The previous iterator was implemented to simply fetch _all_ pathes from
the filesystem, no matter what.
With this implementation, this changes. The iterator now has
functionality to optimize the iteration, if only a subdirectory of the
store is required, for example `$STORE/foo`.
This is done via functionality where the underlying iterator gets
altered.
First of all, the interface was changed to return a `Entries` object,
which itself only covers the libimagstore-internal `PathIterator` type.
This type was changed so that the backend implementation provides an
"PathIterBuilder`, which builds the actual iterator object for the
`PathIterator` type.
The intermediate `StoreIdConstructingIterator` was merged into
`PathIterator` for simplicity.
The `Entries` type got functionality similar to the
`StoreIdIteratorWithStore` type for easier transition to the new API.
This should probably be removed at a later point, though.
As the `walkdir::WalkDir` type is not as nice as it could be, iterators
for two collections in the store could be built like this (untested):
store
.entries()?
.in_collection("foo")
.chain(store.entries()?.in_collection("bar"))
Functionality to exclude subdirectories is not possible with the current
`walkdir::WalkDir` implementation and has to be done during iteration,
with filtering (as usual).
Signed-off-by: Matthias Beyer <mail@beyermatthias.de>
This patch changes the filesystem-backend implementation of libimagstore
to open files on each read/write rather than holding the file handle in
memory at all times.
Whenever a lot of imag store entries are read into memory, the imag
process may ran out of file descriptors. With this patch applied, a
`Store::get()` call on an entry which is not yet in the store cache
would cause the file to be read, but the FD being dropped after that.
Likewise, a `Store::update()` (which is also called if the imag entry is
dropped) would re-open the file on the filesystem and write the contents
from the imag store cache back to the file.
With this patch, opening hundrets or thousands of imag entries should be
no problem anymore, only the available memory should be a limit then.
Signed-off-by: Matthias Beyer <mail@beyermatthias.de>
This patch fixes a bug we did not even hit (yet). It is: When deleting
an Entry from the store, this could potentially leave artifacts in the
cache.
Szenario: An Entry, which was loaded (via `Store::get()` for example),
gets `Store::delete()`ed twice. The first call would work as expected,
but leave the Entry in the Store cache. The second call would then fail,
as the Entry is already removed on the FS, but still in the cache. This
would fail - which is the right thing to do here - but with the wrong
error (with a FileError rather than a FileNotFound error).
This patch fixes this.
First of all, the appropriate `PathBuf` object is calculated in all
cases, as this object is needed to check whether the file is actually
there (which could be the case if the Entry is in the cache and if it is
not).
If the entry is in the cache and is borrowed: error. If not, remove the
entry from the cache. Afterwards the file is deleted on disk.
If the entry is not in the cache, but the file exists, the file is removed.
If the file does not exist: error.
Signed-off-by: Matthias Beyer <mail@beyermatthias.de>
As of rustc 1.26, the `flatten()` method on iterators is preserved by
the rust standard library.
This could cause this code to hard-error some time in the future with
the `flatten()` function actually implemented by the standard library.
Hence we move to use the `Itertools::flatten()` function here
explicitely.
Because the iterators need to be able to check whether the file exists
_in the backend_ (not on disk, but in the backend, because of in-memory
test for example), we need to be able to pass the backend to the
iterator intermediate type.
This patch implements this. It does so by changing the internal backend
member of the store from `Box<FileAbstraction>` to
`Arc<FileAbstraction>`, which gives us the ability to clone the
reference to the backend easily without needing to rely on lifetimes
here, because of the Arc.
Also, less boxes are always good.
This patch reimplements the iterator extensions.
As we iterate (in StoreIdIterator) over Result<StoreId> now anyways, we
don't need the extensions for Result iterators anymore.
This patch rewrites the extensions to be more simple in every way and
generic over the error type in the iterator.
All the errors have to do is implement From<StoreError>, which is what
they do when linking the generated error types with error_chain to the
libimagstore error types.
This patch rewrites the Store::entries() function to not be collecting
the iterator.
It therefore introduces a new, internal, iterator type which creates the
StoreId objects from the pathes the PathIterator yields internally.
With this patch, the Store iterator interface changes, as the iterators
now yield `Result<StoreId, StoreError>` instead of `StoreId`.
This is necessary, as the internal conversion errors shouldn't be
hidden.
Of course, the iterator types (like the StoreGetIterator and so on)
should hold a Result<StoreId> internally as well, and also yield
appropritely. This was changed in this commit, too.
Before we had the problem that when iterating over _a lot_ (like 5k)
entries and also fetching them, at some point the OS would return with
"Too many files open".
That is because the store internally caches a lot.
With this change, the Store gets an API to query how big the cache is,
how much the cache can currently hold and (and that's the main thing in
this patch) to flush the cache to disk.
A function to simply ask the store whether its cache should be flushed
(which would us require to ask the OS how many files we can open...
which would be possible with `libc::getrlimit`) does not yet exist,
though, but could be added easily if desired.
The rust compiler does some fancy things for us: It automatically finds
the right fields if the name of the variable and the file is the same.
Lets use that to reduce boilerplate with this patch.
