feat: No more kernel page table switching upon interrupt.

[lunaix-os.git] / lunaix-os / kernel / process.c

diff --git a/lunaix-os/kernel/process.c b/lunaix-os/kernel/process.c
index de6eaf6625ef0136f21e7aa0f94af7b95cc8ef4e..5112ae19e04a7e78ddf2cc36587e5684799ff5a0 100644 (file)
--- a/lunaix-os/kernel/process.c
+++ b/lunaix-os/kernel/process.c
@@ -1,8 +1,12 @@
 #include <lunaix/process.h>
 #include <lunaix/mm/vmm.h>
 #include <lunaix/process.h>
 #include <lunaix/mm/vmm.h>

+#include <lunaix/mm/region.h>

 #include <lunaix/clock.h>
 #include <lunaix/syslog.h>
 #include <lunaix/common.h>
 #include <lunaix/clock.h>
 #include <lunaix/syslog.h>
 #include <lunaix/common.h>

+#include <lunaix/syscall.h>
+#include <lunaix/spike.h>
+#include <klibc/string.h>

 
 LOG_MODULE("PROC")
 
 
 LOG_MODULE("PROC")
 


@@ -25,22 +29,9 @@ void* __dup_pagetable(pid_t pid, uintptr_t mount_point) {
 
         for (size_t j = 0; j < PG_MAX_ENTRIES; j++)
         {
 
         for (size_t j = 0; j < PG_MAX_ENTRIES; j++)
         {

-            uintptr_t va = ((i << 10) | j) << 12;
-            x86_pte_t ppte = ppt->entry[j];
-            if (!ppte || !(ppte & PG_PRESENT)) {
-                pt->entry[j] = ppte;
-                continue;
-            }
-
-            // FIXME: 根据 mm_region 将读共享的页（如堆）标为只读，而私有的页（如栈），则复制；而写共享的页则无需更改flags
-            if (va >= KSTACK_START) {
-                void* ppa = vmm_dup_page(pid, PG_ENTRY_ADDR(ppte));
-                ppte = ppte & 0xfff | (uintptr_t)ppa;
-            }
-            pt->entry[j] = ppte;
-            // ppte = ppte & ~PG_WRITE;
-            // pt->entry[j] = ppte;
-            // ppt->entry[j] = ppte;
+            x86_pte_t pte = ppt->entry[j];
+            pmm_ref_page(pid, PG_ENTRY_ADDR(pte));
+            pt->entry[j] = pte;

         }
 
         ptd->entry[i] = (uintptr_t)pt_pp | PG_PREM_RW;
         }
 
         ptd->entry[i] = (uintptr_t)pt_pp | PG_PREM_RW;


@@ -51,39 +42,152 @@ void* __dup_pagetable(pid_t pid, uintptr_t mount_point) {
     return ptd_pp;
 }
 
     return ptd_pp;
 }
 

+void __del_pagetable(pid_t pid, uintptr_t mount_point) {
+    x86_page_table* pptd = (x86_page_table*) (mount_point | (0x3FF << 12));
+
+    for (size_t i = 0; i < PG_MAX_ENTRIES - 1; i++)
+    {
+        x86_pte_t ptde = pptd->entry[i];
+        if (!ptde || !(ptde & PG_PRESENT)) {
+            continue;
+        }
+        
+        x86_page_table* ppt = (x86_page_table*) (mount_point | (i << 12));
+
+        for (size_t j = 0; j < PG_MAX_ENTRIES; j++)
+        {
+            x86_pte_t pte = ppt->entry[j];
+            // free the 4KB data page
+            if ((pte & PG_PRESENT)) {
+                pmm_free_page(pid,PG_ENTRY_ADDR(pte));
+            }
+        }
+        // free the L2 page table
+        pmm_free_page(pid, PG_ENTRY_ADDR(ptde));
+    }
+    // free the L1 directory
+    pmm_free_page(pid, PG_ENTRY_ADDR(pptd->entry[PG_MAX_ENTRIES - 1]));
+}
+

 void* dup_pagetable(pid_t pid) {
 void* dup_pagetable(pid_t pid) {

-    return __dup_pagetable(pid, L2_BASE_VADDR);
+    return __dup_pagetable(pid, PD_REFERENCED);

 }
 
 }
 

-void dup_proc() {
-    pid_t pid = alloc_pid();
+__DEFINE_LXSYSCALL(pid_t, fork) {
+    return dup_proc();
+}

 
 

-    /* 
-        FIXME: Problematic! It should mount the page table of process then copy it.
-        The current implementation copy the CURRENTLY loaded pgt.
-        However, dup_pagetable is designed to copy current loaded pgt.
-        
-    */
+__DEFINE_LXSYSCALL(pid_t, getpid) {
+    return __current->pid;
+}
+
+__DEFINE_LXSYSCALL(pid_t, getppid) {
+    return __current->parent->pid;
+}

