xref: /openbmc/qemu/hw/vfio/migration.c (revision 53c7c924)
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_PRE_COPY:
75         return "PRE_COPY";
76     default:
77         return "UNKNOWN STATE";
78     }
79 }
80 
81 static int vfio_migration_set_state(VFIODevice *vbasedev,
82                                     enum vfio_device_mig_state new_state,
83                                     enum vfio_device_mig_state recover_state)
84 {
85     VFIOMigration *migration = vbasedev->migration;
86     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
87                               sizeof(struct vfio_device_feature_mig_state),
88                               sizeof(uint64_t))] = {};
89     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
90     struct vfio_device_feature_mig_state *mig_state =
91         (struct vfio_device_feature_mig_state *)feature->data;
92     int ret;
93 
94     feature->argsz = sizeof(buf);
95     feature->flags =
96         VFIO_DEVICE_FEATURE_SET | VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE;
97     mig_state->device_state = new_state;
98     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
99         /* Try to set the device in some good state */
100         ret = -errno;
101 
102         if (recover_state == VFIO_DEVICE_STATE_ERROR) {
103             error_report("%s: Failed setting device state to %s, err: %s. "
104                          "Recover state is ERROR. Resetting device",
105                          vbasedev->name, mig_state_to_str(new_state),
106                          strerror(errno));
107 
108             goto reset_device;
109         }
110 
111         error_report(
112             "%s: Failed setting device state to %s, err: %s. Setting device in recover state %s",
113                      vbasedev->name, mig_state_to_str(new_state),
114                      strerror(errno), mig_state_to_str(recover_state));
115 
116         mig_state->device_state = recover_state;
117         if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
118             ret = -errno;
119             error_report(
120                 "%s: Failed setting device in recover state, err: %s. Resetting device",
121                          vbasedev->name, strerror(errno));
122 
123             goto reset_device;
124         }
125 
126         migration->device_state = recover_state;
127 
128         return ret;
129     }
130 
131     migration->device_state = new_state;
132     if (mig_state->data_fd != -1) {
133         if (migration->data_fd != -1) {
134             /*
135              * This can happen if the device is asynchronously reset and
136              * terminates a data transfer.
137              */
138             error_report("%s: data_fd out of sync", vbasedev->name);
139             close(mig_state->data_fd);
140 
141             return -EBADF;
142         }
143 
144         migration->data_fd = mig_state->data_fd;
145     }
146 
147     trace_vfio_migration_set_state(vbasedev->name, mig_state_to_str(new_state));
148 
149     return 0;
150 
151 reset_device:
152     if (ioctl(vbasedev->fd, VFIO_DEVICE_RESET)) {
153         hw_error("%s: Failed resetting device, err: %s", vbasedev->name,
154                  strerror(errno));
155     }
156 
157     migration->device_state = VFIO_DEVICE_STATE_RUNNING;
158 
159     return ret;
160 }
161 
162 static int vfio_load_buffer(QEMUFile *f, VFIODevice *vbasedev,
163                             uint64_t data_size)
164 {
165     VFIOMigration *migration = vbasedev->migration;
166     int ret;
167 
168     ret = qemu_file_get_to_fd(f, migration->data_fd, data_size);
169     trace_vfio_load_state_device_data(vbasedev->name, data_size, ret);
170 
171     return ret;
172 }
173 
174 static int vfio_save_device_config_state(QEMUFile *f, void *opaque)
175 {
176     VFIODevice *vbasedev = opaque;
177 
178     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_CONFIG_STATE);
179 
180     if (vbasedev->ops && vbasedev->ops->vfio_save_config) {
181         vbasedev->ops->vfio_save_config(vbasedev, f);
