iz (i@z.x) on 1/20/11 wrote:
---------------------------
>No, what GC does is moving used pages out of mostly free blocks so the block can
>be erased and reused. What you describe is simply erasing free but dirty blocks.
>It's perhaps part of doing GC, but a very small and simple part.
>

There is no erase in marking a PBA free/dirty. It is merely record keeping. The GC routines have to come in later and move any allocated PBAs within an erase block and perform the erase. But that is what the GC routine would have had to do even if the device didn't support trim. trim doesn't erase anything directly, it is just an LBA to PBA mapping and PBA state update.

I think we are also talking a bit crosswise with each other.

So just to clarify...
PBA = physical block address, aka the smallest write to the physical flash and is generally 512B
LBA = logical block address, aka what the OS/FS sees. Also generally 512B.
LBA_TO_PBA Table = mechanism/scheme by which LBAs to mapped to PBAs.
PBA State = current state of a data block reference by a PBA. PBA states are generally: Allocated, free/dirty, free/clean.

Allocated = there is a valid LBA to PBA mapping for the PBA data
free/dirty = there is no valid LBA to PBA mapping for the PBA data but data has already been written and cannot be written again
free/clean = there is no valid LBA to PBA mapping for the PBA data and data can be written

There is also potential to divide the Allocated state into two substates: Allocated/Clean and Allocated/Dirty. It is an optimization that can allow for write de-amplification. This can be used for instance in the case where the data is all zeros. Generally would be used with a higher granularity LBA<>PBA mapping structure.



>Avoiding unnecessary page copying is the sole reason for TRIM to exist. Pre-erasing
>blocks is a nice side-effect, but not that important. So I wouldn't call TRIM and GC independent at all.
>

pre-erasing is exactly what you do not want trim to do. You merely want it to change the state of the PBA.

I guess it would be more correct to say that trim is dependent on GC but GC is not dependent on trim.