Skip to contentSkip to conversation

elf: Relax assumptions about vaddr ordering

PATCH In review Magnus Groß 4 participants 13 messages Jan 26, 2022 – Feb 2, 2022

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 elf: Relax assumptions about vaddr ordering

From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <[email protected]>
Date: Wed, 26 Jan 2022 16:35:07 +0100
Subject: [PATCH] elf: Relax assumptions about vaddr ordering
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().

Specifically consider an ELF binary with the following PT_LOAD segments:

Type  Offset   VirtAddr   PhysAddr   FileSiz  MemSiz    Flg Align
LOAD  0x000000 0x08000000 0x08000000 0x474585 0x474585  R E 0x1000
LOAD  0x475000 0x08475000 0x08475000 0x090a4  0xc6c10   RW  0x1000
LOAD  0x47f000 0x00010000 0x00010000 0x00000  0x7ff0000     0x1000

Note how the last segment is actually the first segment and vice versa.

Since total_mapping_size() only computes the difference between the first and the last segment in the order that they appear, it will return a size of 0 in this case, thus causing load_elf_binary() to fail, which did not happen before that change.

Strictly speaking total_mapping_size() made that assumption already before that patch, but the issue did not appear because the old load_addr_set guards never allowed this call to total_mapping_size().

Instead of fixing this by reverting to the old load_addr_set logic, we fix this by comparing the correct first and last segments in total_mapping_size().

---
 fs/binfmt_elf.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
fs/binfmt_elf.c +14 -4
index f8c7f26f1fbb..0caaad9eddd1 100644
@@ -402,19 +402,29 @@ static unsigned long elf_map(struct file *filep, unsigned long addr,
 static unsigned long total_mapping_size(const struct elf_phdr *cmds, int nr)
 {
 	int i, first_idx = -1, last_idx = -1;
+	unsigned long min_vaddr = ULONG_MAX, max_vaddr = 0;
 
 	for (i = 0; i < nr; i++) {
 		if (cmds[i].p_type == PT_LOAD) {
-			last_idx = i;
-			if (first_idx == -1)
+			/*
+			 * The PT_LOAD segments are not necessarily ordered
+			 * by vaddr. Make sure that we get the segment with
+			 * minimum vaddr (maximum vaddr respectively)
+			 */
+			if (cmds[i].p_vaddr <= min_vaddr) {
 				first_idx = i;
+				min_vaddr = cmds[i].p_vaddr;
+			}
+			if (cmds[i].p_vaddr >= max_vaddr) {
+				last_idx = i;
+				max_vaddr = cmds[i].p_vaddr;
+			}
 		}
 	}
 	if (first_idx == -1)
 		return 0;
 
-	return cmds[last_idx].p_vaddr + cmds[last_idx].p_memsz -
-				ELF_PAGESTART(cmds[first_idx].p_vaddr);
+	return max_vaddr + cmds[last_idx].p_memsz - ELF_PAGESTART(min_vaddr);
 }
 
 static int elf_read(struct file *file, void *buf, size_t len, loff_t pos)
Signature · 1 line
2.34.1
Acked-by on v1 1/1
#1 lore raw
Quoted 26 lines · Magnus Groß

From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <[email protected]>
Date: Wed, 26 Jan 2022 16:35:07 +0100
Subject: [PATCH] elf: Relax assumptions about vaddr ordering
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().

Specifically consider an ELF binary with the following PT_LOAD segments:

Type  Offset   VirtAddr   PhysAddr   FileSiz  MemSiz    Flg Align
LOAD  0x000000 0x08000000 0x08000000 0x474585 0x474585  R E 0x1000
LOAD  0x475000 0x08475000 0x08475000 0x090a4  0xc6c10   RW  0x1000
LOAD  0x47f000 0x00010000 0x00010000 0x00000  0x7ff0000     0x1000

Note how the last segment is actually the first segment and vice versa.

Since total_mapping_size() only computes the difference between the first and the last segment in the order that they appear, it will return a size of 0 in this case, thus causing load_elf_binary() to fail, which did not happen before that change.

Quoted 7 lines

Strictly speaking total_mapping_size() made that assumption already before that patch, but the issue did not appear because the old load_addr_set guards never allowed this call to total_mapping_size().