182     }
183 
184     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
185 
186     trace_vfio_save_device_config_state(vbasedev->name);
187 
188     return qemu_file_get_error(f);
189 }
190 
191 static int vfio_load_device_config_state(QEMUFile *f, void *opaque)
192 {
193     VFIODevice *vbasedev = opaque;
194     uint64_t data;
195 
196     if (vbasedev->ops && vbasedev->ops->vfio_load_config) {
197         int ret;
198 
199         ret = vbasedev->ops->vfio_load_config(vbasedev, f);
200         if (ret) {
201             error_report("%s: Failed to load device config space",
202                          vbasedev->name);
203             return ret;
204         }
205     }
206 
207     data = qemu_get_be64(f);
208     if (data != VFIO_MIG_FLAG_END_OF_STATE) {
209         error_report("%s: Failed loading device config space, "
210                      "end flag incorrect 0x%"PRIx64, vbasedev->name, data);
211         return -EINVAL;
212     }
213 
214     trace_vfio_load_device_config_state(vbasedev->name);
215     return qemu_file_get_error(f);
216 }
217 
218 static void vfio_migration_cleanup(VFIODevice *vbasedev)
219 {
220     VFIOMigration *migration = vbasedev->migration;
221 
222     close(migration->data_fd);
223     migration->data_fd = -1;
224 }
225 
226 static int vfio_query_stop_copy_size(VFIODevice *vbasedev,
227                                      uint64_t *stop_copy_size)
228 {
229     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
230                               sizeof(struct vfio_device_feature_mig_data_size),
231                               sizeof(uint64_t))] = {};
232     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
233     struct vfio_device_feature_mig_data_size *mig_data_size =
234         (struct vfio_device_feature_mig_data_size *)feature->data;
235 
236     feature->argsz = sizeof(buf);
237     feature->flags =
238         VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_MIG_DATA_SIZE;
239 
240     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
241         return -errno;
242     }
243 
244     *stop_copy_size = mig_data_size->stop_copy_length;
245 
246     return 0;
247 }
248 
249 static int vfio_query_precopy_size(VFIOMigration *migration)
250 {
251     struct vfio_precopy_info precopy = {
252         .argsz = sizeof(precopy),
253     };
254 
255     migration->precopy_init_size = 0;
256     migration->precopy_dirty_size = 0;
257 
258     if (ioctl(migration->data_fd, VFIO_MIG_GET_PRECOPY_INFO, &precopy)) {
259         return -errno;
260     }
261 
262     migration->precopy_init_size = precopy.initial_bytes;
263     migration->precopy_dirty_size = precopy.dirty_bytes;
264 
265     return 0;
266 }
267 
268 /* Returns the size of saved data on success and -errno on error */
269 static ssize_t vfio_save_block(QEMUFile *f, VFIOMigration *migration)
270 {
271     ssize_t data_size;
272 
273     data_size = read(migration->data_fd, migration->data_buffer,
274                      migration->data_buffer_size);
275     if (data_size < 0) {
276         /*
277          * Pre-copy emptied all the device state for now. For more information,
278          * please refer to the Linux kernel VFIO uAPI.
279          */
280         if (errno == ENOMSG) {
281             return 0;
282         }
283 
284         return -errno;
285     }
286     if (data_size == 0) {
287         return 0;
288     }
289 
290     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_DATA_STATE);
291     qemu_put_be64(f, data_size);
292     qemu_put_buffer(f, migration->data_buffer, data_size);
293     bytes_transferred += data_size;
294 
295     trace_vfio_save_block(migration->vbasedev->name, data_size);
296 
297     return qemu_file_get_error(f) ?: data_size;
298 }
299 
300 static void vfio_update_estimated_pending_data(VFIOMigration *migration,
301                                                uint64_t data_size)
302 {
303     if (!data_size) {
304         /*
305          * Pre-copy emptied all the device state for now, update estimated sizes
306          * accordingly.
307          */
308         migration->precopy_init_size = 0;
309         migration->precopy_dirty_size = 0;
310 
311         return;
312     }
313 
314     if (migration->precopy_init_size) {
315         uint64_t init_size = MIN(migration->precopy_init_size, data_size);
316 
317         migration->precopy_init_size -= init_size;
318         data_size -= init_size;
319     }
320 
321     migration->precopy_dirty_size -= MIN(migration->precopy_dirty_size,
322                                          data_size);
323 }
324 
325 static bool vfio_precopy_supported(VFIODevice *vbasedev)
326 {
327     VFIOMigration *migration = vbasedev->migration;
328 
329     return migration->mig_flags & VFIO_MIGRATION_PRE_COPY;
330 }
331 
332 /* ---------------------------------------------------------------------- */
333 
334 static int vfio_save_setup(QEMUFile *f, void *opaque)
335 {
336     VFIODevice *vbasedev = opaque;
337     VFIOMigration *migration = vbasedev->migration;
338     uint64_t stop_copy_size = VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE;
339 
340     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_SETUP_STATE);
341 
342     vfio_query_stop_copy_size(vbasedev, &stop_copy_size);
343     migration->data_buffer_size = MIN(VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE,
344                                       stop_copy_size);
345     migration->data_buffer = g_try_malloc0(migration->data_buffer_size);
346     if (!migration->data_buffer) {
347         error_report("%s: Failed to allocate migration data buffer",
348                      vbasedev->name);
349         return -ENOMEM;
350     }
351 
352     if (vfio_precopy_supported(vbasedev)) {
353         int ret;
354 
355         switch (migration->device_state) {
356         case VFIO_DEVICE_STATE_RUNNING:
357             ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_PRE_COPY,
358                                            VFIO_DEVICE_STATE_RUNNING);
359             if (ret) {
360                 return ret;
361             }
362 
363             vfio_query_precopy_size(migration);
364 
365             break;
366         case VFIO_DEVICE_STATE_STOP:
367             /* vfio_save_complete_precopy() will go to STOP_COPY */
368             break;
369         default:
370             return -EINVAL;
371         }
372     }
373 
374     trace_vfio_save_setup(vbasedev->name, migration->data_buffer_size);
375 
376     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
377 
378     return qemu_file_get_error(f);
379 }
380 
381 static void vfio_save_cleanup(void *opaque)
382 {
383     VFIODevice *vbasedev = opaque;
384     VFIOMigration *migration = vbasedev->migration;
385 
386     g_free(migration->data_buffer);
387     migration->data_buffer = NULL;
388     migration->precopy_init_size = 0;
389     migration->precopy_dirty_size = 0;
390     migration->initial_data_sent = false;
391     vfio_migration_cleanup(vbasedev);
392     trace_vfio_save_cleanup(vbasedev->name);
393 }
394 
395 static void vfio_state_pending_estimate(void *opaque, uint64_t *must_precopy,
396                                         uint64_t *can_postcopy)
397 {
398     VFIODevice *vbasedev = opaque;
399     VFIOMigration *migration = vbasedev->migration;
400 
401     if (migration->device_state != VFIO_DEVICE_STATE_PRE_COPY) {
402         return;
403     }
404 
405     *must_precopy +=
406         migration->precopy_init_size + migration->precopy_dirty_size;
407 
408     trace_vfio_state_pending_estimate(vbasedev->name, *must_precopy,
409                                       *can_postcopy,
410                                       migration->precopy_init_size,
411                                       migration->precopy_dirty_size);
412 }
413 
414 /*
415  * Migration size of VFIO devices can be as little as a few KBs or as big as
416  * many GBs. This value should be big enough to cover the worst case.
417  */
418 #define VFIO_MIG_STOP_COPY_SIZE (100 * GiB)
419 
420 static void vfio_state_pending_exact(void *opaque, uint64_t *must_precopy,
421                                      uint64_t *can_postcopy)
422 {
423     VFIODevice *vbasedev = opaque;
424     VFIOMigration *migration = vbasedev->migration;
425     uint64_t stop_copy_size = VFIO_MIG_STOP_COPY_SIZE;
426 
427     /*
428      * If getting pending migration size fails, VFIO_MIG_STOP_COPY_SIZE is
429      * reported so downtime limit won't be violated.
