generic-ipi: remove kmalloc()
Peter Zijlstra [Wed, 25 Feb 2009 12:59:47 +0000 (13:59 +0100)]
Remove the use of kmalloc() from the smp_call_function_*()
calls.

Steven's generic-ipi patch (d7240b98: generic-ipi: use per cpu
data for single cpu ipi calls) started the discussion on the use
of kmalloc() in this code and fixed the
smp_call_function_single(.wait=0) fallback case.

In this patch we complete this by also providing means for the
_many() call, which fully removes the need for kmalloc() in this
code.

The problem with the _many() call is that other cpus might still
be observing our entry when we're done with it. It solved this
by dynamically allocating data elements and RCU-freeing it.

We solve it by using a single per-cpu entry which provides
static storage and solves one half of the problem (avoiding
referencing freed data).

The other half, ensuring the queue iteration it still possible,
is done by placing re-used entries at the head of the list. This
means that if someone was still iterating that entry when it got
moved, he will now re-visit the entries on the list he had
already seen, but avoids skipping over entries like would have
happened had we placed the new entry at the end.

Furthermore, visiting entries twice is not a problem, since we
remove our cpu from the entry's cpumask once its called.

Many thanks to Oleg for his suggestions and him poking holes in
my earlier attempts.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Jens Axboe <jens.axboe@oracle.com>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

kernel/smp.c

index 6ecf4b9..7a0ce25 100644 (file)
 #include <linux/rcupdate.h>
 #include <linux/rculist.h>
 #include <linux/smp.h>
+#include <linux/cpu.h>
 
 static DEFINE_PER_CPU(struct call_single_queue, call_single_queue);
-static LIST_HEAD(call_function_queue);
-__cacheline_aligned_in_smp DEFINE_SPINLOCK(call_function_lock);
+
+static struct {
+       struct list_head        queue;
+       spinlock_t              lock;
+} call_function __cacheline_aligned_in_smp = {
+       .queue = LIST_HEAD_INIT(call_function.queue),
+       .lock  = __SPIN_LOCK_UNLOCKED(call_function.lock),
+};
 
 enum {
        CSD_FLAG_WAIT           = 0x01,
-       CSD_FLAG_ALLOC          = 0x02,
-       CSD_FLAG_LOCK           = 0x04,
+       CSD_FLAG_LOCK           = 0x02,
 };
 
 struct call_function_data {
        struct call_single_data csd;
        spinlock_t lock;
        unsigned int refs;
-       struct rcu_head rcu_head;
-       unsigned long cpumask_bits[];
+       cpumask_var_t cpumask;
 };
 
 struct call_single_queue {
@@ -34,8 +39,45 @@ struct call_single_queue {
        spinlock_t lock;
 };
 
+static DEFINE_PER_CPU(struct call_function_data, cfd_data) = {
+       .lock = __SPIN_LOCK_UNLOCKED(cfd_data.lock),
+};
+
+static int
+hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
+{
+       long cpu = (long)hcpu;
+       struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
+
+       switch (action) {
+       case CPU_UP_PREPARE:
+       case CPU_UP_PREPARE_FROZEN:
+               if (!alloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
+                               cpu_to_node(cpu)))
+                       return NOTIFY_BAD;
+               break;
+
+#ifdef CONFIG_CPU_HOTPLUG
+       case CPU_UP_CANCELED:
+       case CPU_UP_CANCELED_FROZEN:
+
+       case CPU_DEAD:
+       case CPU_DEAD_FROZEN:
+               free_cpumask_var(cfd->cpumask);
+               break;
+#endif
+       };
+
+       return NOTIFY_OK;
+}
+
+static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
+       .notifier_call = hotplug_cfd,
+};
+
 static int __cpuinit init_call_single_data(void)
 {
+       void *cpu = (void *)(long)smp_processor_id();
        int i;
 
        for_each_possible_cpu(i) {
@@ -44,18 +86,69 @@ static int __cpuinit init_call_single_data(void)
                spin_lock_init(&q->lock);
                INIT_LIST_HEAD(&q->list);
        }
+
+       hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
+       register_cpu_notifier(&hotplug_cfd_notifier);
+
        return 0;
 }
 early_initcall(init_call_single_data);
 
