From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from gate001.proxmox.com (gate001.proxmox.com [IPv6:2a0f:8001:1:32::40]) by lore.proxmox.com (Postfix) with ESMTPS id A04791FF0DE for ; Thu, 19 Nov 2026 11:42:26 +0100 (CET) Received: from gate001.proxmox.com (localhost.localdomain [127.0.0.1]) by gate001.proxmox.com (Proxmox) with ESMTP id 706B221C14; Thu, 19 Nov 2026 11:37:22 +0100 (CET) From: Thomas Lamprecht To: pve-devel@lists.proxmox.com Subject: [PATCH docs v4 37/41] vzdump: describe the guest identity of VMs and containers Date: Thu, 19 Nov 2026 11:34:23 +0100 Message-ID: <20261119103551.1666964-38-t.lamprecht@proxmox.com> X-Mailer: git-send-email 2.47.3 In-Reply-To: <20261119103551.1666964-1-t.lamprecht@proxmox.com> References: <20261119103551.1666964-1-t.lamprecht@proxmox.com> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-Bm-Milter-Handled: 55990f41-d878-4baa-be0a-ee34c49e34d2 X-Bm-Transport-Timestamp: 1795084557631 X-SPAM-LEVEL: Spam detection results: 0 AWL -0.391 Adjusted score from AWL reputation of From: address DMARC_MISSING 0.1 Missing DMARC policy KAM_ASCII_DIVIDERS 0.8 Email that uses ascii formatting dividers and possible spam tricks KAM_DMARC_STATUS 0.01 Test Rule for DKIM or SPF Failure with Strict Alignment (newer systems) POISEN_SPAM_PILL 0.1 Meta: its spam POISEN_SPAM_PILL_1 0.1 random spam to be learned in bayes POISEN_SPAM_PILL_3 0.1 random spam to be learned in bayes RCVD_IN_DNSWL_MED -2.3 Sender listed at https://www.dnswl.org/, medium trust SPF_HELO_NONE 0.001 SPF: HELO does not publish an SPF Record SPF_PASS -0.001 SPF: sender matches SPF record Message-ID-Hash: H4HC6IP7TRPE7RKY4C4SK3LM6KTMLRWQ X-Message-ID-Hash: H4HC6IP7TRPE7RKY4C4SK3LM6KTMLRWQ X-MailFrom: t.lamprecht@proxmox.com X-Mailman-Rule-Misses: dmarc-mitigation; no-senders; approved; loop; banned-address; emergency; member-moderation; nonmember-moderation; administrivia; implicit-dest; max-recipients; max-size; news-moderation; no-subject; digests; suspicious-header X-Mailman-Version: 3.3.10 Precedence: list List-Id: Proxmox VE development discussion List-Help: List-Owner: List-Post: List-Subscribe: List-Unsubscribe: Since Proxmox VE 9.3, VMs and containers have a guest identity in 'meta': a random uid and a registration time. Backup software and other tools can distinguish a guest from another that reuses its VMID [0]. Administrators need to understand why restores over a different guest are refused, when a restore keeps the uid of the backup, and why migrations and clones to older nodes are blocked. Those nodes drop 'meta', including VM creation information used for hardware defaults. Document the lifecycle, failed-restore retries, and mixed-version restrictions in the backup and restore chapter, where users are most likely to encounter the identity. Include migration and HA behavior there so the rules can be read together. [0]: https://bugzilla.proxmox.com/show_bug.cgi?id=7846 Signed-off-by: Thomas Lamprecht --- changes v3 -> v4: * the part of 'Failed Restores' about locks and leftovers moved to a patch of its own before this one, what stays is 'Retrying a Failed Restore' * lifecycle table and restore section: a guest with a uid keeps it, except in the retry of a failed restore of that backup * drop the sentence that a retry from standard input needs the override, which only holds for containers * the resource API has uid and rtime only for guests with an identity * remote migration: remove the guest that holds the uid only if it is not needed anymore; the refusal of older targets is described there instead of with the mixed versions * shorter descriptions of uid and rtime, one note with the hints for tools, and the reason for the refusals first in the mixed versions section changes v2 -> v3: * backups are still grouped by type and VMID, also for retention * refused remote migration: do not remove the reported guest if it is the source itself * own section for failed restores, with what a killed or late-failing restore leaves behind * remote migration of a guest without uid, HA recovery through the node of the active CRM, and smaller corrections (the uid gets compared, not the identity) changes v1 -> v2: * say that the uid has 100 random bits and how rtime differs from ctime * describe the refusal of a migration to another cluster, its typical cases and 'allow-new-uid' * retry after a failed restore: scope the pending entry, name the warning, correct the placeholder case and add the note for enforced unique guest IDs vzdump.adoc | 312 ++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 312 insertions(+) diff --git a/vzdump.adoc b/vzdump.adoc index 7587580..5a813c4 100644 --- a/vzdump.adoc +++ b/vzdump.adoc @@ -464,6 +464,13 @@ following CLI tools: For details see the corresponding manual pages. +When you restore a backup over an existing guest, {pve} compares the `uid` of +the backup and of the guest. If they differ, the restore is refused, unless you +set the `allow-identity-mismatch` option. The existing guest then keeps its +identity. See +xref:guest_identity_restore[Restore Over an Existing Guest] for details and +exceptions. + Bandwidth Limit ~~~~~~~~~~~~~~~ @@ -591,6 +598,311 @@ Use another VMID if the xref:pvecm_next_id_range[`next-id`] setting of the datacenter has the `unique` and `enforce` options set, as the first VMID then counts as used. +For how a retry handles the guest identity, see +xref:guest_identity_failed_restore[Retrying a Failed Restore]. + +[[guest_identity]] +Guest Identity +-------------- + +Since {pve} 9.3, each virtual machine and container gets a guest identity: a +random identifier, `uid`, and a registration time, `rtime`. Unlike a VMID, +which can be reused after a guest is destroyed, the `uid` lets backup software +and other tools recognize the same logical guest across restores and +migrations, even under a different VMID. + +Backups are still grouped by guest type and VMID, also for retention. If a new +guest reuses the VMID of a removed guest of the same type, pruning handles the +backups of both guests together. + +[[guest_identity_values]] +Identity Values +~~~~~~~~~~~~~~~ + +The guest identity is part of the `meta` property in the guest configuration. +This property also contains the creation time `ctime` and, for VMs, the QEMU +version that the VM was created with, `creation-qemu`: + +---- +meta: creation-qemu=10.1.2,ctime=1772444467,rtime=1794320571482,uid=w5qczgiem26qnwzwsoil +---- + +`uid`:: + +A random identifier of 20 lowercase letters and digits that encode 100 random +bits, so two guests practically never get the same `uid` by chance. + +`rtime`:: + +The time at which the guest was registered in this cluster, in milliseconds +since the UNIX epoch. It is set whenever the guest gets its `uid` in this +cluster, so a guest restored to a free VMID gets a new `rtime`, even if it keeps +the `uid` of the backup. The creation time `ctime`, in contrast, stays with the +guest: a clone has the `ctime` of its source, and a restored guest or a guest +migrated from another cluster has that of the backup or of the source guest. As +a timestamp, the `rtime` is not guaranteed to be unique. + +{pve} manages the `meta` property automatically. It is read-only through the +API, even for `root@pam`. + +IMPORTANT: The `uid` helps to correlate a guest with its backups and copies, +but does not prove where a guest comes from. Whoever restores a backup or +migrates a guest into the cluster supplies the `uid` of that guest, which {pve} +takes over if no other guest in the cluster has it. {pve} keeps a `uid` unique +within a cluster only for operations through its API. A configuration file +copied by hand has the same `uid` as the original, and guests in different +clusters can have the same `uid`. + +[[guest_identity_read]] +Reading the Identity +~~~~~~~~~~~~~~~~~~~~ + +The following places show the guest identity. All of them require the +`VM.Audit` privilege on the guest. + +* In the web interface, the 'Summary' panel of the guest shows the `uid` as + 'Unique ID'. The 'Details' button next to it, which has an info icon, opens + the 'Guest Identity' window. This window additionally shows the `rtime` as + 'Registered' and the creation information. +* The guest configuration, for example with `qm config ` or + `pct config `, and through the API paths + `/nodes//qemu//config` and `/nodes//lxc//config`. +* The guest status at the API paths `/nodes//qemu//status/current` + and `/nodes//lxc//status/current`, which contains the `uid`. +* The cluster resources at the API path `/cluster/resources`, which contain the + `uid` and `rtime` of each guest that has a guest identity, for example: ++ +---- +# pvesh get /cluster/resources --type vm --output-format json-pretty +---- + +The configuration in a backup contains the `meta` property as well. You can +show it with `pvesm extractconfig `, for example. + +[[guest_identity_lifecycle]] +Lifecycle +~~~~~~~~~ + +The following table shows how operations affect the identity of a guest. + +[width="100%",cols="3,4,2",options="header"] +|=========================================================================== +|Operation |`uid` |`rtime` +|Create, including import |new |current time +|Clone, including linked clones and clones from snapshots |new |current time +|Restore to a free VMID |the backup's `uid` if present and not used by another +guest in the cluster; otherwise new |current time +|Restore to a free VMID with `unique` set |new |current time +|Restore over an existing guest that has a `uid`, with or without `unique` set +|kept, see xref:guest_identity_restore[Restore Over an Existing Guest]; the +retry of a failed restore can take over the backup's `uid` instead, see +xref:guest_identity_failed_restore[Retrying a Failed Restore] |kept; current +time if the retry takes over the `uid` +|Restore over an existing guest without a `uid` |as for a free VMID |current +time +|Migration to another cluster, on the target |the source guest's `uid`; new if +another guest in the target cluster has it and `allow-new-uid` is