dm-raid1: not to add newly-allocated RH_CLEAN region to clean_regions
RFC 1/1 dm-raid1: not to add newly-allocated RH_CLEAN region to clean_regions
Hi,
I would like to know whether the attached patch is something acceptable or not.
In dm-raid1.c, 'clean_regions' in region hash is a list used to link regions which are clean AND no longer interested (freeable). So I think it's safe not to add newly-allocated RH_CLEAN entry to the list because someone who allocates it has interest on the region.
In the current code, rh_alloc()/rh_find() is called from:
- rh_inc(), which makes the state RH_DIRTY if it's RH_CLEAN
- __rh_recovery_prepare(), which makes the state RH_RECOVERING
- rh_delay(), which is called when the region is RH_RECOVERING
So none uses the newly allocated RH_CLEAN region as-is. RH_DIRTY region is changed back to RH_CLEAN and then connected to clean_regions when pending I/O completes. RH_RECOVERING region is reclaimed when the recovery on the region completes.
If this patch is acceptable, we can also merge RH_CLEAN and RH_DIRTY into one state.
Comments are welcome.
Thanks,
Jun'ichi "Nick" Nomura
[RFC] dm-raid1: not to add newly-allocated RH_CLEAN region to clean_regions
The clean_regions list is used to maintain reclaimable hash entries. If the region is known to be marked RH_DIRTY or RH_RECOVERING later, it's not necessarily connected to clean_regions even if it's clean.
OTOH, region hash entries are allocated only when the region is going to be written (RH_DIRTY) or recoverred (RH_RECOVERING).
So it's safe not to add RH_CLEAN region to clean_regions in __rh_alloc(). Not adding to clean_regions protects the region from potential race condition that the hash entry is reclaimed by rh_update_states() before use.
drivers/md/dm-raid1.c +0 -5
@@ -255,11 +254,6 @@ static struct region *__rh_alloc(structelse {__rh_insert (rh, nreg); -if (nreg->state ==RH_CLEAN ) { -spin_lock (&rh->region_lock); -list_add (&nreg->list, &rh->clean_regions); -spin_unlock (&rh->region_lock); - } reg = nreg; }write_unlock_irq (&rh->hash_lock);