Recent OpenZFS work, by @robn.

(Summary by the AI 🤖, read more about that).

• Current Focus Areas
• Pull requests: Active · Recently completed
• Day journal

Current Focus Areas

• Mounts / snapshot automounting — A colleague’s report of bind mount failures after zfs recv -F opened into structural work on ZFS dataset mount handling. #18415 landed a targeted fix for the z_teardown_lock / namespace_sem deadlock. The snapshot-mount-direct branch is the structural replacement — eliminating the userspace round-trip entirely with direct in-kernel vfsmount references. The snapshot-mount-cleanup branch is the current active work, refactoring the ctldir state machine (mounting flag → state enum, unmount split into named/snapentry variants, expiry under entry lock) before further integration.

• Log spacemap condense — #16747 adds on-demand log-spacemap flush and a zpool condense command, allowing operators to flush log spacemaps ahead of export to avoid long export times on large pools. Originally opened in 2023, now re-lifed with rework to close txgs immediately during condense and flush based on a percentage of dirty unflushed blocks rather than a fixed metaslab count.

• NFS4 change cookie — A user report of random EINVAL on NFS4 clients after kernel 6.18.14 (#18487) pointed at a locking gap in ZFS’s STATX_CHANGE_COOKIE path. The change cookie combines coarse ctime with a per-object z_seq counter; z_seq was not always read or written under lock, and was also reset on every znode eviction, allowing regressed cookies. Two complementary fixes are in flight: #18573 (ixhamza) persists z_seq via a new SA_ZPL_SEQ SA attribute so it survives znode eviction; the statx-cookie-seq-lock branch addresses the locking gap. User bisect has since narrowed the issue to kernels 6.19+; 6.18 LTS is unaffected.

• ZAP subsystem — A series of improvements to the ZFS Attribute Processor (ZAP) key-value store. Source reorganisation, internal locking uplift, _by_dnode() API uplift, a µnit-based unit test framework with mocks, and core/cursor/misc function tests have all landed (#18516, #18546, #18551, #18564, #18586, #18603, #18619, #18630, #18654), building a test suite now at 50+ tests. The pipeline ahead fills remaining gaps: zap-unit-fatzap (fatzap: shrink, collision, pointer table growth, cursor cross-leaf) and zap-backend-api (a zap_ops vtable); community contributor @chrislongros has #18638 (mixed-case ZAP tests, approved) and #18639 (uint64-key tests) open; #18643 (namecheck tests) merged by behlendorf. #7633, the old ZTS casenorm integration test tracking issue, is now assigned to robn.

• zstats — New work to unify ZFS statistics across platforms. ZFS currently uses OS-specific kstat/sysctl/procfs APIs throughout, making cross-platform work and statistics introspection inconsistent. The zstats branch introduces a zstats abstraction layer with a unified API for counters, gauges, and atomic operations, with initial conversions of abd_os (Linux and FreeBSD) and dbufstats.

Pull requests

Active

Recently completed

Gone Cold

Day journal

12 Jun (Friday)

• dsl_scan — Continued review discussion on #18652 — acknowledged reviewer suggestion to renumber ZB_* marker constants; agreed current approach is right (renumbering is a large, invasive change for limited benefit)

11 Jun (Thursday)

• dsl_scan — Pushed update to #18652 branch zbookmark-compare-marker: rewrote comparison to handle all three marker bookmark levels explicitly; responded to reviewer’s question about whether ZB_DNODE_LEVEL can reach zbookmark_compare (argues it mostly can’t — it is a sentinel for TRAVERSE_PRE/TRAVERSE_POST rather than a real block reference)
• Log spacemap condense — Commented on #16747: zloop runs consistently locally to completion in ~12 min; noted CI pool uses autotrim which is absent in local testing; investigating whether that is the source of the CI flake
• NFS4 change cookie — Reviewed and approved #18573 — “Persist z_seq across znode eviction” (ixhamza); persists z_seq via new SA_ZPL_SEQ SA attribute so the NFS4 change cookie survives znode eviction; directly addresses the root cause identified in #18487 analysis; ixhamza had briefed robn early
• Reviewed and approved #18638 — “unit/zap: add MT_MATCH_CASE mixed-case test” (chrislongros); adds ZAP mixed-case table test coverage; noted U8_TEXTPREP_TOUPPER is also valid for Unicode normalization forms and MT_MATCH_CASE defangs it

10 Jun (Wednesday)

• Log spacemap condense — Pushed two updates to condense-log-spacemaps for #16747: reworked the debug condense target to properly track and cancel the delay task; added a ztest hook for the regular condense path (ztest exercises the actual condense, not the debug target)

• Assigned #7633 — “Test case: casenorm test group”; old tracking issue for ZTS casenorm integration tests needing update for Linux, directly relevant to ongoing ZAP casenorm work

• #18643 (unit/namecheck: test name validation by chrislongros) merged by behlendorf

9 Jun (Tuesday)

• dsl_scan — Opened #18652 — “zbookmark_compare: handle ‘marker’ bookmarks with negative levels”; correct fix for #14777, superseding #18598 and #18612; marker bookmarks (ZB_ROOT_LEVEL, ZB_ZIL_LEVEL, ZB_DNODE_LEVEL) represent valid dataset-level blocks but negative levels make BP_SPANB unsafe; fix sorts them ahead of “true” object blocks without performing any level arithmetic — they’re typically dataset metadata needed first anyway; ZTS running

• Log spacemap condense — Pushed fix to condense-log-spacemaps and responded to review on #16747: the debug condense target had a quirk where spa_condense_debug_task tasks linger on the delay taskq for up to 1s after condense completes; the task exits without re-arming on the next tick; documented the behaviour in-source

• ZAP — Commented on #18639 — “unit/zap: uint64 keys” (chrislongros); adds test_zap_uint64_keys covering ZAP_FLAG_UINT64_KEY fatzap (the key format used by DDT and BRT); robn had overlapping WIP in zap-unit-uint64-keys and pushed that branch; noted the PR covers the core operations well

• ZAP — Commented on #18638 (chrislongros, unit/zap: zap_lookup_norm) — noted overlap with own zap-unit-casenorm work; was about to post own PR today but #18638 had turned up first; coordinated and posted #18654 as a companion

• Reviewed and approved #18643 — “unit/namecheck: test name validation” (chrislongros); first unit test suite contributed independently of the test framework author — tests pool, dataset, snapshot, bookmark, component, permset, mountpoint, and depth namecheck functions; inline suggestions on naming style, avoiding unnecessary state restoration in forked tests

• ZAP — Opened #18654 — “unit/zap: key case normalization tests”; extends mock ZAP creation to accept normalisation flags; tests MT_NORMALIZE lookups across five Unicode normalization forms (U8_TEXTPREP_TOUPPER, NFC, NFD, NFKC, NFKD); tests conflict flag and real-name return in zap_lookup_norm_by_dnode(); tests zap_remove_norm_by_dnode()

• #18654 (unit/zap: key case normalization tests) applied directly by behlendorf

• #18643 (unit/namecheck: test name validation) approved and merged by behlendorf — “great to see the unit testing getting fleshed out”; suggested test_btree, test_range_tree, or test_raidz as next candidates

• #16747 (On-demand log-spacemap flush; zpool condense command) approved by behlendorf

6 Jun (Saturday)

• dsl_scan — Withdrew #18612 — confirmed with amotin that skipping ZB_ROOT_LEVEL prefetch entirely was wrong; these bookmarks do point to data the prefetcher needs; working on correct approach

