We never expect the offset to be negative, so instead of keeping
it around as off_t, and having to cast to size_t to deal with
signed vs unsigned comparisons, we just use size_t directly.
There's still a few uses of off_t and ssize_t around, but
those will need further refactoring to clean up.
We were using -1 as a sentinel to determine if we had computed the
file descriptor limit, but rlim_t is unsigned. Also, RLIM_INFINITY
is typically -1 as well, so the choice of -1 as a sentinel was
probably not a good idea anyways.
Closes issue #19.
The 'apt' binary is not available on e.g. TravisCI, so lets stick to
using 'apt-get' for a smoother out-of-the-box build process.
The list of packages updated in df7c77a0 have been kept the same.
It's not the bundle that's opened by fuse, that's just our internal
pointer back to where the data lives. As far as fuse is concerned
the only file we have is sparsebundle.dmg, and that's the file that's
opened and released.
Instead of referencing SB_DATA all over the place, and hence also
calling fuse_get_context(), we now do it once of each function call,
through the renamed sparsebundle_current() macro.
This makes the code a bit more transparent and easier on the eye,
versus the more opaque 'data' type.
sparsebundlefs.cpp:171:56: error: 'pread' was not declared in this scope
sparsebundlefs.cpp:172:24: error: 'close' was not declared in this scope
We were missing unistd.h
FUSE 2.9 added support for zero copy reads/writes through two new
callbacks, read_buf and write_buf. We now imeplement read_buf,
which iterates over the bands using the generic code, but instead
of opening and reading the content of the band into memory, we just
open the file and pass a structure back to FUSE with the file
descriptor, size, and offset, and FUSE will then use this on the
kernel side if possible (using splice on GNU/Linux).
For padding with zeroes we open /dev/zero and treat it as any
other file.
Since FUSE does not provide a callback when it's done with the
buffers there's no way for us to know when to close the bands
we opened. To work around this we keep the files open until
either the main image (dmg) file has been closed, or we run
out of file descriptors, at which point we gc the open files
and continue.
If a read was requested at an offset just before the end of a band,
with a length sufficiently large to cross over to the next band, we
would pad the remaining space after the initial band with zeroes,
and never proceed to read the second band.
We now limit the size of each consecutive band read to the end of the
band, so that the while loop will correctly continue with the next
band if there's still more data to be read.
