diff mbox series

[v4,1/9] mbuf: new function to generate raw Tx offload value

Message ID 20190329102726.27716-2-konstantin.ananyev@intel.com (mailing list archive)
State Superseded, archived
Delegated to: akhil goyal
Headers
Series Few small improvements for ipsec library |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/intel-Performance-Testing success Performance Testing PASS
ci/mellanox-Performance-Testing fail Performance Testing issues
ci/Intel-compilation success Compilation OK

Commit Message

Ananyev, Konstantin March 29, 2019, 10:27 a.m. UTC 
  Operations to set/update bit-fields often cause compilers
to generate suboptimal code.
To help avoid such situation for tx_offload fields:
introduce new enum for tx_offload bit-fields lengths and offsets,
and new function to generate raw tx_offload value.

Signed-off-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Acked-by: Akhil Goyal <akhil.goyal@nxp.com>
---
 lib/librte_mbuf/rte_mbuf.h | 79 ++++++++++++++++++++++++++++++++++----
 1 file changed, 72 insertions(+), 7 deletions(-)

Comments

Olivier Matz March 29, 2019, 12:54 p.m. UTC | #1 
Hi Konstantin,

On Fri, Mar 29, 2019 at 10:27:18AM +0000, Konstantin Ananyev wrote:
> Operations to set/update bit-fields often cause compilers
> to generate suboptimal code.
> To help avoid such situation for tx_offload fields:
> introduce new enum for tx_offload bit-fields lengths and offsets,
> and new function to generate raw tx_offload value.
> 
> Signed-off-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> Acked-by: Akhil Goyal <akhil.goyal@nxp.com>

I understand the need. Out of curiosity, do you have any performance
numbers to share?

Few cosmetic questions below.

> ---
>  lib/librte_mbuf/rte_mbuf.h | 79 ++++++++++++++++++++++++++++++++++----
>  1 file changed, 72 insertions(+), 7 deletions(-)
> 
> diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
> index d961ccaf6..0b197e8ce 100644
> --- a/lib/librte_mbuf/rte_mbuf.h
> +++ b/lib/librte_mbuf/rte_mbuf.h
> @@ -479,6 +479,31 @@ struct rte_mbuf_sched {
>  	uint16_t reserved;   /**< Reserved. */
>  }; /**< Hierarchical scheduler */
>  
> +/**
> + * enum for the tx_offload bit-fields lenghts and offsets.
> + * defines the layout of rte_mbuf tx_offload field.
> + */
> +enum {
> +	RTE_MBUF_L2_LEN_BITS = 7,
> +	RTE_MBUF_L3_LEN_BITS = 9,
> +	RTE_MBUF_L4_LEN_BITS = 8,
> +	RTE_MBUF_TSO_SEGSZ_BITS = 16,
> +	RTE_MBUF_OUTL3_LEN_BITS = 9,
> +	RTE_MBUF_OUTL2_LEN_BITS = 7,
> +	RTE_MBUF_L2_LEN_OFS = 0,
> +	RTE_MBUF_L3_LEN_OFS = RTE_MBUF_L2_LEN_OFS + RTE_MBUF_L2_LEN_BITS,
> +	RTE_MBUF_L4_LEN_OFS = RTE_MBUF_L3_LEN_OFS + RTE_MBUF_L3_LEN_BITS,
> +	RTE_MBUF_TSO_SEGSZ_OFS = RTE_MBUF_L4_LEN_OFS + RTE_MBUF_L4_LEN_BITS,
> +	RTE_MBUF_OUTL3_LEN_OFS =
> +		RTE_MBUF_TSO_SEGSZ_OFS + RTE_MBUF_TSO_SEGSZ_BITS,
> +	RTE_MBUF_OUTL2_LEN_OFS =
> +		RTE_MBUF_OUTL3_LEN_OFS + RTE_MBUF_OUTL3_LEN_BITS,
> +	RTE_MBUF_TXOFLD_UNUSED_OFS =
> +		RTE_MBUF_OUTL2_LEN_OFS + RTE_MBUF_OUTL2_LEN_BITS,
> +	RTE_MBUF_TXOFLD_UNUSED_BITS =
> +		sizeof(uint64_t) * CHAR_BIT - RTE_MBUF_TXOFLD_UNUSED_OFS,
> +};
> +

What is the advantage of defining an enum instead of #defines?

