When kzalloc fail, may cause trying to destroy the
workqueue from inside the workqueue.
If num_connections is m (2 < m), and NO.1 ~ NO.n
(1 < n < m) kzalloc are successful. The NO.(n + 1)
failed. Then, nbd_start_device will return ENOMEM
to nbd_start_device_ioctl, and nbd_start_device_ioctl
will return immediately without running flush_workqueue.
However, we still have n recv threads. If nbd_release
run first, recv threads may have to drop the last
config_refs and try to destroy the workqueue from
inside the workqueue.
To fix it, add a flush_workqueue in nbd_start_device.
Fixes: e9e006f5fcf2 ("nbd: fix max number of supported devs")
Signed-off-by: Sun Ke <[email protected]>
---
drivers/block/nbd.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)
diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c
index b4607dd96185..dd1f8c2c6169 100644
--- a/drivers/block/nbd.c
+++ b/drivers/block/nbd.c
@@ -1264,7 +1264,12 @@ static int nbd_start_device(struct nbd_device *nbd)
args = kzalloc(sizeof(*args), GFP_KERNEL);
if (!args) {
- sock_shutdown(nbd);
+ if (i == 0)
+ sock_shutdown(nbd);
+ else {
+ sock_shutdown(nbd);
+ flush_workqueue(nbd->recv_workq);
+ }
return -ENOMEM;
}
sk_set_memalloc(config->socks[i]->sock->sk);
--
2.17.2
On 1/21/20 7:48 AM, Sun Ke wrote:
> When kzalloc fail, may cause trying to destroy the
> workqueue from inside the workqueue.
>
> If num_connections is m (2 < m), and NO.1 ~ NO.n
> (1 < n < m) kzalloc are successful. The NO.(n + 1)
> failed. Then, nbd_start_device will return ENOMEM
> to nbd_start_device_ioctl, and nbd_start_device_ioctl
> will return immediately without running flush_workqueue.
> However, we still have n recv threads. If nbd_release
> run first, recv threads may have to drop the last
> config_refs and try to destroy the workqueue from
> inside the workqueue.
>
> To fix it, add a flush_workqueue in nbd_start_device.
>
> Fixes: e9e006f5fcf2 ("nbd: fix max number of supported devs")
> Signed-off-by: Sun Ke <[email protected]>
> ---
> drivers/block/nbd.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c
> index b4607dd96185..dd1f8c2c6169 100644
> --- a/drivers/block/nbd.c
> +++ b/drivers/block/nbd.c
> @@ -1264,7 +1264,12 @@ static int nbd_start_device(struct nbd_device *nbd)
>
> args = kzalloc(sizeof(*args), GFP_KERNEL);
> if (!args) {
> - sock_shutdown(nbd);
> + if (i == 0)
> + sock_shutdown(nbd);
> + else {
> + sock_shutdown(nbd);
> + flush_workqueue(nbd->recv_workq);
> + }
Just for readability sake why don't we just flush_workqueue() unconditionally,
and add a comment so we know why in the future. Thanks,
Josef
On 1/21/20 7:00 AM, Josef Bacik wrote:
> On 1/21/20 7:48 AM, Sun Ke wrote:
>> When kzalloc fail, may cause trying to destroy the
>> workqueue from inside the workqueue.
>>
>> If num_connections is m (2 < m), and NO.1 ~ NO.n
>> (1 < n < m) kzalloc are successful. The NO.(n + 1)
>> failed. Then, nbd_start_device will return ENOMEM
>> to nbd_start_device_ioctl, and nbd_start_device_ioctl
>> will return immediately without running flush_workqueue.
>> However, we still have n recv threads. If nbd_release
>> run first, recv threads may have to drop the last
>> config_refs and try to destroy the workqueue from
>> inside the workqueue.
>>
>> To fix it, add a flush_workqueue in nbd_start_device.
>>
>> Fixes: e9e006f5fcf2 ("nbd: fix max number of supported devs")
>> Signed-off-by: Sun Ke <[email protected]>
>> ---
>> drivers/block/nbd.c | 7 ++++++-
>> 1 file changed, 6 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c
>> index b4607dd96185..dd1f8c2c6169 100644
>> --- a/drivers/block/nbd.c
>> +++ b/drivers/block/nbd.c
>> @@ -1264,7 +1264,12 @@ static int nbd_start_device(struct nbd_device *nbd)
>>
>> args = kzalloc(sizeof(*args), GFP_KERNEL);
>> if (!args) {
>> - sock_shutdown(nbd);
>> + if (i == 0)
>> + sock_shutdown(nbd);
>> + else {
>> + sock_shutdown(nbd);
>> + flush_workqueue(nbd->recv_workq);
>> + }
>
> Just for readability sake why don't we just flush_workqueue()
> unconditionally, and add a comment so we know why in the future.
Or maybe just make it:
sock_shutdown(nbd);
if (i)
flush_workqueue(nbd->recv_workq);
which does the same thing, but is still readable. The current code with
the shutdown duplication is just a bit odd. Needs a comment either way.
--
Jens Axboe
在 2020/1/22 5:25, Jens Axboe 写道:
> On 1/21/20 7:00 AM, Josef Bacik wrote:
>> On 1/21/20 7:48 AM, Sun Ke wrote:
>>> When kzalloc fail, may cause trying to destroy the
>>> workqueue from inside the workqueue.
>>>
>>> If num_connections is m (2 < m), and NO.1 ~ NO.n
>>> (1 < n < m) kzalloc are successful. The NO.(n + 1)
>>> failed. Then, nbd_start_device will return ENOMEM
>>> to nbd_start_device_ioctl, and nbd_start_device_ioctl
>>> will return immediately without running flush_workqueue.
>>> However, we still have n recv threads. If nbd_release
>>> run first, recv threads may have to drop the last
>>> config_refs and try to destroy the workqueue from
>>> inside the workqueue.
>>>
>>> To fix it, add a flush_workqueue in nbd_start_device.
>>>
>>> Fixes: e9e006f5fcf2 ("nbd: fix max number of supported devs")
>>> Signed-off-by: Sun Ke <[email protected]>
>>> ---
>>> drivers/block/nbd.c | 7 ++++++-
>>> 1 file changed, 6 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c
>>> index b4607dd96185..dd1f8c2c6169 100644
>>> --- a/drivers/block/nbd.c
>>> +++ b/drivers/block/nbd.c
>>> @@ -1264,7 +1264,12 @@ static int nbd_start_device(struct nbd_device *nbd)
>>>
>>> args = kzalloc(sizeof(*args), GFP_KERNEL);
>>> if (!args) {
>>> - sock_shutdown(nbd);
>>> + if (i == 0)
>>> + sock_shutdown(nbd);
>>> + else {
>>> + sock_shutdown(nbd);
>>> + flush_workqueue(nbd->recv_workq);
>>> + }
>>
>> Just for readability sake why don't we just flush_workqueue()
>> unconditionally, and add a comment so we know why in the future.
>
> Or maybe just make it:
>
> sock_shutdown(nbd);
> if (i)
> flush_workqueue(nbd->recv_workq);
>
> which does the same thing, but is still readable. The current code with
> the shutdown duplication is just a bit odd. Needs a comment either way.
>
OK, I will improve it in my v2 patch.
Thanks,
Sun Ke