Instead of fixing this by reverting to the old load_addr_set logic, we fix this by comparing the correct first and last segments in total_mapping_size().

Ah, nice. Yeah, this is good.

Quoted 1 line
Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

-Kees

Quoted 44 lines
---
 fs/binfmt_elf.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
fs/binfmt_elf.c +14 -4
index f8c7f26f1fbb..0caaad9eddd1 100644
@@ -402,19 +402,29 @@ static unsigned long elf_map(struct file *filep, unsigned long addr,
 static unsigned long total_mapping_size(const struct elf_phdr *cmds, int nr)
 {
 	int i, first_idx = -1, last_idx = -1;
+	unsigned long min_vaddr = ULONG_MAX, max_vaddr = 0;
 
 	for (i = 0; i < nr; i++) {
 		if (cmds[i].p_type == PT_LOAD) {
-			last_idx = i;
-			if (first_idx == -1)
+			/*
+			 * The PT_LOAD segments are not necessarily ordered
+			 * by vaddr. Make sure that we get the segment with
+			 * minimum vaddr (maximum vaddr respectively)
+			 */
+			if (cmds[i].p_vaddr <= min_vaddr) {
 				first_idx = i;
+				min_vaddr = cmds[i].p_vaddr;
+			}
+			if (cmds[i].p_vaddr >= max_vaddr) {
+				last_idx = i;
+				max_vaddr = cmds[i].p_vaddr;
+			}
 		}
 	}
 	if (first_idx == -1)
 		return 0;
 
-	return cmds[last_idx].p_vaddr + cmds[last_idx].p_memsz -
-				ELF_PAGESTART(cmds[first_idx].p_vaddr);
+	return max_vaddr + cmds[last_idx].p_memsz - ELF_PAGESTART(min_vaddr);
 }
 
 static int elf_read(struct file *file, void *buf, size_t len, loff_t pos)
Signature · 1 line
2.34.1
Signature · 1 line
Kees Cook
#2 lore raw
Quoted 35 lines · Kees Cook
Quoted 26 lines · Magnus Groß

From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <[email protected]>
Date: Wed, 26 Jan 2022 16:35:07 +0100
Subject: [PATCH] elf: Relax assumptions about vaddr ordering
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().

Specifically consider an ELF binary with the following PT_LOAD segments:

Type  Offset   VirtAddr   PhysAddr   FileSiz  MemSiz    Flg Align
LOAD  0x000000 0x08000000 0x08000000 0x474585 0x474585  R E 0x1000
LOAD  0x475000 0x08475000 0x08475000 0x090a4  0xc6c10   RW  0x1000
LOAD  0x47f000 0x00010000 0x00010000 0x00000  0x7ff0000     0x1000

Note how the last segment is actually the first segment and vice versa.

Since total_mapping_size() only computes the difference between the first and the last segment in the order that they appear, it will return a size of 0 in this case, thus causing load_elf_binary() to fail, which did not happen before that change.

Quoted 7 lines

Strictly speaking total_mapping_size() made that assumption already before that patch, but the issue did not appear because the old load_addr_set guards never allowed this call to total_mapping_size().

Instead of fixing this by reverting to the old load_addr_set logic, we fix this by comparing the correct first and last segments in total_mapping_size().

Quoted 7 lines

Ah, nice. Yeah, this is good.

Quoted 1 line
Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

Andrew, can you pick this up too?

-Kees

Quoted 50 lines

-Kees

Quoted 44 lines
---
 fs/binfmt_elf.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
fs/binfmt_elf.c +14 -4
index f8c7f26f1fbb..0caaad9eddd1 100644
@@ -402,19 +402,29 @@ static unsigned long elf_map(struct file *filep, unsigned long addr,
 static unsigned long total_mapping_size(const struct elf_phdr *cmds, int nr)
 {
 	int i, first_idx = -1, last_idx = -1;
+	unsigned long min_vaddr = ULONG_MAX, max_vaddr = 0;
 
 	for (i = 0; i < nr; i++) {
 		if (cmds[i].p_type == PT_LOAD) {
-			last_idx = i;
-			if (first_idx == -1)
+			/*
+			 * The PT_LOAD segments are not necessarily ordered
+			 * by vaddr. Make sure that we get the segment with
+			 * minimum vaddr (maximum vaddr respectively)
+			 */
+			if (cmds[i].p_vaddr <= min_vaddr) {
 				first_idx = i;
+				min_vaddr = cmds[i].p_vaddr;
+			}
+			if (cmds[i].p_vaddr >= max_vaddr) {
+				last_idx = i;
+				max_vaddr = cmds[i].p_vaddr;
+			}
 		}
 	}
 	if (first_idx == -1)
 		return 0;
 
