Sparse files contain large regions containing only zero bytes. Checking
that a blob only contains zeros is possible with over 100GB/s for modern
x86 CPUs. Calculating sha256 hashes is only possible with 500MB/s (or
2GB/s using hardware acceleration). Thus we can speed up the hash
calculation for all zero blobs (which always have length
chunker.MinSize) by checking for zero bytes and then using the
precomputed hash.
The all zeros check is only performed for blobs with the minimal chunk
size, and thus should add no overhead most of the time. For chunks which
are not all zero but have the minimal chunks size, the overhead will be
below 2% based on the above performance numbers.
This allows reading sparse sections of files as fast as the kernel can
return data to us. On my system using BTRFS this resulted in about
4GB/s.
The restorer can issue multiple calls to WriteAt in parallel. This can
result in unexpected orderings of the Truncate and WriteAt calls and
sometimes too short restored files.
We can either preallocate storage for a file or sparsify it. This
detects a pack file as sparse if it contains an all zero block or
consists of only one block. As the file sparsification is just an
approximation, hide it behind a `--sparse` parameter.
This writes files by using (*os.File).Truncate, which resolves to the
truncate system call on Unix.
Compared to the naive loop,
for _, b := range p {
if b != 0 {
return false
}
}
the optimized allZero is about 10× faster:
name old time/op new time/op delta
AllZero-8 1.09ms ± 1% 0.09ms ± 1% -92.10% (p=0.000 n=10+10)
name old speed new speed delta
AllZero-8 3.84GB/s ± 1% 48.59GB/s ± 1% +1166.51% (p=0.000 n=10+10)
`restic unlock` now only shows `successfully removed locks` if there were locks to be removed.
In addition, it also reports the number of the removed lock files.
Sending data through a channel at very high frequency is extremely
inefficient. Thus use simple callbacks instead of channels.
> name old time/op new time/op delta
> MasterIndexEach-16 6.68s ±24% 0.96s ± 2% -85.64% (p=0.008 n=5+5)
This is especially useful if ssh asks for a password or if closing the
initial connection could return an error due to a problematic server
implementation.
bazil/fuse expects us to return context.Canceled to signal that a
syscall was successfully interrupted. Returning a wrapped version of
that error however causes the fuse library to signal an EIO (input/output
error). Thus unwrap context.Canceled errors before returning them.
rclone can exit early for example when the connection to rclone is
relayed for example via ssh: `-o rclone.program='ssh user@example.org
forced-command'`
For backends which are able to atomically replace files, we just can
overwrite the old copy, if it is necessary to retry an upload. This has
the benefit of issuing one operation less and might be beneficial if a
backend storage, due to bugs or similar, could mix up the order of the
upload and delete calls.
When hard deleting the latest file version on B2, this uncovers earlier
versions. If an upload required retries, multiple version might exist
for a file. Thus to reliably delete a file, we have to remove all
versions of it.
Ignored packs were reported as an empty pack by EachByPack. The most
immediate effect of this is that the progress bar for rebuilding the
index reports processing more packs than actually exist.
Previously the buffer was grown incrementally inside `repo.LoadUnpacked`.
But we can do better as we already know how large the index will be.
Allocate a bit more memory to increase the chance that the buffer can be
reused in the future.
Instead of first checking whether a file is in the repository cache and
then opening it, we just can open the file. This saves one stat call. If
the file is in the cache, everything is fine and otherwise the code
follows its normal fallback path.
sort.Sort is not guaranteed to be stable. Go 1.19 has changed the
sorting algorithm which resulted in changes of the sort order. When
comparing snapshots with identical timestamp but different paths and
tags lists, there is not meaningful order among them. So just keep their
order stable.
Cleanly separate the directory presentation and the snapshot directory
structure. SnapshotsDir now translates the dirStruct into a format
usable by the fuse library and contains only minimal special case rules.
All decisions have moved into SnapshotsDirStructure which now creates a
fully preassembled tree data structure.
For large pack sizes we might be only interested in the first and last
blob of a pack file. Thus stream a pack file in multiple parts if the
gaps between requested blobs grow too large.
Also make the errors a bit less verbose by not prepending the operation,
since pkg/xattr already does that. Old errors looked like
Listxattr: xattr.list /myfiles/.zfs/snapshot: invalid argument
pkg/sftp.Client.MkdirAll(d) does a Stat to determine if d exists and is
a directory, then a recursive call to create the parent, so the calls
for data/?? each take three round trips. Doing a Mkdir first should
eliminate two round trips for 255/256 data directories as well as all
but one of the top-level directories.
Also, we can do all of the calls concurrently. This may reintroduce some
of the Stat calls when multiple goroutines try to create the same
parent, but at the default number of connections, that should not be
much of a problem.
FutureBlob now uses a Take() method as a more memory-efficient way to
retrieve the futures result. In addition, futures are now collected
while saving the file. As only a limited number of blobs can be queued
for uploading, for a large file nearly all FutureBlobs already have
their result ready, such that the FutureBlob object just consumes
memory.
There is no real difference between the FutureTree and FutureFile
structs. However, differentiating both increases the size of the
FutureNode struct.
The FutureNode struct is now only 16 bytes large on 64bit platforms.
That way is has a very low overhead if the corresponding file/directory
was not processed yet.
There is a special case for nodes that were reused from the parent
snapshot, as a go channel seems to have 96 bytes overhead which would
result in a memory usage regression.