-static void csd_flag_wait(struct call_single_data *data)
+/*
+ * csd_wait/csd_complete are used for synchronous ipi calls
+ */
+static void csd_wait_prepare(struct call_single_data *data)
 {
-       /* Wait for response */
-       do {
-               if (!(data->flags & CSD_FLAG_WAIT))
-                       break;
+       data->flags |= CSD_FLAG_WAIT;
+}
+
+static void csd_complete(struct call_single_data *data)
+{
+       if (data->flags & CSD_FLAG_WAIT) {
+               /*
+                * ensure we're all done before saying we are
+                */
+               smp_mb();
+               data->flags &= ~CSD_FLAG_WAIT;
+       }
+}
+
+static void csd_wait(struct call_single_data *data)
+{
+       while (data->flags & CSD_FLAG_WAIT)
                cpu_relax();
-       } while (1);
+}
+
+/*
+ * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
+ *
+ * For non-synchronous ipi calls the csd can still be in use by the previous
+ * function call. For multi-cpu calls its even more interesting as we'll have
+ * to ensure no other cpu is observing our csd.
+ */
+static void csd_lock(struct call_single_data *data)
+{
+       while (data->flags & CSD_FLAG_LOCK)
+               cpu_relax();
+       data->flags = CSD_FLAG_LOCK;
+
+       /*
+        * prevent CPU from reordering the above assignment to ->flags
+        * with any subsequent assignments to other fields of the
+        * specified call_single_data structure.
+        */
+
+       smp_mb();
+}
+
+static void csd_unlock(struct call_single_data *data)
+{
+       WARN_ON(!(data->flags & CSD_FLAG_LOCK));
+       /*
+        * ensure we're all done before releasing data
+        */
+       smp_mb();
+       data->flags &= ~CSD_FLAG_LOCK;
 }
 
 /*
@@ -89,16 +182,7 @@ static void generic_exec_single(int cpu, struct call_single_data *data)
                arch_send_call_function_single_ipi(cpu);
 
        if (wait)
-               csd_flag_wait(data);
-}
-
-static void rcu_free_call_data(struct rcu_head *head)
-{
-       struct call_function_data *data;
-
-       data = container_of(head, struct call_function_data, rcu_head);
-
-       kfree(data);
+               csd_wait(data);
 }
 
 /*
@@ -122,41 +206,35 @@ void generic_smp_call_function_interrupt(void)
         * It's ok to use list_for_each_rcu() here even though we may delete
         * 'pos', since list_del_rcu() doesn't clear ->next
         */
-       rcu_read_lock();
-       list_for_each_entry_rcu(data, &call_function_queue, csd.list) {
+       list_for_each_entry_rcu(data, &call_function.queue, csd.list) {
                int refs;
 
-               if (!cpumask_test_cpu(cpu, to_cpumask(data->cpumask_bits)))
+               spin_lock(&data->lock);
+               if (!cpumask_test_cpu(cpu, data->cpumask)) {
+                       spin_unlock(&data->lock);
                        continue;
+               }
+               cpumask_clear_cpu(cpu, data->cpumask);
+               spin_unlock(&data->lock);
 
                data->csd.func(data->csd.info);
 
                spin_lock(&data->lock);
-               cpumask_clear_cpu(cpu, to_cpumask(data->cpumask_bits));
                WARN_ON(data->refs == 0);
-               data->refs--;
-               refs = data->refs;
+               refs = --data->refs;
+               if (!refs) {
+                       spin_lock(&call_function.lock);
+                       list_del_rcu(&data->csd.list);
+                       spin_unlock(&call_function.lock);
+               }
                spin_unlock(&data->lock);
 
                if (refs)
                        continue;
 
