mirror of
https://github.com/Ysurac/openmptcprouter.git
synced 2025-03-09 15:40:20 +00:00
Update ECF patch
This commit is contained in:
Ycarus (Yannick Chabanois)
parent
29e7137a83
commit
6551f60889
1 changed files with 563 additions and 9 deletions
|
|
@ -1,7 +1,7 @@
|
|||
From 35f41229b58cb8c2611207827aa4f658b82db67e Mon Sep 17 00:00:00 2001
|
||||
From 025619486cf04c0beb9f395609d7711726fd63c6 Mon Sep 17 00:00:00 2001
|
||||
From: Daniel Weber <weberdaniel@gmx.net>
|
||||
Date: Mon, 5 Aug 2019 14:02:30 +0200
|
||||
Subject: [PATCH] mptcp: Earliest Completion First (ECF) Scheduler
|
||||
Subject: [PATCH 1/3] mptcp: Earliest Completion First (ECF) Scheduler
|
||||
|
||||
This scheduler works much like the default MPTCP scheduler. It always
|
||||
prefers the subflow with the smallest round-trip-time that is available.
|
||||
|
|
@ -15,7 +15,7 @@ Signed-off-by: Daniel Weber <weberdaniel@gmx.net>
|
|||
create mode 100644 net/mptcp/mptcp_ecf.c
|
||||
|
||||
diff --git a/net/mptcp/Kconfig b/net/mptcp/Kconfig
|
||||
index d22b7b47860f..dd1f859f1070 100644
|
||||
index 37f3af3db2a6..829ea084cf70 100644
|
||||
--- a/net/mptcp/Kconfig
|
||||
+++ b/net/mptcp/Kconfig
|
||||
@@ -109,6 +109,12 @@ config MPTCP_REDUNDANT
|
||||
|
|
@ -30,7 +30,7 @@ index d22b7b47860f..dd1f859f1070 100644
|
|||
+
|
||||
choice
|
||||
prompt "Default MPTCP Scheduler"
|
||||
default DEFAULT
|
||||
default DEFAULT_SCHEDULER
|
||||
diff --git a/net/mptcp/Makefile b/net/mptcp/Makefile
|
||||
index 82a2d4d945ae..369248a2f68e 100644
|
||||
--- a/net/mptcp/Makefile
|
||||
|
|
@ -44,7 +44,7 @@ index 82a2d4d945ae..369248a2f68e 100644
|
|||
mptcp-$(subst m,y,$(CONFIG_IPV6)) += mptcp_ipv6.o
|
||||
diff --git a/net/mptcp/mptcp_ecf.c b/net/mptcp/mptcp_ecf.c
|
||||
new file mode 100644
|
||||
index 000000000000..d61f4d2ad375
|
||||
index 000000000000..e0bd430a8943
|
||||
--- /dev/null
|
||||
+++ b/net/mptcp/mptcp_ecf.c
|
||||
@@ -0,0 +1,384 @@
|
||||
|
|
@ -73,9 +73,9 @@ index 000000000000..d61f4d2ad375
|
|||
+#include <net/mptcp.h>
|
||||
+#include <trace/events/tcp.h>
|
||||
+
|
||||
+static unsigned int r_beta __read_mostly = 4; /* beta = 1/r_beta = 0.25 */
|
||||
+module_param(r_beta, int, 0644);
|
||||
+MODULE_PARM_DESC(r_beta, "beta for ECF");
|
||||
+static unsigned int mptcp_ecf_r_beta __read_mostly = 4; /* beta = 1/r_beta = 0.25 */
|
||||
+module_param(mptcp_ecf_r_beta, int, 0644);
|
||||
+MODULE_PARM_DESC(mptcp_ecf_r_beta, "beta for ECF");
|
||||
+
|
||||
+struct ecfsched_priv {
|
||||
+ u32 last_rbuf_opti;
|
||||
|
|
@ -185,7 +185,7 @@ index 000000000000..d61f4d2ad375
|
|||
+ lhs = srtt_f * (x_f + cwnd_f * mss);
|
||||
+ rhs = cwnd_f * mss * (srtt_s + delta);
|
||||
+
|
||||
+ if (r_beta * lhs < r_beta * rhs + ecf_cb->switching_margin * rhs) {
|
||||
+ if (mptcp_ecf_r_beta * lhs < mptcp_ecf_r_beta * rhs + ecf_cb->switching_margin * rhs) {
|
||||
+ u32 x_s = sndbuf > cwnd_s * mss ? sndbuf : cwnd_s * mss;
|
||||
+ u64 lhs_s = srtt_s * x_s;
|
||||
+ u64 rhs_s = cwnd_s * mss * (2 * srtt_f + delta);
|
||||
|
|
@ -432,3 +432,557 @@ index 000000000000..d61f4d2ad375
|
|||
+MODULE_LICENSE("GPL");
|
||||
+MODULE_DESCRIPTION("ECF (Earliest Completion First) scheduler for MPTCP, based on default minimum RTT scheduler");
