[RFC,net-next,v5,07/14] page_pool: devmem support

Convert netmem to be a union of struct page and struct netmem. Overload
the LSB of struct netmem* to indicate that it's a net_iov, otherwise
it's a page.

Currently these entries in struct page are rented by the page_pool and
used exclusively by the net stack:

struct {
	unsigned long pp_magic;
	struct page_pool *pp;
	unsigned long _pp_mapping_pad;
	unsigned long dma_addr;
	atomic_long_t pp_ref_count;
};

Mirror these (and only these) entries into struct net_iov and implement
netmem helpers that can access these common fields regardless of
whether the underlying type is page or net_iov.

Implement checks for net_iov in netmem helpers which delegate to mm
APIs, to ensure net_iov are never passed to the mm stack.

Signed-off-by: Mina Almasry <almasrymina@google.com>

---

RFCv5:
- Use netmem instead of page* with LSB set.
- Use pp_ref_count for refcounting net_iov.
- Removed many of the custom checks for netmem.

v1:
- Disable fragmentation support for iov properly.
- fix napi_pp_put_page() path (Yunsheng).
- Use pp_frag_count for devmem refcounting.

---
 include/net/netmem.h            | 145 ++++++++++++++++++++++++++++++--
 include/net/page_pool/helpers.h |  25 +++---
 net/core/page_pool.c            |  26 +++---
 net/core/skbuff.c               |   9 +-
 4 files changed, 164 insertions(+), 41 deletions(-)

On 12/18/23 02:40, Mina Almasry wrote:
> Convert netmem to be a union of struct page and struct netmem. Overload
> the LSB of struct netmem* to indicate that it's a net_iov, otherwise
> it's a page.
> 
> Currently these entries in struct page are rented by the page_pool and
> used exclusively by the net stack:
> 
> struct {
> 	unsigned long pp_magic;
> 	struct page_pool *pp;
> 	unsigned long _pp_mapping_pad;
> 	unsigned long dma_addr;
> 	atomic_long_t pp_ref_count;
> };
> 
> Mirror these (and only these) entries into struct net_iov and implement
> netmem helpers that can access these common fields regardless of
> whether the underlying type is page or net_iov.
> Implement checks for net_iov in netmem helpers which delegate to mm
> APIs, to ensure net_iov are never passed to the mm stack.
> 
> Signed-off-by: Mina Almasry <almasrymina@google.com>
> 
> ---
> 
> RFCv5:
> - Use netmem instead of page* with LSB set.
> - Use pp_ref_count for refcounting net_iov.
> - Removed many of the custom checks for netmem.
> 
> v1:
> - Disable fragmentation support for iov properly.
> - fix napi_pp_put_page() path (Yunsheng).
> - Use pp_frag_count for devmem refcounting.
> 
> ---
>   include/net/netmem.h            | 145 ++++++++++++++++++++++++++++++--
>   include/net/page_pool/helpers.h |  25 +++---
>   net/core/page_pool.c            |  26 +++---
>   net/core/skbuff.c               |   9 +-
>   4 files changed, 164 insertions(+), 41 deletions(-)
> 
> diff --git a/include/net/netmem.h b/include/net/netmem.h
> index 31f338f19da0..7557aecc0f78 100644
> --- a/include/net/netmem.h
> +++ b/include/net/netmem.h
> @@ -12,11 +12,47 @@
>   
>   /* net_iov */
>   
> +DECLARE_STATIC_KEY_FALSE(page_pool_mem_providers);
> +
> +/*  We overload the LSB of the struct page pointer to indicate whether it's
> + *  a page or net_iov.
> + */
> +#define NET_IOV 0x01UL
> +
>   struct net_iov {
> +	unsigned long __unused_padding;
> +	unsigned long pp_magic;
> +	struct page_pool *pp;
>   	struct dmabuf_genpool_chunk_owner *owner;
>   	unsigned long dma_addr;
> +	atomic_long_t pp_ref_count;
>   };

I wonder if it would be better to extract a common sub-struct
used in struct page, struct_group_tagged can help to avoid
touching old code:

struct page {
	unsigned long flags;
	union {
		...
		struct_group_tagged(<struct_name>, ...,
			/**
			 * @pp_magic: magic value to avoid recycling non
			 * page_pool allocated pages.
			 */
			unsigned long pp_magic;
			struct page_pool *pp;
			unsigned long _pp_mapping_pad;
			unsigned long dma_addr;
			atomic_long_t pp_ref_count;
		);
	};
}

struct net_iov {
	unsigned long pad;
	struct <struct_name> p;
};


A bit of a churn with the padding and nesting net_iov but looks
sturdier. No duplication, and you can just check positions of the
structure instead of per-field NET_IOV_ASSERT_OFFSET, which you
have to not forget to update e.g. when adding a new field. Also,
with the change __netmem_clear_lsb can return a pointer to that
structure, casting struct net_iov when it's a page is a bit iffy.