set, or if the +source node runs a {pve} version older than 9.3, see +xref:guest_identity_remote_migration[Migration to Another Cluster] |current time +|Migration to another cluster, on the source if the guest is kept |kept |kept +|Snapshot rollback |kept, not taken from the snapshot |kept +|Migration within the cluster, HA recovery, configuration change, backup, +replication, conversion to a template |kept |kept +|Destroy |removed |removed +|=========================================================================== + +Restores and migrations from another cluster report the outcome in their task +log, in a line with `guest identity:`, for example: + +---- +guest identity: backup uid w5qczgiem26qnwzwsoil is in use by guest 100, keeping uid er3n3wqoqgcl362ep3da +---- + +Guests that existed before the upgrade to {pve} 9.3 get their identity the +first time a node running {pve} 9.3 or later writes their configuration. This +happens, for example, when their configuration changes, when they are backed +up, or when they are migrated to such a node. Their `rtime` is the time of that +write. Backups made before that have no `uid`. + +NOTE: For tools that track guests by their `uid`: a restore that takes over the +`uid` of the backup first assigns a new `uid` to the guest and replaces it at +its end, which can be after the guest got unlocked. A migration from another +cluster takes over the `uid` of the source guest right at its start. A changed +`uid`, for example in `/cluster/resources`, therefore does not mean that the +task has finished. A guest that a tool restores by removing it and creating a +new one through the API gets a new `uid`. To keep the identity, change the +existing guest in place, or let {pve} restore the guest. + +[[guest_identity_restore]] +Restore Over an Existing Guest +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +If the existing guest already has a guest identity, restoring a backup over it +keeps the `uid` and `rtime` of that guest, unless the guest is left over from a +failed restore of that backup, see +xref:guest_identity_failed_restore[Retrying a Failed Restore]. When both the +guest and the backup have a `uid` and the values differ, {pve} treats the backup +as belonging to another guest. This can happen when a VMID was reused. {pve} +refuses such a restore before it overwrites the guest: + +---- +backup belongs to another guest: its uid 'w5qczgiem26qnwzwsoil' differs from uid 'er3n3wqoqgcl362ep3da' of the existing VM 100 - use 'allow-identity-mismatch' to overwrite the VM anyway +---- + +To restore such a backup anyway, set the `allow-identity-mismatch` option. The +existing guest then keeps its identity. In the web interface, select 'Allow +restoring a backup of another guest' in the restore window, which you open from +the 'Backup' panel of the guest. On the command line, use for example: + +---- +# qmrestore --force 1 --allow-identity-mismatch 1 +# pct restore --force 1 --allow-identity-mismatch 1 +---- + +Backups without a `uid`, such as backups made with {pve} versions older than +9.3, are not refused. Container restores from standard input +(`pct restore -`) are an exception: the `uid` of such a backup is only +known once the container has been overwritten. A restore from standard input +over a container with a `uid` therefore always requires +`allow-identity-mismatch`. + +[[guest_identity_failed_restore]] +Retrying a Failed Restore +~~~~~~~~~~~~~~~~~~~~~~~~~ + +For locks and leftovers of a failed restore, see +xref:vzdump_restore_failed[Failed Restores]. + +If the failed restore was to a free VMID or over a guest without a `uid`, and it +had already read a backup that has a `uid`, `qm pending ` or +`pct pending ` lists an entry for `meta` with the `uid` of the backup. As +long as this entry exists, retrying the same restore over the guest left behind +is accepted. Starting the guest or changing its options removes the entry and +logs a warning that names the `uid`. The retry is then refused like a restore of +another guest's backup. + +A successful retry takes over the `uid` of the backup and sets a new `rtime`, as +a restore to a free VMID does. The guest keeps its `uid` and `rtime` instead if +another guest in the cluster has the `uid` of the backup, or if the retry sets +`unique`. + +A guest that had a `uid` and no such entry before the failed restore keeps its +`uid` and gets no entry. A retry is checked like the first attempt: a backup +with the same `uid` is accepted, also after the guest was started or changed, +and a backup with another `uid` requires `allow-identity-mismatch`. + +If a restore over an existing guest without a `uid` stopped before it read the +backup, for example because the backup storage was unreachable, the guest now +has a new `uid`. Retry with `allow-identity-mismatch`. The guest then keeps the +new `uid`. + +A restore can also complete with a warning starting with +`guest identity: unable to adopt backup uid`, for example if the cluster-wide +lock for this step could