- 07 Jan, 2022 6 commits
-
-
Filipe Manana authored
Right now generic_bin_search() always uses a low boundary slot of 0, but in the next patch we'll want to often skip slot 0 when searching for a key. So make generic_bin_search() have the low boundary slot specified as an argument, and move the check for the extent buffer level from btrfs_bin_search() to generic_bin_search() to avoid adding another wrapper around generic_bin_search(). Reviewed-by:
Josef Bacik <josef@toxicpanda.com> Signed-off-by:
Filipe Manana <fdmanana@suse.com> Reviewed-by:
David Sterba <dsterba@suse.com> Signed-off-by:
-
Josef Bacik authored
Now that we clear the extent buffer uptodate if we fail to write it out we need to check to see if our root node is uptodate before we search down it. Otherwise we could return stale data (or potentially corrupt data that was caught by the write verification step) and think that the path is OK to search down. CC: stable@vger.kernel.org # 5.4+ Reviewed-by:
Nikolay Borisov <nborisov@suse.com> Signed-off-by:
-
Nikolay Borisov authored
Currently paused balance precludes adding a device since they are both considered exclusive ops and we can have at most one running at a time. This is problematic in case a filesystem encounters an ENOSPC situation while balance is running, in this case the only thing the user can do is mount the fs with "skip_balance" which pauses balance and delete some data to free up space for balance. However, it should be possible to add a new device when balance is paused. Fix this by allowing device add to proceed when balance is paused. Signed-off-by:
-
Nikolay Borisov authored
This is needed to enable device add to work in cases when a file system has been mounted with 'skip_balance' mount option. Signed-off-by:
-
Nikolay Borisov authored
Current set of exclusive operation states is not sufficient to handle all practical use cases. In particular there is a need to be able to add a device to a filesystem that have paused balance. Currently there is no way to distinguish between a running and a paused balance. Fix this by introducing BTRFS_EXCLOP_BALANCE_PAUSED which is going to be set in 2 occasions: 1. When a filesystem is mounted with skip_balance and there is an unfinished balance it will now be into BALANCE_PAUSED instead of simply BALANCE state. 2. When a running balance is paused. Signed-off-by:
-
Filipe Manana authored
We don't allow send and balance/relocation to run in parallel in order to prevent send failing or silently producing some bad stream. This is because while send is using an extent (specially metadata) or about to read a metadata extent and expecting it belongs to a specific parent node, relocation can run, the transaction used for the relocation is committed and the extent gets reallocated while send is still using the extent, so it ends up with a different content than expected. This can result in just failing to read a metadata extent due to failure of the validation checks (parent transid, level, etc), failure to find a backreference for a data extent, and other unexpected failures. Besides reallocation, there's also a similar problem of an extent getting discarded when it's unpinned after the transaction used for block group relocation is committed. The restriction between balance and send was added in commit 9e967495 ("Btrfs: prevent send failures and crashes due to concurrent relocation"), kernel 5.3, while the more general restriction between send and relocation was added in commit 1cea5cf0 ("btrfs: ensure relocation never runs while we have send operations running"), kernel 5.14. Both send and relocation can be very long running operations. Relocation because it has to do a lot of IO and expensive backreference lookups in case there are many snapshots, and send due to read IO when operating on very large trees. This makes it inconvenient for users and tools to deal with scheduling both operations. For zoned filesystem we also have automatic block group relocation, so send can fail with -EAGAIN when users least expect it or send can end up delaying the block group relocation for too long. In the future we might also get the automatic block group relocation for non zoned filesystems. This change makes it possible for send and relocation to run in parallel. This is achieved the following way: 1) For all