xref: /openbmc/qemu/hw/vfio/migration.c (revision 259ebed4)
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/migration.h"
21 #include "migration/options.h"
22 #include "migration/savevm.h"
23 #include "migration/vmstate.h"
24 #include "migration/qemu-file.h"
25 #include "migration/register.h"
26 #include "migration/blocker.h"
27 #include "migration/misc.h"
28 #include "qapi/error.h"
29 #include "exec/ramlist.h"
30 #include "exec/ram_addr.h"
31 #include "pci.h"
32 #include "trace.h"
33 #include "hw/hw.h"
34 
35 /*
36  * Flags to be used as unique delimiters for VFIO devices in the migration
37  * stream. These flags are composed as:
38  * 0xffffffff => MSB 32-bit all 1s
39  * 0xef10     => Magic ID, represents emulated (virtual) function IO
40  * 0x0000     => 16-bits reserved for flags
41  *
42  * The beginning of state information is marked by _DEV_CONFIG_STATE,
43  * _DEV_SETUP_STATE, or _DEV_DATA_STATE, respectively. The end of a
44  * certain state information is marked by _END_OF_STATE.
45  */
46 #define VFIO_MIG_FLAG_END_OF_STATE      (0xffffffffef100001ULL)
47 #define VFIO_MIG_FLAG_DEV_CONFIG_STATE  (0xffffffffef100002ULL)
48 #define VFIO_MIG_FLAG_DEV_SETUP_STATE   (0xffffffffef100003ULL)
49 #define VFIO_MIG_FLAG_DEV_DATA_STATE    (0xffffffffef100004ULL)
50 #define VFIO_MIG_FLAG_DEV_INIT_DATA_SENT (0xffffffffef100005ULL)
51 
52 /*
53  * This is an arbitrary size based on migration of mlx5 devices, where typically
54  * total device migration size is on the order of 100s of MB. Testing with
55  * larger values, e.g. 128MB and 1GB, did not show a performance improvement.
56  */
57 #define VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE (1 * MiB)
58 
59 static int64_t bytes_transferred;
60 
61 static const char *mig_state_to_str(enum vfio_device_mig_state state)
62 {
63     switch (state) {
64     case VFIO_DEVICE_STATE_ERROR:
65         return "ERROR";
66     case VFIO_DEVICE_STATE_STOP:
67         return "STOP";
68     case VFIO_DEVICE_STATE_RUNNING:
69         return "RUNNING";
70     case VFIO_DEVICE_STATE_STOP_COPY:
71         return "STOP_COPY";
72     case VFIO_DEVICE_STATE_RESUMING:
73         return "RESUMING";
74     case VFIO_DEVICE_STATE_RUNNING_P2P:
75         return "RUNNING_P2P";
76     case VFIO_DEVICE_STATE_PRE_COPY:
77         return "PRE_COPY";
78     case VFIO_DEVICE_STATE_PRE_COPY_P2P:
79         return "PRE_COPY_P2P";
80     default:
81         return "UNKNOWN STATE";
82     }
83 }
84 
85 static int vfio_migration_set_state(VFIODevice *vbasedev,
86                                     enum vfio_device_mig_state new_state,
87                                     enum vfio_device_mig_state recover_state)
88 {
89     VFIOMigration *migration = vbasedev->migration;
90     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
91                               sizeof(struct vfio_device_feature_mig_state),
92                               sizeof(uint64_t))] = {};
93     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
94     struct vfio_device_feature_mig_state *mig_state =
95         (struct vfio_device_feature_mig_state *)feature->data;
96     int ret;
97 
98     feature->argsz = sizeof(buf);
99     feature->flags =
100         VFIO_DEVICE_FEATURE_SET | VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE;
101     mig_state->device_state = new_state;
102     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
103         /* Try to set the device in some good state */
104         ret = -errno;
105 
106         if (recover_state == VFIO_DEVICE_STATE_ERROR) {
107             error_report("%s: Failed setting device state to %s, err: %s. "
108                          "Recover state is ERROR. Resetting device",
109                          vbasedev->name, mig_state_to_str(new_state),
110                          strerror(errno));
111 
112             goto reset_device;
113         }
114 
115         error_report(
116             "%s: Failed setting device state to %s, err: %s. Setting device in recover state %s",
117                      vbasedev->name, mig_state_to_str(new_state),
118                      strerror(errno), mig_state_to_str(recover_state));
119 
120         mig_state->device_state = recover_state;
121         if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
122             ret = -errno;
123             error_report(
124                 "%s: Failed setting device in recover state, err: %s. Resetting device",
125                          vbasedev->name, strerror(errno));
126 
127             goto reset_device;
128         }
129 
130         migration->device_state = recover_state;
131 
132         return ret;
133     }
134 
135     migration->device_state = new_state;
136     if (mig_state->data_fd != -1) {
137         if (migration->data_fd != -1) {
138             /*
139              * This can happen if the device is asynchronously reset and
140              * terminates a data transfer.
