* [PATCH net-next v2 1/2] ip: add helpers to process in-order fragments faster.
2018-08-11 20:27 [PATCH net-next v2 0/2] ip: faster in-order IP fragments Peter Oskolkov
@ 2018-08-11 20:27 ` Peter Oskolkov
2018-08-11 20:27 ` [PATCH net-next v2 2/2] ip: process in-order fragments efficiently Peter Oskolkov
2018-08-12 0:54 ` [PATCH net-next v2 0/2] ip: faster in-order IP fragments David Miller
2 siblings, 0 replies; 4+ messages in thread
From: Peter Oskolkov @ 2018-08-11 20:27 UTC (permalink / raw)
To: David Miller, netdev; +Cc: Peter Oskolkov, Eric Dumazet, Florian Westphal
This patch introduces several helper functions/macros that will be
used in the follow-up patch. No runtime changes yet.
The new logic (fully implemented in the second patch) is as follows:
* Nodes in the rb-tree will now contain not single fragments, but lists
of consecutive fragments ("runs").
* At each point in time, the current "active" run at the tail is
maintained/tracked. Fragments that arrive in-order, adjacent
to the previous tail fragment, are added to this tail run without
triggering the re-balancing of the rb-tree.
* If a fragment arrives out of order with the offset _before_ the tail run,
it is inserted into the rb-tree as a single fragment.
* If a fragment arrives after the current tail fragment (with a gap),
it starts a new "tail" run, as is inserted into the rb-tree
at the end as the head of the new run.
skb->cb is used to store additional information
needed here (suggested by Eric Dumazet).
Reported-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Peter Oskolkov <posk@google.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Florian Westphal <fw@strlen.de>
---
include/net/inet_frag.h | 6 ++++
net/ipv4/ip_fragment.c | 73 +++++++++++++++++++++++++++++++++++++++++
2 files changed, 79 insertions(+)
diff --git a/include/net/inet_frag.h b/include/net/inet_frag.h
index b86d14528188..1662cbc0b46b 100644
--- a/include/net/inet_frag.h
+++ b/include/net/inet_frag.h
@@ -57,7 +57,9 @@ struct frag_v6_compare_key {
* @lock: spinlock protecting this frag
* @refcnt: reference count of the queue
* @fragments: received fragments head
+ * @rb_fragments: received fragments rb-tree root
* @fragments_tail: received fragments tail
+ * @last_run_head: the head of the last "run". see ip_fragment.c
* @stamp: timestamp of the last received fragment
* @len: total length of the original datagram
* @meat: length of received fragments so far
@@ -78,6 +80,7 @@ struct inet_frag_queue {
struct sk_buff *fragments; /* Used in IPv6. */
struct rb_root rb_fragments; /* Used in IPv4. */
struct sk_buff *fragments_tail;
+ struct sk_buff *last_run_head;
ktime_t stamp;
int len;
int meat;
@@ -113,6 +116,9 @@ void inet_frag_kill(struct inet_frag_queue *q);
void inet_frag_destroy(struct inet_frag_queue *q);
struct inet_frag_queue *inet_frag_find(struct netns_frags *nf, void *key);
+/* Free all skbs in the queue; return the sum of their truesizes. */
+unsigned int inet_frag_rbtree_purge(struct rb_root *root);
+
static inline void inet_frag_put(struct inet_frag_queue *q)
{
if (refcount_dec_and_test(&q->refcnt))
diff --git a/net/ipv4/ip_fragment.c b/net/ipv4/ip_fragment.c
index 7cb7ed761d8c..26ace9d2d976 100644
--- a/net/ipv4/ip_fragment.c
+++ b/net/ipv4/ip_fragment.c
@@ -57,6 +57,57 @@
*/
static const char ip_frag_cache_name[] = "ip4-frags";
+/* Use skb->cb to track consecutive/adjacent fragments coming at
+ * the end of the queue. Nodes in the rb-tree queue will
+ * contain "runs" of one or more adjacent fragments.
+ *
+ * Invariants:
+ * - next_frag is NULL at the tail of a "run";
+ * - the head of a "run" has the sum of all fragment lengths in frag_run_len.
+ */
+struct ipfrag_skb_cb {
+ struct inet_skb_parm h;
+ struct sk_buff *next_frag;
+ int frag_run_len;
+};
+
+#define FRAG_CB(skb) ((struct ipfrag_skb_cb *)((skb)->cb))
+
+static void ip4_frag_init_run(struct sk_buff *skb)
+{
+ BUILD_BUG_ON(sizeof(struct ipfrag_skb_cb) > sizeof(skb->cb));
+
+ FRAG_CB(skb)->next_frag = NULL;
+ FRAG_CB(skb)->frag_run_len = skb->len;
+}
+
+/* Append skb to the last "run". */
+static void ip4_frag_append_to_last_run(struct inet_frag_queue *q,
+ struct sk_buff *skb)
+{
+ RB_CLEAR_NODE(&skb->rbnode);
+ FRAG_CB(skb)->next_frag = NULL;
+
+ FRAG_CB(q->last_run_head)->frag_run_len += skb->len;
+ FRAG_CB(q->fragments_tail)->next_frag = skb;
+ q->fragments_tail = skb;
+}
+
+/* Create a new "run" with the skb. */
+static void ip4_frag_create_run(struct inet_frag_queue *q, struct sk_buff *skb)
+{
+ if (q->last_run_head)
+ rb_link_node(&skb->rbnode, &q->last_run_head->rbnode,
+ &q->last_run_head->rbnode.rb_right);
+ else
+ rb_link_node(&skb->rbnode, NULL, &q->rb_fragments.rb_node);
+ rb_insert_color(&skb->rbnode, &q->rb_fragments);
+
+ ip4_frag_init_run(skb);
+ q->fragments_tail = skb;
+ q->last_run_head = skb;
+}
+
/* Describe an entry in the "incomplete datagrams" queue. */
struct ipq {
struct inet_frag_queue q;
@@ -654,6 +705,28 @@ struct sk_buff *ip_check_defrag(struct net *net, struct sk_buff *skb, u32 user)
}
EXPORT_SYMBOL(ip_check_defrag);
+unsigned int inet_frag_rbtree_purge(struct rb_root *root)
+{
+ struct rb_node *p = rb_first(root);
+ unsigned int sum = 0;
+
+ while (p) {
+ struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
+
+ p = rb_next(p);
+ rb_erase(&skb->rbnode, root);
+ while (skb) {
+ struct sk_buff *next = FRAG_CB(skb)->next_frag;
+
+ sum += skb->truesize;
+ kfree_skb(skb);
+ skb = next;
+ }
+ }
+ return sum;
+}
+EXPORT_SYMBOL(inet_frag_rbtree_purge);
+
#ifdef CONFIG_SYSCTL
static int dist_min;
--
2.18.0.597.ga71716f1ad-goog
^ permalink raw reply related [flat|nested] 4+ messages in thread* [PATCH net-next v2 2/2] ip: process in-order fragments efficiently
2018-08-11 20:27 [PATCH net-next v2 0/2] ip: faster in-order IP fragments Peter Oskolkov
2018-08-11 20:27 ` [PATCH net-next v2 1/2] ip: add helpers to process in-order fragments faster Peter Oskolkov
@ 2018-08-11 20:27 ` Peter Oskolkov
2018-08-12 0:54 ` [PATCH net-next v2 0/2] ip: faster in-order IP fragments David Miller
2 siblings, 0 replies; 4+ messages in thread
From: Peter Oskolkov @ 2018-08-11 20:27 UTC (permalink / raw)
To: David Miller, netdev; +Cc: Peter Oskolkov, Eric Dumazet, Florian Westphal
This patch changes the runtime behavior of IP defrag queue:
incoming in-order fragments are added to the end of the current
list/"run" of in-order fragments at the tail.
