Date:   Tue, 11 Jan 2022 00:17:18 -0800
From:   John Hubbard <jhubbard@...dia.com>
To:     Matthew Wilcox <willy@...radead.org>, linux-kernel@...r.kernel.org
Cc:     Christoph Hellwig <hch@....de>, Jason Gunthorpe <jgg@...dia.com>,
        Joao Martins <joao.m.martins@...cle.com>,
        Logan Gunthorpe <logang@...tatee.com>,
        Ming Lei <ming.lei@...hat.com>, linux-block@...r.kernel.org,
        netdev@...r.kernel.org, linux-mm@...ck.org,
        linux-rdma@...r.kernel.org, dri-devel@...ts.freedesktop.org,
        nvdimm@...ts.linux.dev
Subject: Re: Phyr Starter

On 1/10/22 11:34, Matthew Wilcox wrote:
> TLDR: I want to introduce a new data type:
> 
> struct phyr {
>          phys_addr_t addr;
>          size_t len;
> };
> 
> and use it to replace bio_vec as well as using it to replace the array
> of struct pages used by get_user_pages() and friends.
> 
> ---

This would certainly solve quite a few problems at once. Very compelling.

Zooming in on the pinning aspect for a moment: last time I attempted to
convert O_DIRECT callers from gup to pup, I recall wanting very much to
record, in each bio_vec, whether these pages were acquired via FOLL_PIN,
or some non-FOLL_PIN method. Because at the end of the IO, it is not
easy to disentangle which pages require put_page() and which require
unpin_user_page*().

And changing the bio_vec for *that* purpose was not really acceptable.

But now that you're looking to change it in a big way (and with some
spare bits avaiable...oohh!), maybe I can go that direction after all.

Or, are you looking at a design in which any phyr is implicitly FOLL_PIN'd
if it exists at all?

Or any other thoughts in this area are very welcome.

> 
> There are two distinct problems I want to address: doing I/O to memory
> which does not have a struct page and efficiently doing I/O to large
> blobs of physically contiguous memory, regardless of whether it has a
> struct page.  There are some other improvements which I regard as minor.
> 
> There are many types of memory that one might want to do I/O to that do
> not have a struct page, some examples:
>   - Memory on a graphics card (or other PCI card, but gfx seems to be
>     the primary provider of DRAM on the PCI bus today)
>   - DAX, or other pmem (there are some fake pages today, but this is
>     mostly a workaround for the IO problem today)
>   - Guest memory being accessed from the hypervisor (KVM needs to
>     create structpages to make this happen.  Xen doesn't ...)
> All of these kinds of memories can be addressed by the CPU and so also
> by a bus master.  That is, there is a physical address that the CPU
> can use which will address this memory, and there is a way to convert
> that to a DMA address which can be programmed into another device.
> There's no intent here to support memory which can be accessed by a
> complex scheme like writing an address to a control register and then
> accessing the memory through a FIFO; this is for memory which can be
> accessed by DMA and CPU loads and stores.
> 
> For get_user_pages() and friends, we currently fill an array of struct
> pages, each one representing PAGE_SIZE bytes.  For an application that
> is using 1GB hugepages, writing 2^18 entries is a significant overhead.
> It also makes drivers hard to write as they have to recoalesce the
> struct pages, even though the VM can tell it whether those 2^18 pages
> are contiguous.
> 
> On the minor side, struct phyr can represent any mappable chunk of memory.
> A bio_vec is limited to 2^32 bytes, while on 64-bit machines a phyr
> can represent larger than 4GB.  A phyr is the same size as a bio_vec
> on 64 bit (16 bytes), and the same size for 32-bit with PAE (12 bytes).
> It is smaller for 32-bit machines without PAE (8 bytes instead of 12).
> 
> Finally, it may be possible to stop using scatterlist to describe the
> input to the DMA-mapping operation.  We may be able to get struct
> scatterlist down to just dma_address and dma_length, with chaining
> handled through an enclosing struct.
> 
> I would like to see phyr replace bio_vec everywhere it's currently used.
> I don't have time to do that work now because I'm busy with folios.
> If someone else wants to take that on, I shall cheer from the sidelines.

I'm starting to wonder if I should jump in here, in order to get this
as a way to make the O_DIRECT conversion much cleaner. But let's see.

> What I do intend to do is:
> 
>   - Add an interface to gup.c to pin/unpin N phyrs
>   - Add a sg_map_phyrs()
>     This will take an array of phyrs and allocate an sg for them
>   - Whatever else I need to do to make one RDMA driver happy with
>     this scheme
> 
> At that point, I intend to stop and let others more familiar with this
> area of the kernel continue the conversion of drivers.
> 
> P.S. If you've had the Prodigy song running through your head the whole
> time you've been reading this email ... I'm sorry / You're welcome.
> If people insist, we can rename this to phys_range or something boring,
> but I quite like the spelling of phyr with the pronunciation of "fire".

A more conservative or traditional name might look like:

     phys_vec (maintains some resemblance to what it's replacing)
     phys_range
     phys_addr_range

phyr is rather cool, but it also is awfully too close to "phys" for
reading comfort. And there is a lot to be said for self-descriptive names,
which phyr is not.

And of course, you're signing for another huge naming debate with Linus
if you go with the "cool" name here. :)


thanks,
-- 
John Hubbard
NVIDIA

Powered by blists - more mailing lists