- 22 Oct, 2023 40 commits
-
-
Kent Overstreet authored
The sectors argument shouldn't have been a u32 - it can be up to U32_MAX (i.e. fallocate creating persistent reservations), and if replication is enabled we'll overflow when we calculate the real number of sectors to reserve. Oops. Signed-off-by:
Kent Overstreet <kent.overstreet@gmail.com> Signed-off-by:
Kent Overstreet <kent.overstreet@linux.dev>
-
Kent Overstreet authored
All writes prior to a journal write need to be flushed before the journal write itself happens. On single device filesystems, it suffices to mark the write with REQ_PREFLUSH|REQ_FUA, but on multi device filesystems we need to issue flushes to every device - and wait for them to complete - before issuing the journal writes. Previously, we were issuing flushes to every device, but we weren't waiting for them to complete before issuing the journal writes. Signed-off-by:
-
Kent Overstreet authored
This is because we had a bug where we were writing out journal entries with garbage last_seq, and not catching it. Also, completely ignore jset->last_seq when JSET_NO_FLUSH is true, because of aforementioned bug, but change the write path to set last_seq to 0 when JSET_NO_FLUSH is true. Minor other cleanups and comments. Signed-off-by:
-
Kent Overstreet authored
This fixes a bug where mark and sweep gc incorrectly was clearing out the stripes heap and causing assertions to fire later - simpler to just create the stripes heap after gc has finished. Signed-off-by:
-
Kent Overstreet authored
bch2_btree_and_journal_walk() walks the btree overlaying keys from the journal; it was introduced so that we could read in the alloc btree prior to journal replay being done, when journalling of updates to interior btree nodes was introduced. But it didn't have btree node prefetching, which introduced a severe regression with mount times, particularly on spinning rust. This patch implements btree node prefetching for the btree + journal walk, hopefully fixing that. Signed-off-by:
-
Kent Overstreet authored
- Originally bch_extent_stripe_ptr didn't contain the block index, instead we'd have to search through the stripe pointers to figure out which pointer matched. When the block field was added to bch_extent_stripe_ptr, not all of the code was updated to use it. This patch fixes that, and we also now verify that field where it makes sense. - The ec_stripe_buf_init/exit() functions have been improved, and are now used by the bch2_ec_read_extent() (recovery read) path. - get_stripe_key() is now used by bch2_ec_read_extent(). - We now have a getter and setter for checksums within a stripe, like we had previously for block sector counts, and ec_generate_checksums and ec_validate_checksums are now quite a bit smaller and cleaner. ec.c still needs a lot of work, but this patch is slowly moving things in the right direction. Signed-off-by:
-
Kent Overstreet authored
More debugging info is always a good thing. Signed-off-by:
-
Kent Overstreet authored
Signed-off-by: -
Kent Overstreet authored
This fixes a regression introduced by "bcachefs: Refactor filesystem usage accounting". We have to include all the replicas entries that have any of the entries for different journal entries nonzero, we can't skip them if they sum to zero. Signed-off-by:
-
Kent Overstreet authored
Alloc info isn't stored on a particular device, it makes no sense to only be writing it out for rw members - this was causing fsck to not fix alloc info errors, oops. Also, make sure we write out alloc info in other repair paths. Signed-off-by:
-
Kent Overstreet authored
A btree node merge operation deletes a key in the parent node; if when inserting into the parent node we split the parent node, we can end up with a whiteout in the parent node that we don't want. The existing code drops them before doing the split, because they can screw up picking the pivot, but we forgot about the unwritten writeouts area - that needs to be cleared out too. Signed-off-by:
-
Kent Overstreet authored
This reverts part of the change from "bcachefs: Don't use BTREE_INSERT_USE_RESERVE so much" - it turns out we still should be reserving open buckets for btree node allocations, because otherwise data bucket allocations (especially with erasure coding enabled) can use up all our open buckets and we won't be able to do the metadata update that lets us release those open bucket references. Oops. Signed-off-by:
-
Kent Overstreet authored
The zstd compression code seems to have a bug where it will write just past the end of the destination buffer - probably only when the compressed output isn't going to fit in the destination buffer, which will never happen if you're always allocating a bigger buffer than the source buffer which would explain other users not hitting it. But, we size the buffer according to how much contiguous space on disk we have, so... generally, bugs like this don't write more than a word past the end of the buffer, so an easy workaround is to subtract a fudge factor from the buffer size. Signed-off-by:
-
Kent Overstreet authored
We don't want to fail the recovery/mount because of a single error reading from the journal - the relevant journal entry may still be found on other devices, and missing or no journal entries found is already handled later in the recovery process. Signed-off-by:
-
Kent Overstreet authored
It used to be safe to reallocate a buf that the write path owns without holding the journal lock, but now this can trigger an assertion in journal_seq_to_buf(). Signed-off-by:
-
Kent Overstreet authored
With various newer key types - stripe keys, inline data extents - the old approach of calculating the maximum size of the value is becoming more and more error prone. Better to switch to bkey_on_stack, which can dynamically allocate if necessary to handle any size bkey. In particular we also want to get rid of BKEY_EXTENT_VAL_U64s_MAX. Signed-off-by:
-
Kent Overstreet authored
Allocations for copygc have to be kept separate from everything else, so that copygc doesn't get starved. Signed-off-by:
-
Kent Overstreet authored
We don't want writes to not get erasure coded just because the allocator temporarily wasn't keeping up. However, it's not guaranteed that these allocations will ever succeed, we can currently get stuck - especially if devices are different sizes - we still have work to do in this area. Signed-off-by:
-
Kent Overstreet authored
Previously, in the stripe creation path, when reusing an existing stripe we'd read the existing stripe synchronously - ouch. Now, we allocate two stripe bufs if we're using an existing stripe, so that we can do the read asynchronously - and, we read the full stripe so that we can run recovery, if necessary. Signed-off-by:
-
Kent Overstreet authored
Originally, we'd check for -ENOSPC when getting a disk reservation whenever the new extent took up more space on disk than the old extent. Erasure coding screwed this up, because with erasure coding writes are initially replicated, and then in the background the extra replicas are dropped when the stripe is created. This means that with erasure coding enabled, writes will always take up more space on disk than the data they're overwriting - but, according to posix, overwrites aren't supposed to return ENOSPC. So, in this patch we fudge things: if the new extent has more replicas than the _effective_ replicas of the old extent, or if the old extent is compressed and the new one isn't, we check for ENOSPC when getting the disk reservation - otherwise, we don't. Signed-off-by:
-
Kent Overstreet authored
Previously, we were using BTREE_INSERT_RESERVE in a lot of places where it no longer makes sense. - we now have more open_buckets than we used to, and the reserves work better, so we shouldn't need to use BTREE_INSERT_RESERVE just because we're holding open_buckets pinned anymore. - We have the btree key cache for updates to the alloc btree, meaning we no longer need the btree reserve to ensure the allocator can make forward progress. This means that we should only need a reserve for btree updates to ensure that copygc can make forward progress. Since it's now just for copygc, we can also fold RESERVE_BTREE into RESERVE_MOVINGGC (the allocator's freelist reserve). Signed-off-by:
-
Kent Overstreet authored
The move path was calling bch2_bucket_io_time_reset() for cached pointers (which it shouldn't have been), and then not calling bch2_trans_reset() when it got -EINTR (indicating transaction restart). Oops. Signed-off-by:
-
Kent Overstreet authored
__btree_trans_get_iter() was using bch2_btree_iter_upgrade, but it shouldn't have been because on failure bch2_btree_iter_upgrade may drop locks in other iterators, expecting the transaction to be restarted. But __btree_trans_get_iter can't return an error to indicate that we need to restart thet transaction - oops. Signed-off-by:
-
Kent Overstreet authored
We normally avoid having too many dirty keys in the btree key cache, to ensure that we can always shrink our caches to reclaim memory if needed. But this check was causing us to go into an infinite loop on startup, in the btree insert path before journal reclaim was started. Signed-off-by:
-
Kent Overstreet authored
In bch2_btree_interior_update_will_free_node, we copy the journal pins from outstanding writes on the btree node we're about to free. But, this can race with the writes completing, and dropping their journal pins. To guard against this, just use READ_ONCE() in bch2_journal_pin_copy(). Signed-off-by:
-
Kent Overstreet authored
With the btree key cache code, we don't need to update the alloc btree lazily - and this will mean we can remove the bch2_alloc_write() call in the shutdown path. Future work: we really need to expend the bucket IO clocks from 16 to 64 bits, so that we don't have to rescale them. Signed-off-by:
-
Kent Overstreet authored
This now means "the maximum number of pointers within a bkey" - and bch_devs_list is updated to use it instead of BCH_REPLICAS_MAX, since stripes can contain more than BCH_REPLICAS_MAX pointers. Signed-off-by:
-
Kent Overstreet authored
This is something we clearly should be checking for, but weren't - oops. Signed-off-by:
-
Kent Overstreet authored
Particularly on emergency shutdown we can end up having to clean up a lot of dirty cached btree keys here. Signed-off-by:
-
Kent Overstreet authored
If a btree node merger is followed by a split or compact of the parent node, we could end up with the parent btree node iterator pointing to the whiteout inserted by the btree node merge operation - the fix is to ensure that interior btree node iterators always point to the first non whiteout. Signed-off-by:
-
Kent Overstreet authored
With erasure coding, we now have processes in the background that compact data, causing it to take up less space on disk than when it was written, or potentially when it was read. This means that we can't trust the page cache when it says "we have data on disk taking up x amount of space here" - there's always the potential to race with background compaction. To fix this, just check if we need to add to our disk reservation in the bch2_extent_update() path, in the transaction that will do the btree update. Signed-off-by:
-
Kent Overstreet authored
This is needed to fix a bug where we're overflowing iterators within a btree transaction, because we're updating the stripes btree (to update block counts) and the stripes btree trigger is unnecessarily updating the alloc btree - it doesn't need to update the alloc btree when the pointers within a stripe aren't changing. Signed-off-by:
-
Kent Overstreet authored
The stripe creation path was too state-machiney: it would always run the full state machine until it had succesfully created a new stripe. But if we tried to get and reuse an existing stripe after we'd already allocated some buckets, the buckets we'd allocated might have conflicted with the blocks in the existing stripe we need to keep - oops. Signed-off-by:
-
Kent Overstreet authored
Also, print out more information on btree transaction iterator overflow. Signed-off-by:
-
Kent Overstreet authored
When we didn't find a key to delete we were getting a null ptr deref. Signed-off-by:
-
Kent Overstreet authored
Metadata corruption bugs are hard to debug if we can't see exactly what went wrong - try to allocate a bigger buffer so we can print out everything we have. Signed-off-by:
-
Kent Overstreet authored
Without checking if we actually flushed anything, journal reclaim could still go into an infinite loop while trying ot shut down. Signed-off-by:
-
Kent Overstreet authored
Had a type that meant we were triggering journal reclaim _much_ more aggressively than needed. Also, fix a potential integer overflow. Signed-off-by:
-
Kent Overstreet authored
- Try to always keep 1/8th of the journal free, on top of pre-reservations - Move the check for whether the journal is stuck to bch2_journal_space_available, and make it only fire when there aren't any journal writes in flight (that might free up space by updating last_seq) Signed-off-by:
-
Kent Overstreet authored
This patch adds a flag to journal entries which, if set, indicates that they weren't done as flush/fua writes. - non flush/fua journal writes don't update last_seq (i.e. they don't free up space in the journal), thus the journal free space calculations now check whether nonflush journal writes are currently allowed (i.e. are we low on free space, or would doing a flush write free up a lot of space in the journal) - write_delay_ms, the user configurable option for when open journal entries are automatically written, is now interpreted as the max delay between flush journal writes (default 1 second). - bch2_journal_flush_seq_async is changed to ensure a flush write >= the requested sequence number has happened - journal read/replay must now ignore, and blacklist, any journal entries newer than the most recent flush entry in the journal. Also, the way the read_entire_journal option is handled has been improved; struct journal_replay now has an entry, 'ignore', for entries that were read but should not be used. - assorted refactoring and improvements related to journal read in journal_io.c and recovery.c Previously, we'd have to issue a flush/fua write every time we accumulated a full journal entry - typically the bucket size. Now we need to issue them much less frequently: when an fsync is requested, or it's been more than write_delay_ms since the last flush, or when we need to free up space in the journal. This is a significant performance improvement on many write heavy workloads. Signed-off-by:
-
