linux内核源码阅读之facebook硬盘加速利器flashcache

从来没有写过源码阅读，这种感觉越来越强烈，虽然劣于文笔，但还是下定决心认真写一回。

源代码下载请参见上一篇flashcache之我见 http://blog.csdn.net/liumangxiong/article/details/11643473

下面代码对应的是tag下面的1.0版本的。

看内核模块源码，闭着眼睛打开flashcache_init函数，区区百来行代码何足惧也。

1963int __init 

1964flashcache_init(void)

1965{

1966	int r;

1967

1968	r = flashcache_jobs_init();

1969	if (r)

1970		return r;

1971	atomic_set(&nr_cache_jobs, 0);

1972	atomic_set(&nr_pending_jobs, 0);

1973#if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,20)

1974	INIT_WORK(&_kcached_wq, do_work, NULL);

1975#else

1976	INIT_WORK(&_kcached_wq, do_work);

1977#endif

1978	for (r = 0 ; r < 33 ; r++)

1979		size_hist[r] = 0;

1980	r = dm_register_target(&flashcache_target);

1981	if (r < 0) {

1982		DMERR("cache: register failed %d", r);

1983	}

1984#ifdef CONFIG_PROC_FS

1985#if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27)

1986	flashcache_table_header = 

1987		register_sysctl_table(flashcache_root_table, 1);

1988#else

1989	flashcache_table_header = 

1990		register_sysctl_table(flashcache_root_table);

1991#endif

1992	{

1993		struct proc_dir_entry *entry;

1994		

1995		entry = create_proc_entry("flashcache_stats", 0, NULL);

1996		if (entry)

1997			entry->proc_fops =  &flashcache_stats_operations;

1998		entry = create_proc_entry("flashcache_errors", 0, NULL);

1999		if (entry)

2000			entry->proc_fops =  &flashcache_errors_operations;

2001		entry = create_proc_entry("flashcache_iosize_hist", 0, NULL);

2002		if (entry)

2003			entry->proc_fops =  &flashcache_iosize_hist_operations;

2004		entry = create_proc_entry("flashcache_pidlists", 0, NULL);

2005		if (entry)

2006			entry->proc_fops =  &flashcache_pidlists_operations;

2007		entry = create_proc_entry("flashcache_version", 0, NULL);

2008		if (entry)

2009			entry->proc_fops =  &flashcache_version_operations;

2010	}

2011#endif

2012	flashcache_control = (struct flashcache_control_s *)

2013		kmalloc(sizeof(struct flashcache_control_s *), GFP_KERNEL);

2014	flashcache_control->synch_flags = 0;

2015	register_reboot_notifier(&flashcache_notifier);

2016	return r;

2017}

先大致看一眼，flashcache_jobs_init()分配job内存结构的，INIT_WORK初始化WORK的，接下来一看proc字眼就知道是/proc下目录的文件，再后来创建一个flashcache_control_s管理结构，再注册一个关机回调函数。

这样就走马观花地把这个函数看完了，那让写代码的人情何以堪？

再问一下自己，flashcache究竟做了什么？脑子里还是一片空白。那接下来就到每个函数内探个究竟。

441static int 

442flashcache_jobs_init(void)

443{

444#if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27)

445	_job_cache = kmem_cache_create("kcached-jobs",

446	                               sizeof(struct kcached_job),

447	                               __alignof__(struct kcached_job),

448	                               0, NULL, NULL);

449#else

450	_job_cache = kmem_cache_create("kcached-jobs",

451	                               sizeof(struct kcached_job),

452	                               __alignof__(struct kcached_job),

453	                               0, NULL);

454#endif

455	if (!_job_cache)

456		return -ENOMEM;

457

458	_job_pool = mempool_create(MIN_JOBS, mempool_alloc_slab,

459	                           mempool_free_slab, _job_cache);

460	if (!_job_pool) {

461		kmem_cache_destroy(_job_cache);

462		return -ENOMEM;

463	}

464#if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27)

