[v2,12/22] nbd/client: Improve error handling in nbd_negotiate_simple_meta_context()

Always allocate space for the reply returned by the server and
hoist the trace earlier, as it is more interesting to trace the
server's reply (even if it is unexpected) than parroting our
request only on success.  After all, skipping the allocation
for a wrong size was merely a micro-optimization that only
benefitted a broken server, rather than the common case of a
compliant server that meets our expectations.

Then turn the reply handling into a loop (even though we still
never iterate more than once), to make this code easier to use
when later patches do support multiple server replies.  This
changes the error message for a server with two replies (a
corner case we are unlikely to hit in practice) from:

Unexpected reply type 4 (meta context), expected 0 (ack)

to:

Server replied with more than one context

Signed-off-by: Eric Blake <eblake@redhat.com>

---
v2: split patch into easier-to-review pieces [Rich, Vladimir]
---
 nbd/client.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

On Sat, Dec 15, 2018 at 07:53:14AM -0600, Eric Blake wrote:
> Always allocate space for the reply returned by the server and
> hoist the trace earlier, as it is more interesting to trace the
> server's reply (even if it is unexpected) than parroting our
> request only on success.  After all, skipping the allocation
> for a wrong size was merely a micro-optimization that only
> benefitted a broken server, rather than the common case of a
> compliant server that meets our expectations.
> 
> Then turn the reply handling into a loop (even though we still
> never iterate more than once), to make this code easier to use
> when later patches do support multiple server replies.  This
> changes the error message for a server with two replies (a
> corner case we are unlikely to hit in practice) from:
> 
> Unexpected reply type 4 (meta context), expected 0 (ack)
> 
> to:
> 
> Server replied with more than one context
> 
> Signed-off-by: Eric Blake <eblake@redhat.com>
> 
> ---
> v2: split patch into easier-to-review pieces [Rich, Vladimir]
> ---
>  nbd/client.c | 16 ++++++++++++----
>  1 file changed, 12 insertions(+), 4 deletions(-)
> 
> diff --git a/nbd/client.c b/nbd/client.c
> index bcccd5f555e..b6a85fc3ef8 100644
> --- a/nbd/client.c
> +++ b/nbd/client.c
> @@ -684,10 +684,11 @@ static int nbd_negotiate_simple_meta_context(QIOChannel *ioc,
>          return ret;
>      }
> 
> -    if (reply.type == NBD_REP_META_CONTEXT) {
> +    while (reply.type == NBD_REP_META_CONTEXT) {

I'm not sure I understand why this change is safe.

As far as I can see reply.type is only updated in the loop by
nbd_receive_option_reply, and that reads from the server, and so the
server might keep sending NBD_REP_META_CONTEXT packets (instead of the
expected NBD_REP_ACK), so it could now loop forever against a
malicious server?  (This is not taking into account any later patches)

Rich.

>          char *name;
> 
> -        if (reply.length != sizeof(info->context_id) + context_len) {
> +        if (reply.length <= sizeof(info->context_id) ||
> +            reply.length > NBD_MAX_BUFFER_SIZE) {
>              error_setg(errp, "Failed to negotiate meta context '%s', server "
>                         "answered with unexpected length %" PRIu32, context,
>                         reply.length);
> @@ -708,6 +709,15 @@ static int nbd_negotiate_simple_meta_context(QIOChannel *ioc,
>              return -1;
>          }
>          name[reply.length] = '\0';
> +        trace_nbd_opt_meta_reply(name, info->context_id);
> +
> +        if (received) {
> +            error_setg(errp, "Server replied with more than one context");
> +            g_free(name);
> +            nbd_send_opt_abort(ioc);
> +            return -1;
> +        }
> +
>          if (strcmp(context, name)) {
>              error_setg(errp, "Failed to negotiate meta context '%s', server "
>                         "answered with different context '%s'", context,
> @@ -717,8 +727,6 @@ static int nbd_negotiate_simple_meta_context(QIOChannel *ioc,
>              return -1;
>          }
>          g_free(name);
> -
> -        trace_nbd_opt_meta_reply(context, info->context_id);
>          received = true;
> 
>          /* receive NBD_REP_ACK */
> -- 
> 2.17.2

On 12/15/18 9:19 AM, Richard W.M. Jones wrote:
> On Sat, Dec 15, 2018 at 07:53:14AM -0600, Eric Blake wrote:
>> Always allocate space for the reply returned by the server and
>> hoist the trace earlier, as it is more interesting to trace the
>> server's reply (even if it is unexpected) than parroting our
>> request only on success.  After all, skipping the allocation
>> for a wrong size was merely a micro-optimization that only
>> benefitted a broken server, rather than the common case of a
>> compliant server that meets our expectations.
>>
>> Then turn the reply handling into a loop (even though we still
>> never iterate more than once), to make this code easier to use
>> when later patches do support multiple server replies.  This
>> changes the error message for a server with two replies (a
>> corner case we are unlikely to hit in practice) from:
>>
>> Unexpected reply type 4 (meta context), expected 0 (ack)
>>
>> to:
>>
>> Server replied with more than one context
>>
>> Signed-off-by: Eric Blake <eblake@redhat.com>
>>
>> ---
>> v2: split patch into easier-to-review pieces [Rich, Vladimir]
>> ---
>>   nbd/client.c | 16 ++++++++++++----
>>   1 file changed, 12 insertions(+), 4 deletions(-)
>>
>> diff --git a/nbd/client.c b/nbd/client.c
>> index bcccd5f555e..b6a85fc3ef8 100644
>> --- a/nbd/client.c
>> +++ b/nbd/client.c
>> @@ -684,10 +684,11 @@ static int nbd_negotiate_simple_meta_context(QIOChannel *ioc,
>>           return ret;
>>       }
>>
>> -    if (reply.type == NBD_REP_META_CONTEXT) {
>> +    while (reply.type == NBD_REP_META_CONTEXT) {
> 
> I'm not sure I understand why this change is safe.
> 
> As far as I can see reply.type is only updated in the loop by
> nbd_receive_option_reply, and that reads from the server, and so the
> server might keep sending NBD_REP_META_CONTEXT packets (instead of the
> expected NBD_REP_ACK), so it could now loop forever against a
> malicious server?  (This is not taking into account any later patches)

The loop can execute at most twice:

> 
> Rich.
> 
>>           char *name;
>>
>> -        if (reply.length != sizeof(info->context_id) + context_len) {
>> +        if (reply.length <= sizeof(info->context_id) ||
>> +            reply.length > NBD_MAX_BUFFER_SIZE) {
>>               error_setg(errp, "Failed to negotiate meta context '%s', server "
>>                          "answered with unexpected length %" PRIu32, context,
>>                          reply.length);
>> @@ -708,6 +709,15 @@ static int nbd_negotiate_simple_meta_context(QIOChannel *ioc,
>>               return -1;
>>           }
>>           name[reply.length] = '\0';
>> +        trace_nbd_opt_meta_reply(name, info->context_id);
>> +
>> +        if (received) {
>> +            error_setg(errp, "Server replied with more than one context");
>> +            g_free(name);
>> +            nbd_send_opt_abort(ioc);
>> +            return -1;
>> +        }

If the server replies with a second context, we break the loop by 
complaining.

The old code accepted at most one context, by complaining if the 
server's second reply was not ACK; the new code accepts at most one 
context, by complaining if the server sent more than one context, so the 
net effect of killing the connection for a misbehaving server response 
to SET is unchanged.

However, your point about a misbehaving server providing an infinite 
stream of responses to NBD_OPT_LIST or NBD_OPT_LIST_META_CONTEXT is an 
interesting question, and may be worth asking upstream to see if the NBD 
protocol should be tweaked to document any boundaries at how many 
listings a server might send before a client should worry about the 
server being malicious.  (Does not affect this patch, but pre-existing 
when we call nbd_receive_list() for servers that lack NBD_OPT_GO, and 
does impact the later patches in this series that call 
NBD_OPT_LIST_META_CONTEXT for 'qemu-nbd --list').

