On 23/03/2017 23:53, Lars Ellenberg wrote: > Thin does not claim to zero data on discard. which is ok, and correct, > because it only punches holes on full chunks (or whatever you call > them), and leaves the rest in the mapping tree as is. > > And that behaviour would prevent DRBD from exposing discards if > configured on top of thin. (see above) > > But thin *could* easily guarantee zeroing, by simply punching holes > where it can, and zeroing out the not fully-aligned partial start and > end of the range. That's the difference between REQ_OP_DISCARD (only punches holes on full chunks) and REQ_OP_WRITE_ZEROES with the REQ_UNMAP flag (punches holes + zeroes incomplete chunks). dm-thinp's REQ_OP_DISCARD should not do anything for unaligned parts. Instead, layers above should use REQ_OP_WRITE_ZEROES (with or without REQ_UNMAP, as required) if they need zeroes. dm-thinp would have to split off the partial chunks, and zero them in the lower-level device with REQ_OP_WRITE_ZEROES. Paolo
