Storj switched to a single global s3 endpoint backed by a BGP routing.
We want to stop advertizing the former regional endpoints and have the
global one as the only option.
Before this change, when a new object was created s3 returns its
versionID (on a versioned bucket) and rclone recorded it in the
object.
This means that when rclone came to delete the object it would delete
it with the versionID.
However it is common to forbid actions with versionIDs on buckets so
as to preserve the historical record and these operations would fail
whereas they succeeded in pre-v1.60.0 versions.
This patch fixes the problem by not recording versions of objects
supplied by the S3 API on upload unless `--s3-versions` or
`--s3-version-at` is used. This makes rclone behave as it did before
v1.60.0 when version support was introduced.
See: https://forum.rclone.org/t/s3-and-intermittent-403-errors-with-file-renames-and-drag-and-drop-operations-in-windows-explorer/34773
Before this change, we were taking the version ID straight from the
XML blob returned by the SDK and thus pinning the XML into memory
which bulked up the average memory per object from about 400 bytes to
4k.
Copying the string fixes the excess memory usage.
This reverts commit 4f386a1ccd.
It turns out that Alibaba OSS does support list v2 and the detection
code was wrong.
This means that users of the gov version of Alibaba will have to add
`list_version 1` to their config files.
See #6600
In this commit
ab849b3613 s3: fix listing loop when using v2 listing on v1 server
The ContinuationToken was tested for existence, but it is the
NextContinuationToken that we are interested in.
See: #6600
This was caused by
a9bd0c8de6 s3: reduce memory consumption for s3 objects
Which assumed that the StorageClass would always be set, but it isn't
set for Versions.
Before this change, rclone would enter a listing loop if it used v2
listing on a v1 server and the list exceeded 1000 items.
This change detects the problem and gives the user a helpful message.
Fixes#6600
Copying the storageClass string instead of using a pointer to the original string.
This prevents the Go garbage collector from keeping large amounts of
XMLNode structs and references in memory, created by xmlutil.XMLToStruct()
from the aws-sdk-go.
Before this change, some files were giving this error when downloaded
from Cloudflare and other providers.
ERROR corrupted on transfer: sizes differ NNN vs MMM
This is because these providers auto gzips the object when rclone
wasn't expecting it to. (AWS does not gzip objects without their being
uploaded gzipped).
This patch adds a quirk to for fix the problem and a flag to control
it. The quirk `might_gzip` is set to `true` for all providers except
AWS.
See: https://forum.rclone.org/t/s3-error-corrupted-on-transfer-sizes-differ-nnn-vs-mmm/33694/Fixes: #6533
Before this fix it was impossible to stop rclone generating an
X-Amx-Acl: header which is incompatible with GCS with uniform access
control and is generally deprecated at AWS.
The API endpoint GetBucketLocation requires
top level permission.
If we do an authenticated head request to a bucket, the bucket location will be returned in the HTTP headers.
Fixes#5066
Before this change if --user-server-modtime was in use the ModTime
could change for an object as we receive it accurate to the nearest ms
in listings, but only accurate to the nearest second in HEAD and GET
requests.
Normally AWS returns the milliseconds as .000 in listings, but if
versions are in use it may not. Storj S3 also seems to return
milliseconds.
This patch tries to keep the maximum precision in the last modified
time, so it doesn't update a last modified time with a truncated
version if the times were the same to the nearest second.
See: https://forum.rclone.org/t/cache-fingerprint-miss-behavior-leading-to-false-positive-stalen-cache/33404/
Before this fix, the chunksize calculator was using the previous size
of the object, not the new size of the object to calculate the chunk
sizes.
This meant that uploading a replacement object which needed a new
chunk size would fail, using too many parts.
This fix fixes the calculator to take the size explicitly.
Before this change, if an object compressed with "Content-Encoding:
gzip" was downloaded, a length and hash mismatch would occur since the
go runtime automatically decompressed the object on download.
If --s3-decompress is set, this change erases the length and hash on
compressed objects so they can be downloaded successfully, at the cost
of not being able to check the length or the hash of the downloaded
object.
If --s3-decompress is not set the compressed files will be downloaded
as-is providing compressed objects with intact size and hash
information.
See #2658
In
22abd785eb s3: implement reading and writing of metadata #111
The reading information of objects was refactored to use the
s3.HeadObjectOutput structure.
Unfortunately the code branch with `--s3-no-head` was not tested
otherwise this panic would have been discovered.
This shows that this is path is not integration tested, so this adds a
new integration test.
Fixes#6322
The SDK doesn't wrap errors in a Go standard way so they can't be
unwrapped and tested for - eg fatal error.
The code looks for a Serialization or RequestError and returns the
unwrapped underlying error if possible.
This fixes the fs/operations integration tests checking for fatal
errors being returned.
In this commit
e5974ac4b0 s3: use PutObject from the aws SDK to upload single part objects
rclone was made to upload objects to s3 using PUT requests rather than
using signed uploads.
However this change missed the fact that there is a supported way to
do this in the SDK using the SetStreamingBody method on the Request.
This therefore reverts a lot of the previous commit to do with making
an unsigned connection and other complication and uses the SDK
facility.
