Subject: tmpfs: Reclaiming vnodes
To: None <firstname.lastname@example.org>
From: Julio M. Merino Vidal <email@example.com>
Date: 08/03/2005 22:06:53
Hi all (and mentors),
The existing tmpfs code allocates new vnodes for each
access within the filesystem; e.g., the lookup operation gets a
new vnode for each match and returns it. This hasn't caused
problems until now (despite it's plain wrong).
During the past days, I've been implementing the rmdir vnode
operation. As this removes information from the file-system,
problems have started to pop up. Basically, multiple vnodes
were pointing to the same tmpfs_node structure; when the
structure was freed, those vnodes were left with dangling
pointers in them, causing memory faults in future accesses.
So I have to change the code to use a single vnode for a
So far, I've got a fix that avoids crashes, but I doubt it's correct
(comparing it to other file-systems' code). And even if it is,
I have several doubts that the manual pages (vnode(9) and
vnodeops(9)) do not solve.
First of all, I've changed the file-system specific node structure
(tmpfs_node) to keep a pointer to a vnode (let's call this
tn_vnode). When creating the node, tn_vnode is initialized to
NULL, because there is no vnode associated to it yet.
Later on, when tmpfs needs to get a vnode for a specific
tmpfs_node, it checks if tn_vnode is NULL. If it is, it gets a new
vnode (using getnewvnode), attaches it to the tmpfs_node and
uses it. If it's not, it calls vget on tn_vnode and uses that.
First question: several file-systems seem to do the same
(lookup for a existing vnode and return it), but they do _not_
have a call to vget (or at least I can't see it). However, if I do
not add this call, I get a panic (v_usecount =3D=3D -1). Why?
Furthermore, I'm afraid about using vget here because... who
warrants that the vnode is in the free list?
Then, I have problems with the rmdir operation itself. I first
detach a_vp from its tmpfs_node by setting v_data and
tn_node both to NULL. Then I call vput over a_vp, as other
file-systems do. After this, I free the associated tmpfs_node
given that no-one else is pointing to it.
*But* a_vp is not really "released". When I unmount the
file-system, the kernel calls the fsync and reclaim operations
over the value a_vp had (the vnode used during rmdir). As
this vnode has v_data set to NULL, any attempt to access the
structure obviously fails and panics the system.
I can workaround this by checking whether v_data is NULL or
not, but I don't see any other file-system doing this. What
should I do to solve this correctly? Should I tell the system
to completely forget about that vnode in some way?
Thank you very much.
Julio M. Merino Vidal <firstname.lastname@example.org>
The NetBSD Project - http://www.NetBSD.org/