Re: syncing a directory

[Jean-Yves Migeon <jean-yves.migeon%espci.fr@localhost>]

Edgar Fuß wrote:
> Given some combination of filesystem, disk interface et al., there is 
> a certain level of confidence to be achieved by fsync() that, after 
> modifying a bunch of bytes in a specific file, those bytes are to be 
> found in the modified state after a crash.
> So, given the same combination of filesystem etc., is there a method 
> to obtain a comparable level of confidence, that, after moving a bunch 
> of files to a specific directory, those files are still to be found 
> where they have been moved to after a crash? Other than fsync()ing 
> each of those files, I mean.
Sorry for this late arrival, but there is something I do not quite 
understand there: if dovecot is moving file from one directory to 
another, this requires write permission. So question is:
- why open() the directory read only to fsyn() it? From my PoV, there is 
no point in fsync()ing an fd which is opened read only, as you can not 
modify it anyway.
- if the copy/move operation went through rename() (thus, no fd was 
used), IMHO, the directory should be opened RW to fsync() it.

In case the move operation implied some modifications to the file, you 
should fsync() them separately. I am not sure that fsyncing a directory 
will also flush file caches for all inodes found under it.

Feel free to correct me :)

Cheers,

--
Jean-Yves Migeon
jean-yves.migeon%espci.fr@localhost