xref: /openbmc/qemu/hw/vfio/migration.c (revision bf616ce4)
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 "exec/ramlist.h"
28 #include "exec/ram_addr.h"
29 #include "pci.h"
30 #include "trace.h"
31 #include "hw/hw.h"
32 
33 /*
34  * Flags to be used as unique delimiters for VFIO devices in the migration
35  * stream. These flags are composed as:
36  * 0xffffffff => MSB 32-bit all 1s
37  * 0xef10     => Magic ID, represents emulated (virtual) function IO
38  * 0x0000     => 16-bits reserved for flags
39  *
40  * The beginning of state information is marked by _DEV_CONFIG_STATE,
41  * _DEV_SETUP_STATE, or _DEV_DATA_STATE, respectively. The end of a
42  * certain state information is marked by _END_OF_STATE.
43  */
44 #define VFIO_MIG_FLAG_END_OF_STATE      (0xffffffffef100001ULL)
45 #define VFIO_MIG_FLAG_DEV_CONFIG_STATE  (0xffffffffef100002ULL)
46 #define VFIO_MIG_FLAG_DEV_SETUP_STATE   (0xffffffffef100003ULL)
47 #define VFIO_MIG_FLAG_DEV_DATA_STATE    (0xffffffffef100004ULL)
48 #define VFIO_MIG_FLAG_DEV_INIT_DATA_SENT (0xffffffffef100005ULL)
49 
50 /*
51  * This is an arbitrary size based on migration of mlx5 devices, where typically
52  * total device migration size is on the order of 100s of MB. Testing with
53  * larger values, e.g. 128MB and 1GB, did not show a performance improvement.
54  */
55 #define VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE (1 * MiB)
56 
57 static int64_t bytes_transferred;
58 
59 static const char *mig_state_to_str(enum vfio_device_mig_state state)
60 {
61     switch (state) {
62     case VFIO_DEVICE_STATE_ERROR:
63         return "ERROR";
64     case VFIO_DEVICE_STATE_STOP:
65         return "STOP";
66     case VFIO_DEVICE_STATE_RUNNING:
67         return "RUNNING";
68     case VFIO_DEVICE_STATE_STOP_COPY:
69         return "STOP_COPY";
70     case VFIO_DEVICE_STATE_RESUMING:
71         return "RESUMING";
72     case VFIO_DEVICE_STATE_RUNNING_P2P:
73         return "RUNNING_P2P";
74     case VFIO_DEVICE_STATE_PRE_COPY:
75         return "PRE_COPY";
76     case VFIO_DEVICE_STATE_PRE_COPY_P2P:
77         return "PRE_COPY_P2P";
78     default:
79         return "UNKNOWN STATE";
80     }
81 }
82 
83 static int vfio_migration_set_state(VFIODevice *vbasedev,
84                                     enum vfio_device_mig_state new_state,
85                                     enum vfio_device_mig_state recover_state)
86 {
87     VFIOMigration *migration = vbasedev->migration;
88     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
89                               sizeof(struct vfio_device_feature_mig_state),
90                               sizeof(uint64_t))] = {};
91     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
92     struct vfio_device_feature_mig_state *mig_state =
93         (struct vfio_device_feature_mig_state *)feature->data;
94     int ret;
95 
96     feature->argsz = sizeof(buf);
97     feature->flags =
98         VFIO_DEVICE_FEATURE_SET | VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE;
99     mig_state->device_state = new_state;
100     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
101         /* Try to set the device in some good state */
102         ret = -errno;
103 
104         if (recover_state == VFIO_DEVICE_STATE_ERROR) {
105             error_report("%s: Failed setting device state to %s, err: %s. "
106                          "Recover state is ERROR. Resetting device",
107                          vbasedev->name, mig_state_to_str(new_state),
108                          strerror(errno));
109 
110             goto reset_device;
111         }
112 
113         error_report(
114             "%s: Failed setting device state to %s, err: %s. Setting device in recover state %s",
115                      vbasedev->name, mig_state_to_str(new_state),
116                      strerror(errno), mig_state_to_str(recover_state));
117 
118         mig_state->device_state = recover_state;
119         if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
120             ret = -errno;
121             error_report(
122                 "%s: Failed setting device in recover state, err: %s. Resetting device",
123                          vbasedev->name, strerror(errno));
124 
125             goto reset_device;
126         }
127 
128         migration->device_state = recover_state;
129 
130         return ret;
131     }
132 
133     migration->device_state = new_state;
134     if (mig_state->data_fd != -1) {
135         if (migration->data_fd != -1) {
136             /*
137              * This can happen if the device is asynchronously reset and
138              * terminates a data transfer.
