Re: RFR(M): JDK-8061228 Allow JDWP socket connector to accept connections from certain ip addresses only

[serguei.spit...@oracle.com]

Hi Dmitry, Let's get to a consensus for the last item on our plate:   jdwpTransportError AllowPeers(const char* peers);                  .VS.   jdwpTransportError StartListeningWithAllow(const char* address, char** actual_address, char *allow); I still do not like the StartListening() function clone and its name. It looks like a wrong direction to me. This is a fragment of my bug report comment on this issue:   4. The suggested *allow* feature is too destructive for the transport API.      It makes the original function StartListening unusable and its entry slot a garbage.      It will become a big mess if we add more function clones like StartListening11.   . . .   6. The suggested API update makes the transport API VERSION_1_1      incompatible with the initial VERSION_1_0. One more alternate solution to suggest is to add new function:    jdwpTransportError SetTransportConfiguration(jdwpTransportConfiguration config); Where:   typedef struct {       const char* allowed_peers;   } jdwpTransportConfiguration; This approach allows to extend the jdwpTransportConfiguration in the future if necessary. However, I'm not sure we really need this ability. Both alternatives allow for the transport library to support both API versions. It is good by itself and also, it is a way to simplify the transport library. We could formulate the following requirements to the API updates:   - the transport API update must support previous API versions   - the updated API layout must be compatible with previous versions It seems to me, these requirements are achievable and two alternate approaches show it. Please, share your opinions?     Thanks, Serguei On 5/12/17 02:10, serguei.spit...@oracle.com wrote: Hi Dmitry, Thank you for the update with some review comments resolved! I looked at it but will not provide my comments at this point. We will need another update according to the recent design review and items where we still have to reach an agreement. Thanks, Serguei On 5/10/17 08:27, Dmitry Samersoff wrote: Serguei, Fixed minor issues (comments, netmask etc). Added an error for attempt to use allow with an old transport. http://cr.openjdk.java.net/~dsamersoff/JDK-8061228/webrev.17/ see also below. I do not understand this reason for adding more complexity. It seems, there should not be any issues in the future with rejecting all unsupported versions by the transport library. Added a diagram explaining transport version negotiation to CR. I use future versions (1.3; 1.4; 1.5) because all this versioning staff has a long term goal and allow us to develop better transport without breaking existing one. We are not adding extra parameter, we are introducing new function that is a clone of another StartListening function with a version suffix '11' in its name and with an extra parameter. Correct. We changed behavior of StartListening function and 1.1 transport shouldn't care about old one. i.e. when we document 1.1 interface we describe the only function StartListening(env, address, actualAddress, allow) that have to be placed to the StartListening11 slot. Back in 2015 I proposed to separate interfaces entirely (see webrev.04), but we (Alan?) decided that it's an overkill. The original StartListening function is being removed. It is much simpler to introduce new function AllowPeers(char* peers) with the same parameter. This separate function have to be called explicitly before we start listening, It is extra communication step. IMHO, not obvious one. So I would prefer to keep StartListening11 -Dmitry On 2017-05-10 12:37, serguei.spit...@oracle.com wrote: Dmitry, Thank you a lot for the detailed reply! On 5/10/17 01:10, Dmitry Samersoff wrote: Serguei, Please see my comments in-line. On 2017-05-10 00:42, serguei.spit...@oracle.com wrote: Hi Dmitry, I'd like to resolve my questions before the upcoming design discussion on Thu. http://cr.openjdk.java.net/%7Edsamersoff/JDK-8061228/webrev.15/src/jdk.jdwp.agent/share/native/libdt_socket/socketTransport.c.udiff.html (0) The design description from the bug report tells:     > Than we change a negotiation protocol between JDWP and transport.     > We pass maximal supported version to transport initialization routine and expect transport actual version to be returned.     