linux内核源码阅读之facebook硬盘加速flashcache之六
其实到目前为止，如果对读流程已经能轻松地看懂了，那么写流程不需要太多脑细胞。我觉得再写下去没有太大的必要了，后面想想为了保持flashcache完整性，还是写出来吧。接着到写流程：
- 1530static void
- 1531flashcache_write(struct cache_c *dmc, struct bio *bio)
- 1532{
- 1533 int index;
- 1534 int res;
- 1535 struct cacheblock *cacheblk;
- 1536 int queued;
- 1537
- 1538 spin_lock_irq(&dmc->cache_spin_lock);
- 1539 res = flashcache_lookup(dmc, bio, &index);
- 1540 /*
- 1541 * If cache hit and !BUSY, simply redirty page.
- 1542 * If cache hit and BUSY, must wait for IO in prog to complete.
- 1543 * If cache miss and found a block to recycle, we need to
- 1544 * (a) invalidate any partial hits,
- 1545 * (b) write to cache.
- 1546 */
- 1547 if (res != -1) {
- 1548 /* Cache Hit */
- 1549 cacheblk = &dmc->cache[index];
- 1550 if ((cacheblk->cache_state & VALID) &&
- 1551 (cacheblk->dbn == bio->bi_sector)) {
- 1552 /* Cache Hit */
- 1553 flashcache_write_hit(dmc, bio, index);
- 1554 } else {
- 1555 /* Cache Miss, found block to recycle */
- 1556 flashcache_write_miss(dmc, bio, index);
- 1557 }
- 1558 return;
- 1559 }
- 1560 /*
- 1561 * No room in the set. We cannot write to the cache and have to
- 1562 * send the request to disk. Before we do that, we must check
- 1563 * for potential invalidations !
- 1564 */
- 1565 queued = flashcache_inval_blocks(dmc, bio);
- 1566 spin_unlock_irq(&dmc->cache_spin_lock);
- 1567 if (queued) {
- 1568 if (unlikely(queued < 0))
- 1569 flashcache_bio_endio(bio, -EIO);
- 1570 return;
- 1571 }
- 1572 /* Start uncached IO */
- 1573 flashcache_start_uncached_io(dmc, bio);
- 1574 flashcache_clean_set(dmc, hash_block(dmc, bio->bi_sector));
- 1575}
1530static void
1531flashcache_write(struct cache_c *dmc, struct bio *bio)
1532{
1533 int index;
1534 int res;
1535 struct cacheblock *cacheblk;
1536 int queued;
1537
1538 spin_lock_irq(&dmc->cache_spin_lock);
1539 res = flashcache_lookup(dmc, bio, &index);
1540 /*
1541 * If cache hit and !BUSY, simply redirty page.
1542 * If cache hit and BUSY, must wait for IO in prog to complete.
1543 * If cache miss and found a block to recycle, we need to
1544 * (a) invalidate any partial hits,
1545 * (b) write to cache.
1546 */
1547 if (res != -1) {
1548 /* Cache Hit */
1549 cacheblk = &dmc->cache[index];
1550 if ((cacheblk->cache_state & VALID) &&
1551 (cacheblk->dbn == bio->bi_sector)) {
1552 /* Cache Hit */
1553 flashcache_write_hit(dmc, bio, index);
1554 } else {
1555 /* Cache Miss, found block to recycle */
1556 flashcache_write_miss(dmc, bio, index);
1557 }
1558 return;
1559 }
1560 /*
1561 * No room in the set. We cannot write to the cache and have to
1562 * send the request to disk. Before we do that, we must check
1563 * for potential invalidations !