In any case, I wonder if it wouldn't be clearer to change the order like
this:

enum {
	RTE_MBUF_L2_LEN_OFS = 0,
	RTE_MBUF_L2_LEN_BITS = 7,
	RTE_MBUF_L3_LEN_OFS = RTE_MBUF_L2_LEN_OFS + RTE_MBUF_L2_LEN_BITS,
	RTE_MBUF_L3_LEN_BITS = 9,
	RTE_MBUF_L4_LEN_OFS = RTE_MBUF_L3_LEN_OFS + RTE_MBUF_L3_LEN_BITS,
	RTE_MBUF_L4_LEN_BITS = 8,
...


>  /**
>   * The generic rte_mbuf, containing a packet mbuf.
>   */
> @@ -640,19 +665,24 @@ struct rte_mbuf {
>  		uint64_t tx_offload;       /**< combined for easy fetch */
>  		__extension__
>  		struct {
> -			uint64_t l2_len:7;
> +			uint64_t l2_len:RTE_MBUF_L2_LEN_BITS;
>  			/**< L2 (MAC) Header Length for non-tunneling pkt.
>  			 * Outer_L4_len + ... + Inner_L2_len for tunneling pkt.
>  			 */
> -			uint64_t l3_len:9; /**< L3 (IP) Header Length. */
> -			uint64_t l4_len:8; /**< L4 (TCP/UDP) Header Length. */
> -			uint64_t tso_segsz:16; /**< TCP TSO segment size */
> +			uint64_t l3_len:RTE_MBUF_L3_LEN_BITS;
> +			/**< L3 (IP) Header Length. */
> +			uint64_t l4_len:RTE_MBUF_L4_LEN_BITS;
> +			/**< L4 (TCP/UDP) Header Length. */
> +			uint64_t tso_segsz:RTE_MBUF_TSO_SEGSZ_BITS;
> +			/**< TCP TSO segment size */
>  
>  			/* fields for TX offloading of tunnels */
> -			uint64_t outer_l3_len:9; /**< Outer L3 (IP) Hdr Length. */
> -			uint64_t outer_l2_len:7; /**< Outer L2 (MAC) Hdr Length. */
> +			uint64_t outer_l3_len:RTE_MBUF_OUTL3_LEN_BITS;
> +			/**< Outer L3 (IP) Hdr Length. */
> +			uint64_t outer_l2_len:RTE_MBUF_OUTL2_LEN_BITS;
> +			/**< Outer L2 (MAC) Hdr Length. */
>  
> -			/* uint64_t unused:8; */
> +			/* uint64_t unused:RTE_MBUF_TXOFLD_UNUSED_BITS; */
>  		};
>  	};
>  
> @@ -2243,6 +2273,41 @@ static inline int rte_pktmbuf_chain(struct rte_mbuf *head, struct rte_mbuf *tail
>  	return 0;
>  }
>  
> +/*
> + * @warning
> + * @b EXPERIMENTAL: This API may change without prior notice.
> + *
> + * For given input values generate raw tx_offload value.
> + * @param il2
> + *   l2_len value.
> + * @param il3
> + *   l3_len value.
> + * @param il4
> + *   l4_len value.
> + * @param tso
> + *   tso_segsz value.
> + * @param ol3
> + *   outer_l3_len value.
> + * @param ol2
> + *   outer_l2_len value.
> + * @param unused
> + *   unused value.
> + * @return
> + *   raw tx_offload value.
> + */
> +static __rte_always_inline uint64_t
> +rte_mbuf_tx_offload(uint64_t il2, uint64_t il3, uint64_t il4, uint64_t tso,
> +	uint64_t ol3, uint64_t ol2, uint64_t unused)
> +{
> +	return il2 << RTE_MBUF_L2_LEN_OFS |
> +		il3 << RTE_MBUF_L3_LEN_OFS |
> +		il4 << RTE_MBUF_L4_LEN_OFS |
> +		tso << RTE_MBUF_TSO_SEGSZ_OFS |
> +		ol3 << RTE_MBUF_OUTL3_LEN_OFS |
> +		ol2 << RTE_MBUF_OUTL2_LEN_OFS |
> +		unused << RTE_MBUF_TXOFLD_UNUSED_OFS;
> +}
> +
>  /**


