MIT 6.828 - 9. Lab 09: mmap

Tags: MIT 6.828

实验总结

1. 本次实验用时约 20 个小时。
2. 收获是写了一些和 mmap 相关的代码，尝试实现了一个 mmap 的子集，并且通过了课程测试。

实验结束后的全部代码在：https://github.com/monkey2000/xv6-riscv/tree/mmap/

测试结果：

running mmaptest: 
$ make qemu-gdb
(2.3s) 
  mmaptest: mmap_test: OK 
  mmaptest: fork_test: OK 
usertests: 
$ make qemu-gdb
OK (25.5s) 
time: OK 
Score: 100/100

0. 实验准备

实验指导链接

上来直接：

$ cd xv6-riscv-fall19
$ git checkout mmap

1. 实现

Mmap 细节好多…

首先我在这里详述一下都需要实现哪些功能：

1. 实现 mmap 和 munmap 两个系统调用
2. 懒加载，只有触发 page fault 时才真正加载
3. 支持 fork ，和子进程应该共享 Physical Pages （这块还要加引用计数，我从之前 cow 的里面抄了一个）
4. 实现写回，即实现 MAP_SHAERD/MAP_PRIVATE
5. 实现 PROT_READ/PROT_WRITE

系统调用（以及写回）：

uint64 sys_mmap(void) {
  struct proc *p = myproc();
  uint64 addr;
  int length, prot, flags, fd, offset;

  if(argaddr(0, &addr) < 0)
    return -1;
  
  if(argint(1, &length) < 0 || argint(2, &prot) < 0 || argint(3, &flags) < 0
    || argint(4, &fd) < 0 || argint(5, &offset) < 0)
    return -1;
  
  // assumption: addr and offset shall all be zero
  if(addr != 0 || offset != 0)
    return -1;

  if(p->ofile[fd] == 0 || p->ofile[fd]->type != FD_INODE)
    return -1;

  struct file *f = p->ofile[fd];
  if((flags & MAP_SHARED) && (prot & PROT_WRITE) && !f->writable)
    return -1;

  struct map_info *map;
  if((map = map_alloc()) == 0)
    return -1;
  
  map->vend = PGROUNDUP(map->vstart + length);
  map->length = length;
  map->prot = prot;
  map->flags = flags;
  map->file = f;
  map->offset = offset;
  map->used = 1;

  filedup(map->file);

  return map->vstart;
}

uint64 sys_munmap(void) {
  struct proc *p = myproc();
  uint64 addr;
  int length;

  if(argaddr(0, &addr) < 0 || argint(1, &length) < 0)
    return -1;

  if(addr < MMAP_VSTART || addr > MMAP_VEND)
    return -1;
  
  uint64 vpage_base = PGROUNDDOWN(addr);
  int map_sel = (vpage_base - MMAP_VSTART) / MMAP_SIZE;
  struct map_info *map = &p->minfo[map_sel];

  if(map->used == 0)
    return -1;
  
  static pte_t *pte;
  uint64 pa;
  for(uint64 va = vpage_base; va < addr + length; va += PGSIZE) {
    if((pte = walk(p->pagetable, va, 0)) && (*pte & PTE_V)) {
      pa = PTE2PA(*pte);
      if((map->flags & MAP_SHARED) && (*pte & PTE_D)) { // dirty page
        // printf("dirty!\n");
        uint64 file_start = va - vpage_base;
        uint64 write_length = PGSIZE;
        if(write_length > map->length - file_start)
          write_length = map->length - file_start;
        struct file *f = map->file;

        begin_op(f->ip->dev);
        ilock(f->ip);
        writei(f->ip, 0, pa, file_start, write_length);
        // printf("%d bytes wrote, start = %d, length = %d\n", p, file_start, write_length);
        iunlock(f->ip);
        end_op(f->ip->dev);
      }
      uvmunmap(p->pagetable, va, PGSIZE, 0);
      kderef((void*)pa);
    }
  }

