summaryrefslogtreecommitdiff
path: root/Documentation/filesystems/porting
diff options
context:
space:
mode:
Diffstat (limited to 'Documentation/filesystems/porting')
-rw-r--r--Documentation/filesystems/porting686
1 files changed, 0 insertions, 686 deletions
diff --git a/Documentation/filesystems/porting b/Documentation/filesystems/porting
deleted file mode 100644
index 6b7a41cfcaed..000000000000
--- a/Documentation/filesystems/porting
+++ /dev/null
@@ -1,686 +0,0 @@
-Changes since 2.5.0:
-
----
-[recommended]
-
-New helpers: sb_bread(), sb_getblk(), sb_find_get_block(), set_bh(),
- sb_set_blocksize() and sb_min_blocksize().
-
-Use them.
-
-(sb_find_get_block() replaces 2.4's get_hash_table())
-
----
-[recommended]
-
-New methods: ->alloc_inode() and ->destroy_inode().
-
-Remove inode->u.foo_inode_i
-Declare
- struct foo_inode_info {
- /* fs-private stuff */
- struct inode vfs_inode;
- };
- static inline struct foo_inode_info *FOO_I(struct inode *inode)
- {
- return list_entry(inode, struct foo_inode_info, vfs_inode);
- }
-
-Use FOO_I(inode) instead of &inode->u.foo_inode_i;
-
-Add foo_alloc_inode() and foo_destroy_inode() - the former should allocate
-foo_inode_info and return the address of ->vfs_inode, the latter should free
-FOO_I(inode) (see in-tree filesystems for examples).
-
-Make them ->alloc_inode and ->destroy_inode in your super_operations.
-
-Keep in mind that now you need explicit initialization of private data
-typically between calling iget_locked() and unlocking the inode.
-
-At some point that will become mandatory.
-
----
-[mandatory]
-
-Change of file_system_type method (->read_super to ->get_sb)
-
-->read_super() is no more. Ditto for DECLARE_FSTYPE and DECLARE_FSTYPE_DEV.
-
-Turn your foo_read_super() into a function that would return 0 in case of
-success and negative number in case of error (-EINVAL unless you have more
-informative error value to report). Call it foo_fill_super(). Now declare
-
-int foo_get_sb(struct file_system_type *fs_type,
- int flags, const char *dev_name, void *data, struct vfsmount *mnt)
-{
- return get_sb_bdev(fs_type, flags, dev_name, data, foo_fill_super,
- mnt);
-}
-
-(or similar with s/bdev/nodev/ or s/bdev/single/, depending on the kind of
-filesystem).
-
-Replace DECLARE_FSTYPE... with explicit initializer and have ->get_sb set as
-foo_get_sb.
-
----
-[mandatory]
-
-Locking change: ->s_vfs_rename_sem is taken only by cross-directory renames.
-Most likely there is no need to change anything, but if you relied on
-global exclusion between renames for some internal purpose - you need to
-change your internal locking. Otherwise exclusion warranties remain the
-same (i.e. parents and victim are locked, etc.).
-
----
-[informational]
-
-Now we have the exclusion between ->lookup() and directory removal (by
-->rmdir() and ->rename()). If you used to need that exclusion and do
-it by internal locking (most of filesystems couldn't care less) - you
-can relax your locking.
-
----
-[mandatory]
-
-->lookup(), ->truncate(), ->create(), ->unlink(), ->mknod(), ->mkdir(),
-->rmdir(), ->link(), ->lseek(), ->symlink(), ->rename()
-and ->readdir() are called without BKL now. Grab it on entry, drop upon return
-- that will guarantee the same locking you used to have. If your method or its
-parts do not need BKL - better yet, now you can shift lock_kernel() and
-unlock_kernel() so that they would protect exactly what needs to be
-protected.
-
----
-[mandatory]
-
-BKL is also moved from around sb operations. BKL should have been shifted into
-individual fs sb_op functions. If you don't need it, remove it.
-
----
-[informational]
-
-check for ->link() target not being a directory is done by callers. Feel
-free to drop it...
