From: Minep <lunaixsky@qq.com>
Date: Sun, 23 Jul 2023 10:39:16 +0000 (+0100)
Subject: refactor: Optimize the context switch overhead
X-Git-Url: https://scm.lunaixsky.com/lunaix-os.git/commitdiff_plain/946c3fdd31300074cc78841795bd47af908ddddb

refactor: Optimize the context switch overhead
---

diff --git a/lunaix-os/includes/arch/x86/i386_abi.h b/lunaix-os/includes/arch/x86/i386_abi.h
index 8804bc9..8bc8cd2 100644
--- a/lunaix-os/includes/arch/x86/i386_abi.h
+++ b/lunaix-os/includes/arch/x86/i386_abi.h
@@ -1,9 +1,9 @@
 #ifndef __LUNAIX_I386ABI_H
 #define __LUNAIX_I386ABI_H
 
-#define store_retval(retval) __current->intr_ctx.registers.eax = (retval)
+#define store_retval(retval) __current->intr_ctx->registers.eax = (retval)
 
-#define store_retval_to(proc, retval) (proc)->intr_ctx.registers.eax = (retval)
+#define store_retval_to(proc, retval) (proc)->intr_ctx->registers.eax = (retval)
 
 #define j_usr(sp, pc)                                                          \
     asm volatile("movw %0, %%ax\n"                                             \
diff --git a/lunaix-os/includes/arch/x86/interrupts.h b/lunaix-os/includes/arch/x86/interrupts.h
index 6d3c4b1..0d7c815 100644
--- a/lunaix-os/includes/arch/x86/interrupts.h
+++ b/lunaix-os/includes/arch/x86/interrupts.h
@@ -8,24 +8,25 @@
 
 struct exec_param;
 
+struct regcontext
+{
+    reg32 eax;
+    reg32 ebx;
+    reg32 ecx;
+    reg32 edx;
+    reg32 edi;
+    reg32 ebp;
+    reg32 esi;
+    reg32 ds;
+    reg32 es;
+    reg32 fs;
+    reg32 gs;
+} __attribute__((packed));
+
 typedef struct
 {
     unsigned int depth;
-    struct
-    {
-        reg32 eax;
-        reg32 ebx;
-        reg32 ecx;
-        reg32 edx;
-        reg32 edi;
-        reg32 ebp;
-        reg32 esi;
-        reg32 ds;
-        reg32 es;
-        reg32 fs;
-        reg32 gs;
-    } __attribute__((packed)) registers;
-
+    struct regcontext registers;
     union
     {
         reg32 esp;
@@ -35,7 +36,7 @@ typedef struct
 
 struct exec_param
 {
-    isr_param saved_prev_ctx;
+    isr_param* saved_prev_ctx;
     u32_t vector;
     u32_t err_code;
     u32_t eip;
diff --git a/lunaix-os/includes/arch/x86/tss.h b/lunaix-os/includes/arch/x86/tss.h
index c2a2535..931ff26 100644
--- a/lunaix-os/includes/arch/x86/tss.h
+++ b/lunaix-os/includes/arch/x86/tss.h
@@ -1,7 +1,10 @@
 #ifndef __LUNAIX_TSS_H
 #define __LUNAIX_TSS_H
-#include <lunaix/types.h>
 
+#define tss_esp0_off 4
+
+#ifndef __ASM__
+#include <lunaix/types.h>
 struct x86_tss
 {
     u32_t link;
@@ -12,5 +15,6 @@ struct x86_tss
 
 void
 tss_update_esp(u32_t esp0);
+#endif
 
 #endif /* __LUNAIX_TSS_H */
diff --git a/lunaix-os/includes/hal/cpu.h b/lunaix-os/includes/hal/cpu.h
index 694d678..3b77c7e 100644
--- a/lunaix-os/includes/hal/cpu.h
+++ b/lunaix-os/includes/hal/cpu.h
@@ -137,4 +137,10 @@ cpu_rdmsr(u32_t msr_idx, u32_t* reg_high, u32_t* reg_low);
 void
 cpu_wrmsr(u32_t msr_idx, u32_t reg_high, u32_t reg_low);
 
