errp, -ret, "Failed to save state");
279     }
280     return ret;
281 }
282 
283 static int vfio_load_device_config_state(QEMUFile *f, void *opaque)
284 {
285     VFIODevice *vbasedev = opaque;
286     uint64_t data;
287 
288     if (vbasedev->ops && vbasedev->ops->vfio_load_config) {
289         int ret;
290 
291         ret = vbasedev->ops->vfio_load_config(vbasedev, f);
292         if (ret) {
293             error_report("%s: Failed to load device config space",
294                          vbasedev->name);
295             return ret;
296         }
297     }
298 
299     data = qemu_get_be64(f);
300     if (data != VFIO_MIG_FLAG_END_OF_STATE) {
301         error_report("%s: Failed loading device config space, "
302                      "end flag incorrect 0x%"PRIx64, vbasedev->name, data);
303         return -EINVAL;
304     }
305 
306     trace_vfio_load_device_config_state(vbasedev->name);
307     return qemu_file_get_error(f);
308 }
309 
310 static void vfio_migration_cleanup(VFIODevice *vbasedev)
311 {
312     VFIOMigration *migration = vbasedev->migration;
313 
314     close(migration->data_fd);
315     migration->data_fd = -1;
316 }
317 
318 static int vfio_query_stop_copy_size(VFIODevice *vbasedev,
319                                      uint64_t *stop_copy_size)
320 {
321     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
322                               sizeof(struct vfio_device_feature_mig_data_size),
323                               sizeof(uint64_t))] = {};
324     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
325     struct vfio_device_feature_mig_data_size *mig_data_size =
326         (struct vfio_device_feature_mig_data_size *)feature->data;
327 
328     feature->argsz = sizeof(buf);
329     feature->flags =
330         VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_MIG_DATA_SIZE;
331 
332     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
333         return -errno;
334     }
335 
336     *stop_copy_size = mig_data_size->stop_copy_length;
337 
338     return 0;
339 }
340 
341 static int vfio_query_precopy_size(VFIOMigration *migration)
342 {
343     struct vfio_precopy_info precopy = {
344         .argsz = sizeof(precopy),
345     };
346 
347     migration->precopy_init_size = 0;
348     migration->precopy_dirty_size = 0;
349 
350     if (ioctl(migration->data_fd, VFIO_MIG_GET_PRECOPY_INFO, &precopy)) {
351         return -errno;
352     }
353 
354     migration->precopy_init_size = precopy.initial_bytes;
355     migration->precopy_dirty_size = precopy.dirty_bytes;
356 
357     return 0;
358 }
359 
360 /* Returns the size of saved data on success and -errno on error */
361 static ssize_t vfio_save_block(QEMUFile *f, VFIOMigration *migration)
362 {
363     ssize_t data_size;
364 
365     data_size = read(migration->data_fd, migration->data_buffer,
366                      migration->data_buffer_size);
367     if (data_size < 0) {
368         /*
369          * Pre-copy emptied all the device state for now. For more information,
370          * please refer to the Linux kernel VFIO uAPI.
371          */
372         if (errno == ENOMSG) {
373             if (!migration->event_precopy_empty_hit) {
374                 trace_vfio_save_block_precopy_empty_hit(migration->vbasedev->name);
375                 migration->event_precopy_empty_hit = true;
376             }
377             return 0;
378         }
379 
380         return -errno;
381     }
382     if (data_size == 0) {
383         return 0;
384     }
385 
386     /* Non-empty read: re-arm the trace event */
387     migration->event_precopy_empty_hit = false;
388 
389     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_DATA_STATE);
390     qemu_put_be64(f, data_size);
391     qemu_put_buffer(f, migration->data_buffer, data_size);
392     bytes_transferred += data_size;
393 
394     trace_vfio_save_block(migration->vbasedev->name, data_size);
395 
396     return qemu_file_get_error(f) ?: data_size;
397 }
398 
399 static void vfio_update_estimated_pending_data(VFIOMigration *migration,
400                                                uint64_t data_size)
401 {
402     if (!data_size) {
403         /*
404          * Pre-copy emptied all the device state for now, update estimated sizes
405          * accordingly.
