The code piece which would attach a frag to &xdp_buff is almost
identical across the drivers supporting XDP multi-buffer on Rx.
Make it a generic elegant onelner.
Also, I see lots of drivers calculating frags_truesize as
`xdp->frame_sz * nr_frags`. I can't say this is fully correct, since
frags might be backed by chunks of different sizes, especially with
stuff like the header split. Even page_pool_alloc() can give you two
different truesizes on two subsequent requests to allocate the same
buffer size. Add a field to &skb_shared_info (unionized as there's no
free slot currently on x6_64) to track the "true" truesize. It can be
used later when updating an skb.

Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
---
 include/linux/skbuff.h | 16 ++++++--
 include/net/xdp.h      | 90 +++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 101 insertions(+), 5 deletions(-)

diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index c867df5b1051..6ec78c1598fe 100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -607,11 +607,19 @@ struct skb_shared_info {
 	 * Warning : all fields before dataref are cleared in __alloc_skb()
 	 */
 	atomic_t	dataref;
-	unsigned int	xdp_frags_size;
 
-	/* Intermediate layers must ensure that destructor_arg
-	 * remains valid until skb destructor */
-	void *		destructor_arg;
+	union {
+		struct {
+			u32		xdp_frags_size;
+			u32		xdp_frags_truesize;
+		};
+
+		/*
+		 * Intermediate layers must ensure that destructor_arg
+		 * remains valid until skb destructor.
+		 */
+		void		*destructor_arg;
+	};
 
 	/* must be last field, see pskb_expand_head() */
 	skb_frag_t	frags[MAX_SKB_FRAGS];
diff --git a/include/net/xdp.h b/include/net/xdp.h
index c4b408d22669..19d2b283b845 100644
--- a/include/net/xdp.h
+++ b/include/net/xdp.h
@@ -167,6 +167,88 @@ xdp_get_buff_len(const struct xdp_buff *xdp)
 	return len;
 }
 
+/**
+ * __xdp_buff_add_frag - attach a frag to an &xdp_buff
+ * @xdp: XDP buffer to attach the frag to
+ * @page: page containing the frag
+ * @offset: page offset at which the frag starts
+ * @size: size of the frag
+ * @truesize: truesize (page / page frag size) of the frag
+ * @try_coalesce: whether to try coalescing the frags
+ *
+ * Attach a frag to an XDP buffer. If it currently has no frags attached,
+ * initialize the related fields, otherwise check that the frag number
+ * didn't reach the limit of ``MAX_SKB_FRAGS``. If possible, try coalescing
+ * the frag with the previous one.
+ * The function doesn't check/update the pfmemalloc bit. Please use the
+ * non-underscored wrapper in drivers.
+ *
+ * Return: true on success, false if there's no space for the frag in
+ * the shared info struct.
+ */
+static inline bool __xdp_buff_add_frag(struct xdp_buff *xdp, struct page *page,
+				       u32 offset, u32 size, u32 truesize,
+				       bool try_coalesce)
+{
+	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
+	skb_frag_t *prev;
+	u32 nr_frags;
+
+	if (!xdp_buff_has_frags(xdp)) {
+		xdp_buff_set_frags_flag(xdp);
+
+		nr_frags = 0;
+		sinfo->xdp_frags_size = 0;
+		sinfo->xdp_frags_truesize = 0;
+
+		goto fill;
+	}
+
+	nr_frags = sinfo->nr_frags;
+	if (unlikely(nr_frags == MAX_SKB_FRAGS))
+		return false;
+
+	prev = &sinfo->frags[nr_frags - 1];
+	if (try_coalesce && page == skb_frag_page(prev) &&
+	    offset == skb_frag_off(prev) + skb_frag_size(prev))
+		skb_frag_size_add(prev, size);
+	else
+fill:
+		__skb_fill_page_desc_noacc(sinfo, nr_frags++, page,
+					   offset, size);
+
+	sinfo->nr_frags = nr_frags;
+	sinfo->xdp_frags_size += size;
+	sinfo->xdp_frags_truesize += truesize;
+
+	return true;
+}
+
+/**
+ * xdp_buff_add_frag - attach a frag to an &xdp_buff
+ * @xdp: XDP buffer to attach the frag to
+ * @page: page containing the frag
+ * @offset: page offset at which the frag starts
+ * @size: size of the frag
+ * @truesize: truesize (page / page frag size) of the frag
+ *
+ * Version of __xdp_buff_add_frag() which takes care of the pfmemalloc bit.
+ *
+ * Return: true on success, false if there's no space for the frag in
+ * the shared info struct.
+ */
+static inline bool xdp_buff_add_frag(struct xdp_buff *xdp, struct page *page,
+				     u32 offset, u32 size, u32 truesize)
+{
+	if (!__xdp_buff_add_frag(xdp, page, offset, size, truesize, true))
+		return false;
+
+	if (unlikely(page_is_pfmemalloc(page)))
+		xdp_buff_set_frag_pfmemalloc(xdp);
+
+	return true;
+}
+
 struct xdp_frame {
 	void *data;
 	u32 len;
@@ -230,7 +312,13 @@ xdp_update_skb_shared_info(struct sk_buff *skb, u8 nr_frags,
 			   unsigned int size, unsigned int truesize,
 			   bool pfmemalloc)
 {
-	skb_shinfo(skb)->nr_frags = nr_frags;
+	struct skb_shared_info *sinfo = skb_shinfo(skb);
+
+	sinfo->nr_frags = nr_frags;
+	/* ``destructor_arg`` is unionized with ``xdp_frags_{,true}size``,
+	 * reset it after that these fields aren't used anymore.
+	 */
+	sinfo->destructor_arg = NULL;
 
