Skip to contentSkip to conversation

add bpf_stale scheduler

PATCH In review Geliang Tang 3 participants 9 messages Aug 18, 2023 – Sep 9, 2023

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

v3 cover add bpf_stale scheduler

v3:
- init sk_storage in .init, delete it in .release.
We invoke bpf_sk_storage_get() to get the sk_storage map many times. Only the first call is the slow path (see bpf_local_storage_lookup in kernel/bpf/bpf_local_storage.c), alloc the map and cache it. The subsequent calls are all in fast path, the cache hits. So we should first call bpf_sk_storage_get in .init, then call it many times in .get_subflow.
- if no subflow is scheduled, it should return '-1' to indicate an error.

v2:
- store subflow ids instead of storing subflow pointers in sk_storage.

v1:
- This patchset adds the new bpf_stale scheduler. Use sk_storage to save the stale map instead of using subflow->stale flag to manage it.

Geliang Tang (5):
Squash to "selftests/bpf: Add bpf_bkup scheduler" Squash to "selftests/bpf: Add bpf_rr scheduler" Squash to "selftests/bpf: Add bpf_burst scheduler" selftests/bpf: Add bpf_stale scheduler selftests/bpf: Add bpf_stale test

 tools/testing/selftests/bpf/bpf_tcp_helpers.h |   1 +
 .../testing/selftests/bpf/prog_tests/mptcp.c  |  38 +++++
 .../selftests/bpf/progs/mptcp_bpf_bkup.c      |   4 +-
 .../selftests/bpf/progs/mptcp_bpf_burst.c     |   2 +
 .../selftests/bpf/progs/mptcp_bpf_rr.c        |   2 +
 .../selftests/bpf/progs/mptcp_bpf_stale.c     | 152 ++++++++++++++++++
 6 files changed, 198 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c
Signature · 1 line
2.35.3
#1 lore raw

PATCH v3 1/5 Squash to "selftests/bpf: Add bpf_bkup scheduler"

Return -1 when no subflow is selected.

---
 tools/testing/selftests/bpf/progs/mptcp_bpf_bkup.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
tools/testing/selftests/bpf/progs/mptcp_bpf_bkup.c +3 -1
index 7e66ab839d14..904186fb6750 100644
@@ -36,8 +36,10 @@ int BPF_STRUCT_OPS(bpf_bkup_get_subflow, struct mptcp_sock *msk,
 		}
 	}
 
-	if (nr != -1)
+	if (nr != -1) {
 		mptcp_subflow_set_scheduled(mptcp_subflow_ctx_by_pos(data, nr), true);
+		return -1;
+	}
 	return 0;
 }
 
Signature · 1 line
2.35.3
#2 lore raw

PATCH v3 2/5 Squash to "selftests/bpf: Add bpf_rr scheduler"

In reply to Geliang Tang

Alloc and init mptcp_rr_map first in mptcp_sched_rr_init().

---
 tools/testing/selftests/bpf/progs/mptcp_bpf_rr.c | 2 ++
 1 file changed, 2 insertions(+)
tools/testing/selftests/bpf/progs/mptcp_bpf_rr.c +2 -0
index 7a5c058d2408..347ffad90860 100644
@@ -20,6 +20,8 @@ struct {
 SEC("struct_ops/mptcp_sched_rr_init")
 void BPF_PROG(mptcp_sched_rr_init, struct mptcp_sock *msk)
 {
+	bpf_sk_storage_get(&mptcp_rr_map, msk, 0,
+			   BPF_LOCAL_STORAGE_GET_F_CREATE);
 }
 
 SEC("struct_ops/mptcp_sched_rr_release")
Signature · 1 line
2.35.3
#3 lore raw

PATCH v3 3/5 Squash to "selftests/bpf: Add bpf_burst scheduler"

In reply to Geliang Tang

Alloc and init sk_storage in data_init().

