On Mon, 2017-10-02 at 09:35 -0400, Rik van Riel wrote:
> On Wed, 2017-09-27 at 17:45 +0200, Oleg Nesterov wrote:
> > On 09/27, Gargi Sharma wrote:
> > >
> > > -#define find_next_offset(map, off)
> > >
> > > \
> > > - find_next_zero_bit((map)->page, BITS_PER_PAGE,
> > > off)
> > > -
> >
> > this should go into the previous patch, but this is minor...
> >
> > > @@ -208,12 +200,10 @@ struct pid *alloc_pid(struct pid_namespace
> > > *ns)
> > >
> > > upid = pid->numbers + ns->level;
> > > spin_lock_irq(&pidmap_lock);
> > > - if (!(ns->nr_hashed & PIDNS_HASH_ADDING))
> > > + if (!(ns->pid_allocated & PIDNS_ADDING))
> > > goto out_unlock;
> > > for ( ; upid >= pid->numbers; --upid) {
> > > - hlist_add_head_rcu(&upid->pid_chain,
> > > - &pid_hash[pid_hashfn(upid->nr,
> > > upid->ns)]);
> > > - upid->ns->nr_hashed++;
> > > + upid->ns->pid_allocated++;
> >
> > No, this is wrong.
> >
> > It is too late to check PIDNS_HASH_ADDING/PIDNS_ADDING and
> > increment
> > pid_allocated,
> > once we call idr_alloc_cyclic() this pid is already "hashed" in
> > that
> > it can be found
> > by find_pid_ns() with this patch applied.
> >
> > And of course, it is too late to do atomic_set(&pid->count, 1) and
> > initialize
> > pid->tasks[type] lists by the same reason.
>
> Hi Oleg,
>
> Gargi and I are looking at that code, and trying to figure out
> exactly what needs to be done to make all of this correct.
>
> We are thinking something along these lines:
.... and we looked at the code some more. Some of the way the
code is laid out currently looks like an artifact of history.
Lets try this again:
1) pid = kmem_cache_alloc(...)
2) get_pid_ns(ns);
atomic_set(&pid->count, 1);
for (type = 0; ...)
INIT_HLIST(...)
> 3) First, check if this is a new namespace (PIDNS_ADDING), and
> do the call to pid_ns_prepare_proc, before we even call idr_alloc.
> Maybe something like:
>
> if (unlikely(ns->nr_allocated == PIDNS_ADDING)) {
> if (pid_ns_prepare_proc(ns)) {
> disable_pid_allocations(ns);
> goto out_free_ns;
> }
>
> 4) With pid_ns_prepare_proc out of the way, we can put all the code
> from below where the call to pid_ns_prepare_proc is now (except
> error handing) into the main loop of pid allocation, so we can
> do all that stuff under the pidmap_lock:
>
> for (i = ns->level; i >= 0; i--) {
> ...
> idr_alloc_cyclic(...)
> ns->nr_allocated++;
> ...
> }
>
> Would that resolve your objection, or are we barking up the wrong
> tree?
Now the per-pid struct stuff is done once, before the loop, and
the loop deals only with allocated numbers inside each parent
namespace.
The error path would unwind stuff done earlier.
--
All Rights Reversed.
signature.asc
Description: This is a digitally signed message part
