Re: [PATCH 0/3] map multiple blocks in get_block() and mpage_readpages()

[Jeremy Higdon <jeremy@sgi.com>]

On Tue, Feb 21, 2006 at 08:59:53AM +1100, Nathan Scott wrote:
> On Mon, Feb 20, 2006 at 01:21:27PM -0800, Badari Pulavarty wrote:
> 
> I've been running these patches in my development tree for awhile
> and have not seen any problems.  My one (possibly minor) concern
> is that we pass get_block a size in units of bytes, e.g....
> 
> 	bh->b_size = 1 << inode->i_blkbits;
> 	err = get_block(inode, block, bh, 1);
> 
> And b_size is a u32.  We have had the situation in the past where
> people (I'm looking at you, Jeremy ;) have been issuing multiple-
> gigabyte direct reads/writes through XFS.  The syscall interface
> takes an (s)size_t in bytes, which on 64 bit platforms is a 64 bit
> byte count.
> 
> I wonder if this change will end up ruining things for the lunatic
> fringe issuing these kinds of IOs?  Maybe the get_block call could

Hey!  Lunatic fringe indeed.  Harumph!  :-)

Yes, there are a few people out there who will need to issue really
large I/O reads or writes to get maximum I/O bandwidth on large
stripes.  The largest I've done so far is 4GiB, but I expect that
number will likely increase this year, and more likely next year,
if not.

jeremy