This change is caching HMAC hashers for repeated use in subsequent
requests and chunks, so they don't have to be initialized from
scratch every time.
On my local computer this gives me ~5-6 times faster signature
calculation and ~5-6.5% more throughput in S3 requests. The smaller
the payload the better the throughput gets.
A POSIX filesystem does not implicitly create parent directories when
they do not exist. Directories must be explicitly created and
permissions be set.
This also fixes a bug where asynchronous operations would create a
file in the filer before the parent directory was created. If the
file was a symlink or another special type of file the directory
would inherit that type and become unusable in the mounted FS.
When multiple filer requests are in-flight and the current filer
disappears and a new one is selected by the first goroutine, then
there can be a lot of race conditions while retrieving the current
filer.
Therefore, load/save the current filer index atomically.
Sometimes when an unexpected error occurs the cacher would set an
error and return. However, it would not broadcast the condition
signal in that case, therefore leaving the goroutine that runs
readChunkAt stuck forever.
I figured that the condition is unnecessary because readChunkAt is
acquiring a lock that is still held by the cacher goroutine anyway.
Callees of startCaching have to wait for a WaitGroup which makes sure
that readChunkAt can't acquire the lock before startCaching.
This way readChunkAt can execute normally and check for the error.
* Fix FUSE server buffer leaks in file gaps
This change zeros read buffers when encountering file gaps during
file/chunk reads in FUSE mounts.
It prevents leaking internal buffers of the FUSE server which could
otherwise reveal metadata, directory listings, file contents and
other data related to FUSE API calls.
The issue was that buffers are reused, but when a file gap was found
the buffer was not zeroed accordingly and the existing data of the
buffer was kept and returned.
* Move zero logic into its own method
This fixes the calculation of the amount of EC shards a node holds.
Previously a global counter was increased, but also used inside the
loop to apply disk usage deltas. This led to wrong absolute numbers.
The fix is to apply only deltas of single EC shards per iteration.
Currently the file size of only one volume location is taken into
account in the stats. This commit multiplies the disk usages by the
amount of nodes holding a replica of the volume.
This will yield the expected amount of disk usage and matches the
total size calculations from before.
Patrick Schmidt