>>>>> "Jens" == Jens Axboe <jens.axboe@xxxxxxxxxx> writes: Jens> The forward progress reference failed to take stacked drivers into Jens> account, I didn't realize that they need to allocate integrity Jens> data again as well. Yup. We need a bio integrity struct as well as a vector to describe the integrity pages. Ideally I'd like to avoid cloning the integrity bio_vec altogether. The only reason I do it now is because I have to keep the integrity vector in sync with the data vector when that gets sliced and diced. Plus there's the suck of partial completion. If we never changed bio_vecs this wouldn't be an issue. One option would be to add an offset parameter to the bio. That way we could completely avoid cloning bio_vecs. That would mean a bit more complexity in building scatterlists at the bottom of the pile but we'd do fewer memory allocations. I've been tinkering with that approach this evening. I almost have it working for the integrity vector. Doing it for bios is obviously a much bigger task. Jens> Perhaps it would be cleaner to make the integrity allocation more Jens> explicit in the supported paths, instead of hiding it in bio_set? Jens> Dunno, haven't thought much about it, just an alternative approach DM is the only subsystem that manually clones things. MD uses bio_clone and has no idea that the integrity fluff is there. In any case I really think the bio_set approach is a nicer interface than making every stacking driver special-case integrity allocations. -- Martin K. Petersen Oracle Linux Engineering -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