From what I see, the problem is quite similar to what was done with
rte_mbuf_sched_set() recently. So I wondered if it was possible to
declare a structure like this:

	struct rte_mbuf_ol_len {
	        uint64_t l2_len:7;
	        uint64_t l3_len:9; /**< L3 (IP) Header Length. */
	        uint64_t l4_len:8; /**< L4 (TCP/UDP) Header Length. */
		...
	}

And have the set function like this:

        m->l = (struct rte_mbuf_ol_len) {
                .l2_len = l2_len,
                .l3_len = l3_len,
                .l4_len = l4_len,
		...

This would avoid the definition of the offsets and bits, but I didn't
find any way to declare these fields as anonymous in the mbuf structure.
Did you tried that way too?


Thanks,
Olivier
Ananyev, Konstantin March 30, 2019, 2:20 p.m. UTC | #2 
Hi Olivier,

> > Operations to set/update bit-fields often cause compilers
> > to generate suboptimal code.
> > To help avoid such situation for tx_offload fields:
> > introduce new enum for tx_offload bit-fields lengths and offsets,
> > and new function to generate raw tx_offload value.
> >
> > Signed-off-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> > Acked-by: Akhil Goyal <akhil.goyal@nxp.com>
> 
> I understand the need. Out of curiosity, do you have any performance
> numbers to share?

On my board (SKX):
 for micro-benchmark (doing nothing but setting tx_offload for 1M mbufs in a loop) 
the difference is more than 150% -  from ~55 cycles to ~20 cycles per iteration.
For ipsec-secgw - ~3% improvement for tunneled outbound packets.

> 
> Few cosmetic questions below.
> 
> > ---
> >  lib/librte_mbuf/rte_mbuf.h | 79 ++++++++++++++++++++++++++++++++++----
> >  1 file changed, 72 insertions(+), 7 deletions(-)
> >
> > diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
> > index d961ccaf6..0b197e8ce 100644
> > --- a/lib/librte_mbuf/rte_mbuf.h
> > +++ b/lib/librte_mbuf/rte_mbuf.h
> > @@ -479,6 +479,31 @@ struct rte_mbuf_sched {
> >  	uint16_t reserved;   /**< Reserved. */
> >  }; /**< Hierarchical scheduler */
> >
> > +/**
> > + * enum for the tx_offload bit-fields lenghts and offsets.
> > + * defines the layout of rte_mbuf tx_offload field.
> > + */
> > +enum {
> > +	RTE_MBUF_L2_LEN_BITS = 7,
> > +	RTE_MBUF_L3_LEN_BITS = 9,
> > +	RTE_MBUF_L4_LEN_BITS = 8,
> > +	RTE_MBUF_TSO_SEGSZ_BITS = 16,
> > +	RTE_MBUF_OUTL3_LEN_BITS = 9,
> > +	RTE_MBUF_OUTL2_LEN_BITS = 7,
> > +	RTE_MBUF_L2_LEN_OFS = 0,
> > +	RTE_MBUF_L3_LEN_OFS = RTE_MBUF_L2_LEN_OFS + RTE_MBUF_L2_LEN_BITS,
> > +	RTE_MBUF_L4_LEN_OFS = RTE_MBUF_L3_LEN_OFS + RTE_MBUF_L3_LEN_BITS,
> > +	RTE_MBUF_TSO_SEGSZ_OFS = RTE_MBUF_L4_LEN_OFS + RTE_MBUF_L4_LEN_BITS,
> > +	RTE_MBUF_OUTL3_LEN_OFS =
> > +		RTE_MBUF_TSO_SEGSZ_OFS + RTE_MBUF_TSO_SEGSZ_BITS,
> > +	RTE_MBUF_OUTL2_LEN_OFS =
> > +		RTE_MBUF_OUTL3_LEN_OFS + RTE_MBUF_OUTL3_LEN_BITS,
> > +	RTE_MBUF_TXOFLD_UNUSED_OFS =
> > +		RTE_MBUF_OUTL2_LEN_OFS + RTE_MBUF_OUTL2_LEN_BITS,
> > +	RTE_MBUF_TXOFLD_UNUSED_BITS =
> > +		sizeof(uint64_t) * CHAR_BIT - RTE_MBUF_TXOFLD_UNUSED_OFS,
> > +};
> > +
> 
> What is the advantage of defining an enum instead of #defines?

