lkml.org 
[lkml]   [2011]   [Oct]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH][RFC] jump_labels/x86: Use either 5 byte or 2 byte jumps
On Fri, Oct 07, 2011 at 01:09:32PM -0400, Steven Rostedt wrote:
> Note, this is just hacked together and needs to be cleaned up. Please do
> not comment on formatting or other sloppiness of this patch. I know it's
> sloppy and I left debug statements in. I want the comments to be on the
> idea of the patch.
>
> I created a new file called scripts/update_jump_label.[ch] based on some
> of the work of recordmcount.c. This is executed at build time on all
> object files just like recordmcount is. But it does not add any new
> sections, it just modifies the code at build time to convert all jump
> labels into nops.
>
> The idea is in arch/x86/include/asm/jump_label.h to not place a nop, but
> instead to insert a jmp to the label. Depending on how gcc optimizes the
> code, the jmp will be either end up being a 2 byte or 5 byte jump.
>
> After an object is compiled, update_jump_label is executed on this file
> and it reads the ELF relocation table to find the jump label locations
> and examines what jump was used. It then converts the jump into either a
> 2 byte or 5 byte nop that is appropriate.
>
> At boot time, the jump labels no longer need to be converted (although
> we may do so in the future to use better nops depending on the machine
> that is used). When jump labels are enabled, the code is examined to see
> if a two byte or 5 byte version was used, and the appropriate update is
> made.
>
> I just booted this patch and it worked. I was able to enable and disable
> trace points using jump labels. Benchmarks are welcomed :)
>
> Comments and thoughts?
>

Generally, I really like it, I guess b/c I suggested it :) I'll try and
run some workloads on it - A real simple one, I used recently was putting
a single jump label in 'getppid()' and then calling it in a loop - I
wonder if the short nop vs long nop would show up there, as a baseline
test. (fwiw, the jump label vs. no jump label for this test was anywhere
b/w 1-5% improvement).

Anyways, some comments below.

> -- Steve
>
> Sloppy-signed-off-by: Steven Rostedt <rostedt@goodmis.org>
>
> diff --git a/Makefile b/Makefile
> index 31f967c..8368f42 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -245,7 +245,7 @@ CONFIG_SHELL := $(shell if [ -x "$$BASH" ]; then echo $$BASH; \
>
> HOSTCC = gcc
> HOSTCXX = g++
> -HOSTCFLAGS = -Wall -Wmissing-prototypes -Wstrict-prototypes -O2 -fomit-frame-pointer
> +HOSTCFLAGS = -Wall -Wmissing-prototypes -Wstrict-prototypes -g -fomit-frame-pointer
> HOSTCXXFLAGS = -O2
>
> # Decide whether to build built-in, modular, or both.
> @@ -611,6 +611,13 @@ ifdef CONFIG_DYNAMIC_FTRACE
> endif
> endif
>
> +ifdef CONFIG_JUMP_LABEL
> + ifdef CONFIG_HAVE_BUILD_TIME_JUMP_LABEL
> + BUILD_UPDATE_JUMP_LABEL := y
> + export BUILD_UPDATE_JUMP_LABEL
> + endif
> +endif
> +
> # We trigger additional mismatches with less inlining
> ifdef CONFIG_DEBUG_SECTION_MISMATCH
> KBUILD_CFLAGS += $(call cc-option, -fno-inline-functions-called-once)
> diff --git a/arch/Kconfig b/arch/Kconfig
> index 4b0669c..8fa6934 100644
> --- a/arch/Kconfig
> +++ b/arch/Kconfig
> @@ -169,6 +169,12 @@ config HAVE_PERF_EVENTS_NMI
> subsystem. Also has support for calculating CPU cycle events
> to determine how many clock cycles in a given period.
>
> +config HAVE_BUILD_TIME_JUMP_LABEL
> + bool
> + help
> + If an arch uses scripts/update_jump_label to patch in jump nops
> + at build time, then it must enable this option.
> +
> config HAVE_ARCH_JUMP_LABEL
> bool
>
> diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> index 6a47bb2..6de726a 100644
> --- a/arch/x86/Kconfig
> +++ b/arch/x86/Kconfig
> @@ -61,6 +61,7 @@ config X86
> select HAVE_ARCH_KMEMCHECK
> select HAVE_USER_RETURN_NOTIFIER
> select HAVE_ARCH_JUMP_LABEL
> + select HAVE_BUILD_TIME_JUMP_LABEL
> select HAVE_TEXT_POKE_SMP
> select HAVE_GENERIC_HARDIRQS
> select HAVE_SPARSE_IRQ
> diff --git a/arch/x86/include/asm/jump_label.h b/arch/x86/include/asm/jump_label.h
> index a32b18c..872b3e1 100644
> --- a/arch/x86/include/asm/jump_label.h
> +++ b/arch/x86/include/asm/jump_label.h
> @@ -14,7 +14,7 @@
> static __always_inline bool arch_static_branch(struct jump_label_key *key)
> {
> asm goto("1:"
> - JUMP_LABEL_INITIAL_NOP
> + "jmp %l[l_yes]\n"
> ".pushsection __jump_table, \"aw\" \n\t"
> _ASM_ALIGN "\n\t"
> _ASM_PTR "1b, %l[l_yes], %c0 \n\t"
> diff --git a/arch/x86/kernel/jump_label.c b/arch/x86/kernel/jump_label.c
> index 3fee346..1f7f88f 100644
> --- a/arch/x86/kernel/jump_label.c
> +++ b/arch/x86/kernel/jump_label.c
> @@ -16,34 +16,75 @@
>
> #ifdef HAVE_JUMP_LABEL
>
> +static unsigned char nop_short[] = { P6_NOP2 };
> +
> union jump_code_union {
> char code[JUMP_LABEL_NOP_SIZE];
> struct {
> char jump;
> int offset;
> } __attribute__((packed));
> + struct {
> + char jump_short;
> + char offset_short;
> + } __attribute__((packed));
> };
>
> void arch_jump_label_transform(struct jump_entry *entry,
> enum jump_label_type type)
> {
> union jump_code_union code;
> + unsigned char op;
> + unsigned size;
> + unsigned char nop;
> +
> + /* Use probe_kernel_read()? */
> + op = *(unsigned char *)entry->code;
> + nop = ideal_nops[NOP_ATOMIC5][0];
>
> if (type == JUMP_LABEL_ENABLE) {
> - code.jump = 0xe9;
> - code.offset = entry->target -
> - (entry->code + JUMP_LABEL_NOP_SIZE);
> - } else
> - memcpy(&code, ideal_nops[NOP_ATOMIC5], JUMP_LABEL_NOP_SIZE);
> + if (op == 0xe9 || op == 0xeb)
> + /* Already enabled. Warn? */
> + return;
> +

