* [PATCH net-next v2 0/2] Some minor optimization for page pool
@ 2021-08-20 6:56 Yunsheng Lin
2021-08-20 6:56 ` [PATCH net-next v2 1/2] page_pool: use relaxed atomic for release side accounting Yunsheng Lin
2021-08-20 6:56 ` [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping Yunsheng Lin
0 siblings, 2 replies; 9+ messages in thread
From: Yunsheng Lin @ 2021-08-20 6:56 UTC (permalink / raw)
To: davem, kuba; +Cc: hawk, ilias.apalodimas, netdev, linux-kernel, hkallweit1
Patch 1: Use relaxed atomic for release side accounting
Patch 2: Minor optimize for page_pool_dma_map() function
V2: Remove unnecessary unliky() mark as pointed out by
Heiner.
Yunsheng Lin (2):
page_pool: use relaxed atomic for release side accounting
page_pool: optimize the cpu sync operation when DMA mapping
net/core/page_pool.c | 11 ++++++-----
1 file changed, 6 insertions(+), 5 deletions(-)
--
2.7.4
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH net-next v2 1/2] page_pool: use relaxed atomic for release side accounting
2021-08-20 6:56 [PATCH net-next v2 0/2] Some minor optimization for page pool Yunsheng Lin
@ 2021-08-20 6:56 ` Yunsheng Lin
2021-08-20 7:12 ` Jesper Dangaard Brouer
2021-08-20 6:56 ` [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping Yunsheng Lin
1 sibling, 1 reply; 9+ messages in thread
From: Yunsheng Lin @ 2021-08-20 6:56 UTC (permalink / raw)
To: davem, kuba; +Cc: hawk, ilias.apalodimas, netdev, linux-kernel, hkallweit1
There is no need to synchronize the account updating, so
use the relaxed atomic to avoid some memory barrier in the
data path.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
---
net/core/page_pool.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/net/core/page_pool.c b/net/core/page_pool.c
index e140905..1a69784 100644
--- a/net/core/page_pool.c
+++ b/net/core/page_pool.c
@@ -370,7 +370,7 @@ void page_pool_release_page(struct page_pool *pool, struct page *page)
/* This may be the last page returned, releasing the pool, so
* it is not safe to reference pool afterwards.
*/
- count = atomic_inc_return(&pool->pages_state_release_cnt);
+ count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
trace_page_pool_state_release(pool, page, count);
}
EXPORT_SYMBOL(page_pool_release_page);
--
2.7.4
^ permalink raw reply related [flat|nested] 9+ messages in thread
* [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping
2021-08-20 6:56 [PATCH net-next v2 0/2] Some minor optimization for page pool Yunsheng Lin
2021-08-20 6:56 ` [PATCH net-next v2 1/2] page_pool: use relaxed atomic for release side accounting Yunsheng Lin
@ 2021-08-20 6:56 ` Yunsheng Lin
2021-08-20 9:39 ` Ilias Apalodimas
1 sibling, 1 reply; 9+ messages in thread
From: Yunsheng Lin @ 2021-08-20 6:56 UTC (permalink / raw)
To: davem, kuba; +Cc: hawk, ilias.apalodimas, netdev, linux-kernel, hkallweit1
If the DMA_ATTR_SKIP_CPU_SYNC is not set, cpu syncing is
also done in dma_map_page_attrs(), so set the attrs according
to pool->p.flags to avoid calling cpu sync function again.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
---
net/core/page_pool.c | 9 +++++----
1 file changed, 5 insertions(+), 4 deletions(-)
diff --git a/net/core/page_pool.c b/net/core/page_pool.c
index 1a69784..3df5554 100644
--- a/net/core/page_pool.c
+++ b/net/core/page_pool.c
@@ -191,8 +191,12 @@ static void page_pool_dma_sync_for_device(struct page_pool *pool,
static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
{
+ unsigned long attrs = DMA_ATTR_SKIP_CPU_SYNC;
dma_addr_t dma;
+ if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
+ attrs = 0;
+
/* Setup DMA mapping: use 'struct page' area for storing DMA-addr
* since dma_addr_t can be either 32 or 64 bits and does not always fit
* into page private data (i.e 32bit cpu with 64bit DMA caps)
@@ -200,15 +204,12 @@ static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
*/
dma = dma_map_page_attrs(pool->p.dev, page, 0,
(PAGE_SIZE << pool->p.order),
- pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
+ pool->p.dma_dir, attrs);
if (dma_mapping_error(pool->p.dev, dma))
return false;
page_pool_set_dma_addr(page, dma);
- if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