On 12/15/18 9:19 AM, Richard W.M. Jones wrote:
> On Sat, Dec 15, 2018 at 07:53:14AM -0600, Eric Blake wrote:
>> Always allocate space for the reply returned by the server and
>> hoist the trace earlier, as it is more interesting to trace the
>> server's reply (even if it is unexpected) than parroting our
>> request only on success.  After all, skipping the allocation
>> for a wrong size was merely a micro-optimization that only
>> benefitted a broken server, rather than the common case of a
>> compliant server that meets our expectations.
>>
>> Then turn the reply handling into a loop (even though we still
>> never iterate more than once), to make this code easier to use
>> when later patches do support multiple server replies.  This
>> changes the error message for a server with two replies (a
>> corner case we are unlikely to hit in practice) from:
>>
>> Unexpected reply type 4 (meta context), expected 0 (ack)
>>
>> to:
>>
>> Server replied with more than one context
>>
>> Signed-off-by: Eric Blake <eblake@redhat.com>
>>
>> ---
>> v2: split patch into easier-to-review pieces [Rich, Vladimir]
>> ---
>>   nbd/client.c | 16 ++++++++++++----
>>   1 file changed, 12 insertions(+), 4 deletions(-)
>>
>> diff --git a/nbd/client.c b/nbd/client.c
>> index bcccd5f555e..b6a85fc3ef8 100644
>> --- a/nbd/client.c
>> +++ b/nbd/client.c
>> @@ -684,10 +684,11 @@ static int nbd_negotiate_simple_meta_context(QIOChannel *ioc,
>>           return ret;
>>       }
>>
>> -    if (reply.type == NBD_REP_META_CONTEXT) {
>> +    while (reply.type == NBD_REP_META_CONTEXT) {
> 
> I'm not sure I understand why this change is safe.
> 
> As far as I can see reply.type is only updated in the loop by
> nbd_receive_option_reply, and that reads from the server, and so the
> server might keep sending NBD_REP_META_CONTEXT packets (instead of the
> expected NBD_REP_ACK), so it could now loop forever against a
> malicious server?  (This is not taking into account any later patches)

Hmm - now that I've already responded to why the conversion to a loop 
does not change this code, I'm now wondering if I even need this patch. 
In v1 of the series, both SET and LIST shared a common function, and 
since LIST needs the loop, converting SET to use a loop that exits early 
if it executes more than once was needed to make the two actions share a 
common entry point.  But since v2 uses different entry points (because 
it separated the common code into separate helper functions, leaving the 
SET entry point unchanged and adding a new LIST entry point), where only 
the LIST entry point actually has to loop, I might be able to just drop 
this patch entirely and still achieve the same effect.

17.12.2018 18:30, Eric Blake wrote:
> On 12/15/18 9:19 AM, Richard W.M. Jones wrote:
>> On Sat, Dec 15, 2018 at 07:53:14AM -0600, Eric Blake wrote:
>>> Always allocate space for the reply returned by the server and
>>> hoist the trace earlier, as it is more interesting to trace the
>>> server's reply (even if it is unexpected) than parroting our
>>> request only on success.  After all, skipping the allocation
>>> for a wrong size was merely a micro-optimization that only
>>> benefitted a broken server, rather than the common case of a
>>> compliant server that meets our expectations.
>>>
>>> Then turn the reply handling into a loop (even though we still
>>> never iterate more than once), to make this code easier to use
>>> when later patches do support multiple server replies.  This
>>> changes the error message for a server with two replies (a
>>> corner case we are unlikely to hit in practice) from:
>>>
>>> Unexpected reply type 4 (meta context), expected 0 (ack)
>>>
>>> to:
>>>
>>> Server replied with more than one context
>>>
>>> Signed-off-by: Eric Blake <eblake@redhat.com>
>>>
>>> ---
>>> v2: split patch into easier-to-review pieces [Rich, Vladimir]
>>> ---
>>>   nbd/client.c | 16 ++++++++++++----
>>>   1 file changed, 12 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/nbd/client.c b/nbd/client.c
>>> index bcccd5f555e..b6a85fc3ef8 100644
>>> --- a/nbd/client.c
>>> +++ b/nbd/client.c
>>> @@ -684,10 +684,11 @@ static int nbd_negotiate_simple_meta_context(QIOChannel *ioc,
>>>           return ret;
>>>       }
>>>
>>> -    if (reply.type == NBD_REP_META_CONTEXT) {
>>> +    while (reply.type == NBD_REP_META_CONTEXT) {
>>
>> I'm not sure I understand why this change is safe.
>>
>> As far as I can see reply.type is only updated in the loop by
>> nbd_receive_option_reply, and that reads from the server, and so the
>> server might keep sending NBD_REP_META_CONTEXT packets (instead of the
>> expected NBD_REP_ACK), so it could now loop forever against a
>> malicious server?  (This is not taking into account any later patches)
> 
> Hmm - now that I've already responded to why the conversion to a loop does not change this code, I'm now wondering if I even need this patch. In v1 of the series, both SET and LIST shared a common function, and since LIST needs the loop, converting SET to use a loop that exits early if it executes more than once was needed to make the two actions share a common entry point.  But since v2 uses different entry points (because it separated the common code into separate helper functions, leaving the SET entry point unchanged and adding a new LIST entry point), where only the LIST entry point actually has to loop, I might be able to just drop this patch entirely and still achieve the same effect.
> 

Are you going to resend series without this patch? In this case I'd prefer to continue review on already rebased patches. Or it doesn't make much difference?