From 42b9995f88a4e04697bea19d85c9598183446fb1 Mon Sep 17 00:00:00 2001 From: Rob Norris Date: Thu, 1 May 2025 14:07:21 +1000 Subject: [PATCH] linux/super: add tunable to request immediate reclaim of unused inodes Traditionally, unused inodes would be held on the superblock inode cache until the associated on-disk file is removed or the kernel requests reclaim. On filesystems with millions of rarely-used files, this can be a lot of unusable memory. Here we implement the superblock drop_inode method, and add a zfs_delete_inode tunable to control its behaviour. By default it continues the traditional behaviour, but when the tunable is enabled, we signal that the inode should be deleted immediately when the last reference is dropped, rather than cached. This releases the associated data to the dbuf cache and ARC, allowing them to be reclaimed normally. Sponsored-by: Klara, Inc. Sponsored-by: Fastmail Pty Ltd Reviewed-by: Brian Behlendorf Reviewed-by: Alexander Motin Signed-off-by: Rob Norris Closes #17746 --- man/man4/zfs.4 | 23 ++++++++++++++++- module/os/linux/zfs/zpl_super.c | 44 ++++++++++++++++++++++++++++++--- 2 files changed, 62 insertions(+), 5 deletions(-) diff --git a/man/man4/zfs.4 b/man/man4/zfs.4 index 7f1adaceb..60e17ddce 100644 --- a/man/man4/zfs.4 +++ b/man/man4/zfs.4 @@ -18,7 +18,7 @@ .\" own identifying information: .\" Portions Copyright [yyyy] [name of copyright owner] .\" -.Dd August 14, 2025 +.Dd September 15, 2025 .Dt ZFS 4 .Os . @@ -2583,6 +2583,27 @@ the xattr so as to not accumulate duplicates. .It Sy zio_requeue_io_start_cut_in_line Ns = Ns Sy 0 Ns | Ns 1 Pq int Prioritize requeued I/O. . +.It Sy zfs_delete_inode Ns = Ns Sy 0 Ns | Ns 1 Pq int +Sets whether the kernel should free an inode structure when the last reference +is released, or cache it in memory. +Intended for testing/debugging. +.Pp +A live inode structure "pins" versious internal OpenZFS structures in memory, +which can result in large amounts of "unusable" memory on systems with lots of +infrequently-accessed files, until the kernel's memory pressure mechanism +asks OpenZFS to release them. +.Pp +The default value of +.Sy 0 +always caches inodes that appear to still exist on disk. +Setting it to +.Sy 1 +will immediately release unused inodes and their associated memory back to the +dbuf cache or the ARC for reuse, but may reduce performance if inodes are +frequently evicted and reloaded. +.Pp +This parameter is only available on Linux. +. .It Sy zio_taskq_batch_pct Ns = Ns Sy 80 Ns % Pq uint Percentage of online CPUs which will run a worker thread for I/O. These workers are responsible for I/O work such as compression, encryption, diff --git a/module/os/linux/zfs/zpl_super.c b/module/os/linux/zfs/zpl_super.c index 538196286..591a4ee28 100644 --- a/module/os/linux/zfs/zpl_super.c +++ b/module/os/linux/zfs/zpl_super.c @@ -22,6 +22,7 @@ /* * Copyright (c) 2011, Lawrence Livermore National Security, LLC. * Copyright (c) 2023, Datto Inc. All rights reserved. + * Copyright (c) 2025, Klara, Inc. */ @@ -33,6 +34,12 @@ #include #include +/* + * What to do when the last reference to an inode is released. If 0, the kernel + * will cache it on the superblock. If 1, the inode will be freed immediately. + * See zpl_drop_inode(). + */ +int zfs_delete_inode = 0; static struct inode * zpl_inode_alloc(struct super_block *sb) @@ -77,11 +84,36 @@ zpl_dirty_inode(struct inode *ip, int flags) } /* - * When ->drop_inode() is called its return value indicates if the - * inode should be evicted from the inode cache. If the inode is - * unhashed and has no links the default policy is to evict it - * immediately. + * ->drop_inode() is called when the last reference to an inode is released. + * Its return value indicates if the inode should be destroyed immediately, or + * cached on the superblock structure. * + * By default (zfs_delete_inode=0), we call generic_drop_inode(), which returns + * "destroy immediately" if the inode is unhashed and has no links (roughly: no + * longer exists on disk). On datasets with millions of rarely-accessed files, + * this can cause a large amount of memory to be "pinned" by cached inodes, + * which in turn pin their associated dnodes and dbufs, until the kernel starts + * reporting memory pressure and requests OpenZFS release some memory (see + * zfs_prune()). + * + * When set to 1, we call generic_delete_node(), which always returns "destroy + * immediately", resulting in inodes being destroyed immediately, releasing + * their associated dnodes and dbufs to the dbuf cached and the ARC to be + * evicted as normal. + * + * Note that the "last reference" doesn't always mean the last _userspace_ + * reference; the dentry cache also holds a reference, so "busy" inodes will + * still be kept alive that way (subject to dcache tuning). + */ +static int +zpl_drop_inode(struct inode *ip) +{ + if (zfs_delete_inode) + return (generic_delete_inode(ip)); + return (generic_drop_inode(ip)); +} + +/* * The ->evict_inode() callback must minimally truncate the inode pages, * and call clear_inode(). For 2.6.35 and later kernels this will * simply update the inode state, with the sync occurring before the @@ -470,6 +502,7 @@ const struct super_operations zpl_super_operations = { .destroy_inode = zpl_inode_destroy, .dirty_inode = zpl_dirty_inode, .write_inode = NULL, + .drop_inode = zpl_drop_inode, .evict_inode = zpl_evict_inode, .put_super = zpl_put_super, .sync_fs = zpl_sync_fs, @@ -491,3 +524,6 @@ struct file_system_type zpl_fs_type = { .mount = zpl_mount, .kill_sb = zpl_kill_sb, }; + +ZFS_MODULE_PARAM(zfs, zfs_, delete_inode, INT, ZMOD_RW, + "Delete inodes as soon as the last reference is released.");