bcachefs: swap file support via SWP_FS_OPS by matthiasgoergens · Pull Request #1072 · koverstreet/bcachefs

matthiasgoergens · 2026-03-02T00:18:28Z

Add swap file support for bcachefs using the SWP_FS_OPS path (same
mechanism as NFS swap). The filesystem stays in the I/O loop for swap
operations, so swap data gets checksumming, encryption, replication,
compression, and multi-device support.

Unlike btrfs (which disables COW, checksums, and compression for swap
files), bcachefs swap files use the normal COW write path.

How to use

fallocate -l 4G /mnt/bcachefs/swapfile
chmod 600 /mnt/bcachefs/swapfile
mkswap /mnt/bcachefs/swapfile
swapon /mnt/bcachefs/swapfile

Key design decisions

PF_MEMALLOC context: swap_rw sets PF_MEMALLOC for both reads and
writes to prevent reclaim re-entry deadlocks. The BCH_WRITE_swap flag
propagates this to the write index kworker.
Btree node pinning: at swapon, leaf nodes covering the swap file are
marked noevict to avoid disk reads during memory reclaim.
Btree cache pre-reserve: 16 MB of btree node buffers pre-allocated at
swapon (with GFP_NORETRY to avoid OOM on small VMs).
Disk reservation: swap_pages × PAGE_SECTORS reserved at swapon to
prevent ENOSPC during COW writes.
Bkey buffer pre-allocation: 2048-byte buffer allocated with GFP_NOWAIT
before entering PF_MEMALLOC, avoiding __GFP_NOFAIL WARN loops in
bch2_bkey_buf_realloc.
Write-buffer flush noreclaim: upgraded PF_MEMALLOC_NOFS to
PF_MEMALLOC in the write-buffer flush path. The flush holds
wb->flushing.lock while traversing btree nodes; if a node allocation
enters direct reclaim, reclaim can need the journal, which needs
write-buffer flush → deadlock. This is a pre-existing bcachefs bug
not specific to swap, but swap makes it reliably reproducible.
IS_SWAPFILE bypass: swap files skip pagecache locking in the direct I/O
path.
80% fill monitoring: rate-limited warning when extents btree leaves
exceed 80% fill, for early warning before the 67% split threshold.

Test results (22-test ablation matrix)

18/22 PASS — 14/16 core tests pass, 2 expected failures, 2 throttle
timeouts.

Tested at 128–512M RAM covering: pinning on/off, PF_MEMALLOC on/off, btree
cache thrashing, drop_caches, swapoff/swapon cycles, concurrent file I/O
workload, 5-minute sustained pressure, and disk-throttled I/O.

Expected failures (prove hardening is necessary):

small-128m-nothing: all hardening disabled → deadlock
small-128m-noreclaim: PF_MEMALLOC disabled → reclaim deadlock

Throttle timeouts (test harness issue, not kernel bugs):

small-192m-throttled: I/O throttled to 512 KB/s, test times out
large-192m-throttled: same

Design document

Full design rationale in Documentation/filesystems/bcachefs/swap.rst,
including dead ends (pre-fragmentation defeated by extent merging,
GFP flag whack-a-mole) and adversarial analysis (dead-key accumulation,
shared btree nodes, ENOSPC).

Closes #368

Add investigation notes, audit results, and a QEMU-based reproducer for the SRCU-held-across-blocking-operations deadlock. The reproducer boots a 128 MB VM with pre-populated tiered bcachefs (SSD -> HDD), eats 200 MB of memory to force heavy reclaim, and checks whether the system stays responsive while the reconcile thread runs. prepare-vm-disks.sh - format + pre-populate disk images on the host build-initramfs.sh - build static musl Rust init (~430 KB) run-vm-test.sh - run test VM, analyze results vm-init-rs/ - Rust init binary (pedantic clippy clean) INVESTIGATION.md - crash analysis, full audit, GitHub issue survey

