On Fri, Mar 05, 2021 at 11:41:33AM +0900, Keith Busch wrote: > I'll need to think on this to consider if the memory cost is worth it > (8b to 64b), but you could replace nvme_request's 'struct nvme_command' > pointer with the struct itself and not have to allocate anything per IO. > An added bonus is that sync and async handling become more the same. This would solve a lot of mess with the async passthrough, and also more closely match what is done in SCSI. In theory we could use a cut down version without the data and metadata pointers (just 40 bytes), but I'm not sure that is really worth it given that we then need to rearrange things in the I/O path.
