On 8/18/21 8:41 PM, Jeff Layton wrote:
On Wed, 2021-08-18 at 13:18 +0200, Ilya Dryomov wrote:
On Wed, Aug 18, 2021 at 3:25 AM <xiubli@xxxxxxxxxx> wrote:
From: Xiubo Li <xiubli@xxxxxxxxxx>
When force umounting, it will try to remove all the session caps.
If there has any capsnap is in the flushing list, the remove session
caps callback will try to release the capsnap->flush_cap memory to
"ceph_cap_flush_cachep" slab cache, while which is allocated from
kmalloc-256 slab cache.
At the same time switch to list_del_init() because just in case the
force umount has removed it from the lists and the
handle_cap_flushsnap_ack() comes then the seconds list_del_init()
won't crash the kernel.
URL: https://tracker.ceph.com/issues/52283
Signed-off-by: Xiubo Li <xiubli@xxxxxxxxxx>
---
V3:
- rebase to the upstream
fs/ceph/caps.c | 18 ++++++++++++++----
fs/ceph/mds_client.c | 7 ++++---
2 files changed, 18 insertions(+), 7 deletions(-)
diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
index 1b9ca437da92..e239f06babbc 100644
--- a/fs/ceph/caps.c
+++ b/fs/ceph/caps.c
@@ -1712,7 +1712,16 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
struct ceph_cap_flush *ceph_alloc_cap_flush(void)
{
- return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
+ struct ceph_cap_flush *cf;
+
+ cf = kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
+ /*
+ * caps == 0 always means for the capsnap
+ * caps > 0 means dirty caps being flushed
+ * caps == -1 means preallocated, not used yet
+ */
Hi Xiubo,
This comment should be in super.h, on struct ceph_cap_flush
definition.
But more importantly, are you sure that overloading cf->caps this way
is safe? For example, __kick_flushing_caps() tests for cf->caps != 0
and cf->caps == -1 would be interpreted as a cue to call __prep_cap().
Thanks,
Ilya
The cf->caps field should get set to a sane value when it goes onto the
i_cap_flush_list, and I don't see how we'd get into testing against the
->caps field before that point. That said, this mechanism does seem a
bit fragile and subject to later breakage, and the caps code is anything
but clear and easy to follow.
pahole says that there is a 3 byte hole just after the "wake" field in
ceph_cap_flush on x86_64, and that's probably true on other arches as
well. Rather than overloading the caps field with this info, you could
add a new bool there to indicate whether it's embedded or not.
Okay, this also sounds good to me.
I will do that and sent out the V4 later.
Thanks
+ cf->caps = -1;
+ return cf;
}
void ceph_free_cap_flush(struct ceph_cap_flush *cf)
@@ -1747,7 +1756,7 @@ static bool __detach_cap_flush_from_mdsc(struct ceph_mds_client *mdsc,
prev->wake = true;
wake = false;
}
- list_del(&cf->g_list);
+ list_del_init(&cf->g_list);
return wake;
}
@@ -1762,7 +1771,7 @@ static bool __detach_cap_flush_from_ci(struct ceph_inode_info *ci,
prev->wake = true;
wake = false;
}
- list_del(&cf->i_list);
+ list_del_init(&cf->i_list);
return wake;
}
@@ -3642,7 +3651,8 @@ static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
cf = list_first_entry(&to_remove,
struct ceph_cap_flush, i_list);
list_del(&cf->i_list);
- ceph_free_cap_flush(cf);
+ if (cf->caps)
+ ceph_free_cap_flush(cf);
}
if (wake_ci)
diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index 1e013fb09d73..a44adbd1841b 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -1636,7 +1636,7 @@ static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
spin_lock(&mdsc->cap_dirty_lock);
list_for_each_entry(cf, &to_remove, i_list)
- list_del(&cf->g_list);
+ list_del_init(&cf->g_list);
if (!list_empty(&ci->i_dirty_item)) {
pr_warn_ratelimited(
@@ -1688,8 +1688,9 @@ static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
struct ceph_cap_flush *cf;
cf = list_first_entry(&to_remove,
struct ceph_cap_flush, i_list);
- list_del(&cf->i_list);
- ceph_free_cap_flush(cf);
+ list_del_init(&cf->i_list);
+ if (cf->caps)
+ ceph_free_cap_flush(cf);
}
wake_up_all(&ci->i_cap_wq);
--
2.27.0