430      */
431     vfio_query_stop_copy_size(vbasedev, &stop_copy_size);
432     *must_precopy += stop_copy_size;
433 
434     if (migration->device_state == VFIO_DEVICE_STATE_PRE_COPY) {
435         vfio_query_precopy_size(migration);
436 
437         *must_precopy +=
438             migration->precopy_init_size + migration->precopy_dirty_size;
439     }
440 
441     trace_vfio_state_pending_exact(vbasedev->name, *must_precopy, *can_postcopy,
442                                    stop_copy_size, migration->precopy_init_size,
443                                    migration->precopy_dirty_size);
444 }
445 
446 static bool vfio_is_active_iterate(void *opaque)
447 {
448     VFIODevice *vbasedev = opaque;
449     VFIOMigration *migration = vbasedev->migration;
450 
451     return migration->device_state == VFIO_DEVICE_STATE_PRE_COPY;
452 }
453 
454 static int vfio_save_iterate(QEMUFile *f, void *opaque)
455 {
456     VFIODevice *vbasedev = opaque;
457     VFIOMigration *migration = vbasedev->migration;
458     ssize_t data_size;
459 
460     data_size = vfio_save_block(f, migration);
461     if (data_size < 0) {
462         return data_size;
463     }
464 
465     vfio_update_estimated_pending_data(migration, data_size);
466 
467     if (migrate_switchover_ack() && !migration->precopy_init_size &&
468         !migration->initial_data_sent) {
469         qemu_put_be64(f, VFIO_MIG_FLAG_DEV_INIT_DATA_SENT);
470         migration->initial_data_sent = true;
471     } else {
472         qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
473     }
474 
475     trace_vfio_save_iterate(vbasedev->name, migration->precopy_init_size,
476                             migration->precopy_dirty_size);
477 
478     /*
479      * A VFIO device's pre-copy dirty_bytes is not guaranteed to reach zero.
480      * Return 1 so following handlers will not be potentially blocked.
481      */
482     return 1;
483 }
484 
485 static int vfio_save_complete_precopy(QEMUFile *f, void *opaque)
486 {
487     VFIODevice *vbasedev = opaque;
488     ssize_t data_size;
489     int ret;
490 
491     /* We reach here with device state STOP or STOP_COPY only */
492     ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_STOP_COPY,
493                                    VFIO_DEVICE_STATE_STOP);
494     if (ret) {
495         return ret;
496     }
497 
498     do {
499         data_size = vfio_save_block(f, vbasedev->migration);
500         if (data_size < 0) {
501             return data_size;
502         }
503     } while (data_size);
504 
505     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
506     ret = qemu_file_get_error(f);
507     if (ret) {
508         return ret;
509     }
510 
511     /*
512      * If setting the device in STOP state fails, the device should be reset.
513      * To do so, use ERROR state as a recover state.
514      */
515     ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_STOP,
516                                    VFIO_DEVICE_STATE_ERROR);
517     trace_vfio_save_complete_precopy(vbasedev->name, ret);
518 
519     return ret;
520 }
521 
522 static void vfio_save_state(QEMUFile *f, void *opaque)
523 {
524     VFIODevice *vbasedev = opaque;
525     int ret;
526 
527     ret = vfio_save_device_config_state(f, opaque);
528     if (ret) {
529         error_report("%s: Failed to save device config space",
530                      vbasedev->name);
531         qemu_file_set_error(f, ret);
532     }
533 }
534 
535 static int vfio_load_setup(QEMUFile *f, void *opaque)
536 {
537     VFIODevice *vbasedev = opaque;
538 
539     return vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_RESUMING,
540                                    vbasedev->migration->device_state);
541 }
542 
543 static int vfio_load_cleanup(void *opaque)
544 {
545     VFIODevice *vbasedev = opaque;
546 
547     vfio_migration_cleanup(vbasedev);
548     trace_vfio_load_cleanup(vbasedev->name);
549 
550     return 0;
551 }
552 
553 static int vfio_load_state(QEMUFile *f, void *opaque, int version_id)
554 {
555     VFIODevice *vbasedev = opaque;
556     int ret = 0;
557     uint64_t data;
558 
559     data = qemu_get_be64(f);
