| Message ID | 20250106115227.1365643-1-ism.hong@gmail.com (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | bc7584e009c39375294794f7ca751a6b2622c425 |
| Headers | show |
| Series | mips: fix shmctl/semctl/msgctl syscall for o32 | expand |
On Mon, Jan 6, 2025, at 12:52, Ism Hong wrote: > The commit 275f22148e87 ("ipc: rename old-style shmctl/semctl/msgctl > syscalls") switched various architectures to use sys_old_*ctl() with > ipc_parse_version, including mips n32/n64. However, for mips o32, commit > 0d6040d46817 ("arch: add split IPC system calls where needed") added > separate IPC syscalls without properly using the old-style handlers. > > This causes applications using uClibc-ng to fail with -EINVAL when > calling semctl/shmctl/msgctl with IPC_64 flag, as uClibc-ng uses the > syscall numbers from kernel headers to determine whether to use the IPC > multiplexer or split syscalls. In contrast, glibc is unaffected as it > uses a unified feature test macro __ASSUME_DIRECT_SYSVIPC_SYSCALLS > (disabled for mips-o32) to make this decision. > > Fix this by switching the o32 ABI entries for semctl, shmctl and msgctl > to use the old-style handlers, matching the behavior of other > architectures and fixing compatibility with uClibc-ng. > > Signed-off-by: Ism Hong <ism.hong@gmail.com> I just saw this making it into mainline and had another look, sorry I hadn't caught it earlier. It was an intentional decision to use the new-style IPC_64 semantics on architectures that didn't already have the separate system call. You may not like that choice, but it's been done this way for seven years now, and as far as I can tell, glibc relies on this behavior. I think this commit should be reverted, and uclibc be changed to implement the kernel ABI for these syscalls. Arnd
On Mon, Jan 27, 2025 at 10:20:49PM +0100, Arnd Bergmann wrote: > On Mon, Jan 6, 2025, at 12:52, Ism Hong wrote: > > The commit 275f22148e87 ("ipc: rename old-style shmctl/semctl/msgctl > > syscalls") switched various architectures to use sys_old_*ctl() with > > ipc_parse_version, including mips n32/n64. However, for mips o32, commit > > 0d6040d46817 ("arch: add split IPC system calls where needed") added > > separate IPC syscalls without properly using the old-style handlers. > > > > This causes applications using uClibc-ng to fail with -EINVAL when > > calling semctl/shmctl/msgctl with IPC_64 flag, as uClibc-ng uses the > > syscall numbers from kernel headers to determine whether to use the IPC > > multiplexer or split syscalls. In contrast, glibc is unaffected as it > > uses a unified feature test macro __ASSUME_DIRECT_SYSVIPC_SYSCALLS > > (disabled for mips-o32) to make this decision. > > > > Fix this by switching the o32 ABI entries for semctl, shmctl and msgctl > > to use the old-style handlers, matching the behavior of other > > architectures and fixing compatibility with uClibc-ng. > > > > Signed-off-by: Ism Hong <ism.hong@gmail.com> > > I just saw this making it into mainline and had another look, sorry > I hadn't caught it earlier. > > It was an intentional decision to use the new-style IPC_64 > semantics on architectures that didn't already have the > separate system call. > > You may not like that choice, but it's been done this way > for seven years now, and as far as I can tell, glibc relies > on this behavior. > > I think this commit should be reverted, and uclibc be changed > to implement the kernel ABI for these syscalls. I've prepared the revert Thomas.
