Before this change we read sleepTime before acquiring the pacer token
and uses that possibly stale value to schedule the token return. When
many goroutines enter while sleepTime is high (e.g., 10s), each
goroutine caches this 10s value. Even if successful calls rapidly
decay the pacer state to 0, the queued goroutines still schedule 10s
token returns, so the queue drains at 1 req/10s for the entire herd.
This can create multi‑minute delays even after the pacer has dropped
to 0.
After this change we refresh the sleep time after getting the token.
This problem was introduced by the desire to skip reading the pacer
token entirely when sleepTime is 0 in high performance backends (eg
s3, azure blob).
It was possible in the presence of --max-connections and recursive
calls to the pacer to deadlock it leaving all connections waiting on
either a max connection token or a pacer token.
This fixes the problem by making sure we return the pacer token on
schedule if we take it.
This also short circuits the pacer token if sleepTime is 0.
* Adds "aix/ppc64" to the cross-compile target list.
* Including AIX in the build tag of "metadata_other.go".
* Excluding AIX from the main ncdu build tags.
* Marking AIX as an unsupported platform for ncdu.
* Excluding AIX from the fallback redirect implementation.
* Excluding AIX from unix build tags to avoid undefined unix.WNOHANG.
If the pacer was used recursively and --max-connections was in use
then it could deadlock if all the connections were in use at the time
of recursive call (likely).
This affected the azureblob backend because when it receives an
InvalidBlockOrBlob error it attempts to clear the condition before
retrying. This in turn involves recursively calling the pacer.
This fixes the problem by skipping the --max-connections check if the
pacer is called recursively.
The recursive detection is done by stack inspection which isn't ideal,
but the alternative would be to add ctx to all >1,000 pacer calls. The
benchmark reveals stack inspection takes about 55nS per stack level so
it is relatively cheap.
This turned out to be a problem in the tests. The tests used to do
1. allocate
2. increment
3. free
4. decrement
But if one goroutine had just completed 2 and another had just
completed 3 then this can cause the test to register too many
allocations.
This was fixed by doing the test in this order instead:
1. allocate
2. increment
3. decrement
4. free
The 4 operations are atomic.
Fixes#8813
Before this change we used an overcomplicated method of memory
reservations in the pool.RW which caused deadlocks.
This changes it to use a much simpler reservation system where we
actually reserve the memory and store it in the pool.RW. This allows
us to use the semaphore.Weighted to count the actually memory in use
(rather than the memory in use and in the cache). This in turn allows
accurate use of the semaphore by users wanting memory.
Because multipart transfers can need more than one buffer to complete,
if transfers was set very high, it was possible for lots of multipart
transfers to start, grab fewer buffers than chunk size, then deadlock
because no more memory was available.
This fixes the problem by introducing a reservation system which the
multipart transfer uses to ensure it can reserve all the memory for
one chunk before starting.
All user visible Durations should be fs.Duration rather than time.Duration. Suffix is then optional and defaults to s. Additional suffices d, w, M and y are supported, in addition to ms, s, m and h - which are the only ones supported by time.Duration. Absolute times can also be specified, and will be interpreted as duration relative to now.
Before this change, if RetryAfterError was called with a nil err, then
it's Error method would return this when wrapped in a fmt.Errorf
statement
error %!v(PANIC=Error method: runtime error: invalid memory address or nil pointer dereference))
Looking at the code, it looks like RetryAfterError will usually be
called with a nil pointer, so this patch makes sure it has a sensible
error.
This change adds a truncate_bytes mode which counts the number of bytes, as
opposed to the number of UTF-8 characters. This can be useful for ensuring that a
crypt-encoded filename will not exceed the underlying backend's length limits
(see https://forum.rclone.org/t/any-clear-file-name-length-when-using-crypt/36930 ).
This change also adds support for _keep_extension when using truncate and
truncate_bytes.
Before this change the help for convmv was generated by running the
examples each time rclone started up. Unfortunately this involved
running the echo command which did not work on Windows.
This pre-generates the help into `transform.md` and embeds it. It can
be re-generated with `go generate` which is a better solution.
See: https://forum.rclone.org/t/invoke-of-1-70-0-complains-of-echo-not-found/51618
lib/transform adds the transform library, supporting advanced path name
transformations for converting and renaming files and directories by applying
prefixes, suffixes, and other alterations.
It also adds the --name-transform flag for use with sync, copy, and move.
Multiple transformations can be used in sequence, applied in the order they are
specified on the command line.
By default --name-transform will only apply to file names. The means only the leaf
file name will be transformed. However some of the transforms would be better
applied to the whole path or just directories. To choose which which part of the
file path is affected some tags can be added to the --name-transform:
file Only transform the leaf name of files (DEFAULT)
dir Only transform name of directories - these may appear anywhere in the path
all Transform the entire path for files and directories
Example syntax:
--name-transform file,prefix=ABC
--name-transform dir,prefix=DEF
Nick Craig-Wood
yyandGitHub
Nicolas DessartandNick Craig-Wood
dougalandNick Craig-Wood
albertony
Lakshmi-SurekhaandGitHub
viochaandNick Craig-Wood
Oleksandr RedkoandGitHub
vastonusandGitHub
iTroozandGitHub
nielashandNick Craig-Wood
cuiandGitHub
Ed Craig-WoodandGitHub