Skip to contentSkip to conversation

mbuf: replace zero-length marker with unnamed union

PATCH Superseded Gavin Hu 9 participants 24 messages Mar 3, 2020 – May 14, 2020

lore
Share & tools
lore

Gavin Hu

author
#0 lore raw

PATCH 1/1 mbuf: replace zero-length marker with unnamed union

gcc 10.0.1 reports: error: array subscript 0 is outside the bounds of an interior zero-length array 'RTE_MARKER64' {aka 'long unsigned int[0]'} [-Werror=zero-length-bounds] 310 | *(uint64_t *)(&mbuf->rearm_data) = val;

Declaring zero-length arrays in other contexts, including as interior members of structure objects or as non-member objects, is discouraged. Accessing elements of zero-length arrays declared in such contexts is undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")
---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++--------------
 1 file changed, 32 insertions(+), 22 deletions(-)
lib/librte_mbuf/rte_mbuf_core.h +32 -22
index b9a59c879..5390ddcfa 100644
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions (atomic,
-	 * or non-atomic) is controlled by the CONFIG_RTE_MBUF_REFCNT_ATOMIC
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed refcnt */
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data;
+		RTE_STD_C11
+		struct {
+			uint16_t data_off;
+
+			/**
+			 * Reference counter. Its size should at least equal to
+			 * the size of port field (16 bits), to support
+			 * zero-copy broadcast.  It should only be accessed
+			 * using the following functions:
+			 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(),
+			 * and rte_mbuf_refcnt_set(). The functionality of
+			 * these functions (atomic, or non-atomic) is
+			 * controlled by the CONFIG_RTE_MBUF_REFCNT_ATOMIC
+			 * config option.
+			 */
+			RTE_STD_C11
+				union {
+					/**< Atomically accessed refcnt */
+					rte_atomic16_t refcnt_atomic;
+					/** Non-atomically accessed refcnt */
+					uint16_t refcnt;
+				};
+			uint16_t nb_segs;         /**< Number of segments. */
 
-	/** Input port (16 bits to support more than 256 virtual ports).
-	 * The event eth Tx adapter uses this field to specify the output port.
-	 */
-	uint16_t port;
+			/** Input port (16 bits to support more than 256
+			 * virtual ports).  The event eth Tx adapter uses this
+			 * field to specify the output port.
+			 */
+			uint16_t port;
+		};
+	};
 
 	uint64_t ol_flags;        /**< Offload features. */
 
Signature · 1 line
2.25.0
#1 lore raw
Quoted 85 lines · Gavin Hu

gcc 10.0.1 reports: error: array subscript 0 is outside the bounds of an interior zero-length array 'RTE_MARKER64' {aka 'long unsigned int[0]'} [-Werror=zero-length-bounds] 310 | *(uint64_t *)(&mbuf->rearm_data) = val;

Declaring zero-length arrays in other contexts, including as interior members of structure objects or as non-member objects, is discouraged. Accessing elements of zero-length arrays declared in such contexts is undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")
---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++--------------
 1 file changed, 32 insertions(+), 22 deletions(-)
lib/librte_mbuf/rte_mbuf_core.h +25 -22
index b9a59c879..5390ddcfa 100644
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions (atomic,
-	 * or non-atomic) is controlled by the CONFIG_RTE_MBUF_REFCNT_ATOMIC
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed refcnt */
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data;
+		RTE_STD_C11
+		struct {
+			uint16_t data_off;
+
+			/**
+			 * Reference counter. Its size should at least equal to
+			 * the size of port field (16 bits), to support
+			 * zero-copy broadcast.  It should only be accessed
+			 * using the following functions:
+			 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(),
+			 * and rte_mbuf_refcnt_set(). The functionality of
+			 * these functions (atomic, or non-atomic) is
+			 * controlled by the CONFIG_RTE_MBUF_REFCNT_ATOMIC
+			 * config option.
+			 */
+			RTE_STD_C11
+				union {
+					/**< Atomically accessed refcnt */
+					rte_atomic16_t refcnt_atomic;
+					/** Non-atomically accessed refcnt */
+					uint16_t refcnt;
+				};
+			uint16_t nb_segs;         /**< Number of segments. */
 
-	/** Input port (16 bits to support more than 256 virtual ports).
-	 * The event eth Tx adapter uses this field to specify the output port.
-	 */
-	uint16_t port;

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 520

Quoted 10 lines
lib/librte_mbuf/rte_mbuf_core.h +7 -0
+			/** Input port (16 bits to support more than 256
+			 * virtual ports).  The event eth Tx adapter uses this
+			 * field to specify the output port.
+			 */
+			uint16_t port;
+		};
+	};
 
 	uint64_t ol_flags;        /**< Offload features. */
 

Hi Gavin, this causes some errors on x86:

# gcc --version | head -1
gcc (GCC) 10.0.1 20200216 (Red Hat 10.0.1-0.8)

In file included from ../lib/librte_eal/common/include/arch/x86/rte_byteorder.h:13,

                 from ../drivers/net/sfc/sfc_ef10_rx.c:14:

../drivers/net/sfc/sfc_ef10_rx.c: In function ‘sfc_ef10_rx_process_event’:
../drivers/net/sfc/sfc_ef10_rx.c:309:39: error: subscripted value is neither array nor pointer nor vector
309 | RTE_BUILD_BUG_ON(sizeof(m->rearm_data[0]) !=
sizeof(rxq->rearm_data));
| ^
../lib/librte_eal/common/include/rte_common.h:292:65: note: in definition of macro ‘RTE_BUILD_BUG_ON’
292 | #define RTE_BUILD_BUG_ON(condition) ((void)sizeof(char[1 -
2*!!(condition)]))

      |
^~~~~~~~~
../drivers/net/sfc/sfc_ef10_rx.c:310:15: error: subscripted value is
neither array nor pointer nor vector
  310 |  m->rearm_data[0] = rxq->rearm_data;
      |
author
#2 lore raw

Hi Kevin,

Quoted 67 lines

-----Original Message-----
From: Kevin Traynor <[email protected]>
Sent: Wednesday, March 4, 2020 8:33 PM
To: Gavin Hu <[email protected]>; [email protected]
Cc: nd <[email protected]>; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH v1] mbuf: replace zero-length marker with unnamed union

Quoted 21 lines · Gavin Hu

gcc 10.0.1 reports: error: array subscript 0 is outside the bounds of an interior zero-length array 'RTE_MARKER64' {aka 'long unsigned int[0]'} [-Werror=zero-length-bounds] 310 | *(uint64_t *)(&mbuf->rearm_data) = val;

Declaring zero-length arrays in other contexts, including as interior members of structure objects or as non-member objects, is discouraged. Accessing elements of zero-length arrays declared in such contexts is undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")

Review

Quoted 6 lines
---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++--------------
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..5390ddcfa 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 492

Quoted 10 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -10
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions

(atomic,

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed

refcnt */

Quoted 6 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +2 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data;

To address this historical issue, how about changing this line to uint64_t rearm_data[1]?

