On Thu, Dec 03 2020 at 11:27am -0500, Keith Busch <kbusch@xxxxxxxxxx> wrote: > On Thu, Dec 03, 2020 at 09:33:59AM -0500, Mike Snitzer wrote: > > On Wed, Dec 02 2020 at 10:26pm -0500, > > Ming Lei <ming.lei@xxxxxxxxxx> wrote: > > > > > I understand it isn't related with correctness, because the underlying > > > queue can split by its own chunk_sectors limit further. So is the issue > > > too many further-splitting on queue with chunk_sectors 8? then CPU > > > utilization is increased? Or other issue? > > > > No, this is all about correctness. > > > > Seems you're confining the definition of the possible stacking so that > > the top-level device isn't allowed to have its own hard requirements on > > IO sizes it sends to its internal implementation. Just because the > > underlying device can split further doesn't mean that the top-level > > virtual driver can service larger IO sizes (not if the chunk_sectors > > stacking throws away the hint the virtual driver provided because it > > used lcm_not_zero). > > I may be missing something obvious here, but if the lower layers split > to their desired boundary already, why does this limit need to stack? The problematic scenario is when the topmost layer, or layers, are the more constrained. _That_ is why the top-level's chunk_sectors limit cannot be relaxed. For example (in extreme where chunk_sectors is stacked via gcd): dm VDO target (chunk_sectors=4K) on dm-thin (ideally chunk_sectors=1280K, reality chunk_sectors=128K) on 10+2 RAID6 (chunk_sectors=128K, io_opt=1280K) on raid members (chunk_sectors=0) Results in the following bottom up blk_stack_limits() stacking: gcd(128K, 0) = 128K -> but MD just sets chunk_sectors, no stacking is done afaik gcd(1280K, 128K) = 128K -> this one hurts dm-thin, needless splitting gcd(4K, 128K) = 4K -> vdo _must_ receive 4K IOs, hurts but "this is the way" ;) So this is one extreme that shows stacking chunk_sectors is _not_ helpful (if the resulting chunk_sectors were actually used as basis for splitting). Better for each layer to just impose its own chunk_sectors without concern for the layers below. Think I'd be fine with block core removing the chunk_sectors stacking from blk_stack_limits()... (and as you see below, I've been forced to revert to _not_ using stacked chunk_sectors based splitting in DM) > Won't it also work if each layer sets their desired chunk_sectors > without considering their lower layers? The commit that initially > stacked chunk_sectors doesn't provide any explanation. Yes, I think it would work. The current stacking doesn't have the luxury of knowing which layer a blk_stack_limits() maps too. BUT within a layer chunk_sectors really does need to be compatible/symbiotic. So it is unfortunately all or nothing as you build up the stack. And that all-or-nothing stacking of chunk_sectors is why I've now (just last night, based on further review by jdorminy) had to punt on using stacked chunk_sectors and revert DM back to doing its own fine-grained (and varied) splitting on a per DM target basis, see: https://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm.git/commit/?h=dm-5.10-rcX&id=6bb38bcc33bf3093c08bd1b71e4f20c82bb60dd1 Kind of depressing that I went so far down the rabbit hole, of wanting to lean on block core, that I lost sight of an important "tenet of DM": + * Does the target need to split IO even further? + * - varied (per target) IO splitting is a tenet of DM; this + * explains why stacked chunk_sectors based splitting via + * blk_max_size_offset() isn't possible here. And it is because of this that DM is forced to lean on human creation of an optimal IO stack.. which is prone to human error when a particular thinp "blocksize" is selected, etc. Mike