Skip to contentSkip to conversation

nolibc/stdlib: Improve `getauxval(3)` implementation

PATCH In review Charles Mirabile 3 participants 8 messages Jan 16, 2024 Nolibc header file

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 nolibc/stdlib: Improve `getauxval(3)` implementation

Previously the getauxval function checked for a doubly null entry (i.e. one whose type and value both were 0) in the auxv array before exiting the loop.

At least on x86-64, the ABI only specifies that one more long will be present with value 0 (type AT_NULL) after the pairs of auxv entries. Whether or not it has a corresponding value is unspecified. This value is present on linux, but there is no reason to check it as simply seeing an auxv entry whose type value is AT_NULL should be enough.

This is a matter of taste, but I think processing the data in a structured way by coercing it into an array of type value pairs, using multiple return style, and a for loop with a clear exit condition is more readable than the existing infinite loop with multiple exit points and a return value variable.

I also added a call to set errno to ENOENT when the entry is not found as glibc does which allows programs to disambiguate between the case of an auxv that is not present, and one that is, but with value zero.

Fixes c61a078015f3 ("nolibc/stdlib: Implement `getauxval(3)` function")
---
 tools/include/nolibc/stdlib.h | 34 +++++++++++++---------------------
 1 file changed, 13 insertions(+), 21 deletions(-)
tools/include/nolibc/stdlib.h +13 -21
index bacfd35c5156..47be99c5a539 100644
@@ -104,27 +104,19 @@ char *getenv(const char *name)
 static __attribute__((unused))
 unsigned long getauxval(unsigned long type)
 {
-	const unsigned long *auxv = _auxv;
-	unsigned long ret;
-
-	if (!auxv)
-		return 0;
-
-	while (1) {
-		if (!auxv[0] && !auxv[1]) {
-			ret = 0;
-			break;
-		}
-
-		if (auxv[0] == type) {
-			ret = auxv[1];
-			break;
-		}
-
-		auxv += 2;
-	}
-
-	return ret;
+	const struct {
+		unsigned long type, val;
+	} *auxv = (void *)_auxv;
+
+	if (!auxv || !type)
+		goto out;
+
+	for (; auxv->type; ++auxv)
+		if (auxv->type == type)
+			return auxv->val;
+out:
+	SET_ERRNO(ENOENT);
+	return 0;
 }
 
 static __attribute__((unused))
Signature · 1 line
2.41.0
#1 lore raw
Quoted 5 lines · Charles Mirabile

At least on x86-64, the ABI only specifies that one more long will be present with value 0 (type AT_NULL) after the pairs of auxv entries. Whether or not it has a corresponding value is unspecified. This value is present on linux, but there is no reason to check it as simply seeing an auxv entry whose type value is AT_NULL should be enough.

Yeah, I agree with that. I just read the ABI and confirmed that the 'a_un' member is ignored when the type is AT_NULL. Let's stop relying on an unspecified value.

For others who want to check, see page 37 and 38:
https://gitlab.com/x86-psABIs/x86-64-ABI/-/wikis/uploads/221b09355dd540efcbe61b783b6c0ece/x86-64-psABI-2023-09-26.pdf

Quoted 5 lines

This is a matter of taste, but I think processing the data in a structured way by coercing it into an array of type value pairs, using multiple return style, and a for loop with a clear exit condition is more readable than the existing infinite loop with multiple exit points and a return value variable.

Ok. It's more readable using your way. One thing that bothers me a bit is type of 'a_type'. On page 37, the ABI defines the auxv type-val pair as:

  typedef struct
  {
    int a_type;
    union {
      long a_val;
      void *a_ptr;
      void (*a_fnc)();
    } a_un;
  } auxv_t;

Assuming the arch is x86-64 Linux. Note that 'a_type' is an 'int' which is 4 bytes in size, but we use 'unsigned long' instead of 'int' to represent it. However, since 'a_un' needs to be 8 bytes aligned, the compiler will put a 4 bytes padding between 'a_type' and 'a_un', so it ends up just fine (on x86-64).

What do you think about other architectures? Will it potentially be misinterpreted?

I tried to compare it in my head for i386, but it also ends up just fine. I don't know other architectures.

Quoted 3 lines

I also added a call to set errno to ENOENT when the entry is not found as glibc does which allows programs to disambiguate between the case of an auxv that is not present, and one that is, but with value zero.

Good catch!

Signature · 1 line
Ammar Faizi
maintainer
#2 lore raw
Quoted 34 lines · Ammar Faizi
Quoted 5 lines · Charles Mirabile

At least on x86-64, the ABI only specifies that one more long will be present with value 0 (type AT_NULL) after the pairs of auxv entries. Whether or not it has a corresponding value is unspecified. This value is present on linux, but there is no reason to check it as simply seeing an auxv entry whose type value is AT_NULL should be enough.