-	return cmds[last_idx].p_vaddr + cmds[last_idx].p_memsz -
-				ELF_PAGESTART(cmds[first_idx].p_vaddr);
+	return max_vaddr + cmds[last_idx].p_memsz - ELF_PAGESTART(min_vaddr);
 }
 
 static int elf_read(struct file *file, void *buf, size_t len, loff_t pos)
Signature · 1 line
2.34.1
Signature · 1 line
Kees Cook
Signature · 1 line
Kees Cook
#3 lore raw
Quoted 40 lines · On Wed, Jan 26, 2022 at 10:31:42PM -0800 Kees Cook wrote:
Quoted 39 lines · Kees Cook
Quoted 26 lines · Magnus Groß

From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <[email protected]>
Date: Wed, 26 Jan 2022 16:35:07 +0100
Subject: [PATCH] elf: Relax assumptions about vaddr ordering
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().

Specifically consider an ELF binary with the following PT_LOAD segments:

Type  Offset   VirtAddr   PhysAddr   FileSiz  MemSiz    Flg Align
LOAD  0x000000 0x08000000 0x08000000 0x474585 0x474585  R E 0x1000
LOAD  0x475000 0x08475000 0x08475000 0x090a4  0xc6c10   RW  0x1000
LOAD  0x47f000 0x00010000 0x00010000 0x00000  0x7ff0000     0x1000

Note how the last segment is actually the first segment and vice versa.

Since total_mapping_size() only computes the difference between the first and the last segment in the order that they appear, it will return a size of 0 in this case, thus causing load_elf_binary() to fail, which did not happen before that change.

Quoted 7 lines

Strictly speaking total_mapping_size() made that assumption already before that patch, but the issue did not appear because the old load_addr_set guards never allowed this call to total_mapping_size().

Instead of fixing this by reverting to the old load_addr_set logic, we fix this by comparing the correct first and last segments in total_mapping_size().

Ah, nice. Yeah, this is good.

Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

Andrew, can you pick this up too?

-Kees

May I also propose to include this patch in whatever mailing-list corresponds to the 5.16.x bugfix series?
It turns out that almost all native Linux games published by the Virtual Programming company have this kind of weird PT_LOAD ordering including the famous Bioshock Infinite, so right now those games are all completely broken since Linux 5.16.

P.S.: Someone should probably ask Virtual Programming, what kind of tooling they use to create such convoluted ELF binaries.

Quoted 53 lines
Quoted 50 lines

-Kees

Quoted 44 lines
---
 fs/binfmt_elf.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
fs/binfmt_elf.c +14 -4
index f8c7f26f1fbb..0caaad9eddd1 100644
@@ -402,19 +402,29 @@ static unsigned long elf_map(struct file *filep, unsigned long addr,
 static unsigned long total_mapping_size(const struct elf_phdr *cmds, int nr)
 {
 	int i, first_idx = -1, last_idx = -1;
+	unsigned long min_vaddr = ULONG_MAX, max_vaddr = 0;
 
 	for (i = 0; i < nr; i++) {
 		if (cmds[i].p_type == PT_LOAD) {
-			last_idx = i;
-			if (first_idx == -1)
+			/*
+			 * The PT_LOAD segments are not necessarily ordered
+			 * by vaddr. Make sure that we get the segment with
+			 * minimum vaddr (maximum vaddr respectively)
+			 */
+			if (cmds[i].p_vaddr <= min_vaddr) {
 				first_idx = i;
+				min_vaddr = cmds[i].p_vaddr;
+			}
+			if (cmds[i].p_vaddr >= max_vaddr) {
+				last_idx = i;
+				max_vaddr = cmds[i].p_vaddr;
+			}
 		}
 	}
 	if (first_idx == -1)
 		return 0;
 
