Bisync checks file equality before renaming sync conflicts by comparing
checksums. Before this change, backends without checksum support (notably
Crypt) would fall back to --size-only for these checks, which is not a very
safe method (differing files can sometimes have the same size, especially if
they're small.) After this change, Crypt remotes fallback to using Cryptcheck
so that checksums can be compared. As a last resort when neither Check nor
Cryptcheck are available, files are compared using --download so that we can be
certain the files are identical regardless of checksum support.
Before this change, bisync supported `--backup-dir` only when `Path1` and
`Path2` were different paths on the same remote. With this change, bisync
introduces new `--backup-dir1` and `--backup-dir2` flags to support separate
backup-dirs for `Path1` and `Path2`.
`--backup-dir1` and `--backup-dir2` can use different remotes from each other,
but `--backup-dir1` must use the same remote as `Path1`, and `--backup-dir2`
must use the same remote as `Path2`. Each backup directory must not overlap its
respective bisync Path without being excluded by a filter rule.
The standard `--backup-dir` will also work, if both paths use the same remote
(but note that deleted files from both paths would be mixed together in the
same dir). If either `--backup-dir1` and `--backup-dir2` are set, they will
override `--backup-dir`.
Before this change, bisync intentionally ignored Google Docs (albeit in a
buggy way that caused problems during --resync.) After this change, Google Docs
(including Google Sheets, Slides, etc.) are now supported in bisync, subject to
the same options, defaults, and limitations as in `rclone sync`. When bisyncing
drive with non-drive backends, the drive -> non-drive direction is controlled
by `--drive-export-formats` (default `"docx,xlsx,pptx,svg"`) and the non-drive
-> drive direction is controlled by `--drive-import-formats` (default none.)
For example, with the default export/import formats, a Google Sheet on the
drive side will be synced to an `.xlsx` file on the non-drive side. In the
reverse direction, `.xlsx` files with filenames that match an existing Google
Sheet will be synced to that Google Sheet, while `.xlsx` files that do NOT
match an existing Google Sheet will be copied to drive as normal `.xlsx` files
(without conversion to Sheets, although the Google Drive web browser UI may
still give you the option to open it as one.)
If `--drive-import-formats` is set (it's not, by default), then all of the
specified formats will be converted to Google Docs, if there is no existing
Google Doc with a matching name. Caution: such conversion can be quite lossy,
and in most cases it's probably not what you want!
To bisync Google Docs as URL shortcut links (in a manner similar to "Drive for
Desktop"), use: `--drive-export-formats url` (or alternatives.)
Note that these link files cannot be edited on the non-drive side -- you will
get errors if you try to sync an edited link file back to drive. They CAN be
deleted (it will result in deleting the corresponding Google Doc.) If you
create a `.url` file on the non-drive side that does not match an existing
Google Doc, bisyncing it will just result in copying the literal `.url` file
over to drive (no Google Doc will be created.) So, as a general rule of thumb,
think of them as read-only placeholders on the non-drive side, and make all
your changes on the drive side.
Likewise, even with other export-formats, it is best to only move/rename Google
Docs on the drive side. This is because otherwise, bisync will interpret this
as a file deleted and another created, and accordingly, it will delete the
Google Doc and create a new file at the new path. (Whether or not that new file
is a Google Doc depends on `--drive-import-formats`.)
Lastly, take note that all Google Docs on the drive side have a size of `-1`
and no checksum. Therefore, they cannot be reliably synced with the
`--checksum` or `--size-only` flags. (To be exact: they will still get
created/deleted, and bisync's delta engine will notice changes and queue them
for syncing, but the underlying sync function will consider them identical and
skip them.) To work around this, use the default (modtime and size) instead of
`--checksum` or `--size-only`.
To ignore Google Docs entirely, use `--drive-skip-gdocs`.
Nearly all of the Google Docs logic is outsourced to the Drive backend, so
future changes should also be supported by bisync.
Before this change, bisync had no ability to retry in the event of sync errors.
After this change, bisync will retry if --resilient is passed, but only in one
direction at a time. We can safely retry in one direction because the source is
still intact, even if the dest was left in a messy state. If the first
direction still fails after our final retry, we abort and do NOT continue in
the other direction, to prevent the messy dest from polluting the source. If
the first direction succeeds, we do then allow retries in the other direction.
