原文：linux内核笔记之进程地址空间

进程的地址空间由允许进程使用的全部线性地址组成，在32位系统中为0~3GB，每个进程看到的线性地址集合是不同的。

内核通过线性区的资源（数据结构）来表示线性地址区间，线性区是由起始线性地址，长度和一些访问权限来描述的。线性区的大小为页框的整数倍，起始地址为4096的整数倍。

下图展示了x86 Linux 进程的地址空间组织结构：

• 正文段 .text ，这是CPU执行的机器指令部分。通常正文段是共享的，而且是只读的，以防止程序修改其自身的指令。
• 数据段 .data。数据段包含了程序中需要明确赋初值的变量。
• 非初始化数据段 bss。bss 起始于 IBM 704汇编语言中的 Block Storage Start 指令的首字母缩写,并且沿用至今。

线性区描述符

299structvm_area_struct {300/* The first cache line has the info for VMA tree walking. */301302unsignedlongvm_start;/* Our start address within vm_mm. */303unsignedlongvm_end;/* The first byte after our end address304 within vm_mm. */305306/* linked list of VM areas per task, sorted by address */307structvm_area_struct *vm_next, *vm_prev;308309structrb_node vm_rb;310311/*312 * Largest free memory gap in bytes to the left of this VMA.313 * Either between this VMA and vma->vm_prev, or between one of the314 * VMAs below us in the VMA rbtree and its ->vm_prev. This helps315 * get_unmapped_area find a free area of the right size.316 */317unsignedlongrb_subtree_gap;318319/* Second cache line starts here. */320321structmm_struct *vm_mm;/* The address space we belong to. */322pgprot_tvm_page_prot;/* Access permissions of this VMA. */323unsignedlongvm_flags;/* Flags, see mm.h. */324325/*326 * For areas with an address space and backing store,327 * linkage into the address_space->i_mmap interval tree.328 */329struct{330structrb_node rb;331unsignedlongrb_subtree_last;332} shared;333334/*335 * A file‘s MAP_PRIVATE vma can be in both i_mmap tree and anon_vma336 * list, after a COW of one of the file pages. A MAP_SHARED vma337 * can only be in the i_mmap tree. An anonymous MAP_PRIVATE, stack338 * or brk vma (with NULL file) can only be in an anon_vma list.339 */340structlist_head anon_vma_chain;/* Serialized by mmap_sem &341 * page_table_lock */342structanon_vma *anon_vma;/* Serialized by page_table_lock */343344/* Function pointers to deal with this struct. */345conststructvm_operations_struct *vm_ops;346347/* Information about our backing store: */348unsignedlongvm_pgoff;/* Offset (within vm_file) in PAGE_SIZE349 units */350structfile * vm_file;/* File we map to (can be NULL). */351void* vm_private_data;/* was vm_pte (shared mem) */352353#ifndefCONFIG_MMU354structvm_region *vm_region;/* NOMMU mapping region */355#endif356#ifdefCONFIG_NUMA357structmempolicy *vm_policy;/* NUMA policy for the VMA */358#endif359structvm_userfaultfd_ctx vm_userfaultfd_ctx;360};

• vm_start :线性区的起始地址
• vm_end :线性区的结束地址
• vm_rb :作为红黑树中的一个节点使用
• vm_mm :指向所在的内存描述符
• vm_page_prot :线性区中页框的访问权限
• vm_flags :线性区的标志
• vm_next, vm_prev :分别指向线性区链表中的下一个和上一个线性区描述符
• … …