Arnd Bergmann 於 2025/1/28 清晨5:20 寫道: > On Mon, Jan 6, 2025, at 12:52, Ism Hong wrote: >> The commit 275f22148e87 ("ipc: rename old-style shmctl/semctl/msgctl >> syscalls") switched various architectures to use sys_old_*ctl() with >> ipc_parse_version, including mips n32/n64. However, for mips o32, commit >> 0d6040d46817 ("arch: add split IPC system calls where needed") added >> separate IPC syscalls without properly using the old-style handlers. >> >> This causes applications using uClibc-ng to fail with -EINVAL when >> calling semctl/shmctl/msgctl with IPC_64 flag, as uClibc-ng uses the >> syscall numbers from kernel headers to determine whether to use the IPC >> multiplexer or split syscalls. In contrast, glibc is unaffected as it >> uses a unified feature test macro __ASSUME_DIRECT_SYSVIPC_SYSCALLS >> (disabled for mips-o32) to make this decision. >> >> Fix this by switching the o32 ABI entries for semctl, shmctl and msgctl >> to use the old-style handlers, matching the behavior of other >> architectures and fixing compatibility with uClibc-ng. >> >> Signed-off-by: Ism Hong <ism.hong@gmail.com> > > I just saw this making it into mainline and had another look, sorry > I hadn't caught it earlier. > > It was an intentional decision to use the new-style IPC_64 > semantics on architectures that didn't already have the > separate system call. > > You may not like that choice, but it's been done this way > for seven years now, and as far as I can tell, glibc relies > on this behavior. > > I think this commit should be reverted, and uclibc be changed > to implement the kernel ABI for these syscalls. > > Arnd Hi Arnd, Thank you for your explanation, and sorry for the late reply due to the Chinese New Year. From your response, I understand that the preferred solution would be to modify uClibc to align the behavior of semctl/shmctl/msgctl with glibc and musl by adopting a unified feature test macro. If that is the case, I will work on preparing a patch for uClibc accordingly. Please let me know if you have any suggestions or if I have misunderstood anything. Thanks, Ism
On Thu, Jan 30, 2025, at 15:46, Ism Hong wrote: > Arnd Bergmann 於 2025/1/28 清晨5:20 寫道: > > Thank you for your explanation, and sorry for the late reply due to the > Chinese New Year. > > From your response, I understand that the preferred solution would be > to modify uClibc to align the behavior of semctl/shmctl/msgctl with > glibc and musl by adopting a unified feature test macro. If that is the > case, I will work on preparing a patch for uClibc accordingly. Yes, that is correct. I think musl doesn't call the separate syscalls if the architecture provides the old-style socket() syscall, so it's not affected either way, but glibc needs the revert. Arnd
diff --git a/arch/mips/kernel/syscalls/syscall_o32.tbl b/arch/mips/kernel/syscalls/syscall_o32.tbl index 349b8aad1159..e8a57c206758 100644 --- a/arch/mips/kernel/syscalls/syscall_o32.tbl +++ b/arch/mips/kernel/syscalls/syscall_o32.tbl @@ -382,15 +382,15 @@ 368 o32 io_pgetevents sys_io_pgetevents_time32 compat_sys_io_pgetevents # room for arch specific calls 393 o32 semget sys_semget -394 o32 semctl sys_semctl compat_sys_semctl +394 o32 semctl sys_old_semctl compat_sys_old_semctl 395 o32 shmget sys_shmget -396 o32 shmctl sys_shmctl compat_sys_shmctl +396 o32 shmctl sys_old_shmctl compat_sys_old_shmctl 397 o32 shmat sys_shmat compat_sys_shmat 398 o32 shmdt sys_shmdt 399 o32 msgget sys_msgget 400 o32 msgsnd sys_msgsnd compat_sys_msgsnd 401 o32 msgrcv sys_msgrcv compat_sys_msgrcv -402 o32 msgctl sys_msgctl compat_sys_msgctl +402 o32 msgctl sys_old_msgctl compat_sys_old_msgctl 403 o32 clock_gettime64 sys_clock_gettime sys_clock_gettime 404 o32 clock_settime64 sys_clock_settime sys_clock_settime 405 o32 clock_adjtime64 sys_clock_adjtime sys_clock_adjtime
The commit 275f22148e87 ("ipc: rename old-style shmctl/semctl/msgctl syscalls") switched various architectures to use sys_old_*ctl() with ipc_parse_version, including mips n32/n64. However, for mips o32, commit 0d6040d46817 ("arch: add split IPC system calls where needed") added separate IPC syscalls without properly using the old-style handlers. This causes applications using uClibc-ng to fail with -EINVAL when calling semctl/shmctl/msgctl with IPC_64 flag, as uClibc-ng uses the syscall numbers from kernel headers to determine whether to use the IPC multiplexer or split syscalls. In contrast, glibc is unaffected as it uses a unified feature test macro __ASSUME_DIRECT_SYSVIPC_SYSCALLS (disabled for mips-o32) to make this decision. Fix this by switching the o32 ABI entries for semctl, shmctl and msgctl to use the old-style handlers, matching the behavior of other architectures and fixing compatibility with uClibc-ng. Signed-off-by: Ism Hong <ism.hong@gmail.com> --- arch/mips/kernel/syscalls/syscall_o32.tbl | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)