2013-01-07 19:35:47 +00:00
|
|
|
/****************************************************************************
|
2023-07-05 21:43:50 +08:00
|
|
|
* sched/task/task_fork.c
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
2024-09-11 13:45:11 +02:00
|
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
|
|
|
*
|
2020-05-16 09:06:29 -06:00
|
|
|
* Licensed to the Apache Software Foundation (ASF) under one or more
|
|
|
|
|
* contributor license agreements. See the NOTICE file distributed with
|
|
|
|
|
* this work for additional information regarding copyright ownership. The
|
|
|
|
|
* ASF licenses this file to you under the Apache License, Version 2.0 (the
|
|
|
|
|
* "License"); you may not use this file except in compliance with the
|
|
|
|
|
* License. You may obtain a copy of the License at
|
|
|
|
|
*
|
|
|
|
|
* http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
|
*
|
|
|
|
|
* Unless required by applicable law or agreed to in writing, software
|
|
|
|
|
* distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
|
|
|
|
|
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
|
|
|
|
|
* License for the specific language governing permissions and limitations
|
|
|
|
|
* under the License.
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
|
|
|
|
/****************************************************************************
|
|
|
|
|
* Included Files
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
|
|
|
|
#include <nuttx/config.h>
|
|
|
|
|
|
2013-01-27 19:17:56 +00:00
|
|
|
#include <sys/wait.h>
|
2013-01-07 19:35:47 +00:00
|
|
|
#include <stdint.h>
|
2016-11-19 07:30:01 -06:00
|
|
|
#include <sched.h>
|
2014-10-05 16:58:52 -06:00
|
|
|
#include <string.h>
|
2013-01-07 19:35:47 +00:00
|
|
|
#include <assert.h>
|
|
|
|
|
#include <errno.h>
|
2026-04-01 05:11:15 +05:30
|
|
|
#include <nuttx/debug.h>
|
2013-01-07 19:35:47 +00:00
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#include <nuttx/kmalloc.h>
|
2022-09-25 23:08:38 +08:00
|
|
|
#include <nuttx/queue.h>
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#include <nuttx/semaphore.h>
|
2022-09-25 23:08:38 +08:00
|
|
|
|
2014-08-08 17:53:55 -06:00
|
|
|
#include "sched/sched.h"
|
2022-04-04 13:11:06 +03:00
|
|
|
#include "environ/environ.h"
|
2014-08-08 14:06:42 -06:00
|
|
|
#include "group/group.h"
|
2014-08-08 16:44:08 -06:00
|
|
|
#include "task/task.h"
|
2022-05-28 00:43:49 +08:00
|
|
|
#include "tls/tls.h"
|
2013-01-07 19:35:47 +00:00
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
/****************************************************************************
|
|
|
|
|
* Private Function Prototypes
|
|
|
|
|
****************************************************************************/
|
2013-02-08 22:53:14 +00:00
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#if defined(CONFIG_ARCH_ADDRENV) && defined(CONFIG_ARCH_HAVE_FORK)
|
|
|
|
|
static void fork_inherit_stack(FAR struct tcb_s *parent,
|
|
|
|
|
FAR struct tcb_s *child);
|
|
|
|
|
static void fork_inherit_tls(FAR struct tcb_s *child);
|
|
|
|
|
static void fork_restore_parent_env(void);
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
/****************************************************************************
|
|
|
|
|
* Private Functions
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
|
|
|
|
#if defined(CONFIG_ARCH_ADDRENV) && defined(CONFIG_ARCH_HAVE_FORK)
|
|
|
|
|
/****************************************************************************
|
|
|
|
|
* Name: fork_inherit_stack
|
|
|
|
|
*
|
|
|
|
|
* Description:
|
|
|
|
|
* Give the fork() child the parent's stack at the parent's virtual
|
|
|
|
|
* address rather than a relocated copy. The child's address environment
|
|
|
|
|
* is a duplicate, so the parent's stack is already there -- same contents,
|
|
|
|
|
* same address, its own pages -- and nothing needs allocating or copying.
|
|
|
|
|
*
|
|
|
|
|
* A relocated stack would break plain C: a pointer to a local taken
|
|
|
|
|
* before the fork would name the parent's copy, not the child's live
|
|
|
|
|
* object.
|
|
|
|
|
*
|
|
|
|
|
* TCB_FLAG_FREE_STACK is left clear: the stack belongs to the duplicated
|
|
|
|
|
* image and is released with it, so up_release_stack() must not free it.
|
|
|
|
|
*
|
|
|
|
|
* Input Parameters:
|
|
|
|
|
* parent - The parent task's TCB
|
|
|
|
|
* child - The child task's TCB
|
|
|
|
|
*
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
|
|
|
|
static void fork_inherit_stack(FAR struct tcb_s *parent,
|
|
|
|
|
FAR struct tcb_s *child)
|
|
|
|
|
{
|
|
|
|
|
child->stack_alloc_ptr = parent->stack_alloc_ptr;
|
|
|
|
|
child->stack_base_ptr = parent->stack_base_ptr;
|
|
|
|
|
child->adj_stack_size = parent->adj_stack_size;
|
|
|
|
|
child->flags &= ~TCB_FLAG_FREE_STACK;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/****************************************************************************
|
|
|
|
|
* Name: fork_inherit_tls
|
|
|
|
|
*
|
|
|
|
|
* Description:
|
|
|
|
|
* Retarget the thread-local storage the fork() child inherited.