139              */
140             error_report("%s: data_fd out of sync", vbasedev->name);
141             close(mig_state->data_fd);
142 
143             return -EBADF;
144         }
145 
146         migration->data_fd = mig_state->data_fd;
147     }
148 
149     trace_vfio_migration_set_state(vbasedev->name, mig_state_to_str(new_state));
150 
151     return 0;
152 
153 reset_device:
154     if (ioctl(vbasedev->fd, VFIO_DEVICE_RESET)) {
155         hw_error("%s: Failed resetting device, err: %s", vbasedev->name,
156                  strerror(errno));
157     }
158 
159     migration->device_state = VFIO_DEVICE_STATE_RUNNING;
160 
161     return ret;
162 }
163 
164 /*
165  * Some device state transitions require resetting the device if they fail.
166  * This function sets the device in new_state and resets the device if that
167  * fails. Reset is done by using ERROR as the recover state.
168  */
169 static int
170 vfio_migration_set_state_or_reset(VFIODevice *vbasedev,
171                                   enum vfio_device_mig_state new_state)
172 {
173     return vfio_migration_set_state(vbasedev, new_state,
174                                     VFIO_DEVICE_STATE_ERROR);
175 }
176 
177 static int vfio_load_buffer(QEMUFile *f, VFIODevice *vbasedev,
178                             uint64_t data_size)
179 {
180     VFIOMigration *migration = vbasedev->migration;
181     int ret;
182 
183     ret = qemu_file_get_to_fd(f, migration->data_fd, data_size);
184     trace_vfio_load_state_device_data(vbasedev->name, data_size, ret);
185 
186     return ret;
187 }
188 
189 static int vfio_save_device_config_state(QEMUFile *f, void *opaque)
190 {
191     VFIODevice *vbasedev = opaque;
192 
193     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_CONFIG_STATE);
194 
195     if (vbasedev->ops && vbasedev->ops->vfio_save_config) {
196         vbasedev->ops->vfio_save_config(vbasedev, f);
197     }
198 
199     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
200 
201     trace_vfio_save_device_config_state(vbasedev->name);
202 
203     return qemu_file_get_error(f);
204 }
205 
206 static int vfio_load_device_config_state(QEMUFile *f, void *opaque)
207 {
208     VFIODevice *vbasedev = opaque;
209     uint64_t data;
210 
211     if (vbasedev->ops && vbasedev->ops->vfio_load_config) {
212         int ret;
213 
214         ret = vbasedev->ops->vfio_load_config(vbasedev, f);
215         if (ret) {
216             error_report("%s: Failed to load device config space",
217                          vbasedev->name);
218             return ret;
219         }
220     }
221 
222     data = qemu_get_be64(f);
223     if (data != VFIO_MIG_FLAG_END_OF_STATE) {
224         error_report("%s: Failed loading device config space, "
225                      "end flag incorrect 0x%"PRIx64, vbasedev->name, data);
226         return -EINVAL;
227     }
228 
229     trace_vfio_load_device_config_state(vbasedev->name);
230     return qemu_file_get_error(f);
231 }
232 
233 static void vfio_migration_cleanup(VFIODevice *vbasedev)
234 {
235     VFIOMigration *migration = vbasedev->migration;
236 
237     close(migration->data_fd);
238     migration->data_fd = -1;
239 }
240 
241 static int vfio_query_stop_copy_size(VFIODevice *vbasedev,
242                                      uint64_t *stop_copy_size)
243 {
244     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
245                               sizeof(struct vfio_device_feature_mig_data_size),
246                               sizeof(uint64_t))] = {};
