Re: Bio pool & scsi scatter gather pool usage

[David Mansfield]

> > > > > In EVMS, we are adding code to deal with BIO splitting, to
> > > > > enable our feature modules, such as DriveLinking, LVM, & MD
> > > > > Linear, etc to break large BIOs up on chunk size or lower
> > > > > level device boundaries.
> > > >
> > > > Could I suggest that this code not be part of EVMS, but that
> > > > you implement it as a library within the core kernel?  Lots of
> > > > stuff is going to need BIO splitting - software RAID, ataraid,
> > > > XFS, etc.  May as well talk with Jens, Martin Petersen, Arjan,
> > > > Neil Brown.  Do it once, do it right...
> > > >
> > > I take that back.
> > >
> > > We really, really do not want to perform BIO splitting at all.
> > > It requires that the kernel perform GFP_NOIO allocations at
> > > the worst possible time, and it's just broken.
> > >
> > > What I would much prefer is that the top-level BIO assembly
> > > code be able to find out, beforehand, what the maximum
> > > permissible BIO size is at the chosen offset.  It can then
> > > simple restrict the BIO to that size.
> > >

[snipped some ideas]

> 
> Why not just put the smallest required BIO size in a struct for that device?
> Then each read of that struct can be kept in cache...
> 
> Is the BIO max size going to change at different offsets?

My two cents as a non-guru: there are two different reasons for splitting 
a large BIO: 

1) some layer has hit some uncommon boundary condition, like spanning
linearly appended physical volumes in an LV or something like that

2) a fundamental 'maximum-chunkiness' allowed by some layer has been 
exceeded, like stripe size in a raid, or MAX_SECTORS in ide or something 
like that.

It would suck if the system generated large BIOS that needed to be split
for every IO operation, due to #2, but it would also suck to add overhead
to every IO operation for #1.

#1 is an exception, and I think it would be acceptible to have a splitting 
function/mempool for handling what should be a boundary condition only, 
and the concept of a call through the layers to find out #2 at open time 
would handle #2 one time per device or something like that.

David

-- 
/==============================\
| David Mansfield              |
| [e-mail]             |
\==============================/