Using the jump_label_inc/dec interface this shouldn't happen, I would
have it be BUG


> + /* FIXME for all archs */
> + if (op == nop_short[0]) {
> + size = 2;
> + code.jump_short = 0xeb;
> + code.offset = entry->target -
> + (entry->code + 2);
> + /* Check for overflow ? */
> + } else if (op == nop) {
> + size = JUMP_LABEL_NOP_SIZE;
> + code.jump = 0xe9;
> + code.offset = entry->target - (entry->code + size);
> + } else
> + return; /* WARN ? */

same here, at least WARN, more likely BUG()

> +
> + } else {
> + if (op == nop_short[0] || nop)
> + /* Already disabled, warn? */
> + return;
> +

same here.

> + if (op == 0xe9) {
> + size = JUMP_LABEL_NOP_SIZE;
> + memcpy(&code, ideal_nops[NOP_ATOMIC5], size);
> + } else if (op == 0xeb) {
> + size = 2;
> + memcpy(&code, nop_short, size);
> + } else
> + return; /* WARN ? */

same here

> + }
> get_online_cpus();
> mutex_lock(&text_mutex);
> - text_poke_smp((void *)entry->code, &code, JUMP_LABEL_NOP_SIZE);
> + text_poke_smp((void *)entry->code, &code, size);
> mutex_unlock(&text_mutex);
> put_online_cpus();
> }
>
> void arch_jump_label_text_poke_early(jump_label_t addr)
> {
> + return;
> text_poke_early((void *)addr, ideal_nops[NOP_ATOMIC5],
> JUMP_LABEL_NOP_SIZE);
> }

hmmm...we spent a bunch of time selecting the 'ideal' run-time noops I
wouldn't want to drop that work.