-
----
-[informational]
-
-->link() callers hold ->i_mutex on the object we are linking to. Some of your
-problems might be over...
-
----
-[mandatory]
-
-new file_system_type method - kill_sb(superblock). If you are converting
-an existing filesystem, set it according to ->fs_flags:
- FS_REQUIRES_DEV - kill_block_super
- FS_LITTER - kill_litter_super
- neither - kill_anon_super
-FS_LITTER is gone - just remove it from fs_flags.
-
----
-[mandatory]
-
- FS_SINGLE is gone (actually, that had happened back when ->get_sb()
-went in - and hadn't been documented ;-/). Just remove it from fs_flags
-(and see ->get_sb() entry for other actions).
-
----
-[mandatory]
-
-->setattr() is called without BKL now. Caller _always_ holds ->i_mutex, so
-watch for ->i_mutex-grabbing code that might be used by your ->setattr().
-Callers of notify_change() need ->i_mutex now.
-
----
-[recommended]
-
-New super_block field "struct export_operations *s_export_op" for
-explicit support for exporting, e.g. via NFS. The structure is fully
-documented at its declaration in include/linux/fs.h, and in
-Documentation/filesystems/nfs/Exporting.
-
-Briefly it allows for the definition of decode_fh and encode_fh operations
-to encode and decode filehandles, and allows the filesystem to use
-a standard helper function for decode_fh, and provide file-system specific
-support for this helper, particularly get_parent.
-
-It is planned that this will be required for exporting once the code
-settles down a bit.
-
-[mandatory]
-
-s_export_op is now required for exporting a filesystem.
-isofs, ext2, ext3, resierfs, fat
-can be used as examples of very different filesystems.
-
----
-[mandatory]
-
-iget4() and the read_inode2 callback have been superseded by iget5_locked()
-which has the following prototype,
-
- struct inode *iget5_locked(struct super_block *sb, unsigned long ino,
- int (*test)(struct inode *, void *),
- int (*set)(struct inode *, void *),
- void *data);
-
-'test' is an additional function that can be used when the inode
-number is not sufficient to identify the actual file object. 'set'
-should be a non-blocking function that initializes those parts of a
-newly created inode to allow the test function to succeed. 'data' is
-passed as an opaque value to both test and set functions.
-
-When the inode has been created by iget5_locked(), it will be returned with the
-I_NEW flag set and will still be locked. The filesystem then needs to finalize
-the initialization. Once the inode is initialized it must be unlocked by
-calling unlock_new_inode().
-
-The filesystem is responsible for setting (and possibly testing) i_ino
-when appropriate. There is also a simpler iget_locked function that
-just takes the superblock and inode number as arguments and does the
-test and set for you.
-
-e.g.
- inode = iget_locked(sb, ino);
- if (inode->i_state & I_NEW) {
- err = read_inode_from_disk(inode);
- if (err < 0) {
- iget_failed(inode);
- return err;
- }
- unlock_new_inode(inode);
- }
-
-Note that if the process of setting up a new inode fails, then iget_failed()
-should be called on the inode to render it dead, and an appropriate error
-should be passed back to the caller.
-
----
-[recommended]
-
-->getattr() finally getting used. See instances in nfs, minix, etc.
-
----
-[mandatory]
-
-->revalidate() is gone. If your filesystem had it - provide ->getattr()
-and let it call whatever you had as ->revlidate() + (for symlinks that
-had ->revalidate()) add calls in ->follow_link()/->readlink().
-
----
-[mandatory]
-
-->d_parent changes are not protected by BKL anymore. Read access is safe
-if at least one of the following is true:
- * filesystem has no cross-directory rename()
- * we know that parent had been locked (e.g. we are looking at
-->d_parent of ->lookup() argument).
- * we are called from ->rename().
- * the child's ->d_lock is held
-Audit your code and add locking if needed. Notice that any place that is
-not protected by the conditions above is risky even in the old tree - you
-had been relying on BKL and that's prone to screwups. Old tree had quite
-a few holes of that kind - unprotected access to ->d_parent leading to
-anything from oops to silent memory corruption.
