XenLinux balloon的实现分析

balloon的概念来自下面的经典论文

Carl Waldspurger. in Proceedings of the 5th Symposium on Operating Systems Design and Implementation, 2002

简而言之，balloon就是VMM打入guest的一个探子，通常以驱动方式存在，当guest空闲内存比较多时，通过吹气球inflat，让balloon申请若干内存，然后转给VMM。当guest load比较重，可以放气deflat，让VMM塞给balloon若干内存，然后balloon释放给guest。Linux 3.0已经包含了balloon for Xen的实现。

OVERVIEW

balloon的状态由描述，该结构在balloon_init函数中初始化。

struct { //in xen/balloon.h

       /* We aim for 'current allocation' == 'target allocation'. */

       unsigned long current_pages;  //guest当前实际mfn数目

       unsigned long ;   //balloon期待guest拥有的mfn数目

       /* Number of pages in high- and low-memory balloons. guest中被剥夺mfn的页面*/

       unsigned long balloon_low; 

       unsigned long balloon_high;

       unsigned long schedule_delay;

       unsigned long max_schedule_delay;

       unsigned long retry_count;

       unsigned long max_retry_count;

};

工作队列负责具体的操作，实际处理函数balloon_process根据target_pages值决定充气还是放气。外围代码通过调用balloon_set_new_target驱动balloon_worker。

static DECLARE_DELAYED_WORK(, );

 

balloon的外围代码在linux 3.0/drivers/xen/xen-balloon.c中，一个是xenstore的节点memory/target设置watch，另一个是在guest的sysfs文件系统创建子目录xen_memory，该子目录下有几个文件读/写balloon_stats的状态

核心操作inflat和deflat

主要代码中

increase_reservation 对应放气，增加guest的内存。

decrease_reservation对应吹气，减少guest的内存。

 

/* List of , threaded through the mem_map array. ballooned_pages的页面都是有struct page结构，但是其对应的mfn已经被xen回收了

*/

static LIST_HEAD();

 

static enum bp_state increase_reservation(unsigned long nr_pages)

{

       int rc;

       unsigned long  pfn, i;

       struct page   *page;

       struct xen_memory_reservation reservation = {

              .address_bits = 0,

              .extent_order = 0,

              .domid        = DOMID_SELF

       };

      

   /*下面的先取得若干page结构，然后通过超级调用populate_physmap让xen给这些page分配mfn,但是实际并不是如此，返回的mfn放在列表中，但是并没有和开始传递进去的各个page绑定起来，也就是前面传进去的参数只有reservation.nr_extents = nr_pages; 才有作用，page传进去似乎是浪费（只有XENMEMF_populate_on_demand标记传进去page才有作用 fix me）。

*/

 

       page = balloon_first_page();

       for (i = 0; i < nr_pages; i++) {

              if (!page) {

                     nr_pages = i;

                     break;

              }

              frame_list[i] = page_to_pfn(page);

              page = balloon_next_page(page);

       }

       set_xen_guest_handle(reservation.extent_start, );

      

       rc = HYPERVISOR_memory_op(, &reservation);

      

  //将部分page从ballooned pages摘除，和frame_list中取得的mfn绑定，包括P2M表，

//,最后guest使用

       for (i = 0; i < rc; i++) {

              page = balloon_retrieve(false);

 

              pfn = page_to_pfn(page);

              set_phys_to_machine(pfn, );// 更新P2M表

 

              /* Link back into the page tables if not highmem. */

              if (xen_pv_domain() && !PageHighMem(page)) {

                     int ret;

                     ret = HYPERVISOR_update_va_mapping(//更新guest的页表

                            (unsigned long)__va(pfn << PAGE_SHIFT),

                            mfn_pte(frame_list[i], PAGE_KERNEL),

                            0);

              }

 

              /* Relinquish the page back to the allocator. */

              ClearPageReserved(page);

              init_page_count(page);

              ;// 释放给guest使用

       }

       balloon_stats.current_pages += rc;

       return BP_DONE;

}

 

static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)

{

       enum bp_state state = BP_DONE;

       unsigned long  pfn, i;

       struct page   *page;

       int ret;

       struct xen_memory_reservation reservation = {

              .address_bits = 0,

              .extent_order = 0,

              .domid        = DOMID_SELF

       };

 

       for (i = 0; i < nr_pages; i++) {

              if ((page = alloc_page(gfp)) == NULL) {//向guest申请页面

                     nr_pages = i;

                     state = BP_EAGAIN;

                     break;

              }

              pfn = page_to_pfn(page);

              frame_list[i] = pfn_to_mfn(pfn);//取得对应的mfn

 

              scrub_page(page); //为安全起见（可以挪到另一个guest使用），清0

 

              if (xen_pv_domain() && !PageHighMem(page)) {

                     ret = HYPERVISOR_update_va_mapping(//将页表项清0

                            (unsigned long)__va(pfn << PAGE_SHIFT),

                            __pte_ma(0), 0);

                }

 

       }

 

       /* Ensure that ballooned highmem pages don't have kmaps. */

       kmap_flush_unused();

       flush_tlb_all();

 

       /* No more mappings: invalidate P2M and add to balloon. */

       for (i = 0; i < nr_pages; i++) {

              pfn = mfn_to_pfn(frame_list[i]);

              __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);

              balloon_append(pfn_to_page(pfn));//将page加入ballooned_pages链表

       }

 

       set_xen_guest_handle(reservation.extent_start, frame_list);

       reservation.nr_extents   = nr_pages;

    //将mfn转给Xen

       ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);

       balloon_stats.current_pages -= nr_pages;

       return state;

}

alloc_xenballooned_pages和free_xenballooned_pages接口

见[Xen-devel] Re: [PATCH 2/3] xen-balloon: Add interface to retrieve ballooned pages

Pages that have been ballooned are useful for other Xen drivers doing grant table actions, because these pages have valid struct page/PFNs but have no valid MFN so are available for remapping.

Xen的支持

见do_memory_op中的 XENMEM_increase_reservation和XENMEM_decrease_reservation。increase_reservation是利用alloc_domheap_pages从Xen中获取物理页面，而decrease_reservation 是调用guest_remove_page从guest回收页面。