Re: [PATCH 2/2] btrfs: Do first key check under proper lock context

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

On 2018年04月12日 18:59, Nikolay Borisov wrote:
> 
> 
> On 12.04.2018 13:00, Qu Wenruo wrote:
>> The original first key check is handled in
>> btree_read_extent_buffer_pages(), which is good enough for level check,
>> but to verify first key, we need at least read lock on the extent
>> buffer, or we can hit some race and cause false alert.
>>
>> Fix it by adding new verifcation function: btrfs_verify_first_key(),
>> which will first check lock context of both parent and child extent
>> buffer, and only when we have proper lock hold (write or read lock if
> So if the extent buffer is not in commit root, i.e it's active buffer in
> current transaction (am I right?)

That's right.

> then we need write lock? It's not entirely clearly from that sentence.

Not really write lock. Either read or write (both spinning and blocking)
is OK here.

> 
>> extent buffer is not in commit root) we will verify the first key.
>>
>> Most of the verification happens in ctree.c after function call like
>> read_tree_block(), read_node_slot() and read_block_for_search(), and
>> only call btrfs_verify_first_key() after we acquire read/write lock.
>>
>> Also, since first_key check is not as good as level and could easily
>> trigger false alerts due to lock context, put it under
>> CONFIG_BTRFS_FS_CHECK_INTEGRITY to avoid damage to end user.
> 
> Why is first_key check not as good as level

Several reasons.

1) Level tree check can prevent loop/nodes underflow caused by level
   mismatch, but first_key can't
   One problem we're trying to solve is, when level 1 node points to
   another level 1 tree, we can still try to read next level tree block.
   Which could underflow level.
   first_key check can't really handle it.

2) first_key needs extra lock to read out needed info, while level
   check doesn't (hence easier and harder to cause false alert)
   For a given tree block, its level won't change either for active eb
   (modified in current trans) or committed eb.
   But for first_key, we need proper lock to avoid race.

For short, first key check needs a lot of carefully reviewed check
timing and error handlers, while most problem could be exposed by level
check far before it hits first_key check.

> and why could it trigger
> false alert, given this patch allegedly implements it under proper lock
> context, thus eliminating false positives.

I mean the old and merged first_key check patch.