And the next question would be whether it'd be a good idea to encode
iov vs page not by setting a bit but via one of the fields in the
structure, maybe pp_magic.

With that said I'm a bit concerned about the net_iov size. If each
represents 4096 bytes and you're registering 10MB, then you need
30 pages worth of memory just for the iov array. Makes kvmalloc
a must even for relatively small sizes.

And the final bit, I don't believe the overlay is necessary in
this series. Optimisations are great, but this one is a bit more on
the controversial side. Unless I missed something and it does make
things easier, it might make sense to do it separately later.


> +/* These fields in struct page are used by the page_pool and net stack:
> + *
> + *	struct {
> + *		unsigned long pp_magic;
> + *		struct page_pool *pp;
> + *		unsigned long _pp_mapping_pad;
> + *		unsigned long dma_addr;
> + *		atomic_long_t pp_ref_count;
> + *	};
> + *
> + * We mirror the page_pool fields here so the page_pool can access these fields
> + * without worrying whether the underlying fields belong to a page or net_iov.
> + *
> + * The non-net stack fields of struct page are private to the mm stack and must
> + * never be mirrored to net_iov.
> + */
> +#define NET_IOV_ASSERT_OFFSET(pg, iov)             \
> +	static_assert(offsetof(struct page, pg) == \
> +		      offsetof(struct net_iov, iov))
> +NET_IOV_ASSERT_OFFSET(pp_magic, pp_magic);
> +NET_IOV_ASSERT_OFFSET(pp, pp);
> +NET_IOV_ASSERT_OFFSET(dma_addr, dma_addr);
> +NET_IOV_ASSERT_OFFSET(pp_ref_count, pp_ref_count);
> +#undef NET_IOV_ASSERT_OFFSET
> +
>   static inline struct dmabuf_genpool_chunk_owner *
>   net_iov_owner(const struct net_iov *niov)
>   {
> @@ -47,19 +83,25 @@ net_iov_binding(const struct net_iov *niov)
>   struct netmem {
>   	union {
>   		struct page page;
> -
> -		/* Stub to prevent compiler implicitly converting from page*
> -		 * to netmem_t* and vice versa.
> -		 *
> -		 * Other memory type(s) net stack would like to support
> -		 * can be added to this union.
> -		 */
> -		void *addr;
> +		struct net_iov niov;
>   	};
>   };
>   
...

On Tue, Feb 13, 2024 at 5:28 AM Pavel Begunkov <asml.silence@gmail.com> wrote:
>
> On 12/18/23 02:40, Mina Almasry wrote:
> > Convert netmem to be a union of struct page and struct netmem. Overload
> > the LSB of struct netmem* to indicate that it's a net_iov, otherwise
> > it's a page.
> >
> > Currently these entries in struct page are rented by the page_pool and
> > used exclusively by the net stack:
> >
> > struct {
> >       unsigned long pp_magic;
> >       struct page_pool *pp;
> >       unsigned long _pp_mapping_pad;
> >       unsigned long dma_addr;
> >       atomic_long_t pp_ref_count;
> > };
> >
> > Mirror these (and only these) entries into struct net_iov and implement
> > netmem helpers that can access these common fields regardless of
> > whether the underlying type is page or net_iov.
> > Implement checks for net_iov in netmem helpers which delegate to mm
> > APIs, to ensure net_iov are never passed to the mm stack.
> >
> > Signed-off-by: Mina Almasry <almasrymina@google.com>
> >
> > ---
> >
> > RFCv5:
> > - Use netmem instead of page* with LSB set.
> > - Use pp_ref_count for refcounting net_iov.
> > - Removed many of the custom checks for netmem.
> >
> > v1:
> > - Disable fragmentation support for iov properly.
> > - fix napi_pp_put_page() path (Yunsheng).
> > - Use pp_frag_count for devmem refcounting.
> >
> > ---
> >   include/net/netmem.h            | 145 ++++++++++++++++++++++++++++++--
> >   include/net/page_pool/helpers.h |  25 +++---
> >   net/core/page_pool.c            |  26 +++---
> >   net/core/skbuff.c               |   9 +-
> >   4 files changed, 164 insertions(+), 41 deletions(-)
> >
> > diff --git a/include/net/netmem.h b/include/net/netmem.h
> > index 31f338f19da0..7557aecc0f78 100644
> > --- a/include/net/netmem.h
> > +++ b/include/net/netmem.h
> > @@ -12,11 +12,47 @@
> >
> >   /* net_iov */
> >
> > +DECLARE_STATIC_KEY_FALSE(page_pool_mem_providers);
> > +
> > +/*  We overload the LSB of the struct page pointer to indicate whether it's
> > + *  a page or net_iov.
> > + */
> > +#define NET_IOV 0x01UL
> > +
> >   struct net_iov {
> > +     unsigned long __unused_padding;
> > +     unsigned long pp_magic;
> > +     struct page_pool *pp;
> >       struct dmabuf_genpool_chunk_owner *owner;
> >       unsigned long dma_addr;
> > +     atomic_long_t pp_ref_count;
> >   };
>
> I wonder if it would be better to extract a common sub-struct
> used in struct page, struct_group_tagged can help to avoid
> touching old code:
>
> struct page {
>         unsigned long flags;
>         union {
>                 ...
>                 struct_group_tagged(<struct_name>, ...,
>                         /**
>                          * @pp_magic: magic value to avoid recycling non
>                          * page_pool allocated pages.
>                          */
>                         unsigned long pp_magic;
>                         struct page_pool *pp;
>                         unsigned long _pp_mapping_pad;
>                         unsigned long dma_addr;
>                         atomic_long_t pp_ref_count;
>                 );
>         };
> }
>
> struct net_iov {
>         unsigned long pad;
>         struct <struct_name> p;
> };
>
>
> A bit of a churn with the padding and nesting net_iov but looks
> sturdier. No duplication, and you can just check positions of the
> structure instead of per-field NET_IOV_ASSERT_OFFSET, which you
> have to not forget to update e.g. when adding a new field. Also,