465	_pending_job_cache = kmem_cache_create("pending-jobs",

466					       sizeof(struct pending_job),

467					       __alignof__(struct pending_job),

468					       0, NULL, NULL);

469#else

470	_pending_job_cache = kmem_cache_create("pending-jobs",

471					       sizeof(struct pending_job),

472					       __alignof__(struct pending_job),

473					       0, NULL);

474#endif

475	if (!_pending_job_cache) {

476		mempool_destroy(_job_pool);

477		kmem_cache_destroy(_job_cache);

478		return -ENOMEM;

479	}

480

481	_pending_job_pool = mempool_create(MIN_JOBS, mempool_alloc_slab,

482					   mempool_free_slab, _pending_job_cache);

483	if (!_pending_job_pool) {

484		kmem_cache_destroy(_pending_job_cache);

485		mempool_destroy(_job_pool);

486		kmem_cache_destroy(_job_cache);

487		return -ENOMEM;

488	}

489

490	return 0;

491}

首先是flashcache_jobs_init()函数，该函数里创建了两类job和两类的mem_pool，就像双胞胎看起来一样，实际上并不一样。

_job_pool => flashcache_alloc_cache_job => new_kcached_job 调用new_kcached_job 有好多个，有flashcache_dirty_writeback、flashcache_read_hit、flashcache_read_miss、flashcache_write_miss、flashcache_write_hit、flashcache_dirty_writeback_sync、flashcache_start_uncached_io。如果仔细地看一下这些函数的名称，发现这些函数所做的事情正是一个写缓存的基本操作和动作，即writeback, writethrough, hit, miss。

现在就以flashcache_dirty_writeback为例，看看到底在kcacheed_job起了什么作用？

code

首先是用new_kcached_job申请一个kcached_job结构体，接下来判断dmc->fast_remove_in_prog，这个是移除flashcache标志，设备都要删除掉了，显然就没必要再下发命令了。再判断job是否为空，else这里才是干的正事。这里job->action = WRITEDISK;是最重要的一句话，就是前面讲的写缓存基本操作，而这个action就可以看作是一个状态机，对应的状态如下：

245/* kcached/pending job states */

246#define READCACHE	1

247#define WRITECACHE	2

248#define READDISK	3

249#define WRITEDISK	4

250#define READFILL	5	/* Read Cache Miss Fill */

251#define INVALIDATE	6

252#define WRITEDISK_SYNC	7

这里设置的是WRITEDISK，就是写磁盘，那是从哪里写呢？是从写缓存写的，写缓存的数据又是在哪里呢？我们把SSD盘当作写缓存，所以是从SSD盘写到磁盘。那我们是不是要做很多事情，先从SSD读数据然后再往磁盘写呢？是的，但是我们不用做太多的事情，因为linux内核有大名鼎鼎的kcopyd线程，我们只需要把这些烦索的工作交给kcopyd完成就可以了，调用的接口是

int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,

             unsigned int num_dests, struct dm_io_region *dests,
             unsigned int flags, dm_kcopyd_notify_fn fn, void *context)

第一个参数是kcopyd_client，这是是flashcache_ctr即flashcache设备创建的构造函数中创建的，即每一个flashcache设备都对应一个kcopyd_client，那么为什么要创建这个结构体呢？可以简单地理解为使用kcopyd服务的一个句柄。第二参数是数据源，第三个为目的数量，第四个参数为要写的目标，第五个参数为额外标识，这里都设置为0，第六个参数fn是回调函数，设置了回调函数则此函数为异步，不阻塞，如果fn设置为NULL，则会同步等待。最后一个参数context是用于回调函数使用的参数，这里传入的正是我们现在最关心的job。

我们已经把kcached_job派发出去了，接着来看是kcached_job是什么时候回来的，回来又做了什么事情，最后是怎么销毁的？

在dm_kcopyd_copy中设置的回调函数是flashcache_kcopyd_callback。

901static void 

902flashcache_kcopyd_callback(int read_err, unsigned int write_err, void *context)

