lkml.org 
[lkml]   [2017]   [Dec]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] kfree_rcu() should use the new kfree_bulk() interface for freeing rcu structures
On Tue, Dec 19, 2017 at 09:41:58PM +0100, Jesper Dangaard Brouer wrote:
>
> On Tue, 19 Dec 2017 09:52:27 -0800 rao.shoaib@oracle.com wrote:
>
> > +/* Main RCU function that is called to free RCU structures */
> > +static void
> > +__rcu_bulk_free(struct rcu_head *head, rcu_callback_t func, int cpu, bool lazy)
> > +{
> > + unsigned long offset;
> > + void *ptr;
> > + struct rcu_bulk_free *rbf;
> > + struct rcu_bulk_free_container *rbfc = NULL;
> > +
> > + rbf = this_cpu_ptr(&cpu_rbf);
> > +
> > + if (unlikely(!rbf->rbf_init)) {
> > + spin_lock_init(&rbf->rbf_lock);
> > + rbf->rbf_cpu = smp_processor_id();
> > + rbf->rbf_init = true;
> > + }
> > +
> > + /* hold lock to protect against other cpu's */
> > + spin_lock_bh(&rbf->rbf_lock);
>
> I'm not sure this will be faster. Having to take a cross CPU lock here
> (+ BH-disable) could cause scaling issues. Hopefully this lock will
> not be used intensively by other CPUs, right?
>
>
> The current cost of __call_rcu() is a local_irq_save/restore (which is
> quite expensive, but doesn't cause cross CPU chatter).
>
> Later in __rcu_process_callbacks() we have a local_irq_save/restore for
> the entire list, plus a per object cost doing local_bh_disable/enable.
> And for each object we call __rcu_reclaim(), which in some cases
> directly call kfree().

Isn't this lock in a per-CPU object? It -might- go cross-CPU in response
to CPU-hotplug operations, but that should be rare.

Thanx, Paul

> If I had to implement this: I would choose to do the optimization in
> __rcu_process_callbacks() create small on-call-stack ptr-array for
> kfree_bulk(). I would only optimize the case that call kfree()
> directly. In the while(list) loop I would defer calling
> __rcu_reclaim() for __is_kfree_rcu_offset(head->func), and instead add
> them to the ptr-array (and flush if the array is full in loop, and
> kfree_bulk flush after loop).
>
> The real advantage of kfree_bulk() comes from amortizing the per kfree
> (behind-the-scenes) sync cost. There is an additional benefit, because
> objects comes from RCU and will hit a slower path in SLUB. The SLUB
> allocator is very fast for objects that gets recycled quickly (short
> lifetime), non-locked (cpu-local) double-cmpxchg. But slower for
> longer-lived/more-outstanding objects, as this hits a slower code-path,
> fully locked (cross-cpu) double-cmpxchg.
>
> > +
> > + rbfc = rbf->rbf_container;
> > +
> > + if (rbfc == NULL) {
> > + if (rbf->rbf_cached_container == NULL) {
> > + rbf->rbf_container =
> > + kmalloc(sizeof(struct rcu_bulk_free_container),
> > + GFP_ATOMIC);
> > + rbf->rbf_container->rbfc_rbf = rbf;
> > + } else {
> > + rbf->rbf_container = rbf->rbf_cached_container;
> > + rbf->rbf_container->rbfc_rbf = rbf;
> > + cmpxchg(&rbf->rbf_cached_container,
> > + rbf->rbf_cached_container, NULL);
> > + }
> > +
> > + if (unlikely(rbf->rbf_container == NULL)) {
> > +
> > + /* Memory allocation failed maintain a list */
> > +
> > + head->func = (void *)func;
> > + head->next = rbf->rbf_list_head;
> > + rbf->rbf_list_head = head;
> > + rbf->rbf_list_size++;
> > + if (rbf->rbf_list_size == RCU_MAX_ACCUMULATE_SIZE)
> > + __rcu_bulk_schedule_list(rbf);
> > +
> > + goto done;
> > + }
> > +
> > + rbfc = rbf->rbf_container;
> > + rbfc->rbfc_entries = 0;
> > +
> > + if (rbf->rbf_list_head != NULL)
> > + __rcu_bulk_schedule_list(rbf);
> > + }
> > +
> > + offset = (unsigned long)func;
> > + ptr = (void *)head - offset;
> > +
> > + rbfc->rbfc_data[rbfc->rbfc_entries++] = ptr;
> > + if (rbfc->rbfc_entries == RCU_MAX_ACCUMULATE_SIZE) {
> > +
> > + WRITE_ONCE(rbf->rbf_container, NULL);
> > + spin_unlock_bh(&rbf->rbf_lock);
> > + call_rcu(&rbfc->rbfc_rcu, __rcu_bulk_free_impl);
> > + return;
> > + }
> > +
> > +done:
> > + if (!rbf->rbf_monitor) {
> > +
> > + call_rcu(&rbf->rbf_rcu, __rcu_bulk_free_monitor);
> > + rbf->rbf_monitor = true;
> > + }
> > +
> > + spin_unlock_bh(&rbf->rbf_lock);
> > +}
>
>
> --
> Best regards,
> Jesper Dangaard Brouer
> MSc.CS, Principal Kernel Engineer at Red Hat
> LinkedIn: http://www.linkedin.com/in/brouer
>

\
 
 \ /
  Last update: 2017-12-19 21:56    [W:0.896 / U:3.880 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site