-	return cmds[last_idx].p_vaddr + cmds[last_idx].p_memsz -
-				ELF_PAGESTART(cmds[first_idx].p_vaddr);
+	return max_vaddr + cmds[last_idx].p_memsz - ELF_PAGESTART(min_vaddr);
 }
 
 static int elf_read(struct file *file, void *buf, size_t len, loff_t pos)
Signature · 1 line
2.34.1
Signature · 1 line
Kees Cook
Signature · 1 line
Kees Cook
#4 lore raw
Quoted 50 lines · Magnus Groß
Quoted 40 lines · On Wed, Jan 26, 2022 at 10:31:42PM -0800 Kees Cook wrote:
Quoted 39 lines · Kees Cook
Quoted 26 lines · Magnus Groß

From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <[email protected]>
Date: Wed, 26 Jan 2022 16:35:07 +0100
Subject: [PATCH] elf: Relax assumptions about vaddr ordering
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().

Specifically consider an ELF binary with the following PT_LOAD segments:

Type  Offset   VirtAddr   PhysAddr   FileSiz  MemSiz    Flg Align
LOAD  0x000000 0x08000000 0x08000000 0x474585 0x474585  R E 0x1000
LOAD  0x475000 0x08475000 0x08475000 0x090a4  0xc6c10   RW  0x1000
LOAD  0x47f000 0x00010000 0x00010000 0x00000  0x7ff0000     0x1000

Note how the last segment is actually the first segment and vice versa.

Since total_mapping_size() only computes the difference between the first and the last segment in the order that they appear, it will return a size of 0 in this case, thus causing load_elf_binary() to fail, which did not happen before that change.

Quoted 7 lines

Strictly speaking total_mapping_size() made that assumption already before that patch, but the issue did not appear because the old load_addr_set guards never allowed this call to total_mapping_size().

Instead of fixing this by reverting to the old load_addr_set logic, we fix this by comparing the correct first and last segments in total_mapping_size().

Ah, nice. Yeah, this is good.

Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

Andrew, can you pick this up too?

-Kees

Quoted 6 lines

May I also propose to include this patch in whatever mailing-list corresponds to the 5.16.x bugfix series?
It turns out that almost all native Linux games published by the Virtual Programming company have this kind of weird PT_LOAD ordering including the famous Bioshock Infinite, so right now those games are all completely broken since Linux 5.16.

Thanks for additional rationale!

I included the field that would be expected to have this picked up for the 5.16 stable tree (the "Fixes: ..." and "Cc: [email protected]") so once it lands in Linus's tree, it'll get picked up for the v5.16.z series too.

Quoted 2 lines

P.S.: Someone should probably ask Virtual Programming, what kind of tooling they use to create such convoluted ELF binaries.

Does "strings" provide any hints? :)

Signature · 1 line
Kees Cook
Acked-by
#5 lore raw
In reply to Magnus Groß
Quoted 57 lines · Magnus Groß
Quoted 40 lines · On Wed, Jan 26, 2022 at 10:31:42PM -0800 Kees Cook wrote:
Quoted 39 lines · Kees Cook
Quoted 26 lines · Magnus Groß

From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <[email protected]>
Date: Wed, 26 Jan 2022 16:35:07 +0100
Subject: [PATCH] elf: Relax assumptions about vaddr ordering
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().

Specifically consider an ELF binary with the following PT_LOAD segments:

Type  Offset   VirtAddr   PhysAddr   FileSiz  MemSiz    Flg Align
LOAD  0x000000 0x08000000 0x08000000 0x474585 0x474585  R E 0x1000
LOAD  0x475000 0x08475000 0x08475000 0x090a4  0xc6c10   RW  0x1000
LOAD  0x47f000 0x00010000 0x00010000 0x00000  0x7ff0000     0x1000

Note how the last segment is actually the first segment and vice versa.

Since total_mapping_size() only computes the difference between the first and the last segment in the order that they appear, it will return a size of 0 in this case, thus causing load_elf_binary() to fail, which did not happen before that change.