-               spin_lock(&call_function_lock);
-               list_del_rcu(&data->csd.list);
-               spin_unlock(&call_function_lock);
-
-               if (data->csd.flags & CSD_FLAG_WAIT) {
-                       /*
-                        * serialize stores to data with the flag clear
-                        * and wakeup
-                        */
-                       smp_wmb();
-                       data->csd.flags &= ~CSD_FLAG_WAIT;
-               }
-               if (data->csd.flags & CSD_FLAG_ALLOC)
-                       call_rcu(&data->rcu_head, rcu_free_call_data);
+               csd_complete(&data->csd);
+               csd_unlock(&data->csd);
        }
-       rcu_read_unlock();
 
        put_cpu();
 }
@@ -192,14 +270,14 @@ void generic_smp_call_function_single_interrupt(void)
 
                data->func(data->info);
 
-               if (data_flags & CSD_FLAG_WAIT) {
-                       smp_wmb();
-                       data->flags &= ~CSD_FLAG_WAIT;
-               } else if (data_flags & CSD_FLAG_LOCK) {
-                       smp_wmb();
-                       data->flags &= ~CSD_FLAG_LOCK;
-               } else if (data_flags & CSD_FLAG_ALLOC)
-                       kfree(data);
+               if (data_flags & CSD_FLAG_WAIT)
+                       csd_complete(data);
+
+               /*
+                * Unlocked CSDs are valid through generic_exec_single()
+                */
+               if (data_flags & CSD_FLAG_LOCK)
+                       csd_unlock(data);
        }
 }
 
@@ -218,7 +296,9 @@ static DEFINE_PER_CPU(struct call_single_data, csd_data);
 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
                             int wait)
 {
-       struct call_single_data d;
+       struct call_single_data d = {
+               .flags = 0,
+       };
        unsigned long flags;
        /* prevent preemption and reschedule on another processor,
           as well as CPU removal */
@@ -239,13 +319,11 @@ int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
                        /*
                         * We are calling a function on a single CPU
                         * and we are not going to wait for it to finish.
-                        * We first try to allocate the data, but if we
-                        * fail, we fall back to use a per cpu data to pass
-                        * the information to that CPU. Since all callers
-                        * of this code will use the same data, we must
-                        * synchronize the callers to prevent a new caller
-                        * from corrupting the data before the callee
-                        * can access it.
+                        * We use a per cpu data to pass the information to
+                        * that CPU. Since all callers of this code will
+                        * use the same data, we must synchronize the
+                        * callers to prevent a new caller from corrupting
+                        * the data before the callee can access it.
                         *
                         * The CSD_FLAG_LOCK is used to let us know when
                         * the IPI handler is done with the data.
@@ -255,18 +333,11 @@ int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
                         * will make sure the callee is done with the
                         * data before a new caller will use it.
                         */
-                       data = kmalloc(sizeof(*data), GFP_ATOMIC);
-                       if (data)
-                               data->flags = CSD_FLAG_ALLOC;
-                       else {
-                               data = &per_cpu(csd_data, me);
-                               while (data->flags & CSD_FLAG_LOCK)
-                                       cpu_relax();
-                               data->flags = CSD_FLAG_LOCK;
-                       }
+                       data = &__get_cpu_var(csd_data);
+                       csd_lock(data);
                } else {
                        data = &d;
-                       data->flags = CSD_FLAG_WAIT;
+                       csd_wait_prepare(data);
                }
 
                data->func = func;
@@ -326,14 +397,14 @@ void smp_call_function_many(const struct cpumask *mask,
 {
        struct call_function_data *data;
        unsigned long flags;
-       int cpu, next_cpu;
+       int cpu, next_cpu, me = smp_processor_id();
 
        /* Can deadlock when called with interrupts disabled */
        WARN_ON(irqs_disabled());
 