406          */
407         migration->precopy_init_size = 0;
408         migration->precopy_dirty_size = 0;
409 
410         return;
411     }
412 
413     if (migration->precopy_init_size) {
414         uint64_t init_size = MIN(migration->precopy_init_size, data_size);
415 
416         migration->precopy_init_size -= init_size;
417         data_size -= init_size;
418     }
419 
420     migration->precopy_dirty_size -= MIN(migration->precopy_dirty_size,
421                                          data_size);
422 }
423 
424 static bool vfio_precopy_supported(VFIODevice *vbasedev)
425 {
426     VFIOMigration *migration = vbasedev->migration;
427 
428     return migration->mig_flags & VFIO_MIGRATION_PRE_COPY;
429 }
430 
431 /* ---------------------------------------------------------------------- */
432 
433 static int vfio_save_prepare(void *opaque, Error **errp)
434 {
435     VFIODevice *vbasedev = opaque;
436 
437     /*
438      * Snapshot doesn't use postcopy nor background snapshot, so allow snapshot
439      * even if they are on.
440      */
441     if (runstate_check(RUN_STATE_SAVE_VM)) {
442         return 0;
443     }
444 
445     if (migrate_postcopy_ram()) {
446         error_setg(
447             errp, "%s: VFIO migration is not supported with postcopy migration",
448             vbasedev->name);
449         return -EOPNOTSUPP;
450     }
451 
452     if (migrate_background_snapshot()) {
453         error_setg(
454             errp,
455             "%s: VFIO migration is not supported with background snapshot",
456             vbasedev->name);
457         return -EOPNOTSUPP;
458     }
459 
460     return 0;
461 }
462 
463 static int vfio_save_setup(QEMUFile *f, void *opaque, Error **errp)
464 {
465     VFIODevice *vbasedev = opaque;
466     VFIOMigration *migration = vbasedev->migration;
467     uint64_t stop_copy_size = VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE;
468     int ret;
469 
470     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_SETUP_STATE);
471 
472     vfio_query_stop_copy_size(vbasedev, &stop_copy_size);
473     migration->data_buffer_size = MIN(VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE,
474                                       stop_copy_size);
475     migration->data_buffer = g_try_malloc0(migration->data_buffer_size);
476     if (!migration->data_buffer) {
477         error_setg(errp, "%s: Failed to allocate migration data buffer",
478                    vbasedev->name);
479         return -ENOMEM;
480     }
481 
482     migration->event_save_iterate_started = false;
483     migration->event_precopy_empty_hit = false;
484 
485     if (vfio_precopy_supported(vbasedev)) {
486         switch (migration->device_state) {
487         case VFIO_DEVICE_STATE_RUNNING:
488             ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_PRE_COPY,
489                                            VFIO_DEVICE_STATE_RUNNING, errp);
490             if (ret) {
491                 return ret;
492             }
493 
494             vfio_query_precopy_size(migration);
495 
496             break;
497         case VFIO_DEVICE_STATE_STOP:
498             /* vfio_save_complete_precopy() will go to STOP_COPY */
499             break;
500         default:
501             error_setg(errp, "%s: Invalid device state %d", vbasedev->name,
502                        migration->device_state);
503             return -EINVAL;
504         }
505     }
506 
507     trace_vfio_save_setup(vbasedev->name, migration->data_buffer_size);
508 
509     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
510 
511     ret = qemu_file_get_error(f);
512     if (ret < 0) {
513         error_setg_errno(errp, -ret, "%s: save setup failed", vbasedev->name);
514     }
515 
516     return ret;
517 }
518 
519 static void vfio_save_cleanup(void *opaque)
520 {
521     VFIODevice *vbasedev = opaque;
522     VFIOMigration *migration = vbasedev->migration;
523     Error *local_err = NULL;
524     int ret;
525 
526     /*
527      * Changing device state from STOP_COPY to STOP can take time. Do it here,
528      * after migration has completed, so it won't increase downtime.