> diff --git a/scripts/Makefile b/scripts/Makefile
> index df7678f..738b65c 100644
> --- a/scripts/Makefile
> +++ b/scripts/Makefile
> @@ -13,6 +13,7 @@ hostprogs-$(CONFIG_LOGO) += pnmtologo
> hostprogs-$(CONFIG_VT) += conmakehash
> hostprogs-$(CONFIG_IKCONFIG) += bin2c
> hostprogs-$(BUILD_C_RECORDMCOUNT) += recordmcount
> +hostprogs-$(BUILD_UPDATE_JUMP_LABEL) += update_jump_label
>
> always := $(hostprogs-y) $(hostprogs-m)
>
> diff --git a/scripts/Makefile.build b/scripts/Makefile.build
> index a0fd502..bc0d89b 100644
> --- a/scripts/Makefile.build
> +++ b/scripts/Makefile.build
> @@ -258,6 +258,15 @@ cmd_modversions = \
> fi;
> endif
>
> +ifdef BUILD_UPDATE_JUMP_LABEL
> +update_jump_label_source := $(srctree)/scripts/update_jump_label.c \
> + $(srctree)/scripts/update_jump_label.h
> +cmd_update_jump_label = \
> + if [ $(@) != "scripts/mod/empty.o" ]; then \
> + $(objtree)/scripts/update_jump_label "$(@)"; \
> + fi;
> +endif
> +
> ifdef CONFIG_FTRACE_MCOUNT_RECORD
> ifdef BUILD_C_RECORDMCOUNT
> ifeq ("$(origin RECORDMCOUNT_WARN)", "command line")
> @@ -294,6 +303,7 @@ define rule_cc_o_c
> $(cmd_modversions) \
> $(call echo-cmd,record_mcount) \
> $(cmd_record_mcount) \
> + $(cmd_update_jump_label) \
> scripts/basic/fixdep $(depfile) $@ '$(call make-cmd,cc_o_c)' > \
> $(dot-target).tmp; \
> rm -f $(depfile); \
> @@ -301,13 +311,14 @@ define rule_cc_o_c
> endef
>
> # Built-in and composite module parts
> -$(obj)/%.o: $(src)/%.c $(recordmcount_source) FORCE
> +$(obj)/%.o: $(src)/%.c $(recordmcount_source) $(update_jump_label_source) FORCE
> $(call cmd,force_checksrc)
> $(call if_changed_rule,cc_o_c)
>
> # Single-part modules are special since we need to mark them in $(MODVERDIR)
>
> -$(single-used-m): $(obj)/%.o: $(src)/%.c $(recordmcount_source) FORCE
> +$(single-used-m): $(obj)/%.o: $(src)/%.c $(recordmcount_source) \
> + $(update_jump_label_source) FORCE
> $(call cmd,force_checksrc)
> $(call if_changed_rule,cc_o_c)
> @{ echo $(@:.o=.ko); echo $@; } > $(MODVERDIR)/$(@F:.o=.mod)
> diff --git a/scripts/update_jump_label.c b/scripts/update_jump_label.c
> new file mode 100644
> index 0000000..86e17bc
> --- /dev/null
> +++ b/scripts/update_jump_label.c
> @@ -0,0 +1,349 @@
> +/*
> + * update_jump_label.c: replace jmps with nops at compile time.
> + * Copyright 2010 Steven Rostedt <srostedt@redhat.com>, Red Hat Inc.
> + * Parsing of the elf file was influenced by recordmcount.c
> + * originally written by and copyright to John F. Reiser <jreiser@BitWagon.com>.
> + */
> +
> +/*
> + * Note, this code is originally designed for x86, but may be used by
> + * other archs to do the nop updates at compile time instead of at boot time.
> + * X86 uses this as an optimization, as jmps can be either 2 bytes or 5 bytes.
> + * Inserting a 2 byte where possible helps with both CPU performance and
> + * icache strain.
> + */
> +#include <sys/types.h>
> +#include <sys/mman.h>
> +#include <sys/stat.h>
> +#include <getopt.h>
> +#include <elf.h>
> +#include <fcntl.h>
> +#include <setjmp.h>
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include <stdarg.h>
> +#include <string.h>
> +#include <unistd.h>
> +
> +static int fd_map; /* File descriptor for file being modified. */
> +static struct stat sb; /* Remember .st_size, etc. */
> +static int mmap_failed; /* Boolean flag. */
> +
> +static void die(const char *err, const char *fmt, ...)
> +{
> + va_list ap;
> +
> + if (err)
> + perror(err);
> +
> + if (fmt) {
> + va_start(ap, fmt);
> + fprintf(stderr, "Fatal error: ");
> + vfprintf(stderr, fmt, ap);
> + fprintf(stderr, "\n");