1564 */
1565 queued = flashcache_inval_blocks(dmc, bio);
1566 spin_unlock_irq(&dmc->cache_spin_lock);
1567 if (queued) {
1568 if (unlikely(queued < 0))
1569 flashcache_bio_endio(bio, -EIO);
1570 return;
1571 }
1572 /* Start uncached IO */
1573 flashcache_start_uncached_io(dmc, bio);
1574 flashcache_clean_set(dmc, hash_block(dmc, bio->bi_sector));
1575}
第1539行查找是否命中，这里有几种情况：
1）命中且cache空闲，直接写cache块并设置DIRTY标志
2）命中且cache忙，等待上一个请求完成
3）不命中并且找到可用的cache块，invalid有交集的cache块，然后再写到cache
4）没有可用cache块，invalid有次的cache块，写到磁盘
第4种情况在第1573行直接写到磁盘，最后调用的还是dm_io_async_bvec。
再看第1种情况，进入到命中处理分支：
- 1468static void
- 1469flashcache_write_hit(struct cache_c *dmc, struct bio *bio, int index)
- 1470{
- 1471 struct cacheblock *cacheblk;
- 1472 struct pending_job *pjob;
- 1473 struct kcached_job *job;
- 1474
- 1475 cacheblk = &dmc->cache[index];
- 1476 if (!(cacheblk->cache_state & BLOCK_IO_INPROG) && (cacheblk->head == NULL)) {
- 1477 if (cacheblk->cache_state & DIRTY)
- 1478 dmc->dirty_write_hits++;
- 1479 dmc->write_hits++;
- 1480 cacheblk->cache_state |= CACHEWRITEINPROG;
- 1481 spin_unlock_irq(&dmc->cache_spin_lock);
- 1482 job = new_kcached_job(dmc, bio, index);
- 1483 if (unlikely(sysctl_flashcache_error_inject & WRITE_HIT_JOB_ALLOC_FAIL)) {
- 1484 if (job)
- 1485 flashcache_free_cache_job(job);
- 1486 job = NULL;
- 1487 sysctl_flashcache_error_inject &= ~WRITE_HIT_JOB_ALLOC_FAIL;
- 1488 }
- 1489 if (unlikely(job == NULL)) {
- 1490 /*
- 1491 * We have a write hit, and can't allocate a job.
- 1492 * Since we dropped the spinlock, we have to drain any
- 1493 * pending jobs.
- 1494 */
- 1495 DMERR("flashcache: Write (hit) failed ! Can't allocate memory for cache IO, block %lu",
- 1496 cacheblk->dbn);
- 1497 flashcache_bio_endio(bio, -EIO);
- 1498 spin_lock_irq(&dmc->cache_spin_lock);
- 1499 flashcache_free_pending_jobs(dmc, cacheblk, -EIO);
- 1500 cacheblk->cache_state &= ~(BLOCK_IO_INPROG);
- 1501 spin_unlock_irq(&dmc->cache_spin_lock);
- 1502 } else {
- 1503 job->action = WRITECACHE; /* Write data to the source device */
- 1504 DPRINTK("Queue job for %llu", bio->bi_sector);
- 1505 atomic_inc(&dmc->nr_jobs);
- 1506 dmc->ssd_writes++;
- 1507 dm_io_async_bvec(1, &job->cache, WRITE,
- 1508 bio->bi_io_vec + bio->bi_idx,
- 1509 flashcache_io_callback, job);
- 1510 flashcache_unplug_device(dmc->cache_dev->bdev);
- 1511 flashcache_clean_set(dmc, index / dmc->assoc);
- 1512 }
- 1513 } else {
- 1514 pjob = flashcache_alloc_pending_job(dmc);
- 1515 if (unlikely(sysctl_flashcache_error_inject & WRITE_HIT_PENDING_JOB_ALLOC_FAIL)) {
- 1516 if (pjob) {
- 1517 flashcache_free_pending_job(pjob);
- 1518 pjob = NULL;
- 1519 }
- 1520 sysctl_flashcache_error_inject &= ~WRITE_HIT_PENDING_JOB_ALLOC_FAIL;
- 1521 }
- 1522 if (unlikely(pjob == NULL))
- 1523 flashcache_bio_endio(bio, -EIO);
- 1524 else
- 1525 flashcache_enq_pending(dmc, bio, index, WRITECACHE, pjob);
- 1526 spin_unlock_irq(&dmc->cache_spin_lock);
- 1527 }
- 1528}
1468static void
1469flashcache_write_hit(struct cache_c *dmc, struct bio *bio, int index)
1470{
1471 struct cacheblock *cacheblk;