No big difference here, just looks nicer to me.

> 
> In any case, I wonder if it wouldn't be clearer to change the order like
> this:
> 
> enum {
> 	RTE_MBUF_L2_LEN_OFS = 0,
> 	RTE_MBUF_L2_LEN_BITS = 7,
> 	RTE_MBUF_L3_LEN_OFS = RTE_MBUF_L2_LEN_OFS + RTE_MBUF_L2_LEN_BITS,
> 	RTE_MBUF_L3_LEN_BITS = 9,
> 	RTE_MBUF_L4_LEN_OFS = RTE_MBUF_L3_LEN_OFS + RTE_MBUF_L3_LEN_BITS,
> 	RTE_MBUF_L4_LEN_BITS = 8,
> ...

NP, can do this way.

> 
> 
> >  /**
> >   * The generic rte_mbuf, containing a packet mbuf.
> >   */
> > @@ -640,19 +665,24 @@ struct rte_mbuf {
> >  		uint64_t tx_offload;       /**< combined for easy fetch */
> >  		__extension__
> >  		struct {
> > -			uint64_t l2_len:7;
> > +			uint64_t l2_len:RTE_MBUF_L2_LEN_BITS;
> >  			/**< L2 (MAC) Header Length for non-tunneling pkt.
> >  			 * Outer_L4_len + ... + Inner_L2_len for tunneling pkt.
> >  			 */
> > -			uint64_t l3_len:9; /**< L3 (IP) Header Length. */
> > -			uint64_t l4_len:8; /**< L4 (TCP/UDP) Header Length. */
> > -			uint64_t tso_segsz:16; /**< TCP TSO segment size */
> > +			uint64_t l3_len:RTE_MBUF_L3_LEN_BITS;
> > +			/**< L3 (IP) Header Length. */
> > +			uint64_t l4_len:RTE_MBUF_L4_LEN_BITS;
> > +			/**< L4 (TCP/UDP) Header Length. */
> > +			uint64_t tso_segsz:RTE_MBUF_TSO_SEGSZ_BITS;
> > +			/**< TCP TSO segment size */
> >
> >  			/* fields for TX offloading of tunnels */
> > -			uint64_t outer_l3_len:9; /**< Outer L3 (IP) Hdr Length. */
> > -			uint64_t outer_l2_len:7; /**< Outer L2 (MAC) Hdr Length. */
> > +			uint64_t outer_l3_len:RTE_MBUF_OUTL3_LEN_BITS;
> > +			/**< Outer L3 (IP) Hdr Length. */
> > +			uint64_t outer_l2_len:RTE_MBUF_OUTL2_LEN_BITS;
> > +			/**< Outer L2 (MAC) Hdr Length. */
> >
> > -			/* uint64_t unused:8; */
> > +			/* uint64_t unused:RTE_MBUF_TXOFLD_UNUSED_BITS; */
> >  		};
> >  	};
> >
> > @@ -2243,6 +2273,41 @@ static inline int rte_pktmbuf_chain(struct rte_mbuf *head, struct rte_mbuf *tail
> >  	return 0;
> >  }
> >
> > +/*
> > + * @warning
> > + * @b EXPERIMENTAL: This API may change without prior notice.
> > + *
> > + * For given input values generate raw tx_offload value.
> > + * @param il2
> > + *   l2_len value.
> > + * @param il3
> > + *   l3_len value.
> > + * @param il4
> > + *   l4_len value.
> > + * @param tso
> > + *   tso_segsz value.
> > + * @param ol3
> > + *   outer_l3_len value.
> > + * @param ol2
> > + *   outer_l2_len value.
> > + * @param unused
> > + *   unused value.
> > + * @return
> > + *   raw tx_offload value.
> > + */
> > +static __rte_always_inline uint64_t
> > +rte_mbuf_tx_offload(uint64_t il2, uint64_t il3, uint64_t il4, uint64_t tso,
> > +	uint64_t ol3, uint64_t ol2, uint64_t unused)
> > +{
> > +	return il2 << RTE_MBUF_L2_LEN_OFS |
> > +		il3 << RTE_MBUF_L3_LEN_OFS |
> > +		il4 << RTE_MBUF_L4_LEN_OFS |
> > +		tso << RTE_MBUF_TSO_SEGSZ_OFS |
> > +		ol3 << RTE_MBUF_OUTL3_LEN_OFS |
> > +		ol2 << RTE_MBUF_OUTL2_LEN_OFS |
> > +		unused << RTE_MBUF_TXOFLD_UNUSED_OFS;
> > +}
> > +
> >  /**
> 
> 
> From what I see, the problem is quite similar to what was done with
> rte_mbuf_sched_set() recently. So I wondered if it was possible to
> declare a structure like this:
> 
> 	struct rte_mbuf_ol_len {
> 	        uint64_t l2_len:7;
> 	        uint64_t l3_len:9; /**< L3 (IP) Header Length. */
> 	        uint64_t l4_len:8; /**< L4 (TCP/UDP) Header Length. */
> 		...
> 	}
> 
> And have the set function like this:
> 
>         m->l = (struct rte_mbuf_ol_len) {
>                 .l2_len = l2_len,
>                 .l3_len = l3_len,
>                 .l4_len = l4_len,
> 		...
> 
> This would avoid the definition of the offsets and bits, but I didn't
> find any way to declare these fields as anonymous in the mbuf structure.
> Did you tried that way too?