|
||||
+MODULE_VERSION("0.95");
|
||||
|
||||
From 5a9641c84cbb5a49749d7533c20035631985dbe7 Mon Sep 17 00:00:00 2001
|
||||
From: Daniel Weber <weberdaniel@gmx.net>
|
||||
Date: Mon, 9 Mar 2020 11:00:23 +0100
|
||||
Subject: [PATCH 2/3] mptcp: Reduce code-duplication for other schedulers
|
||||
|
||||
'mptcp_next_segment' now honors the function pointer to the actual part
|
||||
that makes the scheduling decision in 'sched_ops->get_subflow'. This
|
||||
allows for a better reuse by other schedulers.
|
||||
|
||||
The BLEST scheduler needs to adapt the direction of lambda value change
|
||||
depending on the occurrence of a retransmission. In order to remove the
|
||||
copied 'mptcp_rcv_buf_optimization' as well the scheduler now checks the
|
||||
tcp 'retrans_stamp' of the meta socket.
|
||||
|
||||
Signed-off-by: Daniel Weber <weberdaniel@gmx.net>
|
||||
---
|
||||
include/net/mptcp.h | 4 +
|
||||
net/mptcp/mptcp_blest.c | 200 +---------------------------------------
|
||||
net/mptcp/mptcp_sched.c | 9 +-
|
||||
3 files changed, 11 insertions(+), 202 deletions(-)
|
||||
|
||||
diff --git a/include/net/mptcp.h b/include/net/mptcp.h
|
||||
index 02312c9ea3a3..82f66ce206cc 100644
|
||||
--- a/include/net/mptcp.h
|
||||
+++ b/include/net/mptcp.h
|
||||
@@ -902,6 +902,10 @@ bool subflow_is_active(const struct tcp_sock *tp);
|
||||
bool subflow_is_backup(const struct tcp_sock *tp);
|
||||
struct sock *get_available_subflow(struct sock *meta_sk, struct sk_buff *skb,
|
||||
bool zero_wnd_test);
|
||||
+struct sk_buff *mptcp_next_segment(struct sock *meta_sk,
|
||||
+ int *reinject,
|
||||
+ struct sock **subsk,
|
||||
+ unsigned int *limit);
|
||||
extern struct mptcp_sched_ops mptcp_sched_default;
|
||||
|
||||
/* Initializes function-pointers and MPTCP-flags */
|
||||
diff --git a/net/mptcp/mptcp_blest.c b/net/mptcp/mptcp_blest.c
|
||||
index 40905a0d1fe5..22e25dd0d44e 100644
|
||||
--- a/net/mptcp/mptcp_blest.c
|
||||
+++ b/net/mptcp/mptcp_blest.c
|
||||
@@ -21,7 +21,6 @@
|
||||
|
||||
#include <linux/module.h>
|
||||
#include <net/mptcp.h>
|
||||
-#include <trace/events/tcp.h>
|
||||
|
||||
static unsigned char lambda __read_mostly = 12;
|
||||
module_param(lambda, byte, 0644);
|
||||
@@ -50,7 +49,6 @@ struct blestsched_priv {
|
||||
};
|
||||
|
||||
struct blestsched_cb {
|
||||
- bool retrans_flag;
|
||||
s16 lambda_1000; /* values range from min_lambda * 100 to max_lambda * 100 */
|
||||
u32 last_lambda_update;
|
||||
};
|
||||
@@ -77,14 +75,13 @@ static void blestsched_update_lambda(struct sock *meta_sk, struct sock *sk)
|
||||
* during the slow flows last RTT => increase lambda
|
||||
* otherwise decrease
|
||||
*/
|
||||
- if (blest_cb->retrans_flag) {
|
||||
+ if (tcp_sk(meta_sk)->retrans_stamp) {
|
||||
/* need to slow down on the slow flow */
|
||||
blest_cb->lambda_1000 += dyn_lambda_bad;
|
||||
} else {
|
||||
/* use the slow flow more */
|
||||
blest_cb->lambda_1000 -= dyn_lambda_good;
|
||||
}
|
||||
- blest_cb->retrans_flag = false;
|
||||
|
||||
/* cap lambda_1000 to its value range */
|
||||
blest_cb->lambda_1000 = min_t(s16, blest_cb->lambda_1000, max_lambda * 100);
|
||||
@@ -240,199 +237,6 @@ struct sock *blest_get_available_subflow(struct sock *meta_sk, struct sk_buff *s