Quoted 69 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 10 lines
lib/librte_mbuf/rte_mbuf_core.h +10 -0
+		RTE_STD_C11
+		struct {
+			uint16_t data_off;
+
+			/**
+			 * Reference counter. Its size should at least equal to
+			 * the size of port field (16 bits), to support
+			 * zero-copy broadcast.  It should only be accessed
+			 * using the following functions:
+			 * rte_mbuf_refcnt_update(),

rte_mbuf_refcnt_read(),

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 3 lines
lib/librte_mbuf/rte_mbuf_core.h +3 -0
+			 * and rte_mbuf_refcnt_set(). The functionality of
+			 * these functions (atomic, or non-atomic) is
+			 * controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 7 lines
lib/librte_mbuf/rte_mbuf_core.h +7 -0
+			 * config option.
+			 */
+			RTE_STD_C11
+				union {
+					/**< Atomically accessed refcnt */
+					rte_atomic16_t refcnt_atomic;
+					/** Non-atomically accessed refcnt

*/

Quoted 9 lines
lib/librte_mbuf/rte_mbuf_core.h +3 -4
+					uint16_t refcnt;
+				};
+			uint16_t nb_segs;         /**< Number of segments. */
 
-	/** Input port (16 bits to support more than 256 virtual ports).
-	 * The event eth Tx adapter uses this field to specify the output port.
-	 */
-	uint16_t port;

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 520

Quoted 10 lines
lib/librte_mbuf/rte_mbuf_core.h +7 -0
+			/** Input port (16 bits to support more than 256
+			 * virtual ports).  The event eth Tx adapter uses this
+			 * field to specify the output port.
+			 */
+			uint16_t port;
+		};
+	};
 
 	uint64_t ol_flags;        /**< Offload features. */
 

Hi Gavin, this causes some errors on x86:

# gcc --version | head -1
gcc (GCC) 10.0.1 20200216 (Red Hat 10.0.1-0.8)

In file included from ../lib/librte_eal/common/include/arch/x86/rte_byteorder.h:13,

                 from ../drivers/net/sfc/sfc_ef10_rx.c:14:

../drivers/net/sfc/sfc_ef10_rx.c: In function 'sfc_ef10_rx_process_event':
../drivers/net/sfc/sfc_ef10_rx.c:309:39: error: subscripted value is neither array nor pointer nor vector
309 | RTE_BUILD_BUG_ON(sizeof(m->rearm_data[0]) !=
sizeof(rxq->rearm_data));
| ^
../lib/librte_eal/common/include/rte_common.h:292:65: note: in definition of macro 'RTE_BUILD_BUG_ON'
292 | #define RTE_BUILD_BUG_ON(condition) ((void)sizeof(char[1 -
2*!!(condition)]))

      |
^~~~~~~~~
../drivers/net/sfc/sfc_ef10_rx.c:310:15: error: subscripted value is
neither array nor pointer nor vector
  310 |  m->rearm_data[0] = rxq->rearm_data;
      |
author
#3 lore raw

PATCH v2 1/1 mbuf: replace zero-length marker with unnamed union

In reply to Gavin Hu

Declaring zero-length arrays in other contexts, including as interior members of structure objects or as non-member objects, is discouraged. Accessing elements of zero-length arrays declared in such contexts is undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to fix the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++--------------
 1 file changed, 32 insertions(+), 22 deletions(-)
lib/librte_mbuf/rte_mbuf_core.h +32 -22
index b9a59c879..34cb152e2 100644
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions (atomic,
-	 * or non-atomic) is controlled by the CONFIG_RTE_MBUF_REFCNT_ATOMIC
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed refcnt */
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data[1];
+		RTE_STD_C11
+		struct {
+			uint16_t data_off;
+
+			/**
+			 * Reference counter. Its size should at least equal to
+			 * the size of port field (16 bits), to support
+			 * zero-copy broadcast.  It should only be accessed
+			 * using the following functions:
+			 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(),
+			 * and rte_mbuf_refcnt_set(). The functionality of
+			 * these functions (atomic, or non-atomic) is
+			 * controlled by the CONFIG_RTE_MBUF_REFCNT_ATOMIC
+			 * config option.
+			 */
+			RTE_STD_C11
+				union {
+					/**< Atomically accessed refcnt */
+					rte_atomic16_t refcnt_atomic;
+					/** Non-atomically accessed refcnt */
+					uint16_t refcnt;
+				};
+			uint16_t nb_segs;         /**< Number of segments. */
 
-	/** Input port (16 bits to support more than 256 virtual ports).
-	 * The event eth Tx adapter uses this field to specify the output port.
-	 */
-	uint16_t port;
+			/** Input port (16 bits to support more than 256
+			 * virtual ports).  The event eth Tx adapter uses this
+			 * field to specify the output port.
+			 */
+			uint16_t port;
+		};
+	};
 
 	uint64_t ol_flags;        /**< Offload features. */
 
Signature · 1 line
2.25.0
#4 lore raw
Quoted 45 lines · Gavin Hu

Declaring zero-length arrays in other contexts, including as interior members of structure objects or as non-member objects, is discouraged. Accessing elements of zero-length arrays declared in such contexts is undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to fix the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++--------------
 1 file changed, 32 insertions(+), 22 deletions(-)
lib/librte_mbuf/rte_mbuf_core.h +0 -12
index b9a59c879..34cb152e2 100644
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions (atomic,
-	 * or non-atomic) is controlled by the CONFIG_RTE_MBUF_REFCNT_ATOMIC
-	 * config option.

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 10 lines
lib/librte_mbuf/rte_mbuf_core.h +2 -6
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed refcnt */
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we are doing with them, what would you think disabling the warning instead of increasing the complexity in mbuf struct?

author
#5 lore raw

Hi Ferruh,

Quoted 68 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM
To: Gavin Hu <[email protected]>; [email protected]
Cc: nd <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

Quoted 21 lines · Gavin Hu

Declaring zero-length arrays in other contexts, including as interior members of structure objects or as non-member objects, is discouraged. Accessing elements of zero-length arrays declared in such contexts is undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to fix the SFC PMD compiling error on x86. <Kevin Traynor>

Review

Quoted 5 lines
---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++--------------
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 8 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -1
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
-	/* next 8 bytes are initialised on RX descriptor rearm */

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 10 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -10
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions (atomic,
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed

refcnt */

Quoted 11 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +2 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we are doing with them,
what would you think disabling the warning instead of increasing the complexity
in mbuf struct?

Okay, I will add -Wno-zero-length-bounds to the compiler toolchain flags. /Gavin

#6 lore raw
Quoted 81 lines · Gavin Hu

Hi Ferruh,

Quoted 68 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM
To: Gavin Hu <[email protected]>; [email protected]
Cc: nd <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

Quoted 21 lines · Gavin Hu

Declaring zero-length arrays in other contexts, including as interior members of structure objects or as non-member objects, is discouraged. Accessing elements of zero-length arrays declared in such contexts is undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to fix the SFC PMD compiling error on x86. <Kevin Traynor>

Review

Quoted 5 lines
---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++--------------
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 8 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -1
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
-	/* next 8 bytes are initialised on RX descriptor rearm */

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 10 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -10
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions (atomic,
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed

refcnt */

Quoted 11 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +2 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we are doing with them,
what would you think disabling the warning instead of increasing the complexity
in mbuf struct?

Quoted 1 line

Okay, I will add -Wno-zero-length-bounds to the compiler toolchain flags.

This would be my preference but I would like to get more input, can you please for more comments before changing the implementation in case there are some strong opinion on it?

#7 lore raw
Quoted 80 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such contexts

is

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to fix the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++----------

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 10 lines
Quoted 10 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 10 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -10
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions

(atomic,

Quoted 8 lines
Quoted 8 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed

refcnt */

Quoted 7 lines
Quoted 7 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +2 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we are

doing

Quoted 4 lines
Quoted 4 lines

with them,
what would you think disabling the warning instead of increasing the complexity
in mbuf struct?

Quoted 9 lines
Quoted 1 line

Okay, I will add -Wno-zero-length-bounds to the compiler toolchain

flags.

This would be my preference but I would like to get more input, can you please
for more comments before changing the implementation in case there are some
strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-length arrays [...] as interior members of structure objects [...] is discouraged.

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques) should be fixed, not the symptom (i.e. getting warnings about using discouraged techniques).

Compiler warnings are here to help, and in my experience they are actually very helpful, although avoiding them often requires somewhat more verbose source code. Disabling this warning not only affects this file, but disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would be optimal if DPDK could be compiled with -Wall, and it would probably reduce the number of released bugs too.

