On 2020/7/6 下午9:41, Josef Bacik wrote:
> On 7/3/20 2:19 AM, Qu Wenruo wrote:
>> [PROBLEM]
>> There are known problem related to how btrfs handles qgroup reserved
>> space.
>> One of the most obvious case is the the test case btrfs/153, which do
>> fallocate, then write into the preallocated range.
>>
>> btrfs/153 1s ... - output mismatch (see
>> xfstests-dev/results//btrfs/153.out.bad)
>> --- tests/btrfs/153.out 2019-10-22 15:18:14.068965341 +0800
>> +++ xfstests-dev/results//btrfs/153.out.bad 2020-07-01
>> 20:24:40.730000089 +0800
>> @@ -1,2 +1,5 @@
>> QA output created by 153
>> +pwrite: Disk quota exceeded
>> +/mnt/scratch/testfile2: Disk quota exceeded
>> +/mnt/scratch/testfile2: Disk quota exceeded
>> Silence is golden
>> ...
>> (Run 'diff -u xfstests-dev/tests/btrfs/153.out
>> xfstests-dev/results//btrfs/153.out.bad' to see the entire diff)
>>
>> [CAUSE]
>> Since commit c6887cd11149 ("Btrfs: don't do nocow check unless we have
>> to"),
>> we always reserve space no matter if it's COW or not.
>>
>> Such behavior change is mostly for performance, and reverting it is not
>> a good idea anyway.
>>
>> For preallcoated extent, we reserve qgroup data space for it already,
>> and since we also reserve data space for qgroup at buffered write time,
>> it needs twice the space for us to write into preallocated space.
>>
>> This leads to the -EDQUOT in buffered write routine.
>>
>> And we can't follow the same solution, unlike data/meta space check,
>> qgroup reserved space is shared between data/meta.
>> The EDQUOT can happen at the metadata reservation, so doing NODATACOW
>> check after qgroup reservation failure is not a solution.
>>
>> [FIX]
>> To solve the problem, we don't return -EDQUOT directly, but every time
>> we got a -EDQUOT, we try to flush qgroup space by:
>> - Flush all inodes of the root
>> NODATACOW writes will free the qgroup reserved at run_dealloc_range().
>> However we don't have the infrastructure to only flush NODATACOW
>> inodes, here we flush all inodes anyway.
>>
>> - Wait ordered extents
>> This would convert the preallocated metadata space into per-trans
>> metadata, which can be freed in later transaction commit.
>>
>> - Commit transaction
>> This will free all per-trans metadata space.
>>
>> Also we don't want to trigger flush too racy, so here we introduce a
>> per-root mutex to ensure if there is a running qgroup flushing, we wait
>> for it to end and don't start re-flush.
>>
>> Fixes: c6887cd11149 ("Btrfs: don't do nocow check unless we have to")
>> Signed-off-by: Qu Wenruo <wqu@xxxxxxxx>
>> ---
>> fs/btrfs/ctree.h | 1 +
>> fs/btrfs/disk-io.c | 1 +
>> fs/btrfs/qgroup.c | 118 ++++++++++++++++++++++++++++++++++++++++-----
>> 3 files changed, 108 insertions(+), 12 deletions(-)
>>
>> diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
>> index 4dd478b4fe3a..891f47c7891f 100644
>> --- a/fs/btrfs/ctree.h
>> +++ b/fs/btrfs/ctree.h
>> @@ -1162,6 +1162,7 @@ struct btrfs_root {
>> spinlock_t qgroup_meta_rsv_lock;
>> u64 qgroup_meta_rsv_pertrans;
>> u64 qgroup_meta_rsv_prealloc;
>> + struct mutex qgroup_flushing_mutex;
>> /* Number of active swapfiles */
>> atomic_t nr_swapfiles;
>> diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
>> index c27022f13150..0116e0b487c9 100644
>> --- a/fs/btrfs/disk-io.c
>> +++ b/fs/btrfs/disk-io.c
>> @@ -1116,6 +1116,7 @@ static void __setup_root(struct btrfs_root
>> *root, struct btrfs_fs_info *fs_info,
>> mutex_init(&root->log_mutex);
>> mutex_init(&root->ordered_extent_mutex);
>> mutex_init(&root->delalloc_mutex);
>> + mutex_init(&root->qgroup_flushing_mutex);
>> init_waitqueue_head(&root->log_writer_wait);
>> init_waitqueue_head(&root->log_commit_wait[0]);
>> init_waitqueue_head(&root->log_commit_wait[1]);
>> diff --git a/fs/btrfs/qgroup.c b/fs/btrfs/qgroup.c
>> index 398de1145d2b..bf9076bc33eb 100644
>> --- a/fs/btrfs/qgroup.c
>> +++ b/fs/btrfs/qgroup.c
>> @@ -3481,13 +3481,14 @@ static int qgroup_revert(struct btrfs_inode
>> *inode,
>> int clear_ret;
>> entry = rb_entry(n, struct ulist_node, rb_node);
>> - if (entry->val >= start + len)
>> - break;
>> - if (entry->val + entry->aux <= start)
>> - goto next;
>> start = entry->val;
>> end = entry->aux;
>> len = end - start + 1;
>> +
>> + if (start >= start + len)
>> + break;
>> + if (start + len <= start)
>> + goto next;
>
> This is unrelated to the change at hand, put this in the 1st patch.