  if(map->vstart == addr && map->length == length) {
    fileclose(map->file);
    map->used = 0;
  } else {
    if(map->vstart == addr)
      map->vstart = addr + length;
    if(map->vend == addr + length)
      map->vend = addr;
  }

  return 0;
}

void mmap_dup(pagetable_t pagetable, struct map_info *m) {
  static pte_t *pte;
  uint64 pa;
  for (uint64 va = m->vstart; va < m->vend; va += PGSIZE) {
    if ((pte = walk(pagetable, va, 0)) && (*pte & PTE_V)) {
      pa = PTE2PA(*pte);
      kref((void *)pa);
    }
  }
}

void mmap_dedup(pagetable_t pagetable, struct map_info *m) {
  static pte_t *pte;
  uint64 pa;
  for (uint64 va = m->vstart; va < m->vend; va += PGSIZE) {
    if ((pte = walk(pagetable, va, 0)) && (*pte & PTE_V)) {
      pa = PTE2PA(*pte);
      uvmunmap(pagetable, va, PGSIZE, 0);
      kderef((void *)pa);
    }
  }
}

懒加载（在 kernel/trap.c > usertrap() 里）：

if(r_scause() == 13 || r_scause() == 15) {
  // 13 for load page fault, 15 for store page fault
  uint64 fault_vaddr = r_stval();
  if(fault_vaddr >= MMAP_VSTART && fault_vaddr < MMAP_VEND) {
    pagetable_t pagetable = p->pagetable;
    uint64 vpage_base = PGROUNDDOWN(fault_vaddr);
    int map_sel = (vpage_base - MMAP_VSTART) / MMAP_SIZE;
    struct map_info *map = &p->minfo[map_sel];
    uint64 map_start = map->vstart;

    // check for invalid mapping & access
    if(map->used == 0 || fault_vaddr >= map->vend) {
      printf("usertrap(): segfault pid=%d epc=%p\n", p->pid, r_sepc());
      p->killed = 1;
      goto end;
    }

    uint64 file_start = vpage_base - map_start;
    uint64 read_length = PGSIZE;
    if(read_length > map->length - file_start)
      read_length = map->length - file_start;
    
    struct file *f = map->file;
    int prot = ((map->prot & PROT_READ) ? PTE_R : 0) | ((map->prot & PROT_WRITE) ? PTE_W : 0);

    char *mem = kalloc();
    if(mem == 0) {
      printf("usertrap(): segfault: no more physical page available, killing process due to a OOM\n");
      p->killed = 1;
      goto end;
    }

    memset(mem, 0, PGSIZE);
    ilock(f->ip);
    readi(f->ip, 0, (uint64)mem, file_start, read_length);
    iunlock(f->ip);

    if(mappages(pagetable, vpage_base, PGSIZE, (uint64)mem, prot|PTE_U) != 0){
      printf("usertrap(): segfault: cannot map a page");
      kfree(mem);
      p->killed = 1;
      goto end;
    }
  } else {
    printf("usertrap(): unexpected scause %p (%s) pid=%d\n", r_scause(), scause_desc(r_scause()), p->pid);
    printf("            sepc=%p stval=%p\n", r_sepc(), r_stval());
    p->killed = 1;
  }
}

支持 fork 和 exit：

// kernel/proc.c > fork()
// share mmap with parent
for(i = 0; i < MMAP_NUM; i++) {
  if(p->minfo[i].used) {
    np->minfo[i] = p->minfo[i];
    mmap_dup(np->pagetable, np->minfo);
  }
}

// kernel/proc.c > exit()
// remove all pte related to mmap
for(int i = 0; i < MMAP_NUM; i++) {
  if(p->minfo[i].used) {
    struct map_info *m = &p->minfo[i];
    m->used = 0;
    mmap_dedup(p->pagetable, m);
  }
}

---

---

2000 - 2099 MonKey's Blog | 自豪地采用 Hexo + Pandoc + KaTeX + Highlight.js