|
|
|
|
|
*
|
|
|
|
|
* tls_dup_info() cannot be used: it carves a fresh TLS block off the
|
|
|
|
|
* stack, which on an inherited stack would carve a second one and shift
|
|
|
|
|
* stack_base_ptr away from the parent's. The child's copy is already in
|
|
|
|
|
* place, so only the fields naming the task itself need correcting.
|
|
|
|
|
*
|
|
|
|
|
* The write lands in user memory at an address the parent also occupies,
|
|
|
|
|
* so the child's address environment must be current for it -- otherwise
|
|
|
|
|
* the parent's own TLS is what gets modified.
|
|
|
|
|
*
|
|
|
|
|
* Input Parameters:
|
|
|
|
|
* child - The child task's TCB
|
|
|
|
|
*
|
|
|
|
|
* Returned Value:
|
|
|
|
|
* Zero (OK) on success; a negated errno value on failure.
|
|
|
|
|
*
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
|
|
|
|
static void fork_inherit_tls(FAR struct tcb_s *child)
|
|
|
|
|
{
|
|
|
|
|
FAR struct tls_info_s *info = (FAR struct tls_info_s *)
|
|
|
|
|
child->stack_alloc_ptr;
|
|
|
|
|
|
|
|
|
|
info->tl_task = child->group->tg_info;
|
|
|
|
|
info->tl_tid = child->pid;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/****************************************************************************
|
|
|
|
|
* Name: fork_restore_parent_env
|
|
|
|
|
*
|
|
|
|
|
* Description:
|
|
|
|
|
* Undo the addrenv_select() that nxtask_setup_fork() made on the child's
|
|
|
|
|
* behalf, putting the caller back in its own address environment. The
|
|
|
|
|
* environment to go back to does not have to be remembered: the caller is
|
|
|
|
|
* the parent, and what was current before was the parent's own.
|
|
|
|
|
*
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
|
|
|
|
static void fork_restore_parent_env(void)
|
|
|
|
|
{
|
|
|
|
|
addrenv_restore(this_task()->addrenv_own);
|
|
|
|
|
}
|
|
|
|
|
#endif /* CONFIG_ARCH_ADDRENV && CONFIG_ARCH_HAVE_FORK */
|
2013-01-07 19:35:47 +00:00
|
|
|
|
|
|
|
|
/****************************************************************************
|
|
|
|
|
* Public Functions
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
|
|
|
|
/****************************************************************************
|
2023-07-05 21:43:50 +08:00
|
|
|
* Name: nxtask_setup_fork
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
|
|
|
|
* Description:
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
* Allocate and initialize the child task's TCB. This is one step in the
|
|
|
|
|
* sequence common to vfork() and fork(); see the comment above the
|
|
|
|
|
* prototype in include/nuttx/sched.h for the whole sequence and for what
|
|
|
|
|
* the two primitives mean.
|
|
|
|
|
*
|
|
|
|
|
* Exactly two things depend on `vfork':
|
|
|
|
|
*
|
|
|
|
|
* - the address environment: vfork() joins the parent's, fork()
|
|
|
|
|
* duplicates it.
|
|
|
|
|
* - the stack: a vfork() child gets its own, which the architecture code
|
|
|
|
|
* fills with a relocated copy; a fork() child inherits the parent's
|
|
|
|
|
* address (fork_inherit_stack()).
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
2014-09-29 10:45:44 -06:00
|
|
|
* Input Parameters:
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
* retaddr - Address at which the child resumes
|
|
|
|
|
* vfork - true for vfork(), false for fork()
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
|
|
|
|
* Returned Value:
|
2023-07-05 21:43:50 +08:00
|
|
|
* Upon successful completion, nxtask_setup_fork() returns a pointer to
|
2014-09-29 10:45:44 -06:00
|
|
|
* newly allocated and initialized child task's TCB. NULL is returned
|
2013-01-07 19:35:47 +00:00
|
|
|
* on any failure and the errno is set appropriately.