560     while (data != VFIO_MIG_FLAG_END_OF_STATE) {
561 
562         trace_vfio_load_state(vbasedev->name, data);
563 
564         switch (data) {
565         case VFIO_MIG_FLAG_DEV_CONFIG_STATE:
566         {
567             return vfio_load_device_config_state(f, opaque);
568         }
569         case VFIO_MIG_FLAG_DEV_SETUP_STATE:
570         {
571             data = qemu_get_be64(f);
572             if (data == VFIO_MIG_FLAG_END_OF_STATE) {
573                 return ret;
574             } else {
575                 error_report("%s: SETUP STATE: EOS not found 0x%"PRIx64,
576                              vbasedev->name, data);
577                 return -EINVAL;
578             }
579             break;
580         }
581         case VFIO_MIG_FLAG_DEV_DATA_STATE:
582         {
583             uint64_t data_size = qemu_get_be64(f);
584 
585             if (data_size) {
586                 ret = vfio_load_buffer(f, vbasedev, data_size);
587                 if (ret < 0) {
588                     return ret;
589                 }
590             }
591             break;
592         }
593         case VFIO_MIG_FLAG_DEV_INIT_DATA_SENT:
594         {
595             if (!vfio_precopy_supported(vbasedev) ||
596                 !migrate_switchover_ack()) {
597                 error_report("%s: Received INIT_DATA_SENT but switchover ack "
598                              "is not used", vbasedev->name);
599                 return -EINVAL;
600             }
601 
602             ret = qemu_loadvm_approve_switchover();
603             if (ret) {
604                 error_report(
605                     "%s: qemu_loadvm_approve_switchover failed, err=%d (%s)",
606                     vbasedev->name, ret, strerror(-ret));
607             }
608 
609             return ret;
610         }
611         default:
612             error_report("%s: Unknown tag 0x%"PRIx64, vbasedev->name, data);
613             return -EINVAL;
614         }
615 
616         data = qemu_get_be64(f);
617         ret = qemu_file_get_error(f);
618         if (ret) {
619             return ret;
620         }
621     }
622     return ret;
623 }
624 
625 static bool vfio_switchover_ack_needed(void *opaque)
626 {
627     VFIODevice *vbasedev = opaque;
628 
629     return vfio_precopy_supported(vbasedev);
630 }
631 
632 static const SaveVMHandlers savevm_vfio_handlers = {
633     .save_setup = vfio_save_setup,
634     .save_cleanup = vfio_save_cleanup,
635     .state_pending_estimate = vfio_state_pending_estimate,
636     .state_pending_exact = vfio_state_pending_exact,
637     .is_active_iterate = vfio_is_active_iterate,
638     .save_live_iterate = vfio_save_iterate,
639     .save_live_complete_precopy = vfio_save_complete_precopy,
640     .save_state = vfio_save_state,
641     .load_setup = vfio_load_setup,
642     .load_cleanup = vfio_load_cleanup,
643     .load_state = vfio_load_state,
644     .switchover_ack_needed = vfio_switchover_ack_needed,
645 };
646 
647 /* ---------------------------------------------------------------------- */
648 
649 static void vfio_vmstate_change(void *opaque, bool running, RunState state)
650 {
651     VFIODevice *vbasedev = opaque;
652     VFIOMigration *migration = vbasedev->migration;
653     enum vfio_device_mig_state new_state;
654     int ret;
655 
656     if (running) {
657         new_state = VFIO_DEVICE_STATE_RUNNING;
658     } else {
659         new_state =
660             (migration->device_state == VFIO_DEVICE_STATE_PRE_COPY &&
661              (state == RUN_STATE_FINISH_MIGRATE || state == RUN_STATE_PAUSED)) ?
662                 VFIO_DEVICE_STATE_STOP_COPY :
663                 VFIO_DEVICE_STATE_STOP;
664     }
665 
666     /*
667      * If setting the device in new_state fails, the device should be reset.
668      * To do so, use ERROR state as a recover state.
669      */
670     ret = vfio_migration_set_state(vbasedev, new_state,
671                                    VFIO_DEVICE_STATE_ERROR);
672     if (ret) {
673         /*
674          * Migration should be aborted in this case, but vm_state_notify()
675          * currently does not support reporting failures.
