[PATCH] btf_encoder: Match ftrace addresses within elf functions

Currently when processing DWARF function, we check its entrypoint
against ftrace addresses, assuming that the ftrace address matches
with function's entrypoint.

This is not the case on some architectures as reported by Nathan
when building kernel on arm [1].

Fixing the check to take into account the whole function not
just the entrypoint.

Most of the is_ftrace_func code was contributed by Andrii.

[1] https://lore.kernel.org/bpf/20210209034416.GA1669105@ubuntu-m3-large-x86/
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---
 btf_encoder.c | 55 +++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 45 insertions(+), 10 deletions(-)

diff --git a/btf_encoder.c b/btf_encoder.c
index b124ec20a689..03242f04c55d 100644
--- a/btf_encoder.c
+++ b/btf_encoder.c

@@ -36,6 +36,7 @@ struct funcs_layout {
 struct elf_function {
 	const char	*name;
 	unsigned long	 addr;
+	unsigned long	 size;
 	unsigned long	 sh_addr;
 	bool		 generated;
 };

@@ -98,6 +99,7 @@ static int collect_function(struct btf_elf *btfe, GElf_Sym *sym,
 
 	functions[functions_cnt].name = name;
 	functions[functions_cnt].addr = elf_sym__value(sym);
+	functions[functions_cnt].size = elf_sym__size(sym);
 	functions[functions_cnt].sh_addr = sh.sh_addr;
 	functions[functions_cnt].generated = false;
 	functions_cnt++;

@@ -236,6 +238,48 @@ get_kmod_addrs(struct btf_elf *btfe, __u64 **paddrs, __u64 *pcount)
 	return 0;
 }
 
+static int is_ftrace_func(struct elf_function *func, __u64 *addrs,
+			  __u64 count, bool kmod)
+{
+	/*
+	 * For vmlinux image both addrs[x] and functions[x]::addr
+	 * values are final address and are comparable.
+	 *
+	 * For kernel module addrs[x] is final address, but
+	 * functions[x]::addr is relative address within section
+	 * and needs to be relocated by adding sh_addr.
+	 */
+	__u64 start = kmod ? func->addr + func->sh_addr : func->addr;
+	__u64 addr, end = func->addr + func->size;
+
+	/*
+	 * The invariant here is addr[r] that is the smallest address
+	 * that is >= than function start addr. Except the corner case
+	 * where there is no such r, but for that we have a final check
+	 * in the return.
+	 */
+	size_t l = 0, r = count - 1, m;
+
+	/* make sure we don't use invalid r */
+	if (count == 0)
+		return false;
+
+	while (l < r) {
+		m = l + (r - l) / 2;
+		addr = addrs[m];
+
+		if (addr >= start) {
+			/* we satisfy invariant, so tighten r */
+			r = m;
+		} else {
+			/* m is not good enough as l, maybe m + 1 will be */
+			l = m + 1;
+		}
+	}
+
+	return start <= addrs[r] && addrs[r] < end;
+}
+
 static int setup_functions(struct btf_elf *btfe, struct funcs_layout *fl)
 {
 	__u64 *addrs, count, i;

@@ -275,18 +319,9 @@ static int setup_functions(struct btf_elf *btfe, struct funcs_layout *fl)
 	 */
 	for (i = 0; i < functions_cnt; i++) {
 		struct elf_function *func = &functions[i];
-		/*
-		 * For vmlinux image both addrs[x] and functions[x]::addr
-		 * values are final address and are comparable.
-		 *
-		 * For kernel module addrs[x] is final address, but
-		 * functions[x]::addr is relative address within section
-		 * and needs to be relocated by adding sh_addr.
-		 */
-		__u64 addr = kmod ? func->addr + func->sh_addr : func->addr;
 
 		/* Make sure function is within ftrace addresses. */
-		if (bsearch(&addr, addrs, count, sizeof(addrs[0]), addrs_cmp)) {
+		if (is_ftrace_func(func, addrs, count, kmod)) {
 			/*
 			 * We iterate over sorted array, so we can easily skip
 			 * not valid item and move following valid field into

-- 
2.29.2

On Fri, Feb 12, 2021 at 2:05 PM Jiri Olsa [off-list ref] wrote:

Currently when processing DWARF function, we check its entrypoint
against ftrace addresses, assuming that the ftrace address matches
with function's entrypoint.

This is not the case on some architectures as reported by Nathan
when building kernel on arm [1].

Fixing the check to take into account the whole function not
just the entrypoint.

Most of the is_ftrace_func code was contributed by Andrii.

[1] https://lore.kernel.org/bpf/20210209034416.GA1669105@ubuntu-m3-large-x86/
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---

LGTM. But see another suggestion below. In either case:

Acked-by: Andrii Nakryiko <andrii@kernel.org>

if we just...

if (kmod)
    func->addr += func->sh_addr;

... here, that would make is_ftrace_func() free of kmod knowledge. If
there are other places that rely on kmod vs non-kmod address of a
function, that would be simplified as well, right?

                /* Make sure function is within ftrace addresses. */
-               if (bsearch(&addr, addrs, count, sizeof(addrs[0]), addrs_cmp)) {
+               if (is_ftrace_func(func, addrs, count, kmod)) {
                        /*
                         * We iterate over sorted array, so we can easily skip
                         * not valid item and move following valid field into
--
2.29.2

On Fri, Feb 12, 2021 at 02:21:04PM -0800, Andrii Nakryiko wrote:

On Fri, Feb 12, 2021 at 2:05 PM Jiri Olsa [off-list ref] wrote:

quoted

Currently when processing DWARF function, we check its entrypoint
against ftrace addresses, assuming that the ftrace address matches
with function's entrypoint.

This is not the case on some architectures as reported by Nathan
when building kernel on arm [1].

Fixing the check to take into account the whole function not
just the entrypoint.

Most of the is_ftrace_func code was contributed by Andrii.

[1] https://lore.kernel.org/bpf/20210209034416.GA1669105@ubuntu-m3-large-x86/
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
---

LGTM. But see another suggestion below. In either case:

Acked-by: Andrii Nakryiko <andrii@kernel.org>

quoted

 btf_encoder.c | 55 +++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 45 insertions(+), 10 deletions(-)

diff --git a/btf_encoder.c b/btf_encoder.c
index b124ec20a689..03242f04c55d 100644
--- a/btf_encoder.c
+++ b/btf_encoder.c

@@ -36,6 +36,7 @@ struct funcs_layout {
 struct elf_function {
        const char      *name;
        unsigned long    addr;
+       unsigned long    size;
        unsigned long    sh_addr;
        bool             generated;
 };

@@ -98,6 +99,7 @@ static int collect_function(struct btf_elf *btfe, GElf_Sym *sym,

        functions[functions_cnt].name = name;
        functions[functions_cnt].addr = elf_sym__value(sym);
+       functions[functions_cnt].size = elf_sym__size(sym);
        functions[functions_cnt].sh_addr = sh.sh_addr;
        functions[functions_cnt].generated = false;
        functions_cnt++;

@@ -236,6 +238,48 @@ get_kmod_addrs(struct btf_elf *btfe, __u64 **paddrs, __u64 *pcount)
        return 0;
 }

+static int is_ftrace_func(struct elf_function *func, __u64 *addrs,
+                         __u64 count, bool kmod)
+{
+       /*
+        * For vmlinux image both addrs[x] and functions[x]::addr
+        * values are final address and are comparable.
+        *
+        * For kernel module addrs[x] is final address, but
+        * functions[x]::addr is relative address within section
+        * and needs to be relocated by adding sh_addr.
+        */
+       __u64 start = kmod ? func->addr + func->sh_addr : func->addr;
+       __u64 addr, end = func->addr + func->size;
+
+       /*
+        * The invariant here is addr[r] that is the smallest address
+        * that is >= than function start addr. Except the corner case
+        * where there is no such r, but for that we have a final check
+        * in the return.
+        */
+       size_t l = 0, r = count - 1, m;
+
+       /* make sure we don't use invalid r */
+       if (count == 0)
+               return false;
+
+       while (l < r) {
+               m = l + (r - l) / 2;
+               addr = addrs[m];
+
+               if (addr >= start) {
+                       /* we satisfy invariant, so tighten r */
+                       r = m;
+               } else {
+                       /* m is not good enough as l, maybe m + 1 will be */
+                       l = m + 1;
+               }
+       }
+
+       return start <= addrs[r] && addrs[r] < end;
+}
+
 static int setup_functions(struct btf_elf *btfe, struct funcs_layout *fl)
 {
        __u64 *addrs, count, i;

@@ -275,18 +319,9 @@ static int setup_functions(struct btf_elf *btfe, struct funcs_layout *fl)
         */
        for (i = 0; i < functions_cnt; i++) {
                struct elf_function *func = &functions[i];
-               /*
-                * For vmlinux image both addrs[x] and functions[x]::addr
-                * values are final address and are comparable.
-                *
-                * For kernel module addrs[x] is final address, but
-                * functions[x]::addr is relative address within section
-                * and needs to be relocated by adding sh_addr.
-                */
-               __u64 addr = kmod ? func->addr + func->sh_addr : func->addr;

if we just...

if (kmod)
    func->addr += func->sh_addr;

... here, that would make is_ftrace_func() free of kmod knowledge. If
there are other places that rely on kmod vs non-kmod address of a
function, that would be simplified as well, right?

yes, this is the only place for now, I'll make the change

thanks,
jirka

quoted

                /* Make sure function is within ftrace addresses. */
-               if (bsearch(&addr, addrs, count, sizeof(addrs[0]), addrs_cmp)) {
+               if (is_ftrace_func(func, addrs, count, kmod)) {
                        /*
                         * We iterate over sorted array, so we can easily skip
                         * not valid item and move following valid field into
--
2.29.2