[PATCH 1/2] erofs-utils: mount: gracefully exit when `erofsmount_nbd()` encounts an error

[zhaoyifan (H)]

On 2025/12/16 15:18, Gao Xiang wrote:
> Hi Yifan,
>
> On 2025/12/16 16:05, Yifan Zhao wrote:
>> If the main process of `erofsmount_nbd()` encounters an error after the
>> nbd device has been successfully set up, it fails to disconnect it
>> before exiting, resulting in the subprocess not being cleaned up and
>> blocked on `ioctl(nbdfd, NBD_DO_IT, 0)`.
>
> Do you have a simple test case (IOWs, how do you test this?)
> And is it possible to move the test case to erofs-utils tests?
>
How to reproduce the issue:


     mount.erofs -t erofs.nbd oci.layer=0 <some_non_erofs_image> /mnt/erofs


After the command fails, a leftover mount.erofs process (forked in 
erofsmount_nbd()) remains uncleaned.

In fact, I don't have a formal test case—I encountered this while trying 
out mount.erofs.
Since this is an end-to-end scenario rather than unit test, would you 
recommend adding a regression test for it (and other discovered 
mount.erofs issues) in our GitHub CI? (I'm happy to implement it.)
>>
>> This patch resolves the issue by invoking `erofs_nbd_disconnect()`
>> before exiting on error.
>
> See below.
>
>>
>> Signed-off-by: Yifan Zhao <zhaoyifan28 at huawei.com>
>> ---
>>   lib/liberofs_nbd.h | 2 +-
>>   mount/main.c       | 8 ++++++++
>>   2 files changed, 9 insertions(+), 1 deletion(-)
>>
>> diff --git a/lib/liberofs_nbd.h b/lib/liberofs_nbd.h
>> index 260605a..93daa24 100644
>> --- a/lib/liberofs_nbd.h
>> +++ b/lib/liberofs_nbd.h
>> @@ -28,7 +28,7 @@ struct erofs_nbd_request {
>>           char   handle[8];    /* older spelling of cookie */
>>       };
>>       u64 from;
>> -        u32 len;
>> +    u32 len;
>>   } __packed;
>>     /* 30-day timeout for NBD recovery */
>> diff --git a/mount/main.c b/mount/main.c
>> index 758e8f8..a093167 100644
>> --- a/mount/main.c
>> +++ b/mount/main.c
>> @@ -1206,6 +1206,14 @@ static int erofsmount_nbd(struct 
>> erofs_nbd_source *source,
>>                   free(id);
>>           }
>>       }
>> +
>> +    if (err < 0) {
>> +        nbdfd = open(nbdpath, O_RDWR);
>
> I'm not sure if it's a best-practice (is it possible
> nbdpath can be reused?)
>
> Could we just kill the subprocess instead?
>
> Also ioctl is discouraged and netlink is preferred now.
>
I will try to give a graceful solution later.

Thanks,

Yifan Zhao

> Thanks,
> Gao Xiang
>
>> +        if (nbdfd > 0) {
>> +            erofs_nbd_disconnect(nbdfd);
>> +            close(nbdfd);
>> +        }
>> +    }
>>       return err;
>>   }
>