|
||||
return bestsk;
|
||||
}
|
||||
|
||||
-/* copy from mptcp_sched.c: mptcp_rcv_buf_optimization */
|
||||
-static struct sk_buff *mptcp_blest_rcv_buf_optimization(struct sock *sk, int penal)
|
||||
-{
|
||||
- struct sock *meta_sk;
|
||||
- const struct tcp_sock *tp = tcp_sk(sk);
|
||||
- struct mptcp_tcp_sock *mptcp;
|
||||
- struct sk_buff *skb_head;
|
||||
- struct blestsched_priv *blest_p = blestsched_get_priv(tp);
|
||||
- struct blestsched_cb *blest_cb;
|
||||
-
|
||||
- meta_sk = mptcp_meta_sk(sk);
|
||||
- skb_head = tcp_rtx_queue_head(meta_sk);
|
||||
-
|
||||
- if (!skb_head)
|
||||
- return NULL;
|
||||
-
|
||||
- /* If penalization is optional (coming from mptcp_next_segment() and
|
||||
- * We are not send-buffer-limited we do not penalize. The retransmission
|
||||
- * is just an optimization to fix the idle-time due to the delay before
|
||||
- * we wake up the application.
|
||||
- */
|
||||
- if (!penal && sk_stream_memory_free(meta_sk))
|
||||
- goto retrans;
|
||||
-
|
||||
- /* Record the occurrence of a retransmission to update the lambda value */
|
||||
- blest_cb = blestsched_get_cb(tcp_sk(meta_sk));
|
||||
- blest_cb->retrans_flag = true;
|
||||
-
|
||||
- /* Only penalize again after an RTT has elapsed */
|
||||
- if (tcp_jiffies32 - blest_p->last_rbuf_opti < usecs_to_jiffies(tp->srtt_us >> 3))
|
||||
- goto retrans;
|
||||
-
|
||||
- /* Half the cwnd of the slow flows */
|
||||
- mptcp_for_each_sub(tp->mpcb, mptcp) {
|
||||
- struct tcp_sock *tp_it = mptcp->tp;
|
||||
-
|
||||
- if (tp_it != tp &&
|
||||
- TCP_SKB_CB(skb_head)->path_mask & mptcp_pi_to_flag(tp_it->mptcp->path_index)) {
|
||||
- if (tp->srtt_us < tp_it->srtt_us && inet_csk((struct sock *)tp_it)->icsk_ca_state == TCP_CA_Open) {
|
||||
- u32 prior_cwnd = tp_it->snd_cwnd;
|
||||
-
|
||||
- tp_it->snd_cwnd = max(tp_it->snd_cwnd >> 1U, 1U);
|
||||
-
|
||||
- /* If in slow start, do not reduce the ssthresh */
|
||||
- if (prior_cwnd >= tp_it->snd_ssthresh)
|
||||
- tp_it->snd_ssthresh = max(tp_it->snd_ssthresh >> 1U, 2U);
|
||||
-
|
||||
- blest_p->last_rbuf_opti = tcp_jiffies32;
|
||||
- }
|
||||
- }
|
||||
- }
|
||||
-
|
||||
-retrans:
|
||||
-
|
||||
- /* Segment not yet injected into this path? Take it!!! */
|
||||
- if (!(TCP_SKB_CB(skb_head)->path_mask & mptcp_pi_to_flag(tp->mptcp->path_index))) {
|
||||
- bool do_retrans = false;
|
||||
- mptcp_for_each_sub(tp->mpcb, mptcp) {
|
||||
- struct tcp_sock *tp_it = mptcp->tp;
|
||||
-
|
||||
- if (tp_it != tp &&
|
||||
- TCP_SKB_CB(skb_head)->path_mask & mptcp_pi_to_flag(tp_it->mptcp->path_index)) {
|
||||
- if (tp_it->snd_cwnd <= 4) {
|
||||
- do_retrans = true;
|
||||
- break;
|
||||
- }
|
||||
-
|
||||
- if (4 * tp->srtt_us >= tp_it->srtt_us) {
|
||||
- do_retrans = false;
|
||||
- break;
|
||||
- } else {
|
||||
- do_retrans = true;
|
||||
- }
|
||||
- }
|
||||
- }
|
||||
-
|
||||
- if (do_retrans && mptcp_is_available(sk, skb_head, false)) {
|
||||
- trace_mptcp_retransmit(sk, skb_head);
|
||||
- return skb_head;
|
||||
- }
|
||||
- }
|
||||
- return NULL;
|
||||
-}
|
||||
-
|
||||
-/* copy from mptcp_sched.c: __mptcp_next_segment */
|
||||
-/* Returns the next segment to be sent from the mptcp meta-queue.
|
||||
- * (chooses the reinject queue if any segment is waiting in it, otherwise,
|
||||
- * chooses the normal write queue).
|
||||
- * Sets *@reinject to 1 if the returned segment comes from the
|
||||
- * reinject queue. Sets it to 0 if it is the regular send-head of the meta-sk,
|
||||
- * and sets it to -1 if it is a meta-level retransmission to optimize the
|
||||
- * receive-buffer.
|
||||
- */
|
||||
-static struct sk_buff *__mptcp_blest_next_segment(struct sock *meta_sk, int *reinject)
|
||||
-{
|
||||
- const struct mptcp_cb *mpcb = tcp_sk(meta_sk)->mpcb;
|
||||
- struct sk_buff *skb = NULL;
|
||||
-
|
||||
- *reinject = 0;
|
||||
-
|
||||
- /* If we are in fallback-mode, just take from the meta-send-queue */
|
||||
- if (mpcb->infinite_mapping_snd || mpcb->send_infinite_mapping)
|
||||
- return tcp_send_head(meta_sk);
|
||||
-
|
||||
- skb = skb_peek(&mpcb->reinject_queue);
|
||||
-
|
||||
- if (skb) {
|
||||
- *reinject = 1;
|
||||
- } else {
|
||||
- skb = tcp_send_head(meta_sk);
|
||||
-
|
||||
- if (!skb && meta_sk->sk_socket &&
|
||||
- test_bit(SOCK_NOSPACE, &meta_sk->sk_socket->flags) &&
|
||||
- sk_stream_wspace(meta_sk) < sk_stream_min_wspace(meta_sk)) {
|
||||
- struct sock *subsk = blest_get_available_subflow(meta_sk, NULL,
|
||||
- false);
|
||||
- if (!subsk)
|
||||
- return NULL;
|
||||
-
|
||||
- skb = mptcp_blest_rcv_buf_optimization(subsk, 0);
|
||||
- if (skb)
|
||||
- *reinject = -1;
|
||||
- }
|
||||
- }
|
||||
- return skb;
|
||||
-}
|
||||
-
|
||||
-/* copy from mptcp_sched.c: mptcp_next_segment */
|
||||
-static struct sk_buff *mptcp_blest_next_segment(struct sock *meta_sk,
|
||||
- int *reinject,
|
||||
- struct sock **subsk,
|
||||
- unsigned int *limit)
|
||||
-{
|
||||
- struct sk_buff *skb = __mptcp_blest_next_segment(meta_sk, reinject);
|
||||
- unsigned int mss_now;
|
||||
- struct tcp_sock *subtp;
|
||||
- u16 gso_max_segs;
|
||||
- u32 max_len, max_segs, window, needed;
|
||||
-
|
||||
- /* As we set it, we have to reset it as well. */
|
||||
- *limit = 0;
|
||||
-
|
||||
- if (!skb)
|
||||
- return NULL;
|
||||
-
|
||||
- *subsk = blest_get_available_subflow(meta_sk, skb, false);
|
||||
- if (!*subsk)
|
||||
- return NULL;
|
||||
-
|
||||
- subtp = tcp_sk(*subsk);
|
||||
- mss_now = tcp_current_mss(*subsk);
|
||||
-
|
||||
- if (!*reinject && unlikely(!tcp_snd_wnd_test(tcp_sk(meta_sk), skb, mss_now))) {
|
||||
- skb = mptcp_blest_rcv_buf_optimization(*subsk, 1);
|
||||
- if (skb)
|
||||
- *reinject = -1;
|
||||
- else
|
||||
- return NULL;
|
||||
- }
|
||||
-
|
||||
- /* No splitting required, as we will only send one single segment */
|
||||
- if (skb->len <= mss_now)
|
||||
- return skb;
|
||||
-
|
||||
- /* The following is similar to tcp_mss_split_point, but
|
||||
- * we do not care about nagle, because we will anyways
|
||||
- * use TCP_NAGLE_PUSH, which overrides this.