529      */
530     if (migration->device_state == VFIO_DEVICE_STATE_STOP_COPY) {
531         ret = vfio_migration_set_state_or_reset(vbasedev,
532                                                 VFIO_DEVICE_STATE_STOP,
533                                                 &local_err);
534         if (ret) {
535             error_report_err(local_err);
536         }
537     }
538 
539     g_free(migration->data_buffer);
540     migration->data_buffer = NULL;
541     migration->precopy_init_size = 0;
542     migration->precopy_dirty_size = 0;
543     migration->initial_data_sent = false;
544     vfio_migration_cleanup(vbasedev);
545     trace_vfio_save_cleanup(vbasedev->name);
546 }
547 
548 static void vfio_state_pending_estimate(void *opaque, uint64_t *must_precopy,
549                                         uint64_t *can_postcopy)
550 {
551     VFIODevice *vbasedev = opaque;
552     VFIOMigration *migration = vbasedev->migration;
553 
554     if (!vfio_device_state_is_precopy(vbasedev)) {
555         return;
556     }
557 
558     *must_precopy +=
559         migration->precopy_init_size + migration->precopy_dirty_size;
560 
561     trace_vfio_state_pending_estimate(vbasedev->name, *must_precopy,
562                                       *can_postcopy,
563                                       migration->precopy_init_size,
564                                       migration->precopy_dirty_size);
565 }
566 
567 /*
568  * Migration size of VFIO devices can be as little as a few KBs or as big as
569  * many GBs. This value should be big enough to cover the worst case.
570  */
571 #define VFIO_MIG_STOP_COPY_SIZE (100 * GiB)
572 
573 static void vfio_state_pending_exact(void *opaque, uint64_t *must_precopy,
574                                      uint64_t *can_postcopy)
575 {
576     VFIODevice *vbasedev = opaque;
577     VFIOMigration *migration = vbasedev->migration;
578     uint64_t stop_copy_size = VFIO_MIG_STOP_COPY_SIZE;
579 
580     /*
581      * If getting pending migration size fails, VFIO_MIG_STOP_COPY_SIZE is
582      * reported so downtime limit won't be violated.
583      */
584     vfio_query_stop_copy_size(vbasedev, &stop_copy_size);
585     *must_precopy += stop_copy_size;
586 
587     if (vfio_device_state_is_precopy(vbasedev)) {
588         vfio_query_precopy_size(migration);
589     }
590 
591     trace_vfio_state_pending_exact(vbasedev->name, *must_precopy, *can_postcopy,
592                                    stop_copy_size, migration->precopy_init_size,
593                                    migration->precopy_dirty_size);
594 }
595 
596 static bool vfio_is_active_iterate(void *opaque)
597 {
598     VFIODevice *vbasedev = opaque;
599 
600     return vfio_device_state_is_precopy(vbasedev);
601 }
602 
603 /*
604  * Note about migration rate limiting: VFIO migration buffer size is currently
605  * limited to 1MB, so there is no need to check if migration rate exceeded (as
606  * in the worst case it will exceed by 1MB). However, if the buffer size is
607  * later changed to a bigger value, migration rate should be enforced here.