> + va_end(ap);
> + }
> +
> + exit(1);
> +}
> +
> +static void usage(char **argv)
> +{
> + char *arg = argv[0];
> + char *p = arg+strlen(arg);
> +
> + while (p >= arg && *p != '/')
> + p--;
> + p++;
> +
> + printf("usage: %s file\n"
> + "\n",p);
> + exit(-1);
> +}
> +
> +/* w8rev, w8nat, ...: Handle endianness. */
> +
> +static uint64_t w8rev(uint64_t const x)
> +{
> + return ((0xff & (x >> (0 * 8))) << (7 * 8))
> + | ((0xff & (x >> (1 * 8))) << (6 * 8))
> + | ((0xff & (x >> (2 * 8))) << (5 * 8))
> + | ((0xff & (x >> (3 * 8))) << (4 * 8))
> + | ((0xff & (x >> (4 * 8))) << (3 * 8))
> + | ((0xff & (x >> (5 * 8))) << (2 * 8))
> + | ((0xff & (x >> (6 * 8))) << (1 * 8))
> + | ((0xff & (x >> (7 * 8))) << (0 * 8));
> +}
> +
> +static uint32_t w4rev(uint32_t const x)
> +{
> + return ((0xff & (x >> (0 * 8))) << (3 * 8))
> + | ((0xff & (x >> (1 * 8))) << (2 * 8))
> + | ((0xff & (x >> (2 * 8))) << (1 * 8))
> + | ((0xff & (x >> (3 * 8))) << (0 * 8));
> +}
> +
> +static uint32_t w2rev(uint16_t const x)
> +{
> + return ((0xff & (x >> (0 * 8))) << (1 * 8))
> + | ((0xff & (x >> (1 * 8))) << (0 * 8));
> +}
> +
> +static uint64_t w8nat(uint64_t const x)
> +{
> + return x;
> +}
> +
> +static uint32_t w4nat(uint32_t const x)
> +{
> + return x;
> +}
> +
> +static uint32_t w2nat(uint16_t const x)
> +{
> + return x;
> +}
> +
> +static uint64_t (*w8)(uint64_t);
> +static uint32_t (*w)(uint32_t);
> +static uint32_t (*w2)(uint16_t);
> +
> +/* ulseek, uread, ...: Check return value for errors. */
> +
> +static off_t
> +ulseek(int const fd, off_t const offset, int const whence)
> +{
> + off_t const w = lseek(fd, offset, whence);
> + if (w == (off_t)-1)
> + die("lseek", NULL);
> +
> + return w;
> +}
> +
> +static size_t
> +uread(int const fd, void *const buf, size_t const count)
> +{
> + size_t const n = read(fd, buf, count);
> + if (n != count)
> + die("read", NULL);
> +
> + return n;
> +}
> +
> +static size_t
> +uwrite(int const fd, void const *const buf, size_t const count)
> +{
> + size_t const n = write(fd, buf, count);
> + if (n != count)
> + die("write", NULL);
> +
> + return n;
> +}
> +
> +static void *
> +umalloc(size_t size)
> +{
> + void *const addr = malloc(size);
> + if (addr == 0)
> + die("malloc", "malloc failed: %zu bytes\n", size);
> +
> + return addr;
> +}
> +
> +/*
> + * Get the whole file as a programming convenience in order to avoid
> + * malloc+lseek+read+free of many pieces. If successful, then mmap
> + * avoids copying unused pieces; else just read the whole file.
> + * Open for both read and write; new info will be appended to the file.
> + * Use MAP_PRIVATE so that a few changes to the in-memory ElfXX_Ehdr
> + * do not propagate to the file until an explicit overwrite at the last.
> + * This preserves most aspects of consistency (all except .st_size)
> + * for simultaneous readers of the file while we are appending to it.
> + * However, multiple writers still are bad. We choose not to use
> + * locking because it is expensive and the use case of kernel build
> + * makes multiple writers unlikely.
> + */
> +static void *mmap_file(char const *fname)
> +{
> + void *addr;
> +
> + fd_map = open(fname, O_RDWR);
> + if (fd_map < 0 || fstat(fd_map, &sb) < 0)
> + die(fname, "failed to open file");
> +
> + if (!S_ISREG(sb.st_mode))
> + die(NULL, "not a regular file: %s\n", fname);
> +
> + addr = mmap(0, sb.st_size, PROT_READ|PROT_WRITE, MAP_PRIVATE,
> + fd_map, 0);
> +
> + mmap_failed = 0;
> + if (addr == MAP_FAILED) {
> + mmap_failed = 1;
> + addr = umalloc(sb.st_size);
> + uread(fd_map, addr, sb.st_size);
> + }
> + return addr;