With that said, sometimes the optimal solution has to give way for the practical solution. And this is a core file, so we should thread lightly.

As for an alternative solution, perhaps we can get rid of the MARKERs in the struct and #define them instead. Not as elegant as Gavin's suggested union based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));

    uint16_t data_off;
    ...
}

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

Med venlig hilsen / kind regards
- Morten Brørup

#8 lore raw
Quoted 143 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Morten Brørup
Sent: Monday, March 9, 2020 1:31 PM
To: Yigit, Ferruh <[email protected]>; Gavin Hu <[email protected]>; [email protected]
Cc: nd <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Ananyev, Konstantin <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

Quoted 50 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such contexts

is

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to fix the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54

+++++++++++++++++++----------

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -6
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the

size

Quoted 13 lines
Quoted 4 lines
Quoted 4 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 4 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -4
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions

(atomic,

Quoted 8 lines
Quoted 8 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed

refcnt */

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +1 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm

*/

Quoted 6 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +1 -0
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we are

doing

Quoted 3 lines
Quoted 3 lines

with them,
what would you think disabling the warning instead of increasing the complexity in mbuf struct?

Quoted 7 lines
Quoted 1 line

Okay, I will add -Wno-zero-length-bounds to the compiler toolchain

flags.

This would be my preference but I would like to get more input, can you please for more comments before changing the implementation in case there are some strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-length arrays [...] as interior members of structure objects [...] is discouraged.

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques) should be fixed, not the symptom (i.e. getting warnings about using discouraged techniques).

Compiler warnings are here to help, and in my experience they are actually very helpful, although avoiding them often requires somewhat more verbose source code. Disabling this warning not only affects this file, but disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would be optimal if DPDK could be compiled with -Wall, and it would probably reduce the number of released bugs too.

With that said, sometimes the optimal solution has to give way for the practical solution. And this is a core file, so we should thread lightly.

As for an alternative solution, perhaps we can get rid of the MARKERs in the struct and #define them instead. Not as elegant as Gavin's suggested union based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));

    uint16_t data_off;
    ...
}
Quoted 3 lines

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

+1 for this, it's very similar to what I was going to suggest, which was:

uint16_t data_off;
#define _REARM_DATA data_off

but your suggestion is probably cleaner than mine.

/Bruce

#9 lore raw

Re: [dpdk-dev] [PATCH v2] mbuf: replace zero-length markerwith unnamed union

Quoted 153 lines

From: dev [mailto:[email protected]] On Behalf Of Richardson, Bruce
Sent: Monday, March 9, 2020 3:16 PM

Quoted 18 lines

From: dev <[email protected]> On Behalf Of Morten Brørup
Sent: Monday, March 9, 2020 1:31 PM

Quoted 16 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such

contexts

Quoted 11 lines
Quoted 11 lines

is

Quoted 10 lines
Quoted 10 lines
Quoted 10 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to

fix

Quoted 23 lines
Quoted 16 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54

+++++++++++++++++++----------

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -6
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the

size

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 490

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following

functions:

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 492

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these

functions

Quoted 8 lines
Quoted 8 lines

(atomic,

Quoted 7 lines
Quoted 7 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically

accessed

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

refcnt */

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +1 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor

rearm

Quoted 3 lines

*/

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +1 -0
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we

are

Quoted 5 lines
Quoted 5 lines

doing

Quoted 4 lines
Quoted 3 lines

with them,
what would you think disabling the warning instead of increasing the complexity in mbuf struct?

Okay, I will add -Wno-zero-length-bounds to the compiler

toolchain

Quoted 10 lines
Quoted 6 lines

flags.

This would be my preference but I would like to get more input, can you please for more comments before changing the implementation in case there are some strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-

length

Quoted 9 lines

arrays [...] as interior members of structure objects [...] is discouraged.

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques) should be fixed, not the symptom (i.e. getting warnings about using discouraged techniques).

Compiler warnings are here to help, and in my experience they are

actually

Quoted 1 line

very helpful, although avoiding them often requires somewhat more

verbose

Quoted 4 lines

source code. Disabling this warning not only affects this file, but disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would

be

Quoted 1 line

optimal if DPDK could be compiled with -Wall, and it would probably

reduce

Quoted 3 lines

the number of released bugs too.

With that said, sometimes the optimal solution has to give way for

the

Quoted 1 line

practical solution. And this is a core file, so we should thread

lightly.

Quoted 3 lines

As for an alternative solution, perhaps we can get rid of the MARKERs

in

Quoted 1 line

the struct and #define them instead. Not as elegant as Gavin's

suggested

Quoted 10 lines

union based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));

    uint16_t data_off;
    ...
}
Quoted 3 lines

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

+1 for this, it's very similar to what I was going to suggest, which was:

Quoted 6 lines

uint16_t data_off;
#define _REARM_DATA data_off

but your suggestion is probably cleaner than mine.

/Bruce

I prefer Gavin's academically correct solution, i.e. using unions.

But check out the code using rearm_data! It's all drivers, and they all type cast it, except the SFC PMD.

So Bruce's suggestion is probably more workable than mine, although the #define'd name is missing an RTE prefix. (And mine is missing an &.)

And a completely alternative route is open: Type cast rearm_data in the SFC PMD, and postpone the MARKER discussion to later.

-Morten

author
#10 lore raw
In reply to Morten Brørup

Hi Morten,

Quoted 141 lines

-----Original Message-----
From: Morten Brørup <[email protected]>
Sent: Monday, March 9, 2020 9:31 PM
To: Ferruh Yigit <[email protected]>; Gavin Hu <[email protected]>; [email protected]
Cc: nd <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

Quoted 80 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such contexts

is

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to fix the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++----------

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 10 lines
Quoted 10 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 10 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -10
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions

(atomic,

Quoted 8 lines
Quoted 8 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed

refcnt */

Quoted 7 lines
Quoted 7 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +2 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we are

doing

Quoted 4 lines
Quoted 4 lines

with them,
what would you think disabling the warning instead of increasing the complexity
in mbuf struct?

Quoted 9 lines
Quoted 1 line

Okay, I will add -Wno-zero-length-bounds to the compiler toolchain

flags.

This would be my preference but I would like to get more input, can you please
for more comments before changing the implementation in case there are some
strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-length arrays [...] as interior members of structure objects [...] is discouraged.

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques) should be fixed, not the symptom (i.e. getting warnings about using discouraged techniques).

Compiler warnings are here to help, and in my experience they are actually very helpful, although avoiding them often requires somewhat more verbose source code. Disabling this warning not only affects this file, but disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would be optimal if DPDK could be compiled with -Wall, and it would probably reduce the number of released bugs too.

With that said, sometimes the optimal solution has to give way for the practical solution. And this is a core file, so we should thread lightly.

As for an alternative solution, perhaps we can get rid of the MARKERs in the struct and #define them instead. Not as elegant as Gavin's suggested union based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));

    uint16_t data_off;
    ...
}
Quoted 1 line

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

This does not work out, it generates new errors:
/root/dpdk/build/include/rte_mbuf_core.h:485:33: error: dereferencing type-punned pointer will break strict-aliasing rules [-Werror=strict-aliasing]
485 | #define rte_mbuf_rearm_data(m) ((uint64_t *)&m->data_off)

Quoted 4 lines

Med venlig hilsen / kind regards
- Morten Brørup

#11 lore raw
Quoted 158 lines

From: dev [mailto:[email protected]] On Behalf Of Gavin Hu
Sent: Wednesday, March 11, 2020 8:50 AM

Hi Morten,

Quoted 19 lines

From: Morten Brørup <[email protected]>
Sent: Monday, March 9, 2020 9:31 PM

Quoted 16 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such

contexts

Quoted 11 lines
Quoted 11 lines

is

Quoted 10 lines
Quoted 10 lines
Quoted 10 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to

fix

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++------

----

Quoted 12 lines
Quoted 12 lines

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -6
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the

size

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 490

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following

functions:

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 492

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these

functions

Quoted 8 lines
Quoted 8 lines

(atomic,

Quoted 7 lines
Quoted 7 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically

accessed

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

refcnt */

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf {

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +1 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor

rearm */

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +1 -0
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we

are

Quoted 3 lines
Quoted 3 lines

doing

Quoted 2 lines
Quoted 2 lines

with them,
what would you think disabling the warning instead of increasing

the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 2 lines

complexity
in mbuf struct?