I thought about such approach, but as you said above it would change
from unnamed struct to named one.
Which, as I understand, means API breakage.
So don't think the hassle will be worth the benefit.
Also the code wouldn't be totally identical - that approach will generate few
extra 'AND' instructions.
Konstantin
diff mbox series

Patch

diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
index d961ccaf6..0b197e8ce 100644
--- a/lib/librte_mbuf/rte_mbuf.h
+++ b/lib/librte_mbuf/rte_mbuf.h
@@ -479,6 +479,31 @@  struct rte_mbuf_sched {
 	uint16_t reserved;   /**< Reserved. */
 }; /**< Hierarchical scheduler */
 
+/**
+ * enum for the tx_offload bit-fields lenghts and offsets.
+ * defines the layout of rte_mbuf tx_offload field.
+ */
+enum {
+	RTE_MBUF_L2_LEN_BITS = 7,
+	RTE_MBUF_L3_LEN_BITS = 9,
+	RTE_MBUF_L4_LEN_BITS = 8,
+	RTE_MBUF_TSO_SEGSZ_BITS = 16,
+	RTE_MBUF_OUTL3_LEN_BITS = 9,
+	RTE_MBUF_OUTL2_LEN_BITS = 7,
+	RTE_MBUF_L2_LEN_OFS = 0,
+	RTE_MBUF_L3_LEN_OFS = RTE_MBUF_L2_LEN_OFS + RTE_MBUF_L2_LEN_BITS,
+	RTE_MBUF_L4_LEN_OFS = RTE_MBUF_L3_LEN_OFS + RTE_MBUF_L3_LEN_BITS,
+	RTE_MBUF_TSO_SEGSZ_OFS = RTE_MBUF_L4_LEN_OFS + RTE_MBUF_L4_LEN_BITS,
+	RTE_MBUF_OUTL3_LEN_OFS =
+		RTE_MBUF_TSO_SEGSZ_OFS + RTE_MBUF_TSO_SEGSZ_BITS,
+	RTE_MBUF_OUTL2_LEN_OFS =
+		RTE_MBUF_OUTL3_LEN_OFS + RTE_MBUF_OUTL3_LEN_BITS,
+	RTE_MBUF_TXOFLD_UNUSED_OFS =
+		RTE_MBUF_OUTL2_LEN_OFS + RTE_MBUF_OUTL2_LEN_BITS,
+	RTE_MBUF_TXOFLD_UNUSED_BITS =
+		sizeof(uint64_t) * CHAR_BIT - RTE_MBUF_TXOFLD_UNUSED_OFS,
+};
+
 /**
  * The generic rte_mbuf, containing a packet mbuf.
  */
@@ -640,19 +665,24 @@  struct rte_mbuf {
 		uint64_t tx_offload;       /**< combined for easy fetch */
 		__extension__
 		struct {
-			uint64_t l2_len:7;
+			uint64_t l2_len:RTE_MBUF_L2_LEN_BITS;
 			/**< L2 (MAC) Header Length for non-tunneling pkt.
 			 * Outer_L4_len + ... + Inner_L2_len for tunneling pkt.
 			 */
