kernel.git/fs/btrfs/delayed-ref.c, branch linux-3.14.y

Btrfs: attach delayed ref updates to delayed ref heads

2014-01-28T21:20:25Z

Currently we have two rb-trees, one for delayed ref heads and one for all of the delayed refs, including the delayed ref heads. When we process the delayed refs we have to hold onto the delayed ref lock for all of the selecting and merging and such, which results in quite a bit of lock contention. This was solved by having a waitqueue and only one flusher at a time, however this hurts if we get a lot of delayed refs queued up. So instead just have an rb tree for the delayed ref heads, and then attach the delayed ref updates to an rb tree that is per delayed ref head. Then we only need to take the delayed ref lock when adding new delayed refs and when selecting a delayed ref head to process, all the rest of the time we deal with a per delayed ref head lock which will be much less contentious. The locking rules for this get a little more complicated since we have to lock up to 3 things to properly process delayed refs, but I will address that problem later. For now this passes all of xfstests and my overnight stress tests. Thanks, Signed-off-by: Josef Bacik Signed-off-by: Chris Mason

Btrfs: skip merge part for delayed data refs

2014-01-28T21:19:23Z

When we have data deduplication on, we'll hang on the merge part because it needs to verify every queued delayed data refs related to this disk offset but we may have millions refs. And in the case of delayed data refs, we don't usually have too much data refs to merge. So it's safe to shut it down for data refs. Signed-off-by: Liu Bo Signed-off-by: Josef Bacik Signed-off-by: Chris Mason

Btrfs: introduce a head ref rbtree

2014-01-28T21:19:22Z

The way how we process delayed refs is 1) get a bunch of head refs, 2) pick up one head ref, 3) go one node back for any delayed ref updates. The head ref is also linked in the same rbtree as the delayed ref is, so in 1) stage, we have to walk one by one including not only head refs, but delayed refs. When we have a great number of delayed refs pending to process, this'll cost time a lot. Here we introduce a head ref specific rbtree, it only has head refs, so troubles go away. Signed-off-by: Liu Bo Signed-off-by: Josef Bacik Signed-off-by: Chris Mason

Btrfs: get rid of sparse warnings

2013-09-01T12:15:50Z

make C=2 fs/btrfs/ CF=-D__CHECK_ENDIAN__ I tried to filter out the warnings for which patches have already been sent to the mailing list, pending for inclusion in btrfs-next. All these changes should be obviously safe. Signed-off-by: Stefan Behrens Signed-off-by: Josef Bacik Signed-off-by: Chris Mason

Btrfs/tracepoint: update delayed ref tracepoints

2013-09-01T11:57:39Z

This shows exactly how btrfs processes the delayed refs onto disks, which is very helpful on understanding delayed ref mechanism and debugging related bugs. Signed-off-by: Liu Bo Signed-off-by: Josef Bacik Signed-off-by: Chris Mason

Btrfs: separate sequence numbers for delayed ref tracking and tree mod log

2013-05-06T19:55:17Z

Sequence numbers for delayed refs have been introduced in the first version of the qgroup patch set. To solve the problem of find_all_roots on a busy file system, the tree mod log was introduced. The sequence numbers for that were simply shared between those two users. However, at one point in qgroup's quota accounting, there's a statement accessing the previous sequence number, that's still just doing (seq - 1) just as it would have to in the very first version. To satisfy that requirement, this patch makes the sequence number counter 64 bit and splits it into a major part (used for qgroup sequence number counting) and a minor part (incremented for each tree modification in the log). This enables us to go exactly one major step backwards, as required for qgroups, while still incrementing the sequence counter for tree mod log insertions to keep track of their order. Keeping them in a single variable means there's no need to change all the code dealing with comparisons of two sequence numbers. The sequence number is reset to 0 on commit (not new in this patch), which ensures we won't overflow the two 32 bit counters. Without this fix, the qgroup tracking can occasionally go wrong and WARN_ONs from the tree mod log code may happen. Signed-off-by: Jan Schmidt Signed-off-by: Josef Bacik