|
||||
- *
|
||||
- * So, we first limit according to the cwnd/gso-size and then according
|
||||
- * to the subflow's window.
|
||||
- */
|
||||
-
|
||||
- gso_max_segs = (*subsk)->sk_gso_max_segs;
|
||||
- if (!gso_max_segs) /* No gso supported on the subflow's NIC */
|
||||
- gso_max_segs = 1;
|
||||
- max_segs = min_t(unsigned int, tcp_cwnd_test(subtp, skb), gso_max_segs);
|
||||
- if (!max_segs)
|
||||
- return NULL;
|
||||
-
|
||||
- max_len = mss_now * max_segs;
|
||||
- window = tcp_wnd_end(subtp) - subtp->write_seq;
|
||||
-
|
||||
- needed = min(skb->len, window);
|
||||
- if (max_len <= skb->len)
|
||||
- /* Take max_win, which is actually the cwnd/gso-size */
|
||||
- *limit = max_len;
|
||||
- else
|
||||
- /* Or, take the window */
|
||||
- *limit = needed;
|
||||
-
|
||||
- return skb;
|
||||
-}
|
||||
-
|
||||
static void blestsched_init(struct sock *sk)
|
||||
{
|
||||
struct blestsched_priv *blest_p = blestsched_get_priv(tcp_sk(sk));
|
||||
@@ -450,7 +254,7 @@ static void blestsched_init(struct sock *sk)
|
||||
|
||||
static struct mptcp_sched_ops mptcp_sched_blest = {
|
||||
.get_subflow = blest_get_available_subflow,
|
||||
- .next_segment = mptcp_blest_next_segment,
|
||||
+ .next_segment = mptcp_next_segment,
|
||||
.init = blestsched_init,
|
||||
.name = "blest",
|
||||
.owner = THIS_MODULE,
|
||||
diff --git a/net/mptcp/mptcp_sched.c b/net/mptcp/mptcp_sched.c
|
||||
index 18c3559b0d48..5bf2946a5caf 100644
|
||||
--- a/net/mptcp/mptcp_sched.c
|
||||
+++ b/net/mptcp/mptcp_sched.c
|
||||
@@ -372,8 +372,8 @@ static struct sk_buff *__mptcp_next_segment(struct sock *meta_sk, int *reinject)
|
||||
if (!skb && meta_sk->sk_socket &&
|
||||
test_bit(SOCK_NOSPACE, &meta_sk->sk_socket->flags) &&
|
||||
sk_stream_wspace(meta_sk) < sk_stream_min_wspace(meta_sk)) {
|
||||
- struct sock *subsk = get_available_subflow(meta_sk, NULL,
|
||||
- false);
|
||||
+ struct sock *subsk = mpcb->sched_ops->get_subflow(meta_sk, NULL,
|
||||
+ false);
|
||||
if (!subsk)
|
||||
return NULL;
|
||||
|
||||
@@ -385,7 +385,7 @@ static struct sk_buff *__mptcp_next_segment(struct sock *meta_sk, int *reinject)
|
||||
return skb;
|
||||
}
|
||||
|
||||
-static struct sk_buff *mptcp_next_segment(struct sock *meta_sk,
|
||||
+struct sk_buff *mptcp_next_segment(struct sock *meta_sk,
|
||||
int *reinject,
|
||||
struct sock **subsk,
|
||||
unsigned int *limit)
|
||||
@@ -402,7 +402,7 @@ static struct sk_buff *mptcp_next_segment(struct sock *meta_sk,
|
||||
if (!skb)
|
||||
return NULL;
|
||||
|
||||