+static inline void
+cpu_ldvmspace(ptr_t vms)
+{
+    cpu_lcr3(vms);
+}
+
 #endif
\ No newline at end of file
diff --git a/lunaix-os/includes/lunaix/process.h b/lunaix-os/includes/lunaix/process.h
index a8ac5d8..bb55f02 100644
--- a/lunaix-os/includes/lunaix/process.h
+++ b/lunaix-os/includes/lunaix/process.h
@@ -90,10 +90,10 @@ struct proc_info
 
     pid_t pid;                // offset = 0
     struct proc_info* parent; // offset = 4
-    isr_param intr_ctx;       // offset = 8
-    ptr_t ustack_top;         // offset = 84 -> 56 -> 60
-    ptr_t page_table;         // offset = 88 -> 60 -> 64
-    void* fxstate;            // offset = 92 -> 64 -> 68
+    isr_param* intr_ctx;      // offset = 8
+    ptr_t ustack_top;         // offset = 84 -> 56 -> 60 -> 12
+    ptr_t page_table;         // offset = 88 -> 60 -> 64 -> 16
+    void* fxstate;            // offset = 92 -> 64 -> 68 -> 20
 
     /* ---- critical section end ---- */
 
@@ -170,7 +170,7 @@ pid_t
 destroy_process(pid_t pid);
 
 void
-setup_proc_mem(struct proc_info* proc, ptr_t kstack_from);
+copy_kernel_stack(struct proc_info* proc, ptr_t kstack_from);
 
 /**
  * @brief 复制当前进程（LunaixOS的类 fork (unix) 实现）
diff --git a/lunaix-os/kernel/asm/x86/interrupt.S b/lunaix-os/kernel/asm/x86/interrupt.S
index 602f793..3007d45 100644
--- a/lunaix-os/kernel/asm/x86/interrupt.S
+++ b/lunaix-os/kernel/asm/x86/interrupt.S
@@ -1,6 +1,7 @@
 #define __ASM__
 #include <arch/x86/interrupts.h>
 #include <arch/x86/i386_asm.h>
+#include <arch/x86/tss.h>
 #include <lunaix/syscall.h>
 #define __ASM_INTR_DIAGNOSIS
 
@@ -29,25 +30,104 @@
 
 */
 
+#define regsize 4
+
+/* stack layout: saved interrupt context */
+    .struct 0
+idepth:
+    .struct idepth + regsize
+ieax:
+    .struct ieax + regsize
+iebx:
+    .struct iebx + regsize
+iecx:
+    .struct iecx + regsize
+iedx:
+    .struct iedx + regsize
+iedi:
+    .struct iedi + regsize
+iebp:
+    .struct iebp + regsize
+iesi:
+    .struct iesi + regsize
+ids:
+    .struct ids + regsize
+ies:
+    .struct ies + regsize
+ifs:
+    .struct ifs + regsize
+igs:
+    .struct igs + regsize
+iesp:
+    .struct iesp + regsize
+isave_prev:
+    .struct isave_prev + regsize
+ivec:
+    .struct ivec + regsize
+iecode:
+    .struct iecode + regsize
+ieip:
+    .struct ieip + regsize
+ics:
+    .struct ics + regsize
+ieflags:
+    .struct ieflags + regsize
+iuesp:
+    .struct iuesp + regsize
+iuss:
+
+
+/* stack layout: execution (flow-control) state context */
+    .struct 0
+exsave_prev:
+    .struct exsave_prev + regsize
+exvec:
+    .struct exvec + regsize
+execode:
+    .struct execode + regsize
+exeip:
+    .struct exeip + regsize
+excs:
+    .struct excs + regsize
+exeflags:
+    .struct exeflags + regsize
+exuesp:
+    .struct exuesp + regsize
+exuss:
+
+/* struct layout: critical section of struct proc_info */
+    .struct 0
+proc_pid:
+    .struct proc_pid + regsize
+proc_parent:
+    .struct proc_parent + regsize
+proc_intr_ctx:
+    .struct proc_intr_ctx + regsize
+proc_ustack_top:
+    .struct proc_ustack_top + regsize
+proc_page_table:
+    .struct proc_page_table + regsize
+proc_fxstate:
+
 .section .text
     .global interrupt_wrapper
     interrupt_wrapper:
         /*
          Stack layout (layout of struct isr_param)
-    msa:   [ss]             > 76
-           [esp]            > 72
-           eflags           > 68
-           cs               > 64
-           eip              > 60
-           err_code         > 56
-           vector           > offset = 52
+    msa:   [ss]             > 76 -> 28
+           [esp]            > 72 -> 24
+           eflags           > 68 -> 20
+           cs               > 64 -> 16
+           eip              > 60 -> 12
+           err_code         > 56 -> 8
+           vector           > offset = 52 -> 4
            [saved_prev_ctx] > offset = 0
            ---
-           esp
+           esp       > 12 * 4 = 48
            gs
            fs
            es
-           ds         > offset = 7 * 4 = 28 + 4
+           ds         > offset = 8 * 4 = 32
            esi
            ebp
            edi
@@ -62,7 +142,7 @@
         */
         cld
 
