On Fri, 2025-06-13 at 08:31 +0100, Al Viro wrote:
> Another series pulled out of tree-in-dcache pile; it massages
> rpc_pipefs to use saner primitives. Lives in
> git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git #work.rpc_pipe
> 6.16-rc1-based, 17 commits, -314 lines of code...
>
> Individual patches in followups.
>
> Folks, please test and review. In various forms this has sat in my tree
> for more than a year and I'd rather get that logjam dealt with.
>
> Overview:
>
> Prep/infrastructure (#1 is shared with #work.simple_recursive_removal)
>
> 1) simple_recursive_removal(): saner interaction with fsnotify
> fsnotify events are triggered before the callback, unlike in real
> unlink(2)/rmdir(2) syscalls. Obviously matters only in case when callback
> is non-empty, which excludes most of the current users in the kernel.
>
> 2) new helper: simple_start_creating()
> Set the things up for kernel-initiated creation of object in a
> tree-in-dcache filesystem. With respect to locking it's an equivalent of
> filename_create() - we either get a negative dentry with locked parent,
> or ERR_PTR() and no locks taken.
> tracefs and debugfs had that open-coded as part of their object
> creation machinery; switched to calling new helper.
>
> rpc_pipefs proper:
>
> 3) rpc_pipe: clean failure exits in fill_super
> ->kill_sb() will be called immediately after a failure
> return anyway, so we don't need to bother with notifier chain and
> rpc_gssd_dummy_depopulate(). What's more, rpc_gssd_dummy_populate()
> failure exits do not need to bother with __rpc_depopulate() - anything
> added to the tree will be taken out by ->kill_sb().
>
> 4) rpc_{rmdir_,}depopulate(): use simple_recursive_removal() instead
> no need to give an exact list of objects to be remove when it's
> simply every child of the victim directory.
>
> 5) rpc_unlink(): use simple_recursive_removal()
> Previous commit was dealing with directories; this one is for
> named pipes (i.e. the thing rpc_pipefs is used for). Note that the
> callback of simple_recursive_removal() is called with the parent locked;
> the victim isn't locked by the caller.
>
> 6) rpc_populate(): lift cleanup into callers
> rpc_populate() is called either from fill_super (where we
> don't need to remove any files on failure - rpc_kill_sb() will take
> them all out anyway) or from rpc_mkdir_populate(), where we need to
> remove the directory we'd been trying to populate along with whatever
> we'd put into it before we failed. Simpler to combine that into
> simple_recursive_removal() there.
>
> 7) rpc_unlink(): saner calling conventions
> * pass it pipe instead of pipe->dentry
> * zero pipe->dentry afterwards
> * it always returns 0; why bother?
>
> 8) rpc_mkpipe_dentry(): saner calling conventions
> Instead of returning a dentry or ERR_PTR(-E...), return 0 and
> store dentry into pipe->dentry on success and return -E... on failure.
> Callers are happier that way...
>
> 9) rpc_pipe: don't overdo directory locking
> Don't try to hold directories locked more than VFS requires;
> lock just before getting a child to be made positive (using
> simple_start_creating()) and unlock as soon as the child is created.
> There's no benefit in keeping the parent locked while populating the
> child - it won't stop dcache lookups anyway.
>
> 10) rpc_pipe: saner primitive for creating subdirectories
> All users of __rpc_mkdir() have the same form - start_creating(),
> followed, in case of success, by __rpc_mkdir() and unlocking parent.
> Combine that into a single helper, expanding __rpc_mkdir() into it, along
> with the call of __rpc_create_common() in it.
> Don't mess with d_drop() + d_add() - just d_instantiate()
> and be done with that. The reason __rpc_create_common() goes for that
> dance is that dentry it gets might or might not be hashed; here we know
> it's hashed.
>
> 11) rpc_pipe: saner primitive for creating regular files
> rpc_new_file(); similar to rpc_new_dir(), except that here we
> pass file_operations as well. Callers don't care about dentry, just
> return 0 or -E...
>
> 12) rpc_mkpipe_dentry(): switch to start_creating()
> ... and make sure we set the rpc_pipe-private part of inode up
> before attaching it to dentry.
>
> 13) rpc_gssd_dummy_populate(): don't bother with rpc_populate()
> Just have it create gssd (in root), clntXX in gssd, then info
> and gssd in clntXX - all with explicit
> rpc_new_dir()/rpc_new_file()/rpc_mkpipe_dentry().
>
> 14) rpc_pipe: expand the calls of rpc_mkdir_populate()
> ... and get rid of convoluted callbacks.
>
> 15) rpc_new_dir(): the last argument is always NULL
> All callers except the one in rpc_populate() pass explicit NULL
> there; rpc_populate() passes its last argument ('private') instead,
> but in the only call of rpc_populate() that creates any subdirectories
> (when creating fixed subdirectories of root) private itself is NULL.
>
> 16) rpc_create_client_dir(): don't bother with rpc_populate()
> not for a single file...
>
> 17) rpc_create_client_dir(): return 0 or -E...
> Callers couldn't care less which dentry did we get - anything
> valid is treated as success.
Aside from a couple of minor nits, this all looks great, Al.
Pity you don't use git format-patch --cover-letter or we'd have the
aggregate diffstat! I imagine this shrinks the rpc_pipefs code
significantly.
You can add this to the pile.
Reviewed-by: Jeff Layton <jlayton@xxxxxxxxxx>
