Context::find_goodbye_entry() is removed and incorporated into the lookup
callback in order to take advantage of the entry_cache and since it is only used
inside this callback.
All entries read on lookup are cached.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
Signed-off-by: Wolfgang Bumiller <w.bumiller@proxmox.com>
By storing the payload start offset in the `DirectoryEntry` and passing this
information to `Decoder::read()`, the payload can be read directly and a repeated
re-reading of the entry information is avoided.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
Signed-off-by: Wolfgang Bumiller <w.bumiller@proxmox.com>
listxattr must only return the name list, no extended attribute values.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
Signed-off-by: Wolfgang Bumiller <w.bumiller@proxmox.com>
ACLs are stored separately in the pxar archive. This implements the functionality
needed to read the ACLs and return them as extended attributes in the getxattr
callback.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
Signed-off-by: Wolfgang Bumiller <w.bumiller@proxmox.com>
The goodbye table of directory entries is cached in a LRU cache to speed up
subsequent accesses.
This is especially important for directories with many entries, as then the
readdirplus callback is called repeatedly because of the limited reply buffer
size.
`DirectoryEntry`s are cached for subsequent access in their own LRU cache,
independent of the goodbye tables.
In order to avoid borrow conflicts, the `Context` provides a fn as_mut_refs
as well as a fn run_with_context_refs.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
Signed-off-by: Wolfgang Bumiller <w.bumiller@proxmox.com>
By passing `&DirectoryEntry` to stat, the function interface is simplified.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
Signed-off-by: Wolfgang Bumiller <w.bumiller@proxmox.com>
This allows to read the target path of a symbolic link in the
Decoder::read_directory_entry() function and stores it in the DirectoryEntry.
By this the Decoder::read_link() function becomes obsolete and is therefore
removed.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
By ambiguously using the Decoder::read_directory_entry() the code is simplified
and reading of the DirectoryEntry is concentrated into Context::run_in_context().
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
Previously it was disciminated based on the entry mode.
For directories, the inode was the offset of the corresponding
goodbye tail mark while for all others it was the offset of the filename.
By simply using the start offset as calculated from the corresponding
goodbye table entry (which yields the archive offset of the filename),
the code is simplified and the more ambiguous read_directory_entry()
function can be used.
The disatvantage of this approach is the need to keep track of the
start and end offsets for each entry, as the end offset is needed in
order to access the goodbye table of directory entries.
The root node still has to be treated special, as it's inode is 1 as per fuse
definition and it has no filename as per the pxar file format definition.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
By not implementing readdir but only readdirplus, the FUSE_CAP_READDIRPLUS flag
is set while the FUSE_CAP_READDIRPLUS_AUTO flag is not set.
Thereby the kernel will issue only readdirplus calls.
Documentation at:
https://libfuse.github.io/doxygen/fuse-3_88_80_2include_2fuse__common_8h.html#a9b90333ad08d0e1c2ed0134d9305ee87
As the expensive part for accessing and reading the attributes is seeking and
decoding each directory entry, it is usefull to force readdirplus calls.
By this a struct `EntryParam` is returned for each entry, therebye avoiding a
subsequent lookup call.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
readdirplus returns the entries together with their `EntryParam`, so subsequent
lookups for each of the entries are avoided.
In order to reduce code duplication, the code for filling the reply buffer is
moved into a macro.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
Remove the current caching of attrs and goodbye tables as it is broken anyway.
This will be replaced with a LRU cache.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
By moving the HashMap into `Context`, the use of lazy_static as well as the
additional Mutex can be avoided (`Context` is already guarded by a Mutex).
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
Cache not only the goodbye table for the last directory but for each opened
directory.
The opendir fuse callback will fill the cache with the goodbye table and
releasedir will remove it from the cache.
This should reduce the number of chuncks fetched from the server in some cases.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
The goodbye table is layed out as binary search tree based on the hash, so use
this to be more efficient when looking up a hash in the table for directories
with a large number of entries.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
And use an extra functzion set_callback() to configure that.
Also rewrite pxar/fuse.rs and implement a generic Session (will get
further cleanups with next patches).
The expensive call to Decoder::read_directory_entry() can be omitted as
Decoder::attributes() returns all the information the fuse response needs.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
and add Session::from_decoder() in order to be able to create a fuse session
with a `Decoder` given as argument.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
This patch introduces `Context` to hold the decoder, ino_offset and caches for
the attributes and the goodbye table.
By caching, certain callbacks can be handled without the need to read additional
data via the decoder, which improves performance.
The searching of the goodbye table is refactored as well, avoiding recursive
function calls in case of a hash collision.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
The hash of the filename in the goodbye table items allows to quickly compare to
a hashed filename.
Unfortunately, a matching hash is no garantee for matching filenames as hash
collisions are possible.
This patch fixes such possible collisions by further checking the filenames once
a matching hash has been found.
This introduces no significant extra cost (except for the filename comparison)
for cases with matching hashes, as the lookup call has to seek and read the file
attributes (including the filename) anyway.
In cases with hash collision, the next matching item is read and treaded
analogously (what means we need at least one extra seek).
As collisions should be not that frequent, this should be an acceptable penalty.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
The functionality of stat is split into smaller sub-functions, which allows
to reuse them more flexible, as the code flow is similar but not always the same.
By this, the ugly and incorrect re-setting of the i-node in the lookup callback
function is avoided.
The correct i-node is now calculated beforehand and stat simply creates a
`libc::stat` struct from the provided parameters.
Also, this fixes incorrect i-node assignments in the readdir callback function.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
The lookup call checks if the given filename is found in the directory referenced
by the i-node by calclulating the filenames hash and looking it up in the
directories goodbye table.
If found, the entries parameters are returned.
In order to be able to lookup the parent offset by a given file offset in the
readdir callback, this also stores the corresponding values in a HashMap.
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