Thanks,
Qu

> 
>>
>> Signed-off-by: Qu Wenruo <wqu@suse.com>
>> ---
>>  fs/btrfs/ctree.c       | 69 ++++++++++++++++++++++++++++++++++++++++++
>>  fs/btrfs/ctree.h       | 58 +++++++++++++++++++++++++++++++++++
>>  fs/btrfs/extent-tree.c |  6 ++++
>>  fs/btrfs/qgroup.c      |  5 +++
>>  fs/btrfs/ref-verify.c  |  6 ++++
>>  fs/btrfs/relocation.c  | 13 ++++++++
>>  6 files changed, 157 insertions(+)
>>
>> diff --git a/fs/btrfs/ctree.c b/fs/btrfs/ctree.c
>> index e044b51a2789..57721aead371 100644
>> --- a/fs/btrfs/ctree.c
>> +++ b/fs/btrfs/ctree.c
>> @@ -1649,6 +1649,11 @@ int btrfs_realloc_node(struct btrfs_trans_handle *trans,
>>  
>>  		btrfs_tree_lock(cur);
>>  		btrfs_set_lock_blocking(cur);
>> +		if (btrfs_verify_first_key(cur, parent, i)) {
>> +			err = -EUCLEAN;
>> +			btrfs_tree_unlock(cur);
>> +			free_extent_buffer(cur);
>> +		}
>>  		err = __btrfs_cow_block(trans, root, cur, parent, i,
>>  					&cur, search_start,
>>  					min(16 * blocksize,
>> @@ -1863,6 +1868,12 @@ static noinline int balance_level(struct btrfs_trans_handle *trans,
>>  
>>  		btrfs_tree_lock(child);
>>  		btrfs_set_lock_blocking(child);
>> +		if (btrfs_verify_first_key(child, mid, 0)) {
>> +			ret = -EUCLEAN;
>> +			btrfs_tree_unlock(child);
>> +			free_extent_buffer(child);
>> +			goto enospc;
>> +		}
>>  		ret = btrfs_cow_block(trans, root, child, mid, 0, &child);
>>  		if (ret) {
>>  			btrfs_tree_unlock(child);
>> @@ -1901,6 +1912,10 @@ static noinline int balance_level(struct btrfs_trans_handle *trans,
>>  	if (left) {
>>  		btrfs_tree_lock(left);
>>  		btrfs_set_lock_blocking(left);
>> +		if (btrfs_verify_first_key(left, parent, pslot - 1)) {
>> +			ret = -EUCLEAN;
>> +			goto enospc;
>> +		}
>>  		wret = btrfs_cow_block(trans, root, left,
>>  				       parent, pslot - 1, &left);
>>  		if (wret) {
>> @@ -1916,6 +1931,10 @@ static noinline int balance_level(struct btrfs_trans_handle *trans,
>>  	if (right) {
>>  		btrfs_tree_lock(right);
>>  		btrfs_set_lock_blocking(right);
>> +		if (btrfs_verify_first_key(right, parent, pslot + 1)) {
>> +			ret = -EUCLEAN;
>> +			goto enospc;
>> +		}
>>  		wret = btrfs_cow_block(trans, root, right,
>>  				       parent, pslot + 1, &right);
>>  		if (wret) {
>> @@ -2079,6 +2098,8 @@ static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
>>  
>>  		btrfs_tree_lock(left);
>>  		btrfs_set_lock_blocking(left);
>> +		if (btrfs_verify_first_key(left, parent, pslot - 1))
>> +			goto left_out;
>>  
>>  		left_nr = btrfs_header_nritems(left);
>>  		if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
>> @@ -2119,6 +2140,7 @@ static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
>>  			}
>>  			return 0;
>>  		}
>> +left_out:
>>  		btrfs_tree_unlock(left);
>>  		free_extent_buffer(left);
>>  	}
>> @@ -2134,6 +2156,8 @@ static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
>>  
>>  		btrfs_tree_lock(right);
>>  		btrfs_set_lock_blocking(right);
>> +		if (btrfs_verify_first_key(right, parent, pslot + 1))
>> +			goto right_out;
>>  
>>  		right_nr = btrfs_header_nritems(right);
>>  		if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(fs_info) - 1) {
>> @@ -2174,6 +2198,7 @@ static noinline int push_nodes_for_insert(struct btrfs_trans_handle *trans,
>>  			}
>>  			return 0;
>>  		}
>> +right_out:
>>  		btrfs_tree_unlock(right);
>>  		free_extent_buffer(right);
>>  	}
>> @@ -2868,6 +2893,11 @@ int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root *root,
>>  					p->locks[level] = BTRFS_READ_LOCK;
>>  				}
>>  				p->nodes[level] = b;
>> +				if (btrfs_verify_first_key(b, p->nodes[level + 1],
>> +							   slot)) {
>> +					ret = -EUCLEAN;
>> +					goto done;
>> +				}
>>  			}
>>  		} else {
>>  			p->slots[level] = slot;
>> @@ -2981,6 +3011,12 @@ int btrfs_search_old_slot(struct btrfs_root *root, const struct btrfs_key *key,
>>  				goto done;
>>  			}
>>  
>> +			/*
>> +			 * NOTE: both parent and child go through
