[PATCHv6 0/5] net/tls: fixes for NVMe-over-TLS

[Sagi Grimberg]

>>>>> Hannes Reinecke <hare at suse.de> wrote:
>>>>>
>>>>>>> 'discover' and 'connect' works, but when I'm trying to transfer data
>>>>>>> (eg by doing a 'mkfs.xfs') the whole thing crashes horribly in
>>>>>>> sock_sendmsg() as it's trying to access invalid pages :-(
>>>>>
>>>>> Can you be more specific about the crash?
>>>>
>>>> Hannes,
>>>>
>>>> See:
>>>> [PATCH net] nvme-tcp: Fix comma-related oops
>>>
>>> Ah, right. That solves _that_ issue.
>>>
>>> But now I'm deadlocking on the tls_rx_reader_lock() (patched as to 
>>> your suggestion). Investigating.
>>
>> Are you sure it is a deadlock? or maybe you returned EAGAIN and nvme-tcp
>> does not interpret this as a transient status and simply returns from
>> io_work?
>>
>>> But it brought up yet another can of worms: what _exactly_ is the 
>>> return value of ->read_sock()?
>>>
>>> There are currently two conflicting use-cases:
>>> -> Ignore the return value, and assume errors etc are signalled
>>>     via 'desc.error'.
>>>     net/strparser/strparser.c
>>>     drivers/infiniband/sw/siw
>>>     drivers/scsi/iscsi_tcp.c
>>> -> use the return value of ->read_sock(), ignoring 'desc.error':
>>>     drivers/nvme/host/tcp.c
>>>     net/ipv4/tcp.c
>>> So which one is it?
>>> Needless to say, implementations following the second style do not
>>> set 'desc.error', causing any errors there to be ignored for callers
>>> from the first style...
>>
>> I don't think ignoring the return value of read_sock makes sense because
>> it can fail outside of the recv_actor failures.
>>
> Oh, but it's not read_actor which is expected to set desc.error.
> Have a look at 'strp_read_sock()':
> 
>          /* sk should be locked here, so okay to do read_sock */
>          sock->ops->read_sock(strp->sk, &desc, strp_recv);
> 
>          desc.error = strp->cb.read_sock_done(strp, desc.error);
> 
> it's the ->read_sock() callback which is expected to set desc.error.

Then it is completely up to the consumer how it wants to interpret the
error.

>> But to be on the safe side, perhaps you can both return an error and set
>> desc.error?
>>
> But why? We can easily make ->read_sock() a void function, then it's 
> obvious that you can't check the return value.

but it returns the consumed byte count, where would this info go?