On 11/25/21 1:25 AM, Arnd Bergmann wrote: > On Thu, Nov 25, 2021 at 12:38 AM Rob Landley <rob@xxxxxxxxxxx> wrote: >> On 11/24/21 1:49 AM, Arnd Bergmann wrote: >> > On Wed, Nov 24, 2021 at 8:31 AM Rob Landley <rob@xxxxxxxxxxx> wrote: > >> > Did you test clone3? >> >> Haven't got anything that's using it (musl-libc doesn't know about it yet) but >> it looked straightforward? (Unlike the #ifdef stack around the previous clone...) >> >> I can try building tools/testing/selftests/clone3 if you like, but for some >> reason the clone3 tests want -lcap which isn't in my cross compiler. (Because to >> test a clone system call, you need to manipulate capability bits. Of course.) >> Right, comment out the LDLIBS line in the makefile and the first 3 built, let's >> try those... Hmmm, it's saying the syscall isn't supported, because it's using >> syscall.h out of the cross compiler headers (not THIS kernel's #includes) which >> of course doesn't have it, and then clone3_selftests.h falls back to: >> >> #ifndef __NR_clone3 >> #define __NR_clone3 -1 >> #endif >> >> Right, stick a 435 in there and... it's still skipping it. Why is it still >> skipping it... because the RUNTIME syscall is returning ENOSYS. Ok, I have to go >> stick printk() calls into the kernel. (Do I have to #define those >> #YES_I_WANT_THIS_SYSCALL_WHY_WOULDNT_I macros? Hmmm...) > > This specific syscall is protected by a macro so it doesn't get implicitly > enabled without architecture specific review for those architectures using > include/uapi/asm-generic/unistd.h. Sigh. >> > This needs a custom wrapper on most architectures >> > to have sensible calling conventions. >> >> Define "sensible" in this context? It's a new 2 argument syscall? (Do you mean a >> libc wrapper?) >> >> > If sh doesn't need it, that should >> > be explained in the changelog text. >> >> I'm happy to try to fix stuff up, but I don't understand the objection. Does it >> do something other than what the old clone did, except without the need to pass >> more arguments than we necessarily have registers defined for? (Calls the same >> clone plumbing, which should call back into arch/sh/kernel/process_32.c already...?) >> >> The most recent clone3 arch addition was commit 59a4e0d5511b which also just >> pulled in the generic version. (Via #define NO_REALLY_I_WANT_THIS_SYSCALL rather >> than editing the tbl file? Looks like I've got some reading to do...) > > The best reference I could find is: > > https://lore.kernel.org/linux-api/20190604160944.4058-2-christian@xxxxxxxxxx/ Does not say what the special handling is. Does not provide an example of said special handling. Implied that only three do NOT need special handling, two of which are x86 and arm, which seems... convenient. Right, let's see what "grep -r clone arch/" says: m68k/kernel/process.c is obviously overriding arc/include/syscalls.h has sys_clone_wrapper() nios2/kernel/process.c has nios2_clone() openrisc/kernel/entry.S has __sys_clone() sparc/kernel/process.c has sparce_clone() h8300/kernel/process.c has its own sys_clone() ia64/kernel/process.c has ia64_clone() user mode linux is just weird. So the architectures that wrap clone are m68k, arc, nios2, openrisc, sparc, h8300, and ia64. Implying that the ones that DON'T are alpha, arm64, hexagon, nds32, parisc, s390, csky, microblaze, powerpc, sh, x86, arm, mips, riscv, and xtensa. Which would mean 2/3 of architectures don't wrap clone, and thus arch/sh not doing so isn't unusual. > If fork() and clone() don't need special handling on arch/sh, then > clone3 shouldn't > need it either, unless the existing ones are also wrong. It looks like > some architectures > override these to avoid leaking register state from the kernel to the > child process. $ cd arch/sh $ grep -r clone tools/Makefile:# Shamelessly cloned from ARM. kernel/process_32.c:int copy_thread(unsigned long clone_flags, unsigned long usp, unsigned long arg, kernel/process_32.c: if (clone_flags & CLONE_SETTLS) kernel/syscalls/syscall.tbl:120 common clone sys_clone kernel/syscalls/syscall.tbl:435 common clone3 sys_clone3 $ grep -r fork include/asm/cacheflush.h: * - flush_cache_dup mm(mm) handles cache flushing when forking kernel/entry-common.S: .globl ret_from_fork kernel/entry-common.S:ret_from_fork: kernel/cpu/init.c: * state prior to hand forking the idle loop. kernel/process_32.c:asmlinkage void ret_from_fork(void); kernel/process_32.c: p->thread.pc = (unsigned long) ret_from_fork; kernel/syscalls/syscall.tbl:2 common fork sys_fork kernel/syscalls/syscall.tbl:190 common vfork sys_vfork Hard to prove a negative, but I'm not seeing any wrappers. It's got some callbacks, but I think the existing plumbing is calling them already? > Arnd Rob