The number of retries is controllable by --retries (default 3)
bisync: high-level retries if --resilient
Before this change, bisync had no ability to retry in the event of sync errors.
After this change, bisync will retry if --resilient is passed, but only in one
direction at a time. We can safely retry in one direction because the source is
still intact, even if the dest was left in a messy state. If the first
direction still fails after our final retry, we abort and do NOT continue in
the other direction, to prevent the messy dest from polluting the source. If
the first direction succeeds, we do then allow retries in the other direction.
The number of retries is controllable by --retries (default 3)
Refactored the case / unicode normalization logic to be much more efficient,
and fix the last outstanding issue from #7270. Before this change, we were
doing lots of for loops and re-normalizing strings we had already normalized
earlier. Now, we leave the normalizing entirely to March and avoid
re-transforming later, which seems to make a large difference in terms of
performance.
Before this change, --resync was handled in three steps, and needed to do a lot
of unnecessary work to implement its own --ignore-existing logic, which also
caused problems with unicode normalization, in addition to being pretty slow.
After this change, it is refactored to produce the same result much more
efficiently, by reducing the three steps to two and letting ci.IgnoreExisting
do the work instead of reinventing the wheel.
The behavior and sync order remain unchanged for now -- just faster (but see
the ongoing lively discussions about potential future changes in #5681!)
Before this change, Bisync sometimes normalized NFD to NFC and sometimes
did not, causing errors in some scenarios (particularly for users of macOS).
It was similarly inconsistent in its handling of case-insensitivity.
There were three main places where Bisync should have normalized, but didn't:
1. When building the list of files that need to be transferred during --resync
2. When building the list of deltas during a non-resync
3. When comparing Path1 to Path2 during --check-sync
After this change, 1 and 3 are resolved, and bisync supports
--no-unicode-normalization and --ignore-case-sync in the same way as sync.
2 will be addressed in a future update.
Before this change, bisync needed to build a full listing for Path1, then a
full listing for Path2, then compare them -- and each of those tasks needed to
finish before the next one could start. In addition to being slow and
inefficient, it also caused real problems if a file changed between the time
bisync checked it on Path1 and the time it checked the corresponding file on
Path2.
This change solves these problems by listing both paths concurrently, using
the same March infrastructure that check and sync use to traverse two
directories in lock-step, optimized by Go's robust concurrency support.
Listings should now be much faster, and any given path is now checked
nearly-instantaneously on both sides, minimizing room for error.
Further discussion:
https://forum.rclone.org/t/bisync-bugs-and-feature-requests/37636#:~:text=4.%20Listings%20should%20alternate%20between%20paths%20to%20minimize%20errors
This introduces a few basic color codings to make the terminal output more
readable (and more fun). Rclone's standard --color flag is supported.
(AUTO|NEVER|ALWAYS)
Only a few lines have colors right now -- more will probably be added in
future versions.
Before this change, bisync had no mechanism for "retrying" a file again next
time, in the event of an unexpected and possibly temporary error. After this
change, bisync is now essentially able to mark a file as needing to be
rechecked next time. Bisync does this by keeping one prior listing on hand at
all times. In a low-confidence situation, bisync can revert a given file row
back to its state at the end of the last known successful sync, ensuring that
any subsequent changes will be re-noticed on the next run.
This can potentially be helpful for a dynamically changing file system, where
files may be changing quickly while bisync is working with them.
Before this change, if --create-empty-src-dirs was specified, bisync would
include directories in the list of deltas to evaluate by their modtime,
relative to the prior sync. This was unnecessary, as rclone does not yet
support setting modtime for directories.
After this change, we skip directories when comparing modtimes. (In other
words, we care only if a directory is created or deleted, not whether it is
newer or older.)
Before this change, if there were changes to sync, bisync listed each path
twice: once before the sync and once after. The second listing caused quite
a lot of problems, in addition to making each run much slower and more
expensive. A serious side-effect was that file changes could slip through
undetected, if they happened to occur while a sync was running (between the
first and second listing snapshots.)
After this change, the second listing is eliminated by getting the underlying
sync operation to report back a list of what it changed. Not only is this more
efficient, but also much more robust to concurrent modifications. It should no
longer be necessary to avoid make changes while it's running -- bisync will
simply learn about those changes next time and handle them on the next run.
