diff mbox series

[rfc,v6,2/4] page_pool: add interface to manipulate frag count in page pool

Message ID 1626752145-27266-3-git-send-email-linyunsheng@huawei.com
State Superseded
Headers show
Series add frag page support in page pool | expand

Commit Message

Yunsheng Lin July 20, 2021, 3:35 a.m. UTC
For 32 bit systems with 64 bit dma, dma_addr[1] is used to
store the upper 32 bit dma addr, those system should be rare
those days.

For normal system, the dma_addr[1] in 'struct page' is not
used, so we can reuse dma_addr[1] for storing frag count,
which means how many frags this page might be splited to.

In order to simplify the page frag support in the page pool,
the PAGE_POOL_DMA_USE_PP_FRAG_COUNT macro is added to indicate
the 32 bit systems with 64 bit dma, and the page frag support
in page pool is disabled for such system.

The newly added page_pool_set_frag_count() is called to reserve
the maximum frag count before any page frag is passed to the
user. The page_pool_atomic_sub_frag_count_return() is called
when user is done with the page frag.

Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
---
 include/linux/mm_types.h | 18 +++++++++++++-----
 include/net/page_pool.h  | 41 ++++++++++++++++++++++++++++++++++-------
 net/core/page_pool.c     |  4 ++++
 3 files changed, 51 insertions(+), 12 deletions(-)

Comments

Yunsheng Lin July 21, 2021, 8:15 a.m. UTC | #1
On 2021/7/20 23:43, Alexander Duyck wrote:
> On Mon, Jul 19, 2021 at 8:36 PM Yunsheng Lin <linyunsheng@huawei.com> wrote:

>>

>> For 32 bit systems with 64 bit dma, dma_addr[1] is used to

>> store the upper 32 bit dma addr, those system should be rare

>> those days.

>>

>> For normal system, the dma_addr[1] in 'struct page' is not

>> used, so we can reuse dma_addr[1] for storing frag count,

>> which means how many frags this page might be splited to.

>>

>> In order to simplify the page frag support in the page pool,

>> the PAGE_POOL_DMA_USE_PP_FRAG_COUNT macro is added to indicate

>> the 32 bit systems with 64 bit dma, and the page frag support

>> in page pool is disabled for such system.

>>

>> The newly added page_pool_set_frag_count() is called to reserve

>> the maximum frag count before any page frag is passed to the

>> user. The page_pool_atomic_sub_frag_count_return() is called

>> when user is done with the page frag.

>>

>> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>

>> ---

>>  include/linux/mm_types.h | 18 +++++++++++++-----

>>  include/net/page_pool.h  | 41 ++++++++++++++++++++++++++++++++++-------

>>  net/core/page_pool.c     |  4 ++++

>>  3 files changed, 51 insertions(+), 12 deletions(-)

>>

> 

> <snip>

> 

>> +static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

>> +                                                         long nr)

>> +{

>> +       long frag_count = atomic_long_read(&page->pp_frag_count);

>> +       long ret;

>> +

>> +       if (frag_count == nr)

>> +               return 0;

>> +

>> +       ret = atomic_long_sub_return(nr, &page->pp_frag_count);

>> +       WARN_ON(ret < 0);

>> +       return ret;

>>  }

>>

> 

> So this should just be an atomic_long_sub_return call. You should get

> rid of the atomic_long_read portion of this as it can cover up

> reference count errors.


atomic_long_sub_return() is used to avoid one possible cache bouncing and
barrrier caused by the last user.

You are right that that may cover up the reference count errors. How about
something like below:

static inline long page_pool_atomic_sub_frag_count_return(struct page *page,
							  long nr)
{
#ifdef CONFIG_DEBUG_PAGE_REF
	long ret = atomic_long_sub_return(nr, &page->pp_frag_count);

	WARN_ON(ret < 0);

	return ret;
#else
	if (atomic_long_read(&page->pp_frag_count) == nr)
		return 0;

	return atomic_long_sub_return(nr, &page->pp_frag_count);
#end
}

Or any better suggestion?


> .

>
Alexander Duyck July 21, 2021, 2:06 p.m. UTC | #2
On Wed, Jul 21, 2021 at 1:15 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:
>

> On 2021/7/20 23:43, Alexander Duyck wrote:

> > On Mon, Jul 19, 2021 at 8:36 PM Yunsheng Lin <linyunsheng@huawei.com> wrote:

> >>

> >> For 32 bit systems with 64 bit dma, dma_addr[1] is used to

> >> store the upper 32 bit dma addr, those system should be rare

> >> those days.

> >>

> >> For normal system, the dma_addr[1] in 'struct page' is not

> >> used, so we can reuse dma_addr[1] for storing frag count,

> >> which means how many frags this page might be splited to.

> >>

> >> In order to simplify the page frag support in the page pool,

> >> the PAGE_POOL_DMA_USE_PP_FRAG_COUNT macro is added to indicate

> >> the 32 bit systems with 64 bit dma, and the page frag support

> >> in page pool is disabled for such system.

> >>

> >> The newly added page_pool_set_frag_count() is called to reserve

> >> the maximum frag count before any page frag is passed to the

> >> user. The page_pool_atomic_sub_frag_count_return() is called

> >> when user is done with the page frag.

> >>

> >> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>

> >> ---

> >>  include/linux/mm_types.h | 18 +++++++++++++-----

> >>  include/net/page_pool.h  | 41 ++++++++++++++++++++++++++++++++++-------

> >>  net/core/page_pool.c     |  4 ++++

> >>  3 files changed, 51 insertions(+), 12 deletions(-)

> >>

> >

> > <snip>

> >

> >> +static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

> >> +                                                         long nr)

> >> +{

> >> +       long frag_count = atomic_long_read(&page->pp_frag_count);

> >> +       long ret;

> >> +

> >> +       if (frag_count == nr)

> >> +               return 0;

> >> +

> >> +       ret = atomic_long_sub_return(nr, &page->pp_frag_count);

> >> +       WARN_ON(ret < 0);

> >> +       return ret;

> >>  }

> >>

> >

> > So this should just be an atomic_long_sub_return call. You should get

> > rid of the atomic_long_read portion of this as it can cover up

> > reference count errors.

>

> atomic_long_sub_return() is used to avoid one possible cache bouncing and

> barrrier caused by the last user.


I assume you mean "atomic_long_read()" here.

> You are right that that may cover up the reference count errors. How about

> something like below:

>

> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

>                                                           long nr)

> {

> #ifdef CONFIG_DEBUG_PAGE_REF

>         long ret = atomic_long_sub_return(nr, &page->pp_frag_count);

>

>         WARN_ON(ret < 0);

>

>         return ret;

> #else

>         if (atomic_long_read(&page->pp_frag_count) == nr)

>                 return 0;

>

>         return atomic_long_sub_return(nr, &page->pp_frag_count);

> #end

> }

>

> Or any better suggestion?


So the one thing I might change would be to make it so that you only
do the atomic_long_read if nr is a constant via __builtin_constant_p.
That way you would be performing the comparison in
__page_pool_put_page and in the cases of freeing or draining the
page_frags you would be using the atomic_long_sub_return which should
be paths where you would not expect it to match or that are slowpath
anyway.

Also I would keep the WARN_ON in both paths just to be on the safe side.
Yunsheng Lin July 22, 2021, 8:07 a.m. UTC | #3
On 2021/7/21 22:06, Alexander Duyck wrote:
> On Wed, Jul 21, 2021 at 1:15 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:

>>

>> On 2021/7/20 23:43, Alexander Duyck wrote:

>>> On Mon, Jul 19, 2021 at 8:36 PM Yunsheng Lin <linyunsheng@huawei.com> wrote:

>>>>

>>>> For 32 bit systems with 64 bit dma, dma_addr[1] is used to

>>>> store the upper 32 bit dma addr, those system should be rare

>>>> those days.

>>>>

>>>> For normal system, the dma_addr[1] in 'struct page' is not

>>>> used, so we can reuse dma_addr[1] for storing frag count,

>>>> which means how many frags this page might be splited to.

>>>>

>>>> In order to simplify the page frag support in the page pool,

>>>> the PAGE_POOL_DMA_USE_PP_FRAG_COUNT macro is added to indicate

>>>> the 32 bit systems with 64 bit dma, and the page frag support

>>>> in page pool is disabled for such system.

>>>>

>>>> The newly added page_pool_set_frag_count() is called to reserve

>>>> the maximum frag count before any page frag is passed to the

>>>> user. The page_pool_atomic_sub_frag_count_return() is called

>>>> when user is done with the page frag.

