__call_rcu如何防止多核竞争

littlenewer

__call_rcu只用了关中断来保护rcp，但是在SMP场景下有可能多个CPU同时把head链到rcp，而rcp又是全局变量rcu_sched_ctrlblk，这样就有可能产生冲突。系统是怎么保证不出问题的呢？

static void __call_rcu(struct rcu_head *head,
                       void (*func)(struct rcu_head *rcu),
                       struct rcu_ctrlblk *rcp)
{
        unsigned long flags;

        debug_rcu_head_queue(head);
        head->func = func;
        head->next = NULL;

        local_irq_save(flags);
        *rcp->curtail = head;
        rcp->curtail = &head->next;
        RCU_TRACE(rcp->qlen++);
        local_irq_restore(flags);
}

Tinnal

回复 1# littlenewer


  你看的代码是rcutiny.c吧。那个是单核的版本，多核的版本应该是rcutree.c


static void
__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
           struct rcu_state *rsp)
{
        unsigned long flags;
        struct rcu_data *rdp;

        head->func = func;
        head->next = NULL;

        smp_mb(); /* Ensure RCU update seen before callback registry. */

        /*
         * Opportunistically note grace-period endings and beginnings.
         * Note that we might see a beginning right after we see an
         * end, but never vice versa, since this CPU has to pass through
         * a quiescent state betweentimes.
         */
        local_irq_save(flags);
        rdp = rsp->rda[smp_processor_id()];
        rcu_process_gp_end(rsp, rdp);
        check_for_new_grace_period(rsp, rdp);

        /* Add the callback to our list. */
        *rdp->nxttail[RCU_NEXT_TAIL] = head;
        rdp->nxttail[RCU_NEXT_TAIL] = &head->next;

        /* Start a new grace period if one not already started. */
        if (!rcu_gp_in_progress(rsp)) {
                unsigned long nestflag;
                struct rcu_node *rnp_root = rcu_get_root(rsp);

                raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
                rcu_start_gp(rsp, nestflag);  /* releases rnp_root->lock. */
        }

        /*
         * Force the grace period if too many callbacks or too long waiting.
         * Enforce hysteresis, and don't invoke force_quiescent_state()
         * if some other CPU has recently done so.  Also, don't bother
         * invoking force_quiescent_state() if the newly enqueued callback
         * is the only one waiting for a grace period to complete.
         */
        if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
                rdp->blimit = LONG_MAX;
                if (rsp->n_force_qs == rdp->n_force_qs_snap &&
                    *rdp->nxttail[RCU_DONE_TAIL] != head)
                        force_quiescent_state(rsp, 0);
                rdp->n_force_qs_snap = rsp->n_force_qs;
                rdp->qlen_last_fqs_check = rdp->qlen;
        } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
                force_quiescent_state(rsp, 1);
        local_irq_restore(flags);
}

Tinnal

偶，要声明一下，我的代码是2.6.34.10的。3.x内核文件改名了，但结论不变。