141              */
142             error_report("%s: data_fd out of sync", vbasedev->name);
143             close(mig_state->data_fd);
144 
145             return -EBADF;
146         }
147 
148         migration->data_fd = mig_state->data_fd;
149     }
150 
151     trace_vfio_migration_set_state(vbasedev->name, mig_state_to_str(new_state));
152 
153     return 0;
154 
155 reset_device:
156     if (ioctl(vbasedev->fd, VFIO_DEVICE_RESET)) {
157         hw_error("%s: Failed resetting device, err: %s", vbasedev->name,
158                  strerror(errno));
159     }
160 
161     migration->device_state = VFIO_DEVICE_STATE_RUNNING;
162 
163     return ret;
164 }
165 
166 static int vfio_load_buffer(QEMUFile *f, VFIODevice *vbasedev,
167                             uint64_t data_size)
168 {
169     VFIOMigration *migration = vbasedev->migration;
170     int ret;
171 
172     ret = qemu_file_get_to_fd(f, migration->data_fd, data_size);
173     trace_vfio_load_state_device_data(vbasedev->name, data_size, ret);
174 
175     return ret;
176 }
177 
178 static int vfio_save_device_config_state(QEMUFile *f, void *opaque)
179 {
180     VFIODevice *vbasedev = opaque;
181 
182     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_CONFIG_STATE);
183 
184     if (vbasedev->ops && vbasedev->ops->vfio_save_config) {
185         vbasedev->ops->vfio_save_config(vbasedev, f);
186     }
187 
188     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
189 
190     trace_vfio_save_device_config_state(vbasedev->name);
191 
192     return qemu_file_get_error(f);
193 }
194 
195 static int vfio_load_device_config_state(QEMUFile *f, void *opaque)
196 {
197     VFIODevice *vbasedev = opaque;
198     uint64_t data;
199 
200     if (vbasedev->ops && vbasedev->ops->vfio_load_config) {
201         int ret;
202 
203         ret = vbasedev->ops->vfio_load_config(vbasedev, f);
204         if (ret) {
205             error_report("%s: Failed to load device config space",
206                          vbasedev->name);
207             return ret;
208         }
209     }
210 
211     data = qemu_get_be64(f);
212     if (data != VFIO_MIG_FLAG_END_OF_STATE) {
213         error_report("%s: Failed loading device config space, "
214                      "end flag incorrect 0x%"PRIx64, vbasedev->name, data);
215         return -EINVAL;
216     }
217 
218     trace_vfio_load_device_config_state(vbasedev->name);
219     return qemu_file_get_error(f);
220 }
221 
222 static void vfio_migration_cleanup(VFIODevice *vbasedev)
223 {
224     VFIOMigration *migration = vbasedev->migration;
225 
226     close(migration->data_fd);
227     migration->data_fd = -1;
228 }
229 
230 static int vfio_query_stop_copy_size(VFIODevice *vbasedev,
231                                      uint64_t *stop_copy_size)
232 {
233     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
234                               sizeof(struct vfio_device_feature_mig_data_size),
235                               sizeof(uint64_t))] = {};
236     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
237     struct vfio_device_feature_mig_data_size *mig_data_size =
238         (struct vfio_device_feature_mig_data_size *)feature->data;
239 
240     feature->argsz = sizeof(buf);
241     feature->flags =
242         VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_MIG_DATA_SIZE;
243 
244     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
245         return -errno;
246     }
247 
248     *stop_copy_size = mig_data_size->stop_copy_length;
249 
250     return 0;
251 }
252 
253 static int vfio_query_precopy_size(VFIOMigration *migration)
254 {
255     struct vfio_precopy_info precopy = {
256         .argsz = sizeof(precopy),
257     };
258 
259     migration->precopy_init_size = 0;
260     migration->precopy_dirty_size = 0;
261 
262     if (ioctl(migration->data_fd, VFIO_MIG_GET_PRECOPY_INFO, &precopy)) {
263         return -errno;
264     }
265 
266     migration->precopy_init_size = precopy.initial_bytes;
267     migration->precopy_dirty_size = precopy.dirty_bytes;
268 
269     return 0;
270 }
271 
272 /* Returns the size of saved data on success and -errno on error */
273 static ssize_t vfio_save_block(QEMUFile *f, VFIOMigration *migration)
274 {
275     ssize_t data_size;
276 
277     data_size = read(migration->data_fd, migration->data_buffer,
278                      migration->data_buffer_size);
279     if (data_size < 0) {
280         /*
281          * Pre-copy emptied all the device state for now. For more information,
282          * please refer to the Linux kernel VFIO uAPI.