 
 

-    void* mnt_pt = vmm_mount_pd(__current->page_table);
+void init_proc(struct proc_info *pcb) {
+    memset(pcb, 0, sizeof(*pcb));

 
 

-    void* pg = __dup_pagetable(pid, mnt_pt);
+    pcb->pid = alloc_pid();
+    pcb->created = clock_systime();
+    pcb->state = PROC_CREATED;
+}

 
 

-    vmm_unmount_pd();
+pid_t dup_proc() {
+    struct proc_info pcb;
+    init_proc(&pcb);
+    pcb.mm = __current->mm;
+    pcb.intr_ctx = __current->intr_ctx;
+    pcb.parent = __current;
+
+#ifdef USE_KERNEL_PG
+    setup_proc_mem(&pcb, PD_MOUNT_1);    //挂载点#1是当前进程的页表
+#else
+    setup_proc_mem(&pcb, PD_REFERENCED);
+#endif
+
+    // 根据 mm_region 进一步配置页表
+    if (!__current->mm.regions) {
+        goto not_copy;
+    }
+    
+    struct mm_region *pos, *n;
+    llist_for_each(pos, n, &__current->mm.regions->head, head) {
+        region_add(&pcb, pos->start, pos->end, pos->attr);

 
 

-    struct proc_info pcb = (struct proc_info) {
-        .created = clock_systime(),
-        .pid = pid,
-        .mm = __current->mm,
-        .page_table = pg,
-        .intr_ctx = __current->intr_ctx,
-        .parent_created = __current->created
-    };
+        // 如果写共享，则不作处理。
+        if ((pos->attr & REGION_WSHARED)) {
+            continue;
+        }
+
+        uintptr_t start_vpn = PG_ALIGN(pos->start) >> 12;
+        uintptr_t end_vpn = PG_ALIGN(pos->end) >> 12;
+        for (size_t i = start_vpn; i < end_vpn; i++)
+        {
+            x86_pte_t *curproc = &PTE_MOUNTED(PD_MOUNT_1, i);
+            x86_pte_t *newproc = &PTE_MOUNTED(PD_MOUNT_2, i);
+            cpu_invplg(curproc);
+            cpu_invplg(newproc);
+
+            if (pos->attr == REGION_RSHARED) {
+                // 如果读共享，则将两者的都标注为只读，那么任何写入都将会应用COW策略。
+                *curproc = *curproc & ~PG_WRITE;
+                *newproc = *newproc & ~PG_WRITE;
+            }
+            else {
+                // 如果是私有页，则将该页从新进程中移除。
+                *newproc = 0;
+            }
+        }
+    }
+    
+not_copy:
+    vmm_unmount_pd(PD_MOUNT_2);

 
     // 正如同fork，返回两次。
     pcb.intr_ctx.registers.eax = 0;
 
     // 正如同fork，返回两次。
     pcb.intr_ctx.registers.eax = 0;

-    __current->intr_ctx.registers.eax = pid;

 
     push_process(&pcb);
 
     push_process(&pcb);

+
+    return pcb.pid;
+}
+
+extern void __kernel_end;
+
+void setup_proc_mem(struct proc_info* proc, uintptr_t usedMnt) {
+    // copy the entire kernel page table
+    pid_t pid = proc->pid;
+    void* pt_copy = __dup_pagetable(pid, usedMnt);
+
+    vmm_mount_pd(PD_MOUNT_2, pt_copy);   // 将新进程的页表挂载到挂载点#2
+
+    // copy the kernel stack
+    for (size_t i = KSTACK_START >> 12; i <= KSTACK_TOP >> 12; i++)
+    {
+        volatile x86_pte_t *ppte = &PTE_MOUNTED(PD_MOUNT_2, i);
+
+        /*
+            This is a fucking nightmare, the TLB caching keep the rewrite to PTE from updating.
+            Even the Nightmare Moon the Evil is far less nasty than this.
+            It took me hours of debugging to figure this out.
+
+            In the name of Celestia our glorious goddess, I will fucking HATE the TLB for the rest of my LIFE!
+        */
+        cpu_invplg(ppte);
+
+        x86_pte_t p = *ppte;
+        void* ppa = vmm_dup_page(pid, PG_ENTRY_ADDR(p));
+        *ppte = (p & 0xfff) | (uintptr_t)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.

     
     

+    // 定义用户栈区域，但是不分配实际的物理页。我们会在Page fault handler里面实现动态分配物理页的逻辑。（虚拟内存的好处！）
+    // FIXME: 这里应该放到spawn_proc里面。
+    // region_add(proc, USTACK_END, USTACK_SIZE, REGION_PRIVATE | REGION_RW);
+
+    // 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
 }
\ No newline at end of file