-
----
-[mandatory]
-
- FS_NOMOUNT is gone. If you use it - just set SB_NOUSER in flags
-(see rootfs for one kind of solution and bdev/socket/pipe for another).
-
----
-[recommended]
-
- Use bdev_read_only(bdev) instead of is_read_only(kdev). The latter
-is still alive, but only because of the mess in drivers/s390/block/dasd.c.
-As soon as it gets fixed is_read_only() will die.
-
----
-[mandatory]
-
-->permission() is called without BKL now. Grab it on entry, drop upon
-return - that will guarantee the same locking you used to have. If
-your method or its parts do not need BKL - better yet, now you can
-shift lock_kernel() and unlock_kernel() so that they would protect
-exactly what needs to be protected.
-
----
-[mandatory]
-
-->statfs() is now called without BKL held. BKL should have been
-shifted into individual fs sb_op functions where it's not clear that
-it's safe to remove it. If you don't need it, remove it.
-
----
-[mandatory]
-
- is_read_only() is gone; use bdev_read_only() instead.
-
----
-[mandatory]
-
- destroy_buffers() is gone; use invalidate_bdev().
-
----
-[mandatory]
-
- fsync_dev() is gone; use fsync_bdev(). NOTE: lvm breakage is
-deliberate; as soon as struct block_device * is propagated in a reasonable
-way by that code fixing will become trivial; until then nothing can be
-done.
-
-[mandatory]
-
- block truncatation on error exit from ->write_begin, and ->direct_IO
-moved from generic methods (block_write_begin, cont_write_begin,
-nobh_write_begin, blockdev_direct_IO*) to callers. Take a look at
-ext2_write_failed and callers for an example.
-
-[mandatory]
-
- ->truncate is gone. The whole truncate sequence needs to be
-implemented in ->setattr, which is now mandatory for filesystems
-implementing on-disk size changes. Start with a copy of the old inode_setattr
-and vmtruncate, and the reorder the vmtruncate + foofs_vmtruncate sequence to
-be in order of zeroing blocks using block_truncate_page or similar helpers,
-size update and on finally on-disk truncation which should not fail.
-setattr_prepare (which used to be inode_change_ok) now includes the size checks
-for ATTR_SIZE and must be called in the beginning of ->setattr unconditionally.
-
-[mandatory]
-
- ->clear_inode() and ->delete_inode() are gone; ->evict_inode() should
-be used instead. It gets called whenever the inode is evicted, whether it has
-remaining links or not. Caller does *not* evict the pagecache or inode-associated
-metadata buffers; the method has to use truncate_inode_pages_final() to get rid
-of those. Caller makes sure async writeback cannot be running for the inode while
-(or after) ->evict_inode() is called.
-
- ->drop_inode() returns int now; it's called on final iput() with
-inode->i_lock held and it returns true if filesystems wants the inode to be
-dropped. As before, generic_drop_inode() is still the default and it's been
-updated appropriately. generic_delete_inode() is also alive and it consists
-simply of return 1. Note that all actual eviction work is done by caller after
-->drop_inode() returns.
-
- As before, clear_inode() must be called exactly once on each call of
-->evict_inode() (as it used to be for each call of ->delete_inode()). Unlike
-before, if you are using inode-associated metadata buffers (i.e.
-mark_buffer_dirty_inode()), it's your responsibility to call
-invalidate_inode_buffers() before clear_inode().
-
- NOTE: checking i_nlink in the beginning of ->write_inode() and bailing out
-if it's zero is not *and* *never* *had* *been* enough. Final unlink() and iput()
-may happen while the inode is in the middle of ->write_inode(); e.g. if you blindly
-free the on-disk inode, you may end up doing that while ->write_inode() is writing
-to it.
-
----
-[mandatory]
-
- .d_delete() now only advises the dcache as to whether or not to cache
-unreferenced dentries, and is now only called when the dentry refcount goes to
-0. Even on 0 refcount transition, it must be able to tolerate being called 0,
-1, or more times (eg. constant, idempotent).
