block级锁的理解

最新推荐文章于 2021-04-12 01:07:28 发布

guogaofeng1219

最新推荐文章于 2021-04-12 01:07:28 发布

阅读量1k

点赞数

文章标签： tree struct insert action delete merge

本文链接：https://blog.csdn.net/guogaofeng1219/article/details/7177318

版权

1 block 里面的sem本质上是一个队列，是对action的排队
目前来说， action包括（io, unmap, remap, merge）
2 block 里面为什么一定要一把状态的锁，因为在某些流程处理中， {检测状态 = >做相应操作} 后面的相应操作必须要依赖前面的状态。如果不用状态锁的话，其它上下文可能会改变该状态，从而使后面的操作没有意义。
3 put_block中为什么用atomic_dec_and_block。这是为了将{ref_cnt等于0，拿树的锁}做成一个原子的事务。即：拿树的锁是基于ref_cnt为0的情况。

这里可以对比理解一下：引用计数可以使用原子变量，而状态需要加锁（而不用原子变量）。因为引用计数通过atomic_dec_and_block保证{检测，操作}是放在一个原子操作中的。

struct block *search_and_get_block(struct sd *sd, sector_t blk)
{
struct block_block *block = NULL;

find:
read_lock_irqsave(&sd->tree_lock, flag);
if ((block = radix_tree_lookup(&sd->active_blocks_tree, blk))) {
get_block(block);
read_unlock_irqrestore(&sd->tree_lock, flag);
goto out;
}
read_unlock_irqrestore(&sd->tree_lock, flag);

block = alloc_and_init_block(sd, blk);

write_lock_irqsave(&sd->tree_lock, flag);

ret = radix_tree_insert(&sd->active_blocks_tree, blk, block);
if (ret == -EEXIST) { /* other ctx has inserted it */
write_unlock_irqrestore(&sd->tree_lock, flag);
free_block(block);
goto find;
} else if (ret == -ENOMEM) { /* NOMEM during insert */
write_unlock_irqrestore(&sd->tree_lock, flag);
free_block(block);
goto out;
}

/* insert successfuly */

get_block(block);
write_unlock_irqrestore(&sd->tree_lock, flag);
out:
return block;
}

void put_block(struct block *block)
{
struct sd *sd = block->sd;

lock_block(block);
if (block->status == PART_IN_LOCAL) {
atomic_dec(&block->ref);
unlock_block(block);
return;
}

if (atomic_dec_and_lock(&block->ref, &sd->tree_lock)) {
radix_tree_delete(&sd->blk_tree, block->blk_no);
set_bitmap(sd, block);
spin_unlock(&sd->tree_lock);

unlock_block(block);
free_block(block);
return;
}
unlock(block);
}