608  */
609 static int vfio_save_iterate(QEMUFile *f, void *opaque)
610 {
611     VFIODevice *vbasedev = opaque;
612     VFIOMigration *migration = vbasedev->migration;
613     ssize_t data_size;
614 
615     if (!migration->event_save_iterate_started) {
616         trace_vfio_save_iterate_start(vbasedev->name);
617         migration->event_save_iterate_started = true;
618     }
619 
620     data_size = vfio_save_block(f, migration);
621     if (data_size < 0) {
622         return data_size;
623     }
624 
625     vfio_update_estimated_pending_data(migration, data_size);
626 
627     if (migrate_switchover_ack() && !migration->precopy_init_size &&
628         !migration->initial_data_sent) {
629         qemu_put_be64(f, VFIO_MIG_FLAG_DEV_INIT_DATA_SENT);
630         migration->initial_data_sent = true;
631     } else {
632         qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
633     }
634 
635     trace_vfio_save_iterate(vbasedev->name, migration->precopy_init_size,
636                             migration->precopy_dirty_size);
637 
638     return !migration->precopy_init_size && !migration->precopy_dirty_size;
639 }
640 
641 static int vfio_save_complete_precopy(QEMUFile *f, void *opaque)
642 {
643     VFIODevice *vbasedev = opaque;
644     ssize_t data_size;
645     int ret;
646     Error *local_err = NULL;
647 
648     trace_vfio_save_complete_precopy_start(vbasedev->name);
649 
650     /* We reach here with device state STOP or STOP_COPY only */
651     ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_STOP_COPY,
652                                    VFIO_DEVICE_STATE_STOP, &local_err);
653     if (ret) {
654         error_report_err(local_err);
655         return ret;
656     }
657 
658     do {
659         data_size = vfio_save_block(f, vbasedev->migration);
660         if (data_size < 0) {
661             return data_size;
662         }
663     } while (data_size);
664 
665     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
666     ret = qemu_file_get_error(f);
667 
668     trace_vfio_save_complete_precopy(vbasedev->name, ret);
669 
670     return ret;
671 }
672 
673 static void vfio_save_state(QEMUFile *f, void *opaque)
674 {
675     VFIODevice *vbasedev = opaque;
676     Error *local_err = NULL;
677     int ret;
678 
679     ret = vfio_save_device_config_state(f, opaque, &local_err);
680     if (ret) {
681         error_prepend(&local_err,
682                       "vfio: Failed to save device config space of %s - ",
683                       vbasedev->name);
684         qemu_file_set_error_obj(f, ret, local_err);
685     }
686 }
687 
688 static int vfio_load_setup(QEMUFile *f, void *opaque, Error **errp)
689 {
690     VFIODevice *vbasedev = opaque;
691 
692     return vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_RESUMING,
693                                     vbasedev->migration->device_state, errp);
694 }
695 
696 static int vfio_load_cleanup(void *opaque)
697 {
698     VFIODevice *vbasedev = opaque;
699 
700     vfio_migration_cleanup(vbasedev);
701     trace_vfio_load_cleanup(vbasedev->name);
702 
703     return 0;
704 }
705 
706 static int vfio_load_state(QEMUFile *f, void *opaque, int version_id)
707 {
708     VFIODevice *vbasedev = opaque;
709     int ret = 0;
710     uint64_t data;
711 
712     data = qemu_get_be64(f);
713     while (data != VFIO_MIG_FLAG_END_OF_STATE) {
714 
715         trace_vfio_load_state(vbasedev->name, data);
716 
717         switch (data) {
718         case VFIO_MIG_FLAG_DEV_CONFIG_STATE:
719         {
720             return vfio_load_device_config_state(f, opaque);
721         }
722         case VFIO_MIG_FLAG_DEV_SETUP_STATE:
723         {
724             data = qemu_get_be64(f);
725             if (data == VFIO_MIG_FLAG_END_OF_STATE) {
726                 return ret;
727             } else {
728                 error_report("%s: SETUP STATE: EOS not found 0x%"PRIx64,
729                              vbasedev->name, data);
730                 return -EINVAL;
731             }
732             break;
733         }
734         case VFIO_MIG_FLAG_DEV_DATA_STATE:
735         {
736             uint64_t data_size = qemu_get_be64(f);
737 
738             if (data_size) {
739                 ret = vfio_load_buffer(f, vbasedev, data_size);
740                 if (ret < 0) {
741                     return ret;
742                 }
743             }
744             break;
745         }
746         case VFIO_MIG_FLAG_DEV_INIT_DATA_SENT:
747         {
748             if (!vfio_precopy_supported(vbasedev) ||
749                 !migrate_switchover_ack()) {
750                 error_report("%s: Received INIT_DATA_SENT but switchover ack "
