Reinstate the splice read/write/move paths removed in 84592a9f, now
gated behind runtime options with negotiated pipe sizing and fallback
logging.
- Transport: msgbuf-piped splice receive, fd-referenced splice reply
(vmsplice + splice), splice-move pipe-bounce, and restored
fuse_buf_copy/fuse_buf_flags plumbing for copy fallbacks
- Negotiation: new pipe-max-size option (auto-raised at INIT when
possible; the sysctl read/write is factored into src/pipe_max_size.cpp
so fuse_init.cpp and option_parser.cpp share one implementation);
fuse-msg-size clamped to what the pipe can hold; splice is disabled
at INIT with a warning if the sysctl can't fit even the minimum,
retracting the FUSE_CAP_SPLICE_* bits already advertised to the
kernel in that case
- Read replies and splice-move are gated at 128 KiB (below that the
syscalls cost more than the copy they avoid); every fallback path
degrades to the copy path and logs once per process via SysLog
- msgbuf pipe cache is process-global and mutex-guarded (mutex_lockguard)
so pipes can move between read and worker threads under
process-thread mode
- API simplified after review: collapsed the three per-axis options
(splice-read/write/move) into a single splice boolean (default on);
no-splice-read/write/move and the bare splice-read/write/move
spellings stay registered as accepted no-ops so existing mount
options don't fail the mount, but they do NOT alter the splice
knob - upgraders with no-splice-* in their config get splice
enabled and must set splice=false explicitly (see
deprecated_options.md)
- Fixed several correctness issues found in parallel review: a
dropped write_buf registration that silently corrupted writes, a
dirty-pipe reuse bug in msgbuf_free, an iov double-send in
fuse_reply_data's splice-move fallback, and pipe-capacity math that
assumed a 4096-byte page size
- Modernized the transport/negotiation code to this codebase's C++
style: RAII (unique_ptr/DEFER) instead of malloc/free and manual fd
close(), std::from_chars/to_chars instead of stoul/snprintf,
_ReplyPipe's pipe ops as methods instead of free functions closing
over a thread_local global, and vendored/libfuse/lib/buffer.cpp (a
near-verbatim port of upstream's buffer.c) brought in line with the
rest of the fork
- Documented splice and pipe-max-size options (mkdocs, FAQ,
deprecated-options list); added tools/bench-splice.sh
- Benchmarked on tmpfs: seq_write +21%, seq_read +52%, small_write
+5%, small_read +12%; validated with a 1h soak (~271 GiB
hash-verified I/O, zero integrity failures), a live-mount round-trip
of I/O above the splice threshold, and the existing unit/smoke
suites at their pre-existing baseline
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- Store cloned read-worker fds on fuse_session and close them there.
- Set up clone fallback through session helpers used by the loop.
- Carry per-request fds through receive, process, and reply paths.
Tests show significant performance increases in highly concurrent
workloads on a high core count system. Upto 15% better on reads and
27% on writes.
- Guard fuse_cfg_t log file and filepath shared pointers with a shared mutex
- Replace atomic shared pointer loads and stores with shared and unique locks
- Add helpers that release path locks and update node state under f.lock.
- Update getattr and setattr stat-cache invalidation while preserving ESTALE handling.
- Rework create, open, tmpfile, and set_path_info to mutate lookup/open state under the shared lock.
- Document the lock/unlock pair as an acquire/release barrier.
- Explain that releasedir waits for in-flight readdir calls before destroy/free.
- Warn against removing the synchronization despite it not being a critical section.
Important changes:
- Switch mergerfs and vendored libfuse Makefiles to -std=c++20.
- Select the latest available GCC/G++ or toolset in container builds.
- Verify compiler C++20 support during package installation.
- Fix dynamic SysLog formatting for fmt C++20 checks.
- Reserve a zero stat CRC to represent an invalid cache fingerprint.
- Include device, size, inode, mtime, and ctime data in the fingerprint.
- Recompute the fingerprint during auto-cache open before keeping cache.
No functional change: libfuse/helper.cpp (fuse_daemonize,
fuse_helper_opt_proc, fuse_parse_cmdline) and mergerfs
fuse_init.cpp (_want helpers) reformatted to match the project's
GNU-style brace placement and aligned parameter lists.
- fuse_opt: reformat process_opt_param to project style; return the
rv from sscanf through a named local so the error branch prints
the failing option cleanly.
- cpu.cpp: when parsing /sys/devices/system/cpu/*/topology/core_id,
bail out cleanly on stream read failure rather than recording an
uninitialized core_id.
- fusermount: free(copy) on the success path of parsing the mount
option (previously leaked on every mount), and replace the fixed-
buffer sprintf into procfile with snprintf bounded by the buffer
size.
- maintenance_thread: widen the iteration count from int to u64
so a long-lived daemon doesn't wrap after ~68 years of ticks
(real effect: silence signed-overflow warnings under UBSan).
- ThreadPool::destroy used to iterate _threads directly while another
thread could be adjusting the same vector. Snapshot the vector
under the mutex and iterate the local copy for pthread_cancel /
pthread_join.
- ObjPool::construct now returns NULL when the pooled allocator
returns NULL instead of invoking placement-new on a null
pointer. In destroy, compute to_node(obj) before running the
destructor - the destructor may scrub bytes the to_node helper
reads.
- In the async/sync workers in fuse_loop, msgbuf_alloc can return
NULL under memory pressure. Fail the worker cleanly with a
diagnostic rather than dereferencing. Also introduce a small
_nproc() helper that treats hardware_concurrency()==0 as 1, so
the downstream min/max arithmetic doesn't produce a zero thread
count.
The signal handler reads the global session pointer from an async-
signal context while another thread installs/removes the session
via set/remove_signal_handlers. Use std::atomic<fuse_session*> with
acquire/release ordering so the handler either sees NULL or a fully
published session, never a torn pointer.
Collapse get_node/get_node_nocheck into a single function that
returns NULL when the node is absent rather than abort()ing.
Callers that previously assumed non-NULL now check the return value
and either reply ESTALE/ENOMEM or fall through gracefully; a few
places where the node is load-bearing (e.g. unhash during remove)
use fatal::abort with a descriptive message.
Switch mutex_lock / mutex_unlock pairs to mutex_lockguard so early
returns on error paths no longer leak the lock, and add NULL-safety
around fuse_dh in readdir/readdir_plus, fuse_dirents_init error
returns, the getattr path when both name and fh are empty, and the
msgbuf allocation in fuse_lib_read.
Pull the new xstrdup/xmalloc wrappers from xalloc.hpp for the root
node name so the common-case OOM path terminates with a diagnostic
instead of a silent corrupted node.
Also guard fuse_session_destroy against a NULL session - called
from error-recovery paths.
The low-level init/destroy callbacks were passed a userdata pointer
that mergerfs never consumed - the caller threaded it through
fuse_lowlevel_new only to hand it back unchanged. Remove the
parameter from the ops struct, from fuse_lowlevel_new(_common),
and from the internal fuse_ll struct; update fuse_lib_init and
fuse_lib_destroy accordingly.
Also drop fuse_reply_ioctl_iov: no caller in-tree invokes it and
the public API is defined by this vendored libfuse fork.
libfuse previously initialized its debug log filepath to
"/dev/stderr", forcing the config layer to reopen stderr on every
load whether or not the user asked for it. Default the libfuse log
path to "" and have the mergerfs config layer only forward the
request to libfuse when the user supplied an explicit path.