> +}
> +
> +static void munmap_file(void *addr)
> +{
> + if (!mmap_failed)
> + munmap(addr, sb.st_size);
> + else
> + free(addr);
> + close(fd_map);
> +}
> +
> +static unsigned char ideal_nop5_x86_64[5] = { 0x0f, 0x1f, 0x44, 0x00, 0x00 };
> +static unsigned char ideal_nop5_x86_32[5] = { 0x3e, 0x8d, 0x74, 0x26, 0x00 };
> +static unsigned char ideal_nop2_x86[2] = { 0x66, 0x99 };
> +static unsigned char *ideal_nop;
> +
> +static int (*make_nop)(void *map, size_t const offset);
> +
> +static int make_nop_x86(void *map, size_t const offset)
> +{
> + unsigned char *op;
> + unsigned char *nop;
> + int size;
> +
> + /* Determine which type of jmp this is 2 byte or 5. */
> + op = map + offset;
> + switch (*op) {
> + case 0xeb: /* 2 byte */
> + size = 2;
> + nop = ideal_nop2_x86;
> + break;
> + case 0xe9: /* 5 byte */
> + size = 5;
> + nop = ideal_nop;
> + break;
> + default:
> + die(NULL, "Bad jump label section\n");
> + }
> +
> + /* convert to nop */
> + ulseek(fd_map, offset, SEEK_SET);
> + uwrite(fd_map, nop, size);
> + return 0;
> +}
> +
> +/* 32 bit and 64 bit are very similar */
> +#include "update_jump_label.h"
> +#define UPDATE_JUMP_LABEL_64
> +#include "update_jump_label.h"
> +
> +static int do_file(const char *fname)
> +{
> + Elf32_Ehdr *const ehdr = mmap_file(fname);
> + unsigned int reltype = 0;
> +
> + w = w4nat;
> + w2 = w2nat;
> + w8 = w8nat;
> + switch (ehdr->e_ident[EI_DATA]) {
> + static unsigned int const endian = 1;
> + default:
> + die(NULL, "unrecognized ELF data encoding %d: %s\n",
> + ehdr->e_ident[EI_DATA], fname);
> + break;
> + case ELFDATA2LSB:
> + if (*(unsigned char const *)&endian != 1) {
> + /* main() is big endian, file.o is little endian. */
> + w = w4rev;
> + w2 = w2rev;
> + w8 = w8rev;
> + }
> + break;
> + case ELFDATA2MSB:
> + if (*(unsigned char const *)&endian != 0) {
> + /* main() is little endian, file.o is big endian. */
> + w = w4rev;
> + w2 = w2rev;
> + w8 = w8rev;
> + }
> + break;
> + } /* end switch */
> +
> + if (memcmp(ELFMAG, ehdr->e_ident, SELFMAG) != 0 ||
> + w2(ehdr->e_type) != ET_REL ||
> + ehdr->e_ident[EI_VERSION] != EV_CURRENT)
> + die(NULL, "unrecognized ET_REL file %s\n", fname);
> +
> + switch (w2(ehdr->e_machine)) {
> + default:
> + die(NULL, "unrecognized e_machine %d %s\n",
> + w2(ehdr->e_machine), fname);
> + break;
> + case EM_386:
> + reltype = R_386_32;
> + make_nop = make_nop_x86;
> + ideal_nop = ideal_nop5_x86_32;
> + break;
> + case EM_ARM: reltype = R_ARM_ABS32;
> + break;
> + case EM_IA_64: reltype = R_IA64_IMM64; break;
> + case EM_MIPS: /* reltype: e_class */ break;
> + case EM_PPC: reltype = R_PPC_ADDR32; break;
> + case EM_PPC64: reltype = R_PPC64_ADDR64; break;
> + case EM_S390: /* reltype: e_class */ break;
> + case EM_SH: reltype = R_SH_DIR32; break;
> + case EM_SPARCV9: reltype = R_SPARC_64; break;
> + case EM_X86_64:
> + make_nop = make_nop_x86;
> + ideal_nop = ideal_nop5_x86_64;
> + reltype = R_X86_64_64;
> + break;
> + } /* end switch */
> +
> + switch (ehdr->e_ident[EI_CLASS]) {
> + default:
> + die(NULL, "unrecognized ELF class %d %s\n",
> + ehdr->e_ident[EI_CLASS], fname);
> + break;
> + case ELFCLASS32:
> + if (w2(ehdr->e_ehsize) != sizeof(Elf32_Ehdr)
> + || w2(ehdr->e_shentsize) != sizeof(Elf32_Shdr))
> + die(NULL, "unrecognized ET_REL file: %s\n", fname);
> +
> + if (w2(ehdr->e_machine) == EM_S390) {
> + reltype = R_390_32;
> + }
> + if (w2(ehdr->e_machine) == EM_MIPS) {
> + reltype = R_MIPS_32;
> + }
> + do_func32(ehdr, fname, reltype);
> + break;
> + case ELFCLASS64: {
> + Elf64_Ehdr *const ghdr = (Elf64_Ehdr *)ehdr;