Audit and fix 25 sites where bch2_trans_unlock() (which only drops btree locks) is used before a potentially long-blocking operation, leaving the SRCU read lock held. If that blocking operation triggers page reclaim, and the shrinker needs to free old btree nodes, it must wait for the SRCU grace period -- which can't complete because we're holding the read lock. This creates a reclaim deadlock under memory pressure. The fix at each site is to use bch2_trans_unlock_long() (which also drops the SRCU read lock) instead of bch2_trans_unlock(). Also introduce a drop_locks_long_do() macro -- the SRCU-dropping variant of drop_locks_do() -- and convert the callers that pass blocking operations (mutex_lock, wait_event, journal reservation, etc.). Also clean up bch2_fsck_ask_yn(): the deferred unlock_long workaround (wait 2s then upgrade to unlock_long) is no longer needed since we now drop SRCU immediately before waiting for user input. Sites fixed: btree/cache.c 4 GFP_KERNEL allocs, sync I/O, wait_on_bit_io btree/commit.c 3 GFP_KERNEL alloc, journal reclaim wait, journal res btree/interior.c 5 mutex, closure_sync, wait_event, down_read btree/locking.c 1 mutex_lock btree/read.c 1 sync disk I/O alloc/foreground.c 1 mutex_lock data/update.c 3 GFP_KERNEL bio alloc, closure_sync (x2) data/ec/io.c 1 kvmalloc + multi-device I/O data/migrate.c 1 closure_wait_event for in-flight writes data/write.c 1 nocow lock + GFP_KERNEL vfs/buffered.c 1 filemap_alloc_folio + GFP_KERNEL vfs/fs.c 1 __wait_on_freeing_inode init/error.c 1 user input wait (+ cleanup) debug/tests.c 1 journal flush Reported on a 128 GB desktop (bcachefs root, tiered SSD+HDD): the reconcile thread held SRCU for 13+ seconds during GFP_KERNEL bio allocation in bch2_data_update_init(), blocking memory reclaim and causing a full system deadlock. Related: koverstreet#934 Related: koverstreet#636 Signed-off-by: Matthias Goergens <matthias.goergens@gmail.com>

matthiasgoergens · 2026-03-08T11:49:49Z

I'm using this branch on my desktop now. It's working quite well so far.

koverstreet · 2026-03-09T18:59:18Z

A few things:

Housekeeping:

This PR includes all the changes from bcachefs: don't hold srcu lock across blocking operations #1069 (the SRCU unlock_long audit). Those should be a separate commit/PR, not bundled here. Please rebase on top of bcachefs: don't hold srcu lock across blocking operations #1069 or drop the duplicate hunks.
There are a lot of files that shouldn't be in a kernel PR: INVESTIGATION.md, plan.markdown, review-findings-*.markdown, writeup.markdown, build-initramfs.sh, prepare-vm-disks.sh, run-vm-test.sh, the vm-init-rs/ Rust binary, spinoff-gfp-noretry.markdown, walkthrough.md, pr-description.markdown. These are useful investigation notes but they don't belong in the kernel tree. Remove them from the PR (keep them in your fork if you like).
module-version.c / version.h with a hardcoded version string — this doesn't belong.
pr_info("bcachefs: custom swap code loaded\n") in init/fs.c — remove.

The swap implementation itself:

The SWP_FS_OPS approach is the right one. Some notes on the implementation:

The BUG_ON / BUG() calls after stall detection (10s in swap_rw, 10s in write index worker) should not be in production kernel code. A WARN_ON_ONCE is fine for diagnostics, but BUG() during swap I/O under memory pressure will make a recoverable situation into a kernel panic. Remove these.
Pinning the entire alloc btree (POS_MIN to SPOS_MAX) in bch2_swap_pin_unpin_nodes is extremely aggressive. On a filesystem with many devices and large allocation spaces, that's potentially thousands of btree nodes pinned in memory permanently. This needs to be scoped more carefully — perhaps just the allocation ranges for the devices that back the swap file's data.
The bch2_swap_noreclaim_enabled global bool checked in the hot write path (data/write.c) is a code smell. Feature toggles via __setup are OK for debugging but shouldn't be the permanent interface. If the noreclaim behavior is correct (and it probably is), just do it unconditionally for swap writes.
The scoped_guard(memalloc_flags, PF_MEMALLOC_NOFS|PF_MEMALLOC) change in btree/write_buffer.c affects all write buffer flushes, not just swap. That needs more thought — you're saying no write buffer flush should ever enter direct reclaim, which may be true but it's a broader change than "swap support."
The debug spew in btree/commit.c (logging when extents leaf is at 80% fill) should be behind a debug flag or removed.

Testing:

You've built your own QEMU test framework — nice work on the dm-delay approach for exposing lock contention. You should know that bcachefs already has a comprehensive test framework called ktest that handles VM provisioning, test orchestration, and integrates with the bcachefs CI. Rather than maintaining a parallel framework, consider adding your swap torture tests as ktest tests — that way they run automatically and benefit the whole project.

— ProofOfConcept

Add copy-on-write swap file support to bcachefs using the SWP_FS_OPS path (same mechanism as NFS and btrfs). This enables checksumming, encryption, compression, replication, and multi-device support for swap data. Key design decisions: PF_MEMALLOC context: swap_rw sets PF_MEMALLOC for both reads and writes. Writes need it because swap writeback runs during reclaim. Reads need it because swap-in page faults could otherwise trigger reclaim that starts competing swap writes, risking btree lock deadlock. The BCH_WRITE_swap flag propagates PF_MEMALLOC to the write index kworker. Btree node pinning: at swapon, leaf nodes for extents, inodes, and alloc btrees covering the swap file are marked noevict. This avoids disk reads during memory reclaim — the hot path only does in-memory btree lookups. Btree cache pre-reserve: 16 MB of btree node buffers are pre-allocated on bc->freeable at swapon using GFP_NORETRY to avoid OOM on small VMs. These cover ~5 concurrent fully-cold btree traversals across extents + inodes + alloc btrees. Disk reservation: swap_pages * PAGE_SECTORS reserved at swapon to prevent ENOSPC during COW writes. Each swap write allocates a new physical block before freeing the old one; without a reservation, concurrent non-swap writers could consume all free space. Bkey buffer pre-allocation: 2048-byte buffer allocated with GFP_NOWAIT before entering PF_MEMALLOC, avoiding __GFP_NOFAIL WARN loops when bch2_bkey_buf_realloc spills from the 96-byte on-stack buffer. IS_SWAPFILE bypass: swap files skip pagecache locking in the direct I/O path since the kernel manages swap file pages directly. Write-buffer flush deadlock fix: upgrade PF_MEMALLOC_NOFS to PF_MEMALLOC in the write-buffer flush path. The flush holds wb->flushing.lock while traversing btree nodes; if a node allocation enters direct reclaim, reclaim can need the journal, which needs write-buffer flush, which is blocked on the mutex — deadlock. PF_MEMALLOC prevents entering reclaim; allocation failures are already handled. Diagnostics: rate-limited warning when extents btree leaves exceed 80% fill (early warning before the 67% split threshold), swap I/O stall detection at 2s (WARN) and 10s (BUG in debug builds). Tested at 128-512M RAM with parallel ablation matrix covering: pinning on/off, PF_MEMALLOC on/off, btree cache thrashing, drop_caches, multi-device tiered storage, lz4/zstd compression, 5-minute sustained pressure, concurrent file I/O workload, and disk-throttled I/O. Signed-off-by: Matthias Schorer <matthias@schorer.dev>

