Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On Mon, Aug 10 2015 at 11:38pm -0400, Kent Overstreet <kent.overstreet at gmail.com> wrote: > On Mon, Aug 10, 2015 at 10:41:55PM -0400, Mike Snitzer wrote: > > On Mon, Aug 10 2015 at 10:00pm -0400, > > Martin K. Petersen <martin.petersen at oracle.com> wrote: > > > > > >>>>> "Ming" == Ming Lin <mlin at kernel.org> writes: > > > > > > Ming> Did you mean still use (UINT_MAX >> 9) in blkdev_issue_discard()? > > > > > > Ming> But that doesn't work for dm-thinp. See Kent's suggestion to use > > > Ming> 1<<31. > > > > > > I'm not sure why things are not working for dm-thinp. Presumably Kent's > > > code would split the discard at a granularity boundary so why would that > > > cause problems for dm? > > > > DM-thinp processes discards internally before it passes them down (if > > configured to do so). If a discard is smaller than the granularity of a > > thinp block (whose size is configurable) or if the start and end of the > > discard's extent is misaligned (relative to the thinp blocks mapped to > > the logical extent) then the discard won't actually discard partial > > thinp blocks. > > This kind of logic really doesn't belong in dm - if it's needed, it really > belongs in bio_split() (which is supposed to work correctly for discards - so if > it is needed, then bio_split() needs fixing...) DM thinp does advertise discard_granularity that reflects the thinp blocksize. blk_queue_split() does look like it'd do the right thing. But the splitting that DM thinp is doing is a long standing implementation (in DM core) that will need to be carefully reviewed/rewritten. We can tackle it after all this late splitting code lands. > IMO though it belongs in the driver - if a discard needs to be dropped because > it's too small and the hardware can't do it, that should be the driver's > responsibility. This isn't about the hardware's limits. This is about the intermediate remapping/stacking driver's own space management hooking off of the discard bio too.