Okay, I will add -Wno-zero-length-bounds to the compiler

toolchain

Quoted 3 lines
Quoted 3 lines

flags.

This would be my preference but I would like to get more input, can

you

Quoted 2 lines
Quoted 2 lines

please
for more comments before changing the implementation in case there

are

Quoted 7 lines
Quoted 3 lines

some
strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-

length

Quoted 1 line

arrays [...] as interior members of structure objects [...] is

discouraged.

Quoted 3 lines

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques)

should

Quoted 1 line

be fixed, not the symptom (i.e. getting warnings about using

discouraged

Quoted 3 lines

techniques).

Compiler warnings are here to help, and in my experience they are

actually

Quoted 2 lines

very helpful, although avoiding them often requires somewhat more verbose source code. Disabling this warning not only affects this

file, but

Quoted 3 lines

disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would

be

Quoted 1 line

optimal if DPDK could be compiled with -Wall, and it would probably

reduce

Quoted 3 lines

the number of released bugs too.

With that said, sometimes the optimal solution has to give way for

the

Quoted 1 line

practical solution. And this is a core file, so we should thread

lightly.

Quoted 3 lines

As for an alternative solution, perhaps we can get rid of the MARKERs

in the

Quoted 1 line

struct and #define them instead. Not as elegant as Gavin's suggested

union

Quoted 10 lines

based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));

    uint16_t data_off;
    ...
}
Quoted 8 lines
Quoted 1 line

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

This does not work out, it generates new errors:
/root/dpdk/build/include/rte_mbuf_core.h:485:33: error: dereferencing type-punned pointer will break strict-aliasing rules [-Werror=strict- aliasing]
485 | #define rte_mbuf_rearm_data(m) ((uint64_t *)&m->data_off)

OK. Then Bruce's suggestion probably won't work either.

I found this article about strict aliasing: https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8

The article basically says that the union based method (i.e. your original suggestion) is valid C (but not C++) and is the common solution.

Alternatives have now been discussed and tested, so we should all support your original suggestion, which seems to be the only correct and viable solution.

Please go ahead with that, and then someone should update the SFC PMD accordingly.

Furthermore, I think that Stephen's suggestion about getting rid of the markers all together is good thinking, but it would require updating a lot of PMDs accordingly. So please also consider removing other markers that can be removed without affecting a whole bunch of other files.

-Morten

#12 lore raw
Quoted 171 lines · Morten Brørup
Quoted 158 lines

From: dev [mailto:[email protected]] On Behalf Of Gavin Hu
Sent: Wednesday, March 11, 2020 8:50 AM

Hi Morten,

Quoted 19 lines

From: Morten Brørup <[email protected]>
Sent: Monday, March 9, 2020 9:31 PM

Quoted 16 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such

contexts

Quoted 11 lines
Quoted 11 lines

is

Quoted 10 lines
Quoted 10 lines
Quoted 10 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to

fix

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++------

----

Quoted 12 lines
Quoted 12 lines

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 488

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -6
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the

size

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 490

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following

functions:

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 492

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these

functions

Quoted 8 lines
Quoted 8 lines

(atomic,

Quoted 7 lines
Quoted 7 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 498

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically

accessed

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

refcnt */

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 485

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +1 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor

rearm */

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +1 -0
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we

are

Quoted 3 lines
Quoted 3 lines

doing

Quoted 2 lines
Quoted 2 lines

with them,
what would you think disabling the warning instead of increasing

the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 2 lines

complexity
in mbuf struct?

Okay, I will add -Wno-zero-length-bounds to the compiler

toolchain

Quoted 3 lines
Quoted 3 lines

flags.

This would be my preference but I would like to get more input, can

you

Quoted 2 lines
Quoted 2 lines

please
for more comments before changing the implementation in case there

are

Quoted 7 lines
Quoted 3 lines

some
strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-

length

Quoted 1 line

arrays [...] as interior members of structure objects [...] is

discouraged.

Quoted 3 lines

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques)

should

Quoted 1 line

be fixed, not the symptom (i.e. getting warnings about using

discouraged

Quoted 3 lines

techniques).

Compiler warnings are here to help, and in my experience they are

actually

Quoted 2 lines

very helpful, although avoiding them often requires somewhat more verbose source code. Disabling this warning not only affects this

file, but

Quoted 3 lines

disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would

be

Quoted 1 line

optimal if DPDK could be compiled with -Wall, and it would probably

reduce

Quoted 3 lines

the number of released bugs too.

With that said, sometimes the optimal solution has to give way for

the

Quoted 1 line

practical solution. And this is a core file, so we should thread

lightly.

Quoted 3 lines

As for an alternative solution, perhaps we can get rid of the MARKERs

in the

Quoted 1 line

struct and #define them instead. Not as elegant as Gavin's suggested

union

Quoted 10 lines

based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));

    uint16_t data_off;
    ...
}
Quoted 8 lines
Quoted 1 line

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

This does not work out, it generates new errors:
/root/dpdk/build/include/rte_mbuf_core.h:485:33: error: dereferencing type-punned pointer will break strict-aliasing rules [-Werror=strict- aliasing]
485 | #define rte_mbuf_rearm_data(m) ((uint64_t *)&m->data_off)

OK. Then Bruce's suggestion probably won't work either.

I found this article about strict aliasing: https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8

Quoted 8 lines

The article basically says that the union based method (i.e. your original suggestion) is valid C (but not C++) and is the common solution.

Alternatives have now been discussed and tested, so we should all support your original suggestion, which seems to be the only correct and viable solution.

Please go ahead with that, and then someone should update the SFC PMD accordingly.

Furthermore, I think that Stephen's suggestion about getting rid of the markers all together is good thinking, but it would require updating a lot of PMDs accordingly. So please also consider removing other markers that can be removed without affecting a whole bunch of other files.

Does it still give errors if we don't have the cast in the macro?

author
#13 lore raw

Hi Bruce,

Quoted 199 lines

-----Original Message-----
From: Bruce Richardson <[email protected]>
Sent: Wednesday, March 11, 2020 8:08 PM
To: Morten Brørup <[email protected]>
Cc: Gavin Hu <[email protected]>; Ferruh Yigit <[email protected]>; [email protected]; nd <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

Quoted 39 lines · Morten Brørup
Quoted 39 lines

From: dev [mailto:[email protected]] On Behalf Of Gavin Hu
Sent: Wednesday, March 11, 2020 8:50 AM

Hi Morten,

Quoted 19 lines

From: Morten Brørup <[email protected]>
Sent: Monday, March 9, 2020 9:31 PM

Quoted 16 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such

contexts

Quoted 14 lines
Quoted 14 lines

is

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")
Quoted 8 lines
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to

fix

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++----

--

Quoted 120 lines
Quoted 111 lines

----

Quoted 12 lines
Quoted 12 lines

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 488

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -6
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the

size

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 490

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following

functions:

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 492

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these

functions

Quoted 8 lines
Quoted 8 lines

(atomic,

Quoted 7 lines
Quoted 7 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 498

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically

accessed

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

refcnt */

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 485

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +1 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor

rearm */

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +1 -0
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we

are

Quoted 3 lines
Quoted 3 lines

doing

Quoted 2 lines
Quoted 2 lines

with them,
what would you think disabling the warning instead of increasing

the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 2 lines

complexity
in mbuf struct?