|
|
|
|
|
*
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
FAR struct tcb_s *nxtask_setup_fork(start_t retaddr, bool vfork)
|
2013-01-07 19:35:47 +00:00
|
|
|
{
|
2021-04-12 23:44:08 +08:00
|
|
|
FAR struct tcb_s *ptcb = this_task();
|
2021-11-29 11:00:27 +08:00
|
|
|
FAR struct tcb_s *parent;
|
2025-02-13 12:12:32 +08:00
|
|
|
FAR struct tcb_s *child;
|
2024-06-09 09:35:34 +08:00
|
|
|
FAR char **argv;
|
2021-04-12 23:44:08 +08:00
|
|
|
size_t stack_size;
|
2013-03-20 18:22:21 +00:00
|
|
|
uint8_t ttype;
|
2013-01-07 19:35:47 +00:00
|
|
|
int priority;
|
|
|
|
|
int ret;
|
|
|
|
|
|
2021-04-12 23:44:08 +08:00
|
|
|
DEBUGASSERT(retaddr != NULL);
|
2013-01-07 19:35:47 +00:00
|
|
|
|
2013-03-20 18:22:21 +00:00
|
|
|
/* Get the type of the fork'ed task (kernel or user) */
|
|
|
|
|
|
2021-04-12 23:44:08 +08:00
|
|
|
if ((ptcb->flags & TCB_FLAG_TTYPE_MASK) == TCB_FLAG_TTYPE_KERNEL)
|
2013-03-20 18:22:21 +00:00
|
|
|
{
|
|
|
|
|
/* Fork'ed from a kernel thread */
|
|
|
|
|
|
|
|
|
|
ttype = TCB_FLAG_TTYPE_KERNEL;
|
2021-11-29 11:00:27 +08:00
|
|
|
parent = ptcb;
|
2013-03-20 18:22:21 +00:00
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
/* Fork'ed from a user task or pthread */
|
|
|
|
|
|
|
|
|
|
ttype = TCB_FLAG_TTYPE_TASK;
|
2021-04-12 23:44:08 +08:00
|
|
|
if ((ptcb->flags & TCB_FLAG_TTYPE_MASK) == TCB_FLAG_TTYPE_TASK)
|
|
|
|
|
{
|
2021-11-29 11:00:27 +08:00
|
|
|
parent = ptcb;
|
2021-04-12 23:44:08 +08:00
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
2021-11-29 11:00:27 +08:00
|
|
|
parent = nxsched_get_tcb(ptcb->group->tg_pid);
|
2021-04-12 23:44:08 +08:00
|
|
|
if (parent == NULL)
|
|
|
|
|
{
|
|
|
|
|
ret = -ENOENT;
|
|
|
|
|
goto errout;
|
|
|
|
|
}
|
|
|
|
|
}
|
2013-03-20 18:22:21 +00:00
|
|
|
}
|
|
|
|
|
|
2013-01-07 19:35:47 +00:00
|
|
|
/* Allocate a TCB for the child task. */
|
|
|
|
|
|
2025-12-30 19:58:12 +08:00
|
|
|
child = kmm_zalloc(sizeof(struct tcb_s));
|
2013-01-07 19:35:47 +00:00
|
|
|
if (!child)
|
|
|
|
|
{
|
2016-06-11 15:50:49 -06:00
|
|
|
serr("ERROR: Failed to allocate TCB\n");
|
2021-04-12 23:44:08 +08:00
|
|
|
ret = -ENOMEM;
|
|
|
|
|
goto errout;
|
2013-01-07 19:35:47 +00:00
|
|
|
}
|
|
|
|
|
|
2025-02-13 12:12:32 +08:00
|
|
|
child->flags |= TCB_FLAG_FREE_TCB;
|
2024-03-04 09:19:27 +08:00
|
|
|
|
sched/pthread/join: refactor pthread join to support join task
1. add support to join main task
| static pthread_t self;
|
| static void *join_task(void *arg)
| {
| int ret;
| ret = pthread_join(self, NULL); <--- /* Fix Task could not be joined */
| return NULL;
| }
|
| int main(int argc, char *argv[])
| {
| pthread_t thread;
|
| self = pthread_self();
|
| pthread_create(&thread, NULL, join_task, NULL);
| sleep(1);
|
| pthread_exit(NULL);
| return 0;
| }
2. Detach active thread will not alloc for additional join, just update the task flag.
3. Remove the return value waiting lock logic (data_sem),
the return value will be stored in the waiting tcb.
4. Revise the return value of pthread_join(), consistent with linux
e.g:
Joining a detached and canceled thread should return EINVAL, not ESRCH
https://pubs.opengroup.org/onlinepubs/009695399/functions/pthread_join.html
[EINVAL]
The value specified by thread does not refer to a joinable thread.
NOTE:
This PR will not increase stack usage, but struct tcb_s will increase 32 bytes.
Signed-off-by: chao an <anchao@lixiang.com>
2024-03-12 15:10:11 +08:00
|
|
|
/* Initialize the task join */
|
|
|
|
|
|
2025-02-13 12:12:32 +08:00
|
|
|
nxtask_joininit(child);
|
sched/pthread/join: refactor pthread join to support join task
1. add support to join main task
| static pthread_t self;
|
| static void *join_task(void *arg)
| {
| int ret;
| ret = pthread_join(self, NULL); <--- /* Fix Task could not be joined */
| return NULL;
| }
|
| int main(int argc, char *argv[])
| {
| pthread_t thread;
|
| self = pthread_self();
|
| pthread_create(&thread, NULL, join_task, NULL);
| sleep(1);
|
| pthread_exit(NULL);
| return 0;
| }
2. Detach active thread will not alloc for additional join, just update the task flag.
3. Remove the return value waiting lock logic (data_sem),
the return value will be stored in the waiting tcb.
4. Revise the return value of pthread_join(), consistent with linux
e.g:
Joining a detached and canceled thread should return EINVAL, not ESRCH
https://pubs.opengroup.org/onlinepubs/009695399/functions/pthread_join.html
[EINVAL]
The value specified by thread does not refer to a joinable thread.
NOTE:
This PR will not increase stack usage, but struct tcb_s will increase 32 bytes.