751                              "is not used", vbasedev->name);
752                 return -EINVAL;
753             }
754 
755             ret = qemu_loadvm_approve_switchover();
756             if (ret) {
757                 error_report(
758                     "%s: qemu_loadvm_approve_switchover failed, err=%d (%s)",
759                     vbasedev->name, ret, strerror(-ret));
760             }
761 
762             return ret;
763         }
764         default:
765             error_report("%s: Unknown tag 0x%"PRIx64, vbasedev->name, data);
766             return -EINVAL;
767         }
768 
769         data = qemu_get_be64(f);
770         ret = qemu_file_get_error(f);
771         if (ret) {
772             return ret;
773         }
774     }
775     return ret;
776 }
777 
778 static bool vfio_switchover_ack_needed(void *opaque)
779 {
780     VFIODevice *vbasedev = opaque;
781 
782     return vfio_precopy_supported(vbasedev);
783 }
784 
785 static const SaveVMHandlers savevm_vfio_handlers = {
786     .save_prepare = vfio_save_prepare,
787     .save_setup = vfio_save_setup,
788     .save_cleanup = vfio_save_cleanup,
789     .state_pending_estimate = vfio_state_pending_estimate,
790     .state_pending_exact = vfio_state_pending_exact,
791     .is_active_iterate = vfio_is_active_iterate,
792     .save_live_iterate = vfio_save_iterate,
793     .save_live_complete_precopy = vfio_save_complete_precopy,
794     .save_state = vfio_save_state,
795     .load_setup = vfio_load_setup,
796     .load_cleanup = vfio_load_cleanup,
797     .load_state = vfio_load_state,
798     .switchover_ack_needed = vfio_switchover_ack_needed,
799 };
800 
801 /* ---------------------------------------------------------------------- */
802 
803 static void vfio_vmstate_change_prepare(void *opaque, bool running,
804                                         RunState state)
805 {
806     VFIODevice *vbasedev = opaque;
807     VFIOMigration *migration = vbasedev->migration;
808     enum vfio_device_mig_state new_state;
809     Error *local_err = NULL;
810     int ret;
811 
812     new_state = migration->device_state == VFIO_DEVICE_STATE_PRE_COPY ?
813                     VFIO_DEVICE_STATE_PRE_COPY_P2P :
814                     VFIO_DEVICE_STATE_RUNNING_P2P;
815 
816     ret = vfio_migration_set_state_or_reset(vbasedev, new_state, &local_err);
817     if (ret) {
818         /*
819          * Migration should be aborted in this case, but vm_state_notify()
820          * currently does not support reporting failures.
821          */
822         migration_file_set_error(ret, local_err);
823     }
824 
825     trace_vfio_vmstate_change_prepare(vbasedev->name, running,
826                                       RunState_str(state),
827                                       mig_state_to_str(new_state));
828 }
829 
830 static void vfio_vmstate_change(void *opaque, bool running, RunState state)
831 {
832     VFIODevice *vbasedev = opaque;
833     enum vfio_device_mig_state new_state;
834     Error *local_err = NULL;
835     int ret;
836 
837     if (running) {
838         new_state = VFIO_DEVICE_STATE_RUNNING;
839     } else {
840         new_state =
841             (vfio_device_state_is_precopy(vbasedev) &&
842              (state == RUN_STATE_FINISH_MIGRATE || state == RUN_STATE_PAUSED)) ?
843                 VFIO_DEVICE_STATE_STOP_COPY :
844                 VFIO_DEVICE_STATE_STOP;
845     }
846 
847     ret = vfio_migration_set_state_or_reset(vbasedev, new_state, &local_err);
848     if (ret) {
849         /*
850          * Migration should be aborted in this case, but vm_state_notify()
851          * currently does not support reporting failures.
852          */
853         migration_file_set_error(ret, local_err);
854     }
855 
856     trace_vfio_vmstate_change(vbasedev->name, running, RunState_str(state),
857                               mig_state_to_str(new_state));
858 }
859 
860 static int vfio_migration_state_notifier(NotifierWithReturn *notifier,
861                                          MigrationEvent *e, Error **errp)
862 {
863     VFIOMigration *migration = container_of(notifier, VFIOMigration,
864                                             migration_state);
865     VFIODevice *vbasedev = migration->vbasedev;
866     Error *local_err = NULL;
867     int ret;
868 
869     trace_vfio_migration_state_notifier(vbasedev->name, e->type);
870 
871     if (e->type == MIG_EVENT_PRECOPY_FAILED) {
872         /*
873          * MigrationNotifyFunc may not return an error code and an Error
874          * object for MIG_EVENT_PRECOPY_FAILED. Hence, report the error
875          * locally and ignore the errp argument.