        /* So, what's a CPU they want?  Ignoring this one. */
        cpu = cpumask_first_and(mask, cpu_online_mask);
-       if (cpu == smp_processor_id())
+       if (cpu == me)
                cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
        /* No online cpus?  We're done. */
        if (cpu >= nr_cpu_ids)
@@ -341,7 +412,7 @@ void smp_call_function_many(const struct cpumask *mask,
 
        /* Do we have another CPU which isn't us? */
        next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
-       if (next_cpu == smp_processor_id())
+       if (next_cpu == me)
                next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
 
        /* Fastpath: do that cpu by itself. */
@@ -350,31 +421,28 @@ void smp_call_function_many(const struct cpumask *mask,
                return;
        }
 
-       data = kmalloc(sizeof(*data) + cpumask_size(), GFP_ATOMIC);
-       if (unlikely(!data)) {
-               /* Slow path. */
-               for_each_online_cpu(cpu) {
-                       if (cpu == smp_processor_id())
-                               continue;
-                       if (cpumask_test_cpu(cpu, mask))
-                               smp_call_function_single(cpu, func, info, wait);
-               }
-               return;
-       }
+       data = &__get_cpu_var(cfd_data);
+       csd_lock(&data->csd);
 
-       spin_lock_init(&data->lock);
-       data->csd.flags = CSD_FLAG_ALLOC;
+       spin_lock_irqsave(&data->lock, flags);
        if (wait)
-               data->csd.flags |= CSD_FLAG_WAIT;
+               csd_wait_prepare(&data->csd);
+
        data->csd.func = func;
        data->csd.info = info;
-       cpumask_and(to_cpumask(data->cpumask_bits), mask, cpu_online_mask);
-       cpumask_clear_cpu(smp_processor_id(), to_cpumask(data->cpumask_bits));
-       data->refs = cpumask_weight(to_cpumask(data->cpumask_bits));
+       cpumask_and(data->cpumask, mask, cpu_online_mask);
+       cpumask_clear_cpu(me, data->cpumask);
+       data->refs = cpumask_weight(data->cpumask);
 
-       spin_lock_irqsave(&call_function_lock, flags);
-       list_add_tail_rcu(&data->csd.list, &call_function_queue);
-       spin_unlock_irqrestore(&call_function_lock, flags);
+       spin_lock(&call_function.lock);
+       /*
+        * Place entry at the _HEAD_ of the list, so that any cpu still
+        * observing the entry in generic_smp_call_function_interrupt() will
+        * not miss any other list entries.
+        */
+       list_add_rcu(&data->csd.list, &call_function.queue);
+       spin_unlock(&call_function.lock);
+       spin_unlock_irqrestore(&data->lock, flags);
 
        /*
         * Make the list addition visible before sending the ipi.
@@ -384,11 +452,11 @@ void smp_call_function_many(const struct cpumask *mask,
        smp_mb();
 
        /* Send a message to all CPUs in the map */
-       arch_send_call_function_ipi_mask(to_cpumask(data->cpumask_bits));
+       arch_send_call_function_ipi_mask(data->cpumask);
 
        /* optionally wait for the CPUs to complete */
        if (wait)
-               csd_flag_wait(&data->csd);
+               csd_wait(&data->csd);
 }
 EXPORT_SYMBOL(smp_call_function_many);
 
@@ -418,20 +486,20 @@ EXPORT_SYMBOL(smp_call_function);
 
 void ipi_call_lock(void)
 {
-       spin_lock(&call_function_lock);
+       spin_lock(&call_function.lock);
 }
 
 void ipi_call_unlock(void)
 {
-       spin_unlock(&call_function_lock);
+       spin_unlock(&call_function.lock);
 }
 
 void ipi_call_lock_irq(void)
 {
-       spin_lock_irq(&call_function_lock);
+       spin_lock_irq(&call_function.lock);
 }
 
 void ipi_call_unlock_irq(void)
 {
-       spin_unlock_irq(&call_function_lock);
+       spin_unlock_irq(&call_function.lock);
 }