Okay, I will add -Wno-zero-length-bounds to the compiler

toolchain

Quoted 3 lines
Quoted 3 lines

flags.

This would be my preference but I would like to get more input, can

you

Quoted 2 lines
Quoted 2 lines

please
for more comments before changing the implementation in case there

are

Quoted 7 lines
Quoted 3 lines

some
strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-

length

Quoted 1 line

arrays [...] as interior members of structure objects [...] is

discouraged.

Quoted 3 lines

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques)

should

Quoted 1 line

be fixed, not the symptom (i.e. getting warnings about using

discouraged

Quoted 3 lines

techniques).

Compiler warnings are here to help, and in my experience they are

actually

Quoted 2 lines

very helpful, although avoiding them often requires somewhat more verbose source code. Disabling this warning not only affects this

file, but

Quoted 3 lines

disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would

be

Quoted 1 line

optimal if DPDK could be compiled with -Wall, and it would probably

reduce

Quoted 3 lines

the number of released bugs too.

With that said, sometimes the optimal solution has to give way for

the

Quoted 1 line

practical solution. And this is a core file, so we should thread

lightly.

Quoted 3 lines

As for an alternative solution, perhaps we can get rid of the MARKERs

in the

Quoted 1 line

struct and #define them instead. Not as elegant as Gavin's suggested

union

Quoted 10 lines

based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));

    uint16_t data_off;
    ...
}
Quoted 8 lines
Quoted 1 line

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

This does not work out, it generates new errors:
/root/dpdk/build/include/rte_mbuf_core.h:485:33: error: dereferencing type-punned pointer will break strict-aliasing rules [-Werror=strict- aliasing]
485 | #define rte_mbuf_rearm_data(m) ((uint64_t *)&m->data_off)

Quoted 3 lines

OK. Then Bruce's suggestion probably won't work either.

I found this article about strict aliasing:

https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8

Quoted 2 lines

The article basically says that the union based method (i.e. your original

suggestion) is valid C (but not C++) and is the common solution.

Quoted 2 lines

Alternatives have now been discussed and tested, so we should all support

your original suggestion, which seems to be the only correct and viable solution.

Quoted 2 lines

Please go ahead with that, and then someone should update the SFC PMD

accordingly.

Quoted 1 line
Quoted 7 lines
Quoted 1 line

Furthermore, I think that Stephen's suggestion about getting rid of the

markers all together is good thinking, but it would require updating a lot of PMDs accordingly. So please also consider removing other markers that can be removed without affecting a whole bunch of other files.

Quoted 1 line

Does it still give errors if we don't have the cast in the macro?

Yes the errors persist if we have the cast in dereferencing.

author
#14 lore raw
In reply to Bruce Richardson

Hi Bruce,

Quoted 199 lines

-----Original Message-----
From: Bruce Richardson <[email protected]>
Sent: Wednesday, March 11, 2020 8:08 PM
To: Morten Brørup <[email protected]>
Cc: Gavin Hu <[email protected]>; Ferruh Yigit <[email protected]>; [email protected]; nd <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

Quoted 39 lines · Morten Brørup
Quoted 39 lines

From: dev [mailto:[email protected]] On Behalf Of Gavin Hu
Sent: Wednesday, March 11, 2020 8:50 AM

Hi Morten,

Quoted 19 lines

From: Morten Brørup <[email protected]>
Sent: Monday, March 9, 2020 9:31 PM

Quoted 16 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such

contexts

Quoted 14 lines
Quoted 14 lines

is

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")
Quoted 8 lines
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to

fix

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++----

--

Quoted 120 lines
Quoted 111 lines

----

Quoted 12 lines
Quoted 12 lines

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 488

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -6
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the

size

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 490

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following

functions:

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 492

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these

functions

Quoted 8 lines
Quoted 8 lines

(atomic,

Quoted 7 lines
Quoted 7 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 498

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically

accessed

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

refcnt */

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 485

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +1 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor

rearm */

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +1 -0
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we

are

Quoted 3 lines
Quoted 3 lines

doing

Quoted 2 lines
Quoted 2 lines

with them,
what would you think disabling the warning instead of increasing

the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 2 lines

complexity
in mbuf struct?

Okay, I will add -Wno-zero-length-bounds to the compiler

toolchain

Quoted 3 lines
Quoted 3 lines

flags.

This would be my preference but I would like to get more input, can

you

Quoted 2 lines
Quoted 2 lines

please
for more comments before changing the implementation in case there

are

Quoted 7 lines
Quoted 3 lines

some
strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-

length

Quoted 1 line

arrays [...] as interior members of structure objects [...] is

discouraged.

Quoted 3 lines

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques)

should

Quoted 1 line

be fixed, not the symptom (i.e. getting warnings about using

discouraged

Quoted 3 lines

techniques).

Compiler warnings are here to help, and in my experience they are

actually

Quoted 2 lines

very helpful, although avoiding them often requires somewhat more verbose source code. Disabling this warning not only affects this

file, but

Quoted 3 lines

disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would

be

Quoted 1 line

optimal if DPDK could be compiled with -Wall, and it would probably

reduce

Quoted 3 lines

the number of released bugs too.

With that said, sometimes the optimal solution has to give way for

the

Quoted 1 line

practical solution. And this is a core file, so we should thread

lightly.

Quoted 3 lines

As for an alternative solution, perhaps we can get rid of the MARKERs

in the

Quoted 1 line

struct and #define them instead. Not as elegant as Gavin's suggested

union

Quoted 10 lines

based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));

    uint16_t data_off;
    ...
}
Quoted 8 lines
Quoted 1 line

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

This does not work out, it generates new errors:
/root/dpdk/build/include/rte_mbuf_core.h:485:33: error: dereferencing type-punned pointer will break strict-aliasing rules [-Werror=strict- aliasing]
485 | #define rte_mbuf_rearm_data(m) ((uint64_t *)&m->data_off)

Quoted 3 lines

OK. Then Bruce's suggestion probably won't work either.

I found this article about strict aliasing:

https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8

Quoted 2 lines

The article basically says that the union based method (i.e. your original

suggestion) is valid C (but not C++) and is the common solution.

Quoted 2 lines

Alternatives have now been discussed and tested, so we should all support

your original suggestion, which seems to be the only correct and viable solution.

Quoted 2 lines

Please go ahead with that, and then someone should update the SFC PMD

accordingly.

Quoted 1 line
Quoted 7 lines
Quoted 1 line

Furthermore, I think that Stephen's suggestion about getting rid of the

markers all together is good thinking, but it would require updating a lot of PMDs accordingly. So please also consider removing other markers that can be removed without affecting a whole bunch of other files.

Quoted 1 line

Does it still give errors if we don't have the cast in the macro?

Yes, it gives errors elsewhere that have the cast.

#15 lore raw
Quoted 202 lines · Gavin Hu

Hi Bruce,

Quoted 200 lines

-----Original Message-----
From: Bruce Richardson <[email protected]>
Sent: Wednesday, March 11, 2020 8:08 PM
To: Morten Brørup <[email protected]>
Cc: Gavin Hu <[email protected]>; Ferruh Yigit <[email protected]>; [email protected]; nd <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

Quoted 39 lines · Morten Brørup
Quoted 39 lines

From: dev [mailto:[email protected]] On Behalf Of Gavin Hu
Sent: Wednesday, March 11, 2020 8:50 AM

Hi Morten,

Quoted 19 lines

From: Morten Brørup <[email protected]>
Sent: Monday, March 9, 2020 9:31 PM

Quoted 16 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such

contexts

Quoted 14 lines
Quoted 14 lines

is

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")
Quoted 8 lines
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to

fix

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++----

--

Quoted 120 lines
Quoted 111 lines

----

Quoted 12 lines
Quoted 12 lines

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 488

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -6
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the

size

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 490

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following

functions:

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 492

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these

functions

Quoted 8 lines
Quoted 8 lines

(atomic,

Quoted 7 lines
Quoted 7 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 498

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically

accessed

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

refcnt */

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 485

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +1 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor

rearm */

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +1 -0
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what we

are

Quoted 3 lines
Quoted 3 lines

doing

Quoted 2 lines
Quoted 2 lines

with them,
what would you think disabling the warning instead of increasing

the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 2 lines

complexity
in mbuf struct?

Okay, I will add -Wno-zero-length-bounds to the compiler

toolchain

Quoted 3 lines
Quoted 3 lines

flags.

This would be my preference but I would like to get more input, can

you

Quoted 2 lines
Quoted 2 lines

please
for more comments before changing the implementation in case there

are

Quoted 7 lines
Quoted 3 lines

some
strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-

length

Quoted 1 line

arrays [...] as interior members of structure objects [...] is

discouraged.

Quoted 3 lines

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques)

should

Quoted 1 line

be fixed, not the symptom (i.e. getting warnings about using

discouraged

Quoted 3 lines

techniques).

Compiler warnings are here to help, and in my experience they are

actually

Quoted 2 lines

very helpful, although avoiding them often requires somewhat more verbose source code. Disabling this warning not only affects this

file, but

Quoted 3 lines

disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would

be

Quoted 1 line

optimal if DPDK could be compiled with -Wall, and it would probably

reduce

Quoted 3 lines

the number of released bugs too.

With that said, sometimes the optimal solution has to give way for

the

Quoted 1 line

practical solution. And this is a core file, so we should thread

lightly.

Quoted 3 lines

As for an alternative solution, perhaps we can get rid of the MARKERs

in the

Quoted 1 line

struct and #define them instead. Not as elegant as Gavin's suggested

union

Quoted 10 lines

based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));

    uint16_t data_off;
    ...
}
Quoted 8 lines
Quoted 1 line

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

This does not work out, it generates new errors:
/root/dpdk/build/include/rte_mbuf_core.h:485:33: error: dereferencing type-punned pointer will break strict-aliasing rules [-Werror=strict- aliasing]
485 | #define rte_mbuf_rearm_data(m) ((uint64_t *)&m->data_off)

Quoted 3 lines

OK. Then Bruce's suggestion probably won't work either.

I found this article about strict aliasing:

https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8

Quoted 2 lines

The article basically says that the union based method (i.e. your original

suggestion) is valid C (but not C++) and is the common solution.

Quoted 2 lines

Alternatives have now been discussed and tested, so we should all support

your original suggestion, which seems to be the only correct and viable solution.

Quoted 2 lines

Please go ahead with that, and then someone should update the SFC PMD

accordingly.

Quoted 2 lines

Furthermore, I think that Stephen's suggestion about getting rid of the

Quoted 9 lines
Quoted 6 lines

markers all together is good thinking, but it would require updating a lot of PMDs accordingly. So please also consider removing other markers that can be removed without affecting a whole bunch of other files.

Quoted 1 line

Does it still give errors if we don't have the cast in the macro?

Yes, it gives errors elsewhere that have the cast.

Hi Gavin, I lost track if v2 is still a candidate for merge. fwiw, it compiles without giving the zero-length-bounds warning on my system.

Kevin.

author
#16 lore raw

Hi Kevin,

Quoted 239 lines

-----Original Message-----
From: Kevin Traynor <[email protected]>
Sent: Wednesday, April 8, 2020 1:14 AM
To: Gavin Hu <[email protected]>; Bruce Richardson <[email protected]>; Morten Brørup <[email protected]>
Cc: Ferruh Yigit <[email protected]>; [email protected]; nd <[email protected]>; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

Quoted 7 lines · Gavin Hu

Hi Bruce,

Quoted 5 lines

-----Original Message-----
From: Bruce Richardson <[email protected]>
Sent: Wednesday, March 11, 2020 8:08 PM
To: Morten Brørup <[email protected]>
Cc: Gavin Hu <[email protected]>; Ferruh Yigit

<[email protected]>;

Quoted 3 lines
Quoted 3 lines

[email protected]; nd <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng

Wang

Quoted 6 lines
Quoted 6 lines

<[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: Re: [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker

with

Quoted 87 lines
Quoted 87 lines

unnamed union

Quoted 39 lines · Morten Brørup
Quoted 39 lines

From: dev [mailto:[email protected]] On Behalf Of Gavin Hu
Sent: Wednesday, March 11, 2020 8:50 AM

Hi Morten,

Quoted 19 lines

From: Morten Brørup <[email protected]>
Sent: Monday, March 9, 2020 9:31 PM

Quoted 16 lines

From: dev [mailto:[email protected]] On Behalf Of Ferruh Yigit
Sent: Monday, March 9, 2020 12:30 PM

Quoted 8 lines · Gavin Hu

Hi Ferruh,

Quoted 6 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Monday, March 9, 2020 4:55 PM

Quoted 1 line · Gavin Hu

Declaring zero-length arrays in other contexts, including as

interior

Quoted 1 line
Quoted 1 line
Quoted 1 line

members of structure objects or as non-member objects, is

discouraged.

Quoted 1 line
Quoted 1 line
Quoted 1 line

Accessing elements of zero-length arrays declared in such

contexts

Quoted 14 lines
Quoted 14 lines

is

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")
Quoted 8 lines
Quoted 8 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to

fix

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++----

--

Quoted 36 lines
Quoted 36 lines

----

Quoted 12 lines
Quoted 12 lines

----

Quoted 11 lines
Quoted 11 lines

Review

Quoted 3 lines
 1 file changed, 32 insertions(+), 22 deletions(-)
(quoted diff) +0 -0

b/lib/librte_mbuf/rte_mbuf_core.h

Quoted 7 lines

index b9a59c879..34cb152e2 100644

lib/librte_mbuf/rte_mbuf_core.h +0 -0
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 488

Quoted 6 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -6
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the

size

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 490

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following

functions:

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 492

Quoted 2 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -2
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these

functions

Quoted 8 lines
Quoted 8 lines

(atomic,

Quoted 7 lines
Quoted 7 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 493

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +0 -1
-	 * or non-atomic) is controlled by the

CONFIG_RTE_MBUF_REFCNT_ATOMIC

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 498

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +0 -3
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically

accessed

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

refcnt */

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 485

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +1 -4
-		/** Non-atomically accessed refcnt */
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor

rearm */

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 1 line
lib/librte_mbuf/rte_mbuf_core.h +1 -0
+		uint64_t rearm_data[1];

We are using zero length array as markers only and know what

we

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

are

Quoted 3 lines
Quoted 3 lines

doing

Quoted 2 lines
Quoted 2 lines

with them,
what would you think disabling the warning instead of increasing

the

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

complexity
in mbuf struct?

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line

Okay, I will add -Wno-zero-length-bounds to the compiler

toolchain

Quoted 3 lines
Quoted 3 lines

flags.

This would be my preference but I would like to get more input, can

you

Quoted 2 lines
Quoted 2 lines

please
for more comments before changing the implementation in case

there

Quoted 48 lines
Quoted 48 lines
Quoted 48 lines
Quoted 48 lines

are

Quoted 7 lines
Quoted 3 lines

some
strong opinion on it?

I have some input to this discussion.

Let me repeat what Gavin's GCC reference states: Declaring zero-

length

Quoted 1 line

arrays [...] as interior members of structure objects [...] is

discouraged.