- page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
-
return true;
}
--
2.7.4
^ permalink raw reply related [flat|nested] 9+ messages in thread
* Re: [PATCH net-next v2 1/2] page_pool: use relaxed atomic for release side accounting
2021-08-20 6:56 ` [PATCH net-next v2 1/2] page_pool: use relaxed atomic for release side accounting Yunsheng Lin
@ 2021-08-20 7:12 ` Jesper Dangaard Brouer
0 siblings, 0 replies; 9+ messages in thread
From: Jesper Dangaard Brouer @ 2021-08-20 7:12 UTC (permalink / raw)
To: Yunsheng Lin, davem, kuba
Cc: brouer, hawk, ilias.apalodimas, netdev, linux-kernel, hkallweit1
On 20/08/2021 08.56, Yunsheng Lin wrote:
> There is no need to synchronize the account updating, so
> use the relaxed atomic to avoid some memory barrier in the
> data path.
>
> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
LGTM
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
> ---
> net/core/page_pool.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/net/core/page_pool.c b/net/core/page_pool.c
> index e140905..1a69784 100644
> --- a/net/core/page_pool.c
> +++ b/net/core/page_pool.c
> @@ -370,7 +370,7 @@ void page_pool_release_page(struct page_pool *pool, struct page *page)
> /* This may be the last page returned, releasing the pool, so
> * it is not safe to reference pool afterwards.
> */
> - count = atomic_inc_return(&pool->pages_state_release_cnt);
> + count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
> trace_page_pool_state_release(pool, page, count);
> }
> EXPORT_SYMBOL(page_pool_release_page);
>
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping
2021-08-20 6:56 ` [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping Yunsheng Lin
@ 2021-08-20 9:39 ` Ilias Apalodimas
2021-08-23 3:56 ` Yunsheng Lin
0 siblings, 1 reply; 9+ messages in thread
From: Ilias Apalodimas @ 2021-08-20 9:39 UTC (permalink / raw)
To: Yunsheng Lin; +Cc: davem, kuba, hawk, netdev, linux-kernel, hkallweit1
On Fri, Aug 20, 2021 at 02:56:51PM +0800, Yunsheng Lin wrote:
> If the DMA_ATTR_SKIP_CPU_SYNC is not set, cpu syncing is
> also done in dma_map_page_attrs(), so set the attrs according
> to pool->p.flags to avoid calling cpu sync function again.
Isn't DMA_ATTR_SKIP_CPU_SYNC checked within dma_map_page_attrs() anyway?
Regards
/Ilias
>
> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
> ---
> net/core/page_pool.c | 9 +++++----
> 1 file changed, 5 insertions(+), 4 deletions(-)
>
> diff --git a/net/core/page_pool.c b/net/core/page_pool.c
> index 1a69784..3df5554 100644
> --- a/net/core/page_pool.c
> +++ b/net/core/page_pool.c
> @@ -191,8 +191,12 @@ static void page_pool_dma_sync_for_device(struct page_pool *pool,
>
> static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
> {
> + unsigned long attrs = DMA_ATTR_SKIP_CPU_SYNC;
> dma_addr_t dma;
>
> + if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
> + attrs = 0;
> +
> /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
> * since dma_addr_t can be either 32 or 64 bits and does not always fit
> * into page private data (i.e 32bit cpu with 64bit DMA caps)
> @@ -200,15 +204,12 @@ static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
> */
> dma = dma_map_page_attrs(pool->p.dev, page, 0,
> (PAGE_SIZE << pool->p.order),
> - pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
> + pool->p.dma_dir, attrs);
> if (dma_mapping_error(pool->p.dev, dma))
> return false;
>
> page_pool_set_dma_addr(page, dma);
>
> - if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
> - page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
> -
> return true;
> }
>
> --
> 2.7.4
>
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping
2021-08-20 9:39 ` Ilias Apalodimas
@ 2021-08-23 3:56 ` Yunsheng Lin
2021-08-23 12:42 ` Ilias Apalodimas
0 siblings, 1 reply; 9+ messages in thread
From: Yunsheng Lin @ 2021-08-23 3:56 UTC (permalink / raw)
To: Ilias Apalodimas; +Cc: davem, kuba, hawk, netdev, linux-kernel, hkallweit1
On 2021/8/20 17:39, Ilias Apalodimas wrote:
> On Fri, Aug 20, 2021 at 02:56:51PM +0800, Yunsheng Lin wrote:
>> If the DMA_ATTR_SKIP_CPU_SYNC is not set, cpu syncing is
>> also done in dma_map_page_attrs(), so set the attrs according
>> to pool->p.flags to avoid calling cpu sync function again.
