Validate that bpf_get_reg_val helper solves the motivating problem of
this patch series: USDT args passed through xmm regs. The userspace
portion of the test forces STAP_PROBE macro to use %xmm0 and %xmm1 regs
to pass a float and an int, which the bpf-side successfully reads using
BPF_USDT.
In the wild I discovered a sanely-configured USDT in Fedora libpthread
using xmm regs to pass scalar values, likely due to register pressure.
urandom_read_lib_xmm mimics this by using -ffixed-$REG flag to mark
r11-r14 unusable and passing many USDT args.
Signed-off-by: Dave Marchevsky <davemarchevsky@xxxxxx>
---
tools/testing/selftests/bpf/Makefile | 8 ++-
tools/testing/selftests/bpf/prog_tests/usdt.c | 7 +++
.../selftests/bpf/progs/test_urandom_usdt.c | 13 ++++
tools/testing/selftests/bpf/urandom_read.c | 3 +
.../selftests/bpf/urandom_read_lib_xmm.c | 62 +++++++++++++++++++
5 files changed, 91 insertions(+), 2 deletions(-)
create mode 100644 tools/testing/selftests/bpf/urandom_read_lib_xmm.c
diff --git a/tools/testing/selftests/bpf/Makefile b/tools/testing/selftests/bpf/Makefile
index 6bbc03161544..19246e34dfe1 100644
--- a/tools/testing/selftests/bpf/Makefile
+++ b/tools/testing/selftests/bpf/Makefile
@@ -172,10 +172,14 @@ $(OUTPUT)/liburandom_read.so: urandom_read_lib1.c urandom_read_lib2.c
$(call msg,LIB,,$@)
$(Q)$(CC) $(CFLAGS) -fPIC $(LDFLAGS) $^ $(LDLIBS) --shared -o $@
-$(OUTPUT)/urandom_read: urandom_read.c urandom_read_aux.c $(OUTPUT)/liburandom_read.so
+$(OUTPUT)/liburandom_read_xmm.so: urandom_read_lib_xmm.c
+ $(call msg,LIB,,$@)
+ $(Q)$(CC) -O0 -ffixed-r11 -ffixed-r12 -ffixed-r13 -ffixed-r14 -fPIC $(LDFLAGS) $^ $(LDLIBS) --shared -o $@
+
+$(OUTPUT)/urandom_read: urandom_read.c urandom_read_aux.c $(OUTPUT)/liburandom_read.so $(OUTPUT)/liburandom_read_xmm.so
$(call msg,BINARY,,$@)
$(Q)$(CC) $(CFLAGS) $(LDFLAGS) $(filter %.c,$^) \
- liburandom_read.so $(LDLIBS) \
+ liburandom_read.so liburandom_read_xmm.so $(LDLIBS) \
-Wl,-rpath=. -Wl,--build-id=sha1 -o $@
$(OUTPUT)/bpf_testmod.ko: $(VMLINUX_BTF) $(wildcard bpf_testmod/Makefile bpf_testmod/*.[ch])
diff --git a/tools/testing/selftests/bpf/prog_tests/usdt.c b/tools/testing/selftests/bpf/prog_tests/usdt.c
index a71f51bdc08d..f98749ac74a7 100644
--- a/tools/testing/selftests/bpf/prog_tests/usdt.c
+++ b/tools/testing/selftests/bpf/prog_tests/usdt.c
@@ -385,6 +385,12 @@ static void subtest_urandom_usdt(bool auto_attach)
goto cleanup;
skel->links.urandlib_read_with_sema = l;
+ l = bpf_program__attach_usdt(skel->progs.urandlib_xmm_reg_read,
+ urand_pid, "./liburandom_read_xmm.so",
+ "urandlib", "xmm_reg_read", NULL);
+ if (!ASSERT_OK_PTR(l, "urandlib_xmm_reg_read"))
+ goto cleanup;
+ skel->links.urandlib_xmm_reg_read = l;
}
/* trigger urandom_read USDTs */
@@ -402,6 +408,7 @@ static void subtest_urandom_usdt(bool auto_attach)
ASSERT_EQ(bss->urandlib_read_with_sema_call_cnt, 1, "urandlib_w_sema_cnt");
ASSERT_EQ(bss->urandlib_read_with_sema_buf_sz_sum, 256, "urandlib_w_sema_sum");
+ ASSERT_EQ(bss->urandlib_xmm_reg_read_buf_sz_sum, 256, "liburandom_read_xmm.so");
cleanup:
if (urand_pipe)
pclose(urand_pipe);
diff --git a/tools/testing/selftests/bpf/progs/test_urandom_usdt.c b/tools/testing/selftests/bpf/progs/test_urandom_usdt.c
index 3539b02bd5f7..575761863eb6 100644
--- a/tools/testing/selftests/bpf/progs/test_urandom_usdt.c
+++ b/tools/testing/selftests/bpf/progs/test_urandom_usdt.c
@@ -67,4 +67,17 @@ int BPF_USDT(urandlib_read_with_sema, int iter_num, int iter_cnt, int buf_sz)
return 0;
}
+int urandlib_xmm_reg_read_buf_sz_sum;
+SEC("usdt/./liburandom_read_xmm.so:urandlib:xmm_reg_read")
+int BPF_USDT(urandlib_xmm_reg_read, int *f1, int *f2, int *f3, int a, int b,
+ int c /*should be float */, int d, int e,
+ int f, int g, int h, int buf_sz)
+{
+ if (urand_pid != (bpf_get_current_pid_tgid() >> 32))
+ return 0;
+
+ __sync_fetch_and_add(&urandlib_xmm_reg_read_buf_sz_sum, buf_sz);
+ return 0;
+}
+
char _license[] SEC("license") = "GPL";
diff --git a/tools/testing/selftests/bpf/urandom_read.c b/tools/testing/selftests/bpf/urandom_read.c
index e92644d0fa75..0ee7aad014e1 100644
--- a/tools/testing/selftests/bpf/urandom_read.c
+++ b/tools/testing/selftests/bpf/urandom_read.c
@@ -20,6 +20,8 @@ void urand_read_without_sema(int iter_num, int iter_cnt, int read_sz);
/* these are coming from urandom_read_lib{1,2}.c */
void urandlib_read_with_sema(int iter_num, int iter_cnt, int read_sz);
void urandlib_read_without_sema(int iter_num, int iter_cnt, int read_sz);
+/* defined in urandom_read_lib_xmm.c */
+void urandlib_read_xmm_args(int iter_num, int iter_cnt, int read_sz);
unsigned short urand_read_with_sema_semaphore SEC(".probes");
@@ -39,6 +41,7 @@ void urandom_read(int fd, int count)
/* trigger USDTs defined in shared lib */
urandlib_read_without_sema(i, count, BUF_SIZE);
urandlib_read_with_sema(i, count, BUF_SIZE);
+ urandlib_read_xmm_args(i, count, BUF_SIZE);
}
}
diff --git a/tools/testing/selftests/bpf/urandom_read_lib_xmm.c b/tools/testing/selftests/bpf/urandom_read_lib_xmm.c
new file mode 100644
index 000000000000..d5f9c9cf74e7
--- /dev/null
+++ b/tools/testing/selftests/bpf/urandom_read_lib_xmm.c
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2022 Meta Platforms, Inc. and affiliates. */
+
+#define STAP_SDT_ARG_CONSTRAINT norx
+/* For x86_64, this was changed from 'nor' default to 'norfxy' in systemtap
+ * commit eaa15b047 ("PR27829: Support floating point values passed through sdt.h markers")
+ * then changed to 'norx' in commit 1d3653936f ("sys/sdt.h fp constraints cont'd, x86-64 edition")
+ * It's not necessary to set STAP_SDT_ARG_CONSTRAINT for newer systemtap to see
+ * xmm regs used in this program
+ */
+
+#include "sdt.h"
+
+int *f1(void)
+{
+ return (int *)100;
+}
+
+int *f2(void)
+{
+ return (int *)200;
+}
+
+int *f3(void)
+{
+ return (int *)300;
+}
+
+/* Compile w/ -ffixed-r11 -ffixed-r12 -ffixed-r13 -ffixed-r14 -O0 */
+void urandlib_read_xmm_args(int iter_num, int iter_cnt, int read_sz)
+{
+ volatile int a, b, d, e, f, g, h, i;
+ a = b = d = e = f = g = h = 300;
+ i = read_sz;
+ volatile float c = 100;
+
+ STAP_PROBE12(urandlib, xmm_reg_read, f1(), f2(), f3(), a, b, c, d, e, f, g, h, i);
+}
+
+/*
+ * `readelf -n` results:
+ * On a test host outside of kernel toolchain (Ubuntu 20.04, 5.13.0-39-generic, gcc 9.4.0-1ubuntu1~20.04.1)
+ * w/ STAP_SDT_ARG_CONSTRAINT norfxy
+ * using system sdt.h:
+ * Provider: urandlib
+ * Name: xmm_reg_read
+ * Location: 0x00000000000011d8, Base: 0x0000000000002008, Semaphore: 0x0000000000000000
+ * Arguments: 8@%rbx 8@%r15 8@%xmm1 -4@%edx -4@%ecx 4@%xmm0 -4@%esi -4@%edi -4@%r8d -4@%r9d -4@%r10d -4@%eax
+ *
+ * Using a newer systemtap's sdt.h (commit acca4ae47 ("Don't run tls tests if debuginfo is missing")):
+ * Provider: urandlib
+ * Name: xmm_reg_read
+ * Location: 0x00000000000011d8, Base: 0x0000000000002008, Semaphore: 0x0000000000000000
+ * Arguments: 8@%rbx 8@%r15 8@%xmm1 -4@%edx -4@%ecx 4f@%xmm0 -4@%esi -4@%edi -4@%r8d -4@%r9d -4@%r10d -4@%eax
+ *
+ * Kernel toolchain:
+ * STAP_SDT_ARG_CONSTRAINT norfxy causes compiler error (due to the 'f'), so using 'norx'
+ * Provider: urandlib
+ * Name: xmm_reg_read
+ * Location: 0x0000000000000717, Base: 0x0000000000000738, Semaphore: 0x0000000000000000
+ * Arguments: 8@%rbx 8@%r15 8@-72(%rbp) -4@%edx -4@%ecx 4f@%xmm0 -4@%esi -4@%edi -4@%r8d -4@%r9d -4@%r10d -4@%xmm1
+ */
--
2.30.2