283          */
284         if (errno == ENOMSG) {
285             return 0;
286         }
287 
288         return -errno;
289     }
290     if (data_size == 0) {
291         return 0;
292     }
293 
294     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_DATA_STATE);
295     qemu_put_be64(f, data_size);
296     qemu_put_buffer(f, migration->data_buffer, data_size);
297     bytes_transferred += data_size;
298 
299     trace_vfio_save_block(migration->vbasedev->name, data_size);
300 
301     return qemu_file_get_error(f) ?: data_size;
302 }
303 
304 static void vfio_update_estimated_pending_data(VFIOMigration *migration,
305                                                uint64_t data_size)
306 {
307     if (!data_size) {
308         /*
309          * Pre-copy emptied all the device state for now, update estimated sizes
310          * accordingly.
311          */
312         migration->precopy_init_size = 0;
313         migration->precopy_dirty_size = 0;
314 
315         return;
316     }
317 
318     if (migration->precopy_init_size) {
319         uint64_t init_size = MIN(migration->precopy_init_size, data_size);
320 
321         migration->precopy_init_size -= init_size;
322         data_size -= init_size;
323     }
324 
325     migration->precopy_dirty_size -= MIN(migration->precopy_dirty_size,
326                                          data_size);
327 }
328 
329 static bool vfio_precopy_supported(VFIODevice *vbasedev)
330 {
331     VFIOMigration *migration = vbasedev->migration;
332 
333     return migration->mig_flags & VFIO_MIGRATION_PRE_COPY;
334 }
335 
336 /* ---------------------------------------------------------------------- */
337 
338 static int vfio_save_prepare(void *opaque, Error **errp)
339 {
340     VFIODevice *vbasedev = opaque;
341 
342     /*
343      * Snapshot doesn't use postcopy nor background snapshot, so allow snapshot
344      * even if they are on.