Quoted 3 lines

Why would we do something that the compiler documentation says is discouraged? I think the problem (i.e. using discouraged techniques)

should

Quoted 1 line

be fixed, not the symptom (i.e. getting warnings about using

discouraged

Quoted 3 lines

techniques).

Compiler warnings are here to help, and in my experience they are

actually

Quoted 2 lines

very helpful, although avoiding them often requires somewhat more verbose source code. Disabling this warning not only affects this

file, but

Quoted 3 lines

disables warnings about potential bugs in other source code too.

Generally, disabling compiler warnings is a slippery slope. It would

be

Quoted 1 line

optimal if DPDK could be compiled with -Wall, and it would probably

reduce

Quoted 3 lines

the number of released bugs too.

With that said, sometimes the optimal solution has to give way for

the

Quoted 1 line

practical solution. And this is a core file, so we should thread

lightly.

Quoted 3 lines

As for an alternative solution, perhaps we can get rid of the MARKERs

in the

Quoted 1 line

struct and #define them instead. Not as elegant as Gavin's suggested

union

Quoted 6 lines

based solution, but it might bring inspiration...

struct rte_mbuf {
    ...
    } __rte_aligned(sizeof(rte_iova_t));
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 5 lines
    uint16_t data_off;
    ...
}

#define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)

This does not work out, it generates new errors:
/root/dpdk/build/include/rte_mbuf_core.h:485:33: error:

dereferencing

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 4 lines

type-punned pointer will break strict-aliasing rules [-Werror=strict- aliasing]
485 | #define rte_mbuf_rearm_data(m) ((uint64_t *)&m->data_off)

OK. Then Bruce's suggestion probably won't work either.

Quoted 7 lines
Quoted 7 lines
Quoted 1 line

I found this article about strict aliasing:

https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8

Quoted 2 lines

The article basically says that the union based method (i.e. your original

suggestion) is valid C (but not C++) and is the common solution.

Quoted 2 lines

Alternatives have now been discussed and tested, so we should all

support

Quoted 1 line
Quoted 1 line

your original suggestion, which seems to be the only correct and viable

solution.

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

Please go ahead with that, and then someone should update the SFC

PMD

Quoted 4 lines
Quoted 4 lines

accordingly.

Quoted 2 lines

Furthermore, I think that Stephen's suggestion about getting rid of the

markers all together is good thinking, but it would require updating a lot

of

Quoted 1 line
Quoted 1 line

PMDs accordingly. So please also consider removing other markers that

can be

Quoted 5 lines
Quoted 4 lines

removed without affecting a whole bunch of other files.

Quoted 1 line

Does it still give errors if we don't have the cast in the macro?

Quoted 7 lines
Quoted 2 lines

Yes, it gives errors elsewhere that have the cast.

Hi Gavin, I lost track if v2 is still a candidate for merge. fwiw, it compiles without giving the zero-length-bounds warning on my system.

Kevin.

Yes, this path alone is a candidate for merge. We brainstormed other solutions but they did not work out.

/Gavin

#17 lore raw
Quoted 8 lines · Gavin Hu
Quoted 6 lines

-----Original Message-----
From: Kevin Traynor <[email protected]>
Hi Gavin, I lost track if v2 is still a candidate for merge. fwiw, it compiles without giving the zero-length-bounds warning on my system.

Kevin.

Yes, this path alone is a candidate for merge.

This patch is not mergeable, it would trigger failures in the ABI checks.

You can see in patchwork that the robot reported a warning in Travis.
http://mails.dpdk.org/archives/test-report/2020-March/119919.html
https://travis-ci.com/github/ovsrobot/dpdk/jobs/295652710#L4476

I opened a bz to libabigail.
https://sourceware.org/bugzilla/show_bug.cgi?id=25661

Either a different solution is found, or your patch will have to deal with this issue (libabigail fix won't be ready soon afaik) and waive this.

Signature · 1 line
David Marchand
author
#18 lore raw

Hi David,

Quoted 24 lines

-----Original Message-----
From: David Marchand <[email protected]>
Sent: Wednesday, April 8, 2020 11:22 PM
To: Gavin Hu <[email protected]>
Cc: Kevin Traynor <[email protected]>; Bruce Richardson <[email protected]>; Morten Brørup <[email protected]>; Ferruh Yigit <[email protected]>; [email protected]; nd <[email protected]>; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>
Subject: Re: [dpdk-stable] [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

Quoted 5 lines · Gavin Hu
Quoted 5 lines

-----Original Message-----
From: Kevin Traynor <[email protected]>
Hi Gavin, I lost track if v2 is still a candidate for merge. fwiw, it compiles without giving the zero-length-bounds warning on my system.

Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Kevin.

Yes, this path alone is a candidate for merge.

This patch is not mergeable, it would trigger failures in the ABI checks.

Isn't it a false failure? If yes, is it ignorable?

Quoted 6 lines

You can see in patchwork that the robot reported a warning in Travis.
http://mails.dpdk.org/archives/test-report/2020-March/119919.html
https://travis-ci.com/github/ovsrobot/dpdk/jobs/295652710#L4476

Quoted 7 lines

I opened a bz to libabigail.
https://sourceware.org/bugzilla/show_bug.cgi?id=25661

Either a different solution is found, or your patch will have to deal with this issue (libabigail fix won't be ready soon afaik) and waive this.

Maybe we come back to 'disable the warning', before the libabigail fix ready? or alternatively ignore this ABI false failure, if it is. I do not have ideas of what otherwise the options are.

Quoted 3 lines

--
David Marchand

#19 lore raw

09/04/2020 11:48, Gavin Hu:

Quoted 23 lines

From: David Marchand <[email protected]>

Quoted 10 lines · Gavin Hu
Quoted 7 lines

From: Kevin Traynor <[email protected]>

Quoted 4 lines

Hi Gavin, I lost track if v2 is still a candidate for merge. fwiw, it compiles without giving the zero-length-bounds warning on my system.

Kevin.

Yes, this path alone is a candidate for merge.

This patch is not mergeable, it would trigger failures in the ABI checks.

Isn't it a false failure? If yes, is it ignorable?

Quoted 9 lines

You can see in patchwork that the robot reported a warning in Travis.
http://mails.dpdk.org/archives/test-report/2020-March/119919.html
https://travis-ci.com/github/ovsrobot/dpdk/jobs/295652710#L4476

I opened a bz to libabigail.
https://sourceware.org/bugzilla/show_bug.cgi?id=25661

Quoted 6 lines
Quoted 3 lines

Either a different solution is found, or your patch will have to deal with this issue (libabigail fix won't be ready soon afaik) and waive this.

Maybe we come back to 'disable the warning', before the libabigail fix ready? or alternatively ignore this ABI false failure, if it is. I do not have ideas of what otherwise the options are.

Gavin,
I did not check this case.
But in general, we do not skip checks, except some checkpatch ones. The policy with ABI checks is "NEVER SKIP".
We prefer postponing patches, waiting for someone to fix tooling.

There is a lack of motivation currently for general concerns. We need to avoid being "write-only" contributors. So two things need to be done:
1/ improve tooling where it needs
2/ review patches from others
I published a review list recently:

	http://mails.dpdk.org/archives/announce/2020-April/000315.html
#20 lore raw
Quoted 28 lines · Thomas Monjalon

09/04/2020 11:48, Gavin Hu:

Quoted 27 lines

From: David Marchand <[email protected]>

Quoted 10 lines · Gavin Hu
Quoted 7 lines

From: Kevin Traynor <[email protected]>

Quoted 4 lines

Hi Gavin, I lost track if v2 is still a candidate for merge. fwiw, it compiles without giving the zero-length-bounds warning on my system.

Kevin.

Yes, this path alone is a candidate for merge.

This patch is not mergeable, it would trigger failures in the ABI checks.

Isn't it a false failure? If yes, is it ignorable?