1472 struct pending_job *pjob;
1473 struct kcached_job *job;
1474
1475 cacheblk = &dmc->cache[index];
1476 if (!(cacheblk->cache_state & BLOCK_IO_INPROG) && (cacheblk->head == NULL)) {
1477 if (cacheblk->cache_state & DIRTY)
1478 dmc->dirty_write_hits++;
1479 dmc->write_hits++;
1480 cacheblk->cache_state |= CACHEWRITEINPROG;
1481 spin_unlock_irq(&dmc->cache_spin_lock);
1482 job = new_kcached_job(dmc, bio, index);
1483 if (unlikely(sysctl_flashcache_error_inject & WRITE_HIT_JOB_ALLOC_FAIL)) {
1484 if (job)
1485 flashcache_free_cache_job(job);
1486 job = NULL;
1487 sysctl_flashcache_error_inject &= ~WRITE_HIT_JOB_ALLOC_FAIL;
1488 }
1489 if (unlikely(job == NULL)) {
1490 /*
1491 * We have a write hit, and can't allocate a job.
1492 * Since we dropped the spinlock, we have to drain any
1493 * pending jobs.
1494 */
1495 DMERR("flashcache: Write (hit) failed ! Can't allocate memory for cache IO, block %lu",
1496 cacheblk->dbn);
1497 flashcache_bio_endio(bio, -EIO);
1498 spin_lock_irq(&dmc->cache_spin_lock);
1499 flashcache_free_pending_jobs(dmc, cacheblk, -EIO);
1500 cacheblk->cache_state &= ~(BLOCK_IO_INPROG);
1501 spin_unlock_irq(&dmc->cache_spin_lock);
1502 } else {
1503 job->action = WRITECACHE; /* Write data to the source device */
1504 DPRINTK("Queue job for %llu", bio->bi_sector);
1505 atomic_inc(&dmc->nr_jobs);
1506 dmc->ssd_writes++;
1507 dm_io_async_bvec(1, &job->cache, WRITE,
1508 bio->bi_io_vec + bio->bi_idx,
1509 flashcache_io_callback, job);
1510 flashcache_unplug_device(dmc->cache_dev->bdev);
1511 flashcache_clean_set(dmc, index / dmc->assoc);
1512 }
1513 } else {
1514 pjob = flashcache_alloc_pending_job(dmc);
1515 if (unlikely(sysctl_flashcache_error_inject & WRITE_HIT_PENDING_JOB_ALLOC_FAIL)) {
1516 if (pjob) {
1517 flashcache_free_pending_job(pjob);
1518 pjob = NULL;
1519 }
1520 sysctl_flashcache_error_inject &= ~WRITE_HIT_PENDING_JOB_ALLOC_FAIL;
1521 }
1522 if (unlikely(pjob == NULL))
1523 flashcache_bio_endio(bio, -EIO);
1524 else
1525 flashcache_enq_pending(dmc, bio, index, WRITECACHE, pjob);
1526 spin_unlock_irq(&dmc->cache_spin_lock);
1527 }
1528}
在1475行获得cache块，在1476行判断是否空闲，在有IO处理或者有pending_job挂着的时候都视为忙。如果cache块空闲，则进入if分支，接下来又是套路了，创建kcached_job，成功的话就在1507行下发写请求。然后接着看写返回时做了哪些处理？进入写回调函数之前，要记住这里设置了两个标志，一个是1480行cache块的CACHEWRITEINPROG，另一个是1503行kcached_job的WRITECACHE，带着这两个标志进入到写回调函数flashcache_io_callback，并直接找到需要的地方：
- 188 case WRITECACHE:
- 189 DPRINTK("flashcache_io_callback: WRITECACHE %d",
- 190 index);
- 191 spin_lock_irqsave(&dmc->cache_spin_lock, flags);
- 192 if (unlikely(sysctl_flashcache_error_inject & WRITECACHE_ERROR)) {
- 193 job->error = error = -EIO;
- 194 sysctl_flashcache_error_inject &= ~WRITECACHE_ERROR;
- 195 }
- 196 VERIFY(cacheblk->cache_state & CACHEWRITEINPROG);
- 197 if (likely(error == 0)) {
- 198#ifdef FLASHCACHE_DO_CHECKSUMS
- 199 dmc->checksum_store++;
- 200 spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);
- 201 flashcache_store_checksum(job);
- 202 /*
- 203 * We need to update the metadata on a DIRTY->DIRTY as well
- 204 * since we save the checksums.
