[PATCH 1/2] reduce symbol table for loaded modules

[PATCH 1/2] reduce symbol table for loaded modules

[Jan Beulich]

Discard all symbols not interesting for kallsyms use: absolute,
section, and in the common case (!KALLSYMS_ALL) data ones.

Signed-off-by: Jan Beulich <jbeulich@novell.com>

---
 include/linux/module.h |    4 +-
 kernel/module.c        |   85 +++++++++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 84 insertions(+), 5 deletions(-)

--- 2.6.31-rc1-module-strip.orig/include/linux/module.h
+++ 2.6.31-rc1-module-strip/include/linux/module.h
@@ -308,8 +308,8 @@ struct module
 
 #ifdef CONFIG_KALLSYMS
 	/* We keep the symbol and string tables for kallsyms. */
-	Elf_Sym *symtab;
-	unsigned int num_symtab;
+	Elf_Sym *symtab, *core_symtab;
+	unsigned int num_symtab, core_num_syms;
 	char *strtab;
 
 	/* Section attributes */
--- 2.6.31-rc1-module-strip.orig/kernel/module.c
+++ 2.6.31-rc1-module-strip/kernel/module.c
@@ -1841,10 +1841,67 @@ static char elf_type(const Elf_Sym *sym,
 	return '?';
 }
 
+static unsigned int filter_symbols(Elf_Sym *dst,
+				   const Elf_Sym *src, unsigned int nsrc,
+				   const Elf_Shdr *sechdrs, unsigned int shnum)
+{
+	unsigned int i, ndst;
+
+	if (dst)
+		*dst = *src;
+	for (ndst = i = 1; i < nsrc; ++i, ++src) {
+		const Elf_Shdr *sec;
+
+		if (src->st_shndx == SHN_UNDEF
+		    || src->st_shndx >= shnum
+		    || !src->st_name)
+			continue;
+		sec = sechdrs + src->st_shndx;
+		if (!(sec->sh_flags & SHF_ALLOC)
+#ifndef CONFIG_KALLSYMS_ALL
+		    || !(sec->sh_flags & SHF_EXECINSTR)
+#endif
+		    || (sec->sh_entsize & INIT_OFFSET_MASK))
+			continue;
+		if (dst)
+			dst[ndst] = *src;
+		++ndst;
+	}
+
+	return ndst;
+}
+
+static unsigned long layout_symtab(struct module *mod,
+				   Elf_Shdr *sechdrs,
+				   unsigned int symindex,
+				   const Elf_Ehdr *hdr,
+				   const char *secstrings)
+{
+	unsigned long symoffs;
+	Elf_Shdr *symsect = sechdrs + symindex;
+
+	symsect->sh_flags |= SHF_ALLOC;
+	symsect->sh_entsize = get_offset(mod, &mod->init_size, symsect,
+					 symindex) | INIT_OFFSET_MASK;
+	DEBUGP("\t%s\n", secstrings + symsect->sh_name);
+
+	symoffs = ALIGN(mod->core_size, symsect->sh_addralign ?: 1);
+	mod->core_size = symoffs
+			 + filter_symbols(NULL,
+			 		  (void *)hdr + symsect->sh_offset,
+					  symsect->sh_size / sizeof(Elf_Sym),
+					  sechdrs, hdr->e_shnum)
+			 * sizeof(Elf_Sym);
+
+	return symoffs;
+}
+
 static void add_kallsyms(struct module *mod,
 			 Elf_Shdr *sechdrs,
+			 unsigned int shnum,
 			 unsigned int symindex,
 			 unsigned int strindex,
+			 unsigned long symoffs,
 			 const char *secstrings)
 {
 	unsigned int i;
@@ -1857,12 +1914,26 @@ static void add_kallsyms(struct module *
 	for (i = 0; i < mod->num_symtab; i++)
 		mod->symtab[i].st_info
 			= elf_type(&mod->symtab[i], sechdrs, secstrings, mod);
+
+	mod->core_symtab = mod->module_core + symoffs;
+	mod->core_num_syms = filter_symbols(mod->core_symtab,
+					    mod->symtab, mod->num_symtab,
+					    sechdrs, shnum);
 }
 #else
+static inline unsigned long layout_symtab(struct module *mod,
+					  Elf_Shdr *sechdrs,
+					  unsigned int symindex,
+					  const Elf_Hdr *hdr,
+					  const char *secstrings)
+{
+}
 static inline void add_kallsyms(struct module *mod,
 				Elf_Shdr *sechdrs,
+				unsigned int shnum,
 				unsigned int symindex,
 				unsigned int strindex,
+				unsigned long symoffs,
 				const char *secstrings)
 {
 }
@@ -1938,6 +2009,9 @@ static noinline struct module *load_modu
 	struct module *mod;
 	long err = 0;
 	void *percpu = NULL, *ptr = NULL; /* Stops spurious gcc warning */
+#ifdef CONFIG_KALLSYMS
+	unsigned long symoffs;
+#endif
 	mm_segment_t old_fs;
 
