Viewing: llite.sync_on_close.4
.TH LLITE.*.SYNC_ON_CLOSE 4 2026-05-19 Lustre "Lustre Kernel Interfaces"
.SH NAME
llite.sync_on_close \- enable (meta)data sync on file close
.SH SYNOPSIS
.SY
.BI "lctl get_param llite." INSTANCE .sync_on_close
.SY
.BI "lctl set_param llite." INSTANCE .sync_on_close= ENABLE
.YS
.SS PROPERTIES
.TP
.B Access Permissions
.br
.BR 644 " | " -rw-r--r--
.TP
.B Scope
.br
Per client mountpoint
.TP
.B Config
.br
Always present on client
.TP
.B Default
.br
.RB llite.*.sync_on_close= 0
.br
.TP
.B Valid Range
.br
.RB llite.*.sync_on_close= 0
.br
.RB llite.*.sync_on_close= 1
.SH DESCRIPTION
The
.BR sync_on_close (4)
tunable parameter allows a client mountpoint to have more NFS-like behavior.
When
.IB ENABLE =1
then new file data is written to the servers immediately when
.BR close (2)
or
.BR fclose (3)
is called and they do not return until the data has been saved persistently.
This avoids issues with applications that do not call
.BR fsync (2)
or
.BR fflush (3)
themselves, but depend on NFS sync-on-close write semantics for consistency.
.PP
With the default
.IB ENABLE =0
setting, writes are flushed to the server when they form a full-sized RPC,
as determined by
.BR osc.max_pages_per_rpc (4),
when
.BR fsync (2)
or equivalent is called,
when the LDLM lock covering the file data is cancelled
(e.g. if another client wants to read the data),
or when the kernel flushes old data under memory pressure or by age.
.PP
It is preferable that applications flush file data explicitly themselves
when this is required, since
.BR sync_on_close (4)
can hurt overall IO performance due to increased close latency,
especially when a large number of small files are being written, like with
.BR tar (1)
or
.BR rsync (1).
.SS NOTE
The "dirty" state of a file is tracked per inode rather than per file
descriptor, so if multiple processes on a client have the same file
open, then the file data and metadata is sync'd only once for the file,
unless it is modified again afterward.
.PP
If multiple threads have the same file open, then the kernel's
VFS-level close will delay calling into Lustre until the last process
using a file closes its descriptor.
.SH MODULES
This parameter is in the following modules:
.EX
.B llite.*.sync_on_close
.EE
.SH EXAMPLES
Enable data fsync on close for all files on the local client:
.EX
.RB "client# " "lctl set_param llite.*.sync_on_close=1"
.EE
.SH AVAILABILITY
.B llite.*.sync_on_close
is part of the
.BR lustre (7)
filesystem package since release 2.18.0.
.\" Added in commit 2.17.5x
.SH SEE ALSO
.BR osc.max_pages_per_rpc (4),
.BR lctl-get_param (8),
.BR lctl-set_param (8)