> + if (w2(ghdr->e_ehsize) != sizeof(Elf64_Ehdr)
> + || w2(ghdr->e_shentsize) != sizeof(Elf64_Shdr))
> + die(NULL, "unrecognized ET_REL file: %s\n", fname);
> +
> + if (w2(ghdr->e_machine) == EM_S390)
> + reltype = R_390_64;
> +
> +#if 0
> + if (w2(ghdr->e_machine) == EM_MIPS) {
> + reltype = R_MIPS_64;
> + Elf64_r_sym = MIPS64_r_sym;
> + }
> +#endif
> + do_func64(ghdr, fname, reltype);
> + break;
> + }
> + } /* end switch */
> +
> + munmap_file(ehdr);
> + return 0;
> +}
> +
> +int main (int argc, char **argv)
> +{
> + if (argc != 2)
> + usage(argv);
> +
> + return do_file(argv[1]);
> +}
> +
> diff --git a/scripts/update_jump_label.h b/scripts/update_jump_label.h
> new file mode 100644
> index 0000000..6ff9846
> --- /dev/null
> +++ b/scripts/update_jump_label.h
> @@ -0,0 +1,322 @@
> +/*
> + * recordmcount.h
> + *
> + * This code was taken out of recordmcount.c written by
> + * Copyright 2009 John F. Reiser <jreiser@BitWagon.com>. All rights reserved.
> + *
> + * The original code had the same algorithms for both 32bit
> + * and 64bit ELF files, but the code was duplicated to support
> + * the difference in structures that were used. This
> + * file creates a macro of everything that is different between
> + * the 64 and 32 bit code, such that by including this header
> + * twice we can create both sets of functions by including this
> + * header once with RECORD_MCOUNT_64 undefined, and again with
> + * it defined.
> + *
> + * This conversion to macros was done by:
> + * Copyright 2010 Steven Rostedt <srostedt@redhat.com>, Red Hat Inc.
> + *
> + * Licensed under the GNU General Public License, version 2 (GPLv2).
> + */
> +
> +#undef EBITS
> +#undef _w
> +#undef _align
> +#undef _size
> +
> +#ifdef UPDATE_JUMP_LABEL_64
> +# define EBITS 64
> +# define _w w8
> +# define _align 7u
> +# define _size 8
> +#else
> +# define EBITS 32
> +# define _w w
> +# define _align 3u
> +# define _size 4
> +#endif
> +
> +#define _FBITS(x, e) x##e
> +#define FBITS(x, e) _FBITS(x,e)
> +#define FUNC(x) FBITS(x,EBITS)
> +
> +#undef Elf_Addr
> +#undef Elf_Ehdr
> +#undef Elf_Shdr
> +#undef Elf_Rel
> +#undef Elf_Rela
> +#undef Elf_Sym
> +#undef ELF_R_SYM
> +#undef ELF_R_TYPE
> +
> +#define __ATTACH(x,y,z) x##y##z
> +#define ATTACH(x,y,z) __ATTACH(x,y,z)
> +
> +#define Elf_Addr ATTACH(Elf,EBITS,_Addr)
> +#define Elf_Ehdr ATTACH(Elf,EBITS,_Ehdr)
> +#define Elf_Shdr ATTACH(Elf,EBITS,_Shdr)
> +#define Elf_Rel ATTACH(Elf,EBITS,_Rel)
> +#define Elf_Rela ATTACH(Elf,EBITS,_Rela)
> +#define Elf_Sym ATTACH(Elf,EBITS,_Sym)
> +#define uint_t ATTACH(uint,EBITS,_t)
> +#define ELF_R_SYM ATTACH(ELF,EBITS,_R_SYM)
> +#define ELF_R_TYPE ATTACH(ELF,EBITS,_R_TYPE)
> +
> +#undef get_shdr
> +#define get_shdr(ehdr) ((Elf_Shdr *)(_w((ehdr)->e_shoff) + (void *)(ehdr)))
> +
> +#undef get_section_loc
> +#define get_section_loc(ehdr, shdr)(_w((shdr)->sh_offset) + (void *)(ehdr))
> +
> +/* Functions and pointers that do_file() may override for specific e_machine. */
> +
> +#if 0
> +static uint_t FUNC(fn_ELF_R_SYM)(Elf_Rel const *rp)
> +{
> + return ELF_R_SYM(_w(rp->r_info));
> +}
> +static uint_t (*FUNC(Elf_r_sym))(Elf_Rel const *rp) = FUNC(fn_ELF_R_SYM);
> +#endif
> +
> +static void FUNC(get_sym_str_and_relp)(Elf_Shdr const *const relhdr,
> + Elf_Ehdr const *const ehdr,
> + Elf_Sym const **sym0,
> + char const **str0,
> + Elf_Rel const **relp)
> +{
> + Elf_Shdr *const shdr0 = get_shdr(ehdr);
> + unsigned const symsec_sh_link = w(relhdr->sh_link);
> + Elf_Shdr const *const symsec = &shdr0[symsec_sh_link];
> + Elf_Shdr const *const strsec = &shdr0[w(symsec->sh_link)];