>>>>

>>>> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>

>>>> ---

>>>>  include/linux/mm_types.h | 18 +++++++++++++-----

>>>>  include/net/page_pool.h  | 41 ++++++++++++++++++++++++++++++++++-------

>>>>  net/core/page_pool.c     |  4 ++++

>>>>  3 files changed, 51 insertions(+), 12 deletions(-)

>>>>

>>>

>>> <snip>

>>>

>>>> +static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

>>>> +                                                         long nr)

>>>> +{

>>>> +       long frag_count = atomic_long_read(&page->pp_frag_count);

>>>> +       long ret;

>>>> +

>>>> +       if (frag_count == nr)

>>>> +               return 0;

>>>> +

>>>> +       ret = atomic_long_sub_return(nr, &page->pp_frag_count);

>>>> +       WARN_ON(ret < 0);

>>>> +       return ret;

>>>>  }

>>>>

>>>

>>> So this should just be an atomic_long_sub_return call. You should get

>>> rid of the atomic_long_read portion of this as it can cover up

>>> reference count errors.

>>

>> atomic_long_sub_return() is used to avoid one possible cache bouncing and

>> barrrier caused by the last user.

> 

> I assume you mean "atomic_long_read()" here.


Yes, sorry for the confusion.

> 

>> You are right that that may cover up the reference count errors. How about

>> something like below:

>>

>> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

>>                                                           long nr)

>> {

>> #ifdef CONFIG_DEBUG_PAGE_REF

>>         long ret = atomic_long_sub_return(nr, &page->pp_frag_count);

>>

>>         WARN_ON(ret < 0);

>>

>>         return ret;

>> #else

>>         if (atomic_long_read(&page->pp_frag_count) == nr)

>>                 return 0;

>>

>>         return atomic_long_sub_return(nr, &page->pp_frag_count);

>> #end

>> }

>>

>> Or any better suggestion?

> 

> So the one thing I might change would be to make it so that you only

> do the atomic_long_read if nr is a constant via __builtin_constant_p.

> That way you would be performing the comparison in

> __page_pool_put_page and in the cases of freeing or draining the

> page_frags you would be using the atomic_long_sub_return which should

> be paths where you would not expect it to match or that are slowpath

> anyway.

> 

> Also I would keep the WARN_ON in both paths just to be on the safe side.


If I understand it correctly, we should change it as below, right?

static inline long page_pool_atomic_sub_frag_count_return(struct page *page,
							  long nr)
{
	long ret;

	/* As suggested by Alexander, atomic_long_read() may cover up the
	 * reference count errors, so avoid calling atomic_long_read() in
	 * the cases of freeing or draining the page_frags, where we would
	 * not expect it to match or that are slowpath anyway.
	 */
	if (__builtin_constant_p(nr) &&
	    atomic_long_read(&page->pp_frag_count) == nr)
		return 0;

	ret = atomic_long_sub_return(nr, &page->pp_frag_count);
	WARN_ON(ret < 0);
	return ret;
}


> .

>
Alexander Duyck July 22, 2021, 3:18 p.m. UTC | #4
> >

> >> You are right that that may cover up the reference count errors. How about

> >> something like below:

> >>

> >> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

> >>                                                           long nr)

> >> {

> >> #ifdef CONFIG_DEBUG_PAGE_REF

> >>         long ret = atomic_long_sub_return(nr, &page->pp_frag_count);

> >>

> >>         WARN_ON(ret < 0);

> >>

> >>         return ret;

> >> #else

> >>         if (atomic_long_read(&page->pp_frag_count) == nr)

> >>                 return 0;

> >>

> >>         return atomic_long_sub_return(nr, &page->pp_frag_count);

> >> #end

> >> }

> >>

> >> Or any better suggestion?

> >

> > So the one thing I might change would be to make it so that you only

> > do the atomic_long_read if nr is a constant via __builtin_constant_p.

> > That way you would be performing the comparison in

> > __page_pool_put_page and in the cases of freeing or draining the

> > page_frags you would be using the atomic_long_sub_return which should

> > be paths where you would not expect it to match or that are slowpath

> > anyway.

> >

> > Also I would keep the WARN_ON in both paths just to be on the safe side.

>

> If I understand it correctly, we should change it as below, right?

>

> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

>                                                           long nr)

> {

>         long ret;

>

>         /* As suggested by Alexander, atomic_long_read() may cover up the

>          * reference count errors, so avoid calling atomic_long_read() in

>          * the cases of freeing or draining the page_frags, where we would

>          * not expect it to match or that are slowpath anyway.

>          */

>         if (__builtin_constant_p(nr) &&

>             atomic_long_read(&page->pp_frag_count) == nr)

>                 return 0;

>

>         ret = atomic_long_sub_return(nr, &page->pp_frag_count);

>         WARN_ON(ret < 0);

>         return ret;

> }


Yes, that is what I had in mind.

One thought I had for a future optimization is that we could look at
reducing the count by 1 so that we could essentially combine the
non-frag and frag cases.Then instead of testing for 1 we would test
for 0 at thee start of the function and test for < 0 to decide if we
want to free it or not instead of testing for 0. With that we can
essentially reduce the calls to the WARN_ON since we should only have
one case where we actually return a value < 0, and we can then check
to see if we overshot -1 which would be the WARN_ON case.

With that a value of 0 instead of 1 would indicate page frag is not in
use for the page *AND/OR* that the page has reached the state where
there are no other frags present so the page can be recycled. In
effect it would allow us to mix page frags and no frags within the
same pool. The added bonus would be we could get rid of the check for
PP_FLAG_PAGE_FRAG flag check in the __page_pool_put_page function and
replace it with a check for PAGE_POOL_DMA_USE_PP_FRAG_COUNT since we
cannot read frag_count in that case.
Yunsheng Lin July 23, 2021, 11:12 a.m. UTC | #5
On 2021/7/22 23:18, Alexander Duyck wrote:
>>>

>>>> You are right that that may cover up the reference count errors. How about

>>>> something like below:

>>>>

>>>> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

>>>>                                                           long nr)

>>>> {

>>>> #ifdef CONFIG_DEBUG_PAGE_REF

>>>>         long ret = atomic_long_sub_return(nr, &page->pp_frag_count);

>>>>

>>>>         WARN_ON(ret < 0);

>>>>

>>>>         return ret;

>>>> #else

>>>>         if (atomic_long_read(&page->pp_frag_count) == nr)

>>>>                 return 0;

>>>>

>>>>         return atomic_long_sub_return(nr, &page->pp_frag_count);

>>>> #end

>>>> }

>>>>

>>>> Or any better suggestion?

>>>

>>> So the one thing I might change would be to make it so that you only

>>> do the atomic_long_read if nr is a constant via __builtin_constant_p.

>>> That way you would be performing the comparison in

>>> __page_pool_put_page and in the cases of freeing or draining the

>>> page_frags you would be using the atomic_long_sub_return which should

>>> be paths where you would not expect it to match or that are slowpath

>>> anyway.

>>>

>>> Also I would keep the WARN_ON in both paths just to be on the safe side.

>>

>> If I understand it correctly, we should change it as below, right?

>>

>> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

>>                                                           long nr)

>> {

>>         long ret;

>>

>>         /* As suggested by Alexander, atomic_long_read() may cover up the

>>          * reference count errors, so avoid calling atomic_long_read() in

>>          * the cases of freeing or draining the page_frags, where we would

>>          * not expect it to match or that are slowpath anyway.

>>          */

>>         if (__builtin_constant_p(nr) &&

>>             atomic_long_read(&page->pp_frag_count) == nr)

>>                 return 0;

>>

>>         ret = atomic_long_sub_return(nr, &page->pp_frag_count);

>>         WARN_ON(ret < 0);

>>         return ret;

>> }

> 

> Yes, that is what I had in mind.

> 

> One thought I had for a future optimization is that we could look at

> reducing the count by 1 so that we could essentially combine the

> non-frag and frag cases.Then instead of testing for 1 we would test

> for 0 at thee start of the function and test for < 0 to decide if we

> want to free it or not instead of testing for 0. With that we can

> essentially reduce the calls to the WARN_ON since we should only have

> one case where we actually return a value < 0, and we can then check

> to see if we overshot -1 which would be the WARN_ON case.

> 

> With that a value of 0 instead of 1 would indicate page frag is not in

> use for the page *AND/OR* that the page has reached the state where

> there are no other frags present so the page can be recycled. In

> effect it would allow us to mix page frags and no frags within the

> same pool. The added bonus would be we could get rid of the check for

> PP_FLAG_PAGE_FRAG flag check in the __page_pool_put_page function and

> replace it with a check for PAGE_POOL_DMA_USE_PP_FRAG_COUNT since we

> cannot read frag_count in that case.