345      */
346     if (runstate_check(RUN_STATE_SAVE_VM)) {
347         return 0;
348     }
349 
350     if (migrate_postcopy_ram()) {
351         error_setg(
352             errp, "%s: VFIO migration is not supported with postcopy migration",
353             vbasedev->name);
354         return -EOPNOTSUPP;
355     }
356 
357     if (migrate_background_snapshot()) {
358         error_setg(
359             errp,
360             "%s: VFIO migration is not supported with background snapshot",
361             vbasedev->name);
362         return -EOPNOTSUPP;
363     }
364 
365     return 0;
366 }
367 
368 static int vfio_save_setup(QEMUFile *f, void *opaque)
369 {
370     VFIODevice *vbasedev = opaque;
371     VFIOMigration *migration = vbasedev->migration;
372     uint64_t stop_copy_size = VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE;
373 
374     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_SETUP_STATE);
375 
376     vfio_query_stop_copy_size(vbasedev, &stop_copy_size);
377     migration->data_buffer_size = MIN(VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE,
378                                       stop_copy_size);
379     migration->data_buffer = g_try_malloc0(migration->data_buffer_size);
380     if (!migration->data_buffer) {
381         error_report("%s: Failed to allocate migration data buffer",
382                      vbasedev->name);
383         return -ENOMEM;
384     }
385 
386     if (vfio_precopy_supported(vbasedev)) {
387         int ret;
388 
389         switch (migration->device_state) {
390         case VFIO_DEVICE_STATE_RUNNING:
391             ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_PRE_COPY,
392                                            VFIO_DEVICE_STATE_RUNNING);
393             if (ret) {
394                 return ret;
395             }
396 
397             vfio_query_precopy_size(migration);
398 
399             break;
400         case VFIO_DEVICE_STATE_STOP:
401             /* vfio_save_complete_precopy() will go to STOP_COPY */
402             break;
403         default:
404             return -EINVAL;
405         }
406     }
407 
408     trace_vfio_save_setup(vbasedev->name, migration->data_buffer_size);
409 
410     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
411 
412     return qemu_file_get_error(f);
413 }
414 
415 static void vfio_save_cleanup(void *opaque)
416 {
417     VFIODevice *vbasedev = opaque;
418     VFIOMigration *migration = vbasedev->migration;
419 
420     /*
421      * Changing device state from STOP_COPY to STOP can take time. Do it here,
422      * after migration has completed, so it won't increase downtime.
423      */
424     if (migration->device_state == VFIO_DEVICE_STATE_STOP_COPY) {
425         /*
426          * If setting the device in STOP state fails, the device should be
427          * reset. To do so, use ERROR state as a recover state.
428          */
429         vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_STOP,
430                                  VFIO_DEVICE_STATE_ERROR);
431     }
432 
433     g_free(migration->data_buffer);
434     migration->data_buffer = NULL;
435     migration->precopy_init_size = 0;
436     migration->precopy_dirty_size = 0;
437     migration->initial_data_sent = false;
438     vfio_migration_cleanup(vbasedev);
439     trace_vfio_save_cleanup(vbasedev->name);
440 }
441 
442 static void vfio_state_pending_estimate(void *opaque, uint64_t *must_precopy,
443                                         uint64_t *can_postcopy)
444 {
445     VFIODevice *vbasedev = opaque;
446     VFIOMigration *migration = vbasedev->migration;
447 
448     if (!vfio_device_state_is_precopy(vbasedev)) {
449         return;
450     }
451 
452     *must_precopy +=
453         migration->precopy_init_size + migration->precopy_dirty_size;
454 
455     trace_vfio_state_pending_estimate(vbasedev->name, *must_precopy,
456                                       *can_postcopy,
457                                       migration->precopy_init_size,
458                                       migration->precopy_dirty_size);
459 }
460 
461 /*
462  * Migration size of VFIO devices can be as little as a few KBs or as big as
463  * many GBs. This value should be big enough to cover the worst case.
464  */
465 #define VFIO_MIG_STOP_COPY_SIZE (100 * GiB)
466 
467 static void vfio_state_pending_exact(void *opaque, uint64_t *must_precopy,
468                                      uint64_t *can_postcopy)
469 {
470     VFIODevice *vbasedev = opaque;
471     VFIOMigration *migration = vbasedev->migration;
472     uint64_t stop_copy_size = VFIO_MIG_STOP_COPY_SIZE;
473 
474     /*
475      * If getting pending migration size fails, VFIO_MIG_STOP_COPY_SIZE is
476      * reported so downtime limit won't be violated.
477      */
478     vfio_query_stop_copy_size(vbasedev, &stop_copy_size);
479     *must_precopy += stop_copy_size;
480 
481     if (vfio_device_state_is_precopy(vbasedev)) {
482         vfio_query_precopy_size(migration);
483 
484         *must_precopy +=
485             migration->precopy_init_size + migration->precopy_dirty_size;
486     }
487 
488     trace_vfio_state_pending_exact(vbasedev->name, *must_precopy, *can_postcopy,
489                                    stop_copy_size, migration->precopy_init_size,
490                                    migration->precopy_dirty_size);
491 }
492 
493 static bool vfio_is_active_iterate(void *opaque)
494 {
495     VFIODevice *vbasedev = opaque;
496 
497     return vfio_device_state_is_precopy(vbasedev);
498 }
499 
500 static int vfio_save_iterate(QEMUFile *f, void *opaque)
501 {
502     VFIODevice *vbasedev = opaque;
503     VFIOMigration *migration = vbasedev->migration;
504     ssize_t data_size;
505 
506     data_size = vfio_save_block(f, migration);
507     if (data_size < 0) {
508         return data_size;
509     }
510 
511     vfio_update_estimated_pending_data(migration, data_size);
512 
513     if (migrate_switchover_ack() && !migration->precopy_init_size &&
514         !migration->initial_data_sent) {
515         qemu_put_be64(f, VFIO_MIG_FLAG_DEV_INIT_DATA_SENT);
516         migration->initial_data_sent = true;
517     } else {
518         qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
519     }
520 
521     trace_vfio_save_iterate(vbasedev->name, migration->precopy_init_size,
522                             migration->precopy_dirty_size);
523 
524     /*
525      * A VFIO device's pre-copy dirty_bytes is not guaranteed to reach zero.
526      * Return 1 so following handlers will not be potentially blocked.
527      */
528     return 1;
529 }
530 
531 static int vfio_save_complete_precopy(QEMUFile *f, void *opaque)
532 {
533     VFIODevice *vbasedev = opaque;
534     ssize_t data_size;
535     int ret;
536 
537     /* We reach here with device state STOP or STOP_COPY only */
538     ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_STOP_COPY,
539                                    VFIO_DEVICE_STATE_STOP);
540     if (ret) {
541         return ret;
542     }
543 
544     do {
545         data_size = vfio_save_block(f, vbasedev->migration);
546         if (data_size < 0) {
547             return data_size;
548         }
549     } while (data_size);
550 
551     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
552     ret = qemu_file_get_error(f);
553     if (ret) {
554         return ret;
555     }
556 
557     trace_vfio_save_complete_precopy(vbasedev->name, ret);
558 
559     return ret;
560 }
561 
562 static void vfio_save_state(QEMUFile *f, void *opaque)
563 {
564     VFIODevice *vbasedev = opaque;
565     int ret;
566 
567     ret = vfio_save_device_config_state(f, opaque);
568     if (ret) {
569         error_report("%s: Failed to save device config space",
570                      vbasedev->name);
571         qemu_file_set_error(f, ret);
572     }
573 }
574 
575 static int vfio_load_setup(QEMUFile *f, void *opaque)
576 {
577     VFIODevice *vbasedev = opaque;
578 
579     return vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_RESUMING,
580                                    vbasedev->migration->device_state);
581 }
582 
583 static int vfio_load_cleanup(void *opaque)
584 {
585     VFIODevice *vbasedev = opaque;
586 
587     vfio_migration_cleanup(vbasedev);
588     trace_vfio_load_cleanup(vbasedev->name);
589 
590     return 0;
591 }
592 
593 static int vfio_load_state(QEMUFile *f, void *opaque, int version_id)
594 {
595     VFIODevice *vbasedev = opaque;
596     int ret = 0;
597     uint64_t data;
598 
599     data = qemu_get_be64(f);
600     while (data != VFIO_MIG_FLAG_END_OF_STATE) {
601 
602         trace_vfio_load_state(vbasedev->name, data);
603 
604         switch (data) {
605         case VFIO_MIG_FLAG_DEV_CONFIG_STATE:
606         {
607             return vfio_load_device_config_state(f, opaque);
608         }
609         case VFIO_MIG_FLAG_DEV_SETUP_STATE:
610         {
611             data = qemu_get_be64(f);
612             if (data == VFIO_MIG_FLAG_END_OF_STATE) {
613                 return ret;
614             } else {
615                 error_report("%s: SETUP STATE: EOS not found 0x%"PRIx64,
616                              vbasedev->name, data);
617                 return -EINVAL;
618             }
619             break;
620         }
621         case VFIO_MIG_FLAG_DEV_DATA_STATE:
622         {
623             uint64_t data_size = qemu_get_be64(f);
624 
625             if (data_size) {
626                 ret = vfio_load_buffer(f, vbasedev, data_size);
627                 if (ret < 0) {
628                     return ret;
629                 }
630             }
631             break;
632         }
633         case VFIO_MIG_FLAG_DEV_INIT_DATA_SENT:
634         {
635             if (!vfio_precopy_supported(vbasedev) ||
636                 !migrate_switchover_ack()) {
637                 error_report("%s: Received INIT_DATA_SENT but switchover ack "
638                              "is not used", vbasedev->name);
639                 return -EINVAL;
640             }
641 
642             ret = qemu_loadvm_approve_switchover();
643             if (ret) {
644                 error_report(
645                     "%s: qemu_loadvm_approve_switchover failed, err=%d (%s)",
646                     vbasedev->name, ret, strerror(-ret));
647             }
648 
649             return ret;
650         }
651         default:
652             error_report("%s: Unknown tag 0x%"PRIx64, vbasedev->name, data);
653             return -EINVAL;
654         }
655 
656         data = qemu_get_be64(f);
657         ret = qemu_file_get_error(f);
658         if (ret) {
659             return ret;
660         }
661     }
662     return ret;
663 }
664 
665 static bool vfio_switchover_ack_needed(void *opaque)
666 {
667     VFIODevice *vbasedev = opaque;
668 
669     return vfio_precopy_supported(vbasedev);
670 }
671 
672 static const SaveVMHandlers savevm_vfio_handlers = {
673     .save_prepare = vfio_save_prepare,
674     .save_setup = vfio_save_setup,
675     .save_cleanup = vfio_save_cleanup,
676     .state_pending_estimate = vfio_state_pending_estimate,
677     .state_pending_exact = vfio_state_pending_exact,
678     .is_active_iterate = vfio_is_active_iterate,
679     .save_live_iterate = vfio_save_iterate,
680     .save_live_complete_precopy = vfio_save_complete_precopy,
681     .save_state = vfio_save_state,
682     .load_setup = vfio_load_setup,
683     .load_cleanup = vfio_load_cleanup,
684     .load_state = vfio_load_state,
685     .switchover_ack_needed = vfio_switchover_ack_needed,
686 };
687 
688 /* ---------------------------------------------------------------------- */
689 
690 static void vfio_vmstate_change_prepare(void *opaque, bool running,
691                                         RunState state)
692 {
693     VFIODevice *vbasedev = opaque;
694     VFIOMigration *migration = vbasedev->migration;
695     enum vfio_device_mig_state new_state;
696     int ret;
697 
698     new_state = migration->device_state == VFIO_DEVICE_STATE_PRE_COPY ?
699                     VFIO_DEVICE_STATE_PRE_COPY_P2P :
700                     VFIO_DEVICE_STATE_RUNNING_P2P;
701 
702     /*
703      * If setting the device in new_state fails, the device should be reset.
704      * To do so, use ERROR state as a recover state.
705      */
706     ret = vfio_migration_set_state(vbasedev, new_state,
707                                    VFIO_DEVICE_STATE_ERROR);
708     if (ret) {
709         /*
710          * Migration should be aborted in this case, but vm_state_notify()
711          * currently does not support reporting failures.
712          */
713         if (migrate_get_current()->to_dst_file) {
714             qemu_file_set_error(migrate_get_current()->to_dst_file, ret);
715         }
716     }
717 
718     trace_vfio_vmstate_change_prepare(vbasedev->name, running,
719                                       RunState_str(state),
720                                       mig_state_to_str(new_state));
721 }
722 
723 static void vfio_vmstate_change(void *opaque, bool running, RunState state)
724 {
725     VFIODevice *vbasedev = opaque;
726     enum vfio_device_mig_state new_state;
727     int ret;
728 
729     if (running) {
730         new_state = VFIO_DEVICE_STATE_RUNNING;
731     } else {
732         new_state =
733             (vfio_device_state_is_precopy(vbasedev) &&
734              (state == RUN_STATE_FINISH_MIGRATE || state == RUN_STATE_PAUSED)) ?
735                 VFIO_DEVICE_STATE_STOP_COPY :
736                 VFIO_DEVICE_STATE_STOP;
737     }
738 
739     /*
740      * If setting the device in new_state fails, the device should be reset.
741      * To do so, use ERROR state as a recover state.
742      */
743     ret = vfio_migration_set_state(vbasedev, new_state,
744                                    VFIO_DEVICE_STATE_ERROR);
745     if (ret) {
746         /*
747          * Migration should be aborted in this case, but vm_state_notify()
748          * currently does not support reporting failures.
