正常流程到flashcache_map的1623行或1625行，按順序先看讀流程：
1221static void
1222flashcache_read(struct cache_c *dmc, struct bio *bio)
1223{
1224 int index;
1225 int res;
1226 struct cacheblock *cacheblk;
1227 int queued;
1228
1229 DPRINTK("Got a %s for %llu %u bytes)",
1230 (bio_rw(bio) == READ ? "READ":"READA"),
1231 bio->bi_sector, bio->bi_size);
1232
1233 spin_lock_irq(&dmc->cache_spin_lock);
1234 res = 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 the
1238 * bio for later.
1239 */
1240 if (res > 0) {
1241 cacheblk = &dmc->cache[index];
1242 if ((cacheblk->cache_state & VALID) &&
1243 (cacheblk->dbn == bio->bi_sector)) {
1244 flashcache_read_hit(dmc, bio, index);
1245 return;
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 */
1252 queued = flashcache_inval_blocks(dmc, bio);
1253 if (queued) {
1254 if (unlikely(queued < 0))
1255 flashcache_bio_endio(bio, -EIO);
1256 spin_unlock_irq(&dmc->cache_spin_lock);
1257 return;
1258 }
1259 if (res == -1 || flashcache_uncacheable(dmc)) {
1260 /* No room or non-cacheable */
1261 spin_unlock_irq(&dmc->cache_spin_lock);
1262 DPRINTK("Cache read: Block %llu(%lu):%s",
1263 bio->bi_sector, bio->bi_size, "CACHE MISS & NO ROOM");
1264 if (res == -1)
1265 flashcache_clean_set(dmc, hash_block(dmc, bio->bi_sector));
1266 /* Start uncached IO */
1267 flashcache_start_uncached_io(dmc, bio);
1268 return;
1269 }
1270 /*
1271 * (res == INVALID) Cache Miss
1272 * And we found cache blocks to replace
1273 * Claim the cache blocks before giving up the spinlock
1274 */
1275 if (dmc->cache[index].cache_state & VALID)
1276 dmc->replace++;
1277 else
1278 dmc->cached_blocks++;
1279 dmc->cache[index].cache_state = VALID | DISKREADINPROG;
1280 dmc->cache[index].dbn = bio->bi_sector;
1281 spin_unlock_irq(&dmc->cache_spin_lock);
1282
1283 DPRINTK("Cache read: Block %llu(%lu), index = %d:%s",
1284 bio->bi_sector, bio->bi_size, index, "CACHE MISS & REPLACE");
1285 flashcache_read_miss(dmc, bio, index);
1286}我非常喜歡flashcache這種小資的感覺，每個函數都比較短，大部分都沒有超過100行的。不像neil大哥寫的md代碼的函數動則幾百行，上千行，總是望啊望啊望不到邊。當然不是說我不喜歡neil的代碼，實際上他的代碼是非常非常好的，因爲md代碼已經有十多年的歷史了，大的框架仍然沒有太大的改變，仍能保持那麼優雅已經是十分難得了。最新版本的md還加了許多新功能，像bad block和replacement機制都是非常實用的。所以對於一名優秀的軟件工程師來說，並不在於寫了多少行代碼，而是編寫的軟件運行在多少臺機器上，爲用戶創造了多少價值。
第1234行是查找bio是否命中，flashcache_lookup函數我們在之前的文章裏已經分析過了。第1244行是命中的情況，我們跟進看看
1119static void
1120flashcache_read_hit(struct cache_c *dmc, struct bio* bio, int index)
1121{
1122 struct cacheblock *cacheblk;
1123 struct pending_job *pjob;
1124
1125 cacheblk = &dmc->cache[index];
1126 if (!(cacheblk->cache_state & BLOCK_IO_INPROG) && (cacheblk->head == NULL)) {
1127 struct kcached_job *job;
1128
1129 cacheblk->cache_state |= CACHEREADINPROG;
1130 dmc->read_hits++;
1131 spin_unlock_irq(&dmc->cache_spin_lock);
1132 DPRINTK("Cache read: Block %llu(%lu), index = %d:%s",
1133 bio->bi_sector, bio->bi_size, index, "CACHE HIT");
1134 job = new_kcached_job(dmc, bio, index);
1135 if (unlikely(sysctl_flashcache_error_inject & READ_HIT_JOB_ALLOC_FAIL)) {
1136 if (job)
1137 flashcache_free_cache_job(job);
1138 job = NULL;
1139 sysctl_flashcache_error_inject &= ~READ_HIT_JOB_ALLOC_FAIL;
1140 }
1141 if (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 */
1147 DMERR("flashcache: Read (hit) failed ! Can't allocate memory for cache IO, block %lu",