Yes, this is nicer. If possible I'll punt it to a minor cleanup as a
follow up change. Logistically I think if this series need-not touch
code outside of net/, that's better.

> with the change __netmem_clear_lsb can return a pointer to that
> structure, casting struct net_iov when it's a page is a bit iffy.
>
> And the next question would be whether it'd be a good idea to encode
> iov vs page not by setting a bit but via one of the fields in the
> structure, maybe pp_magic.
>

I will push back against this, for 2 reasons:

1. I think pp_magic's first 2 bits (and maybe more) are used by mm
code and thus I think extending usage of pp_magic in this series is a
bit iffy and I would like to avoid it. I just don't want to touch the
semantics of struct page if I don't have to.
2. I think this will be a measurable perf regression. Currently we can
tell if a pointer is a page or net_iov without dereferencing the
pointer and dirtying the cache-line. This will cause us to possibly
dereference the pointer in areas where we don't need to. I think I had
an earlier version of this code that required a dereference to tell if
a page was devmem and Eric pointed to me it was a perf regression.

I also don't see any upside of using pp_magic, other than making the
code slightly more readable, maybe.

> With that said I'm a bit concerned about the net_iov size. If each
> represents 4096 bytes and you're registering 10MB, then you need
> 30 pages worth of memory just for the iov array. Makes kvmalloc
> a must even for relatively small sizes.
>

This I think is an age-old challenge with pages. 1.6% of the machine's
memory is 'wasted' on every machine because a struct page needs to be
allocated for each PAGE_SIZE region. We're running into the same issue
here where if we want to refer to PAGE_SIZE regions of memory we need
to allocate some reference to it. Note that net_iov can be relatively
easily extended to support N order pages. Also note that in the devmem
TCP use case it's not really an issue; the minor increase in mem
utilization is more than offset by the saving in memory bw as compared
to using host memory as a bounce buffer. All in all I vote this is
something that can be tuned or improved in the future if someone finds
the extra memory usage a hurdle to using devmem TCP or this net_iov
infra.

> And the final bit, I don't believe the overlay is necessary in
> this series. Optimisations are great, but this one is a bit more on
> the controversial side. Unless I missed something and it does make
> things easier, it might make sense to do it separately later.
>

I completely agree, the overlay is not necessary. I implemented the
overlay in response to Yunsheng's  strong requests for more 'unified'
processing between page and devmem. This is the most unification I can
do IMO without violating the requirements from Jason. I'm prepared to
remove the overlay if it turns out controversial, but so far I haven't
seen any complaints. Jason, please do take a look if you have not
already.