247     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
248     struct vfio_device_feature_mig_data_size *mig_data_size =
249         (struct vfio_device_feature_mig_data_size *)feature->data;
250 
251     feature->argsz = sizeof(buf);
252     feature->flags =
253         VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_MIG_DATA_SIZE;
254 
255     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
256         return -errno;
257     }
258 
259     *stop_copy_size = mig_data_size->stop_copy_length;
260 
261     return 0;
262 }
263 
264 static int vfio_query_precopy_size(VFIOMigration *migration)
265 {
266     struct vfio_precopy_info precopy = {
267         .argsz = sizeof(precopy),
268     };
269 
270     migration->precopy_init_size = 0;
271     migration->precopy_dirty_size = 0;
272 
273     if (ioctl(migration->data_fd, VFIO_MIG_GET_PRECOPY_INFO, &precopy)) {
274         return -errno;
275     }
276 
277     migration->precopy_init_size = precopy.initial_bytes;
278     migration->precopy_dirty_size = precopy.dirty_bytes;
279 
280     return 0;
281 }
282 
283 /* Returns the size of saved data on success and -errno on error */
284 static ssize_t vfio_save_block(QEMUFile *f, VFIOMigration *migration)
285 {
286     ssize_t data_size;
287 
288     data_size = read(migration->data_fd, migration->data_buffer,
289                      migration->data_buffer_size);
290     if (data_size < 0) {
291         /*
292          * Pre-copy emptied all the device state for now. For more information,
293          * please refer to the Linux kernel VFIO uAPI.
294          */
295         if (errno == ENOMSG) {
296             return 0;
297         }
298 
299         return -errno;
300     }
301     if (data_size == 0) {
302         return 0;
303     }
304 
305     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_DATA_STATE);
306     qemu_put_be64(f, data_size);
307     qemu_put_buffer(f, migration->data_buffer, data_size);
308     bytes_transferred += data_size;
309 
310     trace_vfio_save_block(migration->vbasedev->name, data_size);
311 
312     return qemu_file_get_error(f) ?: data_size;
313 }
314 
315 static void vfio_update_estimated_pending_data(VFIOMigration *migration,
316                                                uint64_t data_size)
317 {
318     if (!data_size) {
319         /*
320          * Pre-copy emptied all the device state for now, update estimated sizes
321          * accordingly.
322          */
323         migration->precopy_init_size = 0;
324         migration->precopy_dirty_size = 0;
325 
326         return;
327     }
328 
329     if (migration->precopy_init_size) {
330         uint64_t init_size = MIN(migration->precopy_init_size, data_size);
331 
332         migration->precopy_init_size -= init_size;
333         data_size -= init_size;
334     }
335 
336     migration->precopy_dirty_size -= MIN(migration->precopy_dirty_size,
337                                          data_size);
338 }
339 
340 static bool vfio_precopy_supported(VFIODevice *vbasedev)
341 {
342     VFIOMigration *migration = vbasedev->migration;
343 
344     return migration->mig_flags & VFIO_MIGRATION_PRE_COPY;
345 }
346 
347 /* ---------------------------------------------------------------------- */
348 
349 static int vfio_save_prepare(void *opaque, Error **errp)
350 {
351     VFIODevice *vbasedev = opaque;
352 
353     /*
354      * Snapshot doesn't use postcopy nor background snapshot, so allow snapshot
355      * even if they are on.