Additionally, this also makes --check-sync usable again.
For further discussion, see:
https://forum.rclone.org/t/bisync-bugs-and-feature-requests/37636#:~:text=5.%20Final%20listings%20should%20be%20created%20from%20initial%20snapshot%20%2B%20deltas%2C%20not%20full%20re%2Dscans%2C%20to%20avoid%20errors%20if%20files%20changed%20during%20sync
Allows rclone sync to accept the same output file flags as rclone check,
for the purpose of writing results to a file.
A new --dest-after option is also supported, which writes a list file using
the same ListFormat flags as lsf (including customizable options for hash,
modtime, etc.) Conceptually it is similar to rsync's --itemize-changes, but
not identical -- it should output an accurate list of what will be on the
destination after the sync.
Note that it has a few limitations, and certain scenarios
are not currently supported:
--max-duration / CutoffModeHard
--compare-dest / --copy-dest (because equal() is called multiple times for the
same file)
server-side moves of an entire dir at once (because we never get the individual
file objects in the dir)
High-level retries, because there would be dupes
Possibly some error scenarios that didn't come up on the tests
Note also that each file is logged during the sync, as opposed to after, so it
is most useful as a predictor of what SHOULD happen to each file
(which may or may not match what actually DID.)
Only rclone sync is currently supported -- support for copy and move may be
added in the future.
Before this change, bisync handled copies and deletes in separate operations.
After this change, they are combined in one sync operation, which is faster
and also allows bisync to support --track-renames and --backup-dir.
Bisync uses a --files-from filter containing only the paths bisync has
determined need to be synced. Just like in sync (but in both directions),
if a path is present on the dst but not the src, it's interpreted as a delete
rather than a copy.
Before this change, lsf's time format was hard-coded to "2006-01-02 15:04:05",
regardless of the Fs's precision. After this change, a new optional
--time-format flag is added to allow customizing the format (the default is
unchanged).
Examples:
rclone lsf remote:path --format pt --time-format 'Jan 2, 2006 at 3:04pm (MST)'
rclone lsf remote:path --format pt --time-format '2006-01-02 15:04:05.000000000'
rclone lsf remote:path --format pt --time-format '2006-01-02T15:04:05.999999999Z07:00'
rclone lsf remote:path --format pt --time-format RFC3339
rclone lsf remote:path --format pt --time-format DateOnly
rclone lsf remote:path --format pt --time-format max
--time-format max will automatically truncate '2006-01-02 15:04:05.000000000'
to the maximum precision supported by the remote.
This updates the direct dependencies.
The latest github.com/willscott/go-nfs has changed the interface
slightly so this implements a dummy InvalidateHandle method in order
to satisfy it.
Before this change, listing a subdirectory gave errors like this:
Entry doesn't belong in directory "" (contains subdir) - ignoring
It also did full recursive listings when it didn't need to.
This was caused by the code using the underlying Fs to do recursive
listings on bucket based backends.
Using both the VFS and the underlying Fs is a mistake so this patch
removes the code which uses the underlying Fs and just uses the VFS.
Fixes#7500
- make compile on all unix OSes - this will make the docs appear on linux and rclone.org!
- add --sudo flag for using with mount
- improve error reporting
- fix option handling
Before this change overwriting an existing file with a 0 length file
didn't update the file size.
This change corrects the issue and makes sure the file is truncated
properly.
This was discovered by the full integration tests.
Before this change serve s3 would return NoSuchKey errors when a non
existent prefix was listed.
This change fixes it to return an empty list like AWS does.
This was discovered by the full integration tests.
- Changes
- Rename `--s3-authkey` to `--auth-key` to get it out of the s3 backend namespace
- Enable `Content-MD5` integrity checks
- Remove locking after code audit
- Documentation
- Factor out documentation into seperate file
- Add Quickstart to docs
- Add Bugs section to docs
- Add experimental tag to docs
- Add rclone provider to s3 backend docs
- Fixes
- Correct quirks in s3 backend
- Change fmt.Printlns into fs.Logs
- Make metadata storage per backend not global
- Log on startup if anonymous access is enabled
- Coding style fixes
- rename fs to vfs to save confusion with the rest of rclone code
- rename db to b for *s3Backend
Fixes#7062
- add context to log and fallthrough to error log level
- test: use rclone random lib to generate random strings
- calculate hash from vfs cache if file is uploading
- add server started log with server url
- remove md5 hasher
The upstream library rclone uses for rclone mount no longer supports
freebsd. Not only is it broken, but it no longer compiles.