not be taken in time. The guest is then restored +completely, but keeps its new `uid`. To take over the `uid` of the backup, +repeat the restore over the guest before you start it or change its +configuration. + +[[guest_identity_remote_migration]] +Migration to Another Cluster +~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +A migration to another cluster (`qm remote-migrate`, `pct remote-migrate`) keeps +the `uid` of the guest. The target takes it over at the start of the migration, +before any data is transferred. + +If another guest in the target cluster already has that `uid`, the migration is +refused at that point. Typical cases are: + +* the guest gets migrated back to a cluster that still has the source guest of + an earlier migration, which stays there unless `delete` was set, +* a copy of the guest was restored in the target cluster from a backup, +* an earlier migration of the guest broke off after it had transferred the + configuration and left a locked guest in the target cluster, +* the target is the cluster that the guest is in. + +The error ends with: + +---- +guest 105 in the target cluster already has the uid 'w5qczgiem26qnwzwsoil' of this guest - use 'allow-new-uid' to migrate it with a new uid +---- + +If the reported guest is the source guest itself, do not remove it. To keep the +`uid`, migrate the guest within the cluster instead, with `qm migrate` or +`pct migrate`. + +Otherwise, to keep the `uid`, remove the other guest in the target cluster if +you no longer need it, and migrate again. To migrate anyway, set the +`allow-new-uid` option. The migrated guest then gets a new `uid`, and the other +guest keeps the old one. In both cases, use another target VMID if the +xref:pvecm_next_id_range[`next-id`] setting of the target cluster has the +`unique` and `enforce` options set, as the refused migration already used the +first one. For example: + +---- +# qm remote-migrate --target-bridge --target-storage --allow-new-uid 1 +---- + +A migration to another cluster is refused right after connecting to a target +node with a version older than 9.3, before any data is transferred. The error is +`Remote tunnel endpoint too old, upgrade required`. + +[[guest_identity_mixed_versions]] +Mixed Versions in a Cluster +~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +{pve} versions older than 9.3 ignore the whole `meta` property of a guest that +has a guest identity, including the creation information. A VM started on a node +running such a version can therefore get different virtual hardware, see +xref:qm_machine_type[Machine Type]. When such a node writes the configuration, +the property is lost. + +Therefore, while some nodes of a cluster still run such a version, migrating or +cloning a guest to one of these nodes is refused. The guest keeps running on its +current node. If the target node reports its version, the error message shows +that version and the minimum required version. If the node does not report its +version, the operation is refused because {pve} cannot verify that the node +supports guest identities. The 'Migrate' window of the web interface shows the +reason for the selected target node. + +Moves of HA-managed guests to such nodes are refused in the same way. A running +VM that the HA stack migrates keeps running. The HA stack stops a container, and +a VM that it relocates, before it migrates the guest, so such a guest is stopped +although the move is then refused. The HA stack retries automatic moves as long +as it selects such a node, and such a guest can stay stopped in the meantime. +This includes moves for: + +* the maintenance mode, +* a node shutdown with the `migrate` shutdown policy, +* node affinity rules, +* the CRS load balancer. + +Such moves complete only after the target nodes that the HA stack selects run +{pve} 9.3 or later. + +To avoid refused moves while you upgrade a cluster, do one of the following: + +* Move all guests away from a node before you upgrade it. +* Upgrade the packages on all nodes, one node at a time, before you reboot any + node or enable the maintenance mode. + +The check covers migrations and clones started on a node running {pve} 9.3 or +later. It does not cover: + +* the HA recovery of guests from a fenced node, +* configuration files moved by hand, +* operations started on a node with an older version. + +If an older node writes the configuration of a guest in one of these cases, the +guest loses its `meta` property. In an HA recovery, such a node can also be the +node of the active HA manager (CRM), which writes the configuration when it +removes a leftover lock, for example of a backup. The guest gets a new identity +the next time a node running {pve} 9.3 or later writes its configuration. + +Backups made with {pve} 9.3 or later contain the guest identity. Restore them +only on nodes running {pve} 9.3 or later, because older versions ignore the +`meta` property of such backups in the same way. + [[vzdump_configuration]] Configuration ------------- -- 2.47.3