749          */
750         if (migrate_get_current()->to_dst_file) {
751             qemu_file_set_error(migrate_get_current()->to_dst_file, ret);
752         }
753     }
754 
755     trace_vfio_vmstate_change(vbasedev->name, running, RunState_str(state),
756                               mig_state_to_str(new_state));
757 }
758 
759 static void vfio_migration_state_notifier(Notifier *notifier, void *data)
760 {
761     MigrationState *s = data;
762     VFIOMigration *migration = container_of(notifier, VFIOMigration,
763                                             migration_state);
764     VFIODevice *vbasedev = migration->vbasedev;
765 
766     trace_vfio_migration_state_notifier(vbasedev->name,
767                                         MigrationStatus_str(s->state));
768 
769     switch (s->state) {
770     case MIGRATION_STATUS_CANCELLING:
771     case MIGRATION_STATUS_CANCELLED:
772     case MIGRATION_STATUS_FAILED:
773         /*
774          * If setting the device in RUNNING state fails, the device should
775          * be reset. To do so, use ERROR state as a recover state.
776          */
777         vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_RUNNING,
778                                  VFIO_DEVICE_STATE_ERROR);
779     }
780 }
781 
782 static void vfio_migration_free(VFIODevice *vbasedev)
783 {
784     g_free(vbasedev->migration);
785     vbasedev->migration = NULL;
786 }
787 
788 static int vfio_migration_query_flags(VFIODevice *vbasedev, uint64_t *mig_flags)
789 {
790     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
791                                   sizeof(struct vfio_device_feature_migration),
792                               sizeof(uint64_t))] = {};
793     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
794     struct vfio_device_feature_migration *mig =
795         (struct vfio_device_feature_migration *)feature->data;
796 
797     feature->argsz = sizeof(buf);
798     feature->flags = VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_MIGRATION;
799     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
800         return -errno;
801     }
802 
803     *mig_flags = mig->flags;
804 
805     return 0;
806 }
807 
808 static bool vfio_dma_logging_supported(VFIODevice *vbasedev)
809 {
810     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
811                               sizeof(uint64_t))] = {};
812     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
813 
814     feature->argsz = sizeof(buf);
815     feature->flags = VFIO_DEVICE_FEATURE_PROBE |
816                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
817 
818     return !ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
819 }
820 
821 static int vfio_migration_init(VFIODevice *vbasedev)
822 {
823     int ret;
824     Object *obj;
825     VFIOMigration *migration;
826     char id[256] = "";
827     g_autofree char *path = NULL, *oid = NULL;
828     uint64_t mig_flags = 0;
829     VMChangeStateHandler *prepare_cb;
830 
831     if (!vbasedev->ops->vfio_get_object) {
832         return -EINVAL;
833     }
834 
835     obj = vbasedev->ops->vfio_get_object(vbasedev);
836     if (!obj) {
837         return -EINVAL;
838     }
839 
840     ret = vfio_migration_query_flags(vbasedev, &mig_flags);
841     if (ret) {
842         return ret;
843     }
844 
845     /* Basic migration functionality must be supported */
846     if (!(mig_flags & VFIO_MIGRATION_STOP_COPY)) {
847         return -EOPNOTSUPP;
848     }
849 
850     vbasedev->migration = g_new0(VFIOMigration, 1);
851     migration = vbasedev->migration;
852     migration->vbasedev = vbasedev;
853     migration->device_state = VFIO_DEVICE_STATE_RUNNING;
854     migration->data_fd = -1;
855     migration->mig_flags = mig_flags;
856 
857     vbasedev->dirty_pages_supported = vfio_dma_logging_supported(vbasedev);
858 
859     oid = vmstate_if_get_id(VMSTATE_IF(DEVICE(obj)));
860     if (oid) {
861         path = g_strdup_printf("%s/vfio", oid);
862     } else {
863         path = g_strdup("vfio");
864     }
865     strpadcpy(id, sizeof(id), path, '\0');
866 
867     register_savevm_live(id, VMSTATE_INSTANCE_ID_ANY, 1, &savevm_vfio_handlers,
868                          vbasedev);
869 
870     prepare_cb = migration->mig_flags & VFIO_MIGRATION_P2P ?
