Hi Eric, Happy New Year! Yeah, it's a ceph side bug and I have sent a patch to fix it [1].When unmounting and just before closing the sessions the cephfs server still could send cap message to kceph and it will hold the inodes. So the unmount will skip them.
IMO it still makes sense to improve the vfs code because I hit a crash after this happening but just one time.
[1] https://patchwork.kernel.org/project/ceph-devel/patch/20221221093031.132792-1-xiubli@xxxxxxxxxx/
Thanks - Xiubo On 30/12/2022 07:53, Eric Biggers wrote:
Hi Xiubo, On Fri, Dec 02, 2022 at 03:04:58PM +0800, Xiubo Li wrote:On 02/12/2022 12:19, Eric Biggers wrote:On Fri, Dec 02, 2022 at 09:49:49AM +0800, Xiubo Li wrote:On 02/12/2022 05:10, Eric Biggers wrote:On Thu, Dec 01, 2022 at 11:18:33AM -0800, Eric Biggers wrote:On Thu, Dec 01, 2022 at 02:58:00PM +0800, xiubli@xxxxxxxxxx wrote:From: Xiubo Li <xiubli@xxxxxxxxxx> When close a file it will be deferred to call the fput(), which will hold the inode's i_count. And when unmounting the mountpoint the evict_inodes() may skip evicting some inodes. If encrypt is enabled the kernel generate a warning when removing the encrypt keys when the skipped inodes still hold the keyring:This does not make sense. Unmounting is only possible once all the files on the filesystem have been closed.Specifically, __fput() puts the reference to the dentry (and thus the inode) *before* it puts the reference to the mount. And an unmount cannot be done while the mount still has references. So there should not be any issue here.Eric, When I unmounting I can see the following logs, which I added a debug log in the evcit_inodes(): diff --git a/fs/inode.c b/fs/inode.c index b608528efd3a..f6e69b778d9c 100644 --- a/fs/inode.c +++ b/fs/inode.c @@ -716,8 +716,11 @@ void evict_inodes(struct super_block *sb) again: spin_lock(&sb->s_inode_list_lock); list_for_each_entry_safe(inode, next, &sb->s_inodes, i_sb_list) { - if (atomic_read(&inode->i_count)) + if (atomic_read(&inode->i_count)) { + printk("evict_inodes inode %p, i_count = %d, was skipped!\n", + inode, atomic_read(&inode->i_count)); continue; + } spin_lock(&inode->i_lock); if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { The logs: <4>[ 95.977395] evict_inodes inode 00000000f90aab7b, i_count = 1, was skipped! Any reason could cause this ? Since the inode couldn't be evicted in time and then when removing the master keys it will print this warning.It is expected for evict_inodes() to see some inodes with nonzero refcount, but they should only be filesystem internal inodes. For example, with ext4 this happens with the journal inode. However, filesystem internal inodes cannot be encrypted, so they are irrelevant here. I'd guess that CephFS has a bug where it is leaking a reference to a user inode somewhere.I also added some debug logs to tracker all the inodes in ceph, and all the requests has been finished. I will debug it more to see whether it's leaking a reference here. Thanks Eric.Any progress on tracking this down? - Eric
|