Signed-off-by: chao an <anchao@lixiang.com>
2024-03-12 15:10:11 +08:00
|
|
|
|
2014-09-03 14:58:24 -06:00
|
|
|
/* Allocate a new task group with the same privileges as the parent */
|
2013-02-08 21:42:23 +00:00
|
|
|
|
2025-12-18 19:51:21 +08:00
|
|
|
ret = group_allocate(child, ttype);
|
2013-02-08 21:42:23 +00:00
|
|
|
if (ret < 0)
|
|
|
|
|
{
|
|
|
|
|
goto errout_with_tcb;
|
|
|
|
|
}
|
2025-05-06 21:17:36 +08:00
|
|
|
|
|
|
|
|
#if defined(CONFIG_ARCH_ADDRENV)
|
|
|
|
|
if (ttype != TCB_FLAG_TTYPE_KERNEL)
|
|
|
|
|
{
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
if (vfork)
|
|
|
|
|
{
|
|
|
|
|
/* vfork(): join the parent address environment, exactly as
|
|
|
|
|
* pthread_create() does. The child shares .data, .bss and the
|
|
|
|
|
* heap.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
ret = addrenv_join(parent, child);
|
|
|
|
|
}
|
|
|
|
|
#ifdef CONFIG_ARCH_HAVE_FORK
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
/* POSIX fork(): duplicate the parent's address environment now,
|
|
|
|
|
* before anything else is set up. The duplicate holds a copy of
|
|
|
|
|
* the parent's contents -- including its stack -- at the parent's
|
|
|
|
|
* virtual addresses, which is what lets the child go on to inherit
|
|
|
|
|
* the stack address rather than be given a relocated copy. See
|
|
|
|
|
* fork_inherit_stack().
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
ret = addrenv_fork(parent, child);
|
|
|
|
|
if (ret >= 0)
|
|
|
|
|
{
|
|
|
|
|
/* Make the child's address environment current for the rest of
|
|
|
|
|
* the setup, and for the architecture code that runs after it.
|
|
|
|
|
*
|
|
|
|
|
* From here on, everything written on the child's behalf
|
|
|
|
|
* has to land in the child's image rather than the parent's,
|
|
|
|
|
* because
|
|
|
|
|
* the two occupy the same virtual addresses: its thread-local
|
|
|
|
|
* storage, and -- on architectures that keep the register save
|
|
|
|
|
* area on the user stack rather than on a kernel stack -- the
|
|
|
|
|
* register context the child is resumed from. Writing those
|
|
|
|
|
* under the parent's environment corrupts the parent and
|
|
|
|
|
* leaves the child reading whatever the snapshot happened to
|
|
|
|
|
* contain.
|
|
|
|
|
*
|
|
|
|
|
* Reads are unaffected: everything the setup reads from the
|
|
|
|
|
* parent -- environ, the argument vector -- is legible at the
|
|
|
|
|
* same address in the child, precisely because it is a copy.
|
|
|
|
|
*
|
|
|
|
|
* nxtask_start_fork() puts the parent's environment back.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
FAR struct addrenv_s *oldenv;
|
|
|
|
|
|
|
|
|
|
ret = addrenv_select(child->addrenv_own, &oldenv);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
#else
|
|
|
|
|
/* An address environment without ARCH_HAVE_FORK -- a protected build
|
|
|
|
|
* over an MMU, for instance. There is an address environment to join,
|
|
|
|
|
* but no POSIX fork() to duplicate it for, so the branch above is not
|
|
|
|
|
* compiled and `vfork' is always true here.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
DEBUGASSERT(vfork);
|
|
|
|
|
#endif
|
|
|
|
|
|
2025-05-06 21:17:36 +08:00
|
|
|
if (ret < 0)
|
|
|
|
|
{
|
|
|
|
|
goto errout_with_tcb;
|
|
|
|
|
}
|
|
|
|
|
}
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#else
|
|
|
|
|
/* Without address environments there is only one address space, so
|
|
|
|
|
* everything except the stack is shared no matter which primitive was
|
|
|
|
|
* called. POSIX fork() cannot be provided at all, and CONFIG_ARCH_HAVE_
|
|
|
|
|
* FORK is not selected, so `vfork' is always true here.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
DEBUGASSERT(vfork);
|
2025-05-06 21:17:36 +08:00
|
|
|
#endif
|
2022-04-04 13:11:06 +03:00
|
|
|
|
|
|
|
|
/* Duplicate the parent tasks environment */
|
|
|
|
|
|
2025-02-13 12:12:32 +08:00
|
|
|
ret = env_dup(child->group, environ);
|
2022-04-04 13:11:06 +03:00
|
|
|
if (ret < 0)
|
|
|
|
|
{
|
|
|
|
|
goto errout_with_tcb;
|
|
|
|
|
}
|
2013-02-08 21:42:23 +00:00
|
|
|
|
2013-01-07 19:35:47 +00:00
|
|
|
/* Associate file descriptors with the new task */
|
|
|
|
|
|
2023-11-13 20:59:42 +08:00
|
|
|
ret = group_setuptaskfiles(child, NULL, false);
|
2013-02-08 21:42:23 +00:00
|
|
|
if (ret < OK)
|
2013-01-07 19:35:47 +00:00
|
|
|
{
|
|
|
|
|
goto errout_with_tcb;
|
|
|
|
|
}
|
|
|
|
|
|
2024-09-25 17:14:01 +08:00
|
|
|
/* Set the task name */
|
|
|
|
|
|
|
|
|
|
argv = nxsched_get_stackargs(parent);
|
|
|
|
|
nxtask_setup_name(child, argv[0]);
|
|
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
/* Allocate the stack for the TCB, or inherit the parent's */
|
|
|
|
|
|
|
|
|
|
#if defined(CONFIG_ARCH_ADDRENV) && defined(CONFIG_ARCH_HAVE_FORK)
|
|
|
|
|
if (!vfork && ttype != TCB_FLAG_TTYPE_KERNEL)
|
|
|
|
|
{
|
|
|
|
|
/* The child's copy of the parent's stack is already in place, at the
|
|
|
|
|