Quoted 7 lines

Strictly speaking total_mapping_size() made that assumption already before that patch, but the issue did not appear because the old load_addr_set guards never allowed this call to total_mapping_size().

Instead of fixing this by reverting to the old load_addr_set logic, we fix this by comparing the correct first and last segments in total_mapping_size().

Ah, nice. Yeah, this is good.

Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

Andrew, can you pick this up too?

-Kees

May I also propose to include this patch in whatever mailing-list corresponds to the 5.16.x bugfix series?
It turns out that almost all native Linux games published by the Virtual Programming company have this kind of weird PT_LOAD ordering including the famous Bioshock Infinite, so right now those games are all completely broken since Linux 5.16.

Quoted 2 lines

P.S.: Someone should probably ask Virtual Programming, what kind of tooling they use to create such convoluted ELF binaries.

Oh, actually, this was independently fixed:
https://lore.kernel.org/all/YVmd7D0M6G/[email protected]/

Alexey, you never answered by question about why we can't use a proper type and leave the ELF_PAGESTART() macros alone:
https://lore.kernel.org/all/202110071038.B589687@keescook/

I still don't like the use of "int" in ELF_PAGESTART(), but I agree it shouldn't cause a problem. I just really don't like mixing a signed type with address calculations, from a robustness perspective.

Andrew, can you update elf-fix-overflow-in-total-mapping-size-calculation.patch to include:

Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

Thanks!

-Kees

Quoted 54 lines
Quoted 54 lines
Quoted 51 lines

-Kees

Quoted 44 lines
---
 fs/binfmt_elf.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
fs/binfmt_elf.c +14 -4
index f8c7f26f1fbb..0caaad9eddd1 100644
@@ -402,19 +402,29 @@ static unsigned long elf_map(struct file *filep, unsigned long addr,
 static unsigned long total_mapping_size(const struct elf_phdr *cmds, int nr)
 {
 	int i, first_idx = -1, last_idx = -1;
+	unsigned long min_vaddr = ULONG_MAX, max_vaddr = 0;
 
 	for (i = 0; i < nr; i++) {
 		if (cmds[i].p_type == PT_LOAD) {
-			last_idx = i;
-			if (first_idx == -1)
+			/*
+			 * The PT_LOAD segments are not necessarily ordered
+			 * by vaddr. Make sure that we get the segment with
+			 * minimum vaddr (maximum vaddr respectively)
+			 */
+			if (cmds[i].p_vaddr <= min_vaddr) {
 				first_idx = i;
+				min_vaddr = cmds[i].p_vaddr;
+			}
+			if (cmds[i].p_vaddr >= max_vaddr) {
+				last_idx = i;
+				max_vaddr = cmds[i].p_vaddr;
+			}
 		}
 	}
 	if (first_idx == -1)
 		return 0;
 
-	return cmds[last_idx].p_vaddr + cmds[last_idx].p_memsz -
-				ELF_PAGESTART(cmds[first_idx].p_vaddr);
+	return max_vaddr + cmds[last_idx].p_memsz - ELF_PAGESTART(min_vaddr);
 }
 
 static int elf_read(struct file *file, void *buf, size_t len, loff_t pos)
Signature · 1 line
2.34.1
Signature · 1 line
Kees Cook
Signature · 1 line
Kees Cook
Signature · 1 line
Kees Cook
#6 lore raw
Quoted 58 lines · Kees Cook
Quoted 57 lines · Magnus Groß
Quoted 40 lines · On Wed, Jan 26, 2022 at 10:31:42PM -0800 Kees Cook wrote:
Quoted 39 lines · Kees Cook
Quoted 26 lines · Magnus Groß

From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <[email protected]>
Date: Wed, 26 Jan 2022 16:35:07 +0100
Subject: [PATCH] elf: Relax assumptions about vaddr ordering
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().

Specifically consider an ELF binary with the following PT_LOAD segments:

Type  Offset   VirtAddr   PhysAddr   FileSiz  MemSiz    Flg Align
LOAD  0x000000 0x08000000 0x08000000 0x474585 0x474585  R E 0x1000
LOAD  0x475000 0x08475000 0x08475000 0x090a4  0xc6c10   RW  0x1000
LOAD  0x47f000 0x00010000 0x00010000 0x00000  0x7ff0000     0x1000

Note how the last segment is actually the first segment and vice versa.

