On Sun, Oct 1, 2017 at 2:22 PM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote: > On Sun, Oct 1, 2017 at 2:11 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote: >> On Sun, Oct 01, 2017 at 10:58:06AM -0700, Dan Williams wrote: >>> On Sun, Oct 1, 2017 at 12:57 AM, Christoph Hellwig <hch@xxxxxx> wrote: >>> > While this looks like a really nice cleanup of the code and removes >>> > nasty race conditions I'd like to understand the tradeoffs. >>> > >>> > This now requires every dax device that is used with a file system >>> > to have a struct page backing, which means not only means we'd >>> > break existing setups, but also a sharp turn from previous policy. >>> > >>> > Unless I misremember it was you Intel guys that heavily pushed for >>> > the page-less version, so I'd like to understand why you've changed >>> > your mind. >>> >>> Sure, here's a quick recap of the story so far of how we got here: >>> >>> * In support of page-less I/O operations envisioned by Matthew I >>> introduced pfn_t as a proposal for converting the block layer and >>> other sub-systems to use pfns instead of pages [1]. You helped out on >>> that patch set with some work on the DMA api. [2] >>> >>> * The DMA api conversion effort came to a halt when it came time to >>> touch sparc paths and DaveM said [3]: "Generally speaking, I think >>> that all actual physical memory the kernel operates on should have a >>> struct page backing it." >>> >>> * ZONE_DEVICE was created to solve the DMA problem and in developing / >>> testing that discovered plenty of proof for Dave's assertion (no fork, >>> no ptrace, etc). We should have made the switch to require struct page >>> at that point, but I was persuaded by the argument that changing the >>> dax policy may break existing assumptions, and that there were larger >>> issues to go solve at the time. >>> >>> What changed recently was the discussions around what the dax mount >>> option means and the assertion that we can, in general, make some >>> policy changes on our way to removing the "experimental" designation >>> from filesystem-dax. It is clear that the page-less dax path remains >>> experimental with all the way it fails in several kernel paths, and >>> there has been no patches for several months to revive the effort. >>> Meanwhile the page-less path continues to generate maintenance >>> overhead. The recent gymnastics (new ->post_mmap file_operation) to >>> make sure ->vm_flags are safely manipulated when dynamically changing >>> the dax mode of a file was the final straw for me to pull the trigger >>> on this series. >>> >>> In terms of what breaks by changing this policy it should be noted >>> that we automatically create pages for "legacy" pmem devices, and the >>> default for "ndctl create-namespace" is to allocate pages. I have yet >>> to see a bug report where someone was surprised by fork failing or >>> direct-I/O causing a SIGBUS. So, I think the defaults are working, it >>> is unlikely that there are environments dependent on page-less >>> behavior. >> >> Does this imply that the hardware vendors won't have >> tens of terabytes of pmem in systems in the near to medium term? >> That's what we were originally told to expect by 2018-19 timeframe >> (i.e. 5 years in), and that's kinda what we've been working towards. >> Indeed, supporting systems with a couple of orders of magnitude more >> pmem than ram was the big driver for page-less DAX mappings in the >> first place. i.e. it was needed to avoid the static RAM overhead of >> all the static struct pages for such large amounts of physical >> memory. >> >> If we decide that we must have struct pages for pmem, then we're >> essentially throwing away the ability to support the very systems >> the hardware vendors were telling us we needed to design the pmem >> infrastructure for. If that reality has changed, then I'd suggest >> that we need to determine what the long term replacement for >> pageless IO on large pmem systems will be before we throw what we >> have away. > > No, we can support large pmem with struct page capacity reserved from > pmem itself rather than ram. A 1.5% capacity tax does not appear to be > prohibitive. Dynamic allocation of struct page is another step we can take before deciding we need page-less I/O.