>> +			 * tree_mod_log_rewind(), first key check is no
>> +			 * longer consistent. So no btrfs_verify_first_key()
>> +			 * for this read_block_for_search().
>> +			 */
>>  			err = read_block_for_search(root, p, &b, level,
>>  						    slot, key);
>>  			if (err == -EAGAIN)
>> @@ -3753,6 +3789,8 @@ static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
>>  
>>  	btrfs_tree_lock(right);
>>  	btrfs_set_lock_blocking(right);
>> +	if (btrfs_verify_first_key(right, upper, slot + 1))
>> +		goto out_unlock;
>>  
>>  	free_space = btrfs_leaf_free_space(fs_info, right);
>>  	if (free_space < data_size)
>> @@ -3987,6 +4025,10 @@ static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
>>  
>>  	btrfs_tree_lock(left);
>>  	btrfs_set_lock_blocking(left);
>> +	if (btrfs_verify_first_key(left, path->nodes[1], slot - 1)) {
>> +		ret = 1;
>> +		goto out;
>> +	}
>>  
>>  	free_space = btrfs_leaf_free_space(fs_info, left);
>>  	if (free_space < data_size) {
>> @@ -5205,6 +5247,12 @@ int btrfs_search_forward(struct btrfs_root *root, struct btrfs_key *min_key,
>>  		}
>>  
>>  		btrfs_tree_read_lock(cur);
>> +		if (btrfs_verify_first_key(cur, path->nodes[level], slot)) {
>> +			btrfs_tree_read_unlock(cur);
>> +			free_extent_buffer(cur);
>> +			ret = -EUCLEAN;
>> +			goto out;
>> +		}
>>  
>>  		path->locks[level - 1] = BTRFS_READ_LOCK;
>>  		path->nodes[level - 1] = cur;
>> @@ -5232,6 +5280,11 @@ static int tree_move_down(struct btrfs_fs_info *fs_info,
>>  	if (IS_ERR(eb))
>>  		return PTR_ERR(eb);
>>  
>> +	if (btrfs_verify_first_key(eb, path->nodes[*level],
>> +				   path->slots[*level])) {
>> +		free_extent_buffer(eb);
>> +		return -EUCLEAN;
>> +	}
>>  	path->nodes[*level - 1] = eb;
>>  	path->slots[*level - 1] = 0;
>>  	(*level)--;
>> @@ -5786,6 +5839,14 @@ int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
>>  							  BTRFS_READ_LOCK);
>>  			}
>>  			next_rw_lock = BTRFS_READ_LOCK;
>> +			if (btrfs_verify_first_key(next, path->nodes[level],
>> +						   slot)) {
>> +				ret = -EUCLEAN;
>> +				btrfs_tree_read_unlock(next);
>> +				free_extent_buffer(next);
>> +				btrfs_release_path(path);
>> +				goto done;
>> +			}
>>  		}
>>  		break;
>>  	}
>> @@ -5823,6 +5884,14 @@ int btrfs_next_old_leaf(struct btrfs_root *root, struct btrfs_path *path,
>>  							  BTRFS_READ_LOCK);
>>  			}
>>  			next_rw_lock = BTRFS_READ_LOCK;
>> +			if (btrfs_verify_first_key(next, path->nodes[level],
>> +						   0)) {
>> +				ret = -EUCLEAN;
>> +				btrfs_tree_read_unlock(next);
>> +				free_extent_buffer(next);
>> +				btrfs_release_path(path);
>> +				goto done;
>> +			}
>>  		}
>>  	}
>>  	ret = 0;
>> diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
>> index 0eb55825862a..033d8ae027c3 100644
>> --- a/fs/btrfs/ctree.h
>> +++ b/fs/btrfs/ctree.h
>> @@ -44,6 +44,7 @@
>>  #include "extent_io.h"
>>  #include "extent_map.h"
>>  #include "async-thread.h"
>> +#include "print-tree.h"
>>  
>>  struct btrfs_trans_handle;
>>  struct btrfs_transaction;
>> @@ -3752,4 +3753,61 @@ static inline int btrfs_is_testing(struct btrfs_fs_info *fs_info)
>>  #endif
>>  	return 0;
>>  }
>> +
>> +/*
>> + * Verify the first key of an extent buffer matches with its parent pointer.
>> + *
>> + * For tree blocks in current transaction, caller should hold at least read lock
>> + * for both @eb and @parent to avoid race.
>> + * While for tree blocks in commit root, no such requirement.
>> + */
>> +static inline int btrfs_verify_first_key(struct extent_buffer *eb,
>> +					 struct extent_buffer *parent, int slot)
>> +{
>> +#ifdef CONFIG_BTRFS_FS_CHECK_INTEGRITY
>> +	struct btrfs_key parent_key;
>> +	struct btrfs_key found_key;
>> +	struct btrfs_fs_info *fs_info;
>> +	u64 transid = 0;
>> +	int ret;
>> +
>> +	if (!parent || !eb)
>> +		return 0;
>> +	fs_info = parent->fs_info;
>> +	transid = fs_info->last_trans_committed;
>> +	/*
>> +	 * For first_key check, we need to ensure we have proper lock hold,
>> +	 * or the content may change and cause false alert.