903{

904	struct kcached_job *job = (struct kcached_job *)context;

905	struct cache_c *dmc = job->dmc;

906	int index = job->index;

907	unsigned long flags;

908

909	VERIFY(!in_interrupt());

910	DPRINTK("kcopyd_callback: Index %d", index);

911	VERIFY(job->bio == NULL);

912	spin_lock_irqsave(&dmc->cache_spin_lock, flags);

913	VERIFY(dmc->cache[index].cache_state & (DISKWRITEINPROG | VALID | DIRTY));

914	if (unlikely(sysctl_flashcache_error_inject & KCOPYD_CALLBACK_ERROR)) {

915		read_err = -EIO;

916		sysctl_flashcache_error_inject &= ~KCOPYD_CALLBACK_ERROR;

917	}

918	if (likely(read_err == 0 && write_err == 0)) {

919		spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

920		flashcache_md_write(job);

921	} else {

922		/* Disk write failed. We can not purge this block from flash */

923		DMERR("flashcache: Disk writeback failed ! read error %d write error %d block %lu", 

924		      -read_err, -write_err, job->disk.sector);

925		VERIFY(dmc->cache_sets[index / dmc->assoc].clean_inprog > 0);

926		VERIFY(dmc->clean_inprog > 0);

927		dmc->cache_sets[index / dmc->assoc].clean_inprog--;

928		dmc->clean_inprog--;

929		spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

930		/* Set the error in the job and let do_pending() handle the error */

931		if (read_err) {

932			dmc->ssd_read_errors++;			

933			job->error = read_err;

934		} else {

935			dmc->disk_write_errors++;			

936			job->error = write_err;

937		}

938		flashcache_do_pending(job);

939		flashcache_clean_set(dmc, index / dmc->assoc); /* Kick off more cleanings */

940		dmc->cleanings++;

941	}

942}

到这里就表明写缓存的数据写到磁盘的过程已经完成了。首先检查结果是否成功了，如果都成功的话就调用flashcache_md_write。

860

861/* 

862 * Kick off a cache metadata update (called from workqueue).

863 * Cache metadata update IOs to a given metadata sector are serialized using the 

864 * nr_in_prog bit in the md sector bufhead.

865 * If a metadata IO is already in progress, we queue up incoming metadata updates

866 * on the pending_jobs list of the md sector bufhead. When kicking off an IO, we

867 * cluster all these pending updates and do all of them as 1 flash write (that 

868 * logic is in md_write_kickoff), where it switches out the entire pending_jobs

869 * list and does all of those updates.

870 */

871void

872flashcache_md_write(struct kcached_job *job)

873{

874	struct cache_c *dmc = job->dmc;

875	struct cache_md_sector_head *md_sector_head;

876	unsigned long flags;

877	

878	VERIFY(!in_interrupt());

879	VERIFY(job->action == WRITEDISK || job->action == WRITECACHE || 

880	       job->action == WRITEDISK_SYNC);

881	md_sector_head = &dmc->md_sectors_buf[INDEX_TO_MD_SECTOR(job->index)];

882	spin_lock_irqsave(&dmc->cache_spin_lock, flags);

883	/* If a write is in progress for this metadata sector, queue this update up */

884	if (md_sector_head->nr_in_prog != 0) {

885		struct kcached_job **nodepp;

886		

887		/* A MD update is already in progress, queue this one up for later */

888		nodepp = &md_sector_head->pending_jobs;

889		while (*nodepp != NULL)

890			nodepp = &((*nodepp)->next);

891		job->next = NULL;

892		*nodepp = job;

893		spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

894	} else {

895		md_sector_head->nr_in_prog = 1;

896		spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

897		flashcache_md_write_kickoff(job);

898	}

899}

如果函数有注释还是仔细看一下吧，据个人观察，写linux内核的哥们都是惜字如金，如果他愿意写注释，那看注释绝对比看代码更重要，更有意义，如果有文档的话，那文档就是重中之重。看到这里有注释，真是欣喜万分，基本上看了注释不用看代码都行，但对于我这样的小菜鸟来说，有时还不能完全领会大侠的神意，就会继续读一下代码。

861/* 

862 * Kick off a cache metadata update (called from workqueue).

