Skip to contentSkip to conversation

mbuf: add bulk free function

PATCH Superseded Morten Brørup 3 participants 6 messages Oct 10, 2019

v6 of 9 Superseded Latest is v9 →

lore
Share & tools
lore
#0 lore raw

v6 cover mbuf: add bulk free function

Add function for freeing a bulk of mbufs.

---

v6:
* Remove __rte_always_inline from static function.
The compiler will inline anyway.
v5:
* Rename variables from "free" to "pending" for improved readability.
* Add prefix __ to rte_pktmbuf_free_seg_via_array().
* Add array size parameter to __rte_pktmbuf_free_seg_via_array(). The compiler will optimize the parameter away anyway.
* Add description to __rte_pktmbuf_free_seg_via_array().
* Minor description updates.
v4:
* Mark as experimental by adding __rte_experimental.
* Add function to experimental section of map file.
* Fix source code formatting regarding pointer to pointer.
* Squash multiple modifications into one.
v3:
* Bugfix: Handle pakets with multiple segments.
* Add inline helper function, mainly for readability.
* Fix source code formatting regarding indentation.
v2:
* Function is not inline.
* Optimize to free multible mbufs belonging to the same mempool in bulk. Inspired by ixgbe_tx_free_bufs(), but allowing NULL pointers in the array, just like rte_pktmbuf_free() can take a NULL pointer.
* Use unsigned int instead of unsigned. Passes checkpatch, but mismatches the original coding style of the modified files.
* Fix a typo in the description headline: mempools is plural.

Morten Brørup (2):
mbuf: add bulk free function
mbuf: add bulk free function

 lib/librte_mbuf/rte_mbuf.c           | 66 ++++++++++++++++++++++++++++
 lib/librte_mbuf/rte_mbuf.h           | 15 +++++++
 lib/librte_mbuf/rte_mbuf_version.map |  1 +
 3 files changed, 82 insertions(+)
Signature · 1 line
2.17.1
#1 lore raw

PATCH v6 1/2 mbuf: add bulk free function

Add function for freeing a bulk of mbufs.

---

v6:
* Remove __rte_always_inline from static function.
The compiler will inline anyway.
v5:
* Rename variables from "free" to "pending" for improved readability.
* Add prefix __ to rte_pktmbuf_free_seg_via_array().
* Add array size parameter to __rte_pktmbuf_free_seg_via_array(). The compiler will optimize the parameter away anyway.
* Add description to __rte_pktmbuf_free_seg_via_array().
* Minor description updates.
v4:
* Mark as experimental by adding __rte_experimental.
* Add function to experimental section of map file.
* Fix source code formatting regarding pointer to pointer.
* Squash multiple modifications into one.
v3:
* Bugfix: Handle pakets with multiple segments.
* Add inline helper function, mainly for readability.
* Fix source code formatting regarding indentation.
v2:
* Function is not inline.
* Optimize to free multible mbufs belonging to the same mempool in bulk. Inspired by ixgbe_tx_free_bufs(), but allowing NULL pointers in the array, just like rte_pktmbuf_free() can take a NULL pointer.
* Use unsigned int instead of unsigned. Passes checkpatch, but mismatches the original coding style of the modified files.
* Fix a typo in the description headline: mempools is plural.

---
 lib/librte_mbuf/rte_mbuf.c           | 66 ++++++++++++++++++++++++++++
 lib/librte_mbuf/rte_mbuf.h           | 15 +++++++
 lib/librte_mbuf/rte_mbuf_version.map |  1 +
 3 files changed, 82 insertions(+)
lib/librte_mbuf/rte_mbuf.c +66 -0
index 37718d49c..de5659baa 100644
@@ -245,6 +245,72 @@ int rte_mbuf_check(const struct rte_mbuf *m, int is_header,
 	return 0;
 }
 
