Re: [GIT PULL] PMEM driver for v4.1

[Ingo Molnar]

* Christoph Hellwig <hch@xxxxxx> wrote:

> On Mon, Apr 13, 2015 at 12:45:32PM +0200, Ingo Molnar wrote:
> > Btw., what's the future design plan here? Enable struct page backing, 
> > or provide special codepaths for all DAX uses like the special pte 
> > based approach for mmap()s?
> 
> There are a couple approaches proposed, but we don't have consensus which
> way to go yet (to put it mildly).
> 
>  - the old Intel patches just allocate pages for E820_PMEM regions.
>    I think this is a good way forward for the "old-school" small
>    pmem regions which usually are battery/capacitor + flash backed
>    DRAM anyway.  This could easily be resurrected for the current code,
>    but it couldn't be used for PCI backed pmem regions, and would work
>    although waste a lot of resources for the gigantic pmem devices some
>    Intel people talk about (400TB+ IIRC).

So here's how I calculate the various scenarios:

There are two main usecases visible currently for pmem devices: 'pmem 
as storage' and 'pmem as memory', and they have rather distinct 
characteristics.

1) pmem devices as 'storage':

So the size of 'struct page' is 64 bytes currently.

So even if a pmem device is just 1 TB (small storage), for example to 
replace storage on a system, we'd have to allocate 64 bytes per 4096 
bytes of storage, which would use up 16 GB of main memory just for the 
page struct arrays...

So in this case 'struct page' allocation is not acceptable even for 
relatively small pmem device sizes.

For this usecase I think the current pmem driver is perfectly 
acceptable and in fact ideal:

  - it double buffers, giving higher performance and also protecting
    storage from wear (that is likely flash based)

  - the double buffering makes most struct page based APIs work just 
    fine.

  - it offers the DAX APIs for those weird special cases that really 
    want to do their own cache and memory management (databases and 
    other crazy usecases)

2) pmem devices as 'memory':

Battery backed and similar solutions of nv-dram, these are probably a 
lot smaller (for cost reasons) and are also a lot more RAM-alike, so 
the 'struct page' allocation in main RAM makes sense and possibly 
people would want to avoid the double buffering as well.

Furthermore, in this case we could also do another trick:

>  - Intel has proposed changes that allow block I/O on regions that aren't
>    page backed, by supporting PFN-based scatterlists which would have to be
>    supported all over the I/O path. Reception of that code has been rather
>    mediocre in general, although I wouldn't rule it out.
>
>  - Boaz has shown code that creates pages dynamically for pmem regions.
>    Unlike the old Intel e820 code that would also work for PCI backed
>    pmem regions.  Boaz says he has such a card, but until someone actually
>    publishes specs and/or the trivial pci_driver for them I'm inclined to
>    just ignore that option.
> 
>  - There have been proposals for temporary struct page mappings, or
>    variable sized pages, but as far as I can tell no code to actually
>    implement these schemes.

None of this gives me warm fuzzy feelings...

... has anyone explored the possibility of putting 'struct page' into 
the pmem device itself, essentially using it as metadata?

Since it's directly mapped it should just work for most things if it's 
at least write-through cached (UC would be a horror), and it would 
also solve all the size problems. With write-through caching it should 
also be pretty OK performance-wise. The 64 bytes size is ideal as 
well.

( This would create a bit of a dependency with the kernel version, and 
  would complicate things of how to acquire a fresh page array after 
  bootup, but that could be solved relatively easily IMHO. )

This would eliminate all the negative effects dynamic allocation of 
page structs or sg-lists brings.

Anyway:

Since both the 'pmem as storage' and 'pmem as memory' usecases will 
likely be utilized in practice, IMHO we should hedge our bets by 
supporting both equally well: we should start with the simpler one 
(the current driver) and then support both in the end, with as much 
end user flexibility as possible.

Thanks,

	Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/