This patch disables rclone mount for freebsd.
However all is not lost for freebsd users - compiling rclone with the
`cmount` tag, so `go install -tags cmount` will install a working
`rclone mount` command which uses cgofuse and the libfuse C library
directly.
Note that the binaries from rclone.org will not have mount support as
we don't have a freebsd build machine in CI and it is very hard to
cross compile cmount.
See: https://github.com/bazil/fuse/issues/280Fixes#5843
ncdu stores the position that it was in for each directory. However
doing a rescan can cause those positions to be out of range if the
number of files decreased in a directory. When re-entering the
directory, this causes an index out of range error.
This fixes the problem by detecting the index out of range and
flushing the saved directory position.
See: https://forum.rclone.org/t/slice-bounds-out-of-range-during-ncdu/42492/
This was caused by a change to the upstream library
ProtonMail/go-crypto checking the flags on the keys more strictly.
However the signing key for rclone is very old and does not have those
flags. Adding those flags using `gpg --edit-key` and then the
`change-usage` subcommand to remove, save, quite then re-add, save
quit the signing capabilities caused the key to work.
This also adds tests for the verification and adds the selfupdate
tests into the integration test harness as they had been disabled on
CI because they rely on external sources and are sometimes unreliable.
Fixes#7373
With automount the target mount drive appears twice in /proc/self/mountinfo.
379 27 0:70 / /mnt/rclone rw,relatime shared:433 - autofs systemd-1 rw,fd=57,...
566 379 0:90 / /mnt/rclone rw,nosuid,nodev,relatime shared:488 - fuse.rclone remote: rw,...
Before this fix we only looked for the mount once in
/proc/self/mountinfo. It finds the automount line and since this
doesn't have fs type rclone it concludes the mount isn't ready yet.
This patch makes rclone look through all the mounts and if any of them
have fs type rclone it concludes the mount is ready.
See: https://forum.rclone.org/t/systemd-mount-works-but-automount-does-not/42287/
Before this change, if a hardlink command was issued, rclone would
just ignore it and not return an error.
This changes any unknown operations (including hardlink) to return an
unsupported error.
Summary:
In cases where cmount is not available in macOS, we alias nfsmount to mount command and transparently start the NFS server and mount it to the target dir.
The NFS server is started on localhost on a random port so it is reasonably secure.
Test Plan:
```
go run rclone.go mount --http-url https://beta.rclone.org :http: nfs-test
```
Added mount tests:
```
go test ./cmd/nfsmount
```
Summary:
Adding a new command to serve any remote over NFS. This is only useful for new macOS versions where FUSE mounts are not available.
* Added willscot/go-nfs dependency and updated go.mod and go.sum
Test Plan:
```
go run rclone.go serve nfs --http-url https://beta.rclone.org :http:
```
Test that it is serving correctly by mounting the NFS directory.
```
mkdir nfs-test
mount -oport=58654,mountport=58654 localhost: nfs-test
```
Then we can list the mounted directory to see it is working.
```
ls nfs-test
```
This was always the intention, it was just implemented wrong.
This shortens the s3 docs by 1369 bringing them down to half the size
just about.
Fixes#7325
Before this change, bisync ignored the dryRun parameter (only when specified
via the rc.)
This change fixes the issue, so that the dryRun rc parameter is equivalent to
the --dry-run flag.
If the server returns the MIME type as application/octet-stream we
assume it doesn't really know what the MIME type. This patch tries
matching the MIME type from the file extension instead in this case.
This enables the use of servers (like OneDrive for Business) which
don't allow the setting of MIME types on upload and have a poor
selection of mime types.
Fixes#7259
In this commit we introduced a race condition when using the auth
proxy.
94a320f23c serve ftp: update to goftp.io/server v2.0.1
This was due to the re-organisation of the upstream library which made
the driver be a singleton rather than per session.
This means that when using the auth proxy we need to keep track of
which VFS to use by based on which FTP user is connected.
This also adjusts the locking so that the methods will run
concurrently.
