kernfs: don't depend on d_find_any_alias() when generating notifications
Tejun Heo
maintainerPATCH 1/1 kernfs: don't depend on d_find_any_alias() when generating notifications
kernfs_notify_workfn() sends out file modified events for the scheduled kernfs_nodes. Because the modifications aren't from userland, it doesn't have the matching file struct at hand and can't use fsnotify_modify(). Instead, it looked up the inode and then used d_find_any_alias() to find the dentry and used fsnotify_parent() and fsnotify() directly to generate notifications.
The assumption was that the relevant dentries would have been pinned if there are listeners, which isn't true as inotify doesn't pin dentries at all and watching the parent doesn't pin the child dentries even for dnotify. This led to, for example, inotify watchers not getting notifications if the system is under memory pressure and the matching dentries got reclaimed. It can also be triggered through /proc/sys/vm/drop_caches or a remount attempt which involves shrinking dcache.
fsnotify_parent() only uses the dentry to access the parent inode, which kernfs can do easily. Update kernfs_notify_workfn() so that it uses fsnotify() directly for both the parent and target inodes without going through d_find_any_alias(). While at it, supply the target file name to fsnotify() from kernfs_node->name.
---
Hello,
I'm not sure this is the best way to deal with this but it at least works fine. If there's a better, please let me know. If this approach is okay, in the future, maybe we want to implement a helper on fsnotify side to handle notification generation from back-end side?
Thanks.
fs/kernfs/file.c | 28 +++++++++++++++++++++------- 1 file changed, 21 insertions(+), 7 deletions(-)
fs/kernfs/file.c +21 -7
@@ -840,21 +840,35 @@ static void kernfs_notify_workfn(struct work_struct *work)mutex_lock (&kernfs_mutex);list_for_each_entry (info, &kernfs_root (kn)->supers, node) { +struct kernfs_node *parent;struct inode *inode; -struct dentry *dentry; +/* +* We want fsnotify_modify() on @kn but as the +* modifications aren't originating from userland don't +* have the matching @file available. Look up the inodes +* and generate the events manually. +*/ inode =ilookup (info->sb, kn->ino);if (!inode)continue ; - dentry =d_find_any_alias (inode); -if (dentry) { -fsnotify_parent (NULL , dentry,FS_MODIFY ); -fsnotify (inode,FS_MODIFY , inode,FSNOTIFY_EVENT_INODE , -NULL ,0 ); -dput (dentry); + parent =kernfs_get_parent (kn); +if (parent) { +struct inode *p_inode; + + p_inode =ilookup (info->sb, parent->ino); +if (p_inode) { +fsnotify (p_inode,FS_MODIFY |FS_EVENT_ON_CHILD , + inode,FSNOTIFY_EVENT_INODE , kn->name,0 ); +iput (p_inode); + } + +kernfs_put (parent); } +fsnotify (inode,FS_MODIFY , inode,FSNOTIFY_EVENT_INODE , + kn->name,0 );iput (inode); }
Tejun Heo
maintainerQuoted 29 lines · Tejun Heo
kernfs_notify_workfn()sends out file modified events for the scheduledkernfs_nodes. Because the modifications aren't from userland, it doesn't have the matching file struct at hand and can't usefsnotify_modify(). Instead, it looked up the inode and then usedd_find_any_alias()to find the dentry and usedfsnotify_parent()andfsnotify()directly to generate notifications.The assumption was that the relevant dentries would have been pinned if there are listeners, which isn't true as inotify doesn't pin dentries at all and watching the parent doesn't pin the child dentries even for dnotify. This led to, for example, inotify watchers not getting notifications if the system is under memory pressure and the matching dentries got reclaimed. It can also be triggered through
/proc/sys/vm/drop_cachesor a remount attempt which involves shrinking dcache.
fsnotify_parent()only uses the dentry to access the parent inode, which kernfs can do easily. Updatekernfs_notify_workfn()so that it usesfsnotify()directly for both the parent and target inodes without going throughd_find_any_alias(). While at it, supply the target file name tofsnotify()fromkernfs_node->name.Signed-off-by Tejun Heo <[email protected]>Reported-by Evgeny Vereshchagin <[email protected]>Fixes d911d9874801 ("kernfs: make kernfs_notify() trigger inotify events too")Cc John McCutchan <[email protected]>Cc Robert Love <[email protected]>Cc Eric Paris <[email protected]>Cc [email protected] # v3.16+
Quoted 7 lines
---
Hello,I'm not sure this is the best way to deal with this but it at least works fine. If there's a better, please let me know. If this approach is okay, in the future, maybe we want to implement a helper on fsnotify side to handle notification generation from back-end side?
Greg, can you please pick this one up?
Thanks.
Signature · 1 line
tejun
Greg Kroah-Hartman
maintainerQuoted 33 lines · Tejun Heo
Quoted 32 lines · Tejun Heo
kernfs_notify_workfn()sends out file modified events for the scheduledkernfs_nodes. Because the modifications aren't from userland, it doesn't have the matching file struct at hand and can't usefsnotify_modify(). Instead, it looked up the inode and then usedd_find_any_alias()to find the dentry and usedfsnotify_parent()andfsnotify()directly to generate notifications.The assumption was that the relevant dentries would have been pinned if there are listeners, which isn't true as inotify doesn't pin dentries at all and watching the parent doesn't pin the child dentries even for dnotify. This led to, for example, inotify watchers not getting notifications if the system is under memory pressure and the matching dentries got reclaimed. It can also be triggered through
/proc/sys/vm/drop_cachesor a remount attempt which involves shrinking dcache.
fsnotify_parent()only uses the dentry to access the parent inode, which kernfs can do easily. Updatekernfs_notify_workfn()so that it usesfsnotify()directly for both the parent and target inodes without going throughd_find_any_alias(). While at it, supply the target file name tofsnotify()fromkernfs_node->name.Signed-off-by Tejun Heo <[email protected]>Reported-by Evgeny Vereshchagin <[email protected]>Fixes d911d9874801 ("kernfs: make kernfs_notify() trigger inotify events too")Cc John McCutchan <[email protected]>Cc Robert Love <[email protected]>Cc Eric Paris <[email protected]>Cc [email protected] # v3.16+---
Hello,
Quoted 6 lines
Quoted 4 lines
I'm not sure this is the best way to deal with this but it at least works fine. If there's a better, please let me know. If this approach is okay, in the future, maybe we want to implement a helper on fsnotify side to handle notification generation from back-end side?
Greg, can you please pick this one up?
Will do, thanks.
greg k-h