gfs2: clean up iopen glock mess in gfs2_create_inode
PATCH v3 1/1 gfs2: clean up iopen glock mess in gfs2_create_inode
In reply to a message not in the archiveHi,
Here is my latest version of this patch that fixes some issues Andreas pointed out.
Bob
---
Before this patch, gfs2_create_inode had a use-after-free for the iopen glock in some error paths because it did this:
gfs2_glock_put(io_gl); fail_gunlock2: if (io_gl) clear_bit(GLF_INODE_CREATING, &io_gl->gl_flags);
In some cases, the io_gl was used for create and only had one reference, so the glock might be freed before the clear_bit(). This patch tries to straighten it out by only jumping to the error paths where iopen is properly set, and moving the gfs2_glock_put after the clear_bit.
--- fs/gfs2/inode.c | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-)
fs/gfs2/inode.c +7 -6
@@ -712,7 +712,7 @@ static int gfs2_create_inode(struct inode *dir, struct dentry *dentry, error =gfs2_trans_begin (sdp, blocks,0 );if (error) - goto fail_gunlock2; + goto fail_free_inode;if (blocks >1 ) { ip->i_eattr = ip->i_no_addr +1 ;
@@ -723,7 +723,7 @@ static int gfs2_create_inode(struct inode *dir, struct dentry *dentry, error =gfs2_glock_get (sdp, ip->i_no_addr, &gfs2_iopen_glops,CREATE , &io_gl);if (error) - goto fail_gunlock2; + goto fail_free_inode;BUG_ON (test_and_set_bit (GLF_INODE_CREATING , &io_gl->gl_flags));
@@ -732,7 +732,6 @@ static int gfs2_create_inode(struct inode *dir, struct dentry *dentry, goto fail_gunlock2;glock_set_object (ip->i_iopen_gh.gh_gl, ip); -gfs2_glock_put (io_gl);gfs2_set_iop (inode);insert_inode_hash (inode);
@@ -765,6 +764,8 @@ static int gfs2_create_inode(struct inode *dir, struct dentry *dentry,mark_inode_dirty (inode);d_instantiate (dentry, inode); +/* After instantiate, errors should result in evict which will destroy +* both inode and iopen glocks properly. */ if (file) { file->f_mode |=FMODE_CREATED ; error =finish_open (file, dentry, gfs2_open_common);
@@ -772,15 +773,15 @@ static int gfs2_create_inode(struct inode *dir, struct dentry *dentry,gfs2_glock_dq_uninit (ghs);gfs2_glock_dq_uninit (ghs +1 );clear_bit (GLF_INODE_CREATING , &io_gl->gl_flags); +gfs2_glock_put (io_gl);return error; fail_gunlock3:glock_clear_object (io_gl, ip);gfs2_glock_dq_uninit (&ip->i_iopen_gh); -gfs2_glock_put (io_gl); fail_gunlock2: -if (io_gl) -clear_bit (GLF_INODE_CREATING , &io_gl->gl_flags); +clear_bit (GLF_INODE_CREATING , &io_gl->gl_flags); +gfs2_glock_put (io_gl); fail_free_inode:if (ip->i_gl) {glock_clear_object (ip->i_gl, ip);
Andreas Gruenbacher
maintainerPATCH v3 1/1 gfs2: clean up iopen glock mess in gfs2_create_inode
Bob,
Quoted 8 lines · Bob Peterson
Before this patch,
gfs2_create_inodehad a use-after-free for the iopen glock in some error paths because it did this:gfs2_glock_put(io_gl); fail_gunlock2: if (io_gl) clear_bit(GLF_INODE_CREATING, &io_gl->gl_flags);
Quoted 5 lines
In some cases, the
io_glwas used for create and only had one reference, so the glock might be freed before theclear_bit(). This patch tries to straighten it out by only jumping to the error paths where iopen is properly set, and moving thegfs2_glock_putafter theclear_bit.
this patch fixes a glock refcounting bug and slightly cleans up gfs2_create_inode; it's an obvious improvement. This has passed my testing and I'd like to see this fixed in the upcoming merge window, so I'm adding it to for-next.
Thanks,
Andreas