Let's leave it for a future optimization.
I am not sure if there is use case to support both frag page and non-frag
page for the same page pool. If there is, maybe we can use "page->pp_frag_count
> 0" to indicate that the page is frag page, and "page->pp_frag_count == 0"

to indicate that the page is non-frag page, so that we can support frag page and
non-frag page for the same page pool instead of disabling non-frag page support
when PP_FLAG_PAGE_FRAG flag is set, which might be conflit with the above
optimization?


Also, I am prototyping the tx recycling based on page pool in order to see
if there is any value supporting the tx recycling.
As the busypoll has enable the one-to-one relation between NAPI and sock,
and there is one-to-one relation between NAPI and page pool, perhaps it make
senses that we use page pool to recycle the tx page too?

There are possibly below problems when doing that as I am aware of now:
1. busypoll is for rx, and tx may not be using the same queue as rx even if
   there are *technically* the same flow, so I am not sure it is ok to use
   busypoll infrastructure to get the page pool ptr for a specific sock.

2. There may be multi socks using the same page pool ptr to allocate page for
   multi flow, so we can not assume the same NAPI polling protection as rx,
   which might mean we can only use the recyclable page from pool->ring under the
   r->consumer_lock protection.

3. Right now tcp_sendmsg_locked() use sk_page_frag_refill() to refill the page
   frag for tcp xmit, when implementing a similar sk_page_pool_frag_refill()
   based on page pool, I found that tcp coalesce in tcp_mtu_probe() and
   tcp fragment in tso_fragment() might mess with the page_ref_count directly.

As the above the problem I am aware of(I believe there are other problems I am not
aware of yet), I am not sure if the tcp tx page recycling based on page pool is
doable or not, I would like to hear about your opinion about tcp tx recycling support
based on page pool first, in case it is a dead end to support that.

> .

>
Alexander Duyck July 23, 2021, 4:08 p.m. UTC | #6
On Fri, Jul 23, 2021 at 4:12 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:
>

> On 2021/7/22 23:18, Alexander Duyck wrote:

> >>>

> >>>> You are right that that may cover up the reference count errors. How about

> >>>> something like below:

> >>>>

> >>>> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

> >>>>                                                           long nr)

> >>>> {

> >>>> #ifdef CONFIG_DEBUG_PAGE_REF

> >>>>         long ret = atomic_long_sub_return(nr, &page->pp_frag_count);

> >>>>

> >>>>         WARN_ON(ret < 0);

> >>>>

> >>>>         return ret;

> >>>> #else

> >>>>         if (atomic_long_read(&page->pp_frag_count) == nr)

> >>>>                 return 0;

> >>>>

> >>>>         return atomic_long_sub_return(nr, &page->pp_frag_count);

> >>>> #end

> >>>> }

> >>>>

> >>>> Or any better suggestion?

> >>>

> >>> So the one thing I might change would be to make it so that you only

> >>> do the atomic_long_read if nr is a constant via __builtin_constant_p.

> >>> That way you would be performing the comparison in

> >>> __page_pool_put_page and in the cases of freeing or draining the

> >>> page_frags you would be using the atomic_long_sub_return which should

> >>> be paths where you would not expect it to match or that are slowpath

> >>> anyway.

> >>>

> >>> Also I would keep the WARN_ON in both paths just to be on the safe side.

> >>

> >> If I understand it correctly, we should change it as below, right?

> >>

> >> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

> >>                                                           long nr)

> >> {

> >>         long ret;

> >>

> >>         /* As suggested by Alexander, atomic_long_read() may cover up the

> >>          * reference count errors, so avoid calling atomic_long_read() in

> >>          * the cases of freeing or draining the page_frags, where we would

> >>          * not expect it to match or that are slowpath anyway.

> >>          */

> >>         if (__builtin_constant_p(nr) &&

> >>             atomic_long_read(&page->pp_frag_count) == nr)

> >>                 return 0;

> >>

> >>         ret = atomic_long_sub_return(nr, &page->pp_frag_count);

> >>         WARN_ON(ret < 0);

> >>         return ret;

> >> }

> >

> > Yes, that is what I had in mind.

> >

> > One thought I had for a future optimization is that we could look at

> > reducing the count by 1 so that we could essentially combine the

> > non-frag and frag cases.Then instead of testing for 1 we would test

> > for 0 at thee start of the function and test for < 0 to decide if we

> > want to free it or not instead of testing for 0. With that we can

> > essentially reduce the calls to the WARN_ON since we should only have

> > one case where we actually return a value < 0, and we can then check

> > to see if we overshot -1 which would be the WARN_ON case.

> >

> > With that a value of 0 instead of 1 would indicate page frag is not in

> > use for the page *AND/OR* that the page has reached the state where

> > there are no other frags present so the page can be recycled. In

> > effect it would allow us to mix page frags and no frags within the

> > same pool. The added bonus would be we could get rid of the check for

> > PP_FLAG_PAGE_FRAG flag check in the __page_pool_put_page function and

> > replace it with a check for PAGE_POOL_DMA_USE_PP_FRAG_COUNT since we

> > cannot read frag_count in that case.

>

> Let's leave it for a future optimization.

> I am not sure if there is use case to support both frag page and non-frag

> page for the same page pool. If there is, maybe we can use "page->pp_frag_count

> > 0" to indicate that the page is frag page, and "page->pp_frag_count == 0"

> to indicate that the page is non-frag page, so that we can support frag page and

> non-frag page for the same page pool instead of disabling non-frag page support

> when PP_FLAG_PAGE_FRAG flag is set, which might be conflit with the above

> optimization?


As far as use case I can see a number of potential uses. For example
in the case of drivers that do something like a header/data split I
could see potentially having the header pages be frags while the data
pages being 4K blocks. Basically the big optimization of the count ==
1/0/nr case is that you aren't increasing/decreasing the count and it
is immediately being recycled/reused. So in such a case being able to
add frag count some pages, and not to others would likely be quite
useful.

Basically by shifting the pool values by 1 you can have both in the
same pool with little issue. However the big change is that instead of
testing for count = nr it would end up being pp_frag_count = nr - 1.
So in the case of the standard page pool pages being freed or the last
frag you would be looking at pp_frag_count = 0. In addition we can
mask the WARN_ON overhead as you would be using -1 as the point to
free so you would only have to perform the WARN_ON check for the last
frag instead of every frag.

> Also, I am prototyping the tx recycling based on page pool in order to see

> if there is any value supporting the tx recycling.


Just to clarify here when you say Tx recycling you are talking about
socket to netdev correct? Just want to be certain since the netdev to
netdev case should already have recycling for page pool pages as long
as it follows a 1<->1 path.

> As the busypoll has enable the one-to-one relation between NAPI and sock,

> and there is one-to-one relation between NAPI and page pool, perhaps it make

> senses that we use page pool to recycle the tx page too?

>

> There are possibly below problems when doing that as I am aware of now:

> 1. busypoll is for rx, and tx may not be using the same queue as rx even if

>    there are *technically* the same flow, so I am not sure it is ok to use

>    busypoll infrastructure to get the page pool ptr for a specific sock.

>

> 2. There may be multi socks using the same page pool ptr to allocate page for

>    multi flow, so we can not assume the same NAPI polling protection as rx,

>    which might mean we can only use the recyclable page from pool->ring under the

>    r->consumer_lock protection.

>

> 3. Right now tcp_sendmsg_locked() use sk_page_frag_refill() to refill the page

>    frag for tcp xmit, when implementing a similar sk_page_pool_frag_refill()

>    based on page pool, I found that tcp coalesce in tcp_mtu_probe() and

>    tcp fragment in tso_fragment() might mess with the page_ref_count directly.

>

> As the above the problem I am aware of(I believe there are other problems I am not

> aware of yet), I am not sure if the tcp tx page recycling based on page pool is

> doable or not, I would like to hear about your opinion about tcp tx recycling support

> based on page pool first, in case it is a dead end to support that.


I'm honestly not sure there is much there to gain. Last I knew TCP was
using order 3 pages for transmitting and as a result the overhead for
the pages should already be greatly reduced. In addition one of the
main reasons for page_pool  is the fact that the device has to DMA map
the page and that can have very high overhead on systems with an
IOMMU.