 	skb->len += size;
 	skb->data_len += size;
-- 
2.46.2

On Tue, Oct 15, 2024 at 04:53:43PM +0200, Alexander Lobakin wrote:
> The code piece which would attach a frag to &xdp_buff is almost
> identical across the drivers supporting XDP multi-buffer on Rx.
> Make it a generic elegant onelner.

oneliner

> Also, I see lots of drivers calculating frags_truesize as
> `xdp->frame_sz * nr_frags`. I can't say this is fully correct, since
> frags might be backed by chunks of different sizes, especially with
> stuff like the header split. Even page_pool_alloc() can give you two
> different truesizes on two subsequent requests to allocate the same
> buffer size. Add a field to &skb_shared_info (unionized as there's no
> free slot currently on x6_64) to track the "true" truesize. It can be

x86_64

> used later when updating an skb.

I also agree that xdp->frame_sz * nr_frags for truesize might be an
over-assumption.

Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>

two small nits/questions below that might be ignored.

> 
> Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
> ---
>  include/linux/skbuff.h | 16 ++++++--
>  include/net/xdp.h      | 90 +++++++++++++++++++++++++++++++++++++++++-
>  2 files changed, 101 insertions(+), 5 deletions(-)
> 
> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
> index c867df5b1051..6ec78c1598fe 100644
> --- a/include/linux/skbuff.h
> +++ b/include/linux/skbuff.h
> @@ -607,11 +607,19 @@ struct skb_shared_info {
>  	 * Warning : all fields before dataref are cleared in __alloc_skb()
>  	 */
>  	atomic_t	dataref;
> -	unsigned int	xdp_frags_size;
>  
> -	/* Intermediate layers must ensure that destructor_arg
> -	 * remains valid until skb destructor */
> -	void *		destructor_arg;
> +	union {
> +		struct {
> +			u32		xdp_frags_size;
> +			u32		xdp_frags_truesize;
> +		};
> +
> +		/*
> +		 * Intermediate layers must ensure that destructor_arg
> +		 * remains valid until skb destructor.
> +		 */
> +		void		*destructor_arg;
> +	};
>  
>  	/* must be last field, see pskb_expand_head() */
>  	skb_frag_t	frags[MAX_SKB_FRAGS];
> diff --git a/include/net/xdp.h b/include/net/xdp.h
> index c4b408d22669..19d2b283b845 100644
> --- a/include/net/xdp.h
> +++ b/include/net/xdp.h
> @@ -167,6 +167,88 @@ xdp_get_buff_len(const struct xdp_buff *xdp)
>  	return len;
>  }
>  
> +/**
> + * __xdp_buff_add_frag - attach a frag to an &xdp_buff
> + * @xdp: XDP buffer to attach the frag to
> + * @page: page containing the frag
> + * @offset: page offset at which the frag starts
> + * @size: size of the frag
> + * @truesize: truesize (page / page frag size) of the frag
> + * @try_coalesce: whether to try coalescing the frags
> + *
> + * Attach a frag to an XDP buffer. If it currently has no frags attached,
> + * initialize the related fields, otherwise check that the frag number
> + * didn't reach the limit of ``MAX_SKB_FRAGS``. If possible, try coalescing
> + * the frag with the previous one.
> + * The function doesn't check/update the pfmemalloc bit. Please use the
> + * non-underscored wrapper in drivers.
> + *
> + * Return: true on success, false if there's no space for the frag in
> + * the shared info struct.
> + */
> +static inline bool __xdp_buff_add_frag(struct xdp_buff *xdp, struct page *page,
> +				       u32 offset, u32 size, u32 truesize,
> +				       bool try_coalesce)
> +{
> +	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
> +	skb_frag_t *prev;
> +	u32 nr_frags;
> +
> +	if (!xdp_buff_has_frags(xdp)) {
> +		xdp_buff_set_frags_flag(xdp);
> +
> +		nr_frags = 0;
> +		sinfo->xdp_frags_size = 0;
> +		sinfo->xdp_frags_truesize = 0;
> +
> +		goto fill;
> +	}
> +
> +	nr_frags = sinfo->nr_frags;
> +	if (unlikely(nr_frags == MAX_SKB_FRAGS))
> +		return false;
> +
> +	prev = &sinfo->frags[nr_frags - 1];
> +	if (try_coalesce && page == skb_frag_page(prev) &&
> +	    offset == skb_frag_off(prev) + skb_frag_size(prev))
> +		skb_frag_size_add(prev, size);
> +	else
> +fill:
> +		__skb_fill_page_desc_noacc(sinfo, nr_frags++, page,
> +					   offset, size);
> +
> +	sinfo->nr_frags = nr_frags;