内存描述符

395structmm_struct {396structvm_area_struct *mmap;/* list of VMAs */397structrb_root mm_rb;398u32 vmacache_seqnum;/* per-thread vmacache */399#ifdefCONFIG_MMU400unsignedlong(*get_unmapped_area)(structfile *filp,401unsignedlongaddr,unsignedlonglen,402unsignedlongpgoff,unsignedlongflags);403#endif404unsignedlongmmap_base;/* base of mmap area */405unsignedlongmmap_legacy_base;/* base of mmap area in bottom-up allocations */406unsignedlongtask_size;/* size of task vm space */407unsignedlonghighest_vm_end;/* highest vma end address */408pgd_t* pgd;409atomic_tmm_users;/* How many users with user space? */410atomic_tmm_count;/* How many references to "struct mm_struct" (users count as 1) */411atomic_long_tnr_ptes;/* PTE page table pages */412#ifCONFIG_PGTABLE_LEVELS > 2413atomic_long_tnr_pmds;/* PMD page table pages */414#endif415intmap_count;/* number of VMAs */416417spinlock_tpage_table_lock;/* Protects page tables and some counters */418structrw_semaphore mmap_sem;419420structlist_head mmlist;/* List of maybe swapped mm‘s. These are globally strung421 * together off init_mm.mmlist, and are protected422 * by mmlist_lock423 */424425426unsignedlonghiwater_rss;/* High-watermark of RSS usage */427unsignedlonghiwater_vm;/* High-water virtual memory usage */428429unsignedlongtotal_vm;/* Total pages mapped */430unsignedlonglocked_vm;/* Pages that have PG_mlocked set */431unsignedlongpinned_vm;/* Refcount permanently increased */432unsignedlongdata_vm;/* VM_WRITE & ~VM_SHARED & ~VM_STACK */433unsignedlongexec_vm;/* VM_EXEC & ~VM_WRITE & ~VM_STACK */434unsignedlongstack_vm;/* VM_STACK */435unsignedlongdef_flags;436unsignedlongstart_code, end_code, start_data, end_data;437unsignedlongstart_brk, brk, start_stack;438unsignedlongarg_start, arg_end, env_start, env_end;439440unsignedlongsaved_auxv[AT_VECTOR_SIZE];/* for /proc/PID/auxv */441442/*443 * Special counters, in some configurations protected by the444 * page_table_lock, in other configurations by being atomic.445 */446structmm_rss_stat rss_stat;447448structlinux_binfmt *binfmt;449450cpumask_var_tcpu_vm_mask_var;451452/* Architecture-specific MM context */453mm_context_tcontext;454455unsignedlongflags;/* Must use atomic bitops to access the bits */456457structcore_state *core_state;/* coredumping support */458#ifdefCONFIG_AIO459spinlock_tioctx_lock;460structkioctx_table__rcu *ioctx_table;461#endif462#ifdefCONFIG_MEMCG463/*464 * "owner" points to a task that is regarded as the canonical465 * user/owner of this mm. All of the following must be true in466 * order for it to be changed:467 *468 * current == mm->owner469 * current->mm != mm470 * new_owner->mm == mm471 * new_owner->alloc_lock is held472 */473structtask_struct__rcu *owner;474#endif475476/* store ref to file /proc/<pid>/exe symlink points to */477structfile__rcu *exe_file;478#ifdefCONFIG_MMU_NOTIFIER479structmmu_notifier_mm *mmu_notifier_mm;480#endif481#ifdefined(CONFIG_TRANSPARENT_HUGEPAGE) && !USE_SPLIT_PMD_PTLOCKS482pgtable_tpmd_huge_pte;/* protected by page_table_lock */483#endif484#ifdefCONFIG_CPUMASK_OFFSTACK485structcpumask cpumask_allocation;486#endif487#ifdefCONFIG_NUMA_BALANCING488/*489 * numa_next_scan is the next time that the PTEs will be marked490 * pte_numa. NUMA hinting faults will gather statistics and migrate491 * pages to new nodes if necessary.492 */493unsignedlongnuma_next_scan;494495/* Restart point for scanning and setting pte_numa */496unsignedlongnuma_scan_offset;497498/* numa_scan_seq prevents two threads setting pte_numa */499intnuma_scan_seq;500#endif501#ifdefined(CONFIG_NUMA_BALANCING) || defined(CONFIG_COMPACTION)502/*503 * An operation with batched TLB flushing is going on. Anything that504 * can move process memory needs to flush the TLB when moving a505 * PROT_NONE or PROT_NUMA mapped page.506 */507booltlb_flush_pending;508#endif509structuprobes_state uprobes_state;510#ifdefCONFIG_X86_INTEL_MPX511/* address of the bounds directory */512void__user *bd_addr;513#endif514#ifdefCONFIG_HUGETLB_PAGE515atomic_long_thugetlb_usage;516#endif517#ifdefCONFIG_MMU518structwork_struct async_put_work;519#endif520};

• mmap :线性区描述符链表中的头元素
• mm_rb :线性区描述符所在红黑树的根
• get_unmapped_area :在进程地址空间中搜索有效线性地址区间的方法
• mmap_base :标识第一个分配的匿名线性区或文件内存映射的线性地址
• task_size :进程地址空间的大小
• highest_vm_end :能使用的最高线性地址
• pgd :指向页全局目录
• mm_users :次使用计数器
• mm_count :主使用计数器
• nr_ptes :页表项数量
• map_count ：线性区数量
• mmlist :链接内存描述符链表中的相邻描述符
• … …

线性区相关

进程所拥有的所有线性区通过一个简单地链表链接在一起，链表中的线性区按内存地址升序排列。内核通过进程的内存描述符的 mmap 字段找到线性区链表的第一个线性区。

内核频繁执行的一个操作就是查找包含指定线性地址的的线性区，虽然可以通过遍历链表来查找，但是当线性区数量很庞大时，例如面向对象的数据库，此时的效率会变得非常低效。

Linux2.6把内存描述符存放在红黑树的数据结构中，当插入或删除一个线性区描述符时，内核通过红黑树搜索前后元素，并用搜索结果快速更新链表而不用扫描链表。一般来说，红黑树用来确定含有指定地址的线性区，而链表通常在扫描整个线性区集合的时候使用。

内存描述符相关

进程、内存描述符、线性区描述符、线性地址之间的关系如下:

所有内存描述符存放在一个双向链表中，每个描述符中的 mmlist 字段存放链表中相邻元素的地址。链表的第一个元素是 init_mm 的 mmlist 字段， init_mm 是初始化阶段进程0所使用的内存描述符。

mm_users 字段存放共享 mm_struct 数据结构的轻量级进程（线程）的个数， mm_count 字段是内存描述符的主使用器， mm_users 的所有使用者在 mm_count 中只占有一个单位，也就是说多个线程只使得 mm_count 的值增加了1。假如一个内核线程使用了该内存描述符，则 mm_count 的值增加1。

对于内核线程来说，因为仅运行在内核态，所以永远不会访问低于 TASK_SIZE （3GB）的地址。每个进程的描述符中包含了两种内存描述符指针: mm 和 active_mm。对于普通进程，两者都指向进程的内存描述符，而内核线程的 active_mm 指向进程的内存描述符， mm 为null。（PS:内核线程使用的全局页表存放在主内存描述符的pgd字段中）

总结

本文简单地描述了进程地址空间中的一些主要数据结构以及之间的联系，关于源码以及更多细节内容在日后整理。

【转载】linux内核笔记之进程地址空间