Rather than trying to reuse the devices page pool it might make more
sense to see if you couldn't have TCP just use some sort of circular
buffer of memory that is directly mapped for the device that it is
going to be transmitting to. Essentially what you would be doing is
creating a pre-mapped page and would need to communicate that the
memory is already mapped for the device you want to send it to so that
it could skip that step.
Yunsheng Lin July 24, 2021, 1:07 p.m. UTC | #7
On Fri, Jul 23, 2021 at 09:08:00AM -0700, Alexander Duyck wrote:
> On Fri, Jul 23, 2021 at 4:12 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:

> >

> > On 2021/7/22 23:18, Alexander Duyck wrote:

> > >>>

> > >>>> You are right that that may cover up the reference count errors. How about

> > >>>> something like below:

> > >>>>

> > >>>> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

> > >>>>                                                           long nr)

> > >>>> {

> > >>>> #ifdef CONFIG_DEBUG_PAGE_REF

> > >>>>         long ret = atomic_long_sub_return(nr, &page->pp_frag_count);

> > >>>>

> > >>>>         WARN_ON(ret < 0);

> > >>>>

> > >>>>         return ret;

> > >>>> #else

> > >>>>         if (atomic_long_read(&page->pp_frag_count) == nr)

> > >>>>                 return 0;

> > >>>>

> > >>>>         return atomic_long_sub_return(nr, &page->pp_frag_count);

> > >>>> #end

> > >>>> }

> > >>>>

> > >>>> Or any better suggestion?

> > >>>

> > >>> So the one thing I might change would be to make it so that you only

> > >>> do the atomic_long_read if nr is a constant via __builtin_constant_p.

> > >>> That way you would be performing the comparison in

> > >>> __page_pool_put_page and in the cases of freeing or draining the

> > >>> page_frags you would be using the atomic_long_sub_return which should

> > >>> be paths where you would not expect it to match or that are slowpath

> > >>> anyway.

> > >>>

> > >>> Also I would keep the WARN_ON in both paths just to be on the safe side.

> > >>

> > >> If I understand it correctly, we should change it as below, right?

> > >>

> > >> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

> > >>                                                           long nr)

> > >> {

> > >>         long ret;

> > >>

> > >>         /* As suggested by Alexander, atomic_long_read() may cover up the

> > >>          * reference count errors, so avoid calling atomic_long_read() in

> > >>          * the cases of freeing or draining the page_frags, where we would

> > >>          * not expect it to match or that are slowpath anyway.

> > >>          */

> > >>         if (__builtin_constant_p(nr) &&

> > >>             atomic_long_read(&page->pp_frag_count) == nr)

> > >>                 return 0;

> > >>

> > >>         ret = atomic_long_sub_return(nr, &page->pp_frag_count);

> > >>         WARN_ON(ret < 0);

> > >>         return ret;

> > >> }

> > >

> > > Yes, that is what I had in mind.

> > >

> > > One thought I had for a future optimization is that we could look at

> > > reducing the count by 1 so that we could essentially combine the

> > > non-frag and frag cases.Then instead of testing for 1 we would test

> > > for 0 at thee start of the function and test for < 0 to decide if we

> > > want to free it or not instead of testing for 0. With that we can

> > > essentially reduce the calls to the WARN_ON since we should only have

> > > one case where we actually return a value < 0, and we can then check

> > > to see if we overshot -1 which would be the WARN_ON case.

> > >

> > > With that a value of 0 instead of 1 would indicate page frag is not in

> > > use for the page *AND/OR* that the page has reached the state where

> > > there are no other frags present so the page can be recycled. In

> > > effect it would allow us to mix page frags and no frags within the

> > > same pool. The added bonus would be we could get rid of the check for

> > > PP_FLAG_PAGE_FRAG flag check in the __page_pool_put_page function and

> > > replace it with a check for PAGE_POOL_DMA_USE_PP_FRAG_COUNT since we

> > > cannot read frag_count in that case.

> >

> > Let's leave it for a future optimization.

> > I am not sure if there is use case to support both frag page and non-frag

> > page for the same page pool. If there is, maybe we can use "page->pp_frag_count

> > > 0" to indicate that the page is frag page, and "page->pp_frag_count == 0"

> > to indicate that the page is non-frag page, so that we can support frag page and

> > non-frag page for the same page pool instead of disabling non-frag page support

> > when PP_FLAG_PAGE_FRAG flag is set, which might be conflit with the above

> > optimization?

> 

> As far as use case I can see a number of potential uses. For example

> in the case of drivers that do something like a header/data split I

> could see potentially having the header pages be frags while the data

> pages being 4K blocks. Basically the big optimization of the count ==

> 1/0/nr case is that you aren't increasing/decreasing the count and it

> is immediately being recycled/reused. So in such a case being able to

> add frag count some pages, and not to others would likely be quite

> useful.


I am not sure how the header/data split is implemented in hw, but it
seems the driver is not able to tell which desc will be filled with
header or data in advance, so it might need to allocate 4K block for
all desc?

> 

> Basically by shifting the pool values by 1 you can have both in the

> same pool with little issue. However the big change is that instead of

> testing for count = nr it would end up being pp_frag_count = nr - 1.

> So in the case of the standard page pool pages being freed or the last

> frag you would be looking at pp_frag_count = 0. In addition we can

> mask the WARN_ON overhead as you would be using -1 as the point to

> free so you would only have to perform the WARN_ON check for the last

> frag instead of every frag.


Yes, it seems doable.

> 

> > Also, I am prototyping the tx recycling based on page pool in order to see

> > if there is any value supporting the tx recycling.

> 

> Just to clarify here when you say Tx recycling you are talking about

> socket to netdev correct? Just want to be certain since the netdev to

> netdev case should already have recycling for page pool pages as long

> as it follows a 1<->1 path.


Yes, the above Tx recycling meant socket to netdev.
Also, the above "netdev to netdev" only meant XDP now, but not the IP
forwarding path in the network stack, right?

> 

> > As the busypoll has enable the one-to-one relation between NAPI and sock,

> > and there is one-to-one relation between NAPI and page pool, perhaps it make

> > senses that we use page pool to recycle the tx page too?

> >

> > There are possibly below problems when doing that as I am aware of now:

> > 1. busypoll is for rx, and tx may not be using the same queue as rx even if

> >    there are *technically* the same flow, so I am not sure it is ok to use

> >    busypoll infrastructure to get the page pool ptr for a specific sock.

> >

> > 2. There may be multi socks using the same page pool ptr to allocate page for

> >    multi flow, so we can not assume the same NAPI polling protection as rx,

> >    which might mean we can only use the recyclable page from pool->ring under the

> >    r->consumer_lock protection.

> >

> > 3. Right now tcp_sendmsg_locked() use sk_page_frag_refill() to refill the page

> >    frag for tcp xmit, when implementing a similar sk_page_pool_frag_refill()

> >    based on page pool, I found that tcp coalesce in tcp_mtu_probe() and

> >    tcp fragment in tso_fragment() might mess with the page_ref_count directly.

> >

> > As the above the problem I am aware of(I believe there are other problems I am not

> > aware of yet), I am not sure if the tcp tx page recycling based on page pool is

> > doable or not, I would like to hear about your opinion about tcp tx recycling support

> > based on page pool first, in case it is a dead end to support that.

> 

> I'm honestly not sure there is much there to gain. Last I knew TCP was

> using order 3 pages for transmitting and as a result the overhead for

> the pages should already be greatly reduced. In addition one of the

> main reasons for page_pool  is the fact that the device has to DMA map

> the page and that can have very high overhead on systems with an

> IOMMU.


Yes, avoiding the IOMMU overhead is the main gain. and "order 3 pages"
seems to be disabled on defaut?

> 

> Rather than trying to reuse the devices page pool it might make more

> sense to see if you couldn't have TCP just use some sort of circular

> buffer of memory that is directly mapped for the device that it is

> going to be transmitting to. Essentially what you would be doing is

> creating a pre-mapped page and would need to communicate that the

> memory is already mapped for the device you want to send it to so that

> it could skip that step.


IIUC sk_page_frag_refill() is already doing a similar reusing as the
rx reusing implemented in most driver except for the not pre-mapping
part.

And it seems that even if we pre-map the page and communicate that the
memory is already mapped to the driver, it is likely that we will not
be able to reuse the page when the circular buffer is not big enough
or tx completion/tcp ack is not happening quickly enough, which might
means unmapping/deallocating old circular buffer and allocating/mapping
new circular buffer.

