Commit bec4596b authored by Eric Dumazet's avatar Eric Dumazet Committed by David S. Miller

drop_monitor: dont sleep in atomic context

drop_monitor calls several sleeping functions while in atomic context.

 BUG: sleeping function called from invalid context at mm/slub.c:943
 in_atomic(): 1, irqs_disabled(): 0, pid: 2103, name: kworker/0:2
 Pid: 2103, comm: kworker/0:2 Not tainted 3.5.0-rc1+ #55
 Call Trace:
  [<ffffffff810697ca>] __might_sleep+0xca/0xf0
  [<ffffffff811345a3>] kmem_cache_alloc_node+0x1b3/0x1c0
  [<ffffffff8105578c>] ? queue_delayed_work_on+0x11c/0x130
  [<ffffffff815343fb>] __alloc_skb+0x4b/0x230
  [<ffffffffa00b0360>] ? reset_per_cpu_data+0x160/0x160 [drop_monitor]
  [<ffffffffa00b022f>] reset_per_cpu_data+0x2f/0x160 [drop_monitor]
  [<ffffffffa00b03ab>] send_dm_alert+0x4b/0xb0 [drop_monitor]
  [<ffffffff810568e0>] process_one_work+0x130/0x4c0
  [<ffffffff81058249>] worker_thread+0x159/0x360
  [<ffffffff810580f0>] ? manage_workers.isra.27+0x240/0x240
  [<ffffffff8105d403>] kthread+0x93/0xa0
  [<ffffffff816be6d4>] kernel_thread_helper+0x4/0x10
  [<ffffffff8105d370>] ? kthread_freezable_should_stop+0x80/0x80
  [<ffffffff816be6d0>] ? gs_change+0xb/0xb

Rework the logic to call the sleeping functions in right context.

Use standard timer/workqueue api to let system chose any cpu to perform
the allocation and netlink send.

