Re: [fuse-devel] fallocate for fuse ?

[Han-Wen N. <ha...@gm...>]

On Fri, Sep 9, 2011 at 4:36 PM, Miklos Szeredi <mi...@sz...> wrote:
>>>> Apparently, the program is assuming that fallocate will create space
>>>> for it to write to through the mmap. It doesn't fall back to something
>>>> else, and the write to the mmapped address fails with sigbus.
>>>
>>> Very few filesystems currently implement fallocate, so fuse is probably
>>> not the only fs that this aplicateion will fail with.
>>
>> It's a version of GNU ld. The failure surprised me as well, but maybe
>> they use configure time checks to decide what to do here.
>
> Interesting.  Ext3 doesn't have fallocate, for example, but ld
> obviously does work on ext3.  So apparently ld decided that the fuse
> filesystem in question does support fallocate and didn't even check
> the return value.  I guess this is a bug and should be reported...

I dug around, and there is a bug in ld (more precisely, gold), in that
it misinterprets errorcodes of posix_fallocate(), and tries to
continue, with a crash result.

The root problem here was that my filesystem did not correctly handle
statfs, which posix_fallocate() uses for guessing blocksize. If
statfs() fails,  posix_fallocate() will fail too.

-- 
Han-Wen Nienhuys - ha...@xs... - http://www.xs4all.nl/~hanwen