Add support for XSAVE/XRSTOR to handle x86_64 procesor extended states.
Support for these instructions is required to support AVX instruction set
Signed-off-by: p-szafonimateusz <p-szafonimateusz@xiaomi.com>
purpose:
1 sched_lock is very time-consuming, and reducing its invocations can improve performance.
2 sched_lock is prone to misuse, and narrowing its scope of use is to prevent people from referencing incorrect code and using it
test:
We can use qemu for testing.
compiling
make distclean -j20; ./tools/configure.sh -l qemu-armv8a:nsh_smp ;make -j20
running
qemu-system-aarch64 -cpu cortex-a53 -smp 4 -nographic -machine virt,virtualization=on,gic-version=3 -net none -chardev stdio,id=con,mux=on -serial chardev:con -mon chardev=con,mode=readline -kernel ./nuttx
We have also tested this patch on other ARM hardware platforms.
Signed-off-by: hujun5 <hujun5@xiaomi.com>
Kthreads can share the group data so that to reduce overheads.
This implements shared kthread group via:
- use `tcb_s` instead of `task_tcb_s` for kthreads
- use `g_kthread_group` when creating kthreads
- use stackargs to start tasks and kthreads
see pull/12320 for test logs.
Signed-off-by: Yanfeng Liu <yfliu2008@qq.com>
Thread args have already been saved to stack after the TLS section by
nxtask_setup_stackargs. This is to retrieve it for use.
Signed-off-by: Yanfeng Liu <yfliu2008@qq.com>
This adds pre-allocation and dynamic allocations for sigactions.
Current behavior can be acheived by setting SIG_ALLOC_ACTIONS to
a number larger than 1.
Signed-off-by: Yanfeng Liu <yfliu2008@qq.com>
this_task() is a function call that involves disabling interrupts and this_cpu().
Since restore_critical_section is always called in an interrupt-disabled context,
there's no need to disable interrupts again. Therefore, to save time and achieve
the same effect, I directly call tcb = current_task(me) instead of tcb = this_task().
Signed-off-by: hujun5 <hujun5@xiaomi.com>
we can use g_cpu_lockset to determine whether we are currently in the scheduling lock,
and all accesses and modifications to g_cpu_lockset, g_cpu_irqlock, g_cpu_irqset
are in the critical section, so we can directly operate on it.
test:
We can use qemu for testing.
compiling
make distclean -j20; ./tools/configure.sh -l qemu-armv8a:nsh_smp ;make -j20
running
qemu-system-aarch64 -cpu cortex-a53 -smp 4 -nographic -machine virt,virtualization=on,gic-version=3 -net none -chardev stdio,id=con,mux=on -serial chardev:con -mon chardev=con,mode=readline -kernel ./nuttx
Signed-off-by: hujun5 <hujun5@xiaomi.com>
sched implementation not depends on macro abstraction, so revert below commit:
This reverts commit 4e62d0005a
This reverts commit 0f0c370520
This reverts commit ad0efd04ee
Signed-off-by: chao an <anchao@lixiang.com>
waitpid() cannot be used in kernel mode unless SCHED_HAVE_PARENT is
selected -> add dependency if BUILD_KERNEL is selected.
Why? Because without SCHED_HAVE_PARENT waitpid() works in a non-standard
way, meaning it does not use SIGCHLD to wake the parent, as it should.
Also, returning the child status via stat_loc corrupts memory as stat_loc
points to the parent's address environment:
pid_t nxsched_waitpid(pid_t pid, int *stat_loc, int options)
{
...
group->tg_statloc = stat_loc;
...
}
And later when the status is returned, the child writes to tg_statloc,
which points to the parent's address environment -> random memory
corruption:
static inline void nxtask_exitwakeup(FAR struct tcb_s *tcb, int status)
{
...
if (group->tg_statloc != NULL)
{
*group->tg_statloc = status << 8;
}
...