Document the design rationale for COW swap file support: the reclaim deadlock problem, how other filesystems handle it, and the mechanisms used (PF_MEMALLOC, btree pinning, cache pre-reserve, disk reservation, bkey pre-allocation, write-buffer flush noreclaim). Includes analysis of dead ends (pre-fragmentation defeated by extent merging, GFP flag whack-a-mole, kworker flag propagation) and adversarial analysis (dead-key accumulation, shared btree nodes, ENOSPC scenarios). bcachefs: fix unlimited lock hold times in trans bounds bch2_trans_begin() evaluates whether a transaction has exceeded BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS (10ms) and if so cleanly yields locks via cond_resched(). However, it was resetting trans->last_begin_time unconditionally on *every iteration* of loop macros like for_each_btree_key. This meant that tight loops processing many extents, such as background compression or bch-reconcile, would continuously reset their own yield timer since individual loop iterations execute in microseconds. This permitted single transactions to monopolize crucial btree locks limitlessly, starving other core filesystem operations (e.g. the swap kworker blocking for 120s+ during memory pressure, causing system lockups). This patch introduces a new trans->last_yield_time field which is ONLY cleared when locks are legitimately dropped (in trans_set_unlocked), properly accumulating lock hold times across independent bch2_trans_begin() boundaries until the threshold is organically reached.

0 = always drop SRCU before blocking (bch2_trans_unlock_long), equivalent to the original code path N > 0 = two-phase escalation with (N-1)ms timeout before dropping SRCU This fixes a semantic bug where timeout=0 would still keep SRCU held during sub-jiffy operations, preventing SRCU grace periods from completing and potentially causing stalls. Runtime tunable via: /sys/module/bcachefs/parameters/srcu_escalation_timeout_ms

The move ratelimiter used global in-flight limits (move_ios_in_flight, move_bytes_in_flight) for all devices. This overwhelms rotational devices with deep queues of random writes, causing head thrashing and multi-second write latencies that block btree transaction state and stall journal reclaim. Add per-context max_ios_in_flight and max_sectors_in_flight fields to moving_context, initialized from c->opts. The reconcile phys thread (which already runs per-device for rotational devices) overrides these with much lower limits suitable for spinning disks. Add module parameter move_ios_in_flight_rotational (default 8) to allow runtime tuning without recompilation. Signed-off-by: Matthias <matthias@example.com>

…ng journal writes

dblsaiko · 2026-05-29T04:11:06Z

I rebased this for bcachefs-tools because I wanted to test-drive it myself. From a quick test (yolo running it on my system and then filling up memory), it seems to still work, however no guarantees I made the right decisions merging in the changes from master since I am not familiar with the codebase. I left a few extra comments (search for "Katalin") where my decisions are probably questionable.

This is just a straight rebase with none of the review comments addressed, but hopefully this is something you can pick up future work off of.

The following changes since commit b8f8939fa2437b36208b43206e063ac36266aebe:

  btree commit: btree_key_can_insert_slowpath() (2026-05-28 15:35:21 -0400)

are available in the Git repository at:

  https://git.dblsaiko.net/bcachefs/ swap-files

for you to fetch changes up to 6dd4637e1d615179051cdbb9a35042c8de0f5077:

  bcachefs: make journal bucket discards asynchronous to avoid paralyzing journal writes (2026-05-29 06:00:20 +0200)

Details

Matthias Goergens (10):
      bcachefs: SRCU lock investigation and VM reproducer
      bcachefs: don't hold srcu lock across blocking operations
      docs: note foreground I/O stall resolution in INVESTIGATION.md
      bcachefs: swap file support via SWP_FS_OPS
      Documentation: bcachefs: add swap file design document
      srcu escalation: fix timeout=0 to mean original behavior
      bcachefs: per-device rate limiting for background data moves
      bcachefs: Add checksum_only flag for pure B-Tree metadata replacement
      bcachefs: Elevate B-Tree journal reclaim writes to REQ_PRIO
      bcachefs: make journal bucket discards asynchronous to avoid paralyzing journal writes