Quoted 12 lines

You can see in patchwork that the robot reported a warning in Travis.
http://mails.dpdk.org/archives/test-report/2020-March/119919.html
https://travis-ci.com/github/ovsrobot/dpdk/jobs/295652710#L4476

I opened a bz to libabigail.
https://sourceware.org/bugzilla/show_bug.cgi?id=25661

Either a different solution is found, or your patch will have to deal with this issue (libabigail fix won't be ready soon afaik) and waive this.

Quoted 8 lines
Quoted 2 lines

Maybe we come back to 'disable the warning', before the libabigail fix ready? or alternatively ignore this ABI false failure, if it is. I do not have ideas of what otherwise the options are.

Gavin,
I did not check this case.
But in general, we do not skip checks, except some checkpatch ones. The policy with ABI checks is "NEVER SKIP".
We prefer postponing patches, waiting for someone to fix tooling.

In this case Dave Marchand has more than adequately shown the issue is because of a libabigail failure. and has raised an appropriate BZ in the right place, much kudos.

My read of this, is that libabigail fix, will be complicated.

Quoted 7 lines

There is a lack of motivation currently for general concerns. We need to avoid being "write-only" contributors. So two things need to be done:
1/ improve tooling where it needs
2/ review patches from others
I published a review list recently:

	http://mails.dpdk.org/archives/announce/2020-April/000315.html
author
#21 lore raw
In reply to Thomas Monjalon
Quoted 50 lines

-----Original Message-----
From: Thomas Monjalon <[email protected]>
Sent: Thursday, April 9, 2020 6:50 PM
To: Gavin Hu <[email protected]>
Cc: David Marchand <[email protected]>; Kevin Traynor <[email protected]>; Bruce Richardson <[email protected]>; Morten Brørup <[email protected]>; Ferruh Yigit <[email protected]>; [email protected]; nd <[email protected]>; [email protected]; Honnappa Nagarahalli <[email protected]>; Ruifeng Wang <[email protected]>; Phil Yang <[email protected]>; Joyce Kong <[email protected]>; [email protected]; Olivier MATZ <[email protected]>; Konstantin Ananyev <[email protected]>; Andrew Rybchenko <[email protected]>; nd <[email protected]>; [email protected]
Subject: Re: [dpdk-stable] [dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

09/04/2020 11:48, Gavin Hu:

Quoted 5 lines

From: David Marchand <[email protected]>

Quoted 4 lines · Gavin Hu
Quoted 3 lines

From: Kevin Traynor <[email protected]>

Quoted 2 lines

Hi Gavin, I lost track if v2 is still a candidate for merge. fwiw, it compiles without giving the zero-length-bounds warning on my

system.

Quoted 18 lines
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

Kevin.

Yes, this path alone is a candidate for merge.

This patch is not mergeable, it would trigger failures in the ABI checks.

Isn't it a false failure? If yes, is it ignorable?

Quoted 9 lines

You can see in patchwork that the robot reported a warning in Travis.
http://mails.dpdk.org/archives/test-report/2020-March/119919.html
https://travis-ci.com/github/ovsrobot/dpdk/jobs/295652710#L4476

I opened a bz to libabigail.
https://sourceware.org/bugzilla/show_bug.cgi?id=25661

Quoted 5 lines
Quoted 3 lines

Either a different solution is found, or your patch will have to deal with this issue (libabigail fix won't be ready soon afaik) and waive this.

Maybe we come back to 'disable the warning', before the libabigail fix

Quoted 8 lines

ready? or alternatively ignore this ABI false failure, if it is.

Quoted 1 line

I do not have ideas of what otherwise the options are.

Gavin,
I did not check this case.
But in general, we do not skip checks, except some checkpatch ones. The policy with ABI checks is "NEVER SKIP".
We prefer postponing patches, waiting for someone to fix tooling.

Ok, I am fine with this.

Quoted 7 lines

There is a lack of motivation currently for general concerns. We need to avoid being "write-only" contributors. So two things need to be done:
1/ improve tooling where it needs
2/ review patches from others
I published a review list recently:

	http://mails.dpdk.org/archives/announce/2020-April/000315.html

Thanks!

Quoted 1 line
#22 lore raw
In reply to David Marchand
Quoted 6 lines · David Marchand
Quoted 5 lines · Gavin Hu
Quoted 5 lines

-----Original Message-----
From: Kevin Traynor <[email protected]>
Hi Gavin, I lost track if v2 is still a candidate for merge. fwiw, it compiles without giving the zero-length-bounds warning on my system.

Quoted 6 lines
Quoted 3 lines
Quoted 1 line

Kevin.

Yes, this path alone is a candidate for merge.

This patch is not mergeable, it would trigger failures in the ABI checks.

Sent
http://inbox.dpdk.org/dev/[email protected] as this patch was not reworked/merged.

Kevin.

Quoted 14 lines

You can see in patchwork that the robot reported a warning in Travis.
http://mails.dpdk.org/archives/test-report/2020-March/119919.html
https://travis-ci.com/github/ovsrobot/dpdk/jobs/295652710#L4476

I opened a bz to libabigail.
https://sourceware.org/bugzilla/show_bug.cgi?id=25661

Either a different solution is found, or your patch will have to deal with this issue (libabigail fix won't be ready soon afaik) and waive this.

#23 lore raw
In reply to Ferruh Yigit
Quoted 51 lines · On Mon, 9 Mar 2020 08:55:05 +0000 Ferruh Yigit <[email protected]> wrote:
Quoted 50 lines · Gavin Hu

Declaring zero-length arrays in other contexts, including as interior members of structure objects or as non-member objects, is discouraged. Accessing elements of zero-length arrays declared in such contexts is undefined and may be diagnosed.[1]

Fix by using unnamed union and struct.

https://bugs.dpdk.org/show_bug.cgi?id=396

Bugzilla ID: 396

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html

Fixes 3e6181b07038 ("mbuf: use structure marker from EAL")

---
v2:
* change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to fix the SFC PMD compiling error on x86. <Kevin Traynor>

---
 lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++--------------
 1 file changed, 32 insertions(+), 22 deletions(-)
lib/librte_mbuf/rte_mbuf_core.h +0 -15
index b9a59c879..34cb152e2 100644
@@ -480,31 +480,41 @@ struct rte_mbuf {
 		rte_iova_t buf_physaddr; /**< deprecated */
 	} __rte_aligned(sizeof(rte_iova_t));
 
-	/* next 8 bytes are initialised on RX descriptor rearm */
-	RTE_MARKER64 rearm_data;
-	uint16_t data_off;
-
-	/**
-	 * Reference counter. Its size should at least equal to the size
-	 * of port field (16 bits), to support zero-copy broadcast.
-	 * It should only be accessed using the following functions:
-	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
-	 * rte_mbuf_refcnt_set(). The functionality of these functions (atomic,
-	 * or non-atomic) is controlled by the CONFIG_RTE_MBUF_REFCNT_ATOMIC
-	 * config option.
-	 */
 	RTE_STD_C11
 	union {
-		rte_atomic16_t refcnt_atomic; /**< Atomically accessed refcnt */
-		/** Non-atomically accessed refcnt */
Quoted 8 lines

lib/librte_mbuf/rte_mbuf_core.h struct rte_mbuf { · line 486

Quoted 5 lines
lib/librte_mbuf/rte_mbuf_core.h +2 -3
-		uint16_t refcnt;
-	};
-	uint16_t nb_segs;         /**< Number of segments. */
+		/* next 8 bytes are initialised on RX descriptor rearm */
+		uint64_t rearm_data[1];  

We are using zero length array as markers only and know what we are doing with them, what would you think disabling the warning instead of increasing the complexity in mbuf struct?

No to disabling warnings.

Or get rid of the markers all together, the usage is awkward already.