linux内核源码阅读之facebook硬盘加速flashcache之五

正常流程到flashcache_map的1623行或1625行，按顺序先看读流程：

1221static void1222flashcache_read(struct cache_c *dmc, struct bio *bio)1223{1224int index;1225int res;1226struct cacheblock *cacheblk;1227int queued;12281229DPRINTK("Got a %s for %llu  %u bytes)",1230        (bio_rw(bio) == READ ? "READ":"READA"), 1231bio->bi_sector, bio->bi_size);12321233spin_lock_irq(&dmc->cache_spin_lock);1234res = flashcache_lookup(dmc, bio, &index);1235/* 1236 * Handle Cache Hit case first.1237 * We need to handle 2 cases, BUSY and !BUSY. If BUSY, we enqueue the1238 * bio for later.1239 */1240if (res > 0) {1241cacheblk = &dmc->cache[index];1242if ((cacheblk->cache_state & VALID) && 1243    (cacheblk->dbn == bio->bi_sector)) {1244flashcache_read_hit(dmc, bio, index);1245return;1246}1247}1248/*1249 * In all cases except for a cache hit (and VALID), test for potential 1250 * invalidations that we need to do.1251 */1252queued = flashcache_inval_blocks(dmc, bio);1253if (queued) {1254if (unlikely(queued < 0))1255flashcache_bio_endio(bio, -EIO);1256spin_unlock_irq(&dmc->cache_spin_lock);1257return;1258}1259if (res == -1 || flashcache_uncacheable(dmc)) {1260/* No room or non-cacheable */1261spin_unlock_irq(&dmc->cache_spin_lock);1262DPRINTK("Cache read: Block %llu(%lu):%s",1263bio->bi_sector, bio->bi_size, "CACHE MISS & NO ROOM");1264if (res == -1)1265flashcache_clean_set(dmc, hash_block(dmc, bio->bi_sector));1266/* Start uncached IO */1267flashcache_start_uncached_io(dmc, bio);1268return;1269}1270/* 1271 * (res == INVALID) Cache Miss 1272 * And we found cache blocks to replace1273 * Claim the cache blocks before giving up the spinlock1274 */1275if (dmc->cache[index].cache_state & VALID)1276dmc->replace++;1277else1278dmc->cached_blocks++;1279dmc->cache[index].cache_state = VALID | DISKREADINPROG;1280dmc->cache[index].dbn = bio->bi_sector;1281spin_unlock_irq(&dmc->cache_spin_lock);12821283DPRINTK("Cache read: Block %llu(%lu), index = %d:%s",1284bio->bi_sector, bio->bi_size, index, "CACHE MISS & REPLACE");1285flashcache_read_miss(dmc, bio, index);1286}

我非常喜欢flashcache这种小资的感觉，每个函数都比较短，大部分都没有超过100行的。不像neil大哥写的md代码的函数动则几百行，上千行，总是望啊望啊望不到边。当然不是说我不喜欢neil的代码，实际上他的代码是非常非常好的，因为md代码已经有十多年的历史了，大的框架仍然没有太大的改变，仍能保持那么优雅已经是十分难得了。最新版本的md还加了许多新功能，像bad block和replacement机制都是非常实用的。所以对于一名优秀的软件工程师来说，并不在于写了多少行代码，而是编写的软件运行在多少台机器上，为用户创造了多少价值。