356      */
357     if (runstate_check(RUN_STATE_SAVE_VM)) {
358         return 0;
359     }
360 
361     if (migrate_postcopy_ram()) {
362         error_setg(
363             errp, "%s: VFIO migration is not supported with postcopy migration",
364             vbasedev->name);
365         return -EOPNOTSUPP;
366     }
367 
368     if (migrate_background_snapshot()) {
369         error_setg(
370             errp,
371             "%s: VFIO migration is not supported with background snapshot",
372             vbasedev->name);
373         return -EOPNOTSUPP;
374     }
375 
376     return 0;
377 }
378 
379 static int vfio_save_setup(QEMUFile *f, void *opaque, Error **errp)
380 {
381     VFIODevice *vbasedev = opaque;
382     VFIOMigration *migration = vbasedev->migration;
383     uint64_t stop_copy_size = VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE;
384     int ret;
385 
386     qemu_put_be64(f, VFIO_MIG_FLAG_DEV_SETUP_STATE);
387 
388     vfio_query_stop_copy_size(vbasedev, &stop_copy_size);
389     migration->data_buffer_size = MIN(VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE,
390                                       stop_copy_size);
391     migration->data_buffer = g_try_malloc0(migration->data_buffer_size);
392     if (!migration->data_buffer) {
393         error_setg(errp, "%s: Failed to allocate migration data buffer",
394                    vbasedev->name);
395         return -ENOMEM;
396     }
397 
398     if (vfio_precopy_supported(vbasedev)) {
399         switch (migration->device_state) {
400         case VFIO_DEVICE_STATE_RUNNING:
401             ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_PRE_COPY,
402                                            VFIO_DEVICE_STATE_RUNNING);
403             if (ret) {
404                 error_setg(errp, "%s: Failed to set new PRE_COPY state",
405                            vbasedev->name);
406                 return ret;
407             }
408 
409             vfio_query_precopy_size(migration);
410 
411             break;
412         case VFIO_DEVICE_STATE_STOP:
413             /* vfio_save_complete_precopy() will go to STOP_COPY */
414             break;
415         default:
416             error_setg(errp, "%s: Invalid device state %d", vbasedev->name,
417                        migration->device_state);
418             return -EINVAL;
419         }
420     }
421 
422     trace_vfio_save_setup(vbasedev->name, migration->data_buffer_size);
423 
424     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
425 
426     ret = qemu_file_get_error(f);
427     if (ret < 0) {
428         error_setg_errno(errp, -ret, "%s: save setup failed", vbasedev->name);
429     }
430 
431     return ret;
432 }
433 
434 static void vfio_save_cleanup(void *opaque)
435 {
436     VFIODevice *vbasedev = opaque;
437     VFIOMigration *migration = vbasedev->migration;
438 
439     /*
440      * Changing device state from STOP_COPY to STOP can take time. Do it here,
441      * after migration has completed, so it won't increase downtime.
442      */
443     if (migration->device_state == VFIO_DEVICE_STATE_STOP_COPY) {
444         vfio_migration_set_state_or_reset(vbasedev, VFIO_DEVICE_STATE_STOP);
445     }
446 
447     g_free(migration->data_buffer);
448     migration->data_buffer = NULL;
449     migration->precopy_init_size = 0;
450     migration->precopy_dirty_size = 0;
451     migration->initial_data_sent = false;
452     vfio_migration_cleanup(vbasedev);
453     trace_vfio_save_cleanup(vbasedev->name);
454 }
455 
456 static void vfio_state_pending_estimate(void *opaque, uint64_t *must_precopy,
457                                         uint64_t *can_postcopy)
458 {
459     VFIODevice *vbasedev = opaque;
460     VFIOMigration *migration = vbasedev->migration;
461 
462     if (!vfio_device_state_is_precopy(vbasedev)) {
463         return;
464     }
465 
466     *must_precopy +=
467         migration->precopy_init_size + migration->precopy_dirty_size;
468 
469     trace_vfio_state_pending_estimate(vbasedev->name, *must_precopy,
470                                       *can_postcopy,
471                                       migration->precopy_init_size,
472                                       migration->precopy_dirty_size);
473 }
474 
475 /*
476  * Migration size of VFIO devices can be as little as a few KBs or as big as
477  * many GBs. This value should be big enough to cover the worst case.
478  */
479 #define VFIO_MIG_STOP_COPY_SIZE (100 * GiB)
480 
481 static void vfio_state_pending_exact(void *opaque, uint64_t *must_precopy,
482                                      uint64_t *can_postcopy)
483 {
484     VFIODevice *vbasedev = opaque;
485     VFIOMigration *migration = vbasedev->migration;
486     uint64_t stop_copy_size = VFIO_MIG_STOP_COPY_SIZE;
487 
488     /*
489      * If getting pending migration size fails, VFIO_MIG_STOP_COPY_SIZE is
490      * reported so downtime limit won't be violated.