>
> > +/* These fields in struct page are used by the page_pool and net stack:
> > + *
> > + *   struct {
> > + *           unsigned long pp_magic;
> > + *           struct page_pool *pp;
> > + *           unsigned long _pp_mapping_pad;
> > + *           unsigned long dma_addr;
> > + *           atomic_long_t pp_ref_count;
> > + *   };
> > + *
> > + * We mirror the page_pool fields here so the page_pool can access these fields
> > + * without worrying whether the underlying fields belong to a page or net_iov.
> > + *
> > + * The non-net stack fields of struct page are private to the mm stack and must
> > + * never be mirrored to net_iov.
> > + */
> > +#define NET_IOV_ASSERT_OFFSET(pg, iov)             \
> > +     static_assert(offsetof(struct page, pg) == \
> > +                   offsetof(struct net_iov, iov))
> > +NET_IOV_ASSERT_OFFSET(pp_magic, pp_magic);
> > +NET_IOV_ASSERT_OFFSET(pp, pp);
> > +NET_IOV_ASSERT_OFFSET(dma_addr, dma_addr);
> > +NET_IOV_ASSERT_OFFSET(pp_ref_count, pp_ref_count);
> > +#undef NET_IOV_ASSERT_OFFSET
> > +
> >   static inline struct dmabuf_genpool_chunk_owner *
> >   net_iov_owner(const struct net_iov *niov)
> >   {
> > @@ -47,19 +83,25 @@ net_iov_binding(const struct net_iov *niov)
> >   struct netmem {
> >       union {
> >               struct page page;
> > -
> > -             /* Stub to prevent compiler implicitly converting from page*
> > -              * to netmem_t* and vice versa.
> > -              *
> > -              * Other memory type(s) net stack would like to support
> > -              * can be added to this union.
> > -              */
> > -             void *addr;
> > +             struct net_iov niov;
> >       };
> >   };
> >
> ...
>
> --
> Pavel Begunkov



--
Thanks,
Mina

On 2/13/24 21:11, Mina Almasry wrote:
> On Tue, Feb 13, 2024 at 5:28 AM Pavel Begunkov <asml.silence@gmail.com> wrote:
>>
...
>>
>> A bit of a churn with the padding and nesting net_iov but looks
>> sturdier. No duplication, and you can just check positions of the
>> structure instead of per-field NET_IOV_ASSERT_OFFSET, which you
>> have to not forget to update e.g. when adding a new field. Also,
> 
> Yes, this is nicer. If possible I'll punt it to a minor cleanup as a
> follow up change. Logistically I think if this series need-not touch
> code outside of net/, that's better.

Outside of net it should only be a small change in struct page
layout, but otherwise with struct_group_tagged things like
page->pp_magic would still work. Anyway, I'm not insisting.


>> with the change __netmem_clear_lsb can return a pointer to that
>> structure, casting struct net_iov when it's a page is a bit iffy.
>>
>> And the next question would be whether it'd be a good idea to encode
>> iov vs page not by setting a bit but via one of the fields in the
>> structure, maybe pp_magic.
>>
> 
> I will push back against this, for 2 reasons:
> 
> 1. I think pp_magic's first 2 bits (and maybe more) are used by mm
> code and thus I think extending usage of pp_magic in this series is a
> bit iffy and I would like to avoid it. I just don't want to touch the
> semantics of struct page if I don't have to.
> 2. I think this will be a measurable perf regression. Currently we can
> tell if a pointer is a page or net_iov without dereferencing the
> pointer and dirtying the cache-line. This will cause us to possibly
> dereference the pointer in areas where we don't need to. I think I had
> an earlier version of this code that required a dereference to tell if
> a page was devmem and Eric pointed to me it was a perf regression.

fair enough

> I also don't see any upside of using pp_magic, other than making the
> code slightly more readable, maybe.
> 
>> With that said I'm a bit concerned about the net_iov size. If each
>> represents 4096 bytes and you're registering 10MB, then you need
>> 30 pages worth of memory just for the iov array. Makes kvmalloc
>> a must even for relatively small sizes.
>>
> 
> This I think is an age-old challenge with pages. 1.6% of the machine's
> memory is 'wasted' on every machine because a struct page needs to be
> allocated for each PAGE_SIZE region. We're running into the same issue
> here where if we want to refer to PAGE_SIZE regions of memory we need
> to allocate some reference to it. Note that net_iov can be relatively
> easily extended to support N order pages. Also note that in the devmem
> TCP use case it's not really an issue; the minor increase in mem
> utilization is more than offset by the saving in memory bw as compared
> to using host memory as a bounce buffer.

It's not about memory consumption per se but rather the need
to vmalloc everything because of size.

> All in all I vote this is
> something that can be tuned or improved in the future if someone finds
> the extra memory usage a hurdle to using devmem TCP or this net_iov
> infra.

That's exactly what I was saying about overlaying it with
struct page, where the increase in size came from, but I agree
it's not critical

>> And the final bit, I don't believe the overlay is necessary in
>> this series. Optimisations are great, but this one is a bit more on
>> the controversial side. Unless I missed something and it does make
>> things easier, it might make sense to do it separately later.
>>
> 
> I completely agree, the overlay is not necessary. I implemented the
> overlay in response to Yunsheng's  strong requests for more 'unified'
> processing between page and devmem. This is the most unification I can
> do IMO without violating the requirements from Jason. I'm prepared to
> remove the overlay if it turns out controversial, but so far I haven't
> seen any complaints. Jason, please do take a look if you have not
> already.

Just to be clear, I have no objections to the change but noting
that IMHO it can be removed for now if it'd be dragging down
the set.