is it really necessary to work on local nr_frags instead of directly
update it from sinfo?

> +	sinfo->xdp_frags_size += size;
> +	sinfo->xdp_frags_truesize += truesize;
> +
> +	return true;
> +}
> +
> +/**
> + * xdp_buff_add_frag - attach a frag to an &xdp_buff
> + * @xdp: XDP buffer to attach the frag to
> + * @page: page containing the frag
> + * @offset: page offset at which the frag starts
> + * @size: size of the frag
> + * @truesize: truesize (page / page frag size) of the frag
> + *
> + * Version of __xdp_buff_add_frag() which takes care of the pfmemalloc bit.
> + *
> + * Return: true on success, false if there's no space for the frag in
> + * the shared info struct.
> + */
> +static inline bool xdp_buff_add_frag(struct xdp_buff *xdp, struct page *page,
> +				     u32 offset, u32 size, u32 truesize)
> +{
> +	if (!__xdp_buff_add_frag(xdp, page, offset, size, truesize, true))
> +		return false;
> +
> +	if (unlikely(page_is_pfmemalloc(page)))
> +		xdp_buff_set_frag_pfmemalloc(xdp);
> +
> +	return true;
> +}
> +
>  struct xdp_frame {
>  	void *data;
>  	u32 len;
> @@ -230,7 +312,13 @@ xdp_update_skb_shared_info(struct sk_buff *skb, u8 nr_frags,
>  			   unsigned int size, unsigned int truesize,
>  			   bool pfmemalloc)
>  {
> -	skb_shinfo(skb)->nr_frags = nr_frags;
> +	struct skb_shared_info *sinfo = skb_shinfo(skb);
> +
> +	sinfo->nr_frags = nr_frags;
> +	/* ``destructor_arg`` is unionized with ``xdp_frags_{,true}size``,
> +	 * reset it after that these fields aren't used anymore.
> +	 */
> +	sinfo->destructor_arg = NULL;

wouldn't clearing size and truesize from union be more obvious?
OTOH it's one write vs two :)

>  
>  	skb->len += size;
>  	skb->data_len += size;
> -- 
> 2.46.2
>

From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Date: Thu, 17 Oct 2024 14:26:48 +0200

> On Tue, Oct 15, 2024 at 04:53:43PM +0200, Alexander Lobakin wrote:
>> The code piece which would attach a frag to &xdp_buff is almost
>> identical across the drivers supporting XDP multi-buffer on Rx.
>> Make it a generic elegant onelner.
> 
> oneliner
> 
>> Also, I see lots of drivers calculating frags_truesize as
>> `xdp->frame_sz * nr_frags`. I can't say this is fully correct, since
>> frags might be backed by chunks of different sizes, especially with
>> stuff like the header split. Even page_pool_alloc() can give you two
>> different truesizes on two subsequent requests to allocate the same
>> buffer size. Add a field to &skb_shared_info (unionized as there's no
>> free slot currently on x6_64) to track the "true" truesize. It can be
> 
> x86_64

What a shame from these two typos >_<

> 
>> used later when updating an skb.

[...]

>> +
>> +	prev = &sinfo->frags[nr_frags - 1];
>> +	if (try_coalesce && page == skb_frag_page(prev) &&
>> +	    offset == skb_frag_off(prev) + skb_frag_size(prev))
>> +		skb_frag_size_add(prev, size);
>> +	else
>> +fill:
>> +		__skb_fill_page_desc_noacc(sinfo, nr_frags++, page,
>> +					   offset, size);
>> +
>> +	sinfo->nr_frags = nr_frags;
> 
> is it really necessary to work on local nr_frags instead of directly
> update it from sinfo?

I think you remember the difference when you started to work on ntu and
ntc locally instead of accessing the ring struct all the time? :>

> 
>> +	sinfo->xdp_frags_size += size;
>> +	sinfo->xdp_frags_truesize += truesize;
>> +
>> +	return true;
>> +}

[...]