* parent's address, courtesy of the duplication above.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
fork_inherit_stack(parent, child);
|
|
|
|
|
ret = OK;
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
#endif
|
|
|
|
|
{
|
|
|
|
|
stack_size = (uintptr_t)ptcb->stack_base_ptr -
|
|
|
|
|
(uintptr_t)ptcb->stack_alloc_ptr + ptcb->adj_stack_size;
|
2021-04-12 23:44:08 +08:00
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
ret = up_create_stack(child, stack_size, ttype);
|
|
|
|
|
}
|
2021-04-12 23:44:08 +08:00
|
|
|
|
|
|
|
|
if (ret < OK)
|
|
|
|
|
{
|
|
|
|
|
goto errout_with_tcb;
|
|
|
|
|
}
|
|
|
|
|
|
2024-08-01 12:41:41 +03:00
|
|
|
#if defined(CONFIG_ARCH_ADDRENV) && defined(CONFIG_ARCH_KERNEL_STACK)
|
|
|
|
|
/* Allocate the kernel stack */
|
|
|
|
|
|
|
|
|
|
if (ttype != TCB_FLAG_TTYPE_KERNEL)
|
|
|
|
|
{
|
2025-02-13 12:12:32 +08:00
|
|
|
ret = up_addrenv_kstackalloc(child);
|
2024-08-01 12:41:41 +03:00
|
|
|
if (ret < 0)
|
|
|
|
|
{
|
|
|
|
|
goto errout_with_tcb;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
2013-01-07 19:35:47 +00:00
|
|
|
/* Get the priority of the parent task */
|
|
|
|
|
|
|
|
|
|
#ifdef CONFIG_PRIORITY_INHERITANCE
|
2021-04-12 23:44:08 +08:00
|
|
|
priority = ptcb->base_priority; /* "Normal," unboosted priority */
|
2013-01-07 19:35:47 +00:00
|
|
|
#else
|
2021-04-12 23:44:08 +08:00
|
|
|
priority = ptcb->sched_priority; /* Current priority */
|
2013-01-07 19:35:47 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
/* Initialize the task control block. This calls up_initial_state() */
|
|
|
|
|
|
2016-06-11 11:59:51 -06:00
|
|
|
sinfo("Child priority=%d start=%p\n", priority, retaddr);
|
2021-04-12 23:44:08 +08:00
|
|
|
ret = nxtask_setup_scheduler(child, priority, retaddr,
|
|
|
|
|
ptcb->entry.main, ttype);
|
2013-02-08 21:42:23 +00:00
|
|
|
if (ret < OK)
|
2013-01-07 19:35:47 +00:00
|
|
|
{
|
|
|
|
|
goto errout_with_tcb;
|
|
|
|
|
}
|
|
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
/* Set up thread local storage and the argument vector.
|
|
|
|
|
*
|
|
|
|
|
* A fork() child that inherited its stack already has both, byte for
|
|
|
|
|
* byte, at the addresses the parent has them at -- they came across with
|
|
|
|
|
* the rest of the image. Re-creating them would carve fresh frames off a
|
|
|
|
|
* stack that already contains them, moving stack_base_ptr away from the
|
|
|
|
|
* parent's and undoing the inheritance. Only the TLS fields that name
|
|
|
|
|
* the task itself need correcting.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#if defined(CONFIG_ARCH_ADDRENV) && defined(CONFIG_ARCH_HAVE_FORK)
|
|
|
|
|
if (!vfork && ttype != TCB_FLAG_TTYPE_KERNEL)
|
2025-06-06 18:32:01 +08:00
|
|
|
{
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
fork_inherit_tls(child);
|
2025-06-06 18:32:01 +08:00
|
|
|
}
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
else
|
|
|
|
|
#endif
|
2022-05-31 16:58:07 +08:00
|
|
|
{
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
ret = tls_dup_info(child, parent);
|
|
|
|
|
if (ret < OK)
|
|
|
|
|
{
|
|
|
|
|
goto errout_with_tcb;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ret = nxtask_setup_stackargs(child, argv[0], &argv[1]);
|
|
|
|
|
if (ret < OK)
|
|
|
|
|
{
|
|
|
|
|
goto errout_with_tcb;
|
|
|
|
|
}
|
2022-05-31 16:58:07 +08:00
|
|
|
}
|
2021-04-12 23:44:08 +08:00
|
|
|
|
|
|
|
|
/* Now we have enough in place that we can join the group */
|
2018-08-24 07:43:00 -06:00
|
|
|
|
2025-12-18 19:51:21 +08:00
|
|
|
group_initialize(child);
|
2016-06-11 11:59:51 -06:00
|
|
|
sinfo("parent=%p, returning child=%p\n", parent, child);
|
2013-02-07 16:58:43 +00:00
|
|
|
return child;
|
2013-01-07 19:35:47 +00:00
|
|
|
|
|
|
|
|
errout_with_tcb:
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#if defined(CONFIG_ARCH_ADDRENV) && defined(CONFIG_ARCH_HAVE_FORK)
|
|
|
|
|
/* Get back into the parent's address environment before unwinding. If the
|
|
|
|
|
* duplication above never happened this is the environment we are already
|
|
|
|
|
* in, and addrenv_restore() is then a no-op.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
if (!vfork && ttype != TCB_FLAG_TTYPE_KERNEL)
|
|
|
|
|
{
|
|
|
|
|
fork_restore_parent_env();
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
2020-05-09 08:04:45 -06:00
|
|
|
nxsched_release_tcb((FAR struct tcb_s *)child, ttype);
|
2021-04-12 23:44:08 +08:00
|
|
|
errout:
|
2013-01-07 19:35:47 +00:00
|
|
|
set_errno(-ret);
|
|
|
|
|
return NULL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/****************************************************************************
|
2023-07-05 21:43:50 +08:00
|
|
|
* Name: nxtask_start_fork
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
|
|
|
|
* Description:
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
* The last step of both primitives: finish the child and run it. The
|
|
|
|
|
* architecture-specific code calls this once it has built the child's
|
|
|
|
|
* register context and stack.
