Skip to contentSkip to conversation

fs/binfmt_elf: use ELF_ET_DYN_BASE for PIE (ET_DYN with INTERP) binaries

PATCH In review Azat Khuzhin 2 participants 7 messages Jan 31, 2022 – Feb 3, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fs/binfmt_elf: use ELF_ET_DYN_BASE for PIE (ET_DYN with INTERP) binaries

Since 9630f0d60fec ELF_ET_DYN_BASE is not used as a load_bias anymore and this breaks PIE binaries, since after this change data segment became too nearby the stack:

Before 9630f0d60fec:

    $ strace -febrk /tmp/test-stack |& head
    brk(NULL)                               = 0x555555559000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc5c0
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 7690 (stack diff: 8365664)
    Segmentation fault (core dumped)

After 9630f0d60fec:

    $ strace -ebrk /tmp/test-stack  |& head
    brk(NULL)                               = 0x7ffff7fff000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc640
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 146 (stack diff: 157792)
    Segmentation fault (core dumped)

Found this during compiling with clang, that started to randomly SIGSEGV when it eats some heap.

Reproducer:

    #include <stdio.h>
    #include <stddef.h>
    #include <stdlib.h>

    static int depth = 0;
    static void* bottom_of_stack;

    int inner()
    {
    	char buffer[1024];
    	ptrdiff_t diff;
    	int ret;

    	++depth;
    	diff = bottom_of_stack - __builtin_frame_address(0);

    	fprintf(stderr,
    		"recursion depth: %i (stack diff: %zu)\n",
    		depth, (size_t)diff);

    	for (size_t i = 0; i < 1024; ++i)
    		ret += buffer[i];

    	ret += inner();

    	return ret;
    }

    int main()
    {
    	for (size_t size = 0; size < 128<<20; size += 16<<10)
    		malloc(16<<10);

    	bottom_of_stack = __builtin_frame_address(0);
    	fprintf(stderr, "bottom_of_stack = %p\n", bottom_of_stack);
    	inner();
    }
Fixes 9630f0d60fec ("fs/binfmt_elf: use PT_LOAD p_align values for static PIE")
Cc Chris Kennelly <[email protected]>
Cc Al Viro <[email protected]>
Cc Alexey Dobriyan <[email protected]>
Cc Song Liu <[email protected]>
Cc David Rientjes <[email protected]>
Cc Ian Rogers <[email protected]>
Cc Hugh Dickins <[email protected]>
Cc Suren Baghdasaryan <[email protected]>
Cc Sandeep Patil <[email protected]>
Cc Fangrui Song <[email protected]>
Cc Nick Desaulniers <[email protected]>
Cc Kirill A. Shutemov <[email protected]>
Cc Mike Kravetz <[email protected]>
Cc Shuah Khan <[email protected]>
Cc H.J. Lu <[email protected]>
---
 fs/binfmt_elf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/binfmt_elf.c +1 -1
index 605017eb9349..9e11e6f13e83 100644
@@ -1117,7 +1117,7 @@ static int load_elf_binary(struct linux_binprm *bprm)
 			 * without MAP_FIXED nor MAP_FIXED_NOREPLACE).
 			 */
 			alignment = maximum_alignment(elf_phdata, elf_ex->e_phnum);
-			if (alignment > ELF_MIN_ALIGN) {
+			if (interpreter || alignment > ELF_MIN_ALIGN) {
 				load_bias = ELF_ET_DYN_BASE;
 				if (current->flags & PF_RANDOMIZE)
 					load_bias += arch_mmap_rnd();
Signature · 1 line
2.35.1
#1 lore raw
Quoted 28 lines · Azat Khuzhin

Since 9630f0d60fec ELF_ET_DYN_BASE is not used as a load_bias anymore and this breaks PIE binaries, since after this change data segment became too nearby the stack:

Before 9630f0d60fec:

    $ strace -febrk /tmp/test-stack |& head
    brk(NULL)                               = 0x555555559000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc5c0
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 7690 (stack diff: 8365664)
    Segmentation fault (core dumped)

After 9630f0d60fec:

    $ strace -ebrk /tmp/test-stack  |& head
    brk(NULL)                               = 0x7ffff7fff000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc640
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 146 (stack diff: 157792)
    Segmentation fault (core dumped)
Quoted 2 lines

Found this during compiling with clang, that started to randomly SIGSEGV when it eats some heap.