863 * Cache metadata update IOs to a given metadata sector are serialized using the 

864 * nr_in_prog bit in the md sector bufhead.

865 * If a metadata IO is already in progress, we queue up incoming metadata updates

866 * on the pending_jobs list of the md sector bufhead. When kicking off an IO, we

867 * cluster all these pending updates and do all of them as 1 flash write (that 

868 * logic is in md_write_kickoff), where it switches out the entire pending_jobs

869 * list and does all of those updates.

870 */

派发cache metadata更新（从workqueue调用=》因为这里是从kcopyd回调回来的，所以这里友情提示一下，在内核要十分关心调用的上下文，是看内核代码的必修课，有时也是解决疑难问题的基础）。cache metadata的更新是由结构cache_md_sector_head中nr_in_prog字段来控制更新次序的（就是说更新cache metadata是按次序的，如果前面的更新未完成，后面的更新就排队等候）。排队等候的kcached_job就挂在cache_md_sector_head的pending_jobs上。在前面的更新操作回来时，就一次性把pending_jobs上的所有更新操作一次性派发。（因为所有更新就是对应一个sector中flashcache管理结构的）。

这一段看不明白也没关系，因为这里还没有讲到flashcache的数据组织。但必须明白，我们在flashcache_dirty_writeback中把脏数据从写缓存SSD刷到磁盘，这里要做的事情就是把这个脏数据的的metadata从内存刷到SSD，这样就保证了在异常掉电的情况下元数据可以从SSD中找回。

到这里kcached_job还没有销毁，我们继续跟踪下去 flashcache_md_write=>flashcache_md_write_kickoff。

660static void

661flashcache_md_write_kickoff(struct kcached_job *job)

662{

663	struct cache_c *dmc = job->dmc;	

664	struct flash_cacheblock *md_sector;

665	int md_sector_ix;

666#if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,27)

667	struct io_region where;

668#else

669	struct dm_io_region where;

670#endif

671	int i;

672	struct cache_md_sector_head *md_sector_head;

673	struct kcached_job *orig_job = job;

674	unsigned long flags;

675

676	if (flashcache_alloc_md_sector(job)) {

677		DMERR("flashcache: %d: Cache metadata write failed, cannot alloc page ! block %lu", 

678		      job->action, job->disk.sector);

679		flashcache_md_write_callback(-EIO, job);

680		return;

681	}

682	spin_lock_irqsave(&dmc->cache_spin_lock, flags);

683	/*

684	 * Transfer whatever is on the pending queue to the md_io_inprog queue.

685	 */

686	md_sector_head = &dmc->md_sectors_buf[INDEX_TO_MD_SECTOR(job->index)];

687	md_sector_head->md_io_inprog = md_sector_head->pending_jobs;

688	md_sector_head->pending_jobs = NULL;

689	md_sector = job->md_sector;

690	md_sector_ix = INDEX_TO_MD_SECTOR(job->index) * MD_BLOCKS_PER_SECTOR;

691	/* First copy out the entire sector */

692	for (i = 0 ; 

693	     i < MD_BLOCKS_PER_SECTOR && md_sector_ix < dmc->size ; 

694	     i++, md_sector_ix++) {

695		md_sector[i].dbn = dmc->cache[md_sector_ix].dbn;

696#ifdef FLASHCACHE_DO_CHECKSUMS

697		md_sector[i].checksum = dmc->cache[md_sector_ix].checksum;

698#endif

699		md_sector[i].cache_state = 

700			dmc->cache[md_sector_ix].cache_state & (VALID | INVALID | DIRTY);

701	}

702	/* Then set/clear the DIRTY bit for the "current" index */

703	if (job->action == WRITECACHE) {

704		/* DIRTY the cache block */

705		md_sector[INDEX_TO_MD_SECTOR_OFFSET(job->index)].cache_state = 

706			(VALID | DIRTY);