491      */
492     vfio_query_stop_copy_size(vbasedev, &stop_copy_size);
493     *must_precopy += stop_copy_size;
494 
495     if (vfio_device_state_is_precopy(vbasedev)) {
496         vfio_query_precopy_size(migration);
497 
498         *must_precopy +=
499             migration->precopy_init_size + migration->precopy_dirty_size;
500     }
501 
502     trace_vfio_state_pending_exact(vbasedev->name, *must_precopy, *can_postcopy,
503                                    stop_copy_size, migration->precopy_init_size,
504                                    migration->precopy_dirty_size);
505 }
506 
507 static bool vfio_is_active_iterate(void *opaque)
508 {
509     VFIODevice *vbasedev = opaque;
510 
511     return vfio_device_state_is_precopy(vbasedev);
512 }
513 
514 /*
515  * Note about migration rate limiting: VFIO migration buffer size is currently
516  * limited to 1MB, so there is no need to check if migration rate exceeded (as
517  * in the worst case it will exceed by 1MB). However, if the buffer size is
518  * later changed to a bigger value, migration rate should be enforced here.
519  */
520 static int vfio_save_iterate(QEMUFile *f, void *opaque)
521 {
522     VFIODevice *vbasedev = opaque;
523     VFIOMigration *migration = vbasedev->migration;
524     ssize_t data_size;
525 
526     data_size = vfio_save_block(f, migration);
527     if (data_size < 0) {
528         return data_size;
529     }
530 
531     vfio_update_estimated_pending_data(migration, data_size);
532 
533     if (migrate_switchover_ack() && !migration->precopy_init_size &&
534         !migration->initial_data_sent) {
535         qemu_put_be64(f, VFIO_MIG_FLAG_DEV_INIT_DATA_SENT);
536         migration->initial_data_sent = true;
537     } else {
538         qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
539     }
540 
541     trace_vfio_save_iterate(vbasedev->name, migration->precopy_init_size,
542                             migration->precopy_dirty_size);
543 
544     return !migration->precopy_init_size && !migration->precopy_dirty_size;
545 }
546 
547 static int vfio_save_complete_precopy(QEMUFile *f, void *opaque)
548 {
549     VFIODevice *vbasedev = opaque;
550     ssize_t data_size;
551     int ret;
552 
553     /* We reach here with device state STOP or STOP_COPY only */
554     ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_STOP_COPY,
555                                    VFIO_DEVICE_STATE_STOP);
556     if (ret) {
557         return ret;
558     }
559 
560     do {
561         data_size = vfio_save_block(f, vbasedev->migration);
562         if (data_size < 0) {
563             return data_size;
564         }
565     } while (data_size);
566 
567     qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
568     ret = qemu_file_get_error(f);
569     if (ret) {
570         return ret;
571     }
572 
573     trace_vfio_save_complete_precopy(vbasedev->name, ret);
574 
575     return ret;
576 }
577 
578 static void vfio_save_state(QEMUFile *f, void *opaque)
579 {
580     VFIODevice *vbasedev = opaque;
581     int ret;
582 
583     ret = vfio_save_device_config_state(f, opaque);
584     if (ret) {
585         error_report("%s: Failed to save device config space",
586                      vbasedev->name);
587         qemu_file_set_error(f, ret);
588     }
589 }
590 
591 static int vfio_load_setup(QEMUFile *f, void *opaque, Error **errp)
592 {
593     VFIODevice *vbasedev = opaque;
594     int ret;
595 
596     ret = vfio_migration_set_state(vbasedev, VFIO_DEVICE_STATE_RESUMING,
597                                    vbasedev->migration->device_state);
598     if (ret) {
599         error_setg(errp, "%s: Failed to set RESUMING state", vbasedev->name);
600     }
601     return ret;
602 }
603 
604 static int vfio_load_cleanup(void *opaque)
605 {
606     VFIODevice *vbasedev = opaque;
607 
608     vfio_migration_cleanup(vbasedev);
609     trace_vfio_load_cleanup(vbasedev->name);
610 
611     return 0;
612 }
613 
614 static int vfio_load_state(QEMUFile *f, void *opaque, int version_id)
615 {
616     VFIODevice *vbasedev = opaque;
617     int ret = 0;
618     uint64_t data;
619 
620     data = qemu_get_be64(f);
621     while (data != VFIO_MIG_FLAG_END_OF_STATE) {
622 
623         trace_vfio_load_state(vbasedev->name, data);
624 
625         switch (data) {
626         case VFIO_MIG_FLAG_DEV_CONFIG_STATE:
627         {
628             return vfio_load_device_config_state(f, opaque);
629         }
630         case VFIO_MIG_FLAG_DEV_SETUP_STATE:
631         {
632             data = qemu_get_be64(f);
633             if (data == VFIO_MIG_FLAG_END_OF_STATE) {
634                 return ret;
635             } else {
636                 error_report("%s: SETUP STATE: EOS not found 0x%"PRIx64,
637                              vbasedev->name, data);
638                 return -EINVAL;
639             }
640             break;
641         }
642         case VFIO_MIG_FLAG_DEV_DATA_STATE:
643         {
644             uint64_t data_size = qemu_get_be64(f);
645 
646             if (data_size) {
647                 ret = vfio_load_buffer(f, vbasedev, data_size);
648                 if (ret < 0) {
649                     return ret;
650                 }
651             }
652             break;
653         }
654         case VFIO_MIG_FLAG_DEV_INIT_DATA_SENT:
655         {
656             if (!vfio_precopy_supported(vbasedev) ||
657                 !migrate_switchover_ack()) {
658                 error_report("%s: Received INIT_DATA_SENT but switchover ack "
659                              "is not used", vbasedev->name);
660                 return -EINVAL;
661             }
662 
663             ret = qemu_loadvm_approve_switchover();
664             if (ret) {
665                 error_report(
666                     "%s: qemu_loadvm_approve_switchover failed, err=%d (%s)",
667                     vbasedev->name, ret, strerror(-ret));
668             }
669 
670             return ret;
671         }
672         default:
673             error_report("%s: Unknown tag 0x%"PRIx64, vbasedev->name, data);
674             return -EINVAL;
675         }
676 
677         data = qemu_get_be64(f);
678         ret = qemu_file_get_error(f);
679         if (ret) {
680             return ret;
681         }
682     }
683     return ret;
684 }
685 
686 static bool vfio_switchover_ack_needed(void *opaque)
687 {
688     VFIODevice *vbasedev = opaque;
689 
690     return vfio_precopy_supported(vbasedev);
691 }
692 
693 static const SaveVMHandlers savevm_vfio_handlers = {
694     .save_prepare = vfio_save_prepare,
695     .save_setup = vfio_save_setup,
696     .save_cleanup = vfio_save_cleanup,
697     .state_pending_estimate = vfio_state_pending_estimate,
698     .state_pending_exact = vfio_state_pending_exact,
699     .is_active_iterate = vfio_is_active_iterate,
700     .save_live_iterate = vfio_save_iterate,
701     .save_live_complete_precopy = vfio_save_complete_precopy,
702     .save_state = vfio_save_state,
703     .load_setup = vfio_load_setup,
704     .load_cleanup = vfio_load_cleanup,
705     .load_state = vfio_load_state,
706     .switchover_ack_needed = vfio_switchover_ack_needed,
707 };
708 
709 /* ---------------------------------------------------------------------- */
710 
711 static void vfio_vmstate_change_prepare(void *opaque, bool running,
712                                         RunState state)
713 {
714     VFIODevice *vbasedev = opaque;
715     VFIOMigration *migration = vbasedev->migration;
716     enum vfio_device_mig_state new_state;
717     int ret;
718 
719     new_state = migration->device_state == VFIO_DEVICE_STATE_PRE_COPY ?
720                     VFIO_DEVICE_STATE_PRE_COPY_P2P :
721                     VFIO_DEVICE_STATE_RUNNING_P2P;
722 
723     ret = vfio_migration_set_state_or_reset(vbasedev, new_state);
724     if (ret) {
725         /*
726          * Migration should be aborted in this case, but vm_state_notify()
727          * currently does not support reporting failures.
728          */
729         migration_file_set_error(ret);
730     }
731 
732     trace_vfio_vmstate_change_prepare(vbasedev->name, running,
733                                       RunState_str(state),
734                                       mig_state_to_str(new_state));
735 }
736 
737 static void vfio_vmstate_change(void *opaque, bool running, RunState state)
738 {
739     VFIODevice *vbasedev = opaque;
740     enum vfio_device_mig_state new_state;
741     int ret;
742 
743     if (running) {
744         new_state = VFIO_DEVICE_STATE_RUNNING;
745     } else {
746         new_state =
747             (vfio_device_state_is_precopy(vbasedev) &&
748              (state == RUN_STATE_FINISH_MIGRATE || state == RUN_STATE_PAUSED)) ?