How do I reproduce it on x86-64?

Quoted 77 lines

Reproducer:

    #include <stdio.h>
    #include <stddef.h>
    #include <stdlib.h>

    static int depth = 0;
    static void* bottom_of_stack;

    int inner()
    {
        char buffer[1024];
        ptrdiff_t diff;
        int ret;

        ++depth;
        diff = bottom_of_stack - __builtin_frame_address(0);

        fprintf(stderr,
                "recursion depth: %i (stack diff: %zu)\n",
                depth, (size_t)diff);

        for (size_t i = 0; i < 1024; ++i)
                ret += buffer[i];

        ret += inner();

        return ret;
    }

    int main()
    {
        for (size_t size = 0; size < 128<<20; size += 16<<10)
                malloc(16<<10);

        bottom_of_stack = __builtin_frame_address(0);
        fprintf(stderr, "bottom_of_stack = %p\n", bottom_of_stack);
        inner();
    }
Fixes 9630f0d60fec ("fs/binfmt_elf: use PT_LOAD p_align values for static PIE")
Cc Chris Kennelly <[email protected]>
Cc Al Viro <[email protected]>
Cc Alexey Dobriyan <[email protected]>
Cc Song Liu <[email protected]>
Cc David Rientjes <[email protected]>
Cc Ian Rogers <[email protected]>
Cc Hugh Dickins <[email protected]>
Cc Suren Baghdasaryan <[email protected]>
Cc Sandeep Patil <[email protected]>
Cc Fangrui Song <[email protected]>
Cc Nick Desaulniers <[email protected]>
Cc Kirill A. Shutemov <[email protected]>
Cc Mike Kravetz <[email protected]>
Cc Shuah Khan <[email protected]>
Cc H.J. Lu <[email protected]>
---
 fs/binfmt_elf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/binfmt_elf.c +1 -2
index 605017eb9349..9e11e6f13e83 100644
@@ -1117,7 +1117,7 @@ static int load_elf_binary(struct linux_binprm *bprm)
                         * without MAP_FIXED nor MAP_FIXED_NOREPLACE).
                         */
                        alignment = maximum_alignment(elf_phdata, elf_ex->e_phnum);
-                       if (alignment > ELF_MIN_ALIGN) {
+                       if (interpreter || alignment > ELF_MIN_ALIGN) {
                                load_bias = ELF_ET_DYN_BASE;
                                if (current->flags & PF_RANDOMIZE)
                                        load_bias += arch_mmap_rnd();
--

2.35.1

Signature · 1 line
H.J.
#2 lore raw
Quoted 29 lines · H.J. Lu
Quoted 28 lines · Azat Khuzhin

Since 9630f0d60fec ELF_ET_DYN_BASE is not used as a load_bias anymore and this breaks PIE binaries, since after this change data segment became too nearby the stack:

Before 9630f0d60fec:

    $ strace -febrk /tmp/test-stack |& head
    brk(NULL)                               = 0x555555559000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc5c0
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 7690 (stack diff: 8365664)
    Segmentation fault (core dumped)

After 9630f0d60fec:

    $ strace -ebrk /tmp/test-stack  |& head
    brk(NULL)                               = 0x7ffff7fff000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc640
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 146 (stack diff: 157792)
    Segmentation fault (core dumped)
Quoted 4 lines
Quoted 2 lines

Found this during compiling with clang, that started to randomly SIGSEGV when it eats some heap.

How do I reproduce it on x86-64?

It fails for me for pretty big C++ unit, so I don't have a simple reproducer with clang, but the attached reproducer below should show the problem.

