On Wed June 29 2005 6:30 am, Kevin Corry wrote: > > in such situation, bio's > > bi_size could exceed strip target's chunk size and snap target's chunk > > size, but in current code, we never care about this. so in strip target > > we could read/write bad position after stripe_map() function, and cause > > same effect in dm-snap.c. > > I'm quite certain there is code in dm.c to prevent bio's from spanning the > internal boundaries in the mirror, snapshot, and stripe targets. Have you > read through the code starting in dm.c::dm_request()? The targets specify > where their internal boundaries are, and the core driver is responsible for > splitting up bio's so the targets never get a request that span those > boundaries. And now that I've said this, I've gone back and looked, and indeed the dm-raid1.c code isn't setting the ti->split_io field (it does get set in dm-stripe.c and dm-snap.c). Alasdair, should that field be set for dm-mirror targets? I'm not as familiar with that code, so I'll have to leave that up to you. If it does need to be set, I think the following patch should work. -- Kevin Corry kevcorry@xxxxxxxxxx http://www.ibm.com/linux/ http://evms.sourceforge.net/ Set the target's split_io field when building a dm-mirror device so incoming bio's won't span the mirror's internal regions. --- a/drivers/md/dm-raid1.c 2005-06-29 06:38:16.000000000 -0500 +++ b/drivers/md/dm-raid1.c 2005-06-29 06:41:59.000000000 -0500 @@ -1060,6 +1060,7 @@ } ti->private = ms; + ti->split_io = ms->rh->region_size; r = kcopyd_client_create(DM_IO_PAGES, &ms->kcopyd_client); if (r) {