}
This log appears for NSH internal commands in KERNEL build and
duplicates a lot with task_spawn: log. So drop it to make the
DEBUG_SCHED_INFO logging more readable.
Signed-off-by: Yanfeng Liu <yfliu2008@qq.com>
we removed "select SCHED_RESUMESCHEDULER",
As SCHED_RESUMESCHEDULER is not a necessary feature in SMP,
turning it on by default may affect performance.
Signed-off-by: hujun5 <hujun5@xiaomi.com>
test:
We can use qemu for testing.
compiling
make distclean -j20; ./tools/configure.sh -l qemu-armv8a:nsh_smp ;make -j20 running
qemu-system-aarch64 -cpu cortex-a53 -smp 4 -nographic -machine virt,virtualization=on,gic-version=3 -net none -chardev stdio,id=con,mux=on -serial chardev:con -mon chardev=con,mode=readline -kernel ./nuttx
We have also tested this patch on other ARM hardware platforms.
Signed-off-by: hujun5 <hujun5@xiaomi.com>
purpose:
1 sched_lock is very time-consuming, and reducing its invocations can improve performance.
2 sched_lock is prone to misuse, and narrowing its scope of use is to prevent people from referencing incorrect code and using it
test:
We can use qemu for testing.
compiling
make distclean -j20; ./tools/configure.sh -l qemu-armv8a:nsh_smp ;make -j20
running
qemu-system-aarch64 -cpu cortex-a53 -smp 4 -nographic -machine virt,virtualization=on,gic-version=3 -net none -chardev stdio,id=con,mux=on -serial chardev:con -mon chardev=con,mode=readline -kernel ./nuttx
We have also tested this patch on other ARM hardware platforms.
Signed-off-by: hujun5 <hujun5@xiaomi.com>
purpose:
1 sched_lock is very time-consuming, and reducing its invocations can improve performance.
2 sched_lock is prone to misuse, and narrowing its scope of use is to prevent people from referencing incorrect code and using it
test:
We can use qemu for testing.
compiling
make distclean -j20; ./tools/configure.sh -l qemu-armv8a:nsh_smp ;make -j20
running
qemu-system-aarch64 -cpu cortex-a53 -smp 4 -nographic -machine virt,virtualization=on,gic-version=3 -net none -chardev stdio,id=con,mux=on -serial chardev:con -mon chardev=con,mode=readline -kernel ./nuttx
We have also tested this patch on other ARM hardware platforms.
Signed-off-by: hujun5 <hujun5@xiaomi.com>
1. The critical section does not prevent task scheduling
2. If the critical section is in sched_lock, there is no need to check,
scheduling is not going to happen
3. If sched_lock is in the critical section, sched_unlock will also
trigger scheduling without waiting for the exit of the critical section
4. After exiting the critical section, if there is an interrupt,
the scheduling will be automatically triggered
Signed-off-by: hujun5 <hujun5@xiaomi.com>
In order to ensure the detached thread obtain the correct return
value from pthread_join()/pthread_cancel(), the detached thread
will create joininfo to save the detached status after thread
destroyed. If there are too many of detached threads in the
process group, the joininfo will consume too much memory.
This is not friendly to embedded MCU devices.
This commit keep the semantics as #11898 was introduced,
will no longer save joininfo for detached threads to avoid wasting memory.
Signed-off-by: chao an <anchao@lixiang.com>
Prior to this commit, doswitch is returned uninitialized if
(task_state == TSTATE_TASK_ASSIGNED || task_state == TSTATE_TASK_RUNNING)
&& no context switch occurs
&& (cpu == me).
Signed-off-by: Mingjie Shen <shen497@purdue.edu>
In SMP mode, up_cpu_index()/this_cpu() are the same, both return the index of the physical core.