Using page pool we might be able to alleviate the above problem as it
does for rx?
Alexander Duyck July 25, 2021, 4:49 p.m. UTC | #8
On Sat, Jul 24, 2021 at 6:07 AM Yunsheng Lin <yunshenglin0825@gmail.com> wrote:
>

> On Fri, Jul 23, 2021 at 09:08:00AM -0700, Alexander Duyck wrote:

> > On Fri, Jul 23, 2021 at 4:12 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:

> > >

> > > On 2021/7/22 23:18, Alexander Duyck wrote:

> > > >>>

> > > >>>> You are right that that may cover up the reference count errors. How about

> > > >>>> something like below:

> > > >>>>

> > > >>>> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

> > > >>>>                                                           long nr)

> > > >>>> {

> > > >>>> #ifdef CONFIG_DEBUG_PAGE_REF

> > > >>>>         long ret = atomic_long_sub_return(nr, &page->pp_frag_count);

> > > >>>>

> > > >>>>         WARN_ON(ret < 0);

> > > >>>>

> > > >>>>         return ret;

> > > >>>> #else

> > > >>>>         if (atomic_long_read(&page->pp_frag_count) == nr)

> > > >>>>                 return 0;

> > > >>>>

> > > >>>>         return atomic_long_sub_return(nr, &page->pp_frag_count);

> > > >>>> #end

> > > >>>> }

> > > >>>>

> > > >>>> Or any better suggestion?

> > > >>>

> > > >>> So the one thing I might change would be to make it so that you only

> > > >>> do the atomic_long_read if nr is a constant via __builtin_constant_p.

> > > >>> That way you would be performing the comparison in

> > > >>> __page_pool_put_page and in the cases of freeing or draining the

> > > >>> page_frags you would be using the atomic_long_sub_return which should

> > > >>> be paths where you would not expect it to match or that are slowpath

> > > >>> anyway.

> > > >>>

> > > >>> Also I would keep the WARN_ON in both paths just to be on the safe side.

> > > >>

> > > >> If I understand it correctly, we should change it as below, right?

> > > >>

> > > >> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

> > > >>                                                           long nr)

> > > >> {

> > > >>         long ret;

> > > >>

> > > >>         /* As suggested by Alexander, atomic_long_read() may cover up the

> > > >>          * reference count errors, so avoid calling atomic_long_read() in

> > > >>          * the cases of freeing or draining the page_frags, where we would

> > > >>          * not expect it to match or that are slowpath anyway.

> > > >>          */

> > > >>         if (__builtin_constant_p(nr) &&

> > > >>             atomic_long_read(&page->pp_frag_count) == nr)

> > > >>                 return 0;

> > > >>

> > > >>         ret = atomic_long_sub_return(nr, &page->pp_frag_count);

> > > >>         WARN_ON(ret < 0);

> > > >>         return ret;

> > > >> }

> > > >

> > > > Yes, that is what I had in mind.

> > > >

> > > > One thought I had for a future optimization is that we could look at

> > > > reducing the count by 1 so that we could essentially combine the

> > > > non-frag and frag cases.Then instead of testing for 1 we would test

> > > > for 0 at thee start of the function and test for < 0 to decide if we

> > > > want to free it or not instead of testing for 0. With that we can

> > > > essentially reduce the calls to the WARN_ON since we should only have

> > > > one case where we actually return a value < 0, and we can then check

> > > > to see if we overshot -1 which would be the WARN_ON case.

> > > >

> > > > With that a value of 0 instead of 1 would indicate page frag is not in

> > > > use for the page *AND/OR* that the page has reached the state where

> > > > there are no other frags present so the page can be recycled. In

> > > > effect it would allow us to mix page frags and no frags within the

> > > > same pool. The added bonus would be we could get rid of the check for

> > > > PP_FLAG_PAGE_FRAG flag check in the __page_pool_put_page function and

> > > > replace it with a check for PAGE_POOL_DMA_USE_PP_FRAG_COUNT since we

> > > > cannot read frag_count in that case.

> > >

> > > Let's leave it for a future optimization.

> > > I am not sure if there is use case to support both frag page and non-frag

> > > page for the same page pool. If there is, maybe we can use "page->pp_frag_count

> > > > 0" to indicate that the page is frag page, and "page->pp_frag_count == 0"

> > > to indicate that the page is non-frag page, so that we can support frag page and

> > > non-frag page for the same page pool instead of disabling non-frag page support

> > > when PP_FLAG_PAGE_FRAG flag is set, which might be conflit with the above

> > > optimization?

> >

> > As far as use case I can see a number of potential uses. For example

> > in the case of drivers that do something like a header/data split I

> > could see potentially having the header pages be frags while the data

> > pages being 4K blocks. Basically the big optimization of the count ==

> > 1/0/nr case is that you aren't increasing/decreasing the count and it

> > is immediately being recycled/reused. So in such a case being able to

> > add frag count some pages, and not to others would likely be quite

> > useful.

>

> I am not sure how the header/data split is implemented in hw, but it

> seems the driver is not able to tell which desc will be filled with

> header or data in advance, so it might need to allocate 4K block for

> all desc?


It all depends on the hardware config. In theory you could have
anything from a single use for a page to multiple uses for a page in
the case of headers and/or packets being small. The overhead for
adding/removing the frag count could end up being more than what is
needed if the page is only used once. That is why I was thinking it
might make sense to allow both to coexist in the same pool.

> >

> > Basically by shifting the pool values by 1 you can have both in the

> > same pool with little issue. However the big change is that instead of

> > testing for count = nr it would end up being pp_frag_count = nr - 1.

> > So in the case of the standard page pool pages being freed or the last

> > frag you would be looking at pp_frag_count = 0. In addition we can

> > mask the WARN_ON overhead as you would be using -1 as the point to

> > free so you would only have to perform the WARN_ON check for the last

> > frag instead of every frag.

>

> Yes, it seems doable.

>

> >

> > > Also, I am prototyping the tx recycling based on page pool in order to see

> > > if there is any value supporting the tx recycling.

> >

> > Just to clarify here when you say Tx recycling you are talking about

> > socket to netdev correct? Just want to be certain since the netdev to

> > netdev case should already have recycling for page pool pages as long

> > as it follows a 1<->1 path.

>

> Yes, the above Tx recycling meant socket to netdev.

> Also, the above "netdev to netdev" only meant XDP now, but not the IP

> forwarding path in the network stack, right?

>

> >

> > > As the busypoll has enable the one-to-one relation between NAPI and sock,

> > > and there is one-to-one relation between NAPI and page pool, perhaps it make

> > > senses that we use page pool to recycle the tx page too?

> > >

> > > There are possibly below problems when doing that as I am aware of now:

> > > 1. busypoll is for rx, and tx may not be using the same queue as rx even if

> > >    there are *technically* the same flow, so I am not sure it is ok to use

> > >    busypoll infrastructure to get the page pool ptr for a specific sock.

> > >

> > > 2. There may be multi socks using the same page pool ptr to allocate page for

> > >    multi flow, so we can not assume the same NAPI polling protection as rx,

> > >    which might mean we can only use the recyclable page from pool->ring under the

> > >    r->consumer_lock protection.

> > >

> > > 3. Right now tcp_sendmsg_locked() use sk_page_frag_refill() to refill the page

> > >    frag for tcp xmit, when implementing a similar sk_page_pool_frag_refill()

> > >    based on page pool, I found that tcp coalesce in tcp_mtu_probe() and

> > >    tcp fragment in tso_fragment() might mess with the page_ref_count directly.

> > >

> > > As the above the problem I am aware of(I believe there are other problems I am not

> > > aware of yet), I am not sure if the tcp tx page recycling based on page pool is

> > > doable or not, I would like to hear about your opinion about tcp tx recycling support

> > > based on page pool first, in case it is a dead end to support that.

> >

> > I'm honestly not sure there is much there to gain. Last I knew TCP was

> > using order 3 pages for transmitting and as a result the overhead for

> > the pages should already be greatly reduced. In addition one of the

> > main reasons for page_pool  is the fact that the device has to DMA map

> > the page and that can have very high overhead on systems with an

> > IOMMU.

>

> Yes, avoiding the IOMMU overhead is the main gain. and "order 3 pages"

> seems to be disabled on defaut?

>

> >

> > Rather than trying to reuse the devices page pool it might make more

> > sense to see if you couldn't have TCP just use some sort of circular

> > buffer of memory that is directly mapped for the device that it is

> > going to be transmitting to. Essentially what you would be doing is

> > creating a pre-mapped page and would need to communicate that the

> > memory is already mapped for the device you want to send it to so that

> > it could skip that step.

>

> IIUC sk_page_frag_refill() is already doing a similar reusing as the

> rx reusing implemented in most driver except for the not pre-mapping

> part.