 Documentation/swap.rst   | 262 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 INVESTIGATION.md         | 301 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 build-initramfs.sh       |  38 +++++++++++
 fs/Makefile              |   3 +-
 fs/alloc/foreground.c    |   2 +-
 fs/btree/bkey_buf.h      |   6 ++
 fs/btree/cache.c         |  80 +++++++++++++++++++++--
 fs/btree/cache.h         |   2 +
 fs/btree/commit.c        |  14 +++-
 fs/btree/interior.c      |  14 ++--
 fs/btree/iter.c          |  25 +++++--
 fs/btree/iter.h          |  43 ++++++++++++
 fs/btree/locking.c       |  15 ++++-
 fs/btree/locking.h       |   1 +
 fs/btree/read.c          |   2 +-
 fs/btree/types.h         |   1 +
 fs/btree/write.c         |   6 +-
 fs/btree/write_buffer.c  |  10 +--
 fs/data/ec/io.c          |   2 +-
 fs/data/migrate.c        |   2 +-
 fs/data/move.c           |  15 +++--
 fs/data/move.h           |   8 ++-
 fs/data/reconcile/work.c |  14 ++++
 fs/data/update.c         |  37 ++++++++++-
 fs/data/update.h         |   1 +
 fs/data/write.c          |  72 ++++++++++++++++++--
 fs/data/write.h          |   1 +
 fs/data/write_types.h    |   5 +-
 fs/debug/tests.c         |   2 +
 fs/init/error.c          |  17 +----
 fs/journal/reclaim.c     |   8 ++-
 fs/vfs/buffered.c        |   2 +
 fs/vfs/direct.c          |  34 ++++++++--
 fs/vfs/fs.c              |   6 +-
 fs/vfs/fs.h              |   7 ++
 fs/vfs/swap.c            | 310 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 fs/vfs/swap.h            |  22 +++++++
 prepare-vm-disks.sh      |  66 +++++++++++++++++++
 run-vm-test.sh           |  75 +++++++++++++++++++++
 vm-init-rs/.gitignore    |   2 +
 vm-init-rs/Cargo.toml    |  12 ++++
 vm-init-rs/src/main.rs   | 200 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 42 files changed, 1678 insertions(+), 67 deletions(-)
 create mode 100644 Documentation/swap.rst
 create mode 100644 INVESTIGATION.md
 create mode 100755 build-initramfs.sh
 create mode 100644 fs/vfs/swap.c
 create mode 100644 fs/vfs/swap.h
 create mode 100755 prepare-vm-disks.sh
 create mode 100755 run-vm-test.sh
 create mode 100644 vm-init-rs/.gitignore
 create mode 100644 vm-init-rs/Cargo.toml
 create mode 100644 vm-init-rs/src/main.rs

koverstreet · 2026-05-29T04:38:27Z

oh boy this is going to be a big patchset to review want to join the IRC channel? we should discuss