-
----
-[mandatory]
-
- .d_compare() calling convention and locking rules are significantly
-changed. Read updated documentation in Documentation/filesystems/vfs.rst (and
-look at examples of other filesystems) for guidance.
-
----
-[mandatory]
-
- .d_hash() calling convention and locking rules are significantly
-changed. Read updated documentation in Documentation/filesystems/vfs.rst (and
-look at examples of other filesystems) for guidance.
-
----
-[mandatory]
- dcache_lock is gone, replaced by fine grained locks. See fs/dcache.c
-for details of what locks to replace dcache_lock with in order to protect
-particular things. Most of the time, a filesystem only needs ->d_lock, which
-protects *all* the dcache state of a given dentry.
-
---
-[mandatory]
-
- Filesystems must RCU-free their inodes, if they can have been accessed
-via rcu-walk path walk (basically, if the file can have had a path name in the
-vfs namespace).
-
- Even though i_dentry and i_rcu share storage in a union, we will
-initialize the former in inode_init_always(), so just leave it alone in
-the callback. It used to be necessary to clean it there, but not anymore
-(starting at 3.2).
-
---
-[recommended]
- vfs now tries to do path walking in "rcu-walk mode", which avoids
-atomic operations and scalability hazards on dentries and inodes (see
-Documentation/filesystems/path-lookup.txt). d_hash and d_compare changes
-(above) are examples of the changes required to support this. For more complex
-filesystem callbacks, the vfs drops out of rcu-walk mode before the fs call, so
-no changes are required to the filesystem. However, this is costly and loses
-the benefits of rcu-walk mode. We will begin to add filesystem callbacks that
-are rcu-walk aware, shown below. Filesystems should take advantage of this
-where possible.
-
---
-[mandatory]
- d_revalidate is a callback that is made on every path element (if
-the filesystem provides it), which requires dropping out of rcu-walk mode. This
-may now be called in rcu-walk mode (nd->flags & LOOKUP_RCU). -ECHILD should be
-returned if the filesystem cannot handle rcu-walk. See
-Documentation/filesystems/vfs.rst for more details.
-
- permission is an inode permission check that is called on many or all
-directory inodes on the way down a path walk (to check for exec permission). It
-must now be rcu-walk aware (mask & MAY_NOT_BLOCK). See
-Documentation/filesystems/vfs.rst for more details.
-
---
-[mandatory]
- In ->fallocate() you must check the mode option passed in. If your
-filesystem does not support hole punching (deallocating space in the middle of a
-file) you must return -EOPNOTSUPP if FALLOC_FL_PUNCH_HOLE is set in mode.
-Currently you can only have FALLOC_FL_PUNCH_HOLE with FALLOC_FL_KEEP_SIZE set,
-so the i_size should not change when hole punching, even when puching the end of
-a file off.
-
---
-[mandatory]
- ->get_sb() is gone. Switch to use of ->mount(). Typically it's just
-a matter of switching from calling get_sb_... to mount_... and changing the
-function type. If you were doing it manually, just switch from setting ->mnt_root
-to some pointer to returning that pointer. On errors return ERR_PTR(...).
-
---
-[mandatory]
- ->permission() and generic_permission()have lost flags
-argument; instead of passing IPERM_FLAG_RCU we add MAY_NOT_BLOCK into mask.
- generic_permission() has also lost the check_acl argument; ACL checking
-has been taken to VFS and filesystems need to provide a non-NULL ->i_op->get_acl
-to read an ACL from disk.
-
---
-[mandatory]
- If you implement your own ->llseek() you must handle SEEK_HOLE and
-SEEK_DATA. You can hanle this by returning -EINVAL, but it would be nicer to
-support it in some way. The generic handler assumes that the entire file is
-data and there is a virtual hole at the end of the file. So if the provided
-offset is less than i_size and SEEK_DATA is specified, return the same offset.
-If the above is true for the offset and you are given SEEK_HOLE, return the end
-of the file. If the offset is i_size or greater return -ENXIO in either case.
-
-[mandatory]
- If you have your own ->fsync() you must make sure to call
-filemap_write_and_wait_range() so that all dirty pages are synced out properly.
-You must also keep in mind that ->fsync() is not called with i_mutex held
-anymore, so if you require i_mutex locking you must make sure to take it and
-release it yourself.
-
---
-[mandatory]
- d_alloc_root() is gone, along with a lot of bugs caused by code
-misusing it. Replacement: d_make_root(inode). On success d_make_root(inode)
-allocates and returns a new dentry instantiated with the passed in inode.
-On failure NULL is returned and the passed in inode is dropped so the reference
-to inode is consumed in all cases and failure handling need not do any cleanup
-for the inode. If d_make_root(inode) is passed a NULL inode it returns NULL
-and also requires no further error handling. Typical usage is:
-
- inode = foofs_new_inode(....);
- s->s_root = d_make_root(inode);
- if (!s->s_root)
- /* Nothing needed for the inode cleanup */
- return -ENOMEM;
- ...
-
---
-[mandatory]
- The witch is dead! Well, 2/3 of it, anyway. ->d_revalidate() and
-->lookup() do *not* take struct nameidata anymore; just the flags.
---
-[mandatory]
- ->create() doesn't take struct nameidata *; unlike the previous
-two, it gets "is it an O_EXCL or equivalent?" boolean argument. Note that
-local filesystems can ignore tha argument - they are guaranteed that the
-object doesn't exist. It's remote/distributed ones that might care...
---
-[mandatory]
- FS_REVAL_DOT is gone; if you used to have it, add ->d_weak_revalidate()
-in your dentry operations instead.
---
-[mandatory]
- vfs_readdir() is gone; switch to iterate_dir() instead
---
-[mandatory]
- ->readdir() is gone now; switch to ->iterate()
-[mandatory]
- vfs_follow_link has been removed. Filesystems must use nd_set_link
- from ->follow_link for normal symlinks, or nd_jump_link for magic
- /proc/<pid> style links.
---
-[mandatory]
- iget5_locked()/ilookup5()/ilookup5_nowait() test() callback used to be
- called with both ->i_lock and inode_hash_lock held; the former is *not*
- taken anymore, so verify that your callbacks do not rely on it (none
- of the in-tree instances did). inode_hash_lock is still held,
- of course, so they are still serialized wrt removal from inode hash,
- as well as wrt set() callback of iget5_locked().
---
-[mandatory]
- d_materialise_unique() is gone; d_splice_alias() does everything you
- need now. Remember that they have opposite orders of arguments ;-/
---
-[mandatory]
- f_dentry is gone; use f_path.dentry, or, better yet, see if you can avoid
- it entirely.
---
-[mandatory]
- never call ->read() and ->write() directly; use __vfs_{read,write} or
- wrappers; instead of checking for ->write or ->read being NULL, look for
- FMODE_CAN_{WRITE,READ} in file->f_mode.
---
-[mandatory]
- do _not_ use new_sync_{read,write} for ->read/->write; leave it NULL
- instead.
---
-[mandatory]
- ->aio_read/->aio_write are gone. Use ->read_iter/->write_iter.
----
-[recommended]
- for embedded ("fast") symlinks just set inode->i_link to wherever the
- symlink body is and use simple_follow_link() as ->follow_link().
---
-[mandatory]
- calling conventions for ->follow_link() have changed. Instead of returning
- cookie and using nd_set_link() to store the body to traverse, we return
- the body to traverse and store the cookie using explicit void ** argument.
- nameidata isn't passed at all - nd_jump_link() doesn't need it and
- nd_[gs]et_link() is gone.
---
-[mandatory]
- calling conventions for ->put_link() have changed. It gets inode instead of
- dentry, it does not get nameidata at all and it gets called only when cookie
- is non-NULL. Note that link body isn't available anymore, so if you need it,
- store it as cookie.