>
>> /*
>> * Now the entry is in [start, start + len), revert the
>> * EXTENT_QGROUP_RESERVED bit.
>> @@ -3512,17 +3513,62 @@ static int qgroup_revert(struct btrfs_inode
>> *inode,
>> }
>> /*
>> - * Reserve qgroup space for range [start, start + len).
>> + * Try to free some space for qgroup.
>> *
>> - * This function will either reserve space from related qgroups or doing
>> - * nothing if the range is already reserved.
>> + * For qgroup, there are only 3 ways to free qgroup space:
>> + * - Flush nodatacow write
>> + * Any nodatacow write will free its reserved data space at
>> + * run_delalloc_range().
>> + * In theory, we should only flush nodatacow inodes, but it's
>> + * not yet possible, so we need to flush the whole root.
>> *
>> - * Return 0 for successful reserve
>> - * Return <0 for error (including -EQUOT)
>> + * - Wait for ordered extents
>> + * When ordered extents are finished, their reserved metadata
>> + * is finally converted to per_trans status, which can be freed
>> + * by later commit transaction.
>> *
>> - * NOTE: this function may sleep for memory allocation.
>> + * - Commit transaction
>> + * This would free the meta_per_trans space.
>> + * In theory this shouldn't provide much space, but any more qgroup
>> space
>> + * is needed.
>> */
>> -int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
>> +static int try_flush_qgroup(struct btrfs_root *root)
>> +{
>> + struct btrfs_trans_handle *trans;
>> + int ret;
>> +
>> + if (!is_fstree(root->root_key.objectid))
>> + return 0;
>> +
>
> This will just retry the reservation again without having done
> anything. Is it even possible to end up here with a !fs_tree?
You're right, it's impossible.
The existing btrfs_qgroup_reserve_data() has extra check for root
objectid, and just return 0 for non fs trees.
So it's impossible to hit the !is_fstree() branch.
> If it is
> then maybe we should return -EDQUOT here so we don't retry the
> reservation again.
>
>> + /*
>> + * We don't want to run flush again and again, so if there is a
>> running
>> + * one, we won't try to start a new flush, but exit directly.
>> + */
>> + ret = mutex_trylock(&root->qgroup_flushing_mutex);
>> + if (!ret) {
>> + mutex_lock(&root->qgroup_flushing_mutex);
>> + mutex_unlock(&root->qgroup_flushing_mutex);
>> + return 0;
>> + }
>> +
>> + ret = btrfs_start_delalloc_snapshot(root);
>> + if (ret < 0)
>> + goto unlock;
>> + btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
>> +
>> + trans = btrfs_join_transaction(root);
>> + if (IS_ERR(trans)) {
>> + ret = PTR_ERR(trans);
>> + goto unlock;
>> + }
>> +
>> + ret = btrfs_commit_transaction(trans);
>> +unlock:
>> + mutex_unlock(&root->qgroup_flushing_mutex);
>> + return ret;
>> +}
>
> I still hate this but implementing tickets for quotas probably doesn't
> make sense right now.
Yep. But I'm pretty open to integrating this into the tickets system.
Thanks,
Qu
>
> <snip>
>
>> +
>> + ret = qgroup_reserve_meta(root, num_bytes, type, enforce);
>> + if (!ret)
>> + return ret;
>> + if (ret < 0 && ret != -EDQUOT)
>> + return ret;
>
> if (ret <= 0 && ret != -EDQUOT)
> return ret;
>
> Thanks,
>
> Josef
>