-        subl $52, %esp
+        subl $4, %esp
         pushl %esp
 
         subl $16, %esp
@@ -80,11 +160,11 @@
         pushl %eax
 
         movl __current, %eax
-        movl 8(%eax), %eax
+        movl proc_intr_ctx(%eax), %eax
         incl %eax
         pushl %eax          # nested intr: current depth
 
-        movl 116(%esp), %eax   /* 取出 %cs */
+        movl ics(%esp), %eax   /* 取出 %cs */
         andl $0x3, %eax          /* 判断 RPL */
         jz 1f
 
@@ -98,12 +178,12 @@
 
         # FIXME: Save x87 context to user stack, rather than kernel's memory.
         # 保存x87FPU的状态
-        movl 68(%eax), %ebx
+        movl proc_fxstate(%eax), %ebx
         fxsave (%ebx)
 
         # 保存用户栈顶指针。因为我们允许同级中断的产生，所以需要该手段跟踪用户栈的地址。
-        movl 124(%esp), %ebx     # 取出esp
-        movl %ebx, 60(%eax)     # 存入__current->ustack_top
+        movl iuesp(%esp), %ebx     # 取出esp
+        movl %ebx, proc_ustack_top(%eax)     # 存入__current->ustack_top
 
     1:
         movl %esp, %eax
@@ -121,16 +201,17 @@
 
 #ifdef __ASM_INTR_DIAGNOSIS
         movl %eax, (debug_resv + 8)
-        movl 48(%esp), %eax
-        movl 60(%eax), %eax
+        movl iesp(%esp), %eax
+        movl exeip(%eax), %eax
         movl %eax, (debug_resv + 4) # eip
 #endif
         movl __current, %eax
-        movl 68(%eax), %eax
+        movl proc_fxstate(%eax), %eax
         
         test %eax, %eax     # do we have stored x87 context?
         jz 1f 
-        fxrstor (%eax)      
+        fxrstor (%eax) 
+
 1:
         popl %eax   # discard isr_param::depth
         popl %eax
@@ -150,20 +231,10 @@
 
         movl %eax, tmp_store
         movl __current, %eax
+        
+
         # nested intr: restore saved context
-        popl 8(%eax)       # depth
-        popl 12(%eax)      # eax
-        popl 16(%eax)      # ebx
-        popl 20(%eax)      # ecx
-        popl 24(%eax)      # edx
-        popl 28(%eax)      # edi
-        popl 32(%eax)      # ebp
-        popl 36(%eax)      # esi
-        popl 40(%eax)      # ds
-        popl 44(%eax)      # es
-        popl 48(%eax)      # fs
-        popl 52(%eax)      # gs
-        popl 56(%eax)      # esp
+        popl proc_intr_ctx(%eax)
 
         addl $8, %esp
 
@@ -180,8 +251,9 @@
         shll $3, %eax
         addl $12, %eax
         addl %esp, %eax
-        movl %eax, (_tss + 4)
+        movl %eax, (_tss + tss_esp0_off)
         movl tmp_store, %eax
+
         iret
 
