"Yan, Zheng" <ukernel@xxxxxxxxx> writes:
> On Tue, Mar 12, 2019 at 10:22 PM Luis Henriques <lhenriques@xxxxxxxx> wrote:
...
>> +static struct inode *lookup_quotarealm_inode(struct ceph_mds_client *mdsc,
>> + struct super_block *sb,
>> + struct ceph_snap_realm *realm)
>> +{
>> + struct inode *in;
>> +
>> + in = ceph_lookup_inode(sb, realm->ino);
>> + if (IS_ERR(in)) {
>> + pr_warn("Can't lookup inode %llx (err: %ld)\n",
>> + realm->ino, PTR_ERR(in));
>> + return in;
>> + }
>> +
>> + spin_lock(&mdsc->quotarealms_inodes_lock);
>> + list_add(&ceph_inode(in)->i_quotarealms_inode_item,
>> + &mdsc->quotarealms_inodes_list);
>> + spin_unlock(&mdsc->quotarealms_inodes_lock);
>> +
> Multiple threads can call this function for the same inode at the same
> time. need to handle this. Besides, client needs to record lookupino
> error. Otherwise, client may repeatedly send useless request.
Good point. So, the only way I see to fix this is to drop the
mdsc->quotarealms_inodes_list and instead use an ordered list/tree of
structs that would either point to the corresponding ceph inode or to
NULL if there was an error in the lookup:
struct ceph_realm_inode {
u64 ino;
struct ceph_inode_info *ci;
spinlock_t lock;
unsigned long timeout;
}
The 'timeout' field would be used to try to do the lookup again if the
error occurred long time ago.
The code would then create a new struct for the realm->ino (if one is
not found in the mdsc list), lock it and do the lookupino; if there's a
struct already on the list, it either means there's a lookupino in
progress or there was an error in the last lookup.
This sounds overly complicated so I may be missing the obvious simple
fix. Any ideas?
>> + spin_lock(&realm->inodes_with_caps_lock);
>> + realm->inode = in;
>
> reply of lookup_ino should alreadly set realm->inode
Yes, of course. This was silly.
Cheers,
--
Luis
