Skip to contentSkip to conversation

fuse: do not evict dirty inodes

PATCH In review Maxim Patlasov 2 participants 3 messages Jun 3, 2014 – Aug 15, 2014 FUSE filesystem core

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fuse: do not evict dirty inodes

Commit 1e18bda8 added .write_inode method to the fuse super_operations. This allowed fuse to use the kernel infrastructure for writing out dirty metadata (mtime and ctime for now). However, given that .drop_inode was not redefined from the legacy generic_delete_inode(), on umount(2) generic_shutdown_super() led to the eviction of all inodes disregarding their state.

The patch removes .drop_inode definition from the fuse super_operations. This works because now iput_final() calls generic_drop_inode() and returns w/o evicting inode. This, in turn, allows generic_shutdown_super() to write dirty inodes by calling sync_filesystem().

---
 fs/fuse/inode.c |    1 -
 1 file changed, 1 deletion(-)
fs/fuse/inode.c +0 -1
index 754dcf2..ee017be 100644
@@ -791,7 +791,6 @@ static const struct super_operations fuse_super_operations = {
 	.destroy_inode  = fuse_destroy_inode,
 	.evict_inode	= fuse_evict_inode,
 	.write_inode	= fuse_write_inode,
-	.drop_inode	= generic_delete_inode,
 	.remount_fs	= fuse_remount_fs,
 	.put_super	= fuse_put_super,
 	.umount_begin	= fuse_umount_begin,
maintainer
#1 lore raw
Quoted 22 lines · Maxim Patlasov

Commit 1e18bda8 added .write_inode method to the fuse super_operations. This allowed fuse to use the kernel infrastructure for writing out dirty metadata (mtime and ctime for now). However, given that .drop_inode was not redefined from the legacy generic_delete_inode(), on umount(2) generic_shutdown_super() led to the eviction of all inodes disregarding their state.

The patch removes .drop_inode definition from the fuse super_operations. This works because now iput_final() calls generic_drop_inode() and returns w/o evicting inode. This, in turn, allows generic_shutdown_super() to write dirty inodes by calling sync_filesystem().

---
 fs/fuse/inode.c |    1 -
 1 file changed, 1 deletion(-)
fs/fuse/inode.c +0 -0
index 754dcf2..ee017be 100644

fs/fuse/inode.c static const struct super_operations fuse_super_operations = { · line 796

Quoted 7 lines
fs/fuse/inode.c +0 -1
@@ -791,7 +791,6 @@ static const struct super_operations fuse_super_operations = {
        .destroy_inode  = fuse_destroy_inode,
        .evict_inode    = fuse_evict_inode,
        .write_inode    = fuse_write_inode,
-       .drop_inode     = generic_delete_inode,
        .remount_fs     = fuse_remount_fs,
        .put_super      = fuse_put_super,
        .umount_begin   = fuse_umount_begin,

(Sorry about the late answer)

Big problem with this is that I don't want to make umount(2) and sync(2) wait on userspace filesystem. Generally this would make umount() hang if a fuse daemon was stuck for any reason.

But is this really necessary?

We are talking about just regular files: mtime is only updated by write(2) and friends. ctime is updated by write(2) as well as some other ops. For write, we can sync the times on FLUSH (close), for other ops we could flush the ctime synchronously. E.g. unlink would trigger UNLINK and SETATTR.

Long term, much better solution would be to add a timestamp to fuse_in_header which would remove the duplicate requests and then we could also extend the kernel caching of timestamps from just regular files to everything, which would make the protocol conceptually simpler.

Thanks,
Miklos

#2 lore raw

Hi Miklos,

Quoted 32 lines · Miklos Szeredi
Quoted 22 lines · Maxim Patlasov

Commit 1e18bda8 added .write_inode method to the fuse super_operations. This allowed fuse to use the kernel infrastructure for writing out dirty metadata (mtime and ctime for now). However, given that .drop_inode was not redefined from the legacy generic_delete_inode(), on umount(2) generic_shutdown_super() led to the eviction of all inodes disregarding their state.

The patch removes .drop_inode definition from the fuse super_operations. This works because now iput_final() calls generic_drop_inode() and returns w/o evicting inode. This, in turn, allows generic_shutdown_super() to write dirty inodes by calling sync_filesystem().

---
  fs/fuse/inode.c |    1 -
  1 file changed, 1 deletion(-)
fs/fuse/inode.c +0 -0
index 754dcf2..ee017be 100644

fs/fuse/inode.c static const struct super_operations fuse_super_operations = { · line 796

Quoted 7 lines
fs/fuse/inode.c +0 -1
@@ -791,7 +791,6 @@ static const struct super_operations fuse_super_operations = {
         .destroy_inode  = fuse_destroy_inode,
         .evict_inode    = fuse_evict_inode,
         .write_inode    = fuse_write_inode,
-       .drop_inode     = generic_delete_inode,
         .remount_fs     = fuse_remount_fs,
         .put_super      = fuse_put_super,
         .umount_begin   = fuse_umount_begin,

(Sorry about the late answer)

Quoted 3 lines

Big problem with this is that I don't want to make umount(2) and sync(2) wait on userspace filesystem. Generally this would make umount() hang if a fuse daemon was stuck for any reason.

I think we must honour interests both privileged and unprivileged mounts. In case of trusted environment where only sysad decides which fuse daemons are eligible to run, blocking umount() is completely fine. And more than that, after settling down that sync-close feature, I intend to take on synchronous umount (it's a shame that users have no tools to find out when umount completed better than monitoring of fuse daemon in proc table waiting for daemon termination). So we could put such a behaviour under control of a tunable parameter.

Quoted 9 lines

But is this really necessary?

We are talking about just regular files: mtime is only updated by write(2) and friends. ctime is updated by write(2) as well as some other ops. For write, we can sync the times on FLUSH (close), for other ops we could flush the ctime synchronously. E.g. unlink would trigger UNLINK and SETATTR.

Quoted 5 lines

Long term, much better solution would be to add a timestamp to fuse_in_header which would remove the duplicate requests and then we could also extend the kernel caching of timestamps from just regular files to everything, which would make the protocol conceptually simpler.

I like the idea of extending fuse_in_header. But if we'll extend it, why not go further adding timestamps for mtime and atime as well? Because pushing mtime along with data modifications is more reliable than postponing flush to write_inode call. Also, this would simplify the code a lot -- no more fuse_write_inode and fuse_flush_times needed. And who knows, may be someone will request proper atime handling at some point in future. If you're OK about:

Quoted 5 lines
(quoted diff) +1 -0
@@ -675,10 +675,15 @@ struct fuse_in_header {
        uint32_t        opcode;
        uint64_t        unique;
        uint64_t        nodeid;
+       uint64_t        atime;

Review struct fuse_in_header { · line 687

Quoted 10 lines
(quoted diff) +5 -1
+       uint64_t        mtime;
+       uint64_t        ctime;
+       uint32_t        atimensec;
+       uint32_t        mtimensec;
+       uint32_t        ctimensec;
        uint32_t        uid;
        uint32_t        gid;
        uint32_t        pid;
-       uint32_t        padding;
 };

I could work on the patch.

Thanks,
Maxim