5 Jun (Friday)

• ZAP — Opened #18630 — “zap: misc function removal / uplift / tests”; removes zap_increment_int() (long unused, potentially buggy) and zap_join()/zap_join_key()/zap_join_increment() (last seen 2017, too narrowly focused); adds zap_increment_by_dnode(), _by_dnode() variants for the zap_*_int and zap_*_int_key families, and zap_value_search_by_dnode(); all new API variants get unit tests; adds unit_rand_uint64() and unit_rand_str() helpers to the test framework; ZAP test suite now at 40 tests

• #18630 (zap: misc function removal / uplift / tests) applied directly by behlendorf

4 Jun (Thursday)

• ZAP — Updated #18619 — pushed unit-coverage-pretty branch with lcov 1.x support (lcov 1.14 / Alma 8.10 compatibility); worked around several 1.x quirks to produce the same text report from older tooling; responded to behlendorf’s review

• dsl_scan — Responded to amotin’s inline review on #18612 — agreed with amotin’s analysis; had forgotten that the prefetcher and scanner are separate threads each walking the tree, not a single pass where prefetch is set up ahead of each visit; that means rootbp is meaningful to prefetch and the fix is wrong; planning to withdraw #18612 once amotin weighs in with a better approach

• Reviewed #18625 — “PG_writeback may stuck with writeback_iter” (tiehexue, related to #18454); pushed back on the analysis — the claim that writeback_iter and write_cache_pages coexist and conflict applies only to a narrow HWE window (6.17 on Ubuntu 24.04); the proposed fix applies 6.18+ writeback code back to 6.10 where it isn’t needed

• #18619 (unit: produce a text coverage report) approved by behlendorf

• #18619 (unit: produce a text coverage report) applied directly by behlendorf

3 Jun (Wednesday)

• ZAP — Opened #18619 — “unit: produce a text coverage report”; adds scripts/coverage_report.pl which parses lcov/geninfo tracefiles and emits a text report immediately after make unit-coverage, showing per-file line/branch/function coverage; splits coverage targets into unit-coverage (text) and unit-coverage-html (full HTML); truecolor terminal colouring with fallback; Makefile improvements and doc updates

• #18603 (zap: add zap_cursor_init_by_dnode; cursor unit tests; mock dnode refcounts) applied directly by behlendorf

• amotin raised question on #18612 — whether the ZB_ROOT_LEVEL bookmark’s rootbp points to objset_phys_t and should still be prefetched to allow dsl_scan_prefetch_cb() to queue object prefetches

2 Jun (Tuesday)

• ZAP — Traced zpool_error_scrub_001_pos test hang on #18603 CI to error scrub carrying a live cursor in dsl_scan_t that is reinitialised on resume but not finalised on cancel; added dsl_scan: close errorscrub cursor on pause as a new first commit; final rebase to master; also restored zc_objset/zc_zapobj fields in zap_cursor_t for source compatibility (annotated “do not use” to document the ABI break)

• ZAP — Restructured additional ZAP unit test work into focused branches: zap-unit-misc (test additions for zap_join and zap_value_search), zap-unit-casenorm (case normalisation), zap-unit-fatzap (fatzap: shrink, collision, pointer table growth, cursor cross-leaf)

• Log spacemap condense — Rebased condense-log-spacemaps to master; behlendorf responded to naming concern on #16747 — considered extending zpool sync instead of a new zpool condense subcommand but is leaning toward keeping “condense” as it’s descriptive and likely to accumulate more data structures

• #18612 (dsl_scan_prefetch: don’t prefetch ZB_ROOT_LEVEL bookmarks) approved by behlendorf

1 Jun (Monday)

• Closed #18598 — withdrew zbookmark PR after amotin’s review prompted deeper analysis; the observed UBSAN shift values (-5, -7, -10) are explained by negative level constants (ZB_ROOT_LEVEL = -1, ZB_ZIL_LEVEL = -2, ZB_DNODE_LEVEL = -3) combined with valid ibs values, which bypass the existing IMPLY assertions; the ibs == 0 guard was addressing the wrong root cause

• Opened #18612 — “dsl_scan_prefetch: don’t prefetch ZB_ROOT_LEVEL bookmarks”; correct fix for #14777; dsl_scan_visit_rootbp() creates a marker bookmark with level == ZB_ROOT_LEVEL (-1) and passes it to dsl_scan_prefetch(), which sorts the list via zbookmark_compare() — that calls BP_SPANB() with a negative level, producing the undefined shift; fix: skip prefetch when level == ZB_ROOT_LEVEL (nothing to prefetch anyway, matching how traverse_prefetcher() handles ZB_DNODE_LEVEL); also asserts level >= 0 in zbookmark_compare() for future protection; full ZTS passed

• Published zio-crypt-common — initial work on a cross-platform encryption interface for zio_crypt; establishes an OS abstraction layer, lifts common code, and adds adapters for Linux/userspace (ICP) and FreeBSD

• #18603 (zap: add zap_cursor_init_by_dnode; cursor unit tests) approved by amotin

31 May (Sunday)

• ZAP — Responded to amotin’s review on #18603 about always taking the dnode and zap_t hold in _init() rather than deferring; pushed updated zap-unit-cursor making zap_cursor_init() and friends return int; adopted a deferred-error model where a failed init zeros the cursor enough that zap_cursor_retrieve() returns EIO and zap_cursor_fini() is a no-op, preserving existing call-site patterns while still handling init failures

30 May (Saturday)

• Commented on #18593 — continued discussion on auto-loading the kernel module when zpool list/status is invoked; outlined the behavioural ambiguity introduced if the module is not loaded (“no pools available” currently gives a definitive answer from the kernel; without loading it would be ambiguous between “no pools” and “module not loaded”)

• Reviewed (changes requested) #18605 — “fix: add buffer-length check in rdwr_efi.c” (orbisai0security); pushed back on “high severity” framing — strcpy is not a security issue in itself and the author’s own analysis showed the path is not exploitable; requested appropriate severity description and squash to a single commit

• #18603 (zap: add zap_cursor_init_by_dnode; cursor unit tests) reviewed by amotin — suggested always taking the dnode hold in _init() to allow assertions in cursor operations

29 May (Thursday)

• ZAP — Created zap-unit-cursor branch and opened #18603 — “zap: add zap_cursor_init_by_dnode; cursor unit tests; mock dnode refcounts”; adds zap_cursor_init_by_dnode() and zap_cursor_init_serialized_by_dnode(), which must hold the dnode until the first zap_cursor_retrieve() call; adds light refcounting to mock dnodes so tests can verify all holds are released on cleanup; cursor tests cover walking, multiple scenarios, and serialization (applied directly by behlendorf 3 Jun)
• Pushed splice-read — WIP: Linux splice_read file operation to support sendfile(), splice(), and tee() on ZFS files; reviving earlier exploratory work

28 May (Wednesday)

• Opened #18598 from zbookmark-compare-underflow-fix — “zbookmark_compare: avoid negative shift in block span calculation”; BP_SPANB subtracts SPA_BLKPTRSHIFT from ibs before using it as a shift count; ibs=0 is a legal value and with unsigned math the subtract underflows, making the shift amount undefined behaviour; fix only calls BP_SPANB when ibs is non-zero, otherwise uses the block id directly; ZTS passed; fixes #14777
• bootenv — Rebased bootenv-cleanup onto master for #17985; commented asking FreeBSD loader maintainers (@bsdimp, @kevans91, @tsoome) to weigh in ahead of the 2.5 betas; noted the upcoming FreeBSD dev summit as an opportunity
• Closed #17951 — “compat 6.18 clang - build failing”; already fixed
• Closed #16680 — “block clone and bulk deletions (Veeam)”; concluded the throttling side was addressed by #18043 and remaining design notes are captured elsewhere