+/**
+ * @internal helper function for freeing a bulk of packet mbuf segments
+ * via an array holding the packet mbuf segments from the same mempool
+ * pending to be freed.
+ *
+ * @param m
+ *  The packet mbuf segment to be freed.
+ * @param pending
+ *  Pointer to the array of packet mbuf segments pending to be freed.
+ * @param nb_pending
+ *  Pointer to the number of elements held in the array.
+ * @param pending_sz
+ *  Number of elements the array can hold.
+ *  Note: The compiler should optimize this parameter away when using a
+ *  constant value, such as RTE_PKTMBUF_FREE_PENDING_SZ.
+ */
+static __rte_always_inline void
+__rte_pktmbuf_free_seg_via_array(struct rte_mbuf *m,
+	struct rte_mbuf ** const pending, unsigned int * const nb_pending,
+	const unsigned int pending_sz)
+{
+	m = rte_pktmbuf_prefree_seg(m);
+	if (likely(m != NULL)) {
+		if (*nb_pending == pending_sz ||
+		    (*nb_pending > 0 && m->pool != pending[0]->pool)) {
+			rte_mempool_put_bulk(pending[0]->pool,
+					(void **)pending, *nb_pending);
+			*nb_pending = 0;
+		}
+
+		pending[(*nb_pending)++] = m;
+	}
+}
+
+/**
+ * Size of the array holding mbufs from the same membool pending to be freed
+ * in bulk.
+ */
+#define RTE_PKTMBUF_FREE_PENDING_SZ 64
+
+/* Free a bulk of packet mbufs back into their original mempools. */
+void rte_pktmbuf_free_bulk(struct rte_mbuf **mbufs, unsigned int count)
+{
+	struct rte_mbuf *m, *m_next, *pending[RTE_PKTMBUF_FREE_PENDING_SZ];
+	unsigned int idx, nb_pending = 0;
+
+	for (idx = 0; idx < count; idx++) {
+		m = mbufs[idx];
+		if (unlikely(m == NULL))
+			continue;
+
+		__rte_mbuf_sanity_check(m, 1);
+
+		do {
+			m_next = m->next;
+			__rte_pktmbuf_free_seg_via_array(m,
+					pending, &nb_pending,
+					RTE_PKTMBUF_FREE_PENDING_SZ);
+			m = m_next;
+		} while (m != NULL);
+	}
+
+	if (nb_pending > 0)
+		rte_mempool_put_bulk(pending[0]->pool, (void **)pending, nb_pending);
+}
+
 /* dump a mbuf on console */
 void
 rte_pktmbuf_dump(FILE *f, const struct rte_mbuf *m, unsigned dump_len)
lib/librte_mbuf/rte_mbuf.h +15 -0
index 98225ec80..a02c8b12c 100644
@@ -1907,6 +1907,21 @@ static inline void rte_pktmbuf_free(struct rte_mbuf *m)
 	}
 }
 
+/**
+ * Free a bulk of packet mbufs back into their original mempools.
+ *
+ * Free a bulk of mbufs, and all their segments in case of chained buffers.
+ * Each segment is added back into its original mempool.
+ *
+ *  @param mbufs
+ *    Array of pointers to packet mbufs.
+ *    The array may contain NULL pointers.
+ *  @param count
+ *    Array size.
+ */
+__rte_experimental
+void rte_pktmbuf_free_bulk(struct rte_mbuf **mbufs, unsigned int count);
+
 /**
  * Creates a "clone" of the given packet mbuf.
  *
lib/librte_mbuf/rte_mbuf_version.map +1 -0
index 2662a37bf..cd6154ef2 100644
@@ -50,4 +50,5 @@ EXPERIMENTAL {
 	global:
 
 	rte_mbuf_check;
+	rte_pktmbuf_free_bulk;
 } DPDK_18.08;
Signature · 1 line
2.17.1
#2 lore raw

Hi CI Team,

Although my patch only adds a single function, which is not being called from anywhere, the patch failed the iol-intel-Performance test:
http://mails.dpdk.org/archives/test-report/2019-October/101829.html

So I think something went wrong with the test environment.

Is there any way to trigger a re-run of the test?

Med venlig hilsen / kind regards
- Morten Brørup

Quoted 179 lines

-----Original Message-----
From: Morten Brørup [mailto:[email protected]]
Sent: Thursday, October 10, 2019 11:35 AM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; Morten Brørup
Subject: [PATCH v6 1/2] mbuf: add bulk free function

Add function for freeing a bulk of mbufs.

---

v6:
* Remove __rte_always_inline from static function.
The compiler will inline anyway.
v5:
* Rename variables from "free" to "pending" for improved readability.
* Add prefix __ to rte_pktmbuf_free_seg_via_array().
* Add array size parameter to __rte_pktmbuf_free_seg_via_array(). The compiler will optimize the parameter away anyway.
* Add description to __rte_pktmbuf_free_seg_via_array().
* Minor description updates.
v4:
* Mark as experimental by adding __rte_experimental.
* Add function to experimental section of map file.
* Fix source code formatting regarding pointer to pointer.
* Squash multiple modifications into one.
v3:
* Bugfix: Handle pakets with multiple segments.
* Add inline helper function, mainly for readability.
* Fix source code formatting regarding indentation.
v2:
* Function is not inline.
* Optimize to free multible mbufs belonging to the same mempool in bulk. Inspired by ixgbe_tx_free_bufs(), but allowing NULL pointers in the array, just like rte_pktmbuf_free() can take a NULL pointer.
* Use unsigned int instead of unsigned. Passes checkpatch, but mismatches the original coding style of the modified files.
* Fix a typo in the description headline: mempools is plural.

---
 lib/librte_mbuf/rte_mbuf.c           | 66 ++++++++++++++++++++++++++++
 lib/librte_mbuf/rte_mbuf.h           | 15 +++++++
 lib/librte_mbuf/rte_mbuf_version.map |  1 +
 3 files changed, 82 insertions(+)
lib/librte_mbuf/rte_mbuf.c +2 -0
index 37718d49c..de5659baa 100644
@@ -245,6 +245,72 @@ int rte_mbuf_check(const struct rte_mbuf *m, int
 is_header,
 	return 0;
 }
 
+/**
+ * @internal helper function for freeing a bulk of packet mbuf

segments
+ * via an array holding the packet mbuf segments from the same mempool
+ * pending to be freed.
+ *
+ * @param m
+ * The packet mbuf segment to be freed.
+ * @param pending
+ * Pointer to the array of packet mbuf segments pending to be freed.
+ * @param nb_pending
+ * Pointer to the number of elements held in the array.
+ * @param pending_sz
+ * Number of elements the array can hold.
+ * Note: The compiler should optimize this parameter away when using a
+ * constant value, such as RTE_PKTMBUF_FREE_PENDING_SZ.
+ */

lib/librte_mbuf/rte_mbuf.c +3 -0
+static __rte_always_inline void
+__rte_pktmbuf_free_seg_via_array(struct rte_mbuf *m,
+	struct rte_mbuf ** const pending, unsigned int * const

nb_pending,

lib/librte_mbuf/rte_mbuf.c +17 -0
+	const unsigned int pending_sz)
+{
+	m = rte_pktmbuf_prefree_seg(m);
+	if (likely(m != NULL)) {
+		if (*nb_pending == pending_sz ||
+		    (*nb_pending > 0 && m->pool != pending[0]->pool)) {
+			rte_mempool_put_bulk(pending[0]->pool,
+					(void **)pending, *nb_pending);
+			*nb_pending = 0;
+		}
+
+		pending[(*nb_pending)++] = m;
+	}
+}
+
+/**
+ * Size of the array holding mbufs from the same membool pending to be

freed
+ * in bulk.
+ */

lib/librte_mbuf/rte_mbuf.c +4 -0
+#define RTE_PKTMBUF_FREE_PENDING_SZ 64
+
+/* Free a bulk of packet mbufs back into their original mempools. */
+void rte_pktmbuf_free_bulk(struct rte_mbuf **mbufs, unsigned int

count)

lib/librte_mbuf/rte_mbuf.c +2 -0
+{
+	struct rte_mbuf *m, *m_next,

*pending[RTE_PKTMBUF_FREE_PENDING_SZ];

lib/librte_mbuf/rte_mbuf.c +20 -0
+	unsigned int idx, nb_pending = 0;
+
+	for (idx = 0; idx < count; idx++) {
+		m = mbufs[idx];
+		if (unlikely(m == NULL))
+			continue;
+
+		__rte_mbuf_sanity_check(m, 1);
+
+		do {
+			m_next = m->next;
+			__rte_pktmbuf_free_seg_via_array(m,
+					pending, &nb_pending,
+					RTE_PKTMBUF_FREE_PENDING_SZ);
+			m = m_next;
+		} while (m != NULL);
+	}
+
+	if (nb_pending > 0)
+		rte_mempool_put_bulk(pending[0]->pool, (void **)pending,

nb_pending);

lib/librte_mbuf/rte_mbuf.c +2 -0
+}
+
 /* dump a mbuf on console */
 void
 rte_pktmbuf_dump(FILE *f, const struct rte_mbuf *m, unsigned dump_len)
lib/librte_mbuf/rte_mbuf.h +4 -0
index 98225ec80..a02c8b12c 100644
@@ -1907,6 +1907,21 @@ static inline void rte_pktmbuf_free(struct
 rte_mbuf *m)
 	}
 }
 