|
|
|
|
|
*
|
|
|
|
|
* A vfork() additionally suspends the caller until the child calls _exit()
|
|
|
|
|
* or one of the exec family of functions. The suspension lives here, in
|
|
|
|
|
* the kernel primitive, rather than in a libc waitpid() as it once did.
|
|
|
|
|
* Two things follow from that. The parent is released when the child's
|
|
|
|
|
* TCB is torn down (see nxtask_resume_vfork()), which for an exec()ing
|
|
|
|
|
* child is immediately after exec_swap() has handed the child's pid to the
|
|
|
|
|
* program it loaded -- so the parent resumes at exec(), holding a pid that
|
|
|
|
|
* names the running program, as POSIX requires. And vfork() no longer
|
|
|
|
|
* depends on CONFIG_SCHED_WAITPID.
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
2014-09-29 10:45:44 -06:00
|
|
|
* Input Parameters:
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
* child - The tcb_s struct instance created by nxtask_setup_fork()
|
|
|
|
|
* vfork - true for vfork(), false for fork()
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
|
|
|
|
* Returned Value:
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
* The process ID of the child, or ERROR on failure.
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
pid_t nxtask_start_fork(FAR struct tcb_s *child, bool vfork)
|
2013-01-07 19:35:47 +00:00
|
|
|
{
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#ifdef CONFIG_ARCH_HAVE_VFORK
|
|
|
|
|
/* The rendezvous between the suspended parent and the child lives in this
|
|
|
|
|
* frame: the parent is blocked here for the whole lifetime of the child,
|
|
|
|
|
* so the storage is alive exactly as long as it is needed, and no
|
|
|
|
|
* allocation is required on a path that must not fail.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
sem_t rel;
|
|
|
|
|
int ret;
|
|
|
|
|
#endif
|
2013-01-07 19:35:47 +00:00
|
|
|
pid_t pid;
|
|
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
sinfo("Starting Child TCB=%p vfork=%d\n", child, vfork);
|
2013-01-07 19:35:47 +00:00
|
|
|
DEBUGASSERT(child);
|
|
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#if defined(CONFIG_ARCH_ADDRENV) && defined(CONFIG_ARCH_HAVE_FORK)
|
|
|
|
|
/* The architecture code has finished writing the child's image, so put the
|
|
|
|
|
* parent back in its own address environment. See nxtask_setup_fork().
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
if (!vfork &&
|
|
|
|
|
(child->flags & TCB_FLAG_TTYPE_MASK) != TCB_FLAG_TTYPE_KERNEL)
|
|
|
|
|
{
|
|
|
|
|
fork_restore_parent_env();
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
2013-01-07 19:35:47 +00:00
|
|
|
/* Get the assigned pid before we start the task */
|
|
|
|
|
|
2025-02-13 12:12:32 +08:00
|
|
|
pid = child->pid;
|
2013-01-07 19:35:47 +00:00
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#ifdef CONFIG_ARCH_HAVE_VFORK
|
|
|
|
|
if (vfork)
|
|
|
|
|
{
|
|
|
|
|
nxsem_init(&rel, 0, 0);
|
|
|
|
|
child->vfork_rel = &rel;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
2013-01-07 19:35:47 +00:00
|
|
|
/* Activate the task */
|
|
|
|
|
|
2025-02-13 12:12:32 +08:00
|
|
|
nxtask_activate(child);
|
2013-01-07 19:35:47 +00:00
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#ifdef CONFIG_ARCH_HAVE_VFORK
|
|
|
|
|
if (vfork)
|
|
|
|
|
{
|
|
|
|
|
/* Wait for the child to _exit() or exec(). This is not a cancellation
|
|
|
|
|
* point and must not be interrupted by a signal: the child may be
|
|
|
|
|
* running on our stack, so returning early would corrupt it.