Quoted 39 lines
Quoted 39 lines

Reproducer:

    #include <stdio.h>
    #include <stddef.h>
    #include <stdlib.h>

    static int depth = 0;
    static void* bottom_of_stack;

    int inner()
    {
        char buffer[1024];
        ptrdiff_t diff;
        int ret;

        ++depth;
        diff = bottom_of_stack - __builtin_frame_address(0);

        fprintf(stderr,
                "recursion depth: %i (stack diff: %zu)\n",
                depth, (size_t)diff);

        for (size_t i = 0; i < 1024; ++i)
                ret += buffer[i];

        ret += inner();

        return ret;
    }

    int main()
    {
        for (size_t size = 0; size < 128<<20; size += 16<<10)
                malloc(16<<10);

        bottom_of_stack = __builtin_frame_address(0);
        fprintf(stderr, "bottom_of_stack = %p\n", bottom_of_stack);
        inner();
    }
#3 lore raw
Quoted 31 lines · Azat Khuzhin
Quoted 29 lines · H.J. Lu
Quoted 28 lines · Azat Khuzhin

Since 9630f0d60fec ELF_ET_DYN_BASE is not used as a load_bias anymore and this breaks PIE binaries, since after this change data segment became too nearby the stack:

Before 9630f0d60fec:

    $ strace -febrk /tmp/test-stack |& head
    brk(NULL)                               = 0x555555559000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc5c0
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 7690 (stack diff: 8365664)
    Segmentation fault (core dumped)

After 9630f0d60fec:

    $ strace -ebrk /tmp/test-stack  |& head
    brk(NULL)                               = 0x7ffff7fff000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc640
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 146 (stack diff: 157792)
    Segmentation fault (core dumped)
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

Found this during compiling with clang, that started to randomly SIGSEGV when it eats some heap.

How do I reproduce it on x86-64?

It fails for me for pretty big C++ unit, so I don't have a simple reproducer with clang, but the attached reproducer below should show the problem.

The reproducer doesn't fail for me under 5.17-rc2 on Fedora 35/x86-64 with 32GB RAM. Did you turn off PF_RANDOMIZE?

Quoted 39 lines
Quoted 39 lines
Quoted 39 lines

Reproducer:

    #include <stdio.h>
    #include <stddef.h>
    #include <stdlib.h>

    static int depth = 0;
    static void* bottom_of_stack;

    int inner()
    {
        char buffer[1024];
        ptrdiff_t diff;
        int ret;

        ++depth;
        diff = bottom_of_stack - __builtin_frame_address(0);

        fprintf(stderr,
                "recursion depth: %i (stack diff: %zu)\n",
                depth, (size_t)diff);

        for (size_t i = 0; i < 1024; ++i)
                ret += buffer[i];

        ret += inner();

        return ret;
    }

    int main()
    {
        for (size_t size = 0; size < 128<<20; size += 16<<10)
                malloc(16<<10);

        bottom_of_stack = __builtin_frame_address(0);
        fprintf(stderr, "bottom_of_stack = %p\n", bottom_of_stack);
        inner();
    }
Signature · 1 line
H.J.
#4 lore raw
Quoted 35 lines · H.J. Lu
Quoted 34 lines · Azat Khuzhin
Quoted 32 lines · H.J. Lu
Quoted 30 lines · Azat Khuzhin

Since 9630f0d60fec ELF_ET_DYN_BASE is not used as a load_bias anymore and this breaks PIE binaries, since after this change data segment became too nearby the stack:

Before 9630f0d60fec:

    $ strace -febrk /tmp/test-stack |& head
    brk(NULL)                               = 0x555555559000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc5c0
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 7690 (stack diff: 8365664)
    Segmentation fault (core dumped)

After 9630f0d60fec:

    $ strace -ebrk /tmp/test-stack  |& head
    brk(NULL)                               = 0x7ffff7fff000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc640
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 146 (stack diff: 157792)
    Segmentation fault (core dumped)

Found this during compiling with clang, that started to randomly SIGSEGV when it eats some heap.

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

How do I reproduce it on x86-64?

It fails for me for pretty big C++ unit, so I don't have a simple reproducer with clang, but the attached reproducer below should show the problem.

The reproducer doesn't fail for me under 5.17-rc2 on Fedora 35/x86-64 with 32GB RAM. Did you turn off PF_RANDOMIZE?

Oh, yep, forgot to mention that I have kernel.randomize_va_space=0.

#5 lore raw
Quoted 39 lines · Azat Khuzhin
Quoted 37 lines · H.J. Lu
Quoted 36 lines · Azat Khuzhin
Quoted 33 lines · H.J. Lu
Quoted 28 lines · Azat Khuzhin

Since 9630f0d60fec ELF_ET_DYN_BASE is not used as a load_bias anymore and this breaks PIE binaries, since after this change data segment became too nearby the stack:

Before 9630f0d60fec:

    $ strace -febrk /tmp/test-stack |& head
    brk(NULL)                               = 0x555555559000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc5c0
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 7690 (stack diff: 8365664)
    Segmentation fault (core dumped)

After 9630f0d60fec:

    $ strace -ebrk /tmp/test-stack  |& head
    brk(NULL)                               = 0x7ffff7fff000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc640
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 146 (stack diff: 157792)
    Segmentation fault (core dumped)
Quoted 2 lines

Found this during compiling with clang, that started to randomly SIGSEGV when it eats some heap.

How do I reproduce it on x86-64?

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

It fails for me for pretty big C++ unit, so I don't have a simple reproducer with clang, but the attached reproducer below should show the problem.

The reproducer doesn't fail for me under 5.17-rc2 on Fedora 35/x86-64 with 32GB RAM. Did you turn off PF_RANDOMIZE?

Oh, yep, forgot to mention that I have kernel.randomize_va_space=0.

PIE with interpreter and PIE with alignment > ELF_MIN_ALIGN should always be loaded from ELF_ET_DYN_BASE. Otherwise, either PIE is loaded at an address which is too low or isn't properly aligned.

Signature · 1 line
H.J.
#6 lore raw
Quoted 47 lines · H.J. Lu
Quoted 46 lines · Azat Khuzhin
Quoted 43 lines · H.J. Lu
Quoted 31 lines · Azat Khuzhin
Quoted 29 lines · H.J. Lu
Quoted 28 lines · Azat Khuzhin

Since 9630f0d60fec ELF_ET_DYN_BASE is not used as a load_bias anymore and this breaks PIE binaries, since after this change data segment became too nearby the stack:

Before 9630f0d60fec:

    $ strace -febrk /tmp/test-stack |& head
    brk(NULL)                               = 0x555555559000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc5c0
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 7690 (stack diff: 8365664)
    Segmentation fault (core dumped)

After 9630f0d60fec:

    $ strace -ebrk /tmp/test-stack  |& head
    brk(NULL)                               = 0x7ffff7fff000
    $ /tmp/test-stack
    bottom_of_stack = 0x7fffffffc640
    recursion depth: 1 (stack diff: 32)
    ...
    recursion depth: 146 (stack diff: 157792)
    Segmentation fault (core dumped)
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

Found this during compiling with clang, that started to randomly SIGSEGV when it eats some heap.

How do I reproduce it on x86-64?

It fails for me for pretty big C++ unit, so I don't have a simple reproducer with clang, but the attached reproducer below should show the problem.

The reproducer doesn't fail for me under 5.17-rc2 on Fedora 35/x86-64 with 32GB RAM. Did you turn off PF_RANDOMIZE?

Quoted 6 lines
Quoted 1 line

Oh, yep, forgot to mention that I have kernel.randomize_va_space=0.

PIE with interpreter and PIE with alignment > ELF_MIN_ALIGN should always be loaded from ELF_ET_DYN_BASE. Otherwise, either PIE is loaded at an address which is too low or isn't properly aligned.

So, this is what this patch does, right?
Any news on this patch?