Also avoid a loop if reset_per_cpu_data() cannot allocate memory :
use mod_timer() to wait 1/10 second before next try.
Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Cc: Neil Horman <nhorman@tuxdriver.com>
Reviewed-by: default avatarNeil Horman <nhorman@tuxdriver.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent f8f5701b
...@@ -36,9 +36,6 @@ ...@@ -36,9 +36,6 @@
#define TRACE_ON 1 #define TRACE_ON 1
#define TRACE_OFF 0 #define TRACE_OFF 0
static void send_dm_alert(struct work_struct *unused);
/* /*
* Globals, our netlink socket pointer * Globals, our netlink socket pointer
* and the work handle that will send up * and the work handle that will send up
...@@ -48,11 +45,10 @@ static int trace_state = TRACE_OFF; ...@@ -48,11 +45,10 @@ static int trace_state = TRACE_OFF;
static DEFINE_MUTEX(trace_state_mutex); static DEFINE_MUTEX(trace_state_mutex);
struct per_cpu_dm_data { struct per_cpu_dm_data {
struct work_struct dm_alert_work; spinlock_t lock;
struct sk_buff __rcu *skb; struct sk_buff *skb;
atomic_t dm_hit_count; struct work_struct dm_alert_work;
struct timer_list send_timer; struct timer_list send_timer;
int cpu;
}; };
struct dm_hw_stat_delta { struct dm_hw_stat_delta {
...@@ -78,13 +74,13 @@ static int dm_delay = 1; ...@@ -78,13 +74,13 @@ static int dm_delay = 1;
static unsigned long dm_hw_check_delta = 2*HZ; static unsigned long dm_hw_check_delta = 2*HZ;
static LIST_HEAD(hw_stats_list); static LIST_HEAD(hw_stats_list);
static void reset_per_cpu_data(struct per_cpu_dm_data *data) static struct sk_buff *reset_per_cpu_data(struct per_cpu_dm_data *data)
{ {
size_t al; size_t al;
struct net_dm_alert_msg *msg; struct net_dm_alert_msg *msg;
struct nlattr *nla; struct nlattr *nla;
struct sk_buff *skb; struct sk_buff *skb;
struct sk_buff *oskb = rcu_dereference_protected(data->skb, 1); unsigned long flags;
al = sizeof(struct net_dm_alert_msg); al = sizeof(struct net_dm_alert_msg);
al += dm_hit_limit * sizeof(struct net_dm_drop_point); al += dm_hit_limit * sizeof(struct net_dm_drop_point);
...@@ -99,65 +95,40 @@ static void reset_per_cpu_data(struct per_cpu_dm_data *data) ...@@ -99,65 +95,40 @@ static void reset_per_cpu_data(struct per_cpu_dm_data *data)
sizeof(struct net_dm_alert_msg)); sizeof(struct net_dm_alert_msg));
msg = nla_data(nla); msg = nla_data(nla);
memset(msg, 0, al); memset(msg, 0, al);
} else } else {
schedule_work_on(data->cpu, &data->dm_alert_work); mod_timer(&data->send_timer, jiffies + HZ / 10);
/*
* Don't need to lock this, since we are guaranteed to only
* run this on a single cpu at a time.
* Note also that we only update data->skb if the old and new skb
* pointers don't match. This ensures that we don't continually call
* synchornize_rcu if we repeatedly fail to alloc a new netlink message.
*/
if (skb != oskb) {
rcu_assign_pointer(data->skb, skb);
synchronize_rcu();
atomic_set(&data->dm_hit_count, dm_hit_limit);
} }
spin_lock_irqsave(&data->lock, flags);
swap(data->skb, skb);
spin_unlock_irqrestore(&data->lock, flags);
return skb;
} }
static void send_dm_alert(struct work_struct *unused) static void send_dm_alert(struct work_struct *work)
{ {
struct sk_buff *skb; struct sk_buff *skb;
struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data); struct per_cpu_dm_data *data;
WARN_ON_ONCE(data->cpu != smp_processor_id()); data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
/* skb = reset_per_cpu_data(data);
* Grab the skb we're about to send
*/
skb = rcu_dereference_protected(data->skb, 1);
/*
* Replace it with a new one
*/
reset_per_cpu_data(data);
/*
* Ship it!
*/
if (skb) if (skb)
genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL); genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
put_cpu_var(dm_cpu_data);
} }
/* /*
* This is the timer function to delay the sending of an alert * This is the timer function to delay the sending of an alert
* in the event that more drops will arrive during the * in the event that more drops will arrive during the
* hysteresis period. Note that it operates under the timer interrupt * hysteresis period.
* so we don't need to disable preemption here
*/ */
static void sched_send_work(unsigned long unused) static void sched_send_work(unsigned long _data)
{ {
struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data); struct per_cpu_dm_data *data = (struct per_cpu_dm_data *)_data;
schedule_work_on(smp_processor_id(), &data->dm_alert_work);
put_cpu_var(dm_cpu_data); schedule_work(&data->dm_alert_work);
} }
static void trace_drop_common(struct sk_buff *skb, void *location) static void trace_drop_common(struct sk_buff *skb, void *location)
...@@ -167,33 +138,28 @@ static void trace_drop_common(struct sk_buff *skb, void *location) ...@@ -167,33 +138,28 @@ static void trace_drop_common(struct sk_buff *skb, void *location)
struct nlattr *nla; struct nlattr *nla;
int i; int i;
struct sk_buff *dskb; struct sk_buff *dskb;
struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data); struct per_cpu_dm_data *data;
unsigned long flags;
rcu_read_lock(); local_irq_save(flags);
dskb = rcu_dereference(data->skb); data = &__get_cpu_var(dm_cpu_data);
spin_lock(&data->lock);
dskb = data->skb;
if (!dskb) if (!dskb)
goto out; goto out;
if (!atomic_add_unless(&data->dm_hit_count, -1, 0)) {
/*
* we're already at zero, discard this hit
*/
goto out;
}
nlh = (struct nlmsghdr *)dskb->data; nlh = (struct nlmsghdr *)dskb->data;
nla = genlmsg_data(nlmsg_data(nlh)); nla = genlmsg_data(nlmsg_data(nlh));
msg = nla_data(nla); msg = nla_data(nla);
for (i = 0; i < msg->entries; i++) { for (i = 0; i < msg->entries; i++) {
if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) { if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
msg->points[i].count++; msg->points[i].count++;
atomic_inc(&data->dm_hit_count);
goto out; goto out;
} }
} }
if (msg->entries == dm_hit_limit)
goto out;
/* /*
* We need to create a new entry * We need to create a new entry
*/ */
...@@ -205,13 +171,11 @@ static void trace_drop_common(struct sk_buff *skb, void *location) ...@@ -205,13 +171,11 @@ static void trace_drop_common(struct sk_buff *skb, void *location)
if (!timer_pending(&data->send_timer)) { if (!timer_pending(&data->send_timer)) {
data->send_timer.expires = jiffies + dm_delay * HZ; data->send_timer.expires = jiffies + dm_delay * HZ;
add_timer_on(&data->send_timer, smp_processor_id()); add_timer(&data->send_timer);
} }
out: out:
rcu_read_unlock(); spin_unlock_irqrestore(&data->lock, flags);
put_cpu_var(dm_cpu_data);
return;
} }
static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location) static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
...@@ -418,11 +382,11 @@ static int __init init_net_drop_monitor(void) ...@@ -418,11 +382,11 @@ static int __init init_net_drop_monitor(void)
for_each_possible_cpu(cpu) { for_each_possible_cpu(cpu) {
data = &per_cpu(dm_cpu_data, cpu); data = &per_cpu(dm_cpu_data, cpu);
data->cpu = cpu;
INIT_WORK(&data->dm_alert_work, send_dm_alert); INIT_WORK(&data->dm_alert_work, send_dm_alert);
init_timer(&data->send_timer); init_timer(&data->send_timer);
data->send_timer.data = cpu; data->send_timer.data = (unsigned long)data;
data->send_timer.function = sched_send_work; data->send_timer.function = sched_send_work;
spin_lock_init(&data->lock);
reset_per_cpu_data(data); reset_per_cpu_data(data);
} }
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment