Skip to contentSkip to conversation

samples/bpf: Add xdp_stat sample program

PATCH Superseded Eric Sage 3 participants 7 messages Dec 29, 2019 – Feb 17, 2020

Subsystems (4)

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Eric Sage

author
#0 lore raw

PATCH 1/1 samples/bpf: Add xdp_stat sample program

At Facebook we use tail calls to jump between our firewall filters and our L4LB. This is a program I wrote to estimate per program performance by swapping out the entries in the program array with interceptors that take measurements and then jump to the original entries.

I found the sample programs to be invaluable in understanding how to use the libbpf API (as well as the test env from the xdp-tutorial repo for testing), and want to return the favor. I am currently working on my next iteration that uses fentry/fexit to be less invasive, but I thought it was an interesting PoC of what you can do with program arrays.

---
 samples/bpf/Makefile          |   3 +
 samples/bpf/xdp_stat_common.h |  28 ++
 samples/bpf/xdp_stat_kern.c   | 169 ++++++++
 samples/bpf/xdp_stat_user.c   | 746 ++++++++++++++++++++++++++++++++++
 4 files changed, 946 insertions(+)
 create mode 100644 samples/bpf/xdp_stat_common.h
 create mode 100644 samples/bpf/xdp_stat_kern.c
 create mode 100644 samples/bpf/xdp_stat_user.c
samples/bpf/Makefile +3 -0
index 5b89c0370f33..36469b3344b0 100644
@@ -53,6 +53,7 @@ tprogs-y += task_fd_query
 tprogs-y += xdp_sample_pkts
 tprogs-y += ibumad
 tprogs-y += hbm
+tprogs-y += xdp_stat
 
 # Libbpf dependencies
 LIBBPF = $(TOOLS_PATH)/lib/bpf/libbpf.a
@@ -109,6 +110,7 @@ task_fd_query-objs := bpf_load.o task_fd_query_user.o $(TRACE_HELPERS)
 xdp_sample_pkts-objs := xdp_sample_pkts_user.o $(TRACE_HELPERS)
 ibumad-objs := bpf_load.o ibumad_user.o $(TRACE_HELPERS)
 hbm-objs := bpf_load.o hbm.o $(CGROUP_HELPERS)
+xdp_stat-objs := xdp_stat_user.o
 
 # Tell kbuild to always build the programs
 always := $(tprogs-y)
@@ -170,6 +172,7 @@ always += ibumad_kern.o
 always += hbm_out_kern.o
 always += hbm_edt_kern.o
 always += xdpsock_kern.o
+always += xdp_stat_kern.o
 
 ifeq ($(ARCH), arm)
 # Strip all except -D__LINUX_ARM_ARCH__ option needed to handle linux
samples/bpf/xdp_stat_common.h new file +28 -0
new file mode 100644
index 000000000000..7c0557410704
@@ -0,0 +1,28 @@
+/* SPDX-License-Identifier: GPL-2.0
+ * Copyright (c) 2019 Facebook
+ */
+
+#ifndef __XDP_STAT_COMMON_HEADER_H
+#define __XDP_STAT_COMMON_HEADER_H
+
+/* WARNING: Must match the number of interceptors generated in
+ * xdp_stat_kern.c including xdp/interceptor_0.
+ */
+#define MAX_INTERCEPTORS 10
+
+/* prog_stats_rec contains stats specific to an intercepted BPF program.
+ */
+struct prog_stats_rec {
+	__u64 nr_terminal_runs;
+	__u64 nr_chained_runs;
+	__u64 ns_chained_runtime;
+};
+
+/* interception_info_rec contains global to all interceptors on a single CPU.
+ */
+struct interception_info_rec {
+	__u32 prev_interceptor_nr;
+	__u64 prev_ns_start;
+};
+
+#endif /* __XDP_STAT_COMMON_HEADER_H */
samples/bpf/xdp_stat_kern.c new file +169 -0

169 lines not on this page · raw message

samples/bpf/xdp_stat_user.c new file +746 -0

746 lines not on this page · raw message

Signature · 1 line
2.24.1
#1 lore raw
Quoted 13 lines · Eric Sage

At Facebook we use tail calls to jump between our firewall filters and our L4LB. This is a program I wrote to estimate per program performance by swapping out the entries in the program array with interceptors that take measurements and then jump to the original entries.

I found the sample programs to be invaluable in understanding how to use the libbpf API (as well as the test env from the xdp-tutorial repo for testing), and want to return the favor. I am currently working on my next iteration that uses fentry/fexit to be less invasive, but I thought it was an interesting PoC of what you can do with program arrays.

Quoted 2 lines

---

Hey Eric,

Thanks for contributing this to samples. Ideally this would use BPF skeleton to simplify all the map and program look ups, but given samples/bpf doesn't have a infrastructure set up for this, it might be a bit too much effort at this point. So I think it's ok to do it this way, even though it would be really nice to try.

But please update the old-style bpf_map_def to BTF-defined maps before this gets merged (see below).

Quoted 9 lines
 samples/bpf/Makefile          |   3 +
 samples/bpf/xdp_stat_common.h |  28 ++
 samples/bpf/xdp_stat_kern.c   | 169 ++++++++
 samples/bpf/xdp_stat_user.c   | 746 ++++++++++++++++++++++++++++++++++
 4 files changed, 946 insertions(+)
 create mode 100644 samples/bpf/xdp_stat_common.h
 create mode 100644 samples/bpf/xdp_stat_kern.c
 create mode 100644 samples/bpf/xdp_stat_user.c

[...]

Quoted 7 lines
samples/bpf/xdp_stat_kern.c new file +7 -0
+#include <uapi/linux/bpf.h>
+#include "bpf_helpers.h"
+
+#include "xdp_stat_common.h"
+
+#define MAX_PROG_ARRAY 64
+

samples/bpf/xdp_stat_kern.c line 40

Quoted 10 lines
samples/bpf/xdp_stat_kern.c new file +10 -0
+/* NR is used to map interceptors to the programs that are being intercepted. */
+#define INTERCEPTOR(INDEX)                                                     \
+       SEC("xdp/interceptor_" #INDEX)                                         \
+       int interceptor_##INDEX(struct xdp_md *ctx)                            \
+       {                                                                      \
+               return interceptor_impl(ctx, INDEX);                           \
+       }
+
+       /* Required to use bpf_ktime_get_ns() */
+       char _license[] SEC("license") = "GPL";

Wrong indentation?

samples/bpf/xdp_stat_kern.c line 50

Quoted 10 lines
samples/bpf/xdp_stat_kern.c new file +10 -0
+
+/* interception_info holds a single record per CPU to pass global state between
+ * interceptor programs.
+ */
+struct bpf_map_def SEC("maps") interception_info = {
+       .type = BPF_MAP_TYPE_PERCPU_ARRAY,
+       .key_size = sizeof(__u32),
+       .value_size = sizeof(struct interception_info_rec),
+       .max_entries = 1,
+};

With BTF-define maps this would be

struct {
    __uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
    __uint(max_entries, 1);
    __type(key, __u32);
    __type(value, struct interception_info_rec);
} interception_info SEC(".maps");

Can you please switch all the maps to this new syntax?

Quoted 12 lines
samples/bpf/xdp_stat_kern.c new file +12 -0
+
+/* interceptor_stats maps interceptor indexes to measurements of an intercepted
+ * BPF program. Index 0 maps the interceptor entrypoint to measurements of the
+ * original entrypoint.
+ */
+struct bpf_map_def SEC("maps") prog_stats = {
+       .type = BPF_MAP_TYPE_PERCPU_ARRAY,
+       .key_size = sizeof(__u32),
+       .value_size = sizeof(struct prog_stats_rec),
+       .max_entries = MAX_PROG_ARRAY,
+};
+

[...]

samples/bpf/xdp_stat_user.c line 458

Quoted 13 lines
samples/bpf/xdp_stat_user.c new file +13 -0
+static struct intercept_ctx *intercept__setup(char *mappath, int ifindex)
+{
+       struct intercept_ctx *ctx = malloc(sizeof(struct intercept_ctx));
+
+       ctx->ifindex = ifindex;
+
+       ctx->stats_enabled_oldval =
+               _update_sysctl("/proc/sys/kernel/bpf_stats_enabled", 1);
+       if (ctx->stats_enabled_oldval < 0)
+               perror("ERR: set bpf_stats_enabled sysctl failed\n");
+
+       if (bpf_prog_load(FILENAME_XDP_STAT_KERN, BPF_PROG_TYPE_XDP,
+                         &ctx->bpf_obj, &ctx->entry_fd)) {

bpf_object__open() + bpf_object__load() is preferred over bpf_prog_load() way to open and load BPF object files, would be nice to use that instead.

Alternative is BPF skeleton, but as I said, it will require a bit more set up in Makefile (you'd need to generate skeleton with bpftool gen skeleton <your-bpf-object>.o). Few benefits of BPF skeleton, though:
- you wouldn't need to look up maps and programs by name, they would
be accessible as struct fields of skeleton (e.g., skel->maps.interception_info);
- BPF object file would be embedded into userspace program, so when
distributing this you wouldn't need to drag along a separate .o file.

Quoted 5 lines
samples/bpf/xdp_stat_user.c new file +5 -0
+               fprintf(stderr, "ERR: failed to load %s\n",
+                       FILENAME_XDP_STAT_KERN);
+               return NULL;
+       }
+

[...]

author
#2 lore raw

PATCH v2 1/1 samples/bpf: Add xdp_stat sample program

In reply to Eric Sage

At Facebook we use tail calls to jump between our firewall filters and our L4LB. This is a program I wrote to estimate per program performance by swapping out the entries in the program array with interceptors that take measurements and then jump to the original entries.

I found the sample programs to be invaluable in understanding how to use the libbpf API (as well as the test env from the xdp-tutorial repo for testing), and want to return the favor. I am currently working on my next iteration that uses fentry/fexit to be less invasive, but I thought it was an interesting PoC of what you can do with program arrays.

---
Changes in v2:
- Upped max number of interceptors to 32 and mentioned the max in the help.
- Fixed license formatting.
- Requested change to convert BPF map definitions to BTF.

 samples/bpf/Makefile          |   3 +
 samples/bpf/xdp_stat_common.h |  28 ++
 samples/bpf/xdp_stat_kern.c   | 192 +++++++++
 samples/bpf/xdp_stat_user.c   | 748 ++++++++++++++++++++++++++++++++++
 4 files changed, 971 insertions(+)
 create mode 100644 samples/bpf/xdp_stat_common.h
 create mode 100644 samples/bpf/xdp_stat_kern.c
 create mode 100644 samples/bpf/xdp_stat_user.c
samples/bpf/Makefile +3 -0
index 5b89c0370f33..36469b3344b0 100644
@@ -53,6 +53,7 @@ tprogs-y += task_fd_query
 tprogs-y += xdp_sample_pkts
 tprogs-y += ibumad
 tprogs-y += hbm
+tprogs-y += xdp_stat
 
 # Libbpf dependencies
 LIBBPF = $(TOOLS_PATH)/lib/bpf/libbpf.a
@@ -109,6 +110,7 @@ task_fd_query-objs := bpf_load.o task_fd_query_user.o $(TRACE_HELPERS)
 xdp_sample_pkts-objs := xdp_sample_pkts_user.o $(TRACE_HELPERS)
 ibumad-objs := bpf_load.o ibumad_user.o $(TRACE_HELPERS)
 hbm-objs := bpf_load.o hbm.o $(CGROUP_HELPERS)
+xdp_stat-objs := xdp_stat_user.o
 
 # Tell kbuild to always build the programs
 always := $(tprogs-y)
@@ -170,6 +172,7 @@ always += ibumad_kern.o
 always += hbm_out_kern.o
 always += hbm_edt_kern.o
 always += xdpsock_kern.o
+always += xdp_stat_kern.o
 
 ifeq ($(ARCH), arm)
 # Strip all except -D__LINUX_ARM_ARCH__ option needed to handle linux
samples/bpf/xdp_stat_common.h new file +28 -0
new file mode 100644
index 000000000000..7c0557410704
@@ -0,0 +1,28 @@
+/* SPDX-License-Identifier: GPL-2.0
+ * Copyright (c) 2019 Facebook
+ */
+
+#ifndef __XDP_STAT_COMMON_HEADER_H
+#define __XDP_STAT_COMMON_HEADER_H
+
+/* WARNING: Must match the number of interceptors generated in
+ * xdp_stat_kern.c including xdp/interceptor_0.
+ */
+#define MAX_INTERCEPTORS 10
+
+/* prog_stats_rec contains stats specific to an intercepted BPF program.
+ */
+struct prog_stats_rec {
+	__u64 nr_terminal_runs;
+	__u64 nr_chained_runs;
+	__u64 ns_chained_runtime;
+};
+
+/* interception_info_rec contains global to all interceptors on a single CPU.
+ */
+struct interception_info_rec {
+	__u32 prev_interceptor_nr;
+	__u64 prev_ns_start;
+};
+
+#endif /* __XDP_STAT_COMMON_HEADER_H */
samples/bpf/xdp_stat_kern.c new file +192 -0

192 lines not on this page · raw message

samples/bpf/xdp_stat_user.c new file +748 -0

748 lines not on this page · raw message

Signature · 1 line
2.24.1
#3 lore raw
Quoted 13 lines · Eric Sage

At Facebook we use tail calls to jump between our firewall filters and our L4LB. This is a program I wrote to estimate per program performance by swapping out the entries in the program array with interceptors that take measurements and then jump to the original entries.

I found the sample programs to be invaluable in understanding how to use the libbpf API (as well as the test env from the xdp-tutorial repo for testing), and want to return the favor. I am currently working on my next iteration that uses fentry/fexit to be less invasive, but I thought it was an interesting PoC of what you can do with program arrays.

Quoted 8 lines

---
Changes in v2:
- Upped max number of interceptors to 32 and mentioned the max in the help.
- Fixed license formatting.
- Requested change to convert BPF map definitions to BTF.

I should have mentioned a gotcha with PROG_ARRAY earlier, sorry about that. Use key_size/value_size for it. Other than that, looks good.

[...]

Quoted 11 lines
samples/bpf/xdp_stat_kern.c new file +5 -0
new file mode 100644
index 000000000000..83cff4807f72
@@ -0,0 +1,192 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright(c) 2019 Facebook
+ */
+
+/* Conceptually interception looks like this for a single packet:

samples/bpf/xdp_stat_kern.c line 15

Quoted 10 lines
samples/bpf/xdp_stat_kern.c new file +10 -0
+ *
+ * interceptor_0 -> entrypoint -> interceptor_1 -> prog_1 -> ... ->
+ * interceptor_N -> prog_N -> XDP_ACTION
+ *
+ * At any point in the chain, including in the entrypoint, an XDP_ACTION can
+ * be returned. It is also not assumed that the order of jumps will not change
+ * (except that the entrypoint always comes first).
+ *
+ * Because there is no way to hook into the return of the XDP action, the
+ * entrpoint (interceptor_0) is also used to record the terminal run of the

typo: entrypoint

samples/bpf/xdp_stat_kern.c line 16

Quoted 1 line
samples/bpf/xdp_stat_kern.c new file +1 -0
+ * previous BPF program on the same CPU. Conceputally:

typo: Conceptually

Quoted 7 lines
samples/bpf/xdp_stat_kern.c new file +7 -0
+ *
+ * ... -> prog_N -> XDP_ACTION -> interceptor_0 -> ...
+ *
+ * FIXME: A bad side effect of this is that the reported stats will always be
+ * behind in tracking terminal runs which is confusing to the user.
+ */
+

[...]

samples/bpf/xdp_stat_kern.c line 81

Quoted 8 lines
samples/bpf/xdp_stat_kern.c new file +8 -0
+/* jmp_table_entry has a single entry - the original XDP entrpoint - so that
+ * the intercpetor entrypoint can jump to it.
+ */
+struct {
+       __uint(type, BPF_MAP_TYPE_PROG_ARRAY);
+       __uint(max_entries, 1);
+       __type(key, __u32);
+       __type(value, __u32);

Did this actually work, when you run your tool? Few special maps still don't support specifying type information, I think PROG_ARRAY is one of those. For such cases you have to specify key_size and value_size, instead of key/value types:

__uint(key_size, sizeof(__u32));
__uint(value_size, sizeof(__u32));

Quoted 2 lines
samples/bpf/xdp_stat_kern.c new file +2 -0
+} jmp_table_entrypoint SEC(".maps");
+

[...]

author
#4 lore raw

PATCH v3 1/1 samples/bpf: Add xdp_stat sample program

At Facebook we use tail calls to jump between our firewall filters and our L4LB. This is a program I wrote to estimate per program performance by swapping out the entries in the program array with interceptors that take measurements and then jump to the original entries.

I found the sample programs to be invaluable in understanding how to use the libbpf API (as well as the test env from the xdp-tutorial repo for testing), and want to return the favor. I am currently working on my next iteration that uses fentry/fexit to be less invasive, but I thought it was an interesting PoC of what you can do with program arrays.

