2008-06-06 20:11:50

by Stefan Richter

Subject: [PATCH] firewire: fill_bus_reset_event needs lock protection

Callers of fill_bus_reset_event() have to take card->lock. Otherwise
access to node data may oops if node removal is in progress.

A lockless alternative would be

- event->local_node_id = card->local_node->node_id;
+ tmp = fw_node_get(card->local_node);
+ event->local_node_id = tmp->node_id;
+ fw_node_put(tmp);

and ditto with the other node pointers which fill_bus_reset_event()
accesses. But I went the locked route because one of the two callers
already holds the lock. As a bonus, we don't need the memory barrier
anymore because device->generation and device->node_id are written in
a card->lock protected section.

Signed-off-by: Stefan Richter <[email protected]>
---
drivers/firewire/fw-cdev.c | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)

Index: linux/drivers/firewire/fw-cdev.c
===================================================================
--- linux.orig/drivers/firewire/fw-cdev.c
+++ linux/drivers/firewire/fw-cdev.c
@@ -205,6 +205,7 @@ fw_device_op_read(struct file *file,
return dequeue_event(client, buffer, count);
}

+/* caller must hold card->lock so that node pointers can be dereferenced here */
static void
fill_bus_reset_event(struct fw_cdev_event_bus_reset *event,
struct client *client)
@@ -214,7 +215,6 @@ fill_bus_reset_event(struct fw_cdev_even
event->closure = client->bus_reset_closure;
event->type = FW_CDEV_EVENT_BUS_RESET;
event->generation = client->device->generation;
- smp_rmb(); /* node_id must not be older than generation */
event->node_id = client->device->node_id;
event->local_node_id = card->local_node->node_id;
event->bm_node_id = 0; /* FIXME: We don't track the BM. */
@@ -274,6 +274,7 @@ static int ioctl_get_info(struct client
{
struct fw_cdev_get_info *get_info = buffer;
struct fw_cdev_event_bus_reset bus_reset;
+ struct fw_card *card = client->device->card;
unsigned long ret = 0;

client->version = get_info->version;
@@ -299,13 +300,17 @@ static int ioctl_get_info(struct client
client->bus_reset_closure = get_info->bus_reset_closure;
if (get_info->bus_reset != 0) {
void __user *uptr = u64_to_uptr(get_info->bus_reset);
+ unsigned long flags;

+ spin_lock_irqsave(&card->lock, flags);
fill_bus_reset_event(&bus_reset, client);
+ spin_unlock_irqrestore(&card->lock, flags);
+
if (copy_to_user(uptr, &bus_reset, sizeof(bus_reset)))
return -EFAULT;
}

- get_info->card = client->device->card->index;
+ get_info->card = card->index;

return 0;
}

--
Stefan Richter
-=====-==--- -==- --==-
http://arcgraph.de/sr/

2008-06-13 14:22:17

by Kristian Høgsberg

Subject: Re: [PATCH] firewire: fill_bus_reset_event needs lock protection

On Fri, 2008-06-06 at 22:11 +0200, Stefan Richter wrote:
> Callers of fill_bus_reset_event() have to take card->lock. Otherwise
> access to node data may oops if node removal is in progress.
>
> A lockless alternative would be
>
> - event->local_node_id = card->local_node->node_id;
> + tmp = fw_node_get(card->local_node);
> + event->local_node_id = tmp->node_id;
> + fw_node_put(tmp);
>
> and ditto with the other node pointers which fill_bus_reset_event()
> accesses. But I went the locked route because one of the two callers
> already holds the lock. As a bonus, we don't need the memory barrier
> anymore because device->generation and device->node_id are written in
> a card->lock protected section.

Looks good to me.

Signed-off-by: Kristian Høgsberg <[email protected]>