707	} else { /* job->action == WRITEDISK* */

708		/* un-DIRTY the cache block */

709		md_sector[INDEX_TO_MD_SECTOR_OFFSET(job->index)].cache_state = VALID;

710	}

711

712	for (job = md_sector_head->md_io_inprog ; 

713	     job != NULL ;

714	     job = job->next) {

715		if (job->action == WRITECACHE) {

716			/* DIRTY the cache block */

717			md_sector[INDEX_TO_MD_SECTOR_OFFSET(job->index)].cache_state = 

718				(VALID | DIRTY);

719		} else { /* job->action == WRITEDISK* */

720			/* un-DIRTY the cache block */

721			md_sector[INDEX_TO_MD_SECTOR_OFFSET(job->index)].cache_state = VALID;

722		}

723	}

724	spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

725	where.bdev = dmc->cache_dev->bdev;

726	where.count = 1;

727	where.sector = 1 + INDEX_TO_MD_SECTOR(orig_job->index);

728	dmc->ssd_writes++;

729	dm_io_async_bvec(1, &where, WRITE,

730			 &orig_job->md_io_bvec,

731			 flashcache_md_write_callback, orig_job);

732	flashcache_unplug_device(dmc->cache_dev->bdev);

733}

这里cacheblock 信息保存到job->md_io_bvec的page页中，再调用dm_io_async_bvec将数据写到SSD盘中。我们来看一下该函数原型：

static int dm_io_async_bvec(unsigned int num_regions, 

			    struct dm_io_region *where, int rw, 

			    struct bio_vec *bvec, io_notify_fn fn, 

			    void *context)

该函数与之前的dm_kcopyd_copy类似，我们最关心的是参数where，因为这是人生最重要的一课，你是谁？你要到哪里去？

where的bdev域就是目标设备，而sector域就是起始地址，count表示要写的扇区数。这个函数就是把dmc->cache的管理结构打包到job->md_io_bvec中，然后写到SSD对应位置上。

再接下来看写SSD完成调用flashcache_md_write_callback：

621void 

622flashcache_md_write_callback(unsigned long error, void *context)

623{

624	struct kcached_job *job = (struct kcached_job *)context;

625

626	job->error = error;

627	push_md_complete(job);

628	schedule_work(&_kcached_wq);

629}

该函数只是简单地设置job的返回值，然后放到_md_complete_jobs这个链表里，然后通知workqueue处理。为什么不直接在这个函数里处理，而要放到后面处理呢？这就像每个公司都有个漂亮的前台秘书，这个物流公司送来了大箱的物料，美女秘书当然不会自己搬，随便撒个娇一大群工科男都抢着干活。这里函数是写完成的回调函数，是在软中断中调用的，软中断跟美女秘书一样，干不了重活，只能简单地签收一下，剩下的活就由workqueue来完成了。

要继续我们的跟踪，那就得问workqueue是从哪里来的，workqueue做了什么，或者说对job做了什么？

flashcache_init=>INIT_WORK(&_kcached_wq, do_work);=>process_jobs(&_md_complete_jobs, flashcache_md_write_done);

先看process_jobs

284static void

285process_jobs(struct list_head *jobs,

286	     void (*fn) (struct kcached_job *))

287{

288	struct kcached_job *job;

289

290	while ((job = pop(jobs)))

291		(void)fn(job);

292}

就是从队列中把刚才美女秘书签收的job取出来，然后调用fn，fn就是这里注册的flashcache_md_write_done。

从函数名有个蛋（done），就好像每天下午的5点半，一天的忙碌立马可以收工了，但是悲剧的LZ现在每个月都要加班72个小时，这样想想大家有没有从LZ的不幸中找到自己的幸福？

735void

736flashcache_md_write_done(struct kcached_job *job)

737{

738	struct cache_c *dmc = job->dmc;

739	struct cache_md_sector_head *md_sector_head;

740	int index;

