struct btrfs_delayed_ref_node *ref,
struct btrfs_qgroup_extent_record *qrecord,
u64 bytenr, u64 num_bytes, u64 ref_root, u64 reserved,
- int action, int is_data)
+ int action, int is_data, int *qrecord_inserted_ret)
{
struct btrfs_delayed_ref_head *existing;
struct btrfs_delayed_ref_head *head_ref = NULL;
struct btrfs_delayed_ref_root *delayed_refs;
int count_mod = 1;
int must_insert_reserved = 0;
+ int qrecord_inserted = 0;
/* If reserved is provided, it must be a data extent. */
BUG_ON(!is_data && reserved);
if(btrfs_qgroup_trace_extent_nolock(fs_info,
delayed_refs, qrecord))
kfree(qrecord);
+ else
+ qrecord_inserted = 1;
}
spin_lock_init(&head_ref->lock);
atomic_inc(&delayed_refs->num_entries);
trans->delayed_ref_updates++;
}
+ if (qrecord_inserted_ret)
+ *qrecord_inserted_ret = qrecord_inserted;
return head_ref;
}
struct btrfs_delayed_ref_head *head_ref;
struct btrfs_delayed_ref_root *delayed_refs;
struct btrfs_qgroup_extent_record *record = NULL;
+ int qrecord_inserted;
BUG_ON(extent_op && extent_op->is_data);
ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
* the spin lock
*/
head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
- bytenr, num_bytes, 0, 0, action, 0);
+ bytenr, num_bytes, 0, 0, action, 0,
+ &qrecord_inserted);
add_delayed_tree_ref(fs_info, trans, head_ref, &ref->node, bytenr,
num_bytes, parent, ref_root, level, action);
spin_unlock(&delayed_refs->lock);
+ if (qrecord_inserted)
+ return btrfs_qgroup_trace_extent_post(fs_info, record);
return 0;
free_head_ref:
struct btrfs_delayed_ref_head *head_ref;
struct btrfs_delayed_ref_root *delayed_refs;
struct btrfs_qgroup_extent_record *record = NULL;
+ int qrecord_inserted;
ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
if (!ref)
*/
head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
bytenr, num_bytes, ref_root, reserved,
- action, 1);
+ action, 1, &qrecord_inserted);
add_delayed_data_ref(fs_info, trans, head_ref, &ref->node, bytenr,
num_bytes, parent, ref_root, owner, offset,
action);
spin_unlock(&delayed_refs->lock);
+ if (qrecord_inserted)
+ return btrfs_qgroup_trace_extent_post(fs_info, record);
return 0;
}
add_delayed_ref_head(fs_info, trans, &head_ref->node, NULL, bytenr,
num_bytes, 0, 0, BTRFS_UPDATE_DELAYED_HEAD,
- extent_op->is_data);
+ extent_op->is_data, NULL);
spin_unlock(&delayed_refs->lock);
return 0;
while (node) {
record = rb_entry(node, struct btrfs_qgroup_extent_record,
node);
- ret = btrfs_find_all_roots(NULL, fs_info, record->bytenr, 0,
- &record->old_roots);
+ if (WARN_ON(!record->old_roots))
+ ret = btrfs_find_all_roots(NULL, fs_info,
+ record->bytenr, 0, &record->old_roots);
if (ret < 0)
break;
if (qgroup_to_skip)
return 0;
}
+int btrfs_qgroup_trace_extent_post(struct btrfs_fs_info *fs_info,
+ struct btrfs_qgroup_extent_record *qrecord)
+{
+ struct ulist *old_root;
+ u64 bytenr = qrecord->bytenr;
+ int ret;
+
+ ret = btrfs_find_all_roots(NULL, fs_info, bytenr, 0, &old_root);
+ if (ret < 0)
+ return ret;
+
+ /*
+ * Here we don't need to get the lock of
+ * trans->transaction->delayed_refs, since inserted qrecord won't
+ * be deleted, only qrecord->node may be modified (new qrecord insert)
+ *
+ * So modifying qrecord->old_roots is safe here
+ */
+ qrecord->old_roots = old_root;
+ return 0;
+}
+
int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans,
struct btrfs_fs_info *fs_info, u64 bytenr, u64 num_bytes,
gfp_t gfp_flag)
spin_lock(&delayed_refs->lock);
ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
spin_unlock(&delayed_refs->lock);
- if (ret > 0)
+ if (ret > 0) {
kfree(record);
- return 0;
+ return 0;
+ }
+ return btrfs_qgroup_trace_extent_post(fs_info, record);
}
int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
struct btrfs_fs_info *fs_info);
/*
* Inform qgroup to trace one dirty extent, its info is recorded in @record.
- * So qgroup can account it at commit trans time.
+ * So qgroup can account it at transaction committing time.
*
- * No lock version, caller must acquire delayed ref lock and allocate memory.
+ * No lock version, caller must acquire delayed ref lock and allocated memory,
+ * then call btrfs_qgroup_trace_extent_post() after exiting lock context.
*
* Return 0 for success insert
* Return >0 for existing record, caller can free @record safely.
struct btrfs_delayed_ref_root *delayed_refs,
struct btrfs_qgroup_extent_record *record);
+/*
+ * Post handler after qgroup_trace_extent_nolock().
+ *
+ * NOTE: Current qgroup does the expensive backref walk at transaction
+ * committing time with TRANS_STATE_COMMIT_DOING, this blocks incoming
+ * new transaction.
+ * This is designed to allow btrfs_find_all_roots() to get correct new_roots
+ * result.
+ *
+ * However for old_roots there is no need to do backref walk at that time,
+ * since we search commit roots to walk backref and result will always be
+ * correct.
+ *
+ * Due to the nature of no lock version, we can't do backref there.
+ * So we must call btrfs_qgroup_trace_extent_post() after exiting
+ * spinlock context.
+ *
+ * TODO: If we can fix and prove btrfs_find_all_roots() can get correct result
+ * using current root, then we can move all expensive backref walk out of
+ * transaction committing, but not now as qgroup accounting will be wrong again.
+ */
+int btrfs_qgroup_trace_extent_post(struct btrfs_fs_info *fs_info,
+ struct btrfs_qgroup_extent_record *qrecord);
+
/*
* Inform qgroup to trace one dirty extent, specified by @bytenr and
* @num_bytes.
* So qgroup can account it at commit trans time.
*
- * Better encapsulated version.
+ * Better encapsulated version, with memory allocation and backref walk for
+ * commit roots.
+ * So this can sleep.
*
* Return 0 if the operation is done.
* Return <0 for error, like memory allocation failure or invalid parameter
projects / linux-2.6-block.git / commitdiff