Re: [Bug-ddrescue] Feature request: --direct for output file in fill mode

[Christian Franke]

Antonio Diaz Diaz wrote:
> Christian Franke wrote:
> > when using --fill=- to force reallocation of bad disk blocks, write 
> > behind buffering should be disabled.
> >
> > Therefore, it would make sense to allow "-d, --direct" in fill mode. 
> > It should set O_DIRECT for the output file in this case.
>
> I think buffered write is not a problem if followed by a call to 
> fsync, as ddrescue does.

I'm not sure. The block layer might e.g. legitimately buffer blocks in 
larger chunks. So writing a single block might result in:

1. Read large chunk.
2. Replace block data, mark chunk as dirty.
3. Write chunk later.

Step 1 would fail in case of a bad block within the chunk.

Are you sure that this is not the case for the platforms supported by 
ddrescue?


> OTOH, Linus hates O_DIRECT for writing 
> http://lkml.org/lkml/2007/1/11/129 , and I suppose he knows more about 
> it than me.

These complaints are about "normal" (e.g. database) applications 
bypassing kernel resource management by O_DIRECT. This is IMO unrelated 
to "rescue" operations in presence of bad disk blocks.

Christian