[PATCH v2] fork: generate stub on no-MMU systems
From: Ben Wolsieffer <Ben.Wolsieffer@hefring.com> fork() can be implemented using either the fork or clone syscalls on MMU systems. Therefore the stub is only generated if neither __NR_fork nor __NR_clone are defined. The stub code manually undefines __NR_fork on no-MMU systems in an attempt to enable the stub, but this doesn't work because __NR_clone is still defined. It is not appropriate to undefine __NR_clone because clone is available on no-MMU, it is just not capable of implementing fork. This patch directly enables the fork stub if __ARCH_USE_MMU__ is not defined. This eliminates the need to undefine __NR_fork, so this code is removed Signed-off-by: Ben Wolsieffer <ben.wolsieffer@hefring.com> --- libc/sysdeps/linux/common/stubs.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/libc/sysdeps/linux/common/stubs.c b/libc/sysdeps/linux/common/stubs.c index c17e509b9..634dcde0d 100644 --- a/libc/sysdeps/linux/common/stubs.c +++ b/libc/sysdeps/linux/common/stubs.c @@ -34,10 +34,6 @@ __attribute_used__ static int ret_enosys_stub(void) link_warning(stub, #stub ": this function is not implemented") \ strong_alias(ret_enosys_stub, stub) -#ifndef __ARCH_USE_MMU__ -# undef __NR_fork -#endif - #ifdef __arm__ # define __NR_fadvise64_64 __NR_arm_fadvise64_64 # define __NR_fadvise64 __NR_arm_fadvise64_64 @@ -120,7 +116,7 @@ make_stub(fgetxattr) make_stub(flistxattr) #endif -#if !defined __NR_fork && !defined __NR_clone +#if !defined __ARCH_USE_MMU__ || (!defined __NR_fork && !defined __NR_clone) make_stub(fork) #endif -- 2.42.0
Hi Ben, patch applied and pushed. best regards Waldemar Ben Wolsieffer wrote,
From: Ben Wolsieffer <Ben.Wolsieffer@hefring.com>
fork() can be implemented using either the fork or clone syscalls on MMU systems. Therefore the stub is only generated if neither __NR_fork nor __NR_clone are defined. The stub code manually undefines __NR_fork on no-MMU systems in an attempt to enable the stub, but this doesn't work because __NR_clone is still defined. It is not appropriate to undefine __NR_clone because clone is available on no-MMU, it is just not capable of implementing fork.
This patch directly enables the fork stub if __ARCH_USE_MMU__ is not defined. This eliminates the need to undefine __NR_fork, so this code is removed
Signed-off-by: Ben Wolsieffer <ben.wolsieffer@hefring.com> --- libc/sysdeps/linux/common/stubs.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-)
diff --git a/libc/sysdeps/linux/common/stubs.c b/libc/sysdeps/linux/common/stubs.c index c17e509b9..634dcde0d 100644 --- a/libc/sysdeps/linux/common/stubs.c +++ b/libc/sysdeps/linux/common/stubs.c @@ -34,10 +34,6 @@ __attribute_used__ static int ret_enosys_stub(void) link_warning(stub, #stub ": this function is not implemented") \ strong_alias(ret_enosys_stub, stub)
-#ifndef __ARCH_USE_MMU__ -# undef __NR_fork -#endif - #ifdef __arm__ # define __NR_fadvise64_64 __NR_arm_fadvise64_64 # define __NR_fadvise64 __NR_arm_fadvise64_64 @@ -120,7 +116,7 @@ make_stub(fgetxattr) make_stub(flistxattr) #endif
-#if !defined __NR_fork && !defined __NR_clone +#if !defined __ARCH_USE_MMU__ || (!defined __NR_fork && !defined __NR_clone) make_stub(fork) #endif
-- 2.42.0
"Waldemar" == Waldemar Brodkorb <wbx@uclibc-ng.org> writes:
Hi Ben, patch applied and pushed.
What is the impact of this exactly? As far as I can see, this causes linking something using fork() on nommu to now generate a linker warning but no longer fail with a missing symbol. Is that an improvement? Looking with my Buildroot glasses, I could imagine that such linker warnings could easily be missed in a big build, whereas a linker error naturally not.
best regards Waldemar
Ben Wolsieffer wrote,
From: Ben Wolsieffer <Ben.Wolsieffer@hefring.com>
fork() can be implemented using either the fork or clone syscalls on MMU systems. Therefore the stub is only generated if neither __NR_fork nor __NR_clone are defined. The stub code manually undefines __NR_fork on no-MMU systems in an attempt to enable the stub, but this doesn't work because __NR_clone is still defined. It is not appropriate to undefine __NR_clone because clone is available on no-MMU, it is just not capable of implementing fork.