In AMP mode, up_cpu_index() will return the index of the physical core, and this_cpu() will always return 0
| #ifdef CONFIG_SMP
| # define this_cpu() up_cpu_index()
| #elif defined(CONFIG_AMP)
| # define this_cpu() (0)
| #else
| # define this_cpu() (0)
| #endif
Signed-off-by: chao an <anchao@lixiang.com>
The overhead of spinlok is less than mutext (mutex need to call
enter_critical section.)
After this patch, `down_write_trylock` and `down_read_trylock` can be
use in interrupt context.
The instruction is protected with mutex only one instruction so using
spinlock is better.
Signed-off-by: TaiJu Wu <tjwu1217@gmail.com>
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>
The delete flag is not synchronized with the life cycle of the group,
if the flag set before waitpid(), the tcb will be mistakenly deleted
by group_del_waiter(), use-after-free will happen.
Regression by:
| commit 29e50ffa73 (origin/master, origin/HEAD)
| Author: chao an <anchao@lixiang.com>
| Date: Mon Mar 4 09:19:27 2024 +0800
|
| sched/group: move task group into task_tcb_s to improve performance
|
| move task group into task_tcb_s to avoid access allocator to improve performance
|
| for Task Termination, the time consumption will be reduced ~2us (Tricore TC397 300MHZ):
| 15.97(us) -> 13.55(us)
|
| Signed-off-by: chao an <anchao@lixiang.com>
Signed-off-by: chao an <anchao@lixiang.com>
move task group into task_tcb_s to avoid access allocator to improve performance
for Task Termination, the time consumption will be reduced ~2us (Tricore TC397 300MHZ):
15.97(us) -> 13.55(us)
Signed-off-by: chao an <anchao@lixiang.com>
Skip the child wait if here is only self in member list,
since the members of the task group should be 1 if task exit.
Fix Regression issue that the task could not terminate normally.
Signed-off-by: chao an <anchao@lixiang.com>
Change the type of task group member to single list chain to
avoid accessing the memory allocator to improve the performance
Signed-off-by: chao an <anchao@lixiang.com>
The maximum startup parameters have been checked accordingly in nxtask_setup_stackargs(),
let us save argument counter to avoid limit check.
Signed-off-by: chao an <anchao@lixiang.com>
Task group could find from process id, replace group_findbypid to
task_getgroup to simplify the search logic
Signed-off-by: chao an <anchao@lixiang.com>
Current `CONFIG_PAGING` refers to an experimental implementation
to enable embedded MCUs with some limited RAM space to execute
large programs from some non-random access media.
On-demand paging should be implemented for the kernel mode with
address environment implementation enabled.
Add support for static tcb, applications in some special case can
initialize system resources in advance through static tcb.
| static struct task_tcb_s g_tcb;
|
| memset(&g_tcb, 0, sizeof(struct task_tcb_s));
| g_tcb.cmn.flags = TCB_FLAG_TTYPE_KERNEL;
| nxtask_init(&g_tcb, "PTCB", 101, NULL, 1024, ptcb_task, NULL, NULL, NULL);
|
| ...