27 May (Wednesday)

• Log spacemap condense — Added basic zpool condense tests to #16747: introduced a debug condense type (compiled only with ZFS_DEBUG) that runs a series of delay tasks to bump stats then completes after 10s, used as a hook for test observation; raised naming concern on #16747 — “condense” may be confusing given other unrelated condense operations in ZFS (indirect, metaslab, livelist); proposed the operation might be better described as a log-to-primary flush

• Investigated #18593 — “zpool list/status should not force the zfs kernel to be loaded” (thegushi, FreeBSD 15); located the relevant libzfs_load_module() function on Linux and FreeBSD; noted the behaviour has been present since 2010 on Linux and 2007 on FreeBSD and would be a visible change; offered strace -f zpool list for further diagnosis

• #18586 (unit/zap: test “core” ZAP operations) applied directly by behlendorf

• #18588 (config: detect the right way to get pthreads) merged by behlendorf

26 May (Tuesday)

• Log spacemap condense — Responded to behlendorf’s inline review comments on #16747; rebased condense-log-spacemaps onto current master

25 May (Sunday)

• Opened #18588 from pthread-linkage — “config: detect the right way to get pthreads”; replaces the ad-hoc mix of -pthread/-lpthread flags with the AX_PTHREAD autoconf macro, which produces the correct CFLAGS and LIBS for the target platform; ensures consistent thread state across compilation units regardless of whether they were compiled with -pthread (on Linux, -pthread defines _REENTRANT while -lpthread does not); also avoids undefined behaviour when two units — one compiled with -pthread, one without — call across each other; tested on Linux (gcc+glibc) and FreeBSD (clang) (merged by behlendorf 26 May)
• ZAP — Opened #18586 from zap-unit-core — “unit/zap: test ‘core’ ZAP operations”; extends ZAP unit tests to cover core operations (add, update, remove, count, contains, length); tests are constrained to single-uint64_t values since that’s what both ZAP types share; adds zap_contains_by_dnode() and exposes zap_update_by_dnode(), zap_length_by_dnode(), and zap_get_stats_by_dnode(); also includes two unit framework improvements: zero coverage counters before a coverage run, and a TOPT Makefile arg to pass options through to the test binary (applied directly by behlendorf 26 May)
• Reviewed (changes requested) #18585 — fs_parse API compatibility shim for kernel 5.4.0 (tiehexue); adds m4 macro to detect changed fs_parse signature and wrapper in zpl_super.c; acknowledged as a good start, requested changes and squash to a single commit

23 May (Saturday)

• ZAP — Fixed packaging issues surfaced during #18564 CI runs: added FreeBSD 15.0-RELEASE to the CI matrix, reduced CI scope on push events (skip full run; only run full CI when workflow files change), updated ZTS sanity.run for the new unit test binary; then rebased zap-unit-redux, zap-by-dnode-more, zap-backend-api, and zap-unit-more onto master; deleted zap-unit branch

• #18564 (Unit testing framework: µnit framework, dnode/dbuf/dmu_tx mocks, initial ZAP test suite) applied directly by behlendorf

22 May (Friday)

• ZAP — Triaged #18572 — ZAP-leaf chunk-chain corruption causing soft lockup in zfs_readdir/zap_lookup (happyaron, Debian 2.3.2); le->le_next walks into freelist chunks because l_hash[] entries can point at chunks on the freelist, looping forever with no termination guard; the ASSERT3U guards are debug-only and compile away in production; recognised as a patch from a prior customer recovery scenario that was never upstreamed; pushed zap-chunk-loop-stopper with the defensive fix: zap_leaf: defend against chunk lists with no end marker; noted the root issue appears to be behind #10825, #5346 and several other historical corruption reports
• Reviewed (changes requested) #18568 — zap: TinyZAP for multi-uint64 entries (akashb-22); new on-disk ZAP format between MicroZAP and FatZAP for multi-integer values and long keys (Lustre FID use case); on-board with the concept but the implementation approach is convoluted — requested simplification; behlendorf also engaged positively on direction
• Commented on #18488 — Ubuntu 26.04 ZFS-with-7.0 thread: engaged with john-cabaj (Ubuntu ZFS package maintainer); explained the “experimental” warning mechanism; 2.3.7/2.4.2 releases (in review now) will add official 7.0 support

21 May (Thursday)

• Reviewed and approved #18570 — zpl_xattr: stop heap-allocating prefixed xattr names (anodos325); replaces kmem_asprintf with stack buffer for xattr name construction (bounded by XATTR_NAME_MAX=255), avoiding 3% overhead seen in samba profiling; ixhamza and behlendorf also approved; noted additional refactor opportunity

• #18563 (linux/super: properly apply ro/rw mount option to superblock) applied directly by behlendorf

20 May (Wednesday)

• Mounts — Created mount-ro-rw branch and opened #18563 — “linux/super: properly apply ro/rw mount option to superblock”; regression from #18339: commit f5a9e3a622 changed how SB_RDONLY was applied — it set readonly when a mount had ro, but never cleared it when the mount had rw, so the dataset’s readonly= property always won over an explicit mount -o rw; fix checks fc->sb_flags_mask to detect whether the option set or cleared readonly; includes lifted ZTS helpers and a new consistency test; fixes #18557

• ZAP — Opened #18564 — “Unit testing framework”; introduces µnit as an in-tree unit test framework (two files, no external dependency; each test runs in an isolated fork); three commits: framework itself, dnode/dbuf/dmu_tx mock infrastructure, and an initial ZAP test suite; groundwork from #18516/#18546/#18551 already landed; behlendorf positive: “exactly the kind of infrastructure we want” (applied directly by behlendorf 23 May)

• #18558 (spl_kvmalloc: remove __GFP_COMP before calling vmalloc()) merged by behlendorf

19 May (Tuesday)

• Opened #18558 from spl-kvmalloc-no-gfp-comp — “spl_kvmalloc: remove __GFP_COMP before calling vmalloc()”; fixes #18556; a prior commit stopped passing __GFP_COMP for KM_VMEM allocations, but the fallback path for non-KM_VMEM allocations to vmalloc() still passed it, generating a warning on Linux 7.0; full ZTS on 7.0.8 passed (merged by behlendorf same day)

• Triage on #18557 — mount -o rw not granting write access on readonly=on datasets since 2.4.2; confirmed as own regression from #18339; started working up a test case

• ZAP — Rebased zap-unit, zap-by-dnode-more, zap-backend-api, and zap-unit-redux after #18551 was applied; deleted zap-by-dnode branch

• #18551 (zap: make _by_dnode() API variants be the primary implementation) applied directly by behlendorf

• #18554 (config: show progress output for kernel API checks) merged by behlendorf

18 May (Monday)

• Opened #18554 from configure-kernel-check-progress — “config: show progress output for kernel API checks”; adds a background job in ZFS_LINUX_TEST_COMPILE_ALL that watches build.log and prints a running counter as modules compile; inactive in silent mode or when stdout is not a terminal; prompted by a user asking why configure appeared to hang at checking for available kernel interfaces... (merged by behlendorf same day)