741	unsigned long flags;

742	struct kcached_job *job_list;

743	int error = job->error;

744	struct kcached_job *next;

745	struct cacheblock *cacheblk;

746		

747	VERIFY(!in_interrupt());

748	VERIFY(job->action == WRITEDISK || job->action == WRITECACHE || 

749	       job->action == WRITEDISK_SYNC);

750	flashcache_free_md_sector(job);

751	job->md_sector = NULL;

752	md_sector_head = &dmc->md_sectors_buf[INDEX_TO_MD_SECTOR(job->index)];

753	job_list = job;

754	job->next = md_sector_head->md_io_inprog;

755	md_sector_head->md_io_inprog = NULL;

756	for (job = job_list ; job != NULL ; job = next) {

757		next = job->next;

758		job->error = error;

759		index = job->index;

760		cacheblk = &dmc->cache[index];

761		spin_lock_irqsave(&dmc->cache_spin_lock, flags);

762		if (job->action == WRITECACHE) {

763			if (unlikely(sysctl_flashcache_error_inject & WRITECACHE_MD_ERROR)) {

764				job->error = -EIO;

765				sysctl_flashcache_error_inject &= ~WRITECACHE_MD_ERROR;

766			}

767			if (likely(job->error == 0)) {

768				if ((cacheblk->cache_state & DIRTY) == 0) {

769					dmc->cache_sets[index / dmc->assoc].nr_dirty++;

770					dmc->nr_dirty++;

771				}

772				dmc->md_write_dirty++;

773				cacheblk->cache_state |= DIRTY;

774			} else

775				dmc->ssd_write_errors++;

776			flashcache_bio_endio(job->bio, job->error);

777			if (job->error || cacheblk->head) {

778				if (job->error) {

779					DMERR("flashcache: WRITE: Cache metadata write failed ! error %d block %lu", 

780					      -job->error, cacheblk->dbn);

781				}

782				spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

783				flashcache_do_pending(job);

784			} else {

785				cacheblk->cache_state &= ~BLOCK_IO_INPROG;

786				spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

787				flashcache_free_cache_job(job);

788				if (atomic_dec_and_test(&dmc->nr_jobs))

789					wake_up(&dmc->destroyq);

790			}

791		} else {

792			int action = job->action;

793

794			if (unlikely(sysctl_flashcache_error_inject & WRITEDISK_MD_ERROR)) {

795				job->error = -EIO;

796				sysctl_flashcache_error_inject &= ~WRITEDISK_MD_ERROR;

797			}

798			/*

799			 * If we have an error on a WRITEDISK*, no choice but to preserve the 

800			 * dirty block in cache. Fail any IOs for this block that occurred while

801			 * the block was being cleaned.

802			 */

803			if (likely(job->error == 0)) {

804				dmc->md_write_clean++;

805				cacheblk->cache_state &= ~DIRTY;

806				VERIFY(dmc->cache_sets[index / dmc->assoc].nr_dirty > 0);

807				VERIFY(dmc->nr_dirty > 0);

808				dmc->cache_sets[index / dmc->assoc].nr_dirty--;

809				dmc->nr_dirty--;

810			} else 

811				dmc->ssd_write_errors++;

812			VERIFY(dmc->cache_sets[index / dmc->assoc].clean_inprog > 0);

813			VERIFY(dmc->clean_inprog > 0);

814			dmc->cache_sets[index / dmc->assoc].clean_inprog--;

815			dmc->clean_inprog--;

816			if (job->error || cacheblk->head) {

817				if (job->error) {

818					DMERR("flashcache: CLEAN: Cache metadata write failed ! error %d block %lu", 

819					      -job->error, cacheblk->dbn);

820				}

821				spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

822				flashcache_do_pending(job);

823				/* Kick off more cleanings */

824				if (action == WRITEDISK)

825					flashcache_clean_set(dmc, index / dmc->assoc);

826				else

827					flashcache_sync_blocks(dmc);

828			} else {

829				cacheblk->cache_state &= ~BLOCK_IO_INPROG;