第1234行是查找bio是否命中，flashcache_lookup函数我们在之前的文章里已经分析过了。第1244行是命中的情况，我们跟进看看

1119static void1120flashcache_read_hit(struct cache_c *dmc, struct bio* bio, int index)1121{1122struct cacheblock *cacheblk;1123struct pending_job *pjob;11241125cacheblk = &dmc->cache[index];1126if (!(cacheblk->cache_state & BLOCK_IO_INPROG) && (cacheblk->head == NULL)) {1127struct kcached_job *job;11281129cacheblk->cache_state |= CACHEREADINPROG;1130dmc->read_hits++;1131spin_unlock_irq(&dmc->cache_spin_lock);1132DPRINTK("Cache read: Block %llu(%lu), index = %d:%s",1133bio->bi_sector, bio->bi_size, index, "CACHE HIT");1134job = new_kcached_job(dmc, bio, index);1135if (unlikely(sysctl_flashcache_error_inject & READ_HIT_JOB_ALLOC_FAIL)) {1136if (job)1137flashcache_free_cache_job(job);1138job = NULL;1139sysctl_flashcache_error_inject &= ~READ_HIT_JOB_ALLOC_FAIL;1140}1141if (unlikely(job == NULL)) {1142/* 1143 * We have a read hit, and can't allocate a job.1144 * Since we dropped the spinlock, we have to drain any 1145 * pending jobs.1146 */1147DMERR("flashcache: Read (hit) failed ! Can't allocate memory for cache IO, block %lu", 1148      cacheblk->dbn);1149flashcache_bio_endio(bio, -EIO);1150spin_lock_irq(&dmc->cache_spin_lock);1151flashcache_free_pending_jobs(dmc, cacheblk, -EIO);1152cacheblk->cache_state &= ~(BLOCK_IO_INPROG);1153spin_unlock_irq(&dmc->cache_spin_lock);1154} else {1155job->action = READCACHE; /* Fetch data from cache */1156atomic_inc(&dmc->nr_jobs);1157dmc->ssd_reads++;1158dm_io_async_bvec(1, &job->cache, READ,1159 bio->bi_io_vec + bio->bi_idx,1160 flashcache_io_callback, job);1161flashcache_unplug_device(dmc->cache_dev->bdev);1162}1163} else {1164pjob = flashcache_alloc_pending_job(dmc);1165if (unlikely(sysctl_flashcache_error_inject & READ_HIT_PENDING_JOB_ALLOC_FAIL)) {1166if (pjob) {1167flashcache_free_pending_job(pjob);1168pjob = NULL;1169}1170sysctl_flashcache_error_inject &= ~READ_HIT_PENDING_JOB_ALLOC_FAIL;1171}1172if (pjob == NULL)1173flashcache_bio_endio(bio, -EIO);1174else1175flashcache_enq_pending(dmc, bio, index, READCACHE, pjob);1176spin_unlock_irq(&dmc->cache_spin_lock);1177}1178}

首先获取这个cache块管理结构，第1126行判断cache块不忙的情况，进入1129行设置状态为从cache读，第1134行创建一个kcached_job，在1141行申请kcached_job失败时就对bio返回失败。申请成功到1155行将kcached_job设置为READCACHE，再调用dm_io_async_bvec下发请求，当请求回来时就会调用这里设置的回调函数flashcache_io_callback。再继续看读SSD返回是怎么处理的？我们只看该函数是对READCACHE的处理：

151case READCACHE:152DPRINTK("flashcache_io_callback: READCACHE %d",153index);154spin_lock_irqsave(&dmc->cache_spin_lock, flags);155if (unlikely(sysctl_flashcache_error_inject & READCACHE_ERROR)) {156job->error = error = -EIO;157sysctl_flashcache_error_inject &= ~READCACHE_ERROR;158}159VERIFY(cacheblk->cache_state & CACHEREADINPROG);160spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);161if (unlikely(error))162dmc->ssd_read_errors++;163#ifdef FLASHCACHE_DO_CHECKSUMS164if (likely(error == 0)) {165if (flashcache_validate_checksum(job)) {166DMERR("flashcache_io_callback: Checksum mismatch at disk offset %lu", 167      job->disk.sector);168error = -EIO;169}170}171#endif172flashcache_bio_endio(bio, error);173break;       

实际上真正有意义的就是第172行，将请求done回去了。这样我们就完成的一次读命中的处理。

读命中处理还有一种情况就是第1163行cache块忙的情况，这个时候就申请一个pending_job，挂到cache块上，等cache块上一个请求回来的时候调度。

继续回到flashcache_read，看不命中的情况。

到第1259行flashcache_uncacheable函数是管理命令相关的，加了黑名单之后会跳过cache，直接下发到磁盘。

到1264行res是flashcache_lookup返回的，为-1就表示获取不到可用的cache块，这时就调用flashcache_clean_set清除一下脏块。但获取不到cache块并不是说请求就结束了，还得下发到磁盘，1267行flashcache_start_uncached_io将请求直接下发到磁盘。