| nxtask_activate(&g_tcb.cmn);
Signed-off-by: chao an <anchao@lixiang.com>
Recursive:
25 0x44000e5a in up_cpu_paused (cpu=cpu@entry=1) at armv7-a/arm_cpupause.c:120
26 0x440032f2 in enter_critical_section () at irq/irq_csection.c:275
27 0x44006f24 in nxsched_process_taskload_ticks (tcb=tcb@entry=0x442ba638 <g_idletcb+256>, ticks=0) at sched/sched_cpuload.c:104
28 0x44007310 in nxsched_suspend_critmon (tcb=tcb@entry=0x442ba638 <g_idletcb+256>) at sched/sched_critmonitor.c:303
29 0x44006ef0 in nxsched_suspend_scheduler (tcb=tcb@entry=0x442ba638 <g_idletcb+256>) at sched/sched_suspendscheduler.c:78
30 0x44000e5a in up_cpu_paused (cpu=cpu@entry=1) at armv7-a/arm_cpupause.c:120
31 0x440032f2 in enter_critical_section () at irq/irq_csection.c:275
32 0x44000f1a in arm_pause_handler (irq=<optimized out>, context=<optimized out>, arg=<optimized out>) at armv7-a/arm_cpupause.c:216
33 0x44002ffe in irq_dispatch (irq=irq@entry=2, context=context@entry=0x4449f6b8 <g_cpu1_idlestack+1720>) at irq/irq_dispatch.c:146
34 0x44001612 in arm_doirq (irq=2, irq@entry=0, regs=0x4449f6b8 <g_cpu1_idlestack+1720>) at armv7-a/arm_doirq.c:72
35 0x44000940 in arm_decodeirq (regs=<optimized out>) at armv7-a/arm_gicv2.c:403
36 0x440000b4 in arm_vectorirq () at armv7-a/arm_vectors.S:236
Signed-off-by: ligd <liguiding1@xiaomi.com>
cpu0 thread0: cpu1:
sched_yield()
nxsched_set_priority()
nxsched_running_setpriority()
nxsched_reprioritize_rtr()
nxsched_add_readytorun()
up_cpu_pause()
IRQ enter
arm64_pause_handler()
enter_critical_section() begin
up_cpu_paused() pick thread0
arm64_restorestate() set thread0 tcb->xcp.regs to CURRENT_REGS
up_switch_context()
thread0 -> thread1
arm64_syscall()
case SYS_switch_context
change thread0 tcb->xcp.regs
restore_critical_section()
enter_critical_section() done
leave_critical_section()
IRQ leave with restore CURRENT_REGS
ERROR !!!
Reason:
As descript above, cpu0 swith task: thread0 -> thread1, and the
syscall() execute slowly, this time cpu1 pick thread0 to run at
up_cpu_paused(). Then cpu0 syscall execute, cpu1 IRQ leave error.
Resolve:
Move arm64_restorestate() after enter_critical_section() done
This is a continued fix with:
https://github.com/apache/nuttx/pull/6833
Signed-off-by: ligd <liguiding1@xiaomi.com>
These variables will trigger variable 'ret' set but not used warnings due to different configurations.
Signed-off-by: yinshengkai <yinshengkai@xiaomi.com>
Task activation/exit logs are helpful for device bringup,
especially when user space nsh prompt doesn't show up.
Putting them here will allow cleaning of logs in multiple
up_exit() functions later.
Signed-off-by: Yanfeng Liu <yfliu2008@qq.com>
Usually the startup script is placed under /etc. The contents of the etc directory
are compiled and linked with Nuttx binary in the form of romfs. After startup,
it will be mounted by Nsh.
etc is generated by the different boards, that use genromfs and xxd tools to generate
and compile it into the Nuttx, for example: boards/arm/at32/at32f437-mini/tool/mkromfs.sh
The more common method is etc image generated from the content in the corresponding
board/arch/board/board/src/etc directory, and added by Makefile for example:
boards/sim/sim/sim/src/etc.
But in kernel/protected mode, Nuttx kernel and apps are run in different privileged/
non-privileged mode or the isolated binarys, so as that nsh should use syscall to
access Nuttx kernel by exported API. In this scenario, nsh can not mount the etc image
content, because that is generated in board and as a part of Nuttx kernel.
changes:
- move etc romfs mount from nsh to Nuttx, but keep the script to parse and execute.
- move and rename the related CONFIG, move customized nsh_romfsimg.h to etc_romfs.c
in boards, and no need declaration for romfs_img/romfs_img_len.
This commit changes and updates all configurations in Nuttx arch/board as much as possible,
but if any missing, please refer to the following simple guide:
- rename CONFIG_NSH_ROMFSETC to CONFIG_ETC_ROMFS, and delete CONFIG_NSH_ARCHROMFS in defconfig
- rename the etc romfs mount configs, for example CONFIG_NSH_FATDEVNO to CONFIG_ETC_FATDEVNO
- move customized nsh_romfsimg.h to etc_romfs.c in board/arch/board/board/src and no need
declaration for romfs_img/romfs_img_len.
