so that we can reuse that information
the removal of the adding to the corrupted list is ok, since
'get_chunks_in_order' returns them at the end of the list
and we do the same if the loading fails later in 'verify_index_chunks'
so we still mark them corrupt
(assuming that the load will fail if the stat does)
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
only check every 1024'th, which is cheaper to do than a modulo, as we
can just mask the 10 least-significant-bits and check if the result
is zero.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
Fixes a non-negligible performance regression from commit
7f394c807b
While we skip known-verified chunks in the stat-and-inode-sort loop,
those are only the ones from previous indexes. If there's a repeated
chunk in one index they would get re-verified more often as required.
So, add the check again explicitly to the read+verify loop.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
before reading the chunks from disk in the order of the index file,
stat them first and sort them by inode number.
this can have a very positive impact on read speed on spinning disks,
even with the additional stat'ing of the chunks.
memory footprint should be tolerable, for 1_000_000 chunks
we need about ~16MiB of memory (Vec of 64bit position + 64bit inode)
(assuming 4MiB Chunks, such an index would reference 4TiB of data)
two small benchmarks (single spinner, ext4) here showed an improvement from
~430 seconds to ~330 seconds for a 32GiB fixed index
and from
~160 seconds to ~120 seconds for a 10GiB dynamic index
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
all the verify methods pass along the following:
- task worker
- datastore
- corrupt and verified chunks
might as well pull that out into a common type, with the added bonus of
now having a single point for construction instead of copying the
default capacaties in three different modules..
Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
percentage of verified groups, interpolating based on snapshot count
within the group. in most cases, this will also be closer to 'real'
progress since added snapshots (those which will be verified) in active
backup groups will be roughly evenly distributed, while number of total
snapshots per group will be heavily skewed towards those groups which
have existed the longest, even though most of those old snapshots will
only be re-verified very infrequently.
Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
BackupInfo::list_backup_groups is identical code-wise, and makes more
sense as entry point for listing groups.
Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
log invalid owners to system log, and continue with next group just as
if permission checks fail for the following operations:
- verify store with limited permissions
- list store groups
- list store snapshots
all other call sites either handle it correctly already (sync/pull), or
operate on a single group/snapshot and can bubble up the error.
Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
for verifying a whole datastore. Datastore.Backup now allows verifying
only backups owned by the triggering user.
Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
commit a4915dfc2b made a wrong fix, as
it did not observed that the last expressions was done under the
invariant that we had a last verification result, because if none
could be loaded we already returned true (include).
It thus broke the case for "never re-verify", which is important when
using multiple schedules, a more high frequent one for new,
unverified snapshots, and a low frequency to re-verify older snapshots,
e.g., monthly.
Fix this case again, rework the code to avoid this easy to oversee
invariant. Use a nested match to better express the implication of
each setting, and add some comments.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
Avoid races when updating manifest data by flocking a lock file.
update_manifest is used to ensure updates always happen with the lock
held.
Snapshot deletion also acquires the lock, so it cannot interfere with an
outstanding manifest write.
Signed-off-by: Stefan Reiter <s.reiter@proxmox.com>
There's no point in having that as a seperate method, just parse the
thing into a struct and write it back out correctly.
Also makes further changes to the method simpler.
Signed-off-by: Stefan Reiter <s.reiter@proxmox.com>
If we can't acquire a lock (either because the snapshot disappeared, it
is about to be forgotten/pruned, or it is currently still running) skip
the snapshot. Hold the lock during verification, so that it cannot be
deleted while we are still verifying.
Signed-off-by: Stefan Reiter <s.reiter@proxmox.com>
To untangle the server code from the actual backup
implementation.
It would be ideal if the whole backup/ dir could become its
own crate with minimal dependencies, certainly without
depending on the actual api server. That would then also be
used more easily to create forensic tools for all the data
file types we have in the backup repositories.
Signed-off-by: Wolfgang Bumiller <w.bumiller@proxmox.com>
In theory, one can do std::mem::forget, and ignore the drop handler. With
the lifetime hack, this could result in a crash.
So we simply require 'static lifetime now (futures also needs that).
This ensures that following backups will always upload the chunk,
thereby replacing it with a correct version again.
Format for renaming is <digest>.<counter>.bad where <counter> is used if
a chunk is found to be bad again before a GC cleans it up.
Care has been taken to deliberately only rename a chunk in conditions
where it is guaranteed to be an error in the chunk itself. Otherwise a
broken index file could lead to an unwanted mass-rename of chunks.
Signed-off-by: Stefan Reiter <s.reiter@proxmox.com>
Save the state ("ok" or "failed") and the UPID of the respective
verify task. With this we can easily allow to open the relevant task
log and show when the last verify happened.
As we already load the manifest when listing the snapshots, just add
it there directly.
Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
regular chunks are only decoded when their contents are accessed, in
which case we need to have the key anyway and want to verify the digest.
for blobs we need to verify beforehand, since their checksums are always
calculated based on their raw content, and stored in the manifest.
manifests are also stored as blobs, but don't have a digest in the
traditional sense (they might have a signature covering parts of their
contents, but that is verified already when loading the manifest).
this commit does not cover pull/sync code which copies blobs and chunks
as-is without decoding them.
Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
so that we can print a list at the end of the worker which backups
are corrupt.
this is useful if there are many snapshots and some in between had an
error. Before this patch, the task log simply says to 'look in the logs'
but if the log is very long it makes it hard to see what exactly failed.
Signed-off-by: Dominik Csapak <d.csapak@proxmox.com>