     .global switch_to
@@ -191,8 +263,8 @@
 
         popl %ebx               # next
         movl __current, %eax    
-        movl 64(%eax), %ecx     # __current->pagetable
-        movl 64(%ebx), %eax     # next->pagetable
+        movl proc_page_table(%eax), %ecx     # __current->pagetable
+        movl proc_page_table(%ebx), %eax     # next->pagetable
         
         cmpl %ecx, %eax         # if(next->pagtable != __current->pagetable) {
         jz 1f
@@ -204,13 +276,19 @@
         # 我们已经处在了新的地址空间，为了避免影响其先前的栈布局
         # 需要使用一个临时的栈空间
         movl $tmp_stack, %esp
+        
+        # 更新 tss
+        movl proc_intr_ctx(%ebx), %eax      # proc->intr_ctx
+        movl iesp(%eax), %eax      # intr_ctx->esp
+        movl %eax, (tss_esp0_off + _tss)
+
         call signal_dispatch    # kernel/signal.c
 
         test %eax, %eax         # do we have signal to handle?
         jz 1f
         jmp handle_signal
     1:
-        leal 8(%ebx), %eax
+        movl proc_intr_ctx(%ebx), %eax
         jmp soft_iret
 
     .global handle_signal
diff --git a/lunaix-os/kernel/asm/x86/interrupts.c b/lunaix-os/kernel/asm/x86/interrupts.c
index 3fb24cb..9f86cf2 100644
--- a/lunaix-os/kernel/asm/x86/interrupts.c
+++ b/lunaix-os/kernel/asm/x86/interrupts.c
@@ -20,9 +20,9 @@ void
 intr_handler(isr_param* param)
 {
     param->execp->saved_prev_ctx = __current->intr_ctx;
-    __current->intr_ctx = *param;
+    __current->intr_ctx = param;
 
-    volatile struct exec_param* execp = __current->intr_ctx.execp;
+    volatile struct exec_param* execp = __current->intr_ctx->execp;
 
     if (execp->vector <= 255) {
         isr_cb subscriber = isrm_get(execp->vector);
diff --git a/lunaix-os/kernel/asm/x86/tss.c b/lunaix-os/kernel/asm/x86/tss.c
index 45c01ee..603eca6 100644
--- a/lunaix-os/kernel/asm/x86/tss.c
+++ b/lunaix-os/kernel/asm/x86/tss.c
@@ -1,12 +1,7 @@
 #include <arch/x86/tss.h>
 #include <lunaix/common.h>
+#include <lunaix/process.h>
 
 volatile struct x86_tss _tss = { .link = 0,
                                  .esp0 = KSTACK_TOP,
-                                 .ss0 = KDATA_SEG };
-
-void
-tss_update_esp(u32_t esp0)
-{
-    _tss.esp0 = esp0;
-}
\ No newline at end of file
+                                 .ss0 = KDATA_SEG };
\ No newline at end of file
diff --git a/lunaix-os/kernel/exe/exec.c b/lunaix-os/kernel/exe/exec.c
index 3c4e253..a846126 100644
--- a/lunaix-os/kernel/exe/exec.c
+++ b/lunaix-os/kernel/exe/exec.c
@@ -220,7 +220,7 @@ __DEFINE_LXSYSCALL3(int,
 