876          */
877         ret = vfio_migration_set_state_or_reset(vbasedev,
878                                                 VFIO_DEVICE_STATE_RUNNING,
879                                                 &local_err);
880         if (ret) {
881             error_report_err(local_err);
882         }
883     }
884     return 0;
885 }
886 
887 static void vfio_migration_free(VFIODevice *vbasedev)
888 {
889     g_free(vbasedev->migration);
890     vbasedev->migration = NULL;
891 }
892 
893 static int vfio_migration_query_flags(VFIODevice *vbasedev, uint64_t *mig_flags)
894 {
895     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
896                                   sizeof(struct vfio_device_feature_migration),
897                               sizeof(uint64_t))] = {};
898     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
899     struct vfio_device_feature_migration *mig =
900         (struct vfio_device_feature_migration *)feature->data;
901 
902     feature->argsz = sizeof(buf);
903     feature->flags = VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_MIGRATION;
904     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
905         return -errno;
906     }
907 
908     *mig_flags = mig->flags;
909 
910     return 0;
911 }
912 
913 static bool vfio_dma_logging_supported(VFIODevice *vbasedev)
914 {
915     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
916                               sizeof(uint64_t))] = {};
917     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
918 
919     feature->argsz = sizeof(buf);
920     feature->flags = VFIO_DEVICE_FEATURE_PROBE |
921                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
922 
923     return !ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
924 }
925 
926 static int vfio_migration_init(VFIODevice *vbasedev)
927 {
928     int ret;
929     Object *obj;
930     VFIOMigration *migration;
931     char id[256] = "";
932     g_autofree char *path = NULL, *oid = NULL;
933     uint64_t mig_flags = 0;
934     VMChangeStateHandler *prepare_cb;
935 
936     if (!vbasedev->ops->vfio_get_object) {
937         return -EINVAL;
938     }
939 
940     obj = vbasedev->ops->vfio_get_object(vbasedev);
941     if (!obj) {
942         return -EINVAL;
943     }
944 
945     ret = vfio_migration_query_flags(vbasedev, &mig_flags);
946     if (ret) {
947         return ret;
948     }
949 
950     /* Basic migration functionality must be supported */
951     if (!(mig_flags & VFIO_MIGRATION_STOP_COPY)) {
952         return -EOPNOTSUPP;
953     }
954 
955     vbasedev->migration = g_new0(VFIOMigration, 1);
956     migration = vbasedev->migration;
957     migration->vbasedev = vbasedev;
958     migration->device_state = VFIO_DEVICE_STATE_RUNNING;
959     migration->data_fd = -1;
960     migration->mig_flags = mig_flags;
961 
962     vbasedev->dirty_pages_supported = vfio_dma_logging_supported(vbasedev);
963 
964     oid = vmstate_if_get_id(VMSTATE_IF(DEVICE(obj)));
965     if (oid) {
966         path = g_strdup_printf("%s/vfio", oid);
967     } else {
968         path = g_strdup("vfio");
969     }
970     strpadcpy(id, sizeof(id), path, '\0');
971 
972     register_savevm_live(id, VMSTATE_INSTANCE_ID_ANY, 1, &savevm_vfio_handlers,
973                          vbasedev);
974 
975     prepare_cb = migration->mig_flags & VFIO_MIGRATION_P2P ?