>
> Isn't DMA_ATTR_SKIP_CPU_SYNC checked within dma_map_page_attrs() anyway?
Yes, the checking in dma_map_page_attrs() should save us from
calling dma_sync_single_for_device() again if we set the attrs
according to "pool->p.flags & PP_FLAG_DMA_SYNC_DEV".
As dma_sync_single_for_device() is EXPORT_SYMBOL()'ed, and
should be a no-op for dma coherent device, so there may be a
function calling overhead for dma coherent device, letting
dma_map_page_attrs() handling the sync seems to avoid the stack
pushing/poping overhead:
https://elixir.bootlin.com/linux/latest/source/kernel/dma/direct.h#L104
The one thing I am not sure about is that the pool->p.offset
and pool->p.max_len are used to decide the sync range before this
patch, while the sync range is the same as the map range when doing
the sync in dma_map_page_attrs().
I assumed the above is not a issue? only sync more than we need?
and it won't hurt the performance?
>
> Regards
> /Ilias
>>
>> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
>> ---
>> net/core/page_pool.c | 9 +++++----
>> 1 file changed, 5 insertions(+), 4 deletions(-)
>>
>> diff --git a/net/core/page_pool.c b/net/core/page_pool.c
>> index 1a69784..3df5554 100644
>> --- a/net/core/page_pool.c
>> +++ b/net/core/page_pool.c
>> @@ -191,8 +191,12 @@ static void page_pool_dma_sync_for_device(struct page_pool *pool,
>>
>> static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
>> {
>> + unsigned long attrs = DMA_ATTR_SKIP_CPU_SYNC;
>> dma_addr_t dma;
>>
>> + if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
>> + attrs = 0;
>> +
>> /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
>> * since dma_addr_t can be either 32 or 64 bits and does not always fit
>> * into page private data (i.e 32bit cpu with 64bit DMA caps)
>> @@ -200,15 +204,12 @@ static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
>> */
>> dma = dma_map_page_attrs(pool->p.dev, page, 0,
>> (PAGE_SIZE << pool->p.order),
>> - pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
>> + pool->p.dma_dir, attrs);
>> if (dma_mapping_error(pool->p.dev, dma))
>> return false;
>>
>> page_pool_set_dma_addr(page, dma);
>>
>> - if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
>> - page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
>> -
>> return true;
>> }
>>
>> --
>> 2.7.4
>>
> .
>
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping
2021-08-23 3:56 ` Yunsheng Lin
@ 2021-08-23 12:42 ` Ilias Apalodimas
2021-08-24 7:00 ` Yunsheng Lin
0 siblings, 1 reply; 9+ messages in thread
From: Ilias Apalodimas @ 2021-08-23 12:42 UTC (permalink / raw)
To: Yunsheng Lin; +Cc: davem, kuba, hawk, netdev, linux-kernel, hkallweit1
On Mon, Aug 23, 2021 at 11:56:48AM +0800, Yunsheng Lin wrote:
> On 2021/8/20 17:39, Ilias Apalodimas wrote:
> > On Fri, Aug 20, 2021 at 02:56:51PM +0800, Yunsheng Lin wrote:
> >> If the DMA_ATTR_SKIP_CPU_SYNC is not set, cpu syncing is
> >> also done in dma_map_page_attrs(), so set the attrs according
> >> to pool->p.flags to avoid calling cpu sync function again.