---
-[mandatory]
- any symlink that might use page_follow_link_light/page_put_link() must
- have inode_nohighmem(inode) called before anything might start playing with
- its pagecache. No highmem pages should end up in the pagecache of such
- symlinks. That includes any preseeding that might be done during symlink
- creation. __page_symlink() will honour the mapping gfp flags, so once
- you've done inode_nohighmem() it's safe to use, but if you allocate and
- insert the page manually, make sure to use the right gfp flags.
---
-[mandatory]
- ->follow_link() is replaced with ->get_link(); same API, except that
- * ->get_link() gets inode as a separate argument
- * ->get_link() may be called in RCU mode - in that case NULL
- dentry is passed
---
-[mandatory]
- ->get_link() gets struct delayed_call *done now, and should do
- set_delayed_call() where it used to set *cookie.
- ->put_link() is gone - just give the destructor to set_delayed_call()
- in ->get_link().
---
-[mandatory]
- ->getxattr() and xattr_handler.get() get dentry and inode passed separately.
- dentry might be yet to be attached to inode, so do _not_ use its ->d_inode
- in the instances. Rationale: !@#!@# security_d_instantiate() needs to be
- called before we attach dentry to inode.
---
-[mandatory]
- symlinks are no longer the only inodes that do *not* have i_bdev/i_cdev/
- i_pipe/i_link union zeroed out at inode eviction. As the result, you can't
- assume that non-NULL value in ->i_nlink at ->destroy_inode() implies that
- it's a symlink. Checking ->i_mode is really needed now. In-tree we had
- to fix shmem_destroy_callback() that used to take that kind of shortcut;
- watch out, since that shortcut is no longer valid.
---
-[mandatory]
- ->i_mutex is replaced with ->i_rwsem now. inode_lock() et.al. work as
- they used to - they just take it exclusive. However, ->lookup() may be
- called with parent locked shared. Its instances must not
- * use d_instantiate) and d_rehash() separately - use d_add() or
- d_splice_alias() instead.
- * use d_rehash() alone - call d_add(new_dentry, NULL) instead.
- * in the unlikely case when (read-only) access to filesystem
- data structures needs exclusion for some reason, arrange it
- yourself. None of the in-tree filesystems needed that.
- * rely on ->d_parent and ->d_name not changing after dentry has
- been fed to d_add() or d_splice_alias(). Again, none of the
- in-tree instances relied upon that.
- We are guaranteed that lookups of the same name in the same directory
- will not happen in parallel ("same" in the sense of your ->d_compare()).
- Lookups on different names in the same directory can and do happen in
- parallel now.
---
-[recommended]
- ->iterate_shared() is added; it's a parallel variant of ->iterate().
- Exclusion on struct file level is still provided (as well as that
- between it and lseek on the same struct file), but if your directory
- has been opened several times, you can get these called in parallel.
- Exclusion between that method and all directory-modifying ones is
- still provided, of course.
-
- Often enough ->iterate() can serve as ->iterate_shared() without any
- changes - it is a read-only operation, after all. If you have any
- per-inode or per-dentry in-core data structures modified by ->iterate(),
- you might need something to serialize the access to them. If you
- do dcache pre-seeding, you'll need to switch to d_alloc_parallel() for
- that; look for in-tree examples.
-
- Old method is only used if the new one is absent; eventually it will
- be removed. Switch while you still can; the old one won't stay.
---
-[mandatory]
- ->atomic_open() calls without O_CREAT may happen in parallel.
---
-[mandatory]
- ->setxattr() and xattr_handler.set() get dentry and inode passed separately.
- dentry might be yet to be attached to inode, so do _not_ use its ->d_inode
- in the instances. Rationale: !@#!@# security_d_instantiate() needs to be
- called before we attach dentry to inode and !@#!@##!@$!$#!@#$!@$!@$ smack
- ->d_instantiate() uses not just ->getxattr() but ->setxattr() as well.
---
-[mandatory]
- ->d_compare() doesn't get parent as a separate argument anymore. If you
- used it for finding the struct super_block involved, dentry->d_sb will
- work just as well; if it's something more complicated, use dentry->d_parent.
- Just be careful not to assume that fetching it more than once will yield
- the same value - in RCU mode it could change under you.