The modified negotiation protocol adds extra complexity.     What is a motivation behind this?     Is it really necessary for the transport library to return an actual version instead of rejecting the unmatched version?     I do not see it is really used in the webrev.15 implementation. Transport have to return it's actual version in order to allow agent to perform appropriate action. see libjdwp/transport.c:526 This requirement adds extra complexity to the rules (transport negotiation protocol). It is not really necessary. The loadTransport() already does a lookup of a version that is accepted (not rejected) by the transport library and can save that version. The transport_startTransport() then should use the version found by the loadTransport(). Today it's just a selection of proper API call but in a future it might be too-old-transport-error or deprecation warning or security warning or something else. I do not understand this reason for adding more complexity. It seems, there should not be any issues in the future with rejecting all unsupported versions by the transport library. However, it will be even more simple if one transport library API could support/accept all possible versions (see my alternate suggestion below). (1) The following change in the jdwp transport library will reject theJDWPTRANSPORT_VERSION_1_0 as it is below       the version JDWPTRANSPORT_VERSION_1_1, but will except any version above the JDWPTRANSPORT_VERSION_1_1       (with providing/returning the actual transport version):    jdwpTransport_OnLoad(JavaVM *vm, jdwpTransportCallback* cbTablePtr, - jint version, jdwpTransportEnv** result) + jint version, void** env)    { - if (version != JDWPTRANSPORT_VERSION_1_0) { + if (version < JDWPTRANSPORT_VERSION_1_1) {            return JNI_EVERSION;        } Te following change will also prevent supporting the 1_0 version of the transport library: - interface.StartListening = &socketTransport_startListening; + interface.StartListening = NULL; libdt_socket/socketTransport.c is an implementation of 1.1 *transport* it's not intended to run with 1.0 *backend*. Why not? It would simplifies things if the transport library (and its API) is backward compatible. i.e. 1.1 *backend* can run 1.1 and 1.0 transports but 1.1 *transport* require 1.1 or greater backend. This statement creates a confusion. The truce is that the transport library can support some number of versions. The latest supported version can satisfy the agent if it supports it. see: libjdwp/transport.c:206 for version logic on backend (agent) side The logic at L206 does not require the transport library to return its version. It will work Ok if the library rejects unsupported versions. I'd suggest the following alternate change to the transport API allowing to support both old and new versions at the same time (it would simplify the negotiation rules):     - Add new function:         jdwpTransportError AllowPeers(const char* peers);     - Keep the original StartListening function.       The function uses the allowed peers data if it was previously cached by the AllowPeers().     - It seems, the alternate approach does not require adding the extra_data with version.       But if there is still a real need to get the transport API version then it'd better       to introduce a function named GetTransportVersion() or JDWP_TransportVersion().       This would allow to encapsulate any extra_data that is necessary in such a case.   From pure JDWP hardening point of view we can just add extra parameter *allow* to existing StartListening(). Caller is responsible for stack consistency so old transport continues to work. We are not adding extra parameter, we are introducing new function that is a clone of another StartListening function with a version suffix '11' in its name and with an extra parameter. The original StartListening function is being removed. It is much simpler to introduce new function AllowPeers(char* peers) with the same parameter. The original StartListening function works as before. The updated API can support both versions 1_0 and 1_1. But my goal was to create a versioning in JDWP agent -> transport relations that was missed in initial JDWP design. Further, more complicated, changes (IPv6 support, UDS sockets support etc) require this logic. Would introducing function JdwpTransportVersion() achieve what you wanted? We have a structure jdwpTransportNativeInterface with a