- 205 */
- 206 push_md_io(job);
- 207 schedule_work(&_kcached_wq);
- 208 return;
- 209#else
- 210 spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);
- 211 /* Only do cache metadata update on a non-DIRTY->DIRTY transition */
- 212 if ((cacheblk->cache_state & DIRTY) == 0) {
- 213 push_md_io(job);
- 214 schedule_work(&_kcached_wq);
- 215 return;
- 216 }
- 217#endif
- 218 } else {
- 219 dmc->ssd_write_errors++;
- 220 spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);
- 221 }
- 222 flashcache_bio_endio(bio, error);
- 223 break;
188 case WRITECACHE:
189 DPRINTK("flashcache_io_callback: WRITECACHE %d",
190 index);
191 spin_lock_irqsave(&dmc->cache_spin_lock, flags);
192 if (unlikely(sysctl_flashcache_error_inject & WRITECACHE_ERROR)) {
193 job->error = error = -EIO;
194 sysctl_flashcache_error_inject &= ~WRITECACHE_ERROR;
195 }
196 VERIFY(cacheblk->cache_state & CACHEWRITEINPROG);
197 if (likely(error == 0)) {
198#ifdef FLASHCACHE_DO_CHECKSUMS
199 dmc->checksum_store++;
200 spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);
201 flashcache_store_checksum(job);
202 /*
203 * We need to update the metadata on a DIRTY->DIRTY as well
204 * since we save the checksums.
205 */
206 push_md_io(job);
207 schedule_work(&_kcached_wq);
208 return;
209#else
210 spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);
211 /* Only do cache metadata update on a non-DIRTY->DIRTY transition */
212 if ((cacheblk->cache_state & DIRTY) == 0) {
213 push_md_io(job);
214 schedule_work(&_kcached_wq);
215 return;
216 }
217#endif
218 } else {
219 dmc->ssd_write_errors++;
220 spin_unlock_irqrestore(&dmc->cache_spin_lock, flags);
221 }
222 flashcache_bio_endio(bio, error);
223 break;
写到缓存成功的话，暂不管cache块的校验值，会来到210行，判断原来的cache块是否为脏，如果为脏那就什么事情都不用做了。因为如果cache块本来就是脏，那新来的IO可以直接覆盖到cache块上去。反之如果原来cache块是干净的，那么这个时候要把cache块已经变脏记录到SSD上，于是进入了第213行开始写cache块管理信息。到了这里似乎cache块已经写到缓存中，IO可以返回了，但是到了第215行为什么直接return呢？这里涉及到数据一致性的问题。其实cache块管理结构没有写到缓存中，这个写请求不能算完成。试想如果在这里调用了第222行flashcache_bio_endio把IO返回了，会有什么样的后果？其实大多数情况下是没有什么问题的，但如果在这个时候系统掉电或者宕机了，这时候缓存中记录的cache块状态是干净的，但又已经跟上层返回说IO已经写成功了，那么最后这一次写的数据就丢失了。当然对于大部分用户来说，这一点数据算什么？但对于像银行这样的系统，当你把辛苦了十年的积蓄存到自动取款机，这时自动取款机告诉你存成功了，但不幸的是后台刚好发生了我们上面描述的问题。结果你再查的时候没有你刚才存进去的钱，但你的钱确确实实被取款机收进去了，这时你会有怎样的感受？这里只是举个数据一致性在某些应用中是非常重要的，当然现实中绝大数银行是不会有这样的问题，银行可以有日志查出来，系统也有热备，也是带UPS保护的。
如果原来的cache块为脏的情况就以第222行flashcache_bio_endio结束了。
如果不为脏，那么调用213行将cache块管理结构写到缓存。
- 272void
- 273push_md_io(struct kcached_job *job)
- 274{
- 275 push(&_md_io_jobs, job);
- 276}
272void
273push_md_io(struct kcached_job *job)
274{
275 push(&_md_io_jobs, job);
276}
这里只是简单放到队列中，具体处理的是第214行唤醒的工作队列。该工作队列对应的处理函数是：
- 303 process_jobs(&_md_io_jobs, flashcache_md_write);
303 process_jobs(&_md_io_jobs, flashcache_md_write);
这个函数怎么这么面熟呢？因为在第一小节里已经介绍过了：
http://blog.csdn.net/liumangxiong/article/details/11681787
这里小结一下写命中并且原cache块为干净的数据流程：
1）写命中调用dm_io_async_bvec写缓存
2）写缓存完成回调函数flashcache_io_callback，判断原cache块为干净，需要写cache块管理结构
3）由工作队列_kcached_wq调用flashcache_md_write写cache块管理结构，最终由flashcache_md_write_kickoff调用dm_io_async_bvec将cache块管理结构写到缓存
4）写缓存完成之后调用flashcache_md_write_callback
5）由工作队列_kcached_wq调用flashcache_md_write_done处理
6）在flashcache_md_write_done中判断job类型为WRITECACHE，最后调用flashcache_bio_endio返回
至此，这个IO才完成使命。
接下来讲第3种情况，这种情况就非常简单了。
- 1411static void
- 1412flashcache_write_miss(struct cache_c *dmc, struct bio *bio, int index)
- 1413{
- 1414 struct cacheblock *cacheblk;
- 1415 struct kcached_job *job;
- 1416 int queued;
- 1417
- 1418 cacheblk = &dmc->cache[index];
- 1419 queued = flashcache_inval_blocks(dmc, bio);
- 1420 if (queued) {
- 1421 if (unlikely(queued < 0))
- 1422 flashcache_bio_endio(bio, -EIO);
- 1423 spin_unlock_irq(&dmc->cache_spin_lock);
- 1424 return;
- 1425 }
- 1426 if (cacheblk->cache_state & VALID)
- 1427 dmc->wr_replace++;
- 1428 else
- 1429 dmc->cached_blocks++;
- 1430 cacheblk->cache_state = VALID | CACHEWRITEINPROG;
- 1431 cacheblk->dbn = bio->bi_sector;
- 1432 spin_unlock_irq(&dmc->cache_spin_lock);
- 1433 job = new_kcached_job(dmc, bio, index);
- 1434 if (unlikely(sysctl_flashcache_error_inject & WRITE_MISS_JOB_ALLOC_FAIL)) {
- 1435 if (job)
- 1436 flashcache_free_cache_job(job);
- 1437 job = NULL;
- 1438 sysctl_flashcache_error_inject &= ~WRITE_MISS_JOB_ALLOC_FAIL;
- 1439 }
- 1440 if (unlikely(job == NULL)) {
- 1441 /*
- 1442 * We have a write miss, and can't allocate a job.
- 1443 * Since we dropped the spinlock, we have to drain any
- 1444 * pending jobs.