     // we will jump to new entry point (_u_start) upon syscall's
     // return so execve 'will not return' from the perspective of it's invoker
-    volatile struct exec_param* execp = __current->intr_ctx.execp;
+    volatile struct exec_param* execp = __current->intr_ctx->execp;
     execp->esp = container.stack_top;
     execp->eip = container.exe.entry;
 
diff --git a/lunaix-os/kernel/k_init.c b/lunaix-os/kernel/k_init.c
index 71adb30..8c40b6b 100644
--- a/lunaix-os/kernel/k_init.c
+++ b/lunaix-os/kernel/k_init.c
@@ -126,10 +126,6 @@ spawn_proc0()
      * 目前的解决方案是2
      */
 
-    proc0->intr_ctx = (isr_param){ .registers = { .ds = KDATA_SEG,
-                                                  .es = KDATA_SEG,
-                                                  .fs = KDATA_SEG,
-                                                  .gs = KDATA_SEG } };
     proc0->parent = proc0;
 
     // 方案1：必须在读取eflags之后禁用。否则当进程被调度时，中断依然是关闭的！
@@ -141,7 +137,7 @@ spawn_proc0()
     proc0->page_table = vmm_dup_vmspace(proc0->pid);
 
     // 直接切换到新的拷贝，进行配置。
-    cpu_lcr3(proc0->page_table);
+    cpu_ldvmspace(proc0->page_table);
 
     // 为内核创建一个专属栈空间。
     for (size_t i = 0; i < (KSTACK_SIZE >> PG_SIZE_BITS); i++) {
@@ -155,12 +151,19 @@ spawn_proc0()
 
     struct exec_param* execp =
       (struct exec_param*)(KSTACK_TOP - sizeof(struct exec_param));
+    isr_param* isrp = (isr_param*)((ptr_t)execp - sizeof(isr_param));
 
     *execp = (struct exec_param){ .cs = KCODE_SEG,
                                   .eip = (ptr_t)__proc0,
                                   .ss = KDATA_SEG,
                                   .eflags = cpu_reflags() };
-    proc0->intr_ctx.execp = execp;
+    *isrp = (isr_param){ .registers = { .ds = KDATA_SEG,
+                                        .es = KDATA_SEG,
+                                        .fs = KDATA_SEG,
+                                        .gs = KDATA_SEG },
+                         .execp = execp };
+
+    proc0->intr_ctx = isrp;
 
     // 加载x87默认配置
     asm volatile("fninit\n"
diff --git a/lunaix-os/kernel/process/process.c b/lunaix-os/kernel/process/process.c
index 533c577..4e8fa0a 100644
--- a/lunaix-os/kernel/process/process.c
+++ b/lunaix-os/kernel/process/process.c
@@ -221,7 +221,13 @@ dup_proc()
     __copy_fdtable(pcb);
     region_copy(&__current->mm, &pcb->mm);
 
-    setup_proc_mem(pcb, VMS_SELF);
+    /*
+     *  store the return value for forked process.
+     *  this will be implicit carried over after kernel stack is copied.
+     */
+    store_retval(0);
+
+    copy_kernel_stack(pcb, VMS_SELF);
 
     // 根据 mm_region 进一步配置页表
 
@@ -240,9 +246,6 @@ dup_proc()
 
     vmm_unmount_pd(VMS_MOUNT_1);
 
-    // 正如同fork，返回两次。
-    store_retval_to(pcb, 0);
-
     commit_process(pcb);
 
     return pcb->pid;
@@ -251,7 +254,7 @@ dup_proc()
 extern void __kernel_end;
 
 void
-setup_proc_mem(struct proc_info* proc, ptr_t usedMnt)
+copy_kernel_stack(struct proc_info* proc, ptr_t usedMnt)
 {
     // copy the entire kernel page table
     pid_t pid = proc->pid;
@@ -279,9 +282,5 @@ setup_proc_mem(struct proc_info* proc, ptr_t usedMnt)
         *ppte = (p & 0xfff) | ppa;
     }
 