Btrfs: compare relevant parts of delayed tree refs

2013-05-06T19:54:29Z

A user reported a panic while running a balance. What was happening was he was relocating a block, which added the reference to the relocation tree. Then relocation would walk through the relocation tree and drop that reference and free that block, and then it would walk down a snapshot which referenced the same block and add another ref to the block. The problem is this was all happening in the same transaction, so the parent block was free'ed up when we drop our reference which was immediately available for allocation, and then it was used _again_ to add a reference for the same block from a different snapshot. This resulted in something like this in the delayed ref tree add ref to 90234880, parent=2067398656, ref_root 1766, level 1 del ref to 90234880, parent=2067398656, ref_root 18446744073709551608, level 1 add ref to 90234880, parent=2067398656, ref_root 1767, level 1 as you can see the ref_root's don't match, because when we inc the ref we use the header owner, which is the original tree the block belonged to, instead of the data reloc tree. Then when we remove the extent we use the reloc tree objectid. But none of this matters, since it is a shared reference which means only the parent matters. When the delayed ref stuff runs it adds all the increments first, and then does all the drops, to make sure that we don't delete the ref if we net a positive ref count. But tree blocks aren't allowed to have multiple refs from the same block, so this panics when it tries to add the second ref. We need the add and the drop to cancel each other out in memory so we only do the final add. So to fix this we need to adjust how the delayed refs are added to the tree. Only the ref_root matters when it is a normal backref, and only the parent matters when it is a shared backref. So make our decision based on what ref type we have. This allows us to keep the ref_root in memory in case anybody wants to use it for something else, and it allows the delayed refs to be merged properly so we don't end up with this panic. With this patch the users image no longer panics on mount, and it has a clean fsck after a normal mount/umount cycle. Thanks, Cc: stable@vger.kernel.org Reported-by: Roman Mamedov Signed-off-by: Josef Bacik

Btrfs: make delayed ref lock logic more readable

2013-02-20T14:36:41Z

Locking and unlocking delayed ref mutex are in the different functions, and the name of lock functions is not uniform, so the readability is not so good, this patch optimizes the lock logic and makes it more readable. Signed-off-by: Miao Xie Signed-off-by: Josef Bacik

Btrfs: use slabs for delayed reference allocation

2013-02-20T14:36:34Z

The delayed reference allocation is in the fast path of the IO, so use slabs to improve the speed of the allocation. And besides that, it can do check for leaked objects when the module is removed. Signed-off-by: Miao Xie

Btrfs: allow delayed refs to be merged

2012-08-28T20:53:38Z

Daniel Blueman reported a bug with fio+balance on a ramdisk setup. Basically what happens is the balance relocates a tree block which will drop the implicit refs for all of its children and adds a full backref. Once the block is relocated we have to add the implicit refs back, so when we cow the block again we add the implicit refs for its children back. The problem comes when the original drop ref doesn't get run before we add the implicit refs back. The delayed ref stuff will specifically prefer ADD operations over DROP to keep us from freeing up an extent that will have references to it, so we try to add the implicit ref before it is actually removed and we panic. This worked fine before because the add would have just canceled the drop out and we would have been fine. But the backref walking work needs to be able to freeze the delayed ref stuff in time so we have this ever increasing sequence number that gets attached to all new delayed ref updates which makes us not merge refs and we run into this issue. So to fix this we need to merge delayed refs. So everytime we run a clustered ref we need to try and merge all of its delayed refs. The backref walking stuff locks the delayed ref head before processing, so if we have it locked we are safe to merge any refs inside of the sequence number. If there is no sequence number we can merge all refs. Doing this not only fixes our bug but keeps the delayed ref code from adding and removing useless refs and batching together multiple refs into one search instead of one search per delayed ref, which will really help our commit times. I ran this with Daniels test and 276 and I haven't seen any problems. Thanks, Reported-by: Daniel J Blueman Signed-off-by: Josef Bacik