1148 cacheblk->dbn);
1149 flashcache_bio_endio(bio, -EIO);
1150 spin_lock_irq(&dmc->cache_spin_lock);
1151 flashcache_free_pending_jobs(dmc, cacheblk, -EIO);
1152 cacheblk->cache_state &= ~(BLOCK_IO_INPROG);
1153 spin_unlock_irq(&dmc->cache_spin_lock);
1154 } else {
1155 job->action = READCACHE; /* Fetch data from cache */
1156 atomic_inc(&dmc->nr_jobs);
1157 dmc->ssd_reads++;
1158 dm_io_async_bvec(1, &job->cache, READ,
1159 bio->bi_io_vec + bio->bi_idx,
1160 flashcache_io_callback, job);
1161 flashcache_unplug_device(dmc->cache_dev->bdev);
1162 }
1163 } else {
1164 pjob = flashcache_alloc_pending_job(dmc);
1165 if (unlikely(sysctl_flashcache_error_inject & READ_HIT_PENDING_JOB_ALLOC_FAIL)) {
1166 if (pjob) {
1167 flashcache_free_pending_job(pjob);
1168 pjob = NULL;
1169 }
1170 sysctl_flashcache_error_inject &= ~READ_HIT_PENDING_JOB_ALLOC_FAIL;
1171 }
1172 if (pjob == NULL)
1173 flashcache_bio_endio(bio, -EIO);
1174 else
1175 flashcache_enq_pending(dmc, bio, index, READCACHE, pjob);
1176 spin_unlock_irq(&dmc->cache_spin_lock);
1177 }
1178}首先獲取這個cache塊管理結構，第1126行判斷cache塊不忙的情況，進入1129行設置狀態爲從cache讀，第1134行創建一個kcached_job，在1141行申請kcached_job失敗時就對bio返回失敗。申請成功到1155行將kcached_job設置爲READCACHE，再調用dm_io_async_bvec下發請求，當請求回來時就會調用這裏設置的回調函數flashcache_io_callback。再繼續看讀SSD返回是怎麼處理的？我們只看該函數是對READCACHE的處理：
151 case READCACHE:
152 DPRINTK("flashcache_io_callback: READCACHE %d",
153 index);
154 spin_lock_irqsave(&dmc->cache_spin_lock, flags);
155 if (unlikely(sysctl_flashcache_error_inject & READCACHE_ERROR)) {
156 job->error = error = -EIO;
157 sysctl_flashcache_error_inject &= ~READCACHE_ERROR;
158 }
159 VERIFY(cacheblk->cache_state & CACHEREADINPROG);
160 spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);
161 if (unlikely(error))
162 dmc->ssd_read_errors++;
163#ifdef FLASHCACHE_DO_CHECKSUMS
164 if (likely(error == 0)) {
165 if (flashcache_validate_checksum(job)) {
166 DMERR("flashcache_io_callback: Checksum mismatch at disk offset %lu",
167 job->disk.sector);
168 error = -EIO;
169 }
170 }
171#endif
172 flashcache_bio_endio(bio, error);
173 break;
實際上真正有意義的就是第172行，將請求done回去了。這樣我們就完成的一次讀命中的處理。
讀命中處理還有一種情況就是第1163行cache塊忙的情況，這個時候就申請一個pending_job，掛到cache塊上，等cache塊上一個請求回來的時候調度。
繼續回到flashcache_read，看不命中的情況。
到第1259行flashcache_uncacheable函數是管理命令相關的，加了黑名單之後會跳過cache，直接下發到磁盤。
到1264行res是flashcache_lookup返回的，爲-1就表示獲取不到可用的cache塊，這時就調用flashcache_clean_set清除一下髒塊。但獲取不到cache塊並不是說請求就結束了，還得下發到磁盤，1267行flashcache_start_uncached_io將請求直接下發到磁盤。
第1275行到1278行是統計信息，根據這些信息可以知道flashcache的運行狀況，用於flashcache的性能優化。
接着1279行設置cache塊的狀態，1280行設置cache塊對應磁盤上的扇區，最後調用flashcache_read_miss下發請求：
1180static void
1181flashcache_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];
1186
1187 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行申請了一個kcached_job，申請成功就到1211行，設置job->action=READDISK，調用dm_io_async_bvec直接從磁盤讀取數據。接着調用flashcache_clean_set檢查一下水位線。再看這裏讀磁盤的回調函數flashcache_io_callback，按理說讀完磁盤就可以直接向上層返回數據，但這裏還要把數據緩存起來之後再返回。
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];
122
123 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;歸納一下讀不命中的流程：
1）創建一個kcached_job，直接下發到磁盤
2）讀磁盤返回到flashcache_io_callback，到131行下發READFILL，將讀出來的數據寫到緩存中
3）寫緩存成功並返回到flashcache_io_callback，到174行將數據返回給上層
到這裏已經將讀流程簡單過了一遍，下一個小節介紹寫流程。