-    // 我们不需要分配内核的区域，因为所有的内核代码和数据段只能通过系统调用来访问，任何非法的访问
-    // 都会导致eip落在区域外面，从而segmentation fault.
-
-    // 至于其他的区域我们暂时没有办法知道，因为那需要知道用户程序的信息。我们留到之后在处理。
     proc->page_table = pt_copy;
 }
\ No newline at end of file
diff --git a/lunaix-os/kernel/process/sched.c b/lunaix-os/kernel/process/sched.c
index 5789927..1587848 100644
--- a/lunaix-os/kernel/process/sched.c
+++ b/lunaix-os/kernel/process/sched.c
@@ -1,5 +1,5 @@
+#include <arch/abi.h>
 #include <arch/x86/interrupts.h>
-#include <arch/x86/tss.h>
 
 #include <hal/apic.h>
 #include <hal/cpu.h>
@@ -63,6 +63,8 @@ sched_init_dummy()
     struct exec_param* execp =
       (void*)dummy_stack + DUMMY_STACK_SIZE - sizeof(struct exec_param);
 
+    isr_param* isrp = (void*)execp - sizeof(isr_param);
+
     *execp = (struct exec_param){
         .cs = KCODE_SEG,
         .eflags = cpu_reflags() | 0x0200,
@@ -70,13 +72,15 @@ sched_init_dummy()
         .ss = KDATA_SEG,
     };
 
+    *isrp = (isr_param){ .registers = { .ds = KDATA_SEG,
+                                        .es = KDATA_SEG,
+                                        .fs = KDATA_SEG,
+                                        .gs = KDATA_SEG },
+                         .execp = execp };
+
     // memset to 0
     dummy_proc = (struct proc_info){};
-    dummy_proc.intr_ctx = (isr_param){ .registers = { .ds = KDATA_SEG,
-                                                      .es = KDATA_SEG,
-                                                      .fs = KDATA_SEG,
-                                                      .gs = KDATA_SEG },
-                                       .execp = execp };
+    dummy_proc.intr_ctx = isrp;
 
     dummy_proc.page_table = cpu_rcr3();
     dummy_proc.state = PS_READY;
@@ -99,7 +103,6 @@ run(struct proc_info* proc)
         由于这中间没有进行地址空间的交换，所以第二次跳转使用的是同一个内核栈，而之前默认tss.esp0的值是永远指向最顶部
         这样一来就有可能会覆盖更早的上下文信息（比如嵌套的信号捕获函数）
     */
-    tss_update_esp(proc->intr_ctx.esp);
 
     apic_done_servicing();
 
@@ -226,7 +229,7 @@ __DEFINE_LXSYSCALL1(unsigned int, sleep, unsigned int, seconds)
         llist_append(&root_proc->sleep.sleepers, &__current->sleep.sleepers);
     }
 
-    __current->intr_ctx.registers.eax = seconds;
+    store_retval(seconds);
 
     block_current();
     schedule();
diff --git a/lunaix-os/kernel/process/signal.c b/lunaix-os/kernel/process/signal.c
index a8ec239..15c650d 100644
--- a/lunaix-os/kernel/process/signal.c
+++ b/lunaix-os/kernel/process/signal.c
@@ -76,7 +76,7 @@ signal_dispatch()
         解决办法就是先吧intr_ctx拷贝到一个静态分配的区域里，然后再注入到用户栈。
     */
     static volatile struct proc_sigstate __temp_save;
-    __temp_save.proc_regs = __current->intr_ctx;
+    __temp_save.proc_regs = *__current->intr_ctx;
     memcpy(__temp_save.fxstate, __current->fxstate, 512);
 
     sigframe->sig_num = sig_selected;
@@ -149,7 +149,8 @@ send_single:
 __DEFINE_LXSYSCALL1(int, sigreturn, struct proc_sig, *sig_ctx)
 {
     memcpy(__current->fxstate, sig_ctx->prev_context.fxstate, 512);
-    __current->intr_ctx = sig_ctx->prev_context.proc_regs;
+    // FIXME: Interrupt context is exposed to user space!
+    *__current->intr_ctx = sig_ctx->prev_context.proc_regs;
 
     struct sigact* current = __current->sigctx.inprogress;
     if (current) {