-			uint64_t l3_len:9; /**< L3 (IP) Header Length. */
-			uint64_t l4_len:8; /**< L4 (TCP/UDP) Header Length. */
-			uint64_t tso_segsz:16; /**< TCP TSO segment size */
+			uint64_t l3_len:RTE_MBUF_L3_LEN_BITS;
+			/**< L3 (IP) Header Length. */
+			uint64_t l4_len:RTE_MBUF_L4_LEN_BITS;
+			/**< L4 (TCP/UDP) Header Length. */
+			uint64_t tso_segsz:RTE_MBUF_TSO_SEGSZ_BITS;
+			/**< TCP TSO segment size */
 
 			/* fields for TX offloading of tunnels */
-			uint64_t outer_l3_len:9; /**< Outer L3 (IP) Hdr Length. */
-			uint64_t outer_l2_len:7; /**< Outer L2 (MAC) Hdr Length. */
+			uint64_t outer_l3_len:RTE_MBUF_OUTL3_LEN_BITS;
+			/**< Outer L3 (IP) Hdr Length. */
+			uint64_t outer_l2_len:RTE_MBUF_OUTL2_LEN_BITS;
+			/**< Outer L2 (MAC) Hdr Length. */
 
-			/* uint64_t unused:8; */
+			/* uint64_t unused:RTE_MBUF_TXOFLD_UNUSED_BITS; */
 		};
 	};
 
@@ -2243,6 +2273,41 @@  static inline int rte_pktmbuf_chain(struct rte_mbuf *head, struct rte_mbuf *tail
 	return 0;
 }
 
+/*
+ * @warning
+ * @b EXPERIMENTAL: This API may change without prior notice.
+ *
+ * For given input values generate raw tx_offload value.
+ * @param il2
+ *   l2_len value.
+ * @param il3
+ *   l3_len value.
+ * @param il4
+ *   l4_len value.
+ * @param tso
+ *   tso_segsz value.
+ * @param ol3
+ *   outer_l3_len value.
+ * @param ol2
+ *   outer_l2_len value.
+ * @param unused
+ *   unused value.
+ * @return
+ *   raw tx_offload value.
+ */
+static __rte_always_inline uint64_t
+rte_mbuf_tx_offload(uint64_t il2, uint64_t il3, uint64_t il4, uint64_t tso,
+	uint64_t ol3, uint64_t ol2, uint64_t unused)
+{
+	return il2 << RTE_MBUF_L2_LEN_OFS |
+		il3 << RTE_MBUF_L3_LEN_OFS |
+		il4 << RTE_MBUF_L4_LEN_OFS |
+		tso << RTE_MBUF_TSO_SEGSZ_OFS |
+		ol3 << RTE_MBUF_OUTL3_LEN_OFS |
+		ol2 << RTE_MBUF_OUTL2_LEN_OFS |
+		unused << RTE_MBUF_TXOFLD_UNUSED_OFS;
+}
+
 /**
  * Validate general requirements for Tx offload in mbuf.
  *