>> @@ -230,7 +312,13 @@ xdp_update_skb_shared_info(struct sk_buff *skb, u8 nr_frags,
>>  			   unsigned int size, unsigned int truesize,
>>  			   bool pfmemalloc)
>>  {
>> -	skb_shinfo(skb)->nr_frags = nr_frags;
>> +	struct skb_shared_info *sinfo = skb_shinfo(skb);
>> +
>> +	sinfo->nr_frags = nr_frags;
>> +	/* ``destructor_arg`` is unionized with ``xdp_frags_{,true}size``,
>> +	 * reset it after that these fields aren't used anymore.
>> +	 */
>> +	sinfo->destructor_arg = NULL;
> 
> wouldn't clearing size and truesize from union be more obvious?

But here we actually need to reset the destructor arg pointer.
size/truesize are not needed at this point anymore, but the arg can be
used/tested later, so I thought clearing it here is more clear to the
readers?

> OTOH it's one write vs two :)

Sometimes the compiler can optimize two subsequent writes (e.g. to addr
and addr + 4) into one bigger, but I wouldn't rely on it (that's why in
patch #18 I intensively use casts to u64).

> 
>>  
>>  	skb->len += size;
>>  	skb->data_len += size;
>> -- 
>> 2.46.2

Thanks,
Olek

On Mon, Oct 21, 2024 at 04:10:30PM +0200, Alexander Lobakin wrote:
> From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
> Date: Thu, 17 Oct 2024 14:26:48 +0200
> 
> > On Tue, Oct 15, 2024 at 04:53:43PM +0200, Alexander Lobakin wrote:
> >> The code piece which would attach a frag to &xdp_buff is almost
> >> identical across the drivers supporting XDP multi-buffer on Rx.
> >> Make it a generic elegant onelner.
> > 
> > oneliner
> > 
> >> Also, I see lots of drivers calculating frags_truesize as
> >> `xdp->frame_sz * nr_frags`. I can't say this is fully correct, since
> >> frags might be backed by chunks of different sizes, especially with
> >> stuff like the header split. Even page_pool_alloc() can give you two
> >> different truesizes on two subsequent requests to allocate the same
> >> buffer size. Add a field to &skb_shared_info (unionized as there's no
> >> free slot currently on x6_64) to track the "true" truesize. It can be
> > 
> > x86_64
> 
> What a shame from these two typos >_<
> 
> > 
> >> used later when updating an skb.
> 
> [...]
> 
> >> +
> >> +	prev = &sinfo->frags[nr_frags - 1];
> >> +	if (try_coalesce && page == skb_frag_page(prev) &&
> >> +	    offset == skb_frag_off(prev) + skb_frag_size(prev))
> >> +		skb_frag_size_add(prev, size);
> >> +	else
> >> +fill:
> >> +		__skb_fill_page_desc_noacc(sinfo, nr_frags++, page,
> >> +					   offset, size);
> >> +
> >> +	sinfo->nr_frags = nr_frags;
> > 
> > is it really necessary to work on local nr_frags instead of directly
> > update it from sinfo?
> 
> I think you remember the difference when you started to work on ntu and
> ntc locally instead of accessing the ring struct all the time? :>

Right, although impact there was a bit bigger.
Typos are minor, so:

Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>

unless others have some opinion on union being introduced here.

> 
> > 
> >> +	sinfo->xdp_frags_size += size;
> >> +	sinfo->xdp_frags_truesize += truesize;
> >> +
> >> +	return true;
> >> +}
> 
> [...]
> 
> >> @@ -230,7 +312,13 @@ xdp_update_skb_shared_info(struct sk_buff *skb, u8 nr_frags,
> >>  			   unsigned int size, unsigned int truesize,
> >>  			   bool pfmemalloc)
> >>  {
> >> -	skb_shinfo(skb)->nr_frags = nr_frags;
> >> +	struct skb_shared_info *sinfo = skb_shinfo(skb);
> >> +
> >> +	sinfo->nr_frags = nr_frags;
> >> +	/* ``destructor_arg`` is unionized with ``xdp_frags_{,true}size``,
> >> +	 * reset it after that these fields aren't used anymore.
> >> +	 */
> >> +	sinfo->destructor_arg = NULL;
> > 
> > wouldn't clearing size and truesize from union be more obvious?
> 
> But here we actually need to reset the destructor arg pointer.
> size/truesize are not needed at this point anymore, but the arg can be
> used/tested later, so I thought clearing it here is more clear to the
> readers?
> 
> > OTOH it's one write vs two :)
> 
> Sometimes the compiler can optimize two subsequent writes (e.g. to addr
> and addr + 4) into one bigger, but I wouldn't rely on it (that's why in
> patch #18 I intensively use casts to u64).
> 
> > 
> >>  
> >>  	skb->len += size;
> >>  	skb->data_len += size;
> >> -- 
> >> 2.46.2
> 
> Thanks,
> Olek