- delete default nsh_romfsimg.h, if ROMFSETC is enabled, should generate and compile etc_romfs.c
in board/arch/board/board/src.
Signed-off-by: fangxinyong <fangxinyong@xiaomi.com>
Remove KEEP_ALIVE_HACK to avoid waking up the device periodly.
The workaround it's added by:
commit 6546fa39c7
Author: Gregory Nutt <gnutt@nuttx.org>
Date: Tue Aug 12 11:12:00 2014 -0600
Tickless Stuff: Back out the risky timer operations when the ready-to-run
list is modified. That is unsafe. An ugly workaround is just to keep an
interval timer going all of the time with a minimum duration equal to the
timeslice interval.
But look like it doesn't needed anymore.
Signed-off-by: Xiang Xiao <xiaoxiang@xiaomi.com>
Usage:
1. CONFIG_FS_PROCFS_MAX_STACK_RECORD > 0, such as 32,
2. add '-finstrument-functions' to CFLAGS for What you want to check
stack.
3. mount porcfs
4. cat /proc/<pid>/stack will print backtace & size
Signed-off-by: anjiahao <anjiahao@xiaomi.com>
If the gap between sp and stack_top is too small,
then the stack will not be output,
modify the conditional loop condition, and fix this problem
Signed-off-by: anjiahao <anjiahao@xiaomi.com>
A new locking mechanism: read/write locks
When there is a writer it is not possible to put on a read lock or a write lock; when there is a reader it is possible to reenter the read lock but not the write lock.
Writers are exclusive locks, readers are shared locks.
At the same time through the waiter count to determine whether there is currently a blocked task, if there is then in the unlock time to wake up all the waiter, through the priority of the competition to complete the blocked lock execution.
For example:
When we have a reader blocking two waiter writers, when the reader is unlocked it wakes up both writers. The writer with higher priority wakes up and checks for a successful condition and locks the lock, the second writer wakes up and fails to check for a condition and continues to block the lock.
Signed-off-by: chenrun1 <chenrun1@xiaomi.com>
This moves all the public POSIX semaphore functions into libc and with
this most of the user-space logic is also moved; namely cancel point and
errno handling.
This also removes the need for the _SEM_XX macros used to differentiate
which API is used per user-/kernel mode. Such macros are henceforth
unnecessary.
PR #11165 causes an unnecessary regression; task_delete no longer works,
if the deleted task is from another group.
The logic that prevents this comes from:
nxnotify_cancellation() ->
tls_get_info_pid() ->
nxsched_get_stackinfo()
Which checks for permissions, which does not make sense in this case since
it is the kernel asking for the stack information.
Fix this by partially reverting 11165 and implementing a direct path for
the kernel to query for any tasks TLS.
Commit 9244b5a737 added support
for non-standard field si_user that is useful for passing context
pointers to signal handlers.
This commits makes it work for all signals, not just SA_KERNELHAND.
Previously si_user for normal signals was uninitialized garbage.
The task files should consult the "spawn action" and "O_CLOEXEC flags"
to determine further whether the file should be duplicated.
This PR will further optimize file list duplicating to avoid the performance
regression caused by additional file operations.
Signed-off-by: chao an <anchao@xiaomi.com>
This moves task / thread cancel point logic from the NuttX kernel into
libc, while the data needed by the cancel point logic is moved to TLS.
The change is an enabler to move user-space APIs to libc as well, for
a coherent user/kernel separation.
Some assertions in extreme cases will cause syslog to be unable to
output logs normally, so this PR will restore the input registers
into the array of last registers to ensure that we can also obtain
some important informations.
Signed-off-by: chao an <anchao@xiaomi.com>
