Thanks for your detailed suggestions. 在 2020/1/18 1:32, Mike Christie 写道:
On 01/17/2020 05:50 AM, Sun Ke wrote:Connect and disconnect a nbd device repeatedly, will cause NULL pointer fault. It will appear by the steps: 1. Connect the nbd device and disconnect it, but now nbd device is not disconnected totally. 2. Connect the same nbd device again immediately, it will fail in nbd_start_device with a EBUSY return value. 3. Wait a second to make sure the last config_refs is reduced and run nbd_config_put to disconnect the nbd device totally. 4. Start another process to open the nbd_device, config_refs will increase and at the same time disconnect it.Just to make sure I understood this, for step 4 the process is doing: open(/dev/nbdX); ioctl(NBD_DISCONNECT, /dev/nbdX) or nbd_genl_disconnect(for /dev/nbdX) ?
do nbd_genl_disconnect(for /dev/nbdX); I tested it. Connect /dev/nbdX through ioctl interface by nbd-client -L -N export localhost /dev/nbdX and through netlink interface by nbd-client localhost XXXX /dev/nbdX, disconnect /dev/nbdX by nbd-client -d /dev/nbdX.Both call nbd_genl_disconnect(for /dev/nbdX) and both contain the same null pointer dereference.
There is no successful NBD_DO_IT / nbd_genl_connect between the open and disconnect calls at step #4, because it would normally be done at #2 and that failed. nbd_disconnect_and_put could then reference a null recv_workq. If we are also racing with a close() then that could free the device/config from under nbd_disconnect_and_put.
Yes, nbd_disconnect_and_put could then reference a null recv_workq.
To fix it, add a NBD_HAS_STARTED flag. Set it in nbd_start_device_ioctlI'm not sure if we need the new bit. We could just add a check for a non null task_recv in nbd_genl_disconnect like how nbd_start_device and nbd_genl_disconnect do.
I am also not very sure which is better. because in nbd_config_put, not only recv_workq is null, nbd->task_recv and nbd->config the same.so I doubt that if step 4 do something else will also reference a null pointer.
Yes, I realize it. Just add a check for a non null task_recv in nbd_genl_disconnect is better, right?The new bit might be more clear which is nice. If we got this route, should the new bit be a runtime_flag like other device state bits?
I thought the changes in nbd_genl_reconfigure is necessary althought my test do not call it. but now I think it is superfluous, nbd_genl_reconfigure checks for a non null task_recv.and nbd_genl_connect if nbd device is started successfully. Clear it in nbd_config_put. Test it in nbd_genl_disconnect and nbd_genl_reconfigure. Signed-off-by: Sun Ke <sunke32@huawei.com> --- drivers/block/nbd.c | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c index b4607dd96185..ddd364e208ab 100644 --- a/drivers/block/nbd.c +++ b/drivers/block/nbd.c @@ -83,6 +83,7 @@ struct link_dead_args {#define NBD_DESTROY_ON_DISCONNECT 0#define NBD_DISCONNECT_REQUESTED 1 +#define NBD_HAS_STARTED 2struct nbd_config {u32 flags; @@ -1215,6 +1216,7 @@ static void nbd_config_put(struct nbd_device *nbd) nbd->disk->queue->limits.discard_alignment = 0; blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX); blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue); + clear_bit(NBD_HAS_STARTED, &nbd->flags);mutex_unlock(&nbd->config_lock);nbd_put(nbd); @@ -1290,6 +1292,8 @@ static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *b ret = nbd_start_device(nbd); if (ret) return ret; + else + set_bit(NBD_HAS_STARTED, &nbd->flags);if (max_part)bdev->bd_invalidated = 1; @@ -1961,6 +1965,7 @@ static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info) mutex_unlock(&nbd->config_lock); if (!ret) { set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags); + set_bit(NBD_HAS_STARTED, &nbd->flags); refcount_inc(&nbd->config_refs); nbd_connect_reply(info, nbd->index); } @@ -2008,6 +2013,14 @@ static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info) index); return -EINVAL; } + + if (!test_bit(NBD_HAS_STARTED, &nbd->flags)) { + mutex_unlock(&nbd_index_mutex); + printk(KERN_ERR "nbd: device at index %d failed to start\n", + index); + return -EBUSY; + } + if (!refcount_inc_not_zero(&nbd->refs)) { mutex_unlock(&nbd_index_mutex); printk(KERN_ERR "nbd: device at index %d is going down\n", @@ -2049,6 +2062,14 @@ static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info) index); return -EINVAL; } + + if (!test_bit(NBD_HAS_STARTED, &nbd->flags)) { + mutex_unlock(&nbd_index_mutex); + printk(KERN_ERR "nbd: device at index %d failed to start\n", + index); + return -EBUSY; + } + if (!refcount_inc_not_zero(&nbd->refs)) { mutex_unlock(&nbd_index_mutex); printk(KERN_ERR "nbd: device at index %d is going down\n",
Thanks, Ke
.