> >
> > Isn't DMA_ATTR_SKIP_CPU_SYNC checked within dma_map_page_attrs() anyway?
>
> Yes, the checking in dma_map_page_attrs() should save us from
> calling dma_sync_single_for_device() again if we set the attrs
> according to "pool->p.flags & PP_FLAG_DMA_SYNC_DEV".
But we aren't syncing anything right now when we allocate the pages since
this is called with DMA_ATTR_SKIP_CPU_SYNC. We are syncing the allocated
range on the end of the function, if the pool was created and was requested
to take care of the mappings for us.
>
> As dma_sync_single_for_device() is EXPORT_SYMBOL()'ed, and
> should be a no-op for dma coherent device, so there may be a
> function calling overhead for dma coherent device, letting
> dma_map_page_attrs() handling the sync seems to avoid the stack
> pushing/poping overhead:
>
> https://elixir.bootlin.com/linux/latest/source/kernel/dma/direct.h#L104
>
> The one thing I am not sure about is that the pool->p.offset
> and pool->p.max_len are used to decide the sync range before this
> patch, while the sync range is the same as the map range when doing
> the sync in dma_map_page_attrs().
I am not sure I am following here. We always sync the entire range as well
in the current code as the mapping function is called with max_len.
>
> I assumed the above is not a issue? only sync more than we need?
> and it won't hurt the performance?
We can sync more than we need, but if it's a non-coherent architecture,
there's a performance penalty.
Regards
/Ilias
>
> >
> > Regards
> > /Ilias
> >>
> >> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
> >> ---
> >> net/core/page_pool.c | 9 +++++----
> >> 1 file changed, 5 insertions(+), 4 deletions(-)
> >>
> >> diff --git a/net/core/page_pool.c b/net/core/page_pool.c
> >> index 1a69784..3df5554 100644
> >> --- a/net/core/page_pool.c
> >> +++ b/net/core/page_pool.c
> >> @@ -191,8 +191,12 @@ static void page_pool_dma_sync_for_device(struct page_pool *pool,
> >>
> >> static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
> >> {
> >> + unsigned long attrs = DMA_ATTR_SKIP_CPU_SYNC;
> >> dma_addr_t dma;
> >>
> >> + if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
> >> + attrs = 0;
> >> +
> >> /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
> >> * since dma_addr_t can be either 32 or 64 bits and does not always fit
> >> * into page private data (i.e 32bit cpu with 64bit DMA caps)
> >> @@ -200,15 +204,12 @@ static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
> >> */
> >> dma = dma_map_page_attrs(pool->p.dev, page, 0,
> >> (PAGE_SIZE << pool->p.order),
> >> - pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
> >> + pool->p.dma_dir, attrs);
> >> if (dma_mapping_error(pool->p.dev, dma))
> >> return false;
> >>
> >> page_pool_set_dma_addr(page, dma);
> >>
> >> - if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
> >> - page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
> >> -
> >> return true;
> >> }
> >>
> >> --
> >> 2.7.4
> >>
> > .
> >
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping
2021-08-23 12:42 ` Ilias Apalodimas
@ 2021-08-24 7:00 ` Yunsheng Lin
2021-08-24 9:04 ` Ilias Apalodimas
0 siblings, 1 reply; 9+ messages in thread
From: Yunsheng Lin @ 2021-08-24 7:00 UTC (permalink / raw)
To: Ilias Apalodimas; +Cc: davem, kuba, hawk, netdev, linux-kernel, hkallweit1
On 2021/8/23 20:42, Ilias Apalodimas wrote:
> On Mon, Aug 23, 2021 at 11:56:48AM +0800, Yunsheng Lin wrote:
>> On 2021/8/20 17:39, Ilias Apalodimas wrote:
>>> On Fri, Aug 20, 2021 at 02:56:51PM +0800, Yunsheng Lin wrote:
[..]