---
 tools/testing/selftests/bpf/progs/mptcp_bpf_burst.c | 2 ++
 1 file changed, 2 insertions(+)
tools/testing/selftests/bpf/progs/mptcp_bpf_burst.c +2 -0
index 2962067568e7..b6a8a051741d 100644
@@ -69,6 +69,8 @@ static __always_inline bool sk_stream_memory_free(const struct sock *sk)
 SEC("struct_ops/mptcp_sched_burst_init")
 void BPF_PROG(mptcp_sched_burst_init, struct mptcp_sock *msk)
 {
+	bpf_sk_storage_get(&mptcp_burst_map, msk, 0,
+			   BPF_LOCAL_STORAGE_GET_F_CREATE);
 }
 
 SEC("struct_ops/mptcp_sched_burst_release")
Signature · 1 line
2.35.3
#4 lore raw

PATCH v3 4/5 selftests/bpf: Add bpf_stale scheduler

In reply to Geliang Tang

This patch implements the setting a subflow as stale/unstale in BPF MPTCP scheduler, named bpf_stale. The staled subflow id will be added into a map in sk_storage.

Two helper mptcp_subflow_set_stale() and mptcp_subflow_clear_stale() are added.

In this test, subflow 1 is set as stale in bpf_stale_data_init(). Each subflow is checked whether it's a stale one in bpf_stale_get_subflow() to select a unstale subflow to send data.

---
 tools/testing/selftests/bpf/bpf_tcp_helpers.h |   1 +
 .../selftests/bpf/progs/mptcp_bpf_stale.c     | 152 ++++++++++++++++++
 2 files changed, 153 insertions(+)
 create mode 100644 tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c
tools/testing/selftests/bpf/bpf_tcp_helpers.h +1 -0
index b687f91f2da8..33246629fa36 100644
@@ -239,6 +239,7 @@ struct mptcp_subflow_context {
 	unsigned long avg_pacing_rate;
 	__u32	backup : 1;
 	__u8	stale_count;
+	__u32	subflow_id;
 	struct	sock *tcp_sock;	    /* tcp sk backpointer */
 } __attribute__((preserve_access_index));
 
tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c new file +152 -0

152 lines not on this page · raw message

Signature · 1 line
2.35.3
maintainer
#5 lore raw

Re: [PATCH mptcp-next v3 4/5] selftests/bpf: Add bpf_stale scheduler

Quoted 140 lines · Geliang Tang

This patch implements the setting a subflow as stale/unstale in BPF MPTCP scheduler, named bpf_stale. The staled subflow id will be added into a map in sk_storage.

Two helper mptcp_subflow_set_stale() and mptcp_subflow_clear_stale() are added.

In this test, subflow 1 is set as stale in bpf_stale_data_init(). Each subflow is checked whether it's a stale one in bpf_stale_get_subflow() to select a unstale subflow to send data.

---
tools/testing/selftests/bpf/bpf_tcp_helpers.h | 1 +
.../selftests/bpf/progs/mptcp_bpf_stale.c | 152 ++++++++++++++++++

2 files changed, 153 insertions(+)

create mode 100644 tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c

tools/testing/selftests/bpf/bpf_tcp_helpers.h +1 -0
index b687f91f2da8..33246629fa36 100644
@@ -239,6 +239,7 @@ struct mptcp_subflow_context {
 	unsigned long avg_pacing_rate;
 	__u32	backup : 1;
 	__u8	stale_count;
+	__u32	subflow_id;
 	struct	sock *tcp_sock;	    /* tcp sk backpointer */
 } __attribute__((preserve_access_index));
 
tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c new file +104 -0
new file mode 100644
index 000000000000..08c857f79221
@@ -0,0 +1,152 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2023, SUSE. */
+
+#include <linux/bpf.h>
+#include "bpf_tcp_helpers.h"
+
+char _license[] SEC("license") = "GPL";
+
+struct mptcp_stale_storage {
+	__u8 nr;
+	__u32 ids[MPTCP_SUBFLOWS_MAX];
+};
+
+struct {
+	__uint(type, BPF_MAP_TYPE_SK_STORAGE);
+	__uint(map_flags, BPF_F_NO_PREALLOC);
+	__type(key, int);
+	__type(value, struct mptcp_stale_storage);
+} mptcp_stale_map SEC(".maps");
+
+static void mptcp_subflow_set_stale(struct mptcp_stale_storage *storage,
+				    __u32 subflow_id)
+{
+	if (!subflow_id)
+		return;
+
+	for (int i = 0; i < storage->nr && i < MPTCP_SUBFLOWS_MAX; i++) {
+		if (storage->ids[i] == subflow_id)
+			return;
+	}
+
+	if (storage->nr < MPTCP_SUBFLOWS_MAX - 1)
+		storage->ids[storage->nr++] = subflow_id;
+}
+
+static void mptcp_subflow_clear_stale(struct mptcp_stale_storage *storage,
+				      __u32 subflow_id)
+{
+	if (!subflow_id)
+		return;
+
+	for (int i = 0; i < storage->nr && i < MPTCP_SUBFLOWS_MAX; i++) {
+		if (storage->ids[i] == subflow_id) {
+			for (int j = i; j < MPTCP_SUBFLOWS_MAX - 1; j++) {
+				if (!storage->ids[j + 1])
+					break;
+				storage->ids[j] = storage->ids[j + 1];
+				storage->ids[j + 1] = 0;
+			}
+			storage->nr--;
+			return;
+		}
+	}
+}
+
+static bool mptcp_subflow_is_stale(struct mptcp_stale_storage *storage,
+				   __u32 subflow_id)
+{
+	for (int i = 0; i < storage->nr && i < MPTCP_SUBFLOWS_MAX; i++) {
+		if (storage->ids[i] == subflow_id)
+			return true;
+	}
+
+	return false;
+}
+
+static bool mptcp_subflow_is_active(struct mptcp_sched_data *data,
+				    __u32 subflow_id)
+{
+	for (int i = 0; i < data->subflows && i < MPTCP_SUBFLOWS_MAX; i++) {
+		struct mptcp_subflow_context *subflow;
+
+		subflow = mptcp_subflow_ctx_by_pos(data, i);
+		if (!subflow)
+			break;
+		if (subflow->subflow_id == subflow_id)
+			return true;
+	}
+
+	return false;
+}
+
+SEC("struct_ops/mptcp_sched_stale_init")
+void BPF_PROG(mptcp_sched_stale_init, struct mptcp_sock *msk)
+{
+	struct mptcp_stale_storage *storage;
+
+	storage = bpf_sk_storage_get(&mptcp_stale_map, msk, 0,
+				     BPF_LOCAL_STORAGE_GET_F_CREATE);
+	if (!storage)
+		return;
+
+	storage->nr = 0;
+}
+
+SEC("struct_ops/mptcp_sched_stale_release")
+void BPF_PROG(mptcp_sched_stale_release, struct mptcp_sock *msk)
+{
+	bpf_sk_storage_delete(&mptcp_stale_map, msk);
+}
+
+int BPF_STRUCT_OPS(bpf_stale_get_subflow, struct mptcp_sock *msk,
+		   struct mptcp_sched_data *data)
+{

tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c line 114

Quoted 10 lines
tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c new file +10 -0
+	struct mptcp_subflow_context *subflow;
+	struct mptcp_stale_storage *storage;
+	int nr = -1;
+
+	storage = bpf_sk_storage_get(&mptcp_stale_map, msk, 0,
+				     BPF_LOCAL_STORAGE_GET_F_CREATE);
+	if (!storage)
+		return -1;
+
+	mptcp_sched_data_set_contexts(msk, data);

Should this call be moved to sched.c, right before the calls to msk->sched->get_subflow() in the mptcp_sched_get functions?

It looks like every BPF scheduler example has to call this, so it would be more efficient to set up the data before calling the BPF functions.

tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c line 124

Quoted 10 lines
tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c new file +10 -0
+
+	/* Handle invalid subflow ids for subflows that have been closed */
+	for (int i = 0; i < storage->nr && i < MPTCP_SUBFLOWS_MAX; i++) {
+		if (!mptcp_subflow_is_active(data, storage->ids[i]))
+			mptcp_subflow_clear_stale(storage, storage->ids[i]);
+	}
+
+	subflow = mptcp_subflow_ctx_by_pos(data, 1);
+	if (subflow)
+		mptcp_subflow_set_stale(storage, subflow->subflow_id);

Since this is always marking the subflow in position 1 as stale...

tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c line 133

Quoted 9 lines
tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c new file +9 -0
+
+	for (int i = 0; i < data->subflows && i < MPTCP_SUBFLOWS_MAX; i++) {
+		struct mptcp_subflow_context *subflow;
+
+		subflow = mptcp_subflow_ctx_by_pos(data, i);
+		if (!subflow)
+			break;
+
+		if (mptcp_subflow_is_stale(storage, subflow->subflow_id))

...and then looking up that subflow again, this scheduler is not doing anything with actual stale subflows. The other BPF schedulers are already exercising everything in the BPF infrastructure that is done here (with sk_storage, etc.).

If there's a way to check for real stuck/stale subflows, then I think it would be good to have a sample scheduler that does that. But for now, I don't think this code adds test coverage or helps demonstrate the BPF scheduler.

Quoted 24 lines
tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c new file +19 -0
+			continue;
+
+		nr = i;
+	}
+
+	if (nr != -1) {
+		mptcp_subflow_set_scheduled(mptcp_subflow_ctx_by_pos(data, nr), true);
+		return -1;
+	}
+	return 0;
+}
+
+SEC(".struct_ops")
+struct mptcp_sched_ops stale = {
+	.init		= (void *)mptcp_sched_stale_init,
+	.release	= (void *)mptcp_sched_stale_release,
+	.get_subflow	= (void *)bpf_stale_get_subflow,
+	.name		= "bpf_stale",
+};
Signature · 1 line
2.35.3
#6 lore raw

PATCH v3 5/5 selftests/bpf: Add bpf_stale test

In reply to Geliang Tang

This patch adds the bpf_stale scheduler test: test_stale(). Use sysctl to set net.mptcp.scheduler to use this sched. Add two veth net devices to simulate the multiple addresses case. Use 'ip mptcp endpoint' command to add the new endpoint ADDR_2 to PM netlink. Send data and check bytes_sent of 'ss' output after it to make sure the data has been only sent on ADDR_1 since ADDR_2 is set as stale.

---
 .../testing/selftests/bpf/prog_tests/mptcp.c  | 38 +++++++++++++++++++
 1 file changed, 38 insertions(+)
tools/testing/selftests/bpf/prog_tests/mptcp.c +38 -0
index 3879eaa4fb32..08a6c4c5184a 100644
@@ -15,6 +15,7 @@
 #include "mptcp_bpf_rr.skel.h"
 #include "mptcp_bpf_red.skel.h"
 #include "mptcp_bpf_burst.skel.h"
+#include "mptcp_bpf_stale.skel.h"
 
 #define NS_TEST "mptcp_ns"
 
@@ -651,6 +652,41 @@ static void test_burst(void)
 	mptcp_bpf_burst__destroy(burst_skel);
 }
 
+static void test_stale(void)
+{
+	struct mptcp_bpf_stale *stale_skel;
+	int server_fd, client_fd;
+	struct nstoken *nstoken;
+	struct bpf_link *link;
+
+	stale_skel = mptcp_bpf_stale__open_and_load();
+	if (!ASSERT_OK_PTR(stale_skel, "bpf_stale__open_and_load"))
+		return;
+
+	link = bpf_map__attach_struct_ops(stale_skel->maps.stale);
+	if (!ASSERT_OK_PTR(link, "bpf_map__attach_struct_ops")) {
+		mptcp_bpf_stale__destroy(stale_skel);
+		return;
+	}
+
+	nstoken = sched_init("subflow", "bpf_stale");
+	if (!ASSERT_OK_PTR(nstoken, "sched_init:bpf_stale"))
+		goto fail;
+	server_fd = start_mptcp_server(AF_INET, ADDR_1, PORT_1, 0);
+	client_fd = connect_to_fd(server_fd, 0);
+
+	send_data(server_fd, client_fd, "bpf_stale");
+	ASSERT_OK(has_bytes_sent(ADDR_1), "has_bytes_sent addr_1");
+	ASSERT_GT(has_bytes_sent(ADDR_2), 0, "has_bytes_sent addr_2");
+
+	close(client_fd);
+	close(server_fd);
+fail:
+	cleanup_netns(nstoken);
+	bpf_link__destroy(link);
+	mptcp_bpf_stale__destroy(stale_skel);
+}
+
 void test_mptcp(void)
 {
 	if (test__start_subtest("base"))
@@ -669,4 +705,6 @@ void test_mptcp(void)
 		test_red();
 	if (test__start_subtest("burst"))
 		test_burst();
+	if (test__start_subtest("stale"))
+		test_stale();
 }
Signature · 1 line
2.35.3
maintainer
#7 lore raw
In reply to Geliang Tang
Quoted 9 lines · Geliang Tang

v3:
- init sk_storage in .init, delete it in .release.
We invoke bpf_sk_storage_get() to get the sk_storage map many times. Only the first call is the slow path (see bpf_local_storage_lookup in kernel/bpf/bpf_local_storage.c), alloc the map and cache it. The subsequent calls are all in fast path, the cache hits. So we should first call bpf_sk_storage_get in .init, then call it many times in .get_subflow.
- if no subflow is scheduled, it should return '-1' to indicate an error.

Thanks Geliang, patches 1-3 are ok to squash.

For 4-5, see my reply to patch 4.

- Mat

Quoted 31 lines

v2:
- store subflow ids instead of storing subflow pointers in sk_storage.

v1:
- This patchset adds the new bpf_stale scheduler. Use sk_storage to save the stale map instead of using subflow->stale flag to manage it.

Geliang Tang (5):
Squash to "selftests/bpf: Add bpf_bkup scheduler" Squash to "selftests/bpf: Add bpf_rr scheduler" Squash to "selftests/bpf: Add bpf_burst scheduler" selftests/bpf: Add bpf_stale scheduler selftests/bpf: Add bpf_stale test

tools/testing/selftests/bpf/bpf_tcp_helpers.h |   1 +
.../testing/selftests/bpf/prog_tests/mptcp.c  |  38 +++++
.../selftests/bpf/progs/mptcp_bpf_bkup.c      |   4 +-
.../selftests/bpf/progs/mptcp_bpf_burst.c     |   2 +
.../selftests/bpf/progs/mptcp_bpf_rr.c        |   2 +
.../selftests/bpf/progs/mptcp_bpf_stale.c     | 152 ++++++++++++++++++
6 files changed, 198 insertions(+), 1 deletion(-)

create mode 100644 tools/testing/selftests/bpf/progs/mptcp_bpf_stale.c

Signature · 1 line
2.35.3
#8 lore raw

Hi Geliang, Mat,

Quoted 12 lines · Mat Martineau
Quoted 9 lines · Geliang Tang

v3:
- init sk_storage in .init, delete it in .release.
We invoke bpf_sk_storage_get() to get the sk_storage map many times. Only the first call is the slow path (see bpf_local_storage_lookup in kernel/bpf/bpf_local_storage.c), alloc the map and cache it. The subsequent calls are all in fast path, the cache hits. So we should first call bpf_sk_storage_get in .init, then call it many times in .get_subflow.
- if no subflow is scheduled, it should return '-1' to indicate an error.

Quoted 1 line

Thanks Geliang, patches 1-3 are ok to squash.

Thank you for the patches and your patience!

Now in our tree:

New patches for t/upstream:
- b2afd09b16a1: "squashed" patch 1/5 in "selftests/bpf: Add bpf_bkup scheduler"
- ff989d90d2ea: "squashed" patch 2/5 in "selftests/bpf: Add bpf_rr scheduler"
- ff0afb0f9763: "squashed" patch 3/5 in "selftests/bpf: Add bpf_burst scheduler"
- Results: 86fa15fc9c4f..07a689bb4cbf (export)

Tests are now in progress:

https://cirrus-ci.com/github/multipath-tcp/mptcp_net-next/export/20230909T152638

Cheers,
Matt

Signature · 2 lines
Tessares | Belgium | Hybrid Access Solutions
www.tessares.net