Opinion please

Hello, Serge pointed out something to me on an internal list here and I wanted to poll for opinions. Which would you prefer? I think adding another parameter to audit_send_reply to specify GFP_ATOMIC/KERNEL would be simplest. But is it wise to do so? Here's what he had to say: Hey Tim, it looks like w_master is now adequately protected by the spinlock. Only problem is that you might sleep under rcu_read_lock().  In audit_list_watches(), you call audit_send_watch() from under rcu_read_lock.  audit_send_watch() calls audit_to_transport, which calls kmalloc(GFP_KERNEL).  Switch that to GFP_ATOMIC. audit_send_watch() also calls audit_send_reply(), which calls alloc_skb(GFP_KERNEL), and netlink_unicast.  You could patch audit_send_reply() to take a gfp_mask argument, and send gfp_mask=GFP_ATOMIC from audit_send_watch(). But I'm not sure whether netlink_unicast is a problem (sure looks like it - it puts current on a waitqueue for one thing) or what you should do about it.  You might just need to build a list of things to send under the rcu_read_lock(), then send each element once you rcu_read_unlock(). So something like: struct audit_skb_list {         struct hlist_node list;         void *memblk;         size_t size; }; audit_list_watches(int pid, int seq) {         struct hlist_head skb_list;         struct audit_skb_list listentry;         struct hlist_node *tmp, *pos;         INIT_HLIST_HEAD(&skb_list);         rcu_read_lock();         hlist_for_each_entry_rcu(...) {                 tmp = audit_build_watch(pid, seq, wentry->w_watch);                 if (tmp)                         hlist_add_head(tmp, &skb_list);         }         rcu_read_unlock();         hlist_for_each_entry_safe(listentry, pos, tmp, &skb_list, list) {                 audit_send_reply(pid, seq, AUDIT_WATCH_LIST, 0, 1,                         listentry->memblk, listentry->size);                 kfree(listentry);         }         audit_send_reply(pid, seq, AUDIT_WATCH_LIST, 1, 1, NULL, 0); } -- -tim