- *subsk = get_available_subflow(meta_sk, skb, false);
|
||||
+ *subsk = tcp_sk(meta_sk)->mpcb->sched_ops->get_subflow(meta_sk, skb, false);
|
||||
if (!*subsk)
|
||||
return NULL;
|
||||
|
||||
@@ -449,6 +449,7 @@ static struct sk_buff *mptcp_next_segment(struct sock *meta_sk,
|
||||
|
||||
return skb;
|
||||
}
|
||||
+EXPORT_SYMBOL_GPL(mptcp_next_segment);
|
||||
|
||||
static void defsched_init(struct sock *sk)
|
||||
{
|
||||
|
||||
From 5e8425e43b38e7e0fe566ffd50e197c07807ebdf Mon Sep 17 00:00:00 2001
|
||||
From: Daniel Weber <weberdaniel@gmx.net>
|
||||
Date: Mon, 9 Mar 2020 11:09:27 +0100
|
||||
Subject: [PATCH 3/3] mptcp: Remove code-duplication from ECF scheduler
|
||||
|
||||
The ECF scheduler relies on large parts of the default scheduler. This
|
||||
commit removes the copied blocks and reuses 'mptcp_next_segment' and
|
||||
'mptcp_rcv_buf_optimization' directly from it via function pointers.
|
||||
|
||||
Signed-off-by: Daniel Weber <weberdaniel@gmx.net>
|
||||
---
|
||||
net/mptcp/mptcp_ecf.c | 191 +-----------------------------------------
|
||||
1 file changed, 1 insertion(+), 190 deletions(-)
|
||||
|
||||
diff --git a/net/mptcp/mptcp_ecf.c b/net/mptcp/mptcp_ecf.c
|
||||
index e0bd430a8943..6b976b2b0c72 100644
|
||||
--- a/net/mptcp/mptcp_ecf.c
|
||||
+++ b/net/mptcp/mptcp_ecf.c
|
||||
@@ -21,7 +21,6 @@
|
||||
|
||||
#include <linux/module.h>
|
||||
#include <net/mptcp.h>
|
||||
-#include <trace/events/tcp.h>
|
||||
|
||||
static unsigned int mptcp_ecf_r_beta __read_mostly = 4; /* beta = 1/r_beta = 0.25 */
|
||||
module_param(mptcp_ecf_r_beta, int, 0644);
|
||||
@@ -154,194 +153,6 @@ static struct sock *ecf_get_available_subflow(struct sock *meta_sk,
|
||||
return bestsk;
|
||||
}
|
||||
|
||||
-/* copy from mptcp_sched.c: mptcp_rcv_buf_optimization */
|
||||
-static struct sk_buff *mptcp_ecf_rcv_buf_optimization(struct sock *sk, int penal)
|
||||
-{
|
||||
- struct sock *meta_sk;
|
||||
- const struct tcp_sock *tp = tcp_sk(sk);
|
||||
- struct mptcp_tcp_sock *mptcp;
|
||||
- struct sk_buff *skb_head;
|
||||
- struct ecfsched_priv *ecf_p = ecfsched_get_priv(tp);
|
||||
-
|
||||
- meta_sk = mptcp_meta_sk(sk);
|
||||
- skb_head = tcp_rtx_queue_head(meta_sk);
|
||||
-
|
||||
- if (!skb_head)
|
||||
- return NULL;
|
||||
-
|
||||
- /* If penalization is optional (coming from mptcp_next_segment() and
|
||||
- * We are not send-buffer-limited we do not penalize. The retransmission
|
||||
- * is just an optimization to fix the idle-time due to the delay before
|
||||
- * we wake up the application.
|
||||
- */
|
||||
- if (!penal && sk_stream_memory_free(meta_sk))
|
||||
- goto retrans;
|
||||
-
|
||||
- /* Only penalize again after an RTT has elapsed */
|
||||
- if (tcp_jiffies32 - ecf_p->last_rbuf_opti < usecs_to_jiffies(tp->srtt_us >> 3))
|
||||
- goto retrans;
|
||||
-
|
||||
- /* Half the cwnd of the slow flows */
|
||||
- mptcp_for_each_sub(tp->mpcb, mptcp) {
|
||||
- struct tcp_sock *tp_it = mptcp->tp;
|
||||
-
|
||||
- if (tp_it != tp &&
|
||||
- TCP_SKB_CB(skb_head)->path_mask & mptcp_pi_to_flag(tp_it->mptcp->path_index)) {
|
||||
- if (tp->srtt_us < tp_it->srtt_us && inet_csk((struct sock *)tp_it)->icsk_ca_state == TCP_CA_Open) {
|
||||
- u32 prior_cwnd = tp_it->snd_cwnd;
|
||||
-
|
||||
- tp_it->snd_cwnd = max(tp_it->snd_cwnd >> 1U, 1U);
|
||||
-
|
||||
- /* If in slow start, do not reduce the ssthresh */
|
||||
- if (prior_cwnd >= tp_it->snd_ssthresh)
|
||||
- tp_it->snd_ssthresh = max(tp_it->snd_ssthresh >> 1U, 2U);
|
||||
-
|
||||
- ecf_p->last_rbuf_opti = tcp_jiffies32;
|
||||
- }
|
||||
- }
|
||||
- }
|
||||
-
|
||||
-retrans:
|
||||
-
|
||||
- /* Segment not yet injected into this path? Take it!!! */
|
||||
- if (!(TCP_SKB_CB(skb_head)->path_mask & mptcp_pi_to_flag(tp->mptcp->path_index))) {
|
||||
- bool do_retrans = false;
|
||||
- mptcp_for_each_sub(tp->mpcb, mptcp) {
|
||||
- struct tcp_sock *tp_it = mptcp->tp;
|
||||
-
|
||||
- if (tp_it != tp &&
|
||||
- TCP_SKB_CB(skb_head)->path_mask & mptcp_pi_to_flag(tp_it->mptcp->path_index)) {
|
||||
- if (tp_it->snd_cwnd <= 4) {
|
||||
- do_retrans = true;
|
||||
- break;
|
||||
- }
|
||||
-
|
||||
- if (4 * tp->srtt_us >= tp_it->srtt_us) {
|
||||
- do_retrans = false;
|
||||
- break;
|
||||
- } else {
|
||||
- do_retrans = true;
|
||||
- }
|
||||
- }
|
||||
- }
|
||||
-
|
||||
- if (do_retrans && mptcp_is_available(sk, skb_head, false)) {
|
||||
- trace_mptcp_retransmit(sk, skb_head);
|
||||
- return skb_head;
|
||||
- }
|
||||
- }
|
||||
- return NULL;
|
||||
-}
|
||||
-
|
||||
-/* copy from mptcp_sched.c: __mptcp_next_segment */
|
||||
-/* Returns the next segment to be sent from the mptcp meta-queue.
|
||||
- * (chooses the reinject queue if any segment is waiting in it, otherwise,
|
||||
- * chooses the normal write queue).
|
||||
- * Sets *@reinject to 1 if the returned segment comes from the
|
||||
- * reinject queue. Sets it to 0 if it is the regular send-head of the meta-sk,
|
||||
- * and sets it to -1 if it is a meta-level retransmission to optimize the
|
||||
- * receive-buffer.
|
||||
- */
|
||||
-static struct sk_buff *__mptcp_ecf_next_segment(struct sock *meta_sk, int *reinject)
|
||||
-{
|
||||
- const struct mptcp_cb *mpcb = tcp_sk(meta_sk)->mpcb;
|
||||
- struct sk_buff *skb = NULL;
|
||||
-
|
||||
- *reinject = 0;
|
||||
-
|
||||
- /* If we are in fallback-mode, just take from the meta-send-queue */
|
||||
- if (mpcb->infinite_mapping_snd || mpcb->send_infinite_mapping)
|
||||
- return tcp_send_head(meta_sk);
|
||||
-
|
||||
- skb = skb_peek(&mpcb->reinject_queue);
|
||||
-
|
||||
- if (skb) {
|
||||
- *reinject = 1;
|
||||
- } else {
|
||||
- skb = tcp_send_head(meta_sk);
|
||||
-
|
||||
- if (!skb && meta_sk->sk_socket &&
|
||||
- test_bit(SOCK_NOSPACE, &meta_sk->sk_socket->flags) &&
|
||||
- sk_stream_wspace(meta_sk) < sk_stream_min_wspace(meta_sk)) {
|
||||
- struct sock *subsk = ecf_get_available_subflow(meta_sk, NULL,
|
||||
- false);
|
||||
- if (!subsk)
|
||||
- return NULL;
|
||||
-
|
||||
- skb = mptcp_ecf_rcv_buf_optimization(subsk, 0);
|
||||
- if (skb)
|
||||
- *reinject = -1;
|
||||
- }
|
||||
- }
|
||||
- return skb;
|
||||
-}
|
||||
-
|
||||
-/* copy from mptcp_sched.c: mptcp_next_segment */
|
||||
-static struct sk_buff *mptcp_ecf_next_segment(struct sock *meta_sk,
|
||||
- int *reinject,
|
||||
- struct sock **subsk,
|
||||
- unsigned int *limit)
|
||||
-{
|
||||
- struct sk_buff *skb = __mptcp_ecf_next_segment(meta_sk, reinject);
|
||||
- unsigned int mss_now;
|
||||
- struct tcp_sock *subtp;
|
||||
- u16 gso_max_segs;
|
||||
- u32 max_len, max_segs, window, needed;
|
||||
-
|
||||
- /* As we set it, we have to reset it as well. */
|
||||
- *limit = 0;
|
||||
-
|
||||
- if (!skb)
|
||||
- return NULL;
|
||||
-
|
||||
- *subsk = ecf_get_available_subflow(meta_sk, skb, false);
|
||||
- if (!*subsk)
|
||||
- return NULL;
|
||||
-
|
||||
- subtp = tcp_sk(*subsk);
|
||||
- mss_now = tcp_current_mss(*subsk);
|
||||
-
|
||||
- if (!*reinject && unlikely(!tcp_snd_wnd_test(tcp_sk(meta_sk), skb, mss_now))) {
|
||||
- skb = mptcp_ecf_rcv_buf_optimization(*subsk, 1);
|
||||
- if (skb)
|
||||
- *reinject = -1;
|
||||
- else
|
||||
- return NULL;
|
||||
- }
|
||||
-
|
||||
- /* No splitting required, as we will only send one single segment */
|
||||
- if (skb->len <= mss_now)
|
||||
- return skb;
|
||||
-
|
||||
- /* The following is similar to tcp_mss_split_point, but
|
||||
- * we do not care about nagle, because we will anyways
|
||||
- * use TCP_NAGLE_PUSH, which overrides this.
|
||||
- *
|
||||
- * So, we first limit according to the cwnd/gso-size and then according
|
||||
- * to the subflow's window.
|
||||
- */
|
||||
-
|
||||
- gso_max_segs = (*subsk)->sk_gso_max_segs;
|
||||
- if (!gso_max_segs) /* No gso supported on the subflow's NIC */
|
||||
- gso_max_segs = 1;
|
||||
- max_segs = min_t(unsigned int, tcp_cwnd_test(subtp, skb), gso_max_segs);
|
||||
- if (!max_segs)
|
||||
- return NULL;
|
||||
-
|
||||
- max_len = mss_now * max_segs;
|
||||
- window = tcp_wnd_end(subtp) - subtp->write_seq;
|
||||
-
|
||||
- needed = min(skb->len, window);
|
||||
- if (max_len <= skb->len)
|
||||
- /* Take max_win, which is actually the cwnd/gso-size */
|
||||
- *limit = max_len;
|
||||
- else
|
||||
- /* Or, take the window */
|
||||
- *limit = needed;
|
||||
-
|
||||
- return skb;
|
||||
-}
|
||||
-
|
||||
static void ecfsched_init(struct sock *sk)
|
||||
{
|
||||
struct ecfsched_priv *ecf_p = ecfsched_get_priv(tcp_sk(sk));
|
||||
@@ -353,7 +164,7 @@ static void ecfsched_init(struct sock *sk)
|
||||
|
||||
struct mptcp_sched_ops mptcp_sched_ecf = {
|
||||
.get_subflow = ecf_get_available_subflow,
|
||||
- .next_segment = mptcp_ecf_next_segment,
|
||||
+ .next_segment = mptcp_next_segment,
|
||||
.init = ecfsched_init,
|
||||
.name = "ecf",
|
||||
.owner = THIS_MODULE,
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue