On 4/1/21 11:28 AM, Nick Desaulniers wrote:
On Wed, Mar 31, 2021 at 6:24 PM Yonghong Song <yhs@xxxxxx> wrote:Currently, clang LTO built vmlinux won't work with pahole. LTO introduced cross-cu dwarf tag references and broke current pahole model which handles one cu as a time. The solution is to merge all cu's as one pahole cu as in [1]. We would like to do this merging only if cross-cu dwarf references happens. The LTO build mode is a pretty good indication for that. In earlier version of this patch ([2]), clang flag -grecord-gcc-switches is proposed to add to compilation flags so pahole could detect "-flto" and then merging cu's. This will increate the binary size of 1% without LTO though. Arnaldo suggested to use a note to indicate the vmlinux is built with LTO. Such a cheap way to get whether the vmlinux is built with LTO or not helps pahole but is also useful for tracing as LTO may inline/delete/demote global functions, promote static functions, etc. So this patch added an elfnote with type BUILD_COMPILER_LTO_INFO. The owner of the note is "Linux". With gcc 8.4.1 and clang trunk, without LTO, I got $ readelf -n vmlinux Displaying notes found in: .notes Owner Data size Description ... Linux 0x00000004 func description data: 00 00 00 00 ... With "readelf -x ".notes" vmlinux", I can verify the above "func" with type code 0x101. With clang thin-LTO, I got the same as above except the following: description data: 01 00 00 00 which indicates the vmlinux is built with LTO. [1] https://lore.kernel.org/bpf/20210325065316.3121287-1-yhs@xxxxxx/ [2] https://lore.kernel.org/bpf/20210331001623.2778934-1-yhs@xxxxxx/ Signed-off-by: Yonghong Song <yhs@xxxxxx> --- include/linux/compiler.h | 8 ++++++++ include/linux/elfnote.h | 1 + init/version.c | 2 ++ scripts/mod/modpost.c | 1 + 4 files changed, 12 insertions(+) diff --git a/include/linux/compiler.h b/include/linux/compiler.h index df5b405e6305..b92930877277 100644 --- a/include/linux/compiler.h +++ b/include/linux/compiler.h @@ -245,6 +245,14 @@ static inline void *offset_to_ptr(const int *off) */ #define prevent_tail_call_optimization() mb() +#include <linux/elfnote.h> + +#ifdef CONFIG_LTO +#define BUILD_COMPILER_LTO_INFO ELFNOTE32("Linux", LINUX_ELFNOTE_BUILD_LTO, 1) +#else +#define BUILD_COMPILER_LTO_INFO ELFNOTE32("Linux", LINUX_ELFNOTE_BUILD_LTO, 0) +#endifWith this approach BUILD_COMPILER_LTO_INFO won't be available `#ifdef __ASSEMBLER__`; we don't need it today, and perhaps YAGNI, but I think
That is true. I didn't add it since I don't feel it. BUILD_SALT also added to vdso binary which I feel we don't need it today.
I prefer how include/linux/build-salt.h defines LINUX_ELFNOTE_BUILD_SALT and keeps it isolated there. Similarly, I think it would be better to create a new header, say include/linux/elfnote-lto.h that is basically a copy of include/linux/build-salt.h, but with the relevant defines replaced
Having a separate header like elfnote-lto.h sounds okay. Originally I am reluctant to add a new header file, but maybe a new header file is much cleaner than otherwise.
with the LTO identifiers you add above. Then init/version.c and scripts/mod/modpost.c can include include/linux/elfnote-lto.h and you don't have to touch include/linux/build-salt.h and we can keep the elfnote "types" isolated to their respective headers (otherwise this approach reduces the usefulness of include/linux/build-salt.h even existing, IMO. Feels like it should just be merged into include/linux/elfnote.h entirely at that point).
The only "drawback" is the type values are scattered in different files which I am not really comfortable with it. But with consistent naming convention, all values can be easily searched so we may not have issue at all.
But, this is a much nicer approach! I forgot that elf notes were a thing!+ #include <asm/rwonce.h> #endif /* __LINUX_COMPILER_H */ diff --git a/include/linux/elfnote.h b/include/linux/elfnote.h index 04af7ac40b1a..f5ec2b50ab7d 100644 --- a/include/linux/elfnote.h +++ b/include/linux/elfnote.h @@ -100,5 +100,6 @@ * The types for "Linux" owned notes. */ #define LINUX_ELFNOTE_BUILD_SALT 0x100 +#define LINUX_ELFNOTE_BUILD_LTO 0x101 #endif /* _LINUX_ELFNOTE_H */ diff --git a/init/version.c b/init/version.c index 92afc782b043..a4f74b06fe78 100644 --- a/init/version.c +++ b/init/version.c @@ -9,6 +9,7 @@ #include <generated/compile.h> #include <linux/build-salt.h> +#include <linux/compiler.h> #include <linux/export.h> #include <linux/uts.h> #include <linux/utsname.h> @@ -45,3 +46,4 @@ const char linux_proc_banner[] = " (" LINUX_COMPILER ") %s\n"; BUILD_SALT; +BUILD_COMPILER_LTO_INFO; diff --git a/scripts/mod/modpost.c b/scripts/mod/modpost.c index 24725e50c7b4..713c0d5d5525 100644 --- a/scripts/mod/modpost.c +++ b/scripts/mod/modpost.c @@ -2195,6 +2195,7 @@ static void add_header(struct buffer *b, struct module *mod) buf_printf(b, "#include <linux/compiler.h>\n"); buf_printf(b, "\n"); buf_printf(b, "BUILD_SALT;\n"); + buf_printf(b, "BUILD_COMPILER_LTO_INFO;\n"); buf_printf(b, "\n"); buf_printf(b, "MODULE_INFO(vermagic, VERMAGIC_STRING);\n"); buf_printf(b, "MODULE_INFO(name, KBUILD_MODNAME);\n"); -- 2.30.2