…

On Thu, May 28, 2026, 11:11 PM Katalin Rebhan ***@***.***> wrote: *dblsaiko* left a comment (koverstreet/bcachefs#1072) <#1072 (comment)> I rebased this for bcachefs-tools because I wanted to test-drive it myself. From a quick test (yolo running it on my system and then filling up memory), it seems to still work, however no guarantees I made the right decisions merging in the changes from master since I am not familiar with the codebase. I left a few extra comments (search for "Katalin") where my decisions are probably questionable. This is just a straight rebase with none of the review comments addressed, but hopefully this is something you can pick up future work off of. The following changes since commit b8f8939fa2437b36208b43206e063ac36266aebe: btree commit: btree_key_can_insert_slowpath() (2026-05-28 15:35:21 -0400) are available in the Git repository at: https://git.dblsaiko.net/bcachefs/ swap-files for you to fetch changes up to 6dd4637e1d615179051cdbb9a35042c8de0f5077: bcachefs: make journal bucket discards asynchronous to avoid paralyzing journal writes (2026-05-29 06:00:20 +0200) Details Matthias Goergens (10): bcachefs: SRCU lock investigation and VM reproducer bcachefs: don't hold srcu lock across blocking operations docs: note foreground I/O stall resolution in INVESTIGATION.md bcachefs: swap file support via SWP_FS_OPS Documentation: bcachefs: add swap file design document srcu escalation: fix timeout=0 to mean original behavior bcachefs: per-device rate limiting for background data moves bcachefs: Add checksum_only flag for pure B-Tree metadata replacement bcachefs: Elevate B-Tree journal reclaim writes to REQ_PRIO bcachefs: make journal bucket discards asynchronous to avoid paralyzing journal writes Documentation/swap.rst | 262 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ INVESTIGATION.md | 301 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ build-initramfs.sh | 38 +++++++++++ fs/Makefile | 3 +- fs/alloc/foreground.c | 2 +- fs/btree/bkey_buf.h | 6 ++ fs/btree/cache.c | 80 +++++++++++++++++++++-- fs/btree/cache.h | 2 + fs/btree/commit.c | 14 +++- fs/btree/interior.c | 14 ++-- fs/btree/iter.c | 25 +++++-- fs/btree/iter.h | 43 ++++++++++++ fs/btree/locking.c | 15 ++++- fs/btree/locking.h | 1 + fs/btree/read.c | 2 +- fs/btree/types.h | 1 + fs/btree/write.c | 6 +- fs/btree/write_buffer.c | 10 +-- fs/data/ec/io.c | 2 +- fs/data/migrate.c | 2 +- fs/data/move.c | 15 +++-- fs/data/move.h | 8 ++- fs/data/reconcile/work.c | 14 ++++ fs/data/update.c | 37 ++++++++++- fs/data/update.h | 1 + fs/data/write.c | 72 ++++++++++++++++++-- fs/data/write.h | 1 + fs/data/write_types.h | 5 +- fs/debug/tests.c | 2 + fs/init/error.c | 17 +---- fs/journal/reclaim.c | 8 ++- fs/vfs/buffered.c | 2 + fs/vfs/direct.c | 34 ++++++++-- fs/vfs/fs.c | 6 +- fs/vfs/fs.h | 7 ++ fs/vfs/swap.c | 310 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ fs/vfs/swap.h | 22 +++++++ prepare-vm-disks.sh | 66 +++++++++++++++++++ run-vm-test.sh | 75 +++++++++++++++++++++ vm-init-rs/.gitignore | 2 + vm-init-rs/Cargo.toml | 12 ++++ vm-init-rs/src/main.rs | 200 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++ 42 files changed, 1678 insertions(+), 67 deletions(-) create mode 100644 Documentation/swap.rst create mode 100644 INVESTIGATION.md create mode 100755 build-initramfs.sh create mode 100644 fs/vfs/swap.c create mode 100644 fs/vfs/swap.h create mode 100755 prepare-vm-disks.sh create mode 100755 run-vm-test.sh create mode 100644 vm-init-rs/.gitignore create mode 100644 vm-init-rs/Cargo.toml create mode 100644 vm-init-rs/src/main.rs — Reply to this email directly, view it on GitHub <#1072?email_source=notifications&email_token=AAPGX3UEWT3VEC7I42NZ6QD45EEXDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINJXGA2DKMJQGQ22M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4570451045>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAPGX3TLHYGW6IZO6GABST345EEXDAVCNFSM6AAAAACWDUCG3GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DKNZQGQ2TCMBUGU> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AAPGX3QDGNWLP2YM2DACW3D45EEXDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINJXGA2DKMJQGQ22M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/AAPGX3XF6XOFUFQHEKFJYOD45EEXDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINJXGA2DKMJQGQ22M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you commented.Message ID: ***@***.***>

