summaryrefslogtreecommitdiffstats
path: root/fs/dcache.c
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2019-12-06 18:06:58 +0100
committerLinus Torvalds <torvalds@linux-foundation.org>2019-12-06 18:06:58 +0100
commit0aecba6173216931c436a03183f4759a4fd4c2f2 (patch)
tree7bfc21b84675a72b28de651df2201d397d9aebe4 /fs/dcache.c
parentMerge branch 'next.autofs' of git://git.kernel.org/pub/scm/linux/kernel/git/v... (diff)
parentfs/namei.c: fix missing barriers when checking positivity (diff)
downloadlinux-0aecba6173216931c436a03183f4759a4fd4c2f2.tar.xz
linux-0aecba6173216931c436a03183f4759a4fd4c2f2.zip
Merge branch 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
Pull vfs d_inode/d_flags memory ordering fixes from Al Viro: "Fallout from tree-wide audit for ->d_inode/->d_flags barriers use. Basically, the problem is that negative pinned dentries require careful treatment - unless ->d_lock is locked or parent is held at least shared, another thread can make them positive right under us. Most of the uses turned out to be safe - the main surprises as far as filesystems are concerned were - race in dget_parent() fastpath, that might end up with the caller observing the returned dentry _negative_, due to insufficient barriers. It is positive in memory, but we could end up seeing the wrong value of ->d_inode in CPU cache. Fixed. - manual checks that result of lookup_one_len_unlocked() is positive (and rejection of negatives). Again, insufficient barriers (we might end up with inconsistent observed values of ->d_inode and ->d_flags). Fixed by switching to a new primitive that does the checks itself and returns ERR_PTR(-ENOENT) instead of a negative dentry. That way we get rid of boilerplate converting negatives into ERR_PTR(-ENOENT) in the callers and have a single place to deal with the barrier-related mess - inside fs/namei.c rather than in every caller out there. The guts of pathname resolution *do* need to be careful - the race found by Ritesh is real, as well as several similar races. Fortunately, it turns out that we can take care of that with fairly local changes in there. The tree-wide audit had not been fun, and I hate the idea of repeating it. I think the right approach would be to annotate the places where we are _not_ guaranteed ->d_inode/->d_flags stability and have sparse catch regressions. But I'm still not sure what would be the least invasive way of doing that and it's clearly the next cycle fodder" * 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: fs/namei.c: fix missing barriers when checking positivity fix dget_parent() fastpath race new helper: lookup_positive_unlocked() fs/namei.c: pull positivity check into follow_managed()
Diffstat (limited to 'fs/dcache.c')
-rw-r--r--fs/dcache.c6
1 files changed, 4 insertions, 2 deletions
diff --git a/fs/dcache.c b/fs/dcache.c
index f7931b682a0d..a2749a700230 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -319,7 +319,7 @@ static inline void __d_set_inode_and_type(struct dentry *dentry,
flags = READ_ONCE(dentry->d_flags);
flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
flags |= type_flags;
- WRITE_ONCE(dentry->d_flags, flags);
+ smp_store_release(&dentry->d_flags, flags);
}
static inline void __d_clear_type_and_inode(struct dentry *dentry)
@@ -903,17 +903,19 @@ struct dentry *dget_parent(struct dentry *dentry)
{
int gotref;
struct dentry *ret;
+ unsigned seq;
/*
* Do optimistic parent lookup without any
* locking.
*/
rcu_read_lock();
+ seq = raw_seqcount_begin(&dentry->d_seq);
ret = READ_ONCE(dentry->d_parent);
gotref = lockref_get_not_zero(&ret->d_lockref);
rcu_read_unlock();
if (likely(gotref)) {
- if (likely(ret == READ_ONCE(dentry->d_parent)))
+ if (!read_seqcount_retry(&dentry->d_seq, seq))
return ret;
dput(ret);
}