fixed set of functions (see jdwpTransport.h:153). To add any new function to this structure we have to create a method to detect presence of this function. So we can't use GetTransportVersion(). It is not clear why would you need such a method for any new function? The transport version maps to the whole set of functions. With as separate AllowPeer() function nobody remind an agent writer that they should use it, but extra parameter makes api changes and requirements clear visible (up to compiler warning). I do not see any difference. No compiler warning if NULL is passed in place of the 'allow' parameter. Also I'm against of changing behavior of existing function. (2) The following error messages miss the actual IP address or mask that was found to be illegal: 383 RETURN_ERROR(JDWPTRANSPORT_ERROR_ILLEGAL_ARGUMENT, "invalid ip address for allow"); 392 RETURN_ERROR(JDWPTRANSPORT_ERROR_ILLEGAL_ARGUMENT, "invalid netmask for allow"); Other parameter parsing functions (e.g. "invalid port number specified" at 274) doesn't explain what parameter is bad. It is not good either. But new functionality is added, so the more diagnostic details the better.     I think typical allow would have one or two entry, so verbose error message is not worth significant complication of parsing code. It is still better to print what was not accepted. It should not be a problem to print it anyway. I would prefer to leave it as is. (3) A suggestion on the following: 377 uint32_t mask = 0xFFFFFFFF; 400 mask = 0xFFFFFFFF; // reset mask    I'd suggest a more explicit approach instead of the L400 for a better clarity: 386 if (*s == '/') { 387 // netmask specified 388 s = mask_s2u(s + 1, &mask); 389 if (*(s-1) == '/') { 390 // Input is not consumed, something bad happens 391 _peers_cnt = 0; 392 RETURN_ERROR(JDWPTRANSPORT_ERROR_ILLEGAL_ARGUMENT, "invalid netmask for allow"); 393 } 394 } else { mask = 0xFFFFFFFF; } I'll try it. Ok, thanks. http://cr.openjdk.java.net/%7Edsamersoff/JDK-8061228/webrev.15/src/jdk.jdwp.agent/share/native/libjdwp/transport.c.udiff.html (4) Some confusion with the fragment and its comment: + + /* Pass the latest supported version, + * expect actual transport version in t->extra_data->version + */ + ver = (*onLoad)(jvm, &callback, JDWPTRANSPORT_VERSION_1_1, &t); + if (ver == JNI_EVERSION) {            ver = (*onLoad)(jvm, &callback, JDWPTRANSPORT_VERSION_1_0, &t); + // Special handling for versionless transports + info->transportVersion = JDWPTRANSPORT_VERSION_1_0; + } + else { + info->transportVersion = (*t)->extra_data->version; + } + The term "latest supported version" is ambiguous in this context. Do you mean, supported by the JDWP back-end or by the agent library? Supported by the JDWP backend. I'll update comments. Ok, thanks. Also, it is not clear in what circumstances the agent library would support the version 1_0 only. The JDWP back-end is always coupled with the socket transport library, isn't it? Is it because the libdt_shmem library can be used on Windows or because a different native library path can be used? Could you explain a little bit? It's more generic question: Do we need any backward compatibility here?    I assume *yes* (can't recall why - probably it was a tree-year old decision). If we state that new backend will not support old transports today and in a future - all versioning logic is not necessary. As we see in (1) above the actual transport version can be different from requested only if the requested version is above the latest supported by the transport library. Otherwise, the version is matched or the JNI_EVERSION is returned. The subsequent call to the OnLoad function can succeed only if the library supports just the version 1_0. See answer to (1) above. transport library has exactly one version but JDWP backend supports couple of transport library versions back. I've replied above too. :) (5) Memory allocation for the info->allow field is needed only for the transport version 1_1: + if (allow != NULL) { + info->allow = jvmtiAllocate((int)strlen(allow)+1); + if (info->allow == NULL) { + serror = JDWP_ERROR(OUT_OF_MEMORY); + goto handleError; + } + (void)strcpy(info->allow, allow); + } Correct.  