This isn't doing anything. Channels should get cleaned up by the GC when
the last reference to them disappears, just like all other data
structures. Also inlined BufferPool.Put in Buffer.Release, its only
caller.
This can be caused when the test has uploaded four blobs, then queues
two blobs for upload which are delayed. Then a seventh file can be
opened which lead to a test failure.
Before, the scanner would could files twice if they were included in the
list of backup targets twice, e.g. `restic backup foo foo/bar` would
could the file `foo/bar` twice.
This commit uses the tree structure from the archiver to run the
scanner, so both parts see the same files.
When the tomb is created with a canceled context, then the workers
started via `t.Go` exist nearly immediately. Once for the first time all
started goroutines have been stopped, it is not allowed to issue further
calls to `t.Go`. This is a problem when the started goroutines exit
immediately, as for example the first goroutine might already have
stopped before starting the second one, which is not allowed as once the
first goroutines has stopped no goroutines were running.
To fix this race condition the startup and main task of the archiver now
also run within a `t.Go` function. This also allows unifying the error
handling as it is no longer necessary to distinguish between errors
returned by the workers or the saveTree processing. The tomb now just
returns the first error encountered, which should also be the most
descriptive one.
If the context was canceled then saveTree might receive a treeID or not
depending on the timing. This could cause saveTree to incorrectly return
a nil treeID as valid. Fix this always returning an error when the
context was canceled in the meantime.
A canceled background context lets the blob/tree/fileSavers exit
without reporting an error. The error handling previously replaced
a 'context canceled' error received by the main backup method with
the error reported by the savers. However, in case of a canceled
background context that error is nil, causing restic to loose the
error and save a snapshot with a nil tree.
The archiver first called the Select function for a path before checking
whether the Lstat on that path actually worked. The RejectFuncs in
exclude.go worked around this by checking whether they received a nil
os.FileInfo. Checking first is more obvious and requires less code.
Previously the directory stats were reported immediately after calling
`SaveDir`. However, as the latter method saves the tree asynchronously
the stats were still initialized to their nil value. The stats are now
reported via a callback similar to the one used for the fileSaver.
- The SaveBlob method now checks for duplicates.
- Moves handling of pending blobs to MasterIndex.
-> also cleans up pending index entries when they are saved in the index
-> when using SaveBlob no need to care about index any longer
- Always check for full index and save it when storing packs.
-> removes the need of an index uploader
-> also removes the verbose "uploaded intermediate index" messages
- The Flush method now also saves the index
- Fix race condition when checking and saving full/non-finalized indexes
The Save methods of the BlobSaver, FileSaver and TreeSaver return early
on when the archiver is stopped due to an error. For that they select on
both the tomb.Dying() and context.Done() channels, which can lead to a
race condition when the tomb is killed due to an error: The tomb first
closes its Dying channel before canceling all child contexts.
Archiver.SaveDir only aborts its execution once the context was
canceled. When the tomb killing is paused between closing its Dying
channel and canceling the child contexts, this lets the
FileSaver/TreeSaver.Save methods return immediately, however, ScanDir
still reads further files causing the test case to fail.
As a killed tomb always cancels all child contexts and as the Savers
always use a context bound to the tomb, it is sufficient to just use
context.Done() as escape hatch in the Save functions. This fixes the
mismatch between SaveDir and Save.
Adjust the tests to use contexts bound to the tomb for all interactions
with the Savers.
Archivers TestMetadataChanged incorrectly clears the Extended Attributes
from the expected metadata of the temporary file. This is incorrect as on
SELinux enabled filesystem, as the kernel will automaticly add a SElinux
label. However, since ExtendedAttributes{} != ExtendedAttributes{nil} we
still need to clear them if there are no attributes found.
internal/archiver.readdir and internal/fs.ReadDir were unused.
internal/fs.ReadDirNames and internal/archiver.readdirnames were doing
nearly the same thing, except one sorted its output and opened with
fs.O_NOFOLLOW. Both were only used in internal/archiver.
Each of the random test files was split into the same five blobs. The
test fails once the fifth blob is passed on to `SaveBlob`. That is for
certain interleavings of goroutine execution it would be possible for
the test to trigger the testErr just after storing the first file.
The fixed test uses a different file content for each of the nine files
and fails after writing the fourth blob. The file content is also small
enough to ensure that for each file only a single blob is saved. This
guarantees that the test cannot fail before reading the first four
files. FileReadConcurrency = 2 allows up to two files queued for
processing. Therefore the test can at most open the sixth file before it
has to save the fourth file / blob which triggers the testErr.
In some (rare) cases "fake" UID 51234 may exist in a system running a
test. When this is the case, `cmp.Equal(want, node3)` will fail based on
difference between empty string and an actual username present in a
system.
Fixes github issue #2372
Since I could not remember what the value for `Check` means this commit
renames it to `SameFile`: when set to true, the test should make sure
that `FileChanged` should return false (=file is unmodified).
This commit changes the internal file system implementation for reading
data from stdin, it now returns an error when no bytes could be read. I
think it's worth failing in this case, the user instructed restic to
read some data from stdin, and no data was read at all. Maybe it was in
a pipe and some earlier stage failed.
See #2135 for a short discussion.
When the scanner is slower than the actual backup, the tomb cancels the
context passed to Scan(), which then returns ctx.Err(). In the end, the
main function prints an error message that is not helpful ("Context
cancelled") and exits with an error code although no error occurred.
The code now ignores the error in the context and just uses it for
cancellation. The scanner is not supposed to return an error anyway.
Closes #1978
Adds a SelectByName method to the archive and scanner which only require
the filename as input, and can thus be run before calling lstat on the
file. Can speed up scanning significantly if a lot of filename excludes
are used.
When the archiver is faster than the scanner, restic deadlocks. This
commit adds a `finished` channel to the struct in `ui/backup.go` so that
scanner results are ignored when the archiver is already finished.
Closes #1834
This commit changes how the worker goroutines for saving e.g. blobs
interact. Before, it was possible to get stuck sending an instruction to
archive a file or dir when no worker goroutines were available any more.
This commit introduces a `done` channel for each of the worker pools,
which is set to the channel returned by `tomb.Dying()`, so it is closed
when the first worker returned an error.
This commit changes the archiver so that low-level errors saving data to
the repo are returned to the caller (instead of being handled by the
error callback function). This correctly bubbles up errors like a full
temp file system and makes restic abort early and makes all other worker
goroutines exit.
The previous code tried to be as efficient as possible and only do a
single open() on an item to save, and then fstat() on the fd to find out
what the item is (file, dir, other). For normal files, it would then
start reading the data without opening the file again, so it could not
be exchanged for e.g. a symlink.
This behavior starts the watchdog on my machine when /dev is saved
with restic, and after a few seconds, the machine reboots.
This commit reverts the behavior to the strategy the old archiver code
used: run lstat(), then decide what to do. For normal files, open the
file and then run fstat() on the fd to verify it's still a normal file,
then start reading the data.
The downside is that for normal files we now do two stat() calls
(lstat+fstat) instead of only one. On the upside, this does not start
the watchdog. :)