676          */
677         if (migrate_get_current()->to_dst_file) {
678             qemu_file_set_error(migrate_get_current()->to_dst_file, ret);
679         }
680     }
681 
682     trace_vfio_vmstate_change(vbasedev->name, running, RunState_str(state),
683                               mig_state_to_str(new_state));
684 }
685 
686 static void vfio_migration_state_notifier(Notifier *notifier, void *data)
687 {
688     MigrationState *s = data;
689     VFIOMigration *migration = container_of(notifier, VFIOMigration,
690                                             migration_state);
691     VFIODevice *vbasedev = migration->vbasedev;
692 
693     trace_vfio_migration_state_notifier(vbasedev->name,
694                                         MigrationStatus_str(s->state));
695 
696     switch (s->state) {
697     case MIGRATION_STATUS_CANCELLING:
698     case MIGRATION_STATUS_CANCELLED:
699     case MIGRATION_STATUS_FAILED:
700         /*
701          * If setting the device in RUNNING state fails, the device should
702          * be reset. To do so, use ERROR state as a recover state.
703          */
704         vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_RUNNING,
705                                  VFIO_DEVICE_STATE_ERROR);
706     }
707 }
708 
709 static void vfio_migration_free(VFIODevice *vbasedev)
710 {
711     g_free(vbasedev->migration);
712     vbasedev->migration = NULL;
713 }
714 
715 static int vfio_migration_query_flags(VFIODevice *vbasedev, uint64_t *mig_flags)
716 {
717     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
718                                   sizeof(struct vfio_device_feature_migration),
719                               sizeof(uint64_t))] = {};
720     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
721     struct vfio_device_feature_migration *mig =
722         (struct vfio_device_feature_migration *)feature->data;
723 
724     feature->argsz = sizeof(buf);
725     feature->flags = VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_MIGRATION;
726     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
727         return -errno;
728     }
729 
730     *mig_flags = mig->flags;
731 
732     return 0;
733 }
734 
735 static bool vfio_dma_logging_supported(VFIODevice *vbasedev)
736 {
737     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
738                               sizeof(uint64_t))] = {};
739     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
740 
741     feature->argsz = sizeof(buf);
742     feature->flags = VFIO_DEVICE_FEATURE_PROBE |
743                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
744 
745     return !ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
746 }
747 
748 static int vfio_migration_init(VFIODevice *vbasedev)
749 {
750     int ret;
751     Object *obj;
752     VFIOMigration *migration;
753     char id[256] = "";
754     g_autofree char *path = NULL, *oid = NULL;
755     uint64_t mig_flags = 0;
756 
757     if (!vbasedev->ops->vfio_get_object) {
758         return -EINVAL;
759     }
760 
761     obj = vbasedev->ops->vfio_get_object(vbasedev);
762     if (!obj) {
763         return -EINVAL;
764     }
765 
766     ret = vfio_migration_query_flags(vbasedev, &mig_flags);
767     if (ret) {
768         return ret;
769     }
770 
771     /* Basic migration functionality must be supported */
772     if (!(mig_flags & VFIO_MIGRATION_STOP_COPY)) {
773         return -EOPNOTSUPP;
774     }
775 
776     vbasedev->migration = g_new0(VFIOMigration, 1);
777     migration = vbasedev->migration;
778     migration->vbasedev = vbasedev;
779     migration->device_state = VFIO_DEVICE_STATE_RUNNING;
780     migration->data_fd = -1;
781     migration->mig_flags = mig_flags;
782 
783     vbasedev->dirty_pages_supported = vfio_dma_logging_supported(vbasedev);
784 
785     oid = vmstate_if_get_id(VMSTATE_IF(DEVICE(obj)));
786     if (oid) {
787         path = g_strdup_printf("%s/vfio", oid);
788     } else {
789         path = g_strdup("vfio");
790     }
791     strpadcpy(id, sizeof(id), path, '\0');
792 
793     register_savevm_live(id, VMSTATE_INSTANCE_ID_ANY, 1, &savevm_vfio_handlers,
794                          vbasedev);
795 
796     migration->vm_state = qdev_add_vm_change_state_handler(vbasedev->dev,
797                                                            vfio_vmstate_change,
798                                                            vbasedev);
799     migration->migration_state.notify = vfio_migration_state_notifier;
800     add_migration_state_change_notifier(&migration->migration_state);
801 
802     return 0;
803 }
804 
805 static void vfio_migration_deinit(VFIODevice *vbasedev)
806 {
807     VFIOMigration *migration = vbasedev->migration;
808 
809     remove_migration_state_change_notifier(&migration->migration_state);
810     qemu_del_vm_change_state_handler(migration->vm_state);
811     unregister_savevm(VMSTATE_IF(vbasedev->dev), "vfio", vbasedev);
812     vfio_migration_free(vbasedev);
813     vfio_unblock_multiple_devices_migration();
814 }
815 
816 static int vfio_block_migration(VFIODevice *vbasedev, Error *err, Error **errp)
817 {
818     int ret;
819 
820     if (vbasedev->enable_migration == ON_OFF_AUTO_ON) {
821         error_propagate(errp, err);
822         return -EINVAL;
823     }
824 
825     vbasedev->migration_blocker = error_copy(err);
826     error_free(err);
827 
828     ret = migrate_add_blocker(vbasedev->migration_blocker, errp);
829     if (ret < 0) {
830         error_free(vbasedev->migration_blocker);
831         vbasedev->migration_blocker = NULL;
832     }
833 
834     return ret;
835 }
836 
837 /* ---------------------------------------------------------------------- */
838 
839 int64_t vfio_mig_bytes_transferred(void)
840 {
841     return bytes_transferred;
842 }
843 
844 void vfio_reset_bytes_transferred(void)
845 {
846     bytes_transferred = 0;
847 }
848 
849 /*
850  * Return true when either migration initialized or blocker registered.
851  * Currently only return false when adding blocker fails which will
852  * de-register vfio device.
853  */
854 bool vfio_migration_realize(VFIODevice *vbasedev, Error **errp)
855 {
856     Error *err = NULL;
857     int ret;
858 
859     if (vbasedev->enable_migration == ON_OFF_AUTO_OFF) {
860         error_setg(&err, "%s: Migration is disabled for VFIO device",
861                    vbasedev->name);
862         return !vfio_block_migration(vbasedev, err, errp);
863     }
864 
865     ret = vfio_migration_init(vbasedev);
866     if (ret) {
867         if (ret == -ENOTTY) {
868             error_setg(&err, "%s: VFIO migration is not supported in kernel",
869                        vbasedev->name);
870         } else {
871             error_setg(&err,
872                        "%s: Migration couldn't be initialized for VFIO device, "
873                        "err: %d (%s)",
874                        vbasedev->name, ret, strerror(-ret));
875         }
876 
877         return !vfio_block_migration(vbasedev, err, errp);
878     }
879 
880     if (!vbasedev->dirty_pages_supported) {
881         if (vbasedev->enable_migration == ON_OFF_AUTO_AUTO) {
882             error_setg(&err,
883                        "%s: VFIO device doesn't support device dirty tracking",
884                        vbasedev->name);
885             goto add_blocker;
886         }
887 
888         warn_report("%s: VFIO device doesn't support device dirty tracking",
889                     vbasedev->name);
890     }
891 
892     ret = vfio_block_multiple_devices_migration(vbasedev, errp);
893     if (ret) {
894         goto out_deinit;
895     }
896 
897     if (vfio_viommu_preset(vbasedev)) {
898         error_setg(&err, "%s: Migration is currently not supported "
899                    "with vIOMMU enabled", vbasedev->name);
900         goto add_blocker;
901     }
902 
903     trace_vfio_migration_realize(vbasedev->name);
904     return true;
905 
906 add_blocker:
907     ret = vfio_block_migration(vbasedev, err, errp);
908 out_deinit:
909     if (ret) {
910         vfio_migration_deinit(vbasedev);
911     }
912     return !ret;
913 }
914 
915 void vfio_migration_exit(VFIODevice *vbasedev)
916 {
917     if (vbasedev->migration) {
918         vfio_migration_deinit(vbasedev);
919     }
920 
921     if (vbasedev->migration_blocker) {
922         migrate_del_blocker(vbasedev->migration_blocker);
923         error_free(vbasedev->migration_blocker);
924         vbasedev->migration_blocker = NULL;
925     }
926 }
927