This patch directly enables the fork stub if __ARCH_USE_MMU__ is not defined. This eliminates the need to undefine __NR_fork, so this code is removed
Signed-off-by: Ben Wolsieffer <ben.wolsieffer@hefring.com> --- libc/sysdeps/linux/common/stubs.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-)
diff --git a/libc/sysdeps/linux/common/stubs.c b/libc/sysdeps/linux/common/stubs.c index c17e509b9..634dcde0d 100644 --- a/libc/sysdeps/linux/common/stubs.c +++ b/libc/sysdeps/linux/common/stubs.c @@ -34,10 +34,6 @@ __attribute_used__ static int ret_enosys_stub(void) link_warning(stub, #stub ": this function is not implemented") \ strong_alias(ret_enosys_stub, stub)
-#ifndef __ARCH_USE_MMU__ -# undef __NR_fork -#endif - #ifdef __arm__ # define __NR_fadvise64_64 __NR_arm_fadvise64_64 # define __NR_fadvise64 __NR_arm_fadvise64_64 @@ -120,7 +116,7 @@ make_stub(fgetxattr) make_stub(flistxattr) #endif
-#if !defined __NR_fork && !defined __NR_clone +#if !defined __ARCH_USE_MMU__ || (!defined __NR_fork && !defined __NR_clone) make_stub(fork) #endif
-- 2.42.0
_______________________________________________ devel mailing list -- devel@uclibc-ng.org To unsubscribe send an email to devel-leave@uclibc-ng.org
-- Bye, Peter Korsgaard
On Mon, Sep 18, 2023 at 01:21:32PM +0200, Peter Korsgaard wrote:
"Waldemar" == Waldemar Brodkorb <wbx@uclibc-ng.org> writes:
Hi Ben, patch applied and pushed.
What is the impact of this exactly? As far as I can see, this causes linking something using fork() on nommu to now generate a linker warning but no longer fail with a missing symbol.
Is that an improvement? Looking with my Buildroot glasses, I could imagine that such linker warnings could easily be missed in a big build, whereas a linker error naturally not.
There is a lot of software out there that uses fork() only to implement some optional features, but isn't designed to gracefully disable building those features if fork() isn't available at build time. Think, for example, of software that can either daemonize itself or run in the foreground. Rather than patching each package to remove fork() calls, we can enable the stubs and just be careful to only run the software in a way that doesn't require fork().
best regards Waldemar
Ben Wolsieffer wrote,
From: Ben Wolsieffer <Ben.Wolsieffer@hefring.com>
fork() can be implemented using either the fork or clone syscalls on MMU systems. Therefore the stub is only generated if neither __NR_fork nor __NR_clone are defined. The stub code manually undefines __NR_fork on no-MMU systems in an attempt to enable the stub, but this doesn't work because __NR_clone is still defined. It is not appropriate to undefine __NR_clone because clone is available on no-MMU, it is just not capable of implementing fork.
This patch directly enables the fork stub if __ARCH_USE_MMU__ is not defined. This eliminates the need to undefine __NR_fork, so this code is removed
Signed-off-by: Ben Wolsieffer <ben.wolsieffer@hefring.com> --- libc/sysdeps/linux/common/stubs.c | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-)
diff --git a/libc/sysdeps/linux/common/stubs.c b/libc/sysdeps/linux/common/stubs.c index c17e509b9..634dcde0d 100644 --- a/libc/sysdeps/linux/common/stubs.c +++ b/libc/sysdeps/linux/common/stubs.c @@ -34,10 +34,6 @@ __attribute_used__ static int ret_enosys_stub(void) link_warning(stub, #stub ": this function is not implemented") \ strong_alias(ret_enosys_stub, stub)
-#ifndef __ARCH_USE_MMU__ -# undef __NR_fork -#endif - #ifdef __arm__ # define __NR_fadvise64_64 __NR_arm_fadvise64_64 # define __NR_fadvise64 __NR_arm_fadvise64_64 @@ -120,7 +116,7 @@ make_stub(fgetxattr) make_stub(flistxattr) #endif
-#if !defined __NR_fork && !defined __NR_clone +#if !defined __ARCH_USE_MMU__ || (!defined __NR_fork && !defined __NR_clone) make_stub(fork) #endif
-- 2.42.0
_______________________________________________ devel mailing list -- devel@uclibc-ng.org To unsubscribe send an email to devel-leave@uclibc-ng.org
-- Bye, Peter Korsgaard
participants (3)
-
Ben Wolsieffer -
Peter Korsgaard -
Waldemar Brodkorb