871                      vfio_vmstate_change_prepare :
872                      NULL;
873     migration->vm_state = qdev_add_vm_change_state_handler_full(
874         vbasedev->dev, vfio_vmstate_change, prepare_cb, vbasedev);
875     migration_add_notifier(&migration->migration_state,
876                            vfio_migration_state_notifier);
877 
878     return 0;
879 }
880 
881 static void vfio_migration_deinit(VFIODevice *vbasedev)
882 {
883     VFIOMigration *migration = vbasedev->migration;
884 
885     migration_remove_notifier(&migration->migration_state);
886     qemu_del_vm_change_state_handler(migration->vm_state);
887     unregister_savevm(VMSTATE_IF(vbasedev->dev), "vfio", vbasedev);
888     vfio_migration_free(vbasedev);
889     vfio_unblock_multiple_devices_migration();
890 }
891 
892 static int vfio_block_migration(VFIODevice *vbasedev, Error *err, Error **errp)
893 {
894     if (vbasedev->enable_migration == ON_OFF_AUTO_ON) {
895         error_propagate(errp, err);
896         return -EINVAL;
897     }
898 
899     vbasedev->migration_blocker = error_copy(err);
900     error_free(err);
901 
902     return migrate_add_blocker(&vbasedev->migration_blocker, errp);
903 }
904 
905 /* ---------------------------------------------------------------------- */
906 
907 int64_t vfio_mig_bytes_transferred(void)
908 {
909     return bytes_transferred;
910 }
911 
912 void vfio_reset_bytes_transferred(void)
913 {
914     bytes_transferred = 0;
915 }
916 
917 /*
918  * Return true when either migration initialized or blocker registered.
919  * Currently only return false when adding blocker fails which will
920  * de-register vfio device.
921  */
922 bool vfio_migration_realize(VFIODevice *vbasedev, Error **errp)
923 {
924     Error *err = NULL;
925     int ret;
926 
927     if (vbasedev->enable_migration == ON_OFF_AUTO_OFF) {
928         error_setg(&err, "%s: Migration is disabled for VFIO device",
929                    vbasedev->name);
930         return !vfio_block_migration(vbasedev, err, errp);
931     }
932 
933     ret = vfio_migration_init(vbasedev);
934     if (ret) {
935         if (ret == -ENOTTY) {
936             error_setg(&err, "%s: VFIO migration is not supported in kernel",
937                        vbasedev->name);
938         } else {
939             error_setg(&err,
940                        "%s: Migration couldn't be initialized for VFIO device, "
941                        "err: %d (%s)",
942                        vbasedev->name, ret, strerror(-ret));
943         }
944 
945         return !vfio_block_migration(vbasedev, err, errp);
946     }
947 
948     if (!vbasedev->dirty_pages_supported) {
949         if (vbasedev->enable_migration == ON_OFF_AUTO_AUTO) {
950             error_setg(&err,
951                        "%s: VFIO device doesn't support device dirty tracking",
952                        vbasedev->name);
953             goto add_blocker;
954         }
955 
956         warn_report("%s: VFIO device doesn't support device dirty tracking",
957                     vbasedev->name);
958     }
959 
960     ret = vfio_block_multiple_devices_migration(vbasedev, errp);
961     if (ret) {
962         goto out_deinit;
963     }
964 
965     if (vfio_viommu_preset(vbasedev)) {
966         error_setg(&err, "%s: Migration is currently not supported "
967                    "with vIOMMU enabled", vbasedev->name);
968         goto add_blocker;
969     }
970 
971     trace_vfio_migration_realize(vbasedev->name);
972     return true;
973 
974 add_blocker:
975     ret = vfio_block_migration(vbasedev, err, errp);
976 out_deinit:
977     if (ret) {
978         vfio_migration_deinit(vbasedev);
979     }
980     return !ret;
981 }
982 
983 void vfio_migration_exit(VFIODevice *vbasedev)
984 {
985     if (vbasedev->migration) {
986         vfio_migration_deinit(vbasedev);
987     }
988 
989     migrate_del_blocker(&vbasedev->migration_blocker);
990 }
991