---
Changes in v3:
- Fixed typos in xdp_stat_kern.c
- Switch to using key_size, value_size for prog arrays.

 samples/bpf/Makefile          |   3 +
 samples/bpf/xdp_stat_common.h |  28 ++
 samples/bpf/xdp_stat_kern.c   | 192 +++++++++
 samples/bpf/xdp_stat_user.c   | 748 ++++++++++++++++++++++++++++++++++
 4 files changed, 971 insertions(+)
 create mode 100644 samples/bpf/xdp_stat_common.h
 create mode 100644 samples/bpf/xdp_stat_kern.c
 create mode 100644 samples/bpf/xdp_stat_user.c
samples/bpf/Makefile +3 -0
index 5b89c0370f33..36469b3344b0 100644
@@ -53,6 +53,7 @@ tprogs-y += task_fd_query
 tprogs-y += xdp_sample_pkts
 tprogs-y += ibumad
 tprogs-y += hbm
+tprogs-y += xdp_stat
 
 # Libbpf dependencies
 LIBBPF = $(TOOLS_PATH)/lib/bpf/libbpf.a
@@ -109,6 +110,7 @@ task_fd_query-objs := bpf_load.o task_fd_query_user.o $(TRACE_HELPERS)
 xdp_sample_pkts-objs := xdp_sample_pkts_user.o $(TRACE_HELPERS)
 ibumad-objs := bpf_load.o ibumad_user.o $(TRACE_HELPERS)
 hbm-objs := bpf_load.o hbm.o $(CGROUP_HELPERS)
+xdp_stat-objs := xdp_stat_user.o
 
 # Tell kbuild to always build the programs
 always := $(tprogs-y)
@@ -170,6 +172,7 @@ always += ibumad_kern.o
 always += hbm_out_kern.o
 always += hbm_edt_kern.o
 always += xdpsock_kern.o
+always += xdp_stat_kern.o
 
 ifeq ($(ARCH), arm)
 # Strip all except -D__LINUX_ARM_ARCH__ option needed to handle linux
samples/bpf/xdp_stat_common.h new file +28 -0
new file mode 100644
index 000000000000..7c0557410704
@@ -0,0 +1,28 @@
+/* SPDX-License-Identifier: GPL-2.0
+ * Copyright (c) 2019 Facebook
+ */
+
+#ifndef __XDP_STAT_COMMON_HEADER_H
+#define __XDP_STAT_COMMON_HEADER_H
+
+/* WARNING: Must match the number of interceptors generated in
+ * xdp_stat_kern.c including xdp/interceptor_0.
+ */
+#define MAX_INTERCEPTORS 10
+
+/* prog_stats_rec contains stats specific to an intercepted BPF program.
+ */
+struct prog_stats_rec {
+	__u64 nr_terminal_runs;
+	__u64 nr_chained_runs;
+	__u64 ns_chained_runtime;
+};
+
+/* interception_info_rec contains global to all interceptors on a single CPU.
+ */
+struct interception_info_rec {
+	__u32 prev_interceptor_nr;
+	__u64 prev_ns_start;
+};
+
+#endif /* __XDP_STAT_COMMON_HEADER_H */
samples/bpf/xdp_stat_kern.c new file +192 -0

192 lines not on this page · raw message

samples/bpf/xdp_stat_user.c new file +748 -0

748 lines not on this page · raw message

Signature · 1 line
2.24.1
Acked-by on v3 1/1
#5 lore raw
Quoted 13 lines · Eric Sage

At Facebook we use tail calls to jump between our firewall filters and our L4LB. This is a program I wrote to estimate per program performance by swapping out the entries in the program array with interceptors that take measurements and then jump to the original entries.

I found the sample programs to be invaluable in understanding how to use the libbpf API (as well as the test env from the xdp-tutorial repo for testing), and want to return the favor. I am currently working on my next iteration that uses fentry/fexit to be less invasive, but I thought it was an interesting PoC of what you can do with program arrays.

Quoted 2 lines

---

Looks good, thanks!

Quoted 4 lines

Changes in v3:
- Fixed typos in xdp_stat_kern.c
- Switch to using key_size, value_size for prog arrays.

Just for the future: you should keep entire changelog, starting from v1->v2 changes. Also, netdev patches usually preserve them as part of commit message (for individual patch) or cover letter (for patch set). But in this case I think changes are not substantial enough to record them permanently, so it's fine.

[...]

maintainer
#6 lore raw
In reply to Eric Sage
Quoted 5 lines · Eric Sage

At Facebook we use tail calls to jump between our firewall filters and our L4LB. This is a program I wrote to estimate per program performance by swapping out the entries in the program array with interceptors that take measurements and then jump to the original entries.

Quoted 8 lines

I found the sample programs to be invaluable in understanding how to use the libbpf API (as well as the test env from the xdp-tutorial repo for testing), and want to return the favor. I am currently working on my next iteration that uses fentry/fexit to be less invasive, but I thought it was an interesting PoC of what you can do with program arrays.

Now that bpf-next is back open, this needs a rebase for proceeding to get merged.

Thanks,
Daniel