 	DEBUGP("load_module: umod=%p, len=%lu, uargs=%p\n",
@@ -2020,8 +2094,7 @@ static noinline struct module *load_modu
 	sechdrs[infoindex].sh_flags &= ~(unsigned long)SHF_ALLOC;
 	sechdrs[versindex].sh_flags &= ~(unsigned long)SHF_ALLOC;
 #ifdef CONFIG_KALLSYMS
-	/* Keep symbol and string tables for decoding later. */
-	sechdrs[symindex].sh_flags |= SHF_ALLOC;
+	/* Keep string table for decoding later. */
 	sechdrs[strindex].sh_flags |= SHF_ALLOC;
 #endif
 
@@ -2088,6 +2161,7 @@ static noinline struct module *load_modu
 	   this is done generically; there doesn't appear to be any
 	   special cases for the architectures. */
 	layout_sections(mod, hdr, sechdrs, secstrings);
+	symoffs = layout_symtab(mod, sechdrs, symindex, hdr, secstrings);
 
 	/* Do the allocs. */
 	ptr = module_alloc_update_bounds(mod->core_size);
@@ -2296,7 +2370,8 @@ static noinline struct module *load_modu
 	percpu_modcopy(mod->percpu, (void *)sechdrs[pcpuindex].sh_addr,
 		       sechdrs[pcpuindex].sh_size);
 
-	add_kallsyms(mod, sechdrs, symindex, strindex, secstrings);
+	add_kallsyms(mod, sechdrs, hdr->e_shnum, symindex, strindex,
+		     symoffs, secstrings);
 
 	if (!mod->taints) {
 		struct _ddebug *debug;
@@ -2472,6 +2547,10 @@ SYSCALL_DEFINE3(init_module, void __user
 	/* Drop initial reference. */
 	module_put(mod);
 	trim_init_extable(mod);
+#ifdef CONFIG_KALLSYMS
+	mod->num_symtab = mod->core_num_syms;
+	mod->symtab = mod->core_symtab;
+#endif
 	module_free(mod, mod->module_init);
 	mod->module_init = NULL;
 	mod->init_size = 0;

Re: [PATCH 1/2] reduce symbol table for loaded modules

[Rusty Russell]

On Tue, 30 Jun 2009 09:38:43 pm Jan Beulich wrote:
> Discard all symbols not interesting for kallsyms use: absolute,
> section, and in the common case (!KALLSYMS_ALL) data ones.

I like the idea, but implementation could probably be polished a little bit.

> +static unsigned int filter_symbols(Elf_Sym *dst,
> +				   const Elf_Sym *src, unsigned int nsrc,
> +				   const Elf_Shdr *sechdrs, unsigned int shnum)

How about splitting the test out to an:

	bool is_core_symbol(const Elf_Sym *src, const Elf_Shdr *sechdrs, unsigned int shnum)

Then split filter_symbols into num_core_symbols() and copy_core_symbols().

> +static unsigned long layout_symtab(struct module *mod,

How about append_symbols() instead?

> +				   Elf_Shdr *sechdrs,
> +				   unsigned int symindex,
> +				   const Elf_Ehdr *hdr,
> +				   const char *secstrings)
> +{
> +	unsigned long symoffs;
> +	Elf_Shdr *symsect = sechdrs + symindex;
> +

Comment would help me here:

/* Put symbol section at end of init part of module. */
> +	symsect->sh_flags |= SHF_ALLOC;
> +	symsect->sh_entsize = get_offset(mod, &mod->init_size, symsect,
> +					 symindex) | INIT_OFFSET_MASK;
> +	DEBUGP("\t%s\n", secstrings + symsect->sh_name);
> +

Here too:
/* Append room for core symbols at end of core part. */
> +	symoffs = ALIGN(mod->core_size, symsect->sh_addralign ?: 1);
> +	mod->core_size = symoffs
> +			 + filter_symbols(NULL,
> +			 		  (void *)hdr + symsect->sh_offset,
> +					  symsect->sh_size / sizeof(Elf_Sym),
> +					  sechdrs, hdr->e_shnum)
> +			 * sizeof(Elf_Sym);
> +
> +	return symoffs;
> +}
> +
>  static void add_kallsyms(struct module *mod,
>  			 Elf_Shdr *sechdrs,
> +			 unsigned int shnum,
>  			 unsigned int symindex,
>  			 unsigned int strindex,
> +			 unsigned long symoffs,
>  			 const char *secstrings)
>  {
>  	unsigned int i;
> @@ -1857,12 +1914,26 @@ static void add_kallsyms(struct module *
>  	for (i = 0; i < mod->num_symtab; i++)
>  		mod->symtab[i].st_info
>  			= elf_type(&mod->symtab[i], sechdrs, secstrings, mod);
> +
> +	mod->core_symtab = mod->module_core + symoffs;
> +	mod->core_num_syms = filter_symbols(mod->core_symtab,
> +					    mod->symtab, mod->num_symtab,
> +					    sechdrs, shnum);

Hmm, could avoid symoffs if copy_core_symbols worked backwards.  Probably
not worthwhile tho.

>  	trim_init_extable(mod);
> +#ifdef CONFIG_KALLSYMS
> +	mod->num_symtab = mod->core_num_syms;
> +	mod->symtab = mod->core_symtab;
> +#endif
>  	module_free(mod, mod->module_init);
>  	mod->module_init = NULL;
>  	mod->init_size = 0;

core_num_syms/core_symtab is really a temporary; we should note that in
the module.h header next to those fields.

Thanks!
Rusty.