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
do
|
|
|
|
|
{
|
|
|
|
|
ret = nxsem_wait_uninterruptible(&rel);
|
|
|
|
|
}
|
|
|
|
|
while (ret == -EINTR);
|
|
|
|
|
|
|
|
|
|
nxsem_destroy(&rel);
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
2013-01-07 19:35:47 +00:00
|
|
|
return pid;
|
|
|
|
|
}
|
|
|
|
|
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#ifdef CONFIG_ARCH_HAVE_VFORK
|
|
|
|
|
/****************************************************************************
|
|
|
|
|
* Name: nxtask_resume_vfork
|
|
|
|
|
*
|
|
|
|
|
* Description:
|
|
|
|
|
* Release the vfork() parent suspended on this child, if there is one.
|
|
|
|
|
*
|
|
|
|
|
* Called from nxsched_release_tcb(), the last point in the child's life,
|
|
|
|
|
* by which time an exec()ing child has already handed its pid to the
|
|
|
|
|
* program it loaded. nxtask_abort_fork() reaches it too, so a fork that
|
|
|
|
|
* fails after the rendezvous also releases the parent.
|
|
|
|
|
*
|
|
|
|
|
* Input Parameters:
|
|
|
|
|
* child - The TCB being torn down
|
|
|
|
|
*
|
|
|
|
|
* Returned Value:
|
|
|
|
|
* None
|
|
|
|
|
*
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
|
|
|
|
void nxtask_resume_vfork(FAR struct tcb_s *child)
|
|
|
|
|
{
|
|
|
|
|
FAR sem_t *rel = child->vfork_rel;
|
|
|
|
|
|
|
|
|
|
if (rel != NULL)
|
|
|
|
|
{
|
|
|
|
|
/* Clearing the pointer first is what makes this once-only. */
|
|
|
|
|
|
|
|
|
|
child->vfork_rel = NULL;
|
|
|
|
|
nxsem_post(rel);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
#endif /* CONFIG_ARCH_HAVE_VFORK */
|
|
|
|
|
|
2013-01-07 19:35:47 +00:00
|
|
|
/****************************************************************************
|
2023-07-05 21:43:50 +08:00
|
|
|
* Name: nxtask_abort_fork
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
|
|
|
|
* Description:
|
2023-07-05 21:43:50 +08:00
|
|
|
* Recover from any errors after nxtask_setup_fork() was called.
|
2013-01-07 19:35:47 +00:00
|
|
|
*
|
|
|
|
|
* Returned Value:
|
|
|
|
|
* None
|
|
|
|
|
*
|
|
|
|
|
****************************************************************************/
|
|
|
|
|
|
2025-02-13 12:12:32 +08:00
|
|
|
void nxtask_abort_fork(FAR struct tcb_s *child, int errcode)
|
2013-01-07 19:35:47 +00:00
|
|
|
{
|
!sched/arch/libc: Give fork() and vfork() their real, separate semantics.
NuttX implemented fork() and vfork() as the same function. Both were libc
wrappers around a single up_fork() syscall; vfork() differed only by a
trailing waitpid(). Underneath, the child joined the parent's address
environment -- the same addrenv_join() that pthread_create() uses -- and got
a private copy of the stack. So the child shared .data, .bss and the heap
with its parent and ran concurrently with it.
That is not fork(). It is vfork()-with-a-private-stack under fork()'s name,
and the history says so: today's fork() is NuttX's old vfork(), renamed in
c33d1c9c97 (2023) without any change of behaviour. The failure was silent --
a program written against POSIX fork() compiled, ran, and had its child's
writes land in the parent's variables.
Separate them into two primitives, chosen by which function the caller
called rather than by what the hardware happens to be:
fork() child gets its own copy of the parent's memory at the same
virtual addresses; runs concurrently. Only where an address
environment can be duplicated -- elsewhere it is not declared at
all, so calling it is a build error naming the function.
vfork() child shares the parent's memory; parent suspended until the
child _exit()s or exec()s. Implementable everywhere.
Below libc there is still one syscall. up_fork() gains a bool saying which
primitive the caller used, since the per-architecture register snapshot is
the same for both, and passes it to nxtask_setup_fork(), which is the single
place the memory semantics are decided. The argument arrives in the first
argument register and is never touched: each architecture's snapshot takes
some other call-clobbered register for its scratch, so the flag is simply
still there when the C worker is called.
The vfork() parent suspension moves out of libc into nxtask_start_fork(),
released from nxsched_release_tcb() by nxtask_resume_vfork(). Two things
follow: the parent is resumed at exec(), since exec_swap() has already handed
the child's pid to the loaded program by the time the vfork stub exits, and
vfork() no longer depends on CONFIG_SCHED_WAITPID.
Releasing there requires one fix in nxtask_exit(). It raises rtcb->lockcount
directly rather than through sched_lock() while it tears the TCB down, so the
nxsem_post() that wakes the vfork() parent leaves it queued where a blocked
task collects while pre-emption is off -- g_pendingtasks, or g_readytorun on
SMP -- and the matching raw lockcount-- does not publish it the way
sched_unlock() would, leaving the parent stranded with nothing to move it on.
