feat(pi_lock): feat pi_lock and fix lock

[oeasy1412]

1. 引入 pi_lock 替代 inner_lock，对齐 Linux task_rq_lock() 锁序

原实现使用 RwLock<InnerSchedInfo> 保护进程状态、优先级、调度策略，锁粒度过大且与 Linux 语义不符。

• 状态字段原子化：移除 InnerSchedInfo（含 state + sleep），新增 state_atomic: AtomicU32，使用与 Linux 一致的位编码（TASK_RUNNING=0x0000, TASK_INTERRUPTIBLE=0x0001, TASK_UNINTERRUPTIBLE=0x0002, TASK_STOPPED=0x0004, TASK_DEAD_MARKER=0x0080，退出码存高 20 位）。
• 新增 pi_lock: SpinLock<PiProtected>：保护 cpus_allowed 和 nr_cpus_allowed，集中管理受 pi_lock 保护的字段。
• 调度策略与优先级原子化：sched_policy: RwLock<SchedPolicy> → AtomicU8；prio_data: RwLock<PrioData> → 拆分为 prio: AtomicI32、static_prio: AtomicI32、normal_prio: AtomicI32。
• 锁序规范：pi_lock → rq_lock（对齐 Linux task_rq_lock()），释放时先 rq_lock 后 pi_lock。

2. 修复 rwlock 抢占/中断顺序

原 read_irqsave/write_irqsave/upgradeable_read_irqsave 在自旋循环内部每次迭代都重新关中断，中断状态不一致。

• 在循环外先 save_and_disable_irq() + preempt_disable()，循环内仅尝试获取锁。
• 修复 Guard 转换（downgrade_to_read）：使用 mem::forget(self) 跳过原 Guard 的 preempt_enable，由新 Guard 接管。
• 修正 Drop 顺序：所有 Guard 先恢复中断（irq_guard.take()），再启用抢占（preempt_enable()）。

3. 修复 spinlock 解锁顺序，对齐 Linux spin_unlock_irqrestore / spin_unlock_bh

• 移除 SpinLock::unlock() 中的 preempt_enable()，移至 SpinLockGuard::drop() 中，确保顺序为：先解锁 → 恢复中断 → 启用抢占。
• 调整 SpinLockBhGuard 字段顺序为 guard 在前、bh 在后，确保 Rust Drop 顺序为：先解锁 → 恢复 BH（对齐 spin_unlock_bh）。

4. 重构 __schedule() 对齐 Linux 信号检查语义

• 移除 is_mark_sleep 判断（原 InnerSchedInfo.sleep 标志）。
• 对标 smp_mb__after_spinlock() 语义：获取 rq_lock 后插入 fence(Ordering::SeqCst)。
• 调整 计数位置：nr_uninterruptible++ 在 deactivate_task 之前；nr_iowait++ 在 deactivate_task 之后。
• 调整 activate_task/deactivate_task：移除其中的 nr_iowait/nr_uninterruptible/IDLE_CPUS 维护，移至调用方。

5. 修复 sched_fork() 子进程策略/优先级继承

• 子进程继承父进程的 normal_prio、static_prio、policy
• 使用原子操作替代读写锁。

6. 修复 copy_flags() 子进程标志位继承

子进程不再继承父进程运行时状态标志（NEED_SCHEDULE、EXITING、WAKEKILL、SIGNALED、NEED_MIGRATE 等），仅继承 RANDOMIZE（ASLR 配置）。KTHREAD 标志由创建参数显式设置。

7. 函数重命名（修复拼写）

• check_preempt_currnet → check_preempt_current
• wakeup_new_task → wake_up_new_task

8. Nix 构建系统：新增 GDB 调试支持

• QEMU 动态分配 GDB 端口，支持 QEMU_GDB_WAIT=1 暂停等待连接。
• 新增 gdb-${target} app/package，使用 rust-gdb（x86_64）或 gdb-multiarch（其他架构）。
• 新增 dunitest 测试框架集成。

[fslongjin]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 19d239331a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

在远端 stop 当前任务前先从 rq 摘除或走阻塞调度路径

stop_task() 对“目标任务正运行在远端 CPU”的分支只发送 kick_cpu()，并依赖 __schedule() 中后续 deactivate（见注释“在那里 deactivate”）；但中断路径进入的是 __schedule(SM_PREEMPT)，而该路径不会执行非可运行任务的 deactivate 逻辑（deactivate 仅在 !SM_MASK_PREEMPT 分支）。结果是被置为 Stopped 的当前任务仍可能留在 runqueue 上并再次被选中运行，导致 stop 语义失效（典型场景：对正在远端运行的线程发送作业控制 stop 信号）。

Useful? React with 👍 / 👎.

[oeasy1412]

1. Introduce pi_lock to replace inner_lock, aligning with the Linux task_rq_lock() lock order

The original implementation used RwLock<InnerSchedInfo> to protect process state, priority, and scheduling policy, resulting in overly coarse locking and inconsistency with Linux semantics.