> + Elf_Rel const *const rel0 =
> + (Elf_Rel const *)get_section_loc(ehdr, relhdr);
> +
> + *sym0 = (Elf_Sym const *)get_section_loc(ehdr, symsec);
> +
> + *str0 = (char const *)get_section_loc(ehdr, strsec);
> +
> + *relp = rel0;
> +}
> +
> +/*
> + * Read the relocation table again, but this time its called on sections
> + * that are not going to be traced. The mcount calls here will be converted
> + * into nops.
> + */
> +static void FUNC(nop_jump_label)(Elf_Shdr const *const relhdr,
> + Elf_Ehdr const *const ehdr,
> + const char *const txtname)
> +{
> + Elf_Shdr *const shdr0 = get_shdr(ehdr);
> + Elf_Sym const *sym0;
> + char const *str0;
> + Elf_Rel const *relp;
> + Elf_Rela const *relap;
> + Elf_Shdr const *const shdr = &shdr0[w(relhdr->sh_info)];
> + unsigned rel_entsize = w(relhdr->sh_entsize);
> + unsigned const nrel = _w(relhdr->sh_size) / rel_entsize;
> + int t;
> +
> + FUNC(get_sym_str_and_relp)(relhdr, ehdr, &sym0, &str0, &relp);
> +
> + for (t = nrel; t > 0; t -= 3) {
> + int ret = -1;
> +
> + relap = (Elf_Rela const *)relp;
> + printf("rel offset=%lx info=%lx sym=%lx type=%lx addend=%lx\n",
> + (long)relap->r_offset, (long)relap->r_info,
> + (long)ELF_R_SYM(relap->r_info),
> + (long)ELF_R_TYPE(relap->r_info),
> + (long)relap->r_addend);
> +
> + if (0 && make_nop)
> + ret = make_nop((void *)ehdr, shdr->sh_offset + relp->r_offset);
> +
> + /* jump label sections are paired in threes */
> + relp = (Elf_Rel const *)(rel_entsize * 3 + (void *)relp);
> + }
> +}
> +
> +/* Evade ISO C restriction: no declaration after statement in has_rel_mcount. */
> +static char const *
> +FUNC(__has_rel_jump_table)(Elf_Shdr const *const relhdr, /* is SHT_REL or SHT_RELA */
> + Elf_Shdr const *const shdr0,
> + char const *const shstrtab,
> + char const *const fname)
> +{
> + /* .sh_info depends on .sh_type == SHT_REL[,A] */
> + Elf_Shdr const *const txthdr = &shdr0[w(relhdr->sh_info)];
> + char const *const txtname = &shstrtab[w(txthdr->sh_name)];
> +
> + if (strcmp("__jump_table", txtname) == 0) {
> + fprintf(stderr, "warning: __mcount_loc already exists: %s\n",
> + fname);
> +// succeed_file();
> + }
> + if (w(txthdr->sh_type) != SHT_PROGBITS ||
> + !(w(txthdr->sh_flags) & SHF_EXECINSTR))
> + return NULL;
> + return txtname;
> +}
> +
> +static char const *FUNC(has_rel_jump_table)(Elf_Shdr const *const relhdr,
> + Elf_Shdr const *const shdr0,
> + char const *const shstrtab,
> + char const *const fname)
> +{
> + if (w(relhdr->sh_type) != SHT_REL && w(relhdr->sh_type) != SHT_RELA)
> + return NULL;
> + return FUNC(__has_rel_jump_table)(relhdr, shdr0, shstrtab, fname);
> +}
> +
> +/* Find relocation section hdr for a given section */
> +static const Elf_Shdr *
> +FUNC(find_relhdr)(const Elf_Ehdr *ehdr, const Elf_Shdr *shdr)
> +{
> + const Elf_Shdr *shdr0 = get_shdr(ehdr);
> + int nhdr = w2(ehdr->e_shnum);
> + const Elf_Shdr *hdr;
> + int i;
> +
> + for (hdr = shdr0, i = 0; i < nhdr; hdr = &shdr0[++i]) {
> + if (w(hdr->sh_type) != SHT_REL &&
> + w(hdr->sh_type) != SHT_RELA)
> + continue;
> +
> + /*
> + * The relocation section's info field holds
> + * the section index that it represents.
> + */
> + if (shdr == &shdr0[w(hdr->sh_info)])
> + return hdr;
> + }
> + return NULL;
> +}
> +
> +/* Find a section headr based on name and type */
> +static const Elf_Shdr *
> +FUNC(find_shdr)(const Elf_Ehdr *ehdr, const char *name, uint_t type)
> +{
> + const Elf_Shdr *shdr0 = get_shdr(ehdr);
> + const Elf_Shdr *shstr = &shdr0[w2(ehdr->e_shstrndx)];
> + const char *shstrtab = (char *)get_section_loc(ehdr, shstr);
> + int nhdr = w2(ehdr->e_shnum);
