Re: [PATCH v2 1/3] vfs: add inode iteration methods
From: Christian Brauner
Date: Fri Nov 20 2026 - 06:15:11 EST
On Mon, Nov 16, 2026 at 02:45:44PM +0800, Huang Shijie wrote:
> > struct inode_iter_contex ctx = INODE_ITER_CTX_INIT;
> >
> > while ((inode = sb_inode_iter()))) {
> > /* do stuff */
> > }
> >
> How to manage the sb->s_inode_list_lock in this code?
>
> I have no idea how to hide the sb->s_inode_list_lock in this code.
I think Christoph is right. The lock shouldn't be held across
iterations. Let the iterator take it, grab a reference to the next
inode, drop the lock again and hand the inode out. The reference keeps the
inode on s_inodes so the next call can continue from there.
This is how drop_pagecache_sb() does it today. Sm like (UNTESTED and not
nice-ified):
struct sb_inode_iter {
struct super_block *sb;
struct inode *inode;
bool done;
};
struct inode *sb_inode_iter_next(struct sb_inode_iter *it)
{
struct super_block *sb = it->sb;
struct inode *prev = it->inode, *inode;
if (it->done)
return NULL;
spin_lock(&sb->s_inode_list_lock);
inode = list_prepare_entry(prev, &sb->s_inodes, i_sb_list);
list_for_each_entry_continue(inode, &sb->s_inodes, i_sb_list) {
spin_lock(&inode->i_lock);
if (inode_state_read(inode) & (I_FREEING | I_WILL_FREE | I_NEW)) {
spin_unlock(&inode->i_lock);
continue;
}
__iget(inode);
spin_unlock(&inode->i_lock);
spin_unlock(&sb->s_inode_list_lock);
it->inode = inode;
iput(prev);
return inode;
}
spin_unlock(&sb->s_inode_list_lock);
it->inode = NULL;
it->done = true;
iput(prev);
return NULL;
}
void sb_inode_iter_end(struct sb_inode_iter *it)
{
iput(it->inode);
it->inode = NULL;
it->done = true;
}
and then drop_pagecache_sb() becomes:
struct sb_inode_iter it = { .sb = sb };
struct inode *inode;
while ((inode = sb_inode_iter_next(&it))) {
invalidate_mapping_pages(inode->i_mapping, 0, -1);
cond_resched();
}
Make it a cleanup class and a caller that breaks out early doesn't have
to remember sb_inode_iter_end() either.
This will end because new inodes are added at the head and the cursor
stays on the list. So these assumptions should get
VFS_WARN_ON_ONCE()-es.
Walkers that can't take a reference might need different solution.
I'm not sure if there are any?