On some workloads, UDP stream performance is substantially improved:
RX: ./udp_stream -F 10 -T 2 -l 60
TX: ./udp_stream -c -H <host> -F 10 -T 5 -l 60
with this patchset applied on a 10Gbps receiver:
throughput=9524.18
throughput_units=Mbit/s
upstream (net-next):
throughput=4608.93
throughput_units=Mbit/s
Reported-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Peter Oskolkov <posk@google.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Florian Westphal <fw@strlen.de>
---
net/ipv4/inet_fragment.c | 2 +-
net/ipv4/ip_fragment.c | 110 ++++++++++++++++++++++++---------------
2 files changed, 70 insertions(+), 42 deletions(-)
diff --git a/net/ipv4/inet_fragment.c b/net/ipv4/inet_fragment.c
index 6d258a5669e7..bcb11f3a27c0 100644
--- a/net/ipv4/inet_fragment.c
+++ b/net/ipv4/inet_fragment.c
@@ -146,7 +146,7 @@ void inet_frag_destroy(struct inet_frag_queue *q)
fp = xp;
} while (fp);
} else {
- sum_truesize = skb_rbtree_purge(&q->rb_fragments);
+ sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments);
}
sum = sum_truesize + f->qsize;
diff --git a/net/ipv4/ip_fragment.c b/net/ipv4/ip_fragment.c
index 26ace9d2d976..88281fbce88c 100644
--- a/net/ipv4/ip_fragment.c
+++ b/net/ipv4/ip_fragment.c
@@ -126,8 +126,8 @@ static u8 ip4_frag_ecn(u8 tos)
static struct inet_frags ip4_frags;
-static int ip_frag_reasm(struct ipq *qp, struct sk_buff *prev,
- struct net_device *dev);
+static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
+ struct sk_buff *prev_tail, struct net_device *dev);
static void ip4_frag_init(struct inet_frag_queue *q, const void *a)
@@ -219,7 +219,12 @@ static void ip_expire(struct timer_list *t)
head = skb_rb_first(&qp->q.rb_fragments);
if (!head)
goto out;
- rb_erase(&head->rbnode, &qp->q.rb_fragments);
+ if (FRAG_CB(head)->next_frag)
+ rb_replace_node(&head->rbnode,
+ &FRAG_CB(head)->next_frag->rbnode,
+ &qp->q.rb_fragments);
+ else
+ rb_erase(&head->rbnode, &qp->q.rb_fragments);
memset(&head->rbnode, 0, sizeof(head->rbnode));
barrier();
}
@@ -320,7 +325,7 @@ static int ip_frag_reinit(struct ipq *qp)
return -ETIMEDOUT;
}
- sum_truesize = skb_rbtree_purge(&qp->q.rb_fragments);
+ sum_truesize = inet_frag_rbtree_purge(&qp->q.rb_fragments);
sub_frag_mem_limit(qp->q.net, sum_truesize);
qp->q.flags = 0;
@@ -329,6 +334,7 @@ static int ip_frag_reinit(struct ipq *qp)
qp->q.fragments = NULL;
qp->q.rb_fragments = RB_ROOT;
qp->q.fragments_tail = NULL;
+ qp->q.last_run_head = NULL;
qp->iif = 0;
qp->ecn = 0;
@@ -340,7 +346,7 @@ static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
{
struct net *net = container_of(qp->q.net, struct net, ipv4.frags);
struct rb_node **rbn, *parent;
- struct sk_buff *skb1;
+ struct sk_buff *skb1, *prev_tail;
struct net_device *dev;
unsigned int fragsize;
int flags, offset;
@@ -418,38 +424,41 @@ static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
*/
/* Find out where to put this fragment. */
- skb1 = qp->q.fragments_tail;
- if (!skb1) {
- /* This is the first fragment we've received. */
- rb_link_node(&skb->rbnode, NULL, &qp->q.rb_fragments.rb_node);
- qp->q.fragments_tail = skb;
- } else if ((skb1->ip_defrag_offset + skb1->len) < end) {
- /* This is the common/special case: skb goes to the end. */
+ prev_tail = qp->q.fragments_tail;
+ if (!prev_tail)
+ ip4_frag_create_run(&qp->q, skb); /* First fragment. */
+ else if (prev_tail->ip_defrag_offset + prev_tail->len < end) {
+ /* This is the common case: skb goes to the end. */
/* Detect and discard overlaps. */
- if (offset < (skb1->ip_defrag_offset + skb1->len))
+ if (offset < prev_tail->ip_defrag_offset + prev_tail->len)
goto discard_qp;
- /* Insert after skb1. */
- rb_link_node(&skb->rbnode, &skb1->rbnode, &skb1->rbnode.rb_right);
- qp->q.fragments_tail = skb;
+ if (offset == prev_tail->ip_defrag_offset + prev_tail->len)
+ ip4_frag_append_to_last_run(&qp->q, skb);
+ else
+ ip4_frag_create_run(&qp->q, skb);
} else {
- /* Binary search. Note that skb can become the first fragment, but
- * not the last (covered above). */
+ /* Binary search. Note that skb can become the first fragment,
+ * but not the last (covered above).
+ */
rbn = &qp->q.rb_fragments.rb_node;
do {
parent = *rbn;
skb1 = rb_to_skb(parent);
if (end <= skb1->ip_defrag_offset)
rbn = &parent->rb_left;
- else if (offset >= skb1->ip_defrag_offset + skb1->len)
+ else if (offset >= skb1->ip_defrag_offset +
+ FRAG_CB(skb1)->frag_run_len)
rbn = &parent->rb_right;
else /* Found an overlap with skb1. */
goto discard_qp;
} while (*rbn);
/* Here we have parent properly set, and rbn pointing to
- * one of its NULL left/right children. Insert skb. */
+ * one of its NULL left/right children. Insert skb.
+ */
+ ip4_frag_init_run(skb);
rb_link_node(&skb->rbnode, parent, rbn);
+ rb_insert_color(&skb->rbnode, &qp->q.rb_fragments);
}
- rb_insert_color(&skb->rbnode, &qp->q.rb_fragments);
if (dev)
qp->iif = dev->ifindex;
@@ -476,7 +485,7 @@ static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
unsigned long orefdst = skb->_skb_refdst;
skb->_skb_refdst = 0UL;
- err = ip_frag_reasm(qp, skb, dev);
+ err = ip_frag_reasm(qp, skb, prev_tail, dev);
skb->_skb_refdst = orefdst;
return err;
}
@@ -495,7 +504,7 @@ static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
/* Build a new IP datagram from all its fragments. */
static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
- struct net_device *dev)
+ struct sk_buff *prev_tail, struct net_device *dev)
{
struct net *net = container_of(qp->q.net, struct net, ipv4.frags);
struct iphdr *iph;
@@ -519,10 +528,16 @@ static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
fp = skb_clone(skb, GFP_ATOMIC);
if (!fp)
goto out_nomem;
- rb_replace_node(&skb->rbnode, &fp->rbnode, &qp->q.rb_fragments);
+ FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag;
+ if (RB_EMPTY_NODE(&skb->rbnode))
+ FRAG_CB(prev_tail)->next_frag = fp;
+ else
+ rb_replace_node(&skb->rbnode, &fp->rbnode,
+ &qp->q.rb_fragments);
if (qp->q.fragments_tail == skb)
qp->q.fragments_tail = fp;
skb_morph(skb, head);
+ FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag;
rb_replace_node(&head->rbnode, &skb->rbnode,
&qp->q.rb_fragments);
consume_skb(head);
@@ -558,7 +573,7 @@ static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
clone->len = clone->data_len = head->data_len - plen;
- skb->truesize += clone->truesize;
+ head->truesize += clone->truesize;
clone->csum = 0;
clone->ip_summed = head->ip_summed;
add_frag_mem_limit(qp->q.net, clone->truesize);
@@ -571,24 +586,36 @@ static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
skb_push(head, head->data - skb_network_header(head));
/* Traverse the tree in order, to build frag_list. */
+ fp = FRAG_CB(head)->next_frag;
rbn = rb_next(&head->rbnode);
rb_erase(&head->rbnode, &qp->q.rb_fragments);
- while (rbn) {
- struct rb_node *rbnext = rb_next(rbn);
- fp = rb_to_skb(rbn);
- rb_erase(rbn, &qp->q.rb_fragments);
- rbn = rbnext;
- *nextp = fp;
- nextp = &fp->next;
- fp->prev = NULL;
- memset(&fp->rbnode, 0, sizeof(fp->rbnode));
- head->data_len += fp->len;
- head->len += fp->len;
- if (head->ip_summed != fp->ip_summed)
- head->ip_summed = CHECKSUM_NONE;
- else if (head->ip_summed == CHECKSUM_COMPLETE)
- head->csum = csum_add(head->csum, fp->csum);
- head->truesize += fp->truesize;
+ while (rbn || fp) {
+ /* fp points to the next sk_buff in the current run;
+ * rbn points to the next run.
+ */
+ /* Go through the current run. */
+ while (fp) {
+ *nextp = fp;
+ nextp = &fp->next;
+ fp->prev = NULL;
+ memset(&fp->rbnode, 0, sizeof(fp->rbnode));
+ head->data_len += fp->len;
+ head->len += fp->len;
+ if (head->ip_summed != fp->ip_summed)
+ head->ip_summed = CHECKSUM_NONE;
+ else if (head->ip_summed == CHECKSUM_COMPLETE)
+ head->csum = csum_add(head->csum, fp->csum);
+ head->truesize += fp->truesize;
+ fp = FRAG_CB(fp)->next_frag;
+ }
+ /* Move to the next run. */
+ if (rbn) {
+ struct rb_node *rbnext = rb_next(rbn);
+
+ fp = rb_to_skb(rbn);
+ rb_erase(rbn, &qp->q.rb_fragments);
+ rbn = rbnext;
+ }
}
sub_frag_mem_limit(qp->q.net, head->truesize);
@@ -624,6 +651,7 @@ static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
qp->q.fragments = NULL;
qp->q.rb_fragments = RB_ROOT;
qp->q.fragments_tail = NULL;
+ qp->q.last_run_head = NULL;
return 0;
out_nomem:
--
2.18.0.597.ga71716f1ad-goog
^ permalink raw reply related [flat|nested] 4+ messages in thread