Messages in this thread | | | Date | Sat, 6 Jun 2015 01:36:22 +0200 | From | Oleg Nesterov <> | Subject | Re: [RFC][PATCH 0/5] Optimize percpu-rwsem |
| |
On 06/05, Al Viro wrote: > > On Fri, Jun 05, 2015 at 11:08:57PM +0200, Oleg Nesterov wrote: > > On 06/05, Al Viro wrote: > > > > > > FWIW, I hadn't really looked into stop_machine uses, but fs/locks.c one > > > is really not all that great - there we have a large trashcan of a list > > > (every file_lock on the system) and the only use of that list is /proc/locks > > > output generation. Sure, additions take this CPU's spinlock. And removals > > > take pretty much a random one - losing the timeslice and regaining it on > > > a different CPU is quite likely with the uses there. > > > > > > Why do we need a global lock there, anyway? Why not hold only one for > > > the chain currently being traversed? Sure, we'll need to get and drop > > > them in ->next() that way; so what? > > > > And note that fs/seq_file.c:seq_hlist_next_percpu() has no other users. > > > > And given that locks_delete_global_locks() takes the random lock anyway, > > perhaps the hashed lists/locking makes no sense, I dunno. > > It's not about making life easier for /proc/locks; it's about not screwing > those who add/remove file_lock...
I meant, seq_hlist_next_percpu() could be "static" in fs/locks.c.
> And no, that "random lock" isn't held > when modifying the (per-cpu) lists - it protects the list hanging off each > element of the global list, and /proc/locks scans those lists, so rather > than taking/dropping it in each ->show(), it's taken once in ->start()...
Sure, I understand. I meant that (perhaps) something like
struct { spinlock_t lock; struct list_head *head } file_lock_list[];
locks_insert_global_locks(fl) { int idx = fl->idx = hash(fl); spin_lock(&file_lock_list[idx].lock); hlist_add_head(...); spin_unlock(...); }
seq_hlist_next_percpu() could scan file_lock_list[] and unlock/lock ->lock when it changes the index.
But please forget, this is really minor. Just I think that file_lock_list is not actually "per-cpu", exactly because every locks_delete_global_locks() needs lg_local_lock_cpu(fl->fl_link_cpu) as you pointed out.
Oleg.
| |