17 May (Sunday)

• Commented on #18549 — user report (alex-moch) of UBSAN array-index-out-of-bounds in zap.c on Linux 7.0 with CONFIG_UBSAN; explained no actual memory bug — UBSAN doesn’t understand the full-block allocation pattern where mz_chunks[] covers the rest of the block past the header; pointed to fix
• ZAP — Created mz-chunks-flex branch and opened #18550 — “zap_impl: use flex array field for mzap_phys_t.mz_chunks[]”; mzap_phys_t is always a full-block allocation with mz_chunks[] as an array over the remaining block; UBSAN complains because the declared size is [1]; fix converts to a C99 flexible array member — pure type annotation, no behavior change; fixes #18549 (merged by behlendorf same day)
• ZAP — Opened #18551 — “zap: make _by_dnode() API variants be the primary implementation”; PR 3/N in the ZAP cleanup series following #18546; reorganises so _by_dnode() functions are the actual implementation, with objset+object variants as thin wrappers around them — same pattern as the zap_lock()/zap_lock_by_dnode() change from #18546; rebased zap-unit, zap-by-dnode, zap-by-dnode-more, zap-backend-api, and zap-unit-redux on top

16 May (Saturday)

• ZAP — Continued zap-unit branch: reorganised mock DMU infrastructure — split mock helpers into mock_dmu.c/mock_dmu.h, separating test fixtures from test cases in test_zap.c; rebased dependent branches (zap-by-dnode, zap-by-dnode-more, zap-backend-api)

• #18535 (zstream: fix crashes when refcount tracking enabled) applied directly by behlendorf

• #18546 (zap: internal locking uplift) applied directly by behlendorf

15 May (Friday)

• ZAP — Created zap-lock branch and opened #18546 — “zap: internal locking uplift”: renames lockdir/unlockdir to lock/unlock; makes zap_lock() a simple wrapper around zap_lock_by_dnode(), reducing repetition and error scope; simplifies mzap_create_impl() and the fatzap lock upgrade dance; cleans up refcount tags now that zap_lock_by_dnode() is the sole locking path; PR 2/N in the ZAP cleanup series; rebased zap-unit on top (applied directly by behlendorf 15 May)

• Log spacemap condense — amotin reported a VERIFY0(nvlist_add_nvlist) crash in #16747 testing; fixed the nvlist bug in spa_log_spacemap.c, pushed to condense-log-spacemaps, and responded to amotin’s inline review comment

• zstream — Fixed amotin’s review nit on #18535: missing signal.h include; pushed fix and confirmed it compiles on FreeBSD 15+; PR now has approvals from both amotin and behlendorf

• Responded to behlendorf’s question (on #18536 thread) about enabling reference_tracking_enable=1 by default in debug builds — shared timing comparison from recent ZTS runs

• #18536 (refcount tag fixups: three mismatched hold/rele tags across zap, ddt_log, dsl_bookmark) applied directly by behlendorf

14 May (Thursday)

• Log spacemap condense — Continued inline review discussion on #16747: addressed amotin’s comments about spa_estimate_metaslabs_to_flush() complexity inside MAX() and txg targeting; changed to spa_syncing_txg(spa) + 1 to ensure at least one flush round; rebased and pushed condense-log-spacemaps

• zstream — Created zstream-refcount-crash branch and opened #18535 — running tests with ABD refcount tracking enabled found crashes in zstream’s compress and decompress paths: zfs_refcount_init() had not been called, so the kmem cache used for tracking was not set up; two commits: signal handler setup (matching zdb/ztest) and zfs_refcount_init()/_fini() calls around the main work (applied directly by behlendorf 15 May)

• Created refcount-fixups branch and opened #18536 — “refcount tag fixups”; same ZTS run with refcount tracking surfaced three mismatched hold/rele tags: zap_lookup_length_uint64_impl()/zap_prefetch_uint64_impl() used FTAG for zap_unlockdir() when the matching zap_lockdir() was called by their caller (fixed by threading the caller’s tag through); ddt_log_begin()/ddt_log_commit() both used FTAG for a shared dbuf array (fixed by using the ddt_log_update_t * as tag); dsl_bookmark_create_sync_impl_snap() used FTAG for an internal redaction list hold that didn’t match the caller-provided tag (applied directly by behlendorf 15 May)

• zstats — Created zstats branch: initial work on a cross-platform ZFS statistics abstraction layer — a zstats API to replace the mix of OS-specific kstat/sysctl/procfs implementations with a unified interface; first conversions of abd_os (Linux and FreeBSD) and dbufstats to the new API; adds atomic counter support, grouping, and a macro layer to reduce boilerplate

13 May (Tuesday)

• Approved #18520 — Fix double free for blocks cloned after DDT prune (amotin); amotin incorporated the zfs_valstr.c update and added DDT/BRT CI tests

• Mounts — Opened #18529 from mount-source-option — zpl_super: handle ‘source’ mount option directly; vfs_parse_fs_param_source() was not available before 5.14 and not backported to LTS kernels; handling it directly in zfs removes the need for a configure check; fixes #18528 (merged by behlendorf same day)

• ZAP — Created zap-by-dnode-more branch extending zap-by-dnode with additional _by_dnode() variants; cleaned up zap-api-split branch now that #18516 has landed

• #18520 (Fix double free for blocks cloned after DDT prune) applied directly by behlendorf

12 May (Monday)

• Reviewed (changes requested) #18520 — Fix double free for blocks cloned after DDT prune (amotin); suggested picking a zfs_valstr.c update commit from robn/zfs to go with the fix

• Log spacemap condense — Addressed amotin’s inline review comments on #16747: bounds safety of condense-type array, stats macro approach, txg_kick semantics, metaslab estimation inside MAX(), and a superfluous conditional; pushed updated condense-log-spacemaps branch

• #18516 (zap: reorganise and improve documentation) applied directly by behlendorf

11 May (Sunday)

• ZAP — Created zap-backend-api branch: adds microzap unit test coverage (lookup, length, remove, count, cursor retrieve, stats) and introduces a ZAP backend operations API — a zap_ops vtable with function pointers for open/close/prefetch/add/update/remove/byteswap, making the backend switchable between microzap and fatzap without conditional dispatch through the frontend

10 May (Sunday)

• ZAP — Opened #18516 from zap-api-split — “zap: reorganise and improve documentation”; PR 1/N in a planned series for comprehensive ZAP unit tests and a modular backend; splits the existing zap.c/zap_micro.c into five files: frontend API, common internals, microzap, fatzap core, and fatzap leaf (applied directly by behlendorf 12 May)

9 May (Saturday)

• ZAP — Continued zap-api-split: public and internal interface cleanup

8 May (Friday)

• ZAP — Started zap-api-split branch: split ZAP implementation out from two source files into five, separating frontend, common internals, and the two storage backends (microzap and fatzap)

7 May (Thursday)

• Commented on #18454 — user report of permanent system hangs on kernel 6.12 (Debian 13): folio_end_writeback() not called in all exit paths of zpl_writepages, leaving folios stuck in PG_writeback state; acknowledged and noted it’s in queue

6 May (Wednesday)

• ZAP API — Continued zap-unit-redux branch: expanding fatzap coverage — cursor iteration and cross-leaf transitions, case normalisation, uint64 key type, fatzap_remove, format sanity checks, and upgrade path tests

5 May (Tuesday)

• NFS4 change cookie — Continued triage on #18487: user bisect narrowed issue to kernels 6.19+, 6.18 LTS series unaffected — reduces urgency; planning to set up local reproduction environment
• ZAP API — Restructured ZAP work into named branches: created zap-by-dnode (holding the _by_dnode() API additions: zap_join, zap_value_search, int-key family, zap_contains, zap_stats, zap_remove_norm) and zap-unit-redux (fresh approach to the unit test structure, split from zap-unit)