Before this change uploading files with rclone to:
rclone serve sftp --vfs-cache-mode full
Would return the error:
command "md5sum XXX" failed with error: unexpected non file
This patch detects that the file is still in the VFS cache and reads
the MD5SUM from there rather from the remote.
Fixes#7241
This adds an additional parameter to the creation of each flag. This
specifies one or more flag groups. This **must** be set for global
flags and **must not** be set for local flags.
This causes flags.md to be built with sections to aid comprehension
and it causes the documentation pages for each command (and the
`--help`) to be built showing the flags groups as specified in the
`groups` annotation on the command.
See: https://forum.rclone.org/t/make-docs-for-mortals-not-only-rclone-gurus/39476/
This introduces a new fs.Option flag, Sensitive and uses this along
with IsPassword to redact the info in the config file for support
purposes.
It adds this flag into backends where appropriate. It was necessary to
add oauthutil.SharedOptions to some backends as they were missing
them.
Fixes#5209
When using `rclone cat` to print the contents of several files, the
user may want to inject some separator between the files, such as a
comma or a newline. This patch adds a `--separator` option to the `cat`
command to make that possible. The default value remains an empty
string, `""`, maintaining the prior behavior of `rclone cat`.
Closes#6968
In this commit we accidentally removed the global --rc flags.
0df7466d2b cmd/rcd: Fix command docs to include command specific prefix (#6675)
This re-instates them.
Before this change if both --progress and --interactive were set then
the screen display could become muddled.
This change makes --progress and --interactive use the same lock so
while rclone is asking for interactive questions, the progress will be
paused.
Fixes#6755
This change addresses two issues with commands that re-used
flags from common packages:
1) cobra.Command definitions did not include the command specific
prefix in doc strings.
2) Command specific flag prefixes were added after generating
command doc strings.
Before this change if logs were not redirected, logging would
corrupt the terminal screen.
This commit stores the logs (max ~100 lines) in an array and
print them when the program exits.
This ensures the virtual terminal processing mode is enabled on the rclone process
for Windows 10 consoles (by using Windows Console API functions GetConsoleMode/SetConsoleMode
and flag ENABLE_VIRTUAL_TERMINAL_PROCESSING), which adds native support for ANSI/VT100
escape sequences. This mode is default in many cases, e.g. when using the Windows
Terminal application, but in other cases it is not, and the default can also be
controlled with registry setting (see below), and therefore configuring it on the process
seem to be the only reliable way of ensuring it is enabled when supported.
[HKEY_CURRENT_USER\Console]
"VirtualTerminalLevel"=dword:00000001
Since rclone version 1.61.0 the tree command uses ANSI color sequences in output by
default, but this lead to issues in Windows terminals that were not handling these (#6668).
This commit ensures the tree command uses the terminal package for output. It relies on
go-colorable to properly handle ANSI color sequences: If stdout is connected to a terminal
the escape sequences are decoded and the text are written with color formatting using
Windows Console API. If stdout is not connected to a terminal, e.g. redirected to file,
the escape sequences are stripped off. The tree command has its own method for writing
directly to a file, specified with flag --output, and then the output is not passed
through the terminal package and must therefore be written without ansi codes.
Since version 3 of fuse libfuse no longer does anything when given the
nonempty option and it's default is to allow mounting over non empty
directories like normal mount does.
Some versions of libfuse give an error when using `--allow-non-empty`
which is annoying for the user.
We now do this check ourselves so we no longer need to pass the option
to libfuse.
Fixes#3562
Before this fix, we told cgofuse/WinFSP that the backend was case
insensitive but didn't implement the Getpath backend function to
return the normalised case of a file.
Resently cgofuse started implementing case insensitive files properly
but since we hadn't implemented Getpath, the file names were taking
the default of all in UPPER CASE.
This patch implements Getpath for cgofuse which fixes the case
problems.
This problem came to light when we upgraded cgofuse and WinFSP (to
1.12) which had the code to implement Getpath.
Fixes#6682
Before this change only serve http was Shutting down its server which
was causing other servers such as serve restic to leave behind their
unix sockets.
This change moves the finalisation to lib/http so all servers have it
and removes it from serve http.
Fixes#6648
Before this change, we started the http listener even if --stdio was
supplied.
This also moves the log message so the user won't see the serving via
HTTP message unless they are really using that.
Fixes#6646
In the lib/http refactor
52443c2444 restic: refactor to use lib/http
We forgot to serve the data and wait for the server to finish. This is
not tested in the unit tests as it is part of the command line
handler.
Fixes#6644Fixes#6647
The webdav library was confused by the Path manipulation done by
lib/http when stripping the prefix.
This patch adds the prefix back before calling it.
Fixes#6650
- add support for unix sockets (which skip the auth).
- add support for multiple listeners
- collapse unnecessary internal structure of lib/http so it can all be
imported together
- moves files in sub directories of lib/http into the main lib/http
directory and reworks the code that uses them.
See: https://forum.rclone.org/t/wip-rc-rcd-over-unix-socket/33619Fixes: #6605
The config question "Use auto config?" confused many users and lead to
recurring forum posts from users that were unaware that they were using
a remote or headless machine.
This commit makes the question and possible options more descriptive
and precise.
This commit also adds references to the guide on remote setup in the
documentation of backends using oauth as primary authentication.
* fs: add TerminalColorMode type
* fs: add new config(flags) for TerminalColorMode
* lib/terminal: use TerminalColorMode to determine how to handle colors
* Add documentation for '--terminal-color-mode'
* tree: remove obsolete --color replaced by global --color
This changes the default behaviour of tree. It now displays colors by
default instead of only displaying them when the flag -C/--color was
active. Old behaviour (no color) can be achieved by setting --color to
'never'.
Fixes: #6604
The previous version used values after the maximum Unicode code-point
to encode a key. This could lead to an overflow since a key is a int16,
a rune is int32 and the maximum Unicode code-point is larger than int16.
A better solution is to simply use negative runes for keys.
Before this change if we copied files of unknown size, then they lost
their metadata.
This was particularly noticeable using --s3-decompress.
This change adds metadata to Rcat and RcatSized and changes Copy to
pass the metadata in when it calls Rcat for an unknown sized input.
Fixes#6546
The current default AnnounceInterval is too short, causing the
multicast domain to be flooded with NOTIFY announcements,
which may prevent other dlna devices from sleeping.
This change allows users to set the announcement interval,
and it's default value also increased to 12 minutes.
Even within the interval, rclone can still passively respond to
M-SEARCH requests from other devices.
Verify the http service listening address and the SSDP server
announcement address to prevent accidental listening of IPv6 addresses
that do not support dlna yet and may be globally accessible.
Unlistened addresses on the interface will also be filtered out of the
SSDP announcement to avoid misleading other services in the multicast domain.
Before this change, if the a mount was created via the rc but unmounted
externally with `fusermount -u` say, rclone would still believe the mount
was active when it wasn't.
Changes in github.com/anacrolix/dms changed upnp.ServiceURN to include a
namespace identifier. This identifier was previously hardcoded, but is
now parsed out of the URN. The old SOAP action header parsing logic was
duplicated in rclone and did not handle this field. Resulting responses
included a URN with an empty namespace identifier, breaking clients.
Before this change, rclone serve sftp operating with a new rclone
after the md5sum/sha1sum detection was reworked to just run a plain
`md5sum`/`sha1sum` command in
3ea82032e7 sftp: support md5/sha1 with rsync.net #3254
Failed to signal to the remote that md5sum/sha1sum wasn't supported as
in
71e172a139 serve/sftp: support empty "md5sum" and "sha1sum" commands
We unconditionally return good hashes even if the remote being served
doesn't support the hash type in question.
This fix checks the hash type is supported and returns an error
MD5 hash not supported
When the backend is first contacted this will cause the sftp backend
to detect that the hash type isn't available.
Unfortunately this may have cached the wrong state so editing or
remaking the config may be necessary to fix it.
strings.Title has been deprecated since Go 1.18 and an alternative has been
available since Go 1.0. The rule Title uses for word boundaries does not handle
Unicode punctuation properly. Use golang.org/x/text/cases instead.
This change ensures we call the Shutdown method on backends when
they drop out of the fs/cache and at program exit.
Some backends implement the optional fs.Shutdowner interface. Until now,
Shutdown is only checked and called, when a backend is wrapped (e.g.
crypt, compress, ...).
To have a general way to perform operations at the end of the backend
lifecycle with proper error handling, we can call Shutdown at cache
clear time.
We add a finalize hook to the cache which will be called when values
drop out of the cache.
Previous discussion: https://forum.rclone.org/t/31336