Komzpa · 2026-06-28T04:25:14Z

For easier review against the current bcachefs-tools tree, I refreshed a cleaned carry here:

bcachefs: swap file support via SWP_FS_OPS bcachefs-tools#646

It preserves Matthias as the commit author, keeps the implementation as one focused patch on bcachefs-tools/testing, and drops the investigation notes/VM harness/debug toggles/debug BUG path/hardcoded version files/broad write-buffer PF_MEMALLOC change that were called out above. I also tightened the latest swapon cleanup paths and validated the staged DKMS module build against /usr/src/linux-headers-7.1-amd64; the companion ktest coverage remains koverstreet/ktest#63.

matthiasgoergens force-pushed the swap-files branch 3 times, most recently from d0ab5e4 to 5587b10 Compare March 2, 2026 03:12

matthiasgoergens added 2 commits March 7, 2026 20:08

matthiasgoergens force-pushed the swap-files branch from 5587b10 to 009db27 Compare March 8, 2026 02:17

matthiasgoergens added 3 commits March 11, 2026 19:23

docs: note foreground I/O stall resolution in INVESTIGATION.md

d481acb

matthiasgoergens force-pushed the swap-files branch from 59b05b2 to fee6f09 Compare March 11, 2026 11:28

koverstreet force-pushed the master branch from 01a44e1 to 11271f2 Compare March 12, 2026 02:42

matthiasgoergens added 3 commits March 12, 2026 17:51

Merge branch 'fix/drop-srcu-before-data-update-alloc' into swap-files

91753c1

koverstreet force-pushed the master branch from 11271f2 to d251df1 Compare March 12, 2026 18:55

matthiasgoergens added 3 commits March 13, 2026 11:45

bcachefs: Add checksum_only flag for pure B-Tree metadata replacement

c3a03dd

bcachefs: Elevate B-Tree journal reclaim writes to REQ_PRIO

adb8e6d

bcachefs: make journal bucket discards asynchronous to avoid paralyzi…

3254b65

…ng journal writes

koverstreet force-pushed the master branch from 6303f5b to 990d039 Compare March 14, 2026 03:32

koverstreet force-pushed the master branch from a437918 to 113cc1f Compare March 24, 2026 04:34

koverstreet force-pushed the master branch 2 times, most recently from a6d79f5 to 82c906f Compare April 1, 2026 04:57

koverstreet force-pushed the master branch from 9c59f7d to b931c6f Compare April 26, 2026 00:56

koverstreet force-pushed the master branch 3 times, most recently from c43a605 to 84ce651 Compare May 13, 2026 17:33

This was referenced Jun 28, 2026

tests/bcachefs: cover swapfile reclaim pressure koverstreet/ktest#63

Open

bcachefs: swap file support via SWP_FS_OPS koverstreet/bcachefs-tools#646

Open

This was referenced Jun 28, 2026

RFE: swap file support #368

Open

Swapfile #923

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

bcachefs: swap file support via SWP_FS_OPS#1072

bcachefs: swap file support via SWP_FS_OPS#1072
matthiasgoergens wants to merge 11 commits into
koverstreet:masterfrom
matthiasgoergens:swap-files

matthiasgoergens commented Mar 2, 2026 •

edited

Loading

Uh oh!

matthiasgoergens commented Mar 8, 2026

Uh oh!

koverstreet commented Mar 9, 2026

Uh oh!

dblsaiko commented May 29, 2026

Uh oh!

koverstreet commented May 29, 2026 via email

Uh oh!

Komzpa commented Jun 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

matthiasgoergens commented Mar 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

How to use

Key design decisions

Test results (22-test ablation matrix)

Design document

Uh oh!

matthiasgoergens commented Mar 8, 2026

Uh oh!

koverstreet commented Mar 9, 2026

Uh oh!

dblsaiko commented May 29, 2026

Uh oh!

koverstreet commented May 29, 2026 via email

Uh oh!

Komzpa commented Jun 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

matthiasgoergens commented Mar 2, 2026 •

edited

Loading