4 May (Monday)

• NFS4 change cookie — Continued statx-cookie-seq-lock branch: second commit ensures all ctime/z_seq updates are done atomically under lock; commented on #18487 with full root cause analysis — insufficient locking in the STATX_CHANGE_COOKIE path allows z_seq reads to race with updates, producing a stale change cookie that NFS interprets as “no change”, causing EINVAL on cached directory entries; kernel 6.18.14 change likely shifted timing to expose it more reliably
• ZAP API — Created zap-unit branch: initial unit test infrastructure for ZAP covering fatzap multi-leaf remove, shrink/reclamation, normalization (NFD, NFKC, NFKD), and initial tests for the new _by_dnode() function family

3 May (Sunday)

• Commented on #18487 — NFS4 random EINVAL on clients after kernel ~6.18.14; initial triage, suspecting locking issue around NFS change cookie (STATX_CHANGE_COOKIE / z_seq); created statx-cookie-seq-lock branch with first fix: zfs_getattr_fast now reads z_seq under lock; posted initial patch for community testing
• Closed #18488 — Ubuntu 26.04 shipping ZFS 2.4.1 with kernel 7.0 “SERIOUS DATA LOSS” warning; explained the warning is correct: there are no official OpenZFS releases for kernel 7.0, the package is an unofficial Ubuntu build

2 May (Saturday)

• #18466 (libzfs/mnttab: restore ability to enable/disable cache) applied directly by behlendorf
• #18471 (Linux 7.1 compat: revert d_u.d_alias rename back to d_alias) merged by behlendorf

1 May (Friday)

• Commented on #18471 — responded to reviewer question about including 7.1 patches under a 7.0 META for 2.4.2 build compatibility
• Mounts — Continued snapshot-mount-cleanup branch: snapentry state machine completeness — unmounting entries always transition through the dead state rather than jumping directly to inactive; matched snapentry lifetime to the parent snapdir dentry

30 Apr (Thursday)

• Mounts — Created snapshot-mount-cleanup branch — ctldir locking refactor: mounting flag replaced with a state enum, states added for unmounting and inactive, unmount split into named and snapentry variants, expiry moved under entry lock with snapshot lock reduced to AVL-only scope

• #18475 (AUTHORS: add names of recent new contributors) merged by behlendorf

29 Apr (Wednesday)

• Opened #18475 — AUTHORS: add names of recent new contributors; routine update_authors.pl run (merged 30 Apr)
• Mounts — Pushed final mnttab-cache-enable-disable update for #18466: added repro from #18464 as a libzfs test case, verified fix passes
• Commented on #18464 — explained rationale for keeping libzfs_mnttab_cache() callable but as a no-op for upgrade compatibility

28 Apr (Tuesday)

• Linux 7.1 compat — Opened #18471 — Linux 7.1 compat: struct dentry’s d_u.d_alias field reverts to plain d_alias in 7.1-rc1 (reverting a 3.18 rename); single-commit fix, compile-checked on 4.18+/5.15+/6.8+, ZTS running (merged 2 May)
• Mounts — Continued snapshot-mount-direct: restore refcounts via dentry holds
• Mounts — Started snapentry-dentry-lifecycle branch: ctldir interface cleanup (remove flags arg to zfsctl_snapshot_unmount(), delay param from zfsctl_snapshot_unmount_delay(), unused args from zfsctl_snapshot_alloc); implemented concurrent mounter holding in manage_transit and direct mount creation in the automount graft path

27 Apr (Sunday)

• Mounts — Continued snapshot-mount-direct: tracking pathwalk intent through the VFS layer

• #18466 (libzfs/mnttab: restore ability to enable/disable cache) approved by prakashsurya

25 Apr (Saturday)

• Commented on #18464 — regression in #18296: libzfs_mnttab_cache(hdl, B_FALSE) became a no-op after “always enable the cache” change; Delphix’s multi-handle pool breaks because mounts via one handle populate its AVL, making ZFS_PROP_MOUNTED return wrong answers on other handles; acknowledged regression and referenced #18466 as fix

• Opened #18466 — libzfs/mnttab: restore ability to enable/disable cache; restores disabled mode by flushing AVL on each libzfs_mnttab_find() call when disabled; cache defaults to off, zfs explicitly re-enables (applied directly by behlendorf 2 May)

• Mounts — Continued snapshot-mount-direct: basic teardown function for zfsctl_destroy; separated dentry and parent mount tracking; expiry shape and noop stub

• #18466 (libzfs/mnttab: restore ability to enable/disable cache) approved by ixhamza and behlendorf

24 Apr (Friday)

• Mounts — Continued snapshot-mount-direct: passing back mount flags through manage_transit; lifted and renamed zpl_snapdir_t to zpl_snapentry_t; initial unified snapentry shape

23 Apr (Thursday)

• Mounts — Continued snapshot-mount-direct: manage_transit callback writeup; dentry private management fixup

17 Apr (Friday)

• Mounts — Continued snapshot-mount-direct: working toward hooking automount via the manage_transit kernel callback — initial trigger implementation and struct scaffolding for transit state

• Created casenorm branch — WIP: ZFS case normalization; ZTS helper program and initial test conversion

• #18435 (Linux 7.0 compat: META, by behlendorf) merged; commented confirming all 7.0 release candidates and the final release tested clean

15 Apr (Wednesday)

• Mounts — Continued snapshot-mount-direct: debug on ioctl unmount paths; investigating invalidate flow and delayed-work scheduling; extensive tracing added

14 Apr (Tuesday)

• Mounts — Continued snapshot-mount-direct: extra ref on returned automount; discovered zfs_parse_fs_string() takes 3 args in Linux 6.18 and added a compat fix on the branch; debug on automount lifecycle

11 Apr (Saturday)

• Commented on #18422 — zfs, zpool: ignore SIGPIPE in main() (chrislongros); argued the blanket approach is too broad — interval and long-running modes (zpool iostat <pool> 1, zpool events -f) should not ignore SIGPIPE; questioned whether the motivating scenario actually produces exit 141

10 Apr (Friday)

• Mounts — Continued snapshot-mount-direct: substantial locking and lifecycle rework — restored MNT_INTERNAL with zfsctl_destroy() support; direct snapshot unmount at parent teardown; timer disarm on detach; reimplemented rename around the new locking structure; ref counting fixes and mount ownership checks
• Linux 7.1 compat — Created linux-7.1-compat branch — initial investigation: d_u.d_alias renamed back to d_alias in upcoming kernels

8 Apr (Wednesday)

• Mounts — Continued snapshot-mount-direct: reverted to single lock model, fixing lock bugs, extensive locking debug work; getting close

• Reviewed and approved #18415 — fix snapshot automount deadlock during concurrent zfs recv (ixhamza); drops z_teardown_lock before call_usermodehelper() to break the deadlock cycle with namespace_sem; fixes #18409 and #18073

• Commented on #18409 — snapshot automount deadlock during concurrent zfs recv; this is exactly the deadlock class snapshot-mount-direct eliminates (snapshot automount holds z_teardown_lock(R) across call_usermodehelper, blocking on namespace_sem(W) while /proc/self/mountinfo readers hold namespace_sem(R) and wait on z_teardown_lock); noted #18073 as possible duplicate

• Commented on #18333 — zfs_resume_fs() bail path deadlock and use-after-free; asked whether now resolved by #18310 merging

• #18407 (add ability to set user properties while changing encryption key) merged by behlendorf

• #18415 (fix snapshot automount deadlock during concurrent zfs recv) applied directly by behlendorf

7 Apr (Tuesday)

• Mounts — Continued snapshot-mount-direct: reworking locking model — new locking theory and initial implementation
• Opened #18407 — add ability to set user properties while changing encryption key; port of illumos change by @ahrens; extends zfs change-key -o ... to apply user properties in the same transaction as the key change, using the existing properties nvlist already threaded through to ZFS_IOC_CHANGE_KEY (merged 8 Apr)
• Commented on #18402 — continued discussion on “RFE: speed up SEEK_DATA syscalls”; real improvement requires considering dirty/uncommitted buffers which is very difficult; practical advice is tuning via zfs_dmu_offset_next_sync

6 Apr (Monday)

• Log spacemap condense — Agreed with amotin that metaslabs are the natural unit of flushing; reworked #16747 to count dirty metaslabs at condense start and flush N% per txg (default 5%, ~20s total); pushed log_spacemap: add counter for unflushed metaslabs to condense-log-spacemaps and commented with updated approach
• Commented on #18402 — “RFE: speed up SEEK_DATA syscalls”; explained zfs_dmu_offset_next_sync tunable: disabling it trades hole-detection accuracy for performance by skipping the txg sync

3 Apr (Friday)

• fs_context — Responded to behlendorf’s follow-up question on #18377; pushed rebase of kernel-mount-params-direct

• Mounts — Continued snapshot-mount-direct: mnt mismatch protection, MNT_NOSUID handling, cancel expire time on force unmount

• #18377 (linux/super: reimplement in-kernel mount options parser for fs_context) approved and applied directly by behlendorf

2 Apr (Thursday)

• Mounts — Continued snapshot-mount-direct branch: reworked snapshot entry model for multiple snapentries per snapshot; locking rejig; unmounting by snapentry rather than snapname — progressing toward proper vfsmount-backed snapshot lifecycle management

1 Apr (Wednesday)

• fs_context — Commented on #18377 — responded to behlendorf’s review question about using GFP_KERNEL in the fs_context callback
• Commented on #15918 — kernel BUG at mm/usercopy.c report; pointed to #17129 as potentially related (worker thread spinning in EFAULT loop after parent thread destroyed)

31 Mar (Tuesday)

• fs_context — Rebased #18377 onto current master and added a new commit to kernel-mount-params-direct: workaround for kernels that enforce “forbidden” mount options and reject them before zpl_get_tree() sees them; commented on PR to describe the new commit

• #18376 (Linux 7.0: ensure LSMs get to process mount options) merged by behlendorf

30 Mar (Monday)

• Log spacemap condense — Addressed amotin’s review comments on #16747: fixed metaslab sorting, made -t target argument optional (defaults to all), added -a for all-pools support following scrub/resilver precedent; pushed ABI bump on condense-log-spacemaps
• Log spacemap condense — Asked amotin on #16747 about how to pace incremental spacemap flushing — per-metaslab and per-block approaches both feel too arbitrary; wondering if the right answer is some central IO scheduling layer (the same structural problem FDT had)

28 Mar (Saturday)

• Commented on #18381 — user crash report from USB device removal; user on ZFS 2.1, which is well out of support; advised upgrade to 2.3/2.4 series, noting zed OOM fix and disk detection improvements since 2.2

27 Mar (Friday)

• fs_context — Opened #18377 from kernel-mount-params-direct — linux/super: reimplement in-kernel mount options parser for fs_context; removes old monolithic parser, moves options setup onto fs_context as a structured “intent” before mount/remount; foundational for snapshot automounting without userspace round-trips, libmount integration, and eventual removal of mount.zfs
• fs_context — Opened #18376 from new lsm-mount-opts branch — Linux 7.0: ensure LSMs get to process mount options; when a filesystem provides .parse_monolithic, the kernel skips LSM option processing entirely — breaking SELinux context/fscontext/defcontext/rootcontext properties since #18339; fix by calling security_sb_eat_lsm_opts() before stashing remaining options for zfs_domount() (merged 31 Mar)

26 Mar (Thursday)

• fs_context — Continued kernel-mount-params-direct branch: added vfs_t allocator and made it public; matched vfs_t lifetime to fs_context; removed zfs_mnt_t (passing mount params directly instead); removed parse_monolithic — further simplification of the mount parameter stack, aligning object lifetimes with the fs_context model and removing intermediate structures

25 Mar (Wednesday)

• Mounts — Continued mountset branch: introduced zfs_mountset_iter iterator API and libzfs_mountset_apply; migrated zpool_disable_datasets, zfs_iter_mounted, unshare_unmount, and mount-list callers to use the new iterator; removed zfs_mountset_foreach; marked zfs_mount/zfs_unmount as deprecated — the iterator-based model is becoming the primary interface

23 Mar (Monday)

• #18356 (pyzfs: update license tags/classifiers) merged by behlendorf

22 Mar (Sunday)

• Created header-split branch — WIP: exploring moving ZFS/SPL headers out of sys/ to new zfs/ and spl/ namespaces

21 Mar (Saturday)

• fs_context — Continued kernel-mount-params-direct branch — added better lifetime control and dup-on-copy for context; removed old options parser; mount parameter parsing in the kernel via fs_context params API taking cleaner shape

20 Mar (Friday)

• fs_context — Created kernel-mount-params-direct branch — reworking mount parameter parsing to happen directly in the kernel via the fs_context params API, continuing the flattening work from the always-fs-context branch
• Log spacemap — Re-lifed #16747 — on-demand log-spacemap flush and zpool condense command; reworked to close txg immediately during condense (like scrub) and flush based on % of dirty unflushed blocks rather than a fixed metaslab count
• Opened #18356 — pyzfs: update license tags/classifiers; fix deprecated setuptools license classifier warning by switching to SPDX identifier and adding spdxcheck tag (merged 23 Mar)
• Reviewed and approved #18354 — Linux 7.0: autoconf: remove copy-from-user-inatomic API checks (john-cabaj); PowerPC build fix removing an unnecessary API check that now pulls in a GPL-only symbol

18 Mar (Wednesday)

• Mounts — Created snapshot-mount-direct branch — reworking zfs_ctldir snapshot mounting to hold a vfsmount directly in the snapshot entry; refactoring snap/snapmount paths, removing chroot handling, options allocation, and se_path/mountpoint generation; exploring autoexpiry via dentry invalidation and investigating a leaked zfsvfs reference

• #18339 (linux/super: require fs_context, remove support for old mount API) approved by behlendorf and applied directly

17 Mar (Tuesday)

• fs_context — Opened #18339 — linux/super: require fs_context, remove support for old mount API; first PR from the always-fs-context branch — drops the HAVE_FS_CONTEXT-gated old entry points, flattens the call chain into zpl_get_tree() and zpl_reconfigure(); full ZTS on 6.12.63, compile-checked back to 4.18; note scoped to zpl_super.c only, further cleanup in subsequent PRs

• Responded to amotin’s review comment on #18331 — explained that generic_setlease predates filelock.h: it has always existed but its declaration moved from fs.h to filelock.h at some point, so the #ifdef HAVE_FILELOCK_HEADER guard is for the include only, not the symbol

• #18331 (Linux 7.0: also set setlease handler on directories) approved by amotin and behlendorf; merged by tonyhutter

16 Mar (Monday)

• Mounts — Continued mountset branch: pushed zfs_mountset_apply prototype and initial hacks wiring it into mount callers — exploring an apply-based model for executing a set of mount operations as a unit

• Mounts — Created always-fs-context branch — full init_fs_context adoption: drops HAVE_FS_CONTEXT compatibility shims entirely and refuses to build without fs_context; follow-through on #18260’s groundwork and the 2.5 goal flagged by behlendorf

• Commented on #18316 — statx requires <sys/stat.h> include file (wbx-github/uClibc-ng build failure in Buildroot); suggested using configure to detect statx and skip the test binary if absent; then retracted in favour of @jolivain’s similar-but-better solution

• #18331 (Linux 7.0: also set setlease handler on directories) confirmed working by satmandu on 7.0.0-rc3

15 Mar (Sunday)

• Opened #18331 — Linux 7.0: also set setlease handler on directories; prompted by @satmandu’s report of NFS access failures on 7.0 (worked on 6.19); #18215 added the regular-file setlease handler but missed directory file ops, causing EINVAL when the NFS server tries to take a directory lease

14 Mar (Saturday)

• Mounts — Continued mountset branch: added initial mountbuilder/mountopt APIs — a structured layer for constructing mount operations and managing mount options, building on top of the zfs_mountset abstraction

13 Mar (Friday)

• Mounts — Continued mountset branch: converted zpool_disable_datasets to mountset; reworked zfs_unmount and zfs_unshare for the new API; migration through the remaining libzfs/cmd mount callers continuing

12 Mar (Thursday)

• Mounts — Continued mountset branch: converted unshare_unmount_path, zfs_do_userspace, zfs_for_each, zfs_path_to_zhandle, and zinject to the mountset API; removed legacy getextmntent from libspl; branch migrating steadily through the remaining libzfs/cmd mount callers

• Opened #18312 — config: fix STATX_MNT_ID detection; _GNU_SOURCE was not defined in the configure check, so sys/stat.h wouldn’t produce struct statx or STATX_* defines; missed when a prior change switched from linux/stat.h to sys/stat.h

• Opened #18313 — contrib/debian: add zilstat.1 to Debian packaging install lists; CI fix following #18303 (zilstat: add man page) which caused Debian packaging to fail

• Commented on #18310 — s_active leak fix in zfsvfs_hold() when z_unmounted is true (mischivus); agreed a leak exists but raised questions about whether zfsvfs_rele() would be more appropriate and what happens in the zfsvfs_create() fallback path

• #18312 (config: fix STATX_MNT_ID detection) approved by ixhamza; merged by behlendorf

• #18313 (contrib/debian: add zilstat.1 manpage to installation list) approved by ixhamza; merged by behlendorf

11 Mar (Wednesday)

• Mounts — Continued mountset branch: converted unshare_unmount and zfs_iter_mounted to use the mountset iterator, migrating more of libzfs off the legacy mnttab API
• Reviewed (changes requested) #18304 — README: add FreeBSD 14.4-RELEASE alongside 15.0 (grahamperrin); suggested listing only the latest release of each series rather than accumulating every version

10 Mar (Tuesday)

• Mounts — Started mountset branch — concrete implementation of the zfs_mountset abstraction sketched in the #18296 design discussion: interface definition, libmount-based Linux backend, libzfs global mountset instance, and migration of zfs_is_mounted and the legacy libzfs_mnttab API to use the new interface

• #18295 (Linux: remove minimum version requirement) applied directly by behlendorf

• #18296 (libzfs: mnttab cleanup) applied directly by behlendorf

9 Mar (Monday)

• Linux support — Opened #18295 — remove minimum Linux kernel version requirement; prefer configure checks over a version gate; document actually-supported distros and kernels in README instead

• Mounts — Opened #18296 — libzfs mnttab cleanup; pulls libzfs_mnttab_* functions into libzfs_mnttab.c, general modernisation (SPL mutexes, alloc/free helpers), removes unused legacy struct extmnttab and gettmntany(); preparatory cleanup carried across every mount management prototype, sent separately since it stands on its own

• Mounts — Commented on #18296 in response to @lundman — sketched design direction for mount abstraction: opaque mntent object representing a dataset/mountpoint pair, with OS-specific methods (mount, unmount, property changes, sort); libmount wrapping on Linux, direct implementation on FreeBSD/Windows; libzfs and tools would only ever interact through the abstraction

• Mounts — Started libzfs-mnttab-libmount branch — WIP initial shape of OS-specific mount API, integrating mnttab cleanup changes with the libmount work into a new branch; first cut at the opaque object abstraction described in the design comment

• Reviewed and approved #18290 — FreeBSD: improve dmesg kernel message prefix (concussious); noted Linux mixes ZFS: and zfs: prefixes too, consistency wouldn’t hurt

• Closed #17351 — rename zed binary to zfs-zed; no compelling case made; resolving namespace clashes in a global shared namespace is out of scope

• Linux support — Responded to review nit on #18295 — fixed “prioritise” spelling; pushed updated branch

• Mounts — Responded to review nit from @ixhamza on #18296 — fixed commit title typos (mnntab/gettmntany); pushed updated branch

• #18295 (Linux: remove minimum version requirement) approved by behlendorf

• #18296 (libzfs: mnttab cleanup) approved by behlendorf and ixhamza

8 Mar (Sunday)

• Linux support — Started linux-no-minimum branch — prompted by fs-context adoption work; investigated minimum kernel version requirement; found 4.18 minimum is RHEL8-derived, no longer compiles with modern toolchains, and doesn’t reflect actual support commitments

7 Mar (Saturday)

• Mounts — Split mountcache-initial branch off mountcache — isolating the initial mountcache infrastructure (the /proc/mountinfo parser, dual-tree cache, zfs_is_mounted integration) as a discrete unit for submission as a PR

6 Mar (Friday)

• Commented on #17000 — design discussion on zfs create -pp syntax for creating parents with canmount=no; suggested -O canmount=no as a general “properties for implicitly-created parents” alternative (intelfx); conceded it was a stretch; behlendorf resolved to proceed with -pp (intelfx)

• #18281 (libzpool: lift zfs_file ops out to separate source file) merged by behlendorf

5 Mar (Thursday)

• SHA-512 / SIMD build — Opened #18285 — separate SIMD compile checks for userspace and kernel toolchains; HAVE_SIMD() selector macro picks the right set at compile time; eliminates Ubuntu 24.04 builder failures where objtool rejects SHA512EXT in the kernel module even though the host compiler supports it

• #18285 (separate SIMD compile checks for userspace and kernel) approved by behlendorf and directly applied

4 Mar (Wednesday)

• SHA-512 / SIMD build — Investigated behlendorf’s report of Ubuntu 24.04 build failure in #18233; diagnosed as toolchain/kernel mismatch: compiler supports SHA512EXT but objtool in the kernel rejects it; responded with intent to fix properly; started simd-config branch — refactoring SIMD feature detection to test against both host and kernel toolchains separately

• Opened #18281 — split zfs_file_* ops out of libzpool/kernel.c into a new libzpool/zfs_file_os.c source file; cleanup requested by @lundman to ease porting work (merged 6 Mar)

• #18259 (more consistent TREE_* macro use in AVL comparators) merged by behlendorf

• #18281 (libzpool: lift zfs_file ops out to separate source file) approved by lundman

3 Mar (Tuesday)

• Mounts — Created libmount branch — WIP: reworking ZFS userspace mount management to use Linux’s libmount library; detect and link libmount, initial mount/unmount through it, mnttab_cache remade as a libmount wrapper, zfs_iter_mounted reimplemented over mtab

• Responded to amotin’s review nits on #18259 — removed unneeded lvb != rvb redundancy (good catch); defended belt-and-suspenders wrapping of zbookmark_compare() (AVL comparators must return -1/0/1, non-AVL functions could break at distance if changed); pushed updated branch

• #18259 (more consistent TREE_* macro use in AVL comparators) reviewed by amotin with nits

2 Mar (Monday)

• Commented on #18259 — responded to amotin’s Clang disassembly feedback; pushed rebase to master with || → | in zfs_refcount_compare to eliminate extra branch (tradeoff: load instead, on same cacheline); noted potential for separate search vs insertion comparator as a future idea

27 Feb (Friday)

• Mounts — Created fs-context-uplift branch — WIP: beginning full init_fs_context mount API adoption, next step following #18260
• Commented on #18259 — posted annotated disassembly showing macro changes produce identical codegen in kernel and userspace builds; noted spa_mapping_key_compare() as another candidate for the same treatment

26 Feb (Thursday)

• #18233 (SHA-512 via Intel SHA512 extension) merged by behlendorf
• #18260 (prefer fs_context mount API when available, groundwork for full init_fs_context adoption) merged by behlendorf

25 Feb (Wednesday)

• Mounts — Continued progress on mountcache branch: built out core mountcache infrastructure — alloc/free, locking (rrwlock), /proc/mountinfo refresh, dual-tree lookup (by dataset and by mountpoint), integrated cache into libzfs_handle_t, wired up zfs_is_mounted to use it; then filtered non-ZFS mount entries out of cache, generalised foreach iterator
• SHA-512 performance — Responded to review comment on #18233 — clarified 0.00 CPB values in sha2_test output when the CPU MHz argument is omitted; not a regression
• SHA-512 performance — Posted performance results on Core Ultra 9 285K VM for #18233: sha512ext at 1.82 CPB vs avx2 at 3.18 CPB (~1.75x faster on Arrow Lake desktop)
• Reviewed and approved #18261 — remove deprecated Fedora 41 from CI (tonyhutter)

24 Feb (Tuesday)

• Linux 7.0 compat — Commented on #18216: confirmed full ZTS pass on 7.0-rc1; maintainer flagged init_fs_context full adoption as a 2.5 goal

• SHA-512 performance — Rebased #18233 to master; shared first real hardware test results (Core Ultra 7 258V, ThinkPad X1 gen 13): sha512ext hashes 1GiB in 205ms vs avx2 at 945ms — ~4.6x faster; in-kernel test to follow

• Opened #18259 — more consistent use of TREE_CMP()/TREE_ISIGN()/TREE_PCMP() macros in AVL comparators; ZTS pass on Linux; notes a few curious constructions worth discussing (merged 4 Mar)

• Linux 7.0 compat — Opened #18260 — prefer new fs_context mount API when available (compiles back to 5.2), inverting the 7.0 compat shim check; sending to CI to verify support range before full init_fs_context adoption (merged 26 Feb)

• #18216 (Linux 7.0 compat: blk_queue_nonrot, posix_acl_to_xattr, legacy mount API removal) merged by behlendorf

• #18233 (SHA-512 via Intel SHA512 extension) approved by tonyhutter

23 Feb (Monday)

• Mounts — Started new mountcache branch — fresh approach to mount caching, starting with an initial /proc/mountinfo parser
• Mounts — Started kernel-mount-direct branch — reworking zfs_ctldir snapshot mounting to hold a vfsmount directly in the snapshot entry rather than going through the automount path
• Started superblock-experiments branch — early WIP: reimagining superblock vs inode relationship

22 Feb (Sunday)

• #18215 (F_SETLEASE compatibility fix) merged
• #18233 (SHA-512 via Intel SHA512 extension) approved by AttilaFueloep; nit about .byte directives in zfs_sha512_transform_sha512ext

20 Feb (Friday)

• #18230 (zdb --key crash fix and test harness improvements) merged by behlendorf
• #18216 (blk_queue_nonrot/posix_acl_to_xattr/legacy mount API changes) approved by behlendorf

19 Feb (Thursday)

• SHA-512 performance — Responded to two review comments on #18233 — fixed bugs in cpuid detection logic caught by reviewer; pushed sha512ext (multiple iterations)

• SHA-512 performance — Fixed $NF awk bug; pushed sha2-test-all-correct (final)

• Mounts — Commented on #18205 — investigated full unmount/remount approach; concluded it needs more work; PR left as safety-check only for now

• Mounts — Created mnttab-cache-highlander branch (WIP: mount table caching)

• #18232 (sha2_test correctness checks for all implementations) merged by behlendorf

• #18216 (Linux 7.0 compat) approved by tonyhutter

18 Feb (Wednesday)

• SHA-512 performance — Created sha512ext and sha2-test-all-correct branches
• SHA-512 performance — Opened #18232 — extend sha2_test correctness checks to all implementations; fix output formatting and get_cpu_freq for variable-frequency CPUs (merged 19 Feb)
• SHA-512 performance — Opened #18233 — icp: SHA-512 via Intel SHA512 extension; tested via Intel SDE with Arrow Lake emulation; requesting community testing on real hardware (merged 26 Feb)
• SHA-512 performance — Pushed sha512ext and sha2-test-all-correct branches (iteration)
• Linux 7.0 compat — Pushed setlease branch; commented on #18215 — addressed review feedback on commit structure
• Mounts — Commented on #18207 — discussed canonical vs. use-time normalisation of mountpoint=; noted cross-OS implications
• SHA-512 performance — Responded to review on #18230 — leftover error-threading code in zdb_derive_key()

17 Feb (Tuesday)

• SHA-512 performance — Created sha512-avx512 branch and pushed — deleted same day (AVX-512 approach abandoned)
• SHA-512 performance — Created zdb-load-key-smarter branch; opened #18230 — fix zdb --key crash on unencrypted datasets; fix test harness to catch silent zdb assertion failures (merged 20 Feb)

16 Feb (Monday)

• Linux 7.0 compat — Pushed setlease branch; responded to review on #18215, fixed commit split
• Code review: #18222, #18133

15 Feb (Sunday)

• Linux 7.0 compat — Created setlease branch; opened #18215 — explicitly set .setlease = generic_setlease to survive removal of kernel fallback in 7.0 (merged 22 Feb)
• Linux 7.0 compat — Pushed linux-7.0-compat branch; opened #18216 — handle blk_queue_nonrot rename, new posix_acl_to_xattr signature, legacy mount API removal
• Code review: #18214

14 Feb (Saturday)

• Linux 7.0 compat — Created linux-7.0-compat branch

13 Feb (Friday)

• Mounts — Commented on #18205 — worked out root cause (outstanding long holds prevent destroy); updated PR description

12 Feb (Thursday)

• Mounts — Created flatten-mount-point branch; opened #18207 — flatten mountpoint= parameter before sort in mountpoint_cmp, fixing mount ordering when trailing slashes present
• Mounts — Commented on #18205 responding to CI failures

11 Feb (Wednesday)

• Linux 6.19 compat — Approved #18197 META update to mark 6.19 support (tonyhutter) — merged; 6.19 work now fully landed
• Mounts — Created no-busy-receive-destroy branch; opened #18205 — abort zfs recv -F with EBUSY when dataset has outstanding bind mounts

9 Feb (Monday)

• Linux 6.19 compat — Confirmed ZTS run passed on 6.19 release day; #18053 effectively complete
• Opened and merged #18189 — AUTHORS update

6 Feb (Friday)

• Code review: #18179, #18180