> + const Elf_Shdr *hdr;
> + const char *hdrname;
> + int i;
> +
> + for (hdr = shdr0, i = 0; i < nhdr; hdr = &shdr0[++i]) {
> + if (w(hdr->sh_type) != type)
> + continue;
> +
> + /* If we are just looking for a section by type (ie. SYMTAB) */
> + if (!name)
> + return hdr;
> +
> + hdrname = &shstrtab[w(hdr->sh_name)];
> + if (strcmp(hdrname, name) == 0)
> + return hdr;
> + }
> + return NULL;
> +}
> +
> +static void
> +FUNC(section_update)(const Elf_Ehdr *ehdr, const Elf_Shdr *symhdr,
> + unsigned shtype, const Elf_Rel *rel, void *data)
> +{
> + const Elf_Shdr *shdr0 = get_shdr(ehdr);
> + const Elf_Shdr *targethdr;
> + const Elf_Rela *rela;
> + const Elf_Sym *syment;
> + uint_t offset = _w(rel->r_offset);
> + uint_t info = _w(rel->r_info);
> + uint_t sym = ELF_R_SYM(info);
> + uint_t type = ELF_R_TYPE(info);
> + uint_t addend;
> + uint_t targetloc;
> +
> + if (shtype == SHT_RELA) {
> + rela = (const Elf_Rela *)rel;
> + addend = _w(rela->r_addend);
> + } else
> + addend = _w(*(unsigned short *)(data + offset));
> +
> + syment = (const Elf_Sym *)get_section_loc(ehdr, symhdr);
> + targethdr = &shdr0[w2(syment[sym].st_shndx)];
> + targetloc = _w(targethdr->sh_offset);
> +
> + /* TODO, need a separate function for all archs */
> + if (type != R_386_32)
> + die(NULL, "Arch relocation type %d not supported", type);
> +
> + targetloc += addend;
> +
> +#if 1
> + printf("offset=%x target=%x shoffset=%x add=%x\n",
> + offset, targetloc, _w(targethdr->sh_offset), addend);
> +#endif
> + *(uint_t *)(data + offset) = targetloc;
> +}
> +
> +/* Overall supervision for Elf32 ET_REL file. */
> +static void
> +FUNC(do_func)(Elf_Ehdr *ehdr, char const *const fname, unsigned const reltype)
> +{
> + const Elf_Shdr *jlshdr;
> + const Elf_Shdr *jlrhdr;
> + const Elf_Shdr *symhdr;
> + const Elf_Rel *rel;
> + unsigned size;
> + unsigned cnt;
> + unsigned i;
> + uint_t type;
> + void *jdata;
> + void *data;
> +
> + jlshdr = FUNC(find_shdr)(ehdr, "__jump_table", SHT_PROGBITS);
> + if (!jlshdr)
> + return;
> +
> + jlrhdr = FUNC(find_relhdr)(ehdr, jlshdr);
> + if (!jlrhdr)
> + return;
> +
> + /*
> + * Create and fill in the __jump_table section and use it to
> + * find the offsets into the text that we want to update.
> + * We create it so that we do not depend on the order of the
> + * relocations, and use the table directly, as it is broken
> + * up into sections.
> + */
> + size = _w(jlshdr->sh_size);
> + data = umalloc(size);
> +
> + jdata = (void *)get_section_loc(ehdr, jlshdr);
> + memcpy(data, jdata, size);
> +
> + cnt = _w(jlrhdr->sh_size) / w(jlrhdr->sh_entsize);
> +
> + rel = (const Elf_Rel *)get_section_loc(ehdr, jlrhdr);
> +
> + /* Is this as Rel or Rela? */
> + type = w(jlrhdr->sh_type);
> +
> + symhdr = FUNC(find_shdr)(ehdr, NULL, SHT_SYMTAB);
> +
> + for (i = 0; i < cnt; i++) {
> + FUNC(section_update)(ehdr, symhdr, type, rel, data);
> + rel = (void *)rel + w(jlrhdr->sh_entsize);
> + }
> +
> + /*
> + * This is specific to x86. The jump_table is stored in three
> + * long words. The first is the location of the jmp target we
> + * must update.
> + */
> + cnt = size / sizeof(uint_t);
> +
> + for (i = 0; i < cnt; i += 3)
> + if (0)make_nop((void *)ehdr, *(uint_t *)(data + i * sizeof(uint_t)));
> +

hmmmm, isn't this the line that actually writes in the no-ops? why isn't
it disabled?

> + free(data);
> +}
>
>

Thanks again for doing this...I was still understanding recordmcount.c ;)

-Jason


\
 
 \ /
  Last update: 2011-10-07 20:55    [W:0.108 / U:0.696 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site