>> +	 */
>> +	if (btrfs_header_generation(eb) > transid &&
>> +	    !atomic_read(&eb->read_locks) && !atomic_read(&eb->write_locks)) {
>> +		WARN_ON(1);
>> +		return 0;
>> +	}
>> +	if (btrfs_header_generation(parent) > transid &&
>> +	    !atomic_read(&parent->read_locks) &&
>> +	    !atomic_read(&parent->write_locks)) {
>> +		WARN_ON(1);
>> +		return 0;
>> +	}
>> +	btrfs_node_key_to_cpu(parent, &parent_key, slot);
>> +	if (btrfs_header_level(eb))
>> +		btrfs_node_key_to_cpu(eb, &found_key, 0);
>> +	else
>> +		btrfs_item_key_to_cpu(eb, &found_key, 0);
>> +	ret =  btrfs_comp_cpu_keys(&found_key, &parent_key);
>> +	if (ret) {
>> +		WARN_ON(ret);
>> +		btrfs_err(fs_info,
>> +"tree first key mismatch detected, bytenr=%llu key expected=(%llu, %u, %llu) has=(%llu, %u, %llu)",
>> +			  eb->start, parent_key.objectid, parent_key.type,
>> +			  parent_key.offset, found_key.objectid,
>> +			  found_key.type, found_key.offset);
>> +	}
>> +	return ret;
>> +#else
>> +	return 0;
>> +#endif /* CONFIG_BTRFS_FS_CHECK_INTEGRITY */
>> +}
>> +
>>  #endif
>> diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
>> index f080834e4dd0..17bd22a54281 100644
>> --- a/fs/btrfs/extent-tree.c
>> +++ b/fs/btrfs/extent-tree.c
>> @@ -8807,6 +8807,12 @@ static noinline int do_walk_down(struct btrfs_trans_handle *trans,
>>  		}
>>  		btrfs_tree_lock(next);
>>  		btrfs_set_lock_blocking(next);
>> +		if (btrfs_verify_first_key(next, path->nodes[level],
>> +					   path->slots[level])) {
>> +			btrfs_tree_unlock(next);
>> +			free_extent_buffer(next);
>> +			goto out_unlock;
>> +		}
>>  	}
>>  
>>  	level--;
>> diff --git a/fs/btrfs/qgroup.c b/fs/btrfs/qgroup.c
>> index c7f129356966..4c53954297d8 100644
>> --- a/fs/btrfs/qgroup.c
>> +++ b/fs/btrfs/qgroup.c
>> @@ -1744,6 +1744,11 @@ int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
>>  
>>  			btrfs_tree_read_lock(eb);
>>  			btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
>> +			if (btrfs_verify_first_key(eb, path->nodes[level + 1],
>> +						   parent_slot)) {
>> +				ret = -EUCLEAN;
>> +				goto out;
>> +			}
>>  			path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
>>  
>>  			ret = btrfs_qgroup_trace_extent(trans, fs_info,
>> diff --git a/fs/btrfs/ref-verify.c b/fs/btrfs/ref-verify.c
>> index 6d856eabf34a..b996ca1fd9dc 100644
>> --- a/fs/btrfs/ref-verify.c
>> +++ b/fs/btrfs/ref-verify.c
>> @@ -593,6 +593,12 @@ static int walk_down_tree(struct btrfs_root *root, struct btrfs_path *path,
>>  			}
>>  			btrfs_tree_read_lock(eb);
>>  			btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
>> +			if (btrfs_verify_first_key(eb, path->nodes[level],
>> +			    path->slots[level])) {
>> +				btrfs_tree_read_unlock_blocking(eb);
>> +				free_extent_buffer(eb);
>> +				return -EUCLEAN;
>> +			}
>>  			path->nodes[level-1] = eb;
>>  			path->slots[level-1] = 0;
>>  			path->locks[level-1] = BTRFS_READ_LOCK_BLOCKING;
>> diff --git a/fs/btrfs/relocation.c b/fs/btrfs/relocation.c
>> index 74db4694d776..74a088fbdac4 100644
>> --- a/fs/btrfs/relocation.c
>> +++ b/fs/btrfs/relocation.c
>> @@ -1888,6 +1888,13 @@ int replace_path(struct btrfs_trans_handle *trans,
>>  				break;
>>  			}
>>  			btrfs_tree_lock(eb);
>> +			if (btrfs_verify_first_key(eb, parent, slot)) {
>> +				btrfs_tree_unlock(eb);
>> +				free_extent_buffer(eb);
>> +				ret = -EUCLEAN;
>> +				break;
>> +			}
>> +
>>  			if (cow) {
>>  				ret = btrfs_cow_block(trans, dest, eb, parent,
>>  						      slot, &eb);
>> @@ -2793,6 +2800,12 @@ static int do_relocation(struct btrfs_trans_handle *trans,
>>  		}
>>  		btrfs_tree_lock(eb);
>>  		btrfs_set_lock_blocking(eb);
>> +		if (btrfs_verify_first_key(eb, upper->eb, slot)) {
>> +			btrfs_tree_unlock(eb);
>> +			free_extent_buffer(eb);
>> +			err = -EUCLEAN;
>> +			goto next;
>> +		}
>>  
>>  		if (!node->eb) {
>>  			ret = btrfs_cow_block(trans, root, eb, upper->eb,
>>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>