allocation needed for 1.1 and *greater*. I can change it, but it makes code less readable. The same fragment in a different place should not be less readable, maybe just less elegant. (6) There is no handling for the condition when the *allow* parameter is passed     but the transport version is 1_0 (which does not support the *allow* parameter): Correct. Warning or ever error should be here. I plan to open a separate follow-up CR for this problem - we have to decide how we should handle this situation (warning or error) and change the code, but I can add a plain printf() here if you like. I'm Ok with both error or warning but what is needed from a security point of view? We can avoid filing a separate follow-up CR in this case. Should it be an error if the *allow* parameter is used with the old transport library that does not support it? Thanks, Serguei + /* Interface version 1.0 doesn't support versioning, so we have to + * use global variable and set the version artifically. + * Use (*t)->extra_data->version directly when 1.0 is gone. + */ + switch(info->transportVersion) { + case JDWPTRANSPORT_VERSION_1_0:            err = (*trans)->StartListening(trans, address, &retAddress); + break; + case JDWPTRANSPORT_VERSION_1_1: + err = (*trans)->StartListening11(trans, address, &retAddress, info->allow); + break; + default: + err = JNI_EVERSION; + } -Dmitry Thanks, Serguei On 3/29/17 08:10, Dmitry Samersoff wrote: Robbin, One follow-up issue is that if you start suspended and than connect with an unallowed client the JVM starts and executes the program. Fixed. see http://cr.openjdk.java.net/~dsamersoff/JDK-8061228/webrev.15 -Dmitry On 2017-03-16 15:59, Robbin Ehn wrote: Hi Dmitry, thanks for the update. One follow-up issue is that if you start suspended and than connect with an unallowed client the JVM starts and executes the program. Simple program prints "Hello". [rehn@rehn-ws vanilla-hs]$ java -agentlib:jdwp=transport=dt_socket,server=y,suspend=y,address=*:9999,allow=1.2.3.0/32 -cp . H Listening for transport dt_socket at address: 9999 ERROR: Peer not allowed to connect Listening for transport dt_socket at address: 9999 Hello I think it would be good if the VM stayed suspended when an unallowed client connects. Thanks, Robbin On 03/13/2017 08:07 PM, Dmitry Samersoff wrote: Robbin, Please, see: http://cr.openjdk.java.net/~dsamersoff/JDK-8061228/webrev.11 1: So connecting with an unallowed client terminates the VM. Fixed. 2: Starting with an bad allow filter terminates the VM when connecting a client. Moved allowed parameter (and parser call) to StartListening. -Dmitry On 2017-03-10 15:56, Robbin Ehn wrote: Hi Dmitry, I took a look at this, I have two practical issues: 1: [rehn@rehn-ws dev]$ java -agentlib:jdwp=transport=dt_socket,server=y,suspend=n,address=*:9999,allow=6.6.6.6 -cp runs ForEver Listening for transport dt_socket at address: 9999 ERROR: transport error 202: peer not allowed to connect: Success JDWP exit error JVMTI_ERROR_NONE(0): could not connect, timeout or fatal error [transport.c:358] So connecting with an unallowed client terminates the VM. 2: [rehn@rehn-ws dev]$ java -agentlib:jdwp=transport=dt_socket,server=y,suspend=n,address=*:9999,allow=6.BAD.6.6 -cp runs ForEver Listening for transport dt_socket at address: 9999 ERROR: transport error 202: unable to parse list of allowed peers: Success JDWP exit error JVMTI_ERROR_NONE(0): could not connect, timeout or fatal error [transport.c:358] Starting with an bad allow filter terminates the VM when connecting a client. Connecting with an unallowed ip/port should not terminate the VM and we should verify allow filter directly at startup. Thanks /Robbin On 02/28/2017 10:41 AM, Dmitry Samersoff wrote: Everybody, Please review: http://cr.openjdk.java.net/~dsamersoff/JDK-8061228/webrev.10/ These changes introduce new parameter[1] of the socket transport - allow. Users can explicitly specify a list of hosts that allowed to connect to jdwp server and it's the second part of JDWP hardening[2]. No restrictions are applied by default now but I'll file a separate CR to restrict list of allowed peers to localhost by default. Also these changes implement versioning for jdwp transport and therefor simplify feature development of jdwp. 1. Example command line: -agentlib:jdwp=transport=dt_socket,server=y,suspend=n, address=*,allow="127.0.0.0/8;192.168.0.0/24" Possible values for allow parameter:      *           - accept connections from everywhere.      N.N.N.N     - accept connections from this IP address only      N.N.N.N/nn  - accept connections from particular ip subnet 2. JDK-8052136 JDWP hardening -Dmitry