Checksum verification during a file read

[Jinnan-Guo]

Hello,

I was wondering how the checksum is verified during a file read.

Assume a linux user is reading a file from the dataset. If I understand correctly, the zfs will get the target dnode according to the linux vfs inode. And the zio tree will be constructed in the following order:

root zio -> zio read on dnode -> zio read on indirect node(s) -> zio read on file data

If my understanding is correct, is the checksum verification starts from the target file data and goes up to the dnode?

Thanks for your help.

[amotin]

ZIO tree is not built all at once the way you describe it, but the idea is right. Since checksum is a part of any block pointer, any read can be verified. The moment you import the pool, you implicitly trust one single block pointer stored in uberblock. After that every new block pointer you use to read something is stored in some other block that you have already verified with checksums from its respective block pointer, etc. So the chain of protection is: Data block -> indirect block(s) -> dnodes block -> dnodes indirect block(s) -> objset -> ... -> meta-objset -> ... -> uberblock. But verification obviously starts from uberblock and goes the opposite way to the data block, since you can not trust any information from a block that is not verified yet.

[Jinnan-Guo]

Thank you! That explanation was really helpful.

I have a follow-up question: When a single read operation fetches a data block from disk into memory, does ZFS verify the checksum from the uberblock all the way down to the data block each time, or is the full verification path spread over several operations (like pool import, mounting the dataset, and reading the file)?

Thanks again,
Jinnan