Chris Kirby wrote: > Neil Perrin wrote: >> >> Chris Kirby wrote: >>> But doesn't the rw_enter code block if there's a waiting >>> writer? In other words, can't we cause a deadlock >>> if we already hold the read lock, there's a waiting writer, >>> and we try to grab the read lock again? >>> >>> -Chris >> >> >> I agree. As a thread, there's no way to determine if I already hold >> the read lock and regrabbing it can deadlock as you describe. >> The only options I see are for the callers to grab it or pass >> down an indication that they have it. > > It looks like there are currently 30 callers of dsl_dataset_open_obj. > > The only callers from open context that don't grab > the lock are dsl_dataset_create_root and > dsl_dataset_snapshot_rename_check. > > In sync context, we either have no lock or we have the write lock. > > So if we fix up dsl_dataset_create_root to grab the lock; and > change dsl_dataset_snapshot_rename_check not to do any work > unless we're syncing, I think we can say this: > > need_lock = !RW_WRITE_HELD && dsl_pool_sync_context(dp); > > Does that make sense? It has to be less painful than > adding an indication that the caller already holds the > lock. > > -Chris
Makes sense to me. I don't know the rest of the code well enough to know if this cane be more cleanly or elegantly fixed. Neil.