830				spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

831				flashcache_free_cache_job(job);

832				if (atomic_dec_and_test(&dmc->nr_jobs))

833					wake_up(&dmc->destroyq);

834				/* Kick off more cleanings */

835				if (action == WRITEDISK)

836					flashcache_clean_set(dmc, index / dmc->assoc);

837				else

838					flashcache_sync_blocks(dmc);

839			}

840			dmc->cleanings++;

841			if (action == WRITEDISK_SYNC)

842				flashcache_update_sync_progress(dmc);

843		}

844	}

845	spin_lock_irqsave(&dmc->cache_spin_lock, flags);

846	if (md_sector_head->pending_jobs != NULL) {

847		/* peel off the first job from the pending queue and kick that off */

848		job = md_sector_head->pending_jobs;

849		md_sector_head->pending_jobs = job->next;

850		job->next = NULL;

851		spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

852		VERIFY(job->action == WRITEDISK || job->action == WRITECACHE ||

853		       job->action == WRITEDISK_SYNC);

854		flashcache_md_write_kickoff(job);

855	} else {

856		md_sector_head->nr_in_prog = 0;

857		spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);

858	}

859}

860

首先是flashcache_free_md_sector，这个函数只是简单地把刚才分配的记录cacheblock 的page页释放。哪个刚才啊？就是flashcache_md_write_kickoff中flashcache_alloc_md_sector申请的page页。所以看这个函数时要回头再去看看flashcache_md_write_kickoff，所以前面提到了上下文，那么在这里kickoff是上文，done就是下文，上文种什么因，下文就得到什么果。上文申请了page页，下文就要释放page页；上文把dmc->md_sectors_buf[]中struct kcached_job  *md_io_inprog对应的kcached_job都已经下发了，下文这里才有一个for循环。细心的你可能会问，为什么这里的kcached_job可以一起下发？那首先要来了解一下这里的kcached_job是干什么的。是结构体上的：

/* 

 * We have one of these for *every* cache metadata sector, to keep track

 * of metadata ios in progress for blocks covered in this sector. Only

 * one metadata IO per sector can be in progress at any given point in 

 * time

 */

struct cache_md_sector_head {

	u_int32_t		nr_in_prog;

	struct kcached_job	*pending_jobs, *md_io_inprog;

};

按规矩先看注释，每一个cache metadata扇区都有对应一个cache_md_sector_head结构，用于同步进程（内存中）cacheblock metadata到cache metadata扇区。同时只能有一个IO在同步，对应的是cache_md_sector_head->nr_in_prog。回答上面的问题，就是这些kcached_job是对应同一个扇区内的不同metadata的写，所以可以合并。这个扇区指的是SSD盘上存放flash_block结构的。

再回到flashcache_md_write_done函数中，在for循环中job->action为WRITEDISK，所以直接来到for循环中else，迎面而来的又是一行注释，在WRITEDISK*发生错误时，只有保持cacheblock的DIRTY标志。接下来判断有错误或者cacheblock上还有pending_job，那么继续下发IO，否则的话清除cacheblock的处理标志，这里我们终于见到了kcached_job完成了他的使命，调用flashcache_free_cache_job将该结构返回给内存池。

似乎到这里我们就可以像童话里讲的“从此他们过上了幸福的生活”来结束kcached_job的介绍。然而回归资源池也意味着kcached_job的再生，接着判断action==WRITEDISK，调用flashcache_clean_set，将超过脏水平线的cache块刷回到磁盘。就是说在每次写磁盘返回的时候这个workqueue都会检查一下脏水平线，如果超过就继续往下刷，这就又回到了本文最开始的flashcache_dirty_writeback函数，真是因果联系，环环相扣，kcached_job的再生不是为了自己，而是为cacheblock的再生，所以说人不能只为自己活着，每个人只是万千轮回里的一个元素，都是为了成全其他元素而进入六道轮回。

下面一篇会从flashcache的数据结构和存储设计来分析。