第1275行到1278行是统计信息，根据这些信息可以知道flashcache的运行状况，用于flashcache的性能优化。

接着1279行设置cache块的状态，1280行设置cache块对应磁盘上的扇区，最后调用flashcache_read_miss下发请求：

1180static void1181flashcache_read_miss(struct cache_c *dmc, struct bio* bio,1182               int index)1183{1184     struct kcached_job *job;1185     struct cacheblock *cacheblk = &dmc->cache[index];11861187     job = new_kcached_job(dmc, bio, index);1188     if (unlikely(sysctl_flashcache_error_inject & READ_MISS_JOB_ALLOC_FAIL)) {1189          if (job)1190               flashcache_free_cache_job(job);1191          job = NULL;1192          sysctl_flashcache_error_inject &= ~READ_MISS_JOB_ALLOC_FAIL;1193     }1194     if (unlikely(job == NULL)) {1195          /* 1196          * We have a read miss, and can't allocate a job.1197          * Since we dropped the spinlock, we have to drain any 1198          * pending jobs.1199          */1200          DMERR("flashcache: Read (miss) failed ! Can't allocate memory for cache IO, block %lu", 1201                cacheblk->dbn);1202          flashcache_bio_endio(bio, -EIO);1203          spin_lock_irq(&dmc->cache_spin_lock);1204          dmc->cached_blocks--;1205          cacheblk->cache_state &= ~VALID;1206          cacheblk->cache_state |= INVALID;1207          flashcache_free_pending_jobs(dmc, cacheblk, -EIO);1208          cacheblk->cache_state &= ~(BLOCK_IO_INPROG);1209          spin_unlock_irq(&dmc->cache_spin_lock);1210     } else {1211          job->action = READDISK; /* Fetch data from the source device */1212          atomic_inc(&dmc->nr_jobs);1213          dmc->disk_reads++;1214          dm_io_async_bvec(1, &job->disk, READ,1215                    bio->bi_io_vec + bio->bi_idx,1216                    flashcache_io_callback, job);1217          flashcache_clean_set(dmc, index / dmc->assoc);1218     }1219}

在第1187行申请了一个kcached_job，申请成功就到1211行，设置job->action=READDISK，调用dm_io_async_bvec直接从磁盘读取数据。接着调用flashcache_clean_set检查一下水位线。再看这里读磁盘的回调函数flashcache_io_callback，按理说读完磁盘就可以直接向上层返回数据，但这里还要把数据缓存起来之后再返回。

113void 114flashcache_io_callback(unsigned long error, void *context)115{116     struct kcached_job *job = (struct kcached_job *) context;117     struct cache_c *dmc = job->dmc;118     struct bio *bio;119     unsigned long flags;120     int index = job->index;121     struct cacheblock *cacheblk = &dmc->cache[index];122123     VERIFY(index != -1);          124     bio = job->bio;125     VERIFY(bio != NULL);126     if (error)127          DMERR("flashcache_io_callback: io error %ld block %lu action %d", 128                error, job->disk.sector, job->action);129     job->error = error;130     switch (job->action) {131     case READDISK:132          DPRINTK("flashcache_io_callback: READDISK  %d",133               index);134          spin_lock_irqsave(&dmc->cache_spin_lock, flags);135          if (unlikely(sysctl_flashcache_error_inject & READDISK_ERROR)) {136               job->error = error = -EIO;137               sysctl_flashcache_error_inject &= ~READDISK_ERROR;138          }139          VERIFY(cacheblk->cache_state & DISKREADINPROG);140          spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);141          if (likely(error == 0)) {142               /* Kick off the write to the cache */143               job->action = READFILL;144               flashcache_enqueue_readfill(dmc, job);145               return;146          } else {147               dmc->disk_read_errors++;               148               flashcache_bio_endio(bio, error);149          }150          break;          174     case READFILL:175          DPRINTK("flashcache_io_callback: READFILL %d",176               index);177          spin_lock_irqsave(&dmc->cache_spin_lock, flags);178          if (unlikely(sysctl_flashcache_error_inject & READFILL_ERROR)) {179               job->error = error = -EIO;180               sysctl_flashcache_error_inject &= ~READFILL_ERROR;181          }182          if (unlikely(error))183               dmc->ssd_write_errors++;184          VERIFY(cacheblk->cache_state & DISKREADINPROG);185          spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);186          flashcache_bio_endio(bio, error);187          break;

归纳一下读不命中的流程：

1）创建一个kcached_job，直接下发到磁盘

2）读磁盘返回到flashcache_io_callback，到131行下发READFILL，将读出来的数据写到缓存中

3）写缓存成功并返回到flashcache_io_callback，到174行将数据返回给上层

到这里已经将读流程简单过了一遍，下一个小节介绍写流程。