>

> And it seems that even if we pre-map the page and communicate that the

> memory is already mapped to the driver, it is likely that we will not

> be able to reuse the page when the circular buffer is not big enough

> or tx completion/tcp ack is not happening quickly enough, which might

> means unmapping/deallocating old circular buffer and allocating/mapping

> new circular buffer.

>

> Using page pool we might be able to alleviate the above problem as it

> does for rx?


I would say that instead of looking at going straight for the page
pool it might make more sense to look at seeing if we can coalesce the
DMA mapping of the pages first at the socket layer rather than trying
to introduce the overhead for the page pool. In the case of sockets we
already have the destructors that are called when the memory is freed,
so instead of making sockets use page pool it might make more sense to
extend the socket buffer allocation/freeing to incorporate bulk
mapping and unmapping of pages to optimize the socket Tx path in the
32K page case.
Yunsheng Lin July 27, 2021, 7:54 a.m. UTC | #9
On 2021/7/26 0:49, Alexander Duyck wrote:
> On Sat, Jul 24, 2021 at 6:07 AM Yunsheng Lin <yunshenglin0825@gmail.com> wrote:

>>

>> On Fri, Jul 23, 2021 at 09:08:00AM -0700, Alexander Duyck wrote:

>>> On Fri, Jul 23, 2021 at 4:12 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:

>>>>

>>>> On 2021/7/22 23:18, Alexander Duyck wrote:

>>>>>>>

>>>>>>>> You are right that that may cover up the reference count errors. How about

>>>>>>>> something like below:

>>>>>>>>

>>>>>>>> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

>>>>>>>>                                                           long nr)

>>>>>>>> {

>>>>>>>> #ifdef CONFIG_DEBUG_PAGE_REF

>>>>>>>>         long ret = atomic_long_sub_return(nr, &page->pp_frag_count);

>>>>>>>>

>>>>>>>>         WARN_ON(ret < 0);

>>>>>>>>

>>>>>>>>         return ret;

>>>>>>>> #else

>>>>>>>>         if (atomic_long_read(&page->pp_frag_count) == nr)

>>>>>>>>                 return 0;

>>>>>>>>

>>>>>>>>         return atomic_long_sub_return(nr, &page->pp_frag_count);

>>>>>>>> #end

>>>>>>>> }

>>>>>>>>

>>>>>>>> Or any better suggestion?

>>>>>>>

>>>>>>> So the one thing I might change would be to make it so that you only

>>>>>>> do the atomic_long_read if nr is a constant via __builtin_constant_p.

>>>>>>> That way you would be performing the comparison in

>>>>>>> __page_pool_put_page and in the cases of freeing or draining the

>>>>>>> page_frags you would be using the atomic_long_sub_return which should

>>>>>>> be paths where you would not expect it to match or that are slowpath

>>>>>>> anyway.

>>>>>>>

>>>>>>> Also I would keep the WARN_ON in both paths just to be on the safe side.

>>>>>>

>>>>>> If I understand it correctly, we should change it as below, right?

>>>>>>

>>>>>> static inline long page_pool_atomic_sub_frag_count_return(struct page *page,

>>>>>>                                                           long nr)

>>>>>> {

>>>>>>         long ret;

>>>>>>

>>>>>>         /* As suggested by Alexander, atomic_long_read() may cover up the

>>>>>>          * reference count errors, so avoid calling atomic_long_read() in

>>>>>>          * the cases of freeing or draining the page_frags, where we would

>>>>>>          * not expect it to match or that are slowpath anyway.

>>>>>>          */

>>>>>>         if (__builtin_constant_p(nr) &&

>>>>>>             atomic_long_read(&page->pp_frag_count) == nr)

>>>>>>                 return 0;

>>>>>>

>>>>>>         ret = atomic_long_sub_return(nr, &page->pp_frag_count);

>>>>>>         WARN_ON(ret < 0);

>>>>>>         return ret;

>>>>>> }

>>>>>

>>>>> Yes, that is what I had in mind.

>>>>>

>>>>> One thought I had for a future optimization is that we could look at

>>>>> reducing the count by 1 so that we could essentially combine the

>>>>> non-frag and frag cases.Then instead of testing for 1 we would test

>>>>> for 0 at thee start of the function and test for < 0 to decide if we

>>>>> want to free it or not instead of testing for 0. With that we can

>>>>> essentially reduce the calls to the WARN_ON since we should only have

>>>>> one case where we actually return a value < 0, and we can then check

>>>>> to see if we overshot -1 which would be the WARN_ON case.

>>>>>

>>>>> With that a value of 0 instead of 1 would indicate page frag is not in

>>>>> use for the page *AND/OR* that the page has reached the state where

>>>>> there are no other frags present so the page can be recycled. In

>>>>> effect it would allow us to mix page frags and no frags within the

>>>>> same pool. The added bonus would be we could get rid of the check for

>>>>> PP_FLAG_PAGE_FRAG flag check in the __page_pool_put_page function and

>>>>> replace it with a check for PAGE_POOL_DMA_USE_PP_FRAG_COUNT since we

>>>>> cannot read frag_count in that case.

>>>>

>>>> Let's leave it for a future optimization.

>>>> I am not sure if there is use case to support both frag page and non-frag

>>>> page for the same page pool. If there is, maybe we can use "page->pp_frag_count

>>>>> 0" to indicate that the page is frag page, and "page->pp_frag_count == 0"

>>>> to indicate that the page is non-frag page, so that we can support frag page and

>>>> non-frag page for the same page pool instead of disabling non-frag page support

>>>> when PP_FLAG_PAGE_FRAG flag is set, which might be conflit with the above

>>>> optimization?

>>>

>>> As far as use case I can see a number of potential uses. For example

>>> in the case of drivers that do something like a header/data split I

>>> could see potentially having the header pages be frags while the data

>>> pages being 4K blocks. Basically the big optimization of the count ==

>>> 1/0/nr case is that you aren't increasing/decreasing the count and it

>>> is immediately being recycled/reused. So in such a case being able to

>>> add frag count some pages, and not to others would likely be quite

>>> useful.

>>

>> I am not sure how the header/data split is implemented in hw, but it

>> seems the driver is not able to tell which desc will be filled with

>> header or data in advance, so it might need to allocate 4K block for

>> all desc?

> 

> It all depends on the hardware config. In theory you could have

> anything from a single use for a page to multiple uses for a page in

> the case of headers and/or packets being small. The overhead for

> adding/removing the frag count could end up being more than what is

> needed if the page is only used once. That is why I was thinking it

> might make sense to allow both to coexist in the same pool.


I am agreed that there may be usecase of using both frag page and non-frag
page of the same page pool. Let's leave it for now.

> 

>>>

>>> Basically by shifting the pool values by 1 you can have both in the

>>> same pool with little issue. However the big change is that instead of

>>> testing for count = nr it would end up being pp_frag_count = nr - 1.

>>> So in the case of the standard page pool pages being freed or the last

>>> frag you would be looking at pp_frag_count = 0. In addition we can

>>> mask the WARN_ON overhead as you would be using -1 as the point to

>>> free so you would only have to perform the WARN_ON check for the last

>>> frag instead of every frag.

>>

>> Yes, it seems doable.

>>

>>>

>>>> Also, I am prototyping the tx recycling based on page pool in order to see

>>>> if there is any value supporting the tx recycling.

>>>

>>> Just to clarify here when you say Tx recycling you are talking about

>>> socket to netdev correct? Just want to be certain since the netdev to

>>> netdev case should already have recycling for page pool pages as long

>>> as it follows a 1<->1 path.

>>

>> Yes, the above Tx recycling meant socket to netdev.

>> Also, the above "netdev to netdev" only meant XDP now, but not the IP

>> forwarding path in the network stack, right?

>>

>>>

>>>> As the busypoll has enable the one-to-one relation between NAPI and sock,

>>>> and there is one-to-one relation between NAPI and page pool, perhaps it make

>>>> senses that we use page pool to recycle the tx page too?

>>>>

>>>> There are possibly below problems when doing that as I am aware of now:

>>>> 1. busypoll is for rx, and tx may not be using the same queue as rx even if

>>>>    there are *technically* the same flow, so I am not sure it is ok to use

>>>>    busypoll infrastructure to get the page pool ptr for a specific sock.

>>>>

>>>> 2. There may be multi socks using the same page pool ptr to allocate page for

>>>>    multi flow, so we can not assume the same NAPI polling protection as rx,

>>>>    which might mean we can only use the recyclable page from pool->ring under the

>>>>    r->consumer_lock protection.

>>>>

>>>> 3. Right now tcp_sendmsg_locked() use sk_page_frag_refill() to refill the page

>>>>    frag for tcp xmit, when implementing a similar sk_page_pool_frag_refill()

>>>>    based on page pool, I found that tcp coalesce in tcp_mtu_probe() and

>>>>    tcp fragment in tso_fragment() might mess with the page_ref_count directly.

>>>>

>>>> As the above the problem I am aware of(I believe there are other problems I am not

>>>> aware of yet), I am not sure if the tcp tx page recycling based on page pool is

>>>> doable or not, I would like to hear about your opinion about tcp tx recycling support

>>>> based on page pool first, in case it is a dead end to support that.

>>>

>>> I'm honestly not sure there is much there to gain. Last I knew TCP was

>>> using order 3 pages for transmitting and as a result the overhead for

>>> the pages should already be greatly reduced. In addition one of the

>>> main reasons for page_pool  is the fact that the device has to DMA map

>>> the page and that can have very high overhead on systems with an

>>> IOMMU.

>>

>> Yes, avoiding the IOMMU overhead is the main gain. and "order 3 pages"

>> seems to be disabled on defaut?

>>

>>>

>>> Rather than trying to reuse the devices page pool it might make more

>>> sense to see if you couldn't have TCP just use some sort of circular

>>> buffer of memory that is directly mapped for the device that it is

>>> going to be transmitting to. Essentially what you would be doing is

>>> creating a pre-mapped page and would need to communicate that the

>>> memory is already mapped for the device you want to send it to so that

>>> it could skip that step.

>>

>> IIUC sk_page_frag_refill() is already doing a similar reusing as the

>> rx reusing implemented in most driver except for the not pre-mapping

>> part.

>>

>> And it seems that even if we pre-map the page and communicate that the

>> memory is already mapped to the driver, it is likely that we will not

>> be able to reuse the page when the circular buffer is not big enough

>> or tx completion/tcp ack is not happening quickly enough, which might

>> means unmapping/deallocating old circular buffer and allocating/mapping

>> new circular buffer.

>>

>> Using page pool we might be able to alleviate the above problem as it

>> does for rx?

> 

> I would say that instead of looking at going straight for the page

> pool it might make more sense to look at seeing if we can coalesce the

> DMA mapping of the pages first at the socket layer rather than trying

> to introduce the overhead for the page pool. In the case of sockets we

> already have the destructors that are called when the memory is freed,

> so instead of making sockets use page pool it might make more sense to

> extend the socket buffer allocation/freeing to incorporate bulk

> mapping and unmapping of pages to optimize the socket Tx path in the

> 32K page case.


I was able to enable tx recycling prototyping based on page pool to
run some performance test, the performance improvement is about +20%
(30Gbit -> 38Gbit) for single thread iperf tcp flow when IOMMU is in
strict mode. And CPU usage descreases about 10% for four threads iperf
tcp flow for line speed of 100Gbit when IOMMU is in strict mode.

Looking at the prototyping code, I am agreed that it is a bit controversial
to use the page pool for tx as the page pool is assuming NAPI polling
protection for allocation side.

So I will take a deeper look about your suggestion above to see how to
implement it.

Also, I am assuming the "destructors" means tcp_wfree() for TCP, right?
It seems tcp_wfree() is mainly used to do memory accounting and free
"struct sock" if necessary.
I am not so familiar with socket layer to understand how the "destructors"
will be helpful here, any detailed idea how to use "destructors" here?

> .

>
Alexander Duyck July 27, 2021, 6:38 p.m. UTC | #10
On Tue, Jul 27, 2021 at 12:54 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:
>

> On 2021/7/26 0:49, Alexander Duyck wrote:

> > On Sat, Jul 24, 2021 at 6:07 AM Yunsheng Lin <yunshenglin0825@gmail.com> wrote:

> >>

> >> On Fri, Jul 23, 2021 at 09:08:00AM -0700, Alexander Duyck wrote:

> >>> On Fri, Jul 23, 2021 at 4:12 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:

> >>>>

> >>>> On 2021/7/22 23:18, Alexander Duyck wrote:

> >>>>>>>


<snip>

> >>>

> >>> Rather than trying to reuse the devices page pool it might make more

> >>> sense to see if you couldn't have TCP just use some sort of circular

> >>> buffer of memory that is directly mapped for the device that it is

> >>> going to be transmitting to. Essentially what you would be doing is

> >>> creating a pre-mapped page and would need to communicate that the

> >>> memory is already mapped for the device you want to send it to so that

> >>> it could skip that step.

> >>

> >> IIUC sk_page_frag_refill() is already doing a similar reusing as the

> >> rx reusing implemented in most driver except for the not pre-mapping

> >> part.

> >>

> >> And it seems that even if we pre-map the page and communicate that the

> >> memory is already mapped to the driver, it is likely that we will not

> >> be able to reuse the page when the circular buffer is not big enough

> >> or tx completion/tcp ack is not happening quickly enough, which might

> >> means unmapping/deallocating old circular buffer and allocating/mapping

> >> new circular buffer.

> >>

> >> Using page pool we might be able to alleviate the above problem as it

> >> does for rx?

> >

> > I would say that instead of looking at going straight for the page

> > pool it might make more sense to look at seeing if we can coalesce the

> > DMA mapping of the pages first at the socket layer rather than trying

> > to introduce the overhead for the page pool. In the case of sockets we

> > already have the destructors that are called when the memory is freed,

> > so instead of making sockets use page pool it might make more sense to

> > extend the socket buffer allocation/freeing to incorporate bulk

> > mapping and unmapping of pages to optimize the socket Tx path in the

> > 32K page case.

>

> I was able to enable tx recycling prototyping based on page pool to

> run some performance test, the performance improvement is about +20%

> (30Gbit -> 38Gbit) for single thread iperf tcp flow when IOMMU is in

> strict mode. And CPU usage descreases about 10% for four threads iperf

> tcp flow for line speed of 100Gbit when IOMMU is in strict mode.


That isn't surprising given that for most devices the IOMMU will be
called per frag which can add a fair bit of overhead.

> Looking at the prototyping code, I am agreed that it is a bit controversial

> to use the page pool for tx as the page pool is assuming NAPI polling

> protection for allocation side.

>

> So I will take a deeper look about your suggestion above to see how to

> implement it.

>

> Also, I am assuming the "destructors" means tcp_wfree() for TCP, right?

> It seems tcp_wfree() is mainly used to do memory accounting and free

> "struct sock" if necessary.


Yes, that is what I was thinking. If we had some way to add something
like an argument or way to push the information about where the skbs
are being freed back to the socket the socket could then be looking at
pre-mapping the pages for the device if we assume a 1:1 mapping from
the socket to the device.

> I am not so familiar with socket layer to understand how the "destructors"

> will be helpful here, any detailed idea how to use "destructors" here?


The basic idea is the destructors are called when the skb is orphaned
or freed. So it might be a good spot to put in any logic to free pages
from your special pool. The only thing you would need to sort out is
making certain to bump reference counts appropriately if the skb is
cloned and the destructor is copied.
Yunsheng Lin Aug. 2, 2021, 9:17 a.m. UTC | #11
On 2021/7/28 2:38, Alexander Duyck wrote:
> On Tue, Jul 27, 2021 at 12:54 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:

>>

>> On 2021/7/26 0:49, Alexander Duyck wrote:

>>> On Sat, Jul 24, 2021 at 6:07 AM Yunsheng Lin <yunshenglin0825@gmail.com> wrote:

>>>>

>>>> On Fri, Jul 23, 2021 at 09:08:00AM -0700, Alexander Duyck wrote:

>>>>> On Fri, Jul 23, 2021 at 4:12 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:

>>>>>>

>>>>>> On 2021/7/22 23:18, Alexander Duyck wrote:

>>>>>>>>>

> 

> <snip>

> 

>>>>>

>>>>> Rather than trying to reuse the devices page pool it might make more

>>>>> sense to see if you couldn't have TCP just use some sort of circular

>>>>> buffer of memory that is directly mapped for the device that it is

>>>>> going to be transmitting to. Essentially what you would be doing is

>>>>> creating a pre-mapped page and would need to communicate that the

>>>>> memory is already mapped for the device you want to send it to so that

>>>>> it could skip that step.

>>>>

>>>> IIUC sk_page_frag_refill() is already doing a similar reusing as the

>>>> rx reusing implemented in most driver except for the not pre-mapping

>>>> part.

>>>>

>>>> And it seems that even if we pre-map the page and communicate that the

>>>> memory is already mapped to the driver, it is likely that we will not

>>>> be able to reuse the page when the circular buffer is not big enough

>>>> or tx completion/tcp ack is not happening quickly enough, which might

>>>> means unmapping/deallocating old circular buffer and allocating/mapping

>>>> new circular buffer.

>>>>

>>>> Using page pool we might be able to alleviate the above problem as it

>>>> does for rx?

>>>

>>> I would say that instead of looking at going straight for the page

>>> pool it might make more sense to look at seeing if we can coalesce the

>>> DMA mapping of the pages first at the socket layer rather than trying

>>> to introduce the overhead for the page pool. In the case of sockets we

>>> already have the destructors that are called when the memory is freed,

>>> so instead of making sockets use page pool it might make more sense to

>>> extend the socket buffer allocation/freeing to incorporate bulk

>>> mapping and unmapping of pages to optimize the socket Tx path in the

>>> 32K page case.

>>

>> I was able to enable tx recycling prototyping based on page pool to

>> run some performance test, the performance improvement is about +20%

>> (30Gbit -> 38Gbit) for single thread iperf tcp flow when IOMMU is in

>> strict mode. And CPU usage descreases about 10% for four threads iperf

>> tcp flow for line speed of 100Gbit when IOMMU is in strict mode.

> 

> That isn't surprising given that for most devices the IOMMU will be

> called per frag which can add a fair bit of overhead.

> 

>> Looking at the prototyping code, I am agreed that it is a bit controversial

>> to use the page pool for tx as the page pool is assuming NAPI polling

>> protection for allocation side.

>>

>> So I will take a deeper look about your suggestion above to see how to

>> implement it.

>>

>> Also, I am assuming the "destructors" means tcp_wfree() for TCP, right?

>> It seems tcp_wfree() is mainly used to do memory accounting and free

>> "struct sock" if necessary.

> 

> Yes, that is what I was thinking. If we had some way to add something

> like an argument or way to push the information about where the skbs

> are being freed back to the socket the socket could then be looking at

> pre-mapping the pages for the device if we assume a 1:1 mapping from

> the socket to the device.

> 

>> I am not so familiar with socket layer to understand how the "destructors"

>> will be helpful here, any detailed idea how to use "destructors" here?

> 

> The basic idea is the destructors are called when the skb is orphaned

> or freed. So it might be a good spot to put in any logic to free pages

> from your special pool. The only thing you would need to sort out is

> making certain to bump reference counts appropriately if the skb is

> cloned and the destructor is copied.


It seems the destructor is not copied when a skb is cloned, see:
https://elixir.bootlin.com/linux/latest/source/net/core/skbuff.c#L1050

For IPv4 TCP, tcp_write_xmit() calls __tcp_transmit_skb() to send the
new cloned skb using ip_queue_xmit(), and the original skb is kept in
sk->tcp_rtx_queue to wait for the ack packet. The destructor is assigned
to the new cloned skb in __tcp_transmit_skb(), and when destructor is
called in the tx completion process, it seems the frag page might be
still used by the retransmiting process, which means it is better not to
unmap or recycle that frag page in skb->destructor?

Also I tried to implement a frag pool to replace the page pool, but it
seems the feature needed for frag pool is already implemented by the
page pool, so implementing a new frag pool does not make sense.

For Rx, we have 1 : 1 relation between struct napi_struct instance and
struct page_pool instance, It seems we have the below options if the
recycling pool makes sense for Tx too:
1. 1 : 1 relation between struct net_device instance and struct page_pool
   instance.
2. 1 : 1 relation between struct napi_struct instance and struct page_pool
   instance.
3. 1 : 1 relation between struct sock instance and struct page_pool instance.

For me, the option 2 seems to make more sense to me if we can reuse the same
page pool for both Tx and Rx.

As the where or when to "bump reference counts appropriately", it seems the
__skb_frag_ref() might be a good spot to increment frag count if the frag is
copied, and the bit 0 for "struct page" ptr in frag->bv_page is reversed as
the lower 12 bits for dma address, so we can use bit 0 in frag->bv_page to
indicate the corresponding page is from a page pool or not? If the page is
from a page pool, then __skb_frag_ref() can do a atomic_inc(pp_frag_count)
instead of get_page(). This also might means that skb->pp_recycle is only used
to indicate if the head data page is from page pool or not when the bit 0
of frag->bv_page can be used to indicate if the corresponding frag page is
from page pool or not.

And doing atomic_inc(pp_frag_count) in __skb_frag_ref() seems to match the
semantics of "recycle after all users of page is done with the page", at least
for most users in the netstack?

> .

>
diff mbox series

Patch

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index d33d97c..3e339dd 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -103,11 +103,19 @@  struct page {
 			unsigned long pp_magic;
 			struct page_pool *pp;
 			unsigned long _pp_mapping_pad;
-			/**
-			 * @dma_addr: might require a 64-bit value on
-			 * 32-bit architectures.
-			 */
-			unsigned long dma_addr[2];
+			unsigned long dma_addr;
+			union {
+				/**
+				 * dma_addr_upper: might require a 64-bit
+				 * value on 32-bit architectures.
+				 */
+				unsigned long dma_addr_upper;
+				/**
+				 * For frag page support, not supported in
+				 * 32-bit architectures with 64-bit DMA.
+				 */
+				atomic_long_t pp_frag_count;
+			};
 		};
 		struct {	/* slab, slob and slub */
 			union {
diff --git a/include/net/page_pool.h b/include/net/page_pool.h
index 8d7744d..517d89f 100644
--- a/include/net/page_pool.h
+++ b/include/net/page_pool.h
@@ -45,7 +45,10 @@ 
 					* Please note DMA-sync-for-CPU is still
 					* device driver responsibility
 					*/
-#define PP_FLAG_ALL		(PP_FLAG_DMA_MAP | PP_FLAG_DMA_SYNC_DEV)
+#define PP_FLAG_PAGE_FRAG	BIT(2) /* for page frag feature */
+#define PP_FLAG_ALL		(PP_FLAG_DMA_MAP |\
+				 PP_FLAG_DMA_SYNC_DEV |\
+				 PP_FLAG_PAGE_FRAG)
 
 /*
  * Fast allocation side cache array/stack
@@ -198,19 +201,43 @@  static inline void page_pool_recycle_direct(struct page_pool *pool,
 	page_pool_put_full_page(pool, page, true);
 }
 
+#define PAGE_POOL_DMA_USE_PP_FRAG_COUNT	\
+		(sizeof(dma_addr_t) > sizeof(unsigned long))
+
 static inline dma_addr_t page_pool_get_dma_addr(struct page *page)
 {
-	dma_addr_t ret = page->dma_addr[0];
-	if (sizeof(dma_addr_t) > sizeof(unsigned long))
-		ret |= (dma_addr_t)page->dma_addr[1] << 16 << 16;
+	dma_addr_t ret = page->dma_addr;
+
+	if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT)
+		ret |= (dma_addr_t)page->dma_addr_upper << 16 << 16;
+
 	return ret;
 }
 
 static inline void page_pool_set_dma_addr(struct page *page, dma_addr_t addr)
 {
-	page->dma_addr[0] = addr;
-	if (sizeof(dma_addr_t) > sizeof(unsigned long))
-		page->dma_addr[1] = upper_32_bits(addr);
+	page->dma_addr = addr;
+	if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT)
+		page->dma_addr_upper = upper_32_bits(addr);
+}
+
+static inline void page_pool_set_frag_count(struct page *page, long nr)
+{
+	atomic_long_set(&page->pp_frag_count, nr);
+}
+
+static inline long page_pool_atomic_sub_frag_count_return(struct page *page,
+							  long nr)
+{
+	long frag_count = atomic_long_read(&page->pp_frag_count);
+	long ret;
+
+	if (frag_count == nr)
+		return 0;
+
+	ret = atomic_long_sub_return(nr, &page->pp_frag_count);
+	WARN_ON(ret < 0);
+	return ret;
 }
 
 static inline bool is_page_pool_compiled_in(void)
diff --git a/net/core/page_pool.c b/net/core/page_pool.c
index 78838c6..68fab94 100644
--- a/net/core/page_pool.c
+++ b/net/core/page_pool.c
@@ -67,6 +67,10 @@  static int page_pool_init(struct page_pool *pool,
 		 */
 	}
 
+	if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT &&
+	    pool->p.flags & PP_FLAG_PAGE_FRAG)
+		return -EINVAL;
+
 	if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
 		return -ENOMEM;