976                      vfio_vmstate_change_prepare :
977                      NULL;
978     migration->vm_state = qdev_add_vm_change_state_handler_full(
979         vbasedev->dev, vfio_vmstate_change, prepare_cb, vbasedev);
980     migration_add_notifier(&migration->migration_state,
981                            vfio_migration_state_notifier);
982 
983     return 0;
984 }
985 
986 static void vfio_migration_deinit(VFIODevice *vbasedev)
987 {
988     VFIOMigration *migration = vbasedev->migration;
989 
990     migration_remove_notifier(&migration->migration_state);
991     qemu_del_vm_change_state_handler(migration->vm_state);
992     unregister_savevm(VMSTATE_IF(vbasedev->dev), "vfio", vbasedev);
993     vfio_migration_free(vbasedev);
994     vfio_unblock_multiple_devices_migration();
995 }
996 
997 static int vfio_block_migration(VFIODevice *vbasedev, Error *err, Error **errp)
998 {
999     if (vbasedev->enable_migration == ON_OFF_AUTO_ON) {
1000         error_propagate(errp, err);
1001         return -EINVAL;
1002     }
1003 
1004     vbasedev->migration_blocker = error_copy(err);
1005     error_free(err);
1006 
1007     return migrate_add_blocker_normal(&vbasedev->migration_blocker, errp);
1008 }
1009 
1010 /* ---------------------------------------------------------------------- */
1011 
1012 int64_t vfio_mig_bytes_transferred(void)
1013 {
1014     return bytes_transferred;
1015 }
1016 
1017 void vfio_reset_bytes_transferred(void)
1018 {
1019     bytes_transferred = 0;
1020 }
1021 
1022 /*
1023  * Return true when either migration initialized or blocker registered.
1024  * Currently only return false when adding blocker fails which will
1025  * de-register vfio device.
1026  */
1027 bool vfio_migration_realize(VFIODevice *vbasedev, Error **errp)
1028 {
1029     Error *err = NULL;
1030     int ret;
1031 
1032     if (vbasedev->enable_migration == ON_OFF_AUTO_OFF) {
1033         error_setg(&err, "%s: Migration is disabled for VFIO device",
1034                    vbasedev->name);
1035         return !vfio_block_migration(vbasedev, err, errp);
1036     }
1037 
1038     ret = vfio_migration_init(vbasedev);
1039     if (ret) {
1040         if (ret == -ENOTTY) {
1041             error_setg(&err, "%s: VFIO migration is not supported in kernel",
1042                        vbasedev->name);
1043         } else {
1044             error_setg(&err,
1045                        "%s: Migration couldn't be initialized for VFIO device, "
1046                        "err: %d (%s)",
1047                        vbasedev->name, ret, strerror(-ret));
1048         }
1049 
1050         return !vfio_block_migration(vbasedev, err, errp);
1051     }
1052 
1053     if ((!vbasedev->dirty_pages_supported ||
1054          vbasedev->device_dirty_page_tracking == ON_OFF_AUTO_OFF) &&
1055         !vbasedev->iommu_dirty_tracking) {
1056         if (vbasedev->enable_migration == ON_OFF_AUTO_AUTO) {
1057             error_setg(&err,
1058                        "%s: VFIO device doesn't support device and "
1059                        "IOMMU dirty tracking", vbasedev->name);
1060             goto add_blocker;
1061         }
1062 
1063         warn_report("%s: VFIO device doesn't support device and "
1064                     "IOMMU dirty tracking", vbasedev->name);
1065     }
1066 
1067     ret = vfio_block_multiple_devices_migration(vbasedev, errp);
1068     if (ret) {
1069         goto out_deinit;
1070     }
1071 
1072     if (vfio_viommu_preset(vbasedev)) {
1073         error_setg(&err, "%s: Migration is currently not supported "
1074                    "with vIOMMU enabled", vbasedev->name);
1075         goto add_blocker;
1076     }
1077 
1078     trace_vfio_migration_realize(vbasedev->name);
1079     return true;
1080 
1081 add_blocker:
1082     ret = vfio_block_migration(vbasedev, err, errp);
1083 out_deinit:
1084     if (ret) {
1085         vfio_migration_deinit(vbasedev);
1086     }
1087     return !ret;
1088 }
1089 
1090 void vfio_migration_exit(VFIODevice *vbasedev)
1091 {
1092     if (vbasedev->migration) {
1093         vfio_migration_deinit(vbasedev);
1094     }
1095 
1096     migrate_del_blocker(&vbasedev->migration_blocker);
1097 }
1098