Since total_mapping_size() only computes the difference between the first and the last segment in the order that they appear, it will return a size of 0 in this case, thus causing load_elf_binary() to fail, which did not happen before that change.

Quoted 7 lines

Strictly speaking total_mapping_size() made that assumption already before that patch, but the issue did not appear because the old load_addr_set guards never allowed this call to total_mapping_size().

Instead of fixing this by reverting to the old load_addr_set logic, we fix this by comparing the correct first and last segments in total_mapping_size().

Ah, nice. Yeah, this is good.

Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

Andrew, can you pick this up too?

-Kees

May I also propose to include this patch in whatever mailing-list corresponds to the 5.16.x bugfix series?
It turns out that almost all native Linux games published by the Virtual Programming company have this kind of weird PT_LOAD ordering including the famous Bioshock Infinite, so right now those games are all completely broken since Linux 5.16.

Quoted 5 lines
Quoted 2 lines

P.S.: Someone should probably ask Virtual Programming, what kind of tooling they use to create such convoluted ELF binaries.

Oh, actually, this was independently fixed:
https://lore.kernel.org/all/YVmd7D0M6G/[email protected]/

Oh wow, I accidently fixed real bug.

Quoted 7 lines

Alexey, you never answered by question about why we can't use a proper type and leave the ELF_PAGESTART() macros alone:
https://lore.kernel.org/all/202110071038.B589687@keescook/

I still don't like the use of "int" in ELF_PAGESTART(), but I agree it shouldn't cause a problem. I just really don't like mixing a signed type with address calculations, from a robustness perspective.

It is very robust. There are 2 ways to mask pointers

unsigned long & ~(unsigned long)-1

or

unsigned long & ~(int)-1

Both work. Second variant works for uint32_t too.

As I wrote in that thread, this macro

	#define ELF_PAGESTART(_v) ((_v) & ~(unsigned long)(ELF_MIN_ALIGN-1))

is slightly incorrect because type of the expression can be (unsigned long) but it logically should be typeof(v). Now fixing by switching to ALIGN doesn't do anything because ALIGN has the same problem.
And fixing ALIGN requires to go through thousands of usages, which is way too much for one localised ELF fix.

PT_ALEXEY

#7 lore raw
In reply to Kees Cook
Quoted 61 lines · On Fri, Jan 28, 2022 at 02:30:12PM -0800 Kees Cook wrote:
Quoted 57 lines · Magnus Groß
Quoted 40 lines · On Wed, Jan 26, 2022 at 10:31:42PM -0800 Kees Cook wrote:
Quoted 39 lines · Kees Cook
Quoted 26 lines · Magnus Groß

From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <[email protected]>
Date: Wed, 26 Jan 2022 16:35:07 +0100
Subject: [PATCH] elf: Relax assumptions about vaddr ordering
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().

Specifically consider an ELF binary with the following PT_LOAD segments:

Type  Offset   VirtAddr   PhysAddr   FileSiz  MemSiz    Flg Align
LOAD  0x000000 0x08000000 0x08000000 0x474585 0x474585  R E 0x1000
LOAD  0x475000 0x08475000 0x08475000 0x090a4  0xc6c10   RW  0x1000
LOAD  0x47f000 0x00010000 0x00010000 0x00000  0x7ff0000     0x1000

Note how the last segment is actually the first segment and vice versa.

Since total_mapping_size() only computes the difference between the first and the last segment in the order that they appear, it will return a size of 0 in this case, thus causing load_elf_binary() to fail, which did not happen before that change.

Quoted 7 lines

Strictly speaking total_mapping_size() made that assumption already before that patch, but the issue did not appear because the old load_addr_set guards never allowed this call to total_mapping_size().

Instead of fixing this by reverting to the old load_addr_set logic, we fix this by comparing the correct first and last segments in total_mapping_size().

Ah, nice. Yeah, this is good.

Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

Andrew, can you pick this up too?

-Kees

May I also propose to include this patch in whatever mailing-list corresponds to the 5.16.x bugfix series?
It turns out that almost all native Linux games published by the Virtual Programming company have this kind of weird PT_LOAD ordering including the famous Bioshock Infinite, so right now those games are all completely broken since Linux 5.16.

Quoted 2 lines

P.S.: Someone should probably ask Virtual Programming, what kind of tooling they use to create such convoluted ELF binaries.

Quoted 6 lines

Oh, actually, this was independently fixed:
https://lore.kernel.org/all/YVmd7D0M6G/[email protected]/

Alexey, you never answered by question about why we can't use a proper type and leave the ELF_PAGESTART() macros alone:
https://lore.kernel.org/all/202110071038.B589687@keescook/

Oh sorry, I didn't see that there was already a patch floating around that fixed the issue, otherwise I would have not wasted so much time on debugging this.
Oh well, doesn't matter now, I still learned a lot about kernel development and debugging with kgdb, I will probably be able to make some use of that knowledge in the future for another kernel patch.

Quoted 4 lines
Quoted 2 lines

P.S.: Someone should probably ask Virtual Programming, what kind of tooling they use to create such convoluted ELF binaries.

Does "strings" provide any hints? :)

It seems to be crosstool-ng
(https://github.com/crosstool-ng/crosstool-ng):

readelf -p .comment bioshock.i386

String dump of section '.comment':

[  0]  GCC: (crosstool-NG 1.17.0) 4.6.3
[ 21]  GCC: (Ubuntu 4.9.1-16ubuntu6) 4.9.1
[ 45]  GCC: (Ubuntu 4.8.2-19ubuntu1) 4.8.2

Not sure though if crosstool-ng outputs these weird ELF binaries in general or if it's just a bug.

--
Magnus

Quoted 57 lines
Quoted 54 lines
Quoted 54 lines
Quoted 51 lines

-Kees

Quoted 44 lines
---
 fs/binfmt_elf.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
fs/binfmt_elf.c +14 -4
index f8c7f26f1fbb..0caaad9eddd1 100644
@@ -402,19 +402,29 @@ static unsigned long elf_map(struct file *filep, unsigned long addr,
 static unsigned long total_mapping_size(const struct elf_phdr *cmds, int nr)
 {
 	int i, first_idx = -1, last_idx = -1;
+	unsigned long min_vaddr = ULONG_MAX, max_vaddr = 0;
 
 	for (i = 0; i < nr; i++) {
 		if (cmds[i].p_type == PT_LOAD) {
-			last_idx = i;
-			if (first_idx == -1)
+			/*
+			 * The PT_LOAD segments are not necessarily ordered
+			 * by vaddr. Make sure that we get the segment with
+			 * minimum vaddr (maximum vaddr respectively)
+			 */
+			if (cmds[i].p_vaddr <= min_vaddr) {
 				first_idx = i;
+				min_vaddr = cmds[i].p_vaddr;
+			}
+			if (cmds[i].p_vaddr >= max_vaddr) {
+				last_idx = i;
+				max_vaddr = cmds[i].p_vaddr;
+			}
 		}
 	}
 	if (first_idx == -1)
 		return 0;
 
-	return cmds[last_idx].p_vaddr + cmds[last_idx].p_memsz -
-				ELF_PAGESTART(cmds[first_idx].p_vaddr);
+	return max_vaddr + cmds[last_idx].p_memsz - ELF_PAGESTART(min_vaddr);
 }
 
 static int elf_read(struct file *file, void *buf, size_t len, loff_t pos)
Signature · 1 line
2.34.1
Signature · 1 line
Kees Cook
Signature · 1 line
Kees Cook
Signature · 1 line
Kees Cook
maintainer
#8 lore raw
In reply to Kees Cook
Quoted 6 lines · On Fri, 28 Jan 2022 14:30:12 -0800 Kees Cook <[email protected]> wrote:

Andrew, can you update elf-fix-overflow-in-total-mapping-size-calculation.patch to include:

Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

Done.

I'm taking it that we can omit this patch ("elf: Relax assumptions about vaddr ordering") and that Alexey's "ELF: fix overflow in total mapping size calculation" will suffice?

#9 lore raw
Quoted 9 lines · Andrew Morton
Quoted 6 lines · On Fri, 28 Jan 2022 14:30:12 -0800 Kees Cook <[email protected]> wrote:

Andrew, can you update elf-fix-overflow-in-total-mapping-size-calculation.patch to include:

Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")
Quoted 5 lines

Done.

I'm taking it that we can omit this patch ("elf: Relax assumptions about vaddr ordering") and that Alexey's "ELF: fix overflow in total mapping size calculation" will suffice?

Yes, it has the same result. Thanks!

Signature · 1 line
Kees Cook
#10 lore raw
In reply to Andrew Morton
Quoted 9 lines · Andrew Morton
Quoted 6 lines · On Fri, 28 Jan 2022 14:30:12 -0800 Kees Cook <[email protected]> wrote:

Andrew, can you update elf-fix-overflow-in-total-mapping-size-calculation.patch to include:

Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")
Quoted 5 lines

Done.

I'm taking it that we can omit this patch ("elf: Relax assumptions about vaddr ordering") and that Alexey's "ELF: fix overflow in total mapping size calculation" will suffice?

Yes, it is same patch conceptually.
It should work, but those who can't play Bioshock are better test it.

#11 lore raw

Am Wed, Feb 02, 2022 at 06:15:58PM +0300 schrieb Alexey Dobriyan:

Quoted 10 lines · Andrew Morton
Quoted 9 lines · On Fri, 28 Jan 2022 14:30:12 -0800 Kees Cook <[email protected]> wrote:
Quoted 7 lines

Andrew, can you update elf-fix-overflow-in-total-mapping-size-calculation.patch to include:

Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")
Quoted 8 lines
Quoted 5 lines

Done.

I'm taking it that we can omit this patch ("elf: Relax assumptions about vaddr ordering") and that Alexey's "ELF: fix overflow in total mapping size calculation" will suffice?

Yes, it is same patch conceptually.
It should work, but those who can't play Bioshock are better test it.

Yes it works.

Although the change from unsigned int to int is not necessary in the first place, as you can avoid the -1 initialization for min_addr by simply using ULONG_MAX, as can be seen in my patch.

maintainer
#12 lore raw
In reply to Kees Cook
Quoted 38 lines · On Wed, 26 Jan 2022 22:31:42 -0800 Kees Cook <[email protected]> wrote:
Quoted 37 lines · Kees Cook
Quoted 26 lines · Magnus Groß

From ff4dde97e82727727bda711f2367c05663498b24 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Magnus=20Gro=C3=9F?= <[email protected]>
Date: Wed, 26 Jan 2022 16:35:07 +0100
Subject: [PATCH] elf: Relax assumptions about vaddr ordering
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE") introduced a regression, where the kernel now assumes that PT_LOAD segments are ordered by vaddr in load_elf_binary().

Specifically consider an ELF binary with the following PT_LOAD segments:

Type  Offset   VirtAddr   PhysAddr   FileSiz  MemSiz    Flg Align
LOAD  0x000000 0x08000000 0x08000000 0x474585 0x474585  R E 0x1000
LOAD  0x475000 0x08475000 0x08475000 0x090a4  0xc6c10   RW  0x1000
LOAD  0x47f000 0x00010000 0x00010000 0x00000  0x7ff0000     0x1000

Note how the last segment is actually the first segment and vice versa.

Since total_mapping_size() only computes the difference between the first and the last segment in the order that they appear, it will return a size of 0 in this case, thus causing load_elf_binary() to fail, which did not happen before that change.

Quoted 7 lines

Strictly speaking total_mapping_size() made that assumption already before that patch, but the issue did not appear because the old load_addr_set guards never allowed this call to total_mapping_size().

Instead of fixing this by reverting to the old load_addr_set logic, we fix this by comparing the correct first and last segments in total_mapping_size().

Ah, nice. Yeah, this is good.

Quoted 8 lines
Quoted 5 lines
Quoted 1 line
Fixes 5f501d555653 ("binfmt_elf: reintroduce using MAP_FIXED_NOREPLACE")

Andrew, can you pick this up too?

It conflicts significantly with Alexey's "ELF: fix overflow in total mapping size calculation".