Remove the POSIX_FADV_DONTNEED and POSIX_FADV_SEQUENTIAL calls
from the local backend. The DONTNEED calls cause severe spinlock
contention on parallel file systems (and any system with many
concurrent transfers), because each call triggers per-page cache
teardown under a global lock.
Observed on a 256-core system running rclone with 64 parallel
transfers over Lustre: 69% of all CPU cycles were spent in
kernel spinlock contention from the fadvise path, with effective
throughput well below hardware capability.
The kernel's own page reclaim (kswapd) handles eviction more
efficiently from a single context. Since rclone does not always
read files sequentially (e.g. multipart uploads rewind and
re-read blocks), FADV_SEQUENTIAL was also not reliably correct.
This is consistent with the non-Linux behavior (which never
called fadvise) and with restic's decision to remove identical
code (restic/restic#670).
Fixes#7886
This PR optimizes the PROPFIND requests in the webdav backend to only ask for
the specific properties rclone actually needs.
Currently, the generic webdav backend sends an empty XML body during directory
listing (listAll), which causes the server to fall back to allprops by default.
This forces the server to return properties we never use, such as
getcontenttype.
Fetching getcontenttype can be a very expensive operation on the server side.
For instance, in the official golang.org/x/net/webdav library, determining the
content type requires the server to open the file and read the first 500 bytes.
For a directory with 1,300 files in my environment, rclone ls time dropped from
~30s to ~4s (as fast as native ls).
This only applies to the other vendor for backwards compatibility which could be
expanded.
AWS S3 requires Content-MD5 for PutObject with Object Lock parameters.
Since rclone passes a non-seekable io.Reader, the SDK cannot compute
checksums automatically. Buffer the body and compute MD5 manually for
singlepart PutObject and presigned request uploads when Object Lock
parameters are set. Multipart uploads are unaffected as Object Lock
headers go on CreateMultipartUpload which has no body.
Add object_lock_supported provider quirk (default true) to allow
skipping Object Lock integration tests on providers with incomplete
S3 API support. Set to false for GCS which uses non-standard
x-goog-bypass-governance-retention header and doesn't implement
PutObjectLegalHold/GetObjectLegalHold.
Add Multipart and Presigned subtests to Object Lock integration tests
to cover all three upload paths.
Fixes#9199
URLPathEscapeAll was only passing [A-Za-z0-9/] through unencoded, causing
it to percent-encode RFC 3986 unreserved characters (-, ., _, ~). Per RFC
3986 §2.3, unreserved characters MUST NOT be percent-encoded, and a URI
that unnecessarily encodes them is not equivalent to one that does not.
Servers that perform strict path matching without normalising
percent-encoded characters will reject the over-encoded form with a 404.
Before: /files/my-report.pdf → /files/my%2Dreport%2Epdf
After: /files/my-report.pdf → /files/my-report.pdf
Reserved characters (spaces, semicolons, colons, etc.) continue to be
encoded as before.
When extsort.Strings() cannot create temporary files (e.g. due to
apparmor restrictions or permission denied), it returns a nil sorter
with the error on errChan. The code then called Sort() on the nil
sorter, causing a panic.
Check for nil sorter and return the error instead of panicking.
Fixes#9244
The format field was read in Handle() without synchronization while
setFormat() could write it concurrently from InitLogging(). This
caused a data race detected by the race detector, failing
TestListBucketsAuthProxy in cmd/serve/s3.
Fix by protecting all access to format with the existing mutex.
This adds 11 previously-missing rc params for newer bisync features.
It also makes optional parameters truly optional. (Previously, callers were
required to supply every single one, even if using the default value.)
This reduces the number of go routines which can get out of hand when
using large --transfers and --multi-thread-streams from potentially
--multi-thread-streams * --transfers Go routines to --max-memory /
--multi-thread-chunk-size
It serializes the memory allocator in each transfer which should be
good for performance and reduce lock contention.
Before this change server side copies would show at 0% until they were
done then show at 100%.
With support from the backend, server side copies can now be accounted
in real time. This will only work for backends which have been
modified and themselves get feedback about how copies are going.
If the transfer fails, the bytes accounted will be reversed.
Add gzip compression for directory listings and text assets served over HTTP.
This reduces the rclone repository file listing from 40 kB to 8 kB and reduces
the rclone MANUAL.txt from 2.7 MB to 700 kB.
This makes listings and assets served across the network load faster.
The compression level of 5 should be a good balance between size and speed.
