On Fri, Feb 17, 2017 at 09:25:27AM +0800, Ming Lei wrote: > Hi Shaohua, > > On Fri, Feb 17, 2017 at 6:16 AM, Shaohua Li <shli@xxxxxxxxxx> wrote: > > On Thu, Feb 16, 2017 at 07:45:30PM +0800, Ming Lei wrote: > >> In MD's resync I/O path, there are lots of direct access to bio's > >> bvec table. This patchset kills most of them, and the conversion > >> is quite straightforward. > > > > I don't like this approach. The MD uses a hacky way to manage pages allocated, > > this is the root of the problem. The patches add another hack way to do the > > Yes, I agree, and bio_iov_iter_get_pages() uses this kind of hacky way too > actually. > > > management. I'd like to see explict management of the pages, for example, add > > data structure in r1bio to manage the pages, then we can use existing API for > > all the stuffes we need. > > Yeah, that is definitely clean, but we have to pay the following cost: > > - allocate at least N * (128 + 4) bytes per each r1_bio/r10_bio > - N is pool_info.raid_disks for raid1, and conf->copies for raid10 > > If we are happy to introduce the cost, I can take this way in V1. It's not a big deal. The inflight bio shouldn't be big, so the r1_bio count isn't big. We don't waste much. Thanks, Shaohua