>>
>> https://elixir.bootlin.com/linux/latest/source/kernel/dma/direct.h#L104
>>
>> The one thing I am not sure about is that the pool->p.offset
>> and pool->p.max_len are used to decide the sync range before this
>> patch, while the sync range is the same as the map range when doing
>> the sync in dma_map_page_attrs().
>
> I am not sure I am following here. We always sync the entire range as well
> in the current code as the mapping function is called with max_len.
>
>>
>> I assumed the above is not a issue? only sync more than we need?
>> and it won't hurt the performance?
>
> We can sync more than we need, but if it's a non-coherent architecture,
> there's a performance penalty.
Since I do not have any performance data to prove if there is a
performance penalty for non-coherent architecture, I will drop it:)
>
> Regards
> /Ilias
>>
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping
2021-08-24 7:00 ` Yunsheng Lin
@ 2021-08-24 9:04 ` Ilias Apalodimas
0 siblings, 0 replies; 9+ messages in thread
From: Ilias Apalodimas @ 2021-08-24 9:04 UTC (permalink / raw)
To: Yunsheng Lin, Lorenzo Bianconi
Cc: David S. Miller, Jakub Kicinski, Jesper Dangaard Brouer,
Networking, open list, Heiner Kallweit
Hi Yunsheng,
+cc Lorenzo, which has done some tests on non-coherent platforms
On Tue, 24 Aug 2021 at 10:00, Yunsheng Lin <linyunsheng@huawei.com> wrote:
>
> On 2021/8/23 20:42, Ilias Apalodimas wrote:
> > On Mon, Aug 23, 2021 at 11:56:48AM +0800, Yunsheng Lin wrote:
> >> On 2021/8/20 17:39, Ilias Apalodimas wrote:
> >>> On Fri, Aug 20, 2021 at 02:56:51PM +0800, Yunsheng Lin wrote:
>
> [..]
> >>
> >> https://elixir.bootlin.com/linux/latest/source/kernel/dma/direct.h#L104
> >>
> >> The one thing I am not sure about is that the pool->p.offset
> >> and pool->p.max_len are used to decide the sync range before this
> >> patch, while the sync range is the same as the map range when doing
> >> the sync in dma_map_page_attrs().
> >
> > I am not sure I am following here. We always sync the entire range as well
> > in the current code as the mapping function is called with max_len.
> >
> >>
> >> I assumed the above is not a issue? only sync more than we need?
> >> and it won't hurt the performance?
> >
> > We can sync more than we need, but if it's a non-coherent architecture,
> > there's a performance penalty.
>
> Since I do not have any performance data to prove if there is a
> performance penalty for non-coherent architecture, I will drop it:)
I am pretty sure it does affect it. Unless I am missing something the
patch simply re-arranges calls to avoid calling dma_map_page_attrs()
right?
However since dma_map_page_attrs() won't do anything sync-related
since it's called with DMA_ATTR_SKIP_CPU_SYNC, I doubt calling it will
have any measurable difference. If there is, we should pick it up.
Regards
/Ilias
>
> >
> > Regards
> > /Ilias
> >>
^ permalink raw reply [flat|nested] 9+ messages in thread
end of thread, other threads:[~2021-08-24 9:05 UTC | newest]
Thread overview: 9+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-08-20 6:56 [PATCH net-next v2 0/2] Some minor optimization for page pool Yunsheng Lin
2021-08-20 6:56 ` [PATCH net-next v2 1/2] page_pool: use relaxed atomic for release side accounting Yunsheng Lin
2021-08-20 7:12 ` Jesper Dangaard Brouer
2021-08-20 6:56 ` [PATCH net-next v2 2/2] page_pool: optimize the cpu sync operation when DMA mapping Yunsheng Lin
2021-08-20 9:39 ` Ilias Apalodimas
2021-08-23 3:56 ` Yunsheng Lin
2021-08-23 12:42 ` Ilias Apalodimas
2021-08-24 7:00 ` Yunsheng Lin
2021-08-24 9:04 ` Ilias Apalodimas
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).