+/**
+ * Free a bulk of packet mbufs back into their original mempools.
+ *
+ * Free a bulk of mbufs, and all their segments in case of chained

buffers.
+ * Each segment is added back into its original mempool.
+ *
+ * @param mbufs
+ * Array of pointers to packet mbufs.
+ * The array may contain NULL pointers.
+ * @param count
+ * Array size.
+ */

lib/librte_mbuf/rte_mbuf.h +2 -0
+__rte_experimental
+void rte_pktmbuf_free_bulk(struct rte_mbuf **mbufs, unsigned int

count);

lib/librte_mbuf/rte_mbuf.h +1 -0
+
 /**
  * Creates a "clone" of the given packet mbuf.
  *
lib/librte_mbuf/rte_mbuf_version.map +1 -1
index 2662a37bf..cd6154ef2 100644
@@ -50,4 +50,5 @@ EXPERIMENTAL {
 	global:
 
 	rte_mbuf_check;
+	rte_pktmbuf_free_bulk;
 } DPDK_18.08;
--

2.17.1

#3 lore raw

Hi Morten,

Sometimes the tests report a false positive. Unfortunately, there isn't a way yet for tests to be reran by contributors. At the moment, only the environment owners (in this case, some members from the Intel team) and I (who maintains the "iol" prefixed CI) can rerun the test.

Depending on the patch, the maintainers should realize that a patch shouldn't be reporting a failure and ignore the result (or rerun if they are able to).

I've reran the test for you. Although, I've noticed a new patch version was sent (and passed), so it won't really make a difference.

Thanks

Quoted 195 lines · Morten Brørup

Hi CI Team,

Although my patch only adds a single function, which is not being called from anywhere, the patch failed the iol-intel-Performance test:
http://mails.dpdk.org/archives/test-report/2019-October/101829.html

So I think something went wrong with the test environment.

Is there any way to trigger a re-run of the test?

Med venlig hilsen / kind regards
- Morten Brørup

Quoted 179 lines

-----Original Message-----
From: Morten Brørup [mailto:[email protected]]
Sent: Thursday, October 10, 2019 11:35 AM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; Morten Brørup
Subject: [PATCH v6 1/2] mbuf: add bulk free function

Add function for freeing a bulk of mbufs.

---

v6:
* Remove __rte_always_inline from static function.
The compiler will inline anyway.
v5:
* Rename variables from "free" to "pending" for improved readability.
* Add prefix __ to rte_pktmbuf_free_seg_via_array().
* Add array size parameter to __rte_pktmbuf_free_seg_via_array(). The compiler will optimize the parameter away anyway.
* Add description to __rte_pktmbuf_free_seg_via_array().
* Minor description updates.
v4:
* Mark as experimental by adding __rte_experimental.
* Add function to experimental section of map file.
* Fix source code formatting regarding pointer to pointer.
* Squash multiple modifications into one.
v3:
* Bugfix: Handle pakets with multiple segments.
* Add inline helper function, mainly for readability.
* Fix source code formatting regarding indentation.
v2:
* Function is not inline.
* Optimize to free multible mbufs belonging to the same mempool in bulk. Inspired by ixgbe_tx_free_bufs(), but allowing NULL pointers in the array, just like rte_pktmbuf_free() can take a NULL pointer.
* Use unsigned int instead of unsigned. Passes checkpatch, but mismatches the original coding style of the modified files.
* Fix a typo in the description headline: mempools is plural.

---
 lib/librte_mbuf/rte_mbuf.c           | 66 ++++++++++++++++++++++++++++
 lib/librte_mbuf/rte_mbuf.h           | 15 +++++++
 lib/librte_mbuf/rte_mbuf_version.map |  1 +
 3 files changed, 82 insertions(+)
lib/librte_mbuf/rte_mbuf.c +2 -0
index 37718d49c..de5659baa 100644
@@ -245,6 +245,72 @@ int rte_mbuf_check(const struct rte_mbuf *m, int
 is_header,
      return 0;
 }
 
+/**
+ * @internal helper function for freeing a bulk of packet mbuf

segments
+ * via an array holding the packet mbuf segments from the same mempool
+ * pending to be freed.
+ *
+ * @param m
+ * The packet mbuf segment to be freed.
+ * @param pending
+ * Pointer to the array of packet mbuf segments pending to be freed.
+ * @param nb_pending
+ * Pointer to the number of elements held in the array.
+ * @param pending_sz
+ * Number of elements the array can hold.
+ * Note: The compiler should optimize this parameter away when using a
+ * constant value, such as RTE_PKTMBUF_FREE_PENDING_SZ.
+ */

lib/librte_mbuf/rte_mbuf.c +3 -0
+static __rte_always_inline void
+__rte_pktmbuf_free_seg_via_array(struct rte_mbuf *m,
+     struct rte_mbuf ** const pending, unsigned int * const

nb_pending,

lib/librte_mbuf/rte_mbuf.c +17 -0
+     const unsigned int pending_sz)
+{
+     m = rte_pktmbuf_prefree_seg(m);
+     if (likely(m != NULL)) {
+             if (*nb_pending == pending_sz ||
+                 (*nb_pending > 0 && m->pool != pending[0]->pool)) {
+                     rte_mempool_put_bulk(pending[0]->pool,
+                                     (void **)pending, *nb_pending);
+                     *nb_pending = 0;
+             }
+
+             pending[(*nb_pending)++] = m;
+     }
+}
+
+/**
+ * Size of the array holding mbufs from the same membool pending to be

freed
+ * in bulk.
+ */

lib/librte_mbuf/rte_mbuf.c +4 -0
+#define RTE_PKTMBUF_FREE_PENDING_SZ 64
+
+/* Free a bulk of packet mbufs back into their original mempools. */
+void rte_pktmbuf_free_bulk(struct rte_mbuf **mbufs, unsigned int

count)

lib/librte_mbuf/rte_mbuf.c +2 -0
+{
+     struct rte_mbuf *m, *m_next,

*pending[RTE_PKTMBUF_FREE_PENDING_SZ];

lib/librte_mbuf/rte_mbuf.c +20 -0
+     unsigned int idx, nb_pending = 0;
+
+     for (idx = 0; idx < count; idx++) {
+             m = mbufs[idx];
+             if (unlikely(m == NULL))
+                     continue;
+
+             __rte_mbuf_sanity_check(m, 1);
+
+             do {
+                     m_next = m->next;
+                     __rte_pktmbuf_free_seg_via_array(m,
+                                     pending, &nb_pending,
+                                     RTE_PKTMBUF_FREE_PENDING_SZ);
+                     m = m_next;
+             } while (m != NULL);
+     }
+
+     if (nb_pending > 0)
+             rte_mempool_put_bulk(pending[0]->pool, (void **)pending,

nb_pending);

lib/librte_mbuf/rte_mbuf.c +2 -0
+}
+
 /* dump a mbuf on console */
 void
 rte_pktmbuf_dump(FILE *f, const struct rte_mbuf *m, unsigned dump_len)
lib/librte_mbuf/rte_mbuf.h +4 -0
index 98225ec80..a02c8b12c 100644
@@ -1907,6 +1907,21 @@ static inline void rte_pktmbuf_free(struct
 rte_mbuf *m)
      }
 }
 
+/**
+ * Free a bulk of packet mbufs back into their original mempools.
+ *
+ * Free a bulk of mbufs, and all their segments in case of chained

buffers.
+ * Each segment is added back into its original mempool.
+ *
+ * @param mbufs
+ * Array of pointers to packet mbufs.
+ * The array may contain NULL pointers.
+ * @param count
+ * Array size.
+ */

lib/librte_mbuf/rte_mbuf.h +2 -0
+__rte_experimental
+void rte_pktmbuf_free_bulk(struct rte_mbuf **mbufs, unsigned int

count);

lib/librte_mbuf/rte_mbuf.h +1 -0
+
 /**
  * Creates a "clone" of the given packet mbuf.
  *
lib/librte_mbuf/rte_mbuf_version.map +1 -1
index 2662a37bf..cd6154ef2 100644
@@ -50,4 +50,5 @@ EXPERIMENTAL {
      global:
 
      rte_mbuf_check;
+     rte_pktmbuf_free_bulk;
 } DPDK_18.08;
--

2.17.1

Signature · 2 lines
Jeremy Plsek
UNH InterOperability Laboratory
#4 lore raw
In reply to Morten Brørup
Quoted 6 lines

Hi CI Team,

Although my patch only adds a single function, which is not being called from anywhere,

Which actually means we forgot that UT would also probably be needed for that addition :(

Quoted 191 lines

the patch failed the iol-intel-Performance test:
http://mails.dpdk.org/archives/test-report/2019-October/101829.html

So I think something went wrong with the test environment.