The fix mirrors sched_unlock() for each case: nxsched_merge_pending(), or
nxsched_deliver_task() under CONFIG_SMP. Both are no-ops while pre-emption is
still disabled, and up_exit() re-reads this_task() afterwards, so a change of
the ready-to-run head is honoured. Without it vfork() deadlocks wherever no
other task happens to call sched_unlock() afterwards -- rv-virt:nsh64 and
rv-virt:pnsh64, where NSH is blocked in waitpid() holding the lock, and
qemu-armv8a:citest_smp, which hangs the moment the vfork() test runs.
fork() is built on a new addrenv_fork(), backed by an up_addrenv_fork() hook
that duplicates an address environment into freshly allocated pages mapped at
the same virtual addresses -- unlike up_addrenv_clone(), which copies only
the representation and leaves both pointing at the same page tables. The
child then adopts the parent's stack geometry rather than being given a
relocated copy: a pointer to a stack local taken before fork() must name the
same object in the child that it named in the parent, and the parent's stack
is already in the duplicate, with its contents, at the parent's address.
No architecture implements up_addrenv_fork() yet, so this commit leaves
fork() unavailable everywhere. That is the intended state. It withdraws
fork() from ARCH_ARM, flat ARCH_ARM64, ARCH_RISCV, ARCH_SIM and ARCH_X86_64,
where until now it named the sharing primitive; per-architecture patches
restore it, with POSIX semantics, as up_addrenv_fork() lands. In the
meantime the sharing primitive is still there under the name that describes
it: vfork() for a child that runs a program, pthread_create() for a second
flow of control that shares memory, posix_spawn() for both at once.
Kconfig: ARCH_HAVE_VFORK inherits ARCH_HAVE_FORK's select lines, conditions
included, so no configuration gains machinery; ARCH_HAVE_FORK is redefined to
mean "can provide POSIX fork() semantics" and now depends on ARCH_ADDRENV.
There is one deliberate departure from "verbatim". ARCH_ARM selected the
fork family unconditionally, BUILD_KERNEL included, and that has never
worked: on a kernel build the architecture's fork entry point sees the
kernel's return address and stack pointer rather than the caller's, so the
child resumes at a kernel address. On qemu-armv7a:knsh master faults in
ostest's fork case with "Child did not run" and then a data abort; without
the condition this change faults the same way through vfork(). ARCH_ARM64
and ARCH_X86_64 already carried "if !BUILD_KERNEL" for exactly this reason --
ARM was the outlier. Conditioning it turns a runtime fault into an honest
absence, which is the whole point of the change; arch/arm takes the condition
off again in the patch that adds its saved-syscall-frame path. Only the
MMU-capable ARM ports are affected, since Cortex-M cannot build BUILD_KERNEL
at all.
Also fixes two latent syntax errors found on the way: a missing comma in
riscv_fork.c and mips_fork.c, both in *_FRAMEPOINTER && !SAVE_GP branches
that are never compiled today.
BREAKING CHANGE: fork() is withdrawn from every architecture. It is no
longer declared in unistd.h, so code that calls it fails to build with an error
naming the function, and the sharing behaviour it used to have is gone rather
than renamed. CONFIG_ARCH_HAVE_FORK no longer means "fork() exists"; it means
"this configuration can provide POSIX fork() semantics", and no architecture
selects it yet.
Quick fix, chosen by why the call was made:
to run a program vfork() + exec*(), or better posix_spawn()
a second flow of control that pthread_create()
shares the caller's memory
a genuinely independent copy keep fork(), and wait for the per-arch patch
of the process that implements up_addrenv_fork() and selects
CONFIG_ARCH_HAVE_FORK
Out-of-tree code that tests CONFIG_ARCH_HAVE_FORK to decide whether a
fork-then-exec path is available wants CONFIG_ARCH_HAVE_VFORK instead, which is
selected in exactly the places CONFIG_ARCH_HAVE_FORK used to be. The full
migration guide is Documentation/guides/fork_vfork_migration.rst.
Assisted-by: Claude Code:claude-opus-5
Signed-off-by: Marco Casaroli <marco.casaroli@gmail.com>
2026-08-07 12:12:59 +02:00
|
|
|
#if defined(CONFIG_ARCH_ADDRENV) && defined(CONFIG_ARCH_HAVE_FORK)
|
|
|
|
|
/* A child holding an address environment of its own, rather than a
|
|
|
|
|
* reference to the caller's, is a fork() child, and nxtask_setup_fork()
|
|
|
|
|
* left that environment selected. Get back into the parent's before
|
|
|
|
|
* unwinding. See nxtask_setup_fork().
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
if (child->addrenv_own != NULL &&
|
|
|
|
|
child->addrenv_own != this_task()->addrenv_own)
|
|
|
|
|
{
|
|
|
|
|
fork_restore_parent_env();
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
|
2020-05-16 09:06:29 -06:00
|
|
|
/* The TCB was added to the active task list by nxtask_setup_scheduler() */
|
2013-01-07 19:35:47 +00:00
|
|
|
|
2024-03-19 12:33:08 +08:00
|
|
|
dq_rem((FAR dq_entry_t *)child, list_inactivetasks());
|
2013-01-07 19:35:47 +00:00
|
|
|
|
|
|
|
|
/* Release the TCB */
|
|
|
|
|
|
2025-02-13 12:12:32 +08:00
|
|
|
nxsched_release_tcb(child, child->flags & TCB_FLAG_TTYPE_MASK);
|
2013-01-07 19:35:47 +00:00
|
|
|
set_errno(errcode);
|
2013-01-15 15:40:18 +00:00
|
|
|
}
|