Yeah, I agree with that. I just read the ABI and confirmed that the 'a_un' member is ignored when the type is AT_NULL. Let's stop relying on an unspecified value.

For others who want to check, see page 37 and 38:
https://gitlab.com/x86-psABIs/x86-64-ABI/-/wikis/uploads/221b09355dd540efcbe61b783b6c0ece/x86-64-psABI-2023-09-26.pdf

Quoted 5 lines

This is a matter of taste, but I think processing the data in a structured way by coercing it into an array of type value pairs, using multiple return style, and a for loop with a clear exit condition is more readable than the existing infinite loop with multiple exit points and a return value variable.

Ok. It's more readable using your way. One thing that bothers me a bit is type of 'a_type'. On page 37, the ABI defines the auxv type-val pair as:

  typedef struct
  {
    int a_type;
    union {
      long a_val;
      void *a_ptr;
      void (*a_fnc)();
    } a_un;
  } auxv_t;
Quoted 8 lines

Assuming the arch is x86-64 Linux. Note that 'a_type' is an 'int' which is 4 bytes in size, but we use 'unsigned long' instead of 'int' to represent it. However, since 'a_un' needs to be 8 bytes aligned, the compiler will put a 4 bytes padding between 'a_type' and 'a_un', so it ends up just fine (on x86-64).

What do you think about other architectures? Will it potentially be misinterpreted?

Indeed, it would fail on a 64-bit big endian architecture. Let's just declare the local variable the same way as it is in the spec, it will be much cleaner and more reliable.

Thanks,
Willy

maintainer
#3 lore raw
Quoted 41 lines · Willy Tarreau
Quoted 40 lines · Ammar Faizi
Quoted 5 lines · Charles Mirabile

At least on x86-64, the ABI only specifies that one more long will be present with value 0 (type AT_NULL) after the pairs of auxv entries. Whether or not it has a corresponding value is unspecified. This value is present on linux, but there is no reason to check it as simply seeing an auxv entry whose type value is AT_NULL should be enough.

Yeah, I agree with that. I just read the ABI and confirmed that the 'a_un' member is ignored when the type is AT_NULL. Let's stop relying on an unspecified value.

For others who want to check, see page 37 and 38:
https://gitlab.com/x86-psABIs/x86-64-ABI/-/wikis/uploads/221b09355dd540efcbe61b783b6c0ece/x86-64-psABI-2023-09-26.pdf

Quoted 5 lines

This is a matter of taste, but I think processing the data in a structured way by coercing it into an array of type value pairs, using multiple return style, and a for loop with a clear exit condition is more readable than the existing infinite loop with multiple exit points and a return value variable.

Ok. It's more readable using your way. One thing that bothers me a bit is type of 'a_type'. On page 37, the ABI defines the auxv type-val pair as:

  typedef struct
  {
    int a_type;
    union {
      long a_val;
      void *a_ptr;
      void (*a_fnc)();
    } a_un;
  } auxv_t;

Assuming the arch is x86-64 Linux. Note that 'a_type' is an 'int' which is 4 bytes in size, but we use 'unsigned long' instead of 'int' to represent it. However, since 'a_un' needs to be 8 bytes aligned, the compiler will put a 4 bytes padding between 'a_type' and 'a_un', so it ends up just fine (on x86-64).

Quoted 6 lines
Quoted 2 lines

What do you think about other architectures? Will it potentially be misinterpreted?

Indeed, it would fail on a 64-bit big endian architecture. Let's just declare the local variable the same way as it is in the spec, it will be much cleaner and more reliable.

With that said, if previous code used to work on such architectures, maybe the definition above is only for x86_64 and differs on other archs. Maybe it's really defined as two longs ?

Willy

#4 lore raw
Quoted 5 lines · Willy Tarreau
Quoted 4 lines · Willy Tarreau
Quoted 2 lines · Ammar Faizi

What do you think about other architectures? Will it potentially be misinterpreted?

Quoted 7 lines
Quoted 3 lines

Indeed, it would fail on a 64-bit big endian architecture. Let's just declare the local variable the same way as it is in the spec, it will be much cleaner and more reliable.

With that said, if previous code used to work on such architectures, maybe the definition above is only for x86_64 and differs on other archs. Maybe it's really defined as two longs ?

I just took a look at the kernel source code:
https://github.com/torvalds/linux/blob/v6.7/fs/binfmt_elf.c#L226-L261

The auxv is stored in elf_info variable, the type is elf_addr_t. Not sure what kind of typedef is that. I'll check.