---
-[mandatory]
- ->rename() has an added flags argument. Any flags not handled by the
- filesystem should result in EINVAL being returned.
---
-[recommended]
- ->readlink is optional for symlinks. Don't set, unless filesystem needs
- to fake something for readlink(2).
---
-[mandatory]
- ->getattr() is now passed a struct path rather than a vfsmount and
- dentry separately, and it now has request_mask and query_flags arguments
- to specify the fields and sync type requested by statx. Filesystems not
- supporting any statx-specific features may ignore the new arguments.
---
-[mandatory]
- ->atomic_open() calling conventions have changed. Gone is int *opened,
- along with FILE_OPENED/FILE_CREATED. In place of those we have
- FMODE_OPENED/FMODE_CREATED, set in file->f_mode. Additionally, return
- value for 'called finish_no_open(), open it yourself' case has become
- 0, not 1. Since finish_no_open() itself is returning 0 now, that part
- does not need any changes in ->atomic_open() instances.
---
-[mandatory]
- alloc_file() has become static now; two wrappers are to be used instead.
- alloc_file_pseudo(inode, vfsmount, name, flags, ops) is for the cases
- when dentry needs to be created; that's the majority of old alloc_file()
- users. Calling conventions: on success a reference to new struct file
- is returned and callers reference to inode is subsumed by that. On
- failure, ERR_PTR() is returned and no caller's references are affected,
- so the caller needs to drop the inode reference it held.
- alloc_file_clone(file, flags, ops) does not affect any caller's references.
- On success you get a new struct file sharing the mount/dentry with the
- original, on failure - ERR_PTR().
---
-[mandatory]
- ->clone_file_range() and ->dedupe_file_range have been replaced with
- ->remap_file_range(). See Documentation/filesystems/vfs.rst for more
- information.
---
-[recommended]
- ->lookup() instances doing an equivalent of
- if (IS_ERR(inode))
- return ERR_CAST(inode);
- return d_splice_alias(inode, dentry);
- don't need to bother with the check - d_splice_alias() will do the
- right thing when given ERR_PTR(...) as inode. Moreover, passing NULL
- inode to d_splice_alias() will also do the right thing (equivalent of
- d_add(dentry, NULL); return NULL;), so that kind of special cases
- also doesn't need a separate treatment.
---
-[strongly recommended]
- take the RCU-delayed parts of ->destroy_inode() into a new method -
- ->free_inode(). If ->destroy_inode() becomes empty - all the better,
- just get rid of it. Synchronous work (e.g. the stuff that can't
- be done from an RCU callback, or any WARN_ON() where we want the
- stack trace) *might* be movable to ->evict_inode(); however,
- that goes only for the things that are not needed to balance something
- done by ->alloc_inode(). IOW, if it's cleaning up the stuff that
- might have accumulated over the life of in-core inode, ->evict_inode()
- might be a fit.
-
- Rules for inode destruction:
- * if ->destroy_inode() is non-NULL, it gets called
- * if ->free_inode() is non-NULL, it gets scheduled by call_rcu()
- * combination of NULL ->destroy_inode and NULL ->free_inode is
- treated as NULL/free_inode_nonrcu, to preserve the compatibility.
-
- Note that the callback (be it via ->free_inode() or explicit call_rcu()
- in ->destroy_inode()) is *NOT* ordered wrt superblock destruction;
- as the matter of fact, the superblock and all associated structures
- might be already gone. The filesystem driver is guaranteed to be still
- there, but that's it. Freeing memory in the callback is fine; doing
- more than that is possible, but requires a lot of care and is best
- avoided.
---
-[mandatory]
- DCACHE_RCUACCESS is gone; having an RCU delay on dentry freeing is the
- default. DCACHE_NORCU opts out, and only d_alloc_pseudo() has any
- business doing so.
---
-[mandatory]
- d_alloc_pseudo() is internal-only; uses outside of alloc_file_pseudo() are
- very suspect (and won't work in modules). Such uses are very likely to
- be misspelled d_alloc_anon().