tree searches, send acquires a read lock on the commit root semaphore; 2) After each tree search, and before releasing the commit root semaphore, the leaf is cloned and placed in the search path (struct btrfs_path); 3) After releasing the commit root semaphore, the changed_cb() callback is invoked, which operates on the leaf and writes commands to the pipe (or file in case send/receive is not used with a pipe). It's important here to not hold a lock on the commit root semaphore, because if we did we could deadlock when sending and receiving to the same filesystem using a pipe - the send task blocks on the pipe because it's full, the receive task, which is the only consumer of the pipe, triggers a transaction commit when attempting to create a subvolume or reserve space for a write operation for example, but the transaction commit blocks trying to write lock the commit root semaphore, resulting in a deadlock; 4) Before moving to the next key, or advancing to the next change in case of an incremental send, check if a transaction used for relocation was committed (or is about to finish its commit). If so, release the search path(s) and restart the search, to where we were before, so that we don't operate on stale extent buffers. The search restarts are always possible because both the send and parent roots are RO, and no one can add, remove of update keys (change their offset) in RO trees - the only exception is deduplication, but that is still not allowed to run in parallel with send; 5) Periodically check if there is contention on the commit root semaphore, which means there is a transaction commit trying to write lock it, and release the semaphore and reschedule if there is contention, so as to avoid causing any significant delays to transaction commits. This leaves some room for optimizations for send to have less path releases and re searching the trees when there's relocation running, but for now it's kept simple as it performs quite well (on very large trees with resulting send streams in the order of a few hundred gigabytes). Test case btrfs/187, from fstests, stresses relocation, send and deduplication attempting to run in parallel, but without verifying if send succeeds and if it produces correct streams. A new test case will be added that exercises relocation happening in parallel with send and then checks that send succeeds and the resulting streams are correct. A final note is that for now this still leaves the mutual exclusion between send operations and deduplication on files belonging to a root used by send operations. A solution for that will be slightly more complex but it will eventually be built on top of this change. Signed-off-by:
-
- 03 Jan, 2022 34 commits
-
-
Nikolay Borisov authored
btrfs_free_space_ctl::private is either unset or it always points to struct btrfs_block_group when it is set. So there's no point in keeping the unhelpful 'private' name and keeping it an untyped pointer. Change both the type and name to be self-describing. No functional changes. Signed-off-by:
-
Nikolay Borisov authored
There is no point in the function taking an fs_info and a btrfs_free_space because the ctl passed always belongs to the block group. Furthermore fs_info can be referenced from the block group. No functional changes. Signed-off-by:
-
Nikolay Borisov authored
The only difference between the two is whether btrfs_free_space::bytes is adjusted. Instead of having 2 separate functions control this behavior via an additional parameter and make them one function instead. No functional changes. Signed-off-by:
-
Nikolay Borisov authored
The only difference is the former adjusts btrfs_free_space::bytes member. Consolidate the two function into 1 and add a bool parameter which controls whether the adjustment is made or not. No functional changes. Signed-off-by:
-
Josef Bacik authored
In the future we are going to have multiple copies of these trees. To facilitate this we need a way to lookup the different roots we are looking for. Handle this by adding a global root rb tree that is indexed on the root->root_key. Then instead of loading the roots at mount time with individually targeted keys, simply search the tree_root for anything with the specific objectid we want. This will make it straightforward to support both old style and new style file systems. Signed-off-by:
-
Josef Bacik authored
We don't set SHAREABLE on the extent root, we don't need to have this safety check here. Signed-off-by:
-
Josef Bacik authored
We're going to have multiple free space roots in the future, so adjust all the users of the free space root to use a helper to access the root. Signed-off-by:
-
Josef Bacik authored
We are going to have multiple csum roots in the future, so convert all users of ->csum_root to btrfs_csum_root() and rename ->csum_root to ->_csum_root so we can easily find remaining users in the future. Signed-off-by:
-
Josef Bacik authored
We have a few places where we skip doing csums if we mounted with one of the rescue options that ignores bad csum roots. In the future when there are multiple csum roots it'll be costly to check and see if there are any missing csum roots, so simply add a flag to indicate the fs should skip loading csums in case of errors. Signed-off-by:
-
Josef Bacik authored
In the future we may have multiple csum roots, so simply check the objectid is for a csum root instead of checking against ->csum_root. Signed-off-by:
-
Josef Bacik authored
When we start having multiple extent roots we'll need to use a helper to get to the correct extent_root. Rename fs_info->extent_root to _extent_root and convert all of the users of the extent root to using the btrfs_extent_root() helper. This will allow us to easily clean up the remaining direct accesses in the future. Signed-off-by:
-
Josef Bacik authored
In the future we're going to have multiple csum and extent root trees, so init the roots block_rsv at setup_root time based on their root key objectid. Signed-off-by:
-
Josef Bacik authored
We only need the root to start a transaction, and since it's a global root we can pick anything, change to the tree_root as we'll have a lot of extent roots in the future. Signed-off-by:
-
Josef Bacik authored
We are going to have many extent_roots soon, and we don't need a root here necessarily as we're not modifying anything, we're just getting the trans handle so we can have an accurate view of references, so use the tree_root here. Signed-off-by:
-
Josef Bacik authored
We're just using the extent_root to set the chunk owner to root_key->objectid, which is BTRFS_EXTENT_TREE_OBJECTID, so use that directly instead of using the root. Signed-off-by:
-
Josef Bacik authored
We only defrag leaves on roots that have SHAREABLE set, so we don't need to check if we're the extent root as it doesn't have SHAREABLE set. Signed-off-by:
-
Josef Bacik authored
This is a leftover from when we used to independently swap the extent root's commit root and the fs tree commit roots. At the time I simply changed the helper to a list_add. There's actually no reason to not add the extent root to the switch commit root at this point, we don't care about the order we do the switching since it's all done under the commit_root_sem. If we re-mark the extent root dirty after adding it to the switch_commits list we'll see that BTRFS_ROOT_DIRTY isn't set and then list_move it back onto the dirty list, and then we'll redo the tree update and everything will be ok. Signed-off-by:
-
Josef Bacik authored
We're only using this to start the transaction with to possibly allocate a chunk. It doesn't really matter which root to use, but with extent tree v2 we'll need a bytenr to look up a extent root which makes the usage of the extent_root awkward here. Simply change it to the chunk_root. Signed-off-by:
-
Josef Bacik authored
With extent tree v2 we'll have a different extent root based on where the bytenr is located, so adjust the remove_extent_backref() helper and it's helpers to pass the extent_root around. Signed-off-by:
-
Josef Bacik authored
With extent tree v2 we will have a separate root to hold the block group items. Add a btrfs_block_group_root() that will return the appropriate root given the flags of the fs, and convert all functions that need to modify block group items to use the helper. Signed-off-by:
-
Josef Bacik authored
If we're looking for leafs that point to a data extent we want to record the extent items that point at our bytenr. At this point we have the reference and we know for a fact that this leaf should have a reference to our bytenr. However if there's some sort of corruption we may not find any references to our leaf, and thus could end up with eie == NULL. Replace this BUG_ON() with an ASSERT() and then return -EUCLEAN for the mortals. Signed-off-by:
-
Josef Bacik authored
We search for an extent entry with .offset = -1, which shouldn't be a thing, but corruption happens. Add an ASSERT() for the developers, return -EUCLEAN for mortals. Signed-off-by:
-
Josef Bacik authored
We define __TRANS_DUMMY always, so this extra ifdef stuff is not needed. Signed-off-by:
-
Josef Bacik authored
This comment was much closer to the related code when it was originally added, but has slowly migrated north far from its ancestral lands. Move it back down with its people. Signed-off-by:
-
Josef Bacik authored
We pass in the path, but use btrfs_next_item() using the root we searched with. Pass the root down to add_keyed_refs() instead of the fs_info so we can continue to use the same root we searched with. Signed-off-by:
-
Josef Bacik authored
Nobody is using this anymore, remove it. Signed-off-by:
-
Josef Bacik authored
The root on the trans->root can be anything, and generally we're committing from the transaction kthread so it's usually the tree_root. Change this to just take an fs_info, and to maintain compatibility simply put the ROOT_TREE_OBJECTID as the root objectid for the tracepoint. This will allow use to remove trans->root. Signed-off-by:
-
Josef Bacik authored
Currently we do this awful thing where we get another ref on a trans handle, async off that handle and commit the transaction from that work. Because we do this we have to mess with current->journal_info and the freeze counting stuff. We already have an async thing to kick for the transaction commit, the transaction kthread. Replace this work struct with a flag on the fs_info to tell the kthread to go ahead and commit even if it's before our timeout. Then we can drastically simplify the async transaction commit path. Note: this can be simplified and functionality based on the pending operation COMMIT. Signed-off-by:
-
Josef Bacik authored
This is no longer used, the -o nobarrier is handled by BTRFS_MOUNT_NOBARRIER. Remove the flag. Signed-off-by:
-
Nikolay Borisov authored
Reshuffle the code inside the first loop of tree_search_offset so that one if() is eliminated and the becomes more linear. Reviewed-by:
-
Qu Wenruo authored
[BUG] When debugging calc_bio_boundaries(), I found that even for RAID1 metadata, we're following stripe length to calculate stripe boundary. # mkfs.btrfs -m raid1 -d raid1 /dev/test/scratch[12] # mount /dev/test/scratch /mnt/btrfs # xfs_io -f -c "pwrite 0 64K" /mnt/btrfs/file # umount Above very basic operations will make calc_bio_boundaries() to report the following result: submit_extent_page: r/i=1/1 file_offset=22036480 len_to_stripe_boundary=49152 submit_extent_page: r/i=1/1 file_offset=30474240 len_to_stripe_boundary=65536 ... submit_extent_page: r/i=1/1 file_offset=30523392 len_to_stripe_boundary=16384 submit_extent_page: r/i=1/1 file_offset=30457856 len_to_stripe_boundary=16384 submit_extent_page: r/i=5/257 file_offset=0 len_to_stripe_boundary=65536 submit_extent_page: r/i=5/257 file_offset=65536 len_to_stripe_boundary=65536 submit_extent_page: r/i=1/1 file_offset=30490624 len_to_stripe_boundary=49152 submit_extent_page: r/i=1/1 file_offset=30507008 len_to_stripe_boundary=32768 Where "r/i" is the rootid and inode, 1/1 means they metadata. The remaining names match the member used in kernel. Even all data/metadata are using RAID1, we're still following stripe length. [CAUSE] This behavior is caused by a wrong condition in btrfs_get_io_geometry(): if (map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) { /* Fill using stripe_len */ len = min_t(u64, em->len - offset, max_len); } else { len = em->len - offset; } This means, only for SINGLE we will not follow stripe_len. However for profiles like RAID1*, DUP, they don't need to bother stripe_len. This can lead to unnecessary bio split for RAID1*/DUP profiles, and can even be a blockage for future zoned RAID support. [FIX] Introduce one single-use macro, BTRFS_BLOCK_GROUP_STRIPE_MASK, and change the condition to only calculate the length using stripe length for stripe based profiles. Reviewed-by:Johannes Thumshirn <johannes.thumshirn@wdc.com> Reviewed-by:
Anand Jain <anand.jain@oracle.com> Signed-off-by:
Qu Wenruo <wqu@suse.com> Signed-off-by:
-
Nikolay Borisov authored
This is a small optimisation since the currently 'entry' is already checked in the if () {} else if {} construct above the loop. In essence the first iteration of the final while loop is redundant. To eliminate this extra check simply get the next entry at the beginning of the loop. Reviewed-by: -
Josef Bacik authored
I noticed a few corner cases when looking at my bytes_index patch for obvious bugs, so add a bunch of tests to validate proper behavior of the bytes_index tree. A couple of basic tests to make sure it puts things in the correct order, and then more complicated tests to make sure it re-arranges bitmap entries properly and does the right thing when we try to make allocations. Signed-off-by:
-
Josef Bacik authored
Currently we index free space on offset only, because usually we have a hint from the allocator that we want to honor for locality reasons. However if we fail to use this hint we have to go back to a brute force search through the free space entries to find a large enough extent. With sufficiently fragmented free space this becomes quite expensive, as we have to linearly search all of the free space entries to find if we have a part that's long enough. To fix this add a cached rb tree to index based on free space entry bytes. This will allow us to quickly look up the largest chunk in the free space tree for this block group, and stop searching once we've found an entry that is too small to satisfy our allocation. We simply choose to use this tree if we're searching from the beginning of the block group, as we know we do not care about locality at that point. I wrote an allocator test that creates a 10TiB ram backed null block device and then fallocates random files until the file system is full. I think go through and delete all of the odd files. Then I spawn 8 threads that fallocate 64MiB files (1/2 our extent size cap) until the file system is full again. I use bcc's funclatency to measure the latency of find_free_extent. The baseline results are nsecs : count distribution 0 -> 1 : 0 | | 2 -> 3 : 0 | | 4 -> 7 : 0 | | 8 -> 15 : 0 | | 16 -> 31 : 0 | | 32 -> 63 : 0 | | 64 -> 127 : 0 | | 128 -> 255 : 0 | | 256 -> 511 : 10356 |**** | 512 -> 1023 : 58242 |************************* | 1024 -> 2047 : 74418 |******************************** | 2048 -> 4095 : 90393 |****************************************| 4096 -> 8191 : 79119 |*********************************** | 8192 -> 16383 : 35614 |*************** | 16384 -> 32767 : 13418 |***** | 32768 -> 65535 : 12811 |***** | 65536 -> 131071 : 17090 |******* | 131072 -> 262143 : 26465 |*********** | 262144 -> 524287 : 40179 |***************** | 524288 -> 1048575 : 55469 |************************ | 1048576 -> 2097151 : 48807 |********************* | 2097152 -> 4194303 : 26744 |*********** | 4194304 -> 8388607 : 35351 |*************** | 8388608 -> 16777215 : 13918 |****** | 16777216 -> 33554431 : 21 | | avg = 908079 nsecs, total: 580889071441 nsecs, count: 639690 And the patch results are nsecs : count distribution 0 -> 1 : 0 | | 2 -> 3 : 0 | | 4 -> 7 : 0 | | 8 -> 15 : 0 | | 16 -> 31 : 0 | | 32 -> 63 : 0 | | 64 -> 127 : 0 | | 128 -> 255 : 0 | | 256 -> 511 : 6883 |** | 512 -> 1023 : 54346 |********************* | 1024 -> 2047 : 79170 |******************************** | 2048 -> 4095 : 98890 |****************************************| 4096 -> 8191 : 81911 |********************************* | 8192 -> 16383 : 27075 |********** | 16384 -> 32767 : 14668 |***** | 32768 -> 65535 : 13251 |***** | 65536 -> 131071 : 15340 |****** | 131072 -> 262143 : 26715 |********** | 262144 -> 524287 : 43274 |***************** | 524288 -> 1048575 : 53870 |********************* | 1048576 -> 2097151 : 55368 |********************** | 2097152 -> 4194303 : 41036 |**************** | 4194304 -> 8388607 : 24927 |********** | 8388608 -> 16777215 : 33 | | 16777216 -> 33554431 : 9 | | avg = 623599 nsecs, total: 397259314759 nsecs, count: 637042 There's a little variation in the amount of calls done because of timing of the threads with metadata requirements, but the avg, total, and count's are relatively consistent between runs (usually within 2-5% of each other). As you can see here we have around a 30% decrease in average latency with a 30% decrease in overall time spent in find_free_extent. Reviewed-by:
-