Is there any way to trigger a re-run of the test?

Med venlig hilsen / kind regards
- Morten Brørup

Quoted 179 lines

-----Original Message-----
From: Morten Brørup [mailto:[email protected]]
Sent: Thursday, October 10, 2019 11:35 AM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; Morten Brørup
Subject: [PATCH v6 1/2] mbuf: add bulk free function

Add function for freeing a bulk of mbufs.

---

v6:
* Remove __rte_always_inline from static function.
The compiler will inline anyway.
v5:
* Rename variables from "free" to "pending" for improved readability.
* Add prefix __ to rte_pktmbuf_free_seg_via_array().
* Add array size parameter to __rte_pktmbuf_free_seg_via_array(). The compiler will optimize the parameter away anyway.
* Add description to __rte_pktmbuf_free_seg_via_array().
* Minor description updates.
v4:
* Mark as experimental by adding __rte_experimental.
* Add function to experimental section of map file.
* Fix source code formatting regarding pointer to pointer.
* Squash multiple modifications into one.
v3:
* Bugfix: Handle pakets with multiple segments.
* Add inline helper function, mainly for readability.
* Fix source code formatting regarding indentation.
v2:
* Function is not inline.
* Optimize to free multible mbufs belonging to the same mempool in bulk. Inspired by ixgbe_tx_free_bufs(), but allowing NULL pointers in the array, just like rte_pktmbuf_free() can take a NULL pointer.
* Use unsigned int instead of unsigned. Passes checkpatch, but mismatches the original coding style of the modified files.
* Fix a typo in the description headline: mempools is plural.

---
 lib/librte_mbuf/rte_mbuf.c           | 66 ++++++++++++++++++++++++++++
 lib/librte_mbuf/rte_mbuf.h           | 15 +++++++
 lib/librte_mbuf/rte_mbuf_version.map |  1 +
 3 files changed, 82 insertions(+)
lib/librte_mbuf/rte_mbuf.c +2 -0
index 37718d49c..de5659baa 100644
@@ -245,6 +245,72 @@ int rte_mbuf_check(const struct rte_mbuf *m, int
 is_header,
 	return 0;
 }
 
+/**
+ * @internal helper function for freeing a bulk of packet mbuf

segments
+ * via an array holding the packet mbuf segments from the same mempool
+ * pending to be freed.
+ *
+ * @param m
+ * The packet mbuf segment to be freed.
+ * @param pending
+ * Pointer to the array of packet mbuf segments pending to be freed.
+ * @param nb_pending
+ * Pointer to the number of elements held in the array.
+ * @param pending_sz
+ * Number of elements the array can hold.
+ * Note: The compiler should optimize this parameter away when using a
+ * constant value, such as RTE_PKTMBUF_FREE_PENDING_SZ.
+ */

lib/librte_mbuf/rte_mbuf.c +3 -0
+static __rte_always_inline void
+__rte_pktmbuf_free_seg_via_array(struct rte_mbuf *m,
+	struct rte_mbuf ** const pending, unsigned int * const

nb_pending,

lib/librte_mbuf/rte_mbuf.c +17 -0
+	const unsigned int pending_sz)
+{
+	m = rte_pktmbuf_prefree_seg(m);
+	if (likely(m != NULL)) {
+		if (*nb_pending == pending_sz ||
+		    (*nb_pending > 0 && m->pool != pending[0]->pool)) {
+			rte_mempool_put_bulk(pending[0]->pool,
+					(void **)pending, *nb_pending);
+			*nb_pending = 0;
+		}
+
+		pending[(*nb_pending)++] = m;
+	}
+}
+
+/**
+ * Size of the array holding mbufs from the same membool pending to be

freed
+ * in bulk.
+ */

lib/librte_mbuf/rte_mbuf.c +4 -0
+#define RTE_PKTMBUF_FREE_PENDING_SZ 64
+
+/* Free a bulk of packet mbufs back into their original mempools. */
+void rte_pktmbuf_free_bulk(struct rte_mbuf **mbufs, unsigned int

count)

lib/librte_mbuf/rte_mbuf.c +2 -0
+{
+	struct rte_mbuf *m, *m_next,

*pending[RTE_PKTMBUF_FREE_PENDING_SZ];

lib/librte_mbuf/rte_mbuf.c +20 -0
+	unsigned int idx, nb_pending = 0;
+
+	for (idx = 0; idx < count; idx++) {
+		m = mbufs[idx];
+		if (unlikely(m == NULL))
+			continue;
+
+		__rte_mbuf_sanity_check(m, 1);
+
+		do {
+			m_next = m->next;
+			__rte_pktmbuf_free_seg_via_array(m,
+					pending, &nb_pending,
+					RTE_PKTMBUF_FREE_PENDING_SZ);
+			m = m_next;
+		} while (m != NULL);
+	}
+
+	if (nb_pending > 0)
+		rte_mempool_put_bulk(pending[0]->pool, (void **)pending,

nb_pending);

lib/librte_mbuf/rte_mbuf.c +2 -0
+}
+
 /* dump a mbuf on console */
 void
 rte_pktmbuf_dump(FILE *f, const struct rte_mbuf *m, unsigned dump_len)
lib/librte_mbuf/rte_mbuf.h +4 -0
index 98225ec80..a02c8b12c 100644
@@ -1907,6 +1907,21 @@ static inline void rte_pktmbuf_free(struct
 rte_mbuf *m)
 	}
 }
 
+/**
+ * Free a bulk of packet mbufs back into their original mempools.
+ *
+ * Free a bulk of mbufs, and all their segments in case of chained

buffers.
+ * Each segment is added back into its original mempool.
+ *
+ * @param mbufs
+ * Array of pointers to packet mbufs.
+ * The array may contain NULL pointers.
+ * @param count
+ * Array size.
+ */

lib/librte_mbuf/rte_mbuf.h +2 -0
+__rte_experimental
+void rte_pktmbuf_free_bulk(struct rte_mbuf **mbufs, unsigned int

count);

lib/librte_mbuf/rte_mbuf.h +1 -0
+
 /**
  * Creates a "clone" of the given packet mbuf.
  *
lib/librte_mbuf/rte_mbuf_version.map +1 -1
index 2662a37bf..cd6154ef2 100644
@@ -50,4 +50,5 @@ EXPERIMENTAL {
 	global:
 
 	rte_mbuf_check;
+	rte_pktmbuf_free_bulk;
 } DPDK_18.08;
--

2.17.1

#5 lore raw

PATCH v6 2/2 mbuf: add bulk free function

In reply to Morten Brørup

Add function for freeing a bulk of mbufs.

---

v6:
* Remove __rte_always_inline from static function.
The compiler will inline anyway.
v5:
* Rename variables from "free" to "pending" for improved readability.
* Add prefix __ to rte_pktmbuf_free_seg_via_array().
* Add array size parameter to __rte_pktmbuf_free_seg_via_array(). The compiler will optimize the parameter away anyway.
* Add description to __rte_pktmbuf_free_seg_via_array().
* Minor description updates.
v4:
* Mark as experimental by adding __rte_experimental.
* Add function to experimental section of map file.
* Fix source code formatting regarding pointer to pointer.
* Squash multiple modifications into one.
v3:
* Bugfix: Handle pakets with multiple segments.
* Add inline helper function, mainly for readability.
* Fix source code formatting regarding indentation.
v2:
* Function is not inline.
* Optimize to free multible mbufs belonging to the same mempool in bulk. Inspired by ixgbe_tx_free_bufs(), but allowing NULL pointers in the array, just like rte_pktmbuf_free() can take a NULL pointer.
* Use unsigned int instead of unsigned. Passes checkpatch, but mismatches the original coding style of the modified files.
* Fix a typo in the description headline: mempools is plural.

---
 lib/librte_mbuf/rte_mbuf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
lib/librte_mbuf/rte_mbuf.c +1 -1
index de5659baa..854af1104 100644
@@ -261,7 +261,7 @@ int rte_mbuf_check(const struct rte_mbuf *m, int is_header,
  *  Note: The compiler should optimize this parameter away when using a
  *  constant value, such as RTE_PKTMBUF_FREE_PENDING_SZ.
  */
-static __rte_always_inline void
+static void
 __rte_pktmbuf_free_seg_via_array(struct rte_mbuf *m,
 	struct rte_mbuf ** const pending, unsigned int * const nb_pending,
 	const unsigned int pending_sz)
Signature · 1 line
2.17.1