Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
> On 4 Aug 2021, at 01:19, Peter Firmstone wrote: > > Excellent, Ron, that's exactly what I'm after. > > I need to be able to implement an authorization layer on top of the JDK, but > reach down into the JDK to use authorization to control access. > > Can we find out how many such checks that OpenJDK is prepared to support, > then we will pick the most important? > > Pick any three you’d like. I’m joking. How many or few — or whether this is the right approach at all — depends, like everything, on value vs. cost. Maybe Sean or Alan can tell you where the best place to discuss this is — there probably won’t be a separate mailing list, as it won’t be a