749                 VFIO_DEVICE_STATE_STOP_COPY :
750                 VFIO_DEVICE_STATE_STOP;
751     }
752 
753     ret = vfio_migration_set_state_or_reset(vbasedev, new_state);
754     if (ret) {
755         /*
756          * Migration should be aborted in this case, but vm_state_notify()
757          * currently does not support reporting failures.
758          */
759         migration_file_set_error(ret);
760     }
761 
762     trace_vfio_vmstate_change(vbasedev->name, running, RunState_str(state),
763                               mig_state_to_str(new_state));
764 }
765 
766 static int vfio_migration_state_notifier(NotifierWithReturn *notifier,
767                                          MigrationEvent *e, Error **errp)
768 {
769     VFIOMigration *migration = container_of(notifier, VFIOMigration,
770                                             migration_state);
771     VFIODevice *vbasedev = migration->vbasedev;
772 
773     trace_vfio_migration_state_notifier(vbasedev->name, e->type);
774 
775     if (e->type == MIG_EVENT_PRECOPY_FAILED) {
776         vfio_migration_set_state_or_reset(vbasedev, VFIO_DEVICE_STATE_RUNNING);
777     }
778     return 0;
779 }
780 
781 static void vfio_migration_free(VFIODevice *vbasedev)
782 {
783     g_free(vbasedev->migration);
784     vbasedev->migration = NULL;
785 }
786 
787 static int vfio_migration_query_flags(VFIODevice *vbasedev, uint64_t *mig_flags)
788 {
789     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
790                                   sizeof(struct vfio_device_feature_migration),
791                               sizeof(uint64_t))] = {};
792     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
793     struct vfio_device_feature_migration *mig =
794         (struct vfio_device_feature_migration *)feature->data;
795 
796     feature->argsz = sizeof(buf);
797     feature->flags = VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_MIGRATION;
798     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
799         return -errno;
800     }
801 
802     *mig_flags = mig->flags;
803 
804     return 0;
805 }
806 
807 static bool vfio_dma_logging_supported(VFIODevice *vbasedev)
808 {
809     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
810                               sizeof(uint64_t))] = {};
811     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
812 
813     feature->argsz = sizeof(buf);
814     feature->flags = VFIO_DEVICE_FEATURE_PROBE |
815                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
816 
817     return !ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
818 }
819 
820 static int vfio_migration_init(VFIODevice *vbasedev)
821 {
822     int ret;
823     Object *obj;
824     VFIOMigration *migration;
825     char id[256] = "";
826     g_autofree char *path = NULL, *oid = NULL;
827     uint64_t mig_flags = 0;
828     VMChangeStateHandler *prepare_cb;
829 
830     if (!vbasedev->ops->vfio_get_object) {
831         return -EINVAL;
832     }
833 
834     obj = vbasedev->ops->vfio_get_object(vbasedev);
835     if (!obj) {
836         return -EINVAL;
837     }
838 
839     ret = vfio_migration_query_flags(vbasedev, &mig_flags);
840     if (ret) {
841         return ret;
842     }
843 
844     /* Basic migration functionality must be supported */
845     if (!(mig_flags & VFIO_MIGRATION_STOP_COPY)) {
846         return -EOPNOTSUPP;
847     }
848 
849     vbasedev->migration = g_new0(VFIOMigration, 1);
850     migration = vbasedev->migration;
851     migration->vbasedev = vbasedev;
852     migration->device_state = VFIO_DEVICE_STATE_RUNNING;
853     migration->data_fd = -1;
854     migration->mig_flags = mig_flags;
855 
856     vbasedev->dirty_pages_supported = vfio_dma_logging_supported(vbasedev);
857 
858     oid = vmstate_if_get_id(VMSTATE_IF(DEVICE(obj)));
859     if (oid) {
860         path = g_strdup_printf("%s/vfio", oid);
861     } else {
862         path = g_strdup("vfio");
863     }
864     strpadcpy(id, sizeof(id), path, '\0');
865 
866     register_savevm_live(id, VMSTATE_INSTANCE_ID_ANY, 1, &savevm_vfio_handlers,
867                          vbasedev);
868 
869     prepare_cb = migration->mig_flags & VFIO_MIGRATION_P2P ?