• Atomic state field: Remove InnerSchedInfo (which contained state + sleep) and add state_atomic: AtomicU32, using the same bit encoding as Linux (TASK_RUNNING=0x0000, TASK_INTERRUPTIBLE=0x0001, TASK_UNINTERRUPTIBLE=0x0002, TASK_STOPPED=0x0004, TASK_DEAD_MARKER=0x0080, with the exit code stored in the upper 20 bits).
• New pi_lock: SpinLock<PiProtected>: Protects cpus_allowed and nr_cpus_allowed, centralizing fields guarded by pi_lock.
• Atomic scheduling policy and priority: sched_policy: RwLock<SchedPolicy> → AtomicU8; prio_data: RwLock<PrioData> → split into prio: AtomicI32, static_prio: AtomicI32, normal_prio: AtomicI32.
• Lock ordering discipline: pi_lock → rq_lock (aligning with Linux task_rq_lock()). On release, rq_lock is dropped before pi_lock.

2. Fix rwlock preemption/interrupt ordering

The original read_irqsave/write_irqsave/upgradeable_read_irqsave re-disabled interrupts on every iteration of the spin loop, leading to inconsistent interrupt states.

• Perform save_and_disable_irq() + preempt_disable() outside the loop, and only attempt to acquire the lock inside the loop.
• Fix guard conversion (downgrade_to_read): Use mem::forget(self) to skip the original guard’s preempt_enable, letting the new guard take over.
• Correct Drop order: All guards restore interrupts first (irq_guard.take()) and then re-enable preemption (preempt_enable()).

3. Fix spinlock unlock order, aligning with Linux spin_unlock_irqrestore / spin_unlock_bh

• Remove preempt_enable() from SpinLock::unlock() and move it to SpinLockGuard::drop() to guarantee the order: unlock → restore interrupts → enable preemption.
• Reorder fields in SpinLockBhGuard so that guard comes before bh, ensuring Rust’s Drop order is: unlock → restore BH (aligning with spin_unlock_bh).

4. Refactor __schedule() to align with Linux signal-check semantics

• Remove the is_mark_sleep check (the old InnerSchedInfo.sleep flag).
• Match smp_mb__after_spinlock() semantics: Insert fence(Ordering::SeqCst) after acquiring rq_lock.
• Reposition counters: nr_uninterruptible++ before deactivate_task; nr_iowait++ after deactivate_task.
• Adjust activate_task/deactivate_task: Remove the nr_iowait/nr_uninterruptible/IDLE_CPUS bookkeeping from these functions and move it to the call sites.

5. Fix child policy/priority inheritance in sched_fork()

• The child process inherits the parent’s normal_prio, static_prio, and policy.
• Use atomic operations instead of reader-writer locks.

6. Fix child flag inheritance in copy_flags()

The child process no longer inherits the parent’s run-time state flags (such as NEED_SCHEDULE, EXITING, WAKEKILL, SIGNALED, NEED_MIGRATE), and only inherits RANDOMIZE (ASLR configuration). The KTHREAD flag is set explicitly through the creation parameters.

7. Rename functions (fix spelling)

• check_preempt_currnet → check_preempt_current
• wakeup_new_task → wake_up_new_task

8. Nix build system: Add GDB debugging support

• QEMU dynamically allocates a GDB port, with support for QEMU_GDB_WAIT=1 to pause and wait for a connection.
• Add gdb-${target} app/package, using rust-gdb (x86_64) or gdb-multiarch (other architectures).
• Add integration with the dunitest test framework.

[fslongjin]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

DragonOS/kernel/src/process/mod.rs

Lines 607 to 614 in b4b51e9

	} else if !update_clock {
	// 正在远端 CPU 运行：dequeue + kick
	// 此处异步 dequeue，远端 CPU 收到 kick 后 __schedule(SM_PREEMPT) 时
	// stopped task 已不在红黑树中，pick_next_task 不会选中它
	rq.deactivate_task(
	pcb.clone(),
	DequeueFlag::DEQUEUE_STOPPED | DequeueFlag::DEQUEUE_NOCLOCK,
	);

避免远端 current 被重复出队

当 SIGSTOP/作业控制停止命中正在其他 CPU 上运行的任务时，这里先把远端 rq.current() 从运行队列出队并发送 kick；但远端 CPU 随后进入 __schedule(SM_NONE) 时会看到同一个任务仍是 Stopped，并在 kernel/src/sched/mod.rs 的非抢占路径再次无条件调用 rq.deactivate_task(prev, DEQUEUE_SLEEP | ...)。这会对同一个调度实体执行两次 dequeue，导致 nr_running/CFS 树等运行队列状态下溢或损坏；远端 current 应只标记需要调度并 kick，让它在自己的 __schedule() 路径完成唯一一次出队。

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[fslongjin]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9694d3eb21

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[fslongjin]

补充一个非阻塞的构建配置 review 点：本 PR 把 flake.nix 里的 qemuConfig.debug 默认从 false 改成了 true，这会让普通 nix run .#start-* / yolo 启动也默认打开 QEMU GDB TCP stub。新增 gdb-${target} 调试能力本身是合理的，但默认启用调试端口会改变普通运行、CI 和安全边界。

建议保持默认 debug = false，把 GDB stub 做成显式 opt-in（例如单独 debug app、参数或环境配置），或者拆成单独 PR 讨论。这个点和本轮调度/flags 根因修复解耦，不需要混在内核修复里。