This changes the interface to NewObject so that if NewObject is called
on a directory then it should return fs.ErrorIsDir if possible without
doing any extra work, otherwise fs.ErrorObjectNotFound.
Tested on integration test server with:
go run integration-test.go -tests backend -run TestIntegration/FsMkdir/FsPutFiles/FsNewObjectDir -branch fix-stat -maxtries 1
Betweeen rclone v1.54 and v1.55 there was an approx 3x performance
regression when transferring to distant SFTP servers (in particular
rsync.net).
This turned out to be due to the library github.com/pkg/sftp rclone
uses. Concurrent writes used to be enabled in this library by default
(for v1.12.0 as used in rclone v1.54) but they are no longer enabled
(for v1.13.0 as used in rclone v1.55) for safety reasons and it is
necessary to enable them specifically.
The safety concerns are due to the uncertainty as to whether writes
come in order and whether a half completed file might have holes in
it. This isn't a problem for rclone since a) it doesn't restart
uploads and b) it has a post-transfer checksum test.
This change introduces a new flag `--sftp-disable-concurrent-writes`
to control the feature which defaults to false, meaning that
concurrent writes are enabled as in v1.54.
However this isn't quite enough to fix the problem as the sftp library
needs to be able to sniff the size of the stream from the reader
passed in, so this also adds a `Size` interface to the reader to
enable this. This involved a patch to the library.
The library was reverted to v1.12.0 for v1.55.1 - this patch installs
v1.13.0+master to fix the Size interface problem.
See: https://github.com/pkg/sftp/issues/426
In
a3fcadddc8 sftp: close idle connections after --sftp-idle-timeout (1m by default)
Idle SFTP connections were closed after 1 minute. However due to the
way SSH multiplexes connections over a single SSH connection this
meant that if uploads or downloads went on for more than one minute
they failed with "EOF errors" as their underlying connection was
closed.
This fixes the problem by not clearing idle connections if there are
any transfers in progress.
Fixes#5197
This reverts the library update done in this commit.
713f8f357d sftp: fix "file not found" errors for read once servers
Reverting this commit triples the performance to a far away sftp server.
See: https://github.com/pkg/sftp/issues/426
Some sftp servers don't allow the user to access the file after upload.
In this case the error message indicates that using
--sftp-set-modtime=false would fix the problem. However it doesn't
because SetModTime does a stat call which can't be disabled.
Update SetModTime failed: SetModTime stat failed: object not found
After upload this patch checks for an `object not found` error if
set_modtime == false and ignores it, returning the expected size of
the object instead.
It also makes SetModTime do nothing if set_modtime = false
https://forum.rclone.org/t/sftp-update-setmodtime-failed/22873
It introduces a new flag --sftp-disable-concurrent-reads to stop the
problematic behaviour in the SFTP library for read-once servers.
This upgrades the sftp library to v1.13.0 which has the fix.
This is done by making fs.Config private and attaching it to the
context instead.
The Config should be obtained with fs.GetConfig and fs.AddConfig
should be used to get a new mutable config that can be changed.
This adds a context.Context parameter to NewFs and related calls.
This is necessary as part of reading config from the context -
backends need to be able to read the global config.
As reported in
https://github.com/rclone/rclone/issues/4660#issuecomment-705502792
After switching to a password callback function, if the ssh connection
aborts and needs to be reconnected then the user is-reprompted for their
password. Instead we now remember the password they entered and just give
that back. We do lose the ability for them to correct mistakes, but that's
the situation from before switching to callbacks. We keep the benefits
of not asking for passwords until the SSH connection succeeds (right
known_hosts entry, for example).
This required a small refactor of how `f := &Fs{}` was built, so we can
store the saved password in the Fs object
Based on Issue 4087
https://github.com/rclone/rclone/issues/4087
Current behaviour is insecure. If the user specifies this value then we
switch to validating the server hostkey and so can detect server changes
or MITM-type attacks.
Before this change rclone used the relative path from the current
working directory.
It appears that WS FTP doesn't like this and the openssh sftp tool
also uses absolute paths which is a good reason for switching to
absolute paths.
This change reads the current working directory at startup and bases
all file requests from there.
See: https://forum.rclone.org/t/sftp-ssh-fx-failure-directory-not-found/17436
For SSH authentication, `key_pem` should both override `key_file`
and not require other SSH authentication methods to be set.
Prior to this fix, rclone would attempt to use an ssh-agent
when `key_pem` was the only SSH authentication method set.
Fixes#4240
Before this change we early exited the SetModTime call which means we
skipped reading the info about the file.
This change reads info about the file in the SetModTime call even if
we are skipping setting the modtime.
See: https://forum.rclone.org/t/sftp-and-set-modtime-false-error/16362
This error started happening after updating golang/x/crypto which was
done as a side effect of:
3801b8109 vendor: update termbox-go to fix ncdu command on FreeBSD
This turned out to be a deliberate policy of making
ssh.ParsePrivateKeyWithPassphrase fail if the passphrase was empty.
See: https://go-review.googlesource.com/c/crypto/+/207599
This fix calls ssh.ParsePrivateKey if the passphrase is empty and
ssh.ParsePrivateKeyWithPassphrase otherwise which fixes the problem.
This also corrects the symlink detection logic to only check symlink
files. Previous to this it was checking all directories too which was
making it do more stat calls than was necessary.
This bug was introduced as part of adding context to the backends and
slipped through the net because the About call did not have an
interface assertion in the sftp backend.
I checked there were no other missing interface assertions on all the
optional methods on all the backends.
- Change rclone/fs interfaces to accept context.Context
- Update interface implementations to use context.Context
- Change top level usage to propagate context to lover level functions
Context propagation is needed for stopping transfers and passing other
request-scoped values.
Before this change we used Remove to remove directories. This works
fine on Unix based systems but not so well on Windows based ones.
Swap to using RemoveDirectory instead.
This unifies the 3 methods of reading config
* command line
* environment variable
* config file
And allows them all to be configured in all places. This is done by
making the []fs.Option in the backend registration be the master
source of what the backend options are.
The backend changes are:
* Use the new configmap.Mapper parameter
* Use configstruct to parse it into an Options struct
* Add all config to []fs.Option including defaults and help
* Remove all uses of pflag
* Remove all uses of config.FileGet
This was caused by using the sftp.File.Read method which resets the
streaming window after each call. Replacing it with sftp.File.WriteTo
and an io.Pipe fixes the problem bringing the speed to the same as the
sftp binary.
If md5sum/sha1sum fails we debug what it outputed on stderr and return
an empty hash indicating we didn't have a hash, rather than
hash.ErrUnsupported indicating that we don't support this hash type.
This fixes lots of ERROR messages for sftp and synology NAS which,
while it supports md5sum the SFTP paths and the SSH paths are
different so md5sum doesn't work.
We also stop disabling md5sum/sha1sum on errors since typically Hashes
is only checked at the start of a sync run and isn't expected to
change dynamically.
The sftp library delivers the attributes of the symlink rather than
the object pointed to in directory listings, however when we use Stat
from the library it points to the objects.
Previous to this fix this caused items pointed to by symlinks to be
unusable.
After the fix both symlinked files and directories work as expected.
A Range request can never request 0 bytes however this change was made
to make a clearer signal that the limit means read to the end.
Add test and more documentation and fixup uses
The purpose of this is to make it easier to maintain and eventually to
allow the rclone backends to be re-used in other projects without
having to use the rclone configuration system.
The new code layout is documented in CONTRIBUTING.