Each auxv entry is added using this macro:

 #define NEW_AUX_ENT(id, val) \
         do { \
                 *elf_info++ = id; \
                 *elf_info++ = val; \
         } while (0)

where id is the type. That clearly implies type and val have the same size on the Linux kernel.

Signature · 1 line
Ammar Faizi
#5 lore raw
Quoted 22 lines · Ammar Faizi
Quoted 5 lines · Willy Tarreau
Quoted 4 lines · Willy Tarreau
Quoted 2 lines · Ammar Faizi

What do you think about other architectures? Will it potentially be misinterpreted?

Quoted 7 lines
Quoted 3 lines

Indeed, it would fail on a 64-bit big endian architecture. Let's just declare the local variable the same way as it is in the spec, it will be much cleaner and more reliable.

With that said, if previous code used to work on such architectures, maybe the definition above is only for x86_64 and differs on other archs. Maybe it's really defined as two longs ?

I just took a look at the kernel source code:
https://github.com/torvalds/linux/blob/v6.7/fs/binfmt_elf.c#L226-L261

The auxv is stored in elf_info variable, the type is elf_addr_t. Not sure what kind of typedef is that. I'll check.

Each auxv entry is added using this macro:

Quoted 8 lines
 #define NEW_AUX_ENT(id, val) \
         do { \
                 *elf_info++ = id; \
                 *elf_info++ = val; \
         } while (0)

where id is the type. That clearly implies type and val have the same size on the Linux kernel.

So here is the result:

1. 'elf_addr_t' defintion ( https://github.com/torvalds/linux/blob/v6.7/include/linux/elf.h#L38-L62 ):

  (simplified)
  #if ELF_CLASS == ELFCLASS32
  #define elf_addr_t	Elf32_Off
  #else
  #define elf_addr_t	Elf64_Off
  #endif

2. 'Elf32_Off' and 'Elf64_Off' typedefs ( https://github.com/torvalds/linux/blob/v6.7/include/uapi/linux/elf.h#L8-L23 )

typedef __u32 Elf32_Off;
typedef __u64 Elf64_Off;

Assuming 'ELFCLASS32' is for 32-bit architectures, then it's two __u64 on 64-bit arch, and two __u32 on 32-bit arch. That is identical to 'unsigned long' for both cases (on Linux). So it's fine to have 'unsigned long' for both 'type' and 'value'.

Signature · 1 line
Ammar Faizi
maintainer
#6 lore raw
Quoted 45 lines · Ammar Faizi
Quoted 22 lines · Ammar Faizi
Quoted 5 lines · Willy Tarreau
Quoted 4 lines · Willy Tarreau
Quoted 2 lines · Ammar Faizi

What do you think about other architectures? Will it potentially be misinterpreted?

Quoted 7 lines
Quoted 3 lines

Indeed, it would fail on a 64-bit big endian architecture. Let's just declare the local variable the same way as it is in the spec, it will be much cleaner and more reliable.

With that said, if previous code used to work on such architectures, maybe the definition above is only for x86_64 and differs on other archs. Maybe it's really defined as two longs ?

I just took a look at the kernel source code:
https://github.com/torvalds/linux/blob/v6.7/fs/binfmt_elf.c#L226-L261

The auxv is stored in elf_info variable, the type is elf_addr_t. Not sure what kind of typedef is that. I'll check.

Each auxv entry is added using this macro:

Quoted 8 lines
 #define NEW_AUX_ENT(id, val) \
         do { \
                 *elf_info++ = id; \
                 *elf_info++ = val; \
         } while (0)

where id is the type. That clearly implies type and val have the same size on the Linux kernel.

So here is the result:

1. 'elf_addr_t' defintion ( https://github.com/torvalds/linux/blob/v6.7/include/linux/elf.h#L38-L62 ):

  (simplified)
  #if ELF_CLASS == ELFCLASS32
  #define elf_addr_t	Elf32_Off
  #else
  #define elf_addr_t	Elf64_Off
  #endif

2. 'Elf32_Off' and 'Elf64_Off' typedefs ( https://github.com/torvalds/linux/blob/v6.7/include/uapi/linux/elf.h#L8-L23 )

Quoted 7 lines

typedef __u32 Elf32_Off;
typedef __u64 Elf64_Off;

Assuming 'ELFCLASS32' is for 32-bit architectures, then it's two __u64 on 64-bit arch, and two __u32 on 32-bit arch. That is identical to 'unsigned long' for both cases (on Linux). So it's fine to have 'unsigned long' for both 'type' and 'value'.

Yeah I agree, thanks for checking.

Willy