- 1445 */
- 1446 DMERR("flashcache: Write (miss) failed ! Can't allocate memory for cache IO, block %lu",
- 1447 cacheblk->dbn);
- 1448 flashcache_bio_endio(bio, -EIO);
- 1449 spin_lock_irq(&dmc->cache_spin_lock);
- 1450 dmc->cached_blocks--;
- 1451 cacheblk->cache_state &= ~VALID;
- 1452 cacheblk->cache_state |= INVALID;
- 1453 flashcache_free_pending_jobs(dmc, cacheblk, -EIO);
- 1454 cacheblk->cache_state &= ~(BLOCK_IO_INPROG);
- 1455 spin_unlock_irq(&dmc->cache_spin_lock);
- 1456 } else {
- 1457 job->action = WRITECACHE;
- 1458 atomic_inc(&dmc->nr_jobs);
- 1459 dmc->ssd_writes++;
- 1460 dm_io_async_bvec(1, &job->cache, WRITE,
- 1461 bio->bi_io_vec + bio->bi_idx,
- 1462 flashcache_io_callback, job);
- 1463 flashcache_unplug_device(dmc->cache_dev->bdev);
- 1464 flashcache_clean_set(dmc, index / dmc->assoc);
- 1465 }
- 1466}
1411static void
1412flashcache_write_miss(struct cache_c *dmc, struct bio *bio, int index)
1413{
1414 struct cacheblock *cacheblk;
1415 struct kcached_job *job;
1416 int queued;
1417
1418 cacheblk = &dmc->cache[index];
1419 queued = flashcache_inval_blocks(dmc, bio);
1420 if (queued) {
1421 if (unlikely(queued < 0))
1422 flashcache_bio_endio(bio, -EIO);
1423 spin_unlock_irq(&dmc->cache_spin_lock);
1424 return;
1425 }
1426 if (cacheblk->cache_state & VALID)
1427 dmc->wr_replace++;
1428 else
1429 dmc->cached_blocks++;
1430 cacheblk->cache_state = VALID | CACHEWRITEINPROG;
1431 cacheblk->dbn = bio->bi_sector;
1432 spin_unlock_irq(&dmc->cache_spin_lock);
1433 job = new_kcached_job(dmc, bio, index);
1434 if (unlikely(sysctl_flashcache_error_inject & WRITE_MISS_JOB_ALLOC_FAIL)) {
1435 if (job)
1436 flashcache_free_cache_job(job);
1437 job = NULL;
1438 sysctl_flashcache_error_inject &= ~WRITE_MISS_JOB_ALLOC_FAIL;
1439 }
1440 if (unlikely(job == NULL)) {
1441 /*
1442 * We have a write miss, and can't allocate a job.
1443 * Since we dropped the spinlock, we have to drain any
1444 * pending jobs.
1445 */
1446 DMERR("flashcache: Write (miss) failed ! Can't allocate memory for cache IO, block %lu",
1447 cacheblk->dbn);
1448 flashcache_bio_endio(bio, -EIO);
1449 spin_lock_irq(&dmc->cache_spin_lock);
1450 dmc->cached_blocks--;
1451 cacheblk->cache_state &= ~VALID;
1452 cacheblk->cache_state |= INVALID;
1453 flashcache_free_pending_jobs(dmc, cacheblk, -EIO);
1454 cacheblk->cache_state &= ~(BLOCK_IO_INPROG);
1455 spin_unlock_irq(&dmc->cache_spin_lock);
1456 } else {
1457 job->action = WRITECACHE;
1458 atomic_inc(&dmc->nr_jobs);
1459 dmc->ssd_writes++;
1460 dm_io_async_bvec(1, &job->cache, WRITE,
1461 bio->bi_io_vec + bio->bi_idx,
1462 flashcache_io_callback, job);
1463 flashcache_unplug_device(dmc->cache_dev->bdev);
1464 flashcache_clean_set(dmc, index / dmc->assoc);
1465 }
1466}
大多数函数都已经是老朋友了。第1430行cache块设置了VALID标志，表示在有效数据，第1431行设置cache块对应的磁盘的bi_sector扇区。接着到第1460行下发写缓存请求，写缓存的情况与写命中的一样就不再继续跟进了。
下一节讲缓存超水位线写回磁盘。