870                      vfio_vmstate_change_prepare :
871                      NULL;
872     migration->vm_state = qdev_add_vm_change_state_handler_full(
873         vbasedev->dev, vfio_vmstate_change, prepare_cb, vbasedev);
874     migration_add_notifier(&migration->migration_state,
875                            vfio_migration_state_notifier);
876 
877     return 0;
878 }
879 
880 static void vfio_migration_deinit(VFIODevice *vbasedev)
881 {
882     VFIOMigration *migration = vbasedev->migration;
883 
884     migration_remove_notifier(&migration->migration_state);
885     qemu_del_vm_change_state_handler(migration->vm_state);
886     unregister_savevm(VMSTATE_IF(vbasedev->dev), "vfio", vbasedev);
887     vfio_migration_free(vbasedev);
888     vfio_unblock_multiple_devices_migration();
889 }
890 
891 static int vfio_block_migration(VFIODevice *vbasedev, Error *err, Error **errp)
892 {
893     if (vbasedev->enable_migration == ON_OFF_AUTO_ON) {
894         error_propagate(errp, err);
895         return -EINVAL;
896     }
897 
898     vbasedev->migration_blocker = error_copy(err);
899     error_free(err);
900 
901     return migrate_add_blocker_normal(&vbasedev->migration_blocker, errp);
902 }
903 
904 /* ---------------------------------------------------------------------- */
905 
906 int64_t vfio_mig_bytes_transferred(void)
907 {
908     return bytes_transferred;
909 }
910 
911 void vfio_reset_bytes_transferred(void)
912 {
913     bytes_transferred = 0;
914 }
915 
916 /*
917  * Return true when either migration initialized or blocker registered.
918  * Currently only return false when adding blocker fails which will
919  * de-register vfio device.
920  */
921 bool vfio_migration_realize(VFIODevice *vbasedev, Error **errp)
922 {
923     Error *err = NULL;
924     int ret;
925 
926     if (vbasedev->enable_migration == ON_OFF_AUTO_OFF) {
927         error_setg(&err, "%s: Migration is disabled for VFIO device",
928                    vbasedev->name);
929         return !vfio_block_migration(vbasedev, err, errp);
930     }
931 
932     ret = vfio_migration_init(vbasedev);
933     if (ret) {
934         if (ret == -ENOTTY) {
935             error_setg(&err, "%s: VFIO migration is not supported in kernel",
936                        vbasedev->name);
937         } else {
938             error_setg(&err,
939                        "%s: Migration couldn't be initialized for VFIO device, "
940                        "err: %d (%s)",
941                        vbasedev->name, ret, strerror(-ret));
942         }
943 
944         return !vfio_block_migration(vbasedev, err, errp);
945     }
946 
947     if (!vbasedev->dirty_pages_supported) {
948         if (vbasedev->enable_migration == ON_OFF_AUTO_AUTO) {
949             error_setg(&err,
950                        "%s: VFIO device doesn't support device dirty tracking",
951                        vbasedev->name);
952             goto add_blocker;
953         }
954 
955         warn_report("%s: VFIO device doesn't support device dirty tracking",
956                     vbasedev->name);
957     }
958 
959     ret = vfio_block_multiple_devices_migration(vbasedev, errp);
960     if (ret) {
961         goto out_deinit;
962     }
963 
964     if (vfio_viommu_preset(vbasedev)) {
965         error_setg(&err, "%s: Migration is currently not supported "
966                    "with vIOMMU enabled", vbasedev->name);
967         goto add_blocker;
968     }
969 
970     trace_vfio_migration_realize(vbasedev->name);
971     return true;
972 
973 add_blocker:
974     ret = vfio_block_migration(vbasedev, err, errp);
975 out_deinit:
976     if (ret) {
977         vfio_migration_deinit(vbasedev);
978     }
979     return !ret;
980 }
981 
982 void vfio_migration_exit(VFIODevice *vbasedev)
983 {
984     if (vbasedev->migration) {
985         vfio_migration_deinit(vbasedev);
986     }
987 
988     migrate_del_blocker(&vbasedev->migration_blocker);
989 }
990