Tom Vier wrote: >(resending) > >are wandering logs like rcu? unless i'm reading it wrong, it sounds like you >create a copy of the node, but with the updated pointer. you go up the tree, >creating updated parents, and the whole new path is finally committed when >the highest node's pointer is updated, right? how big are nodes in v4? 512 >bytes? when updating a tree path, it causes a lot of single sector writes to >different areas (far apart?) but the only dependence is that all the new >updated versions of the nodes must be written before the top node is, right? >does v4 use extents? if not, how are the data blocks finally pointed to? > >sun has (or is writing) a new fs i read about recently that has data crc's >with error correction (i guess that's what the new "self healing" buzzword >means - i've heard ibm using it in commercials). i've thought about writing a >v4 plugin. i'm not sure how fine grained to make it. one crc for the whole >file isn't feasable. are blocks variably sized in reiserfs? perhaps each >extent or whatever could be crc'ed, with the crc's stored as metadata (using >whatever method v4 provides for saving variable length metadata). > > Putting a crc in every node would be easy to code and useful to users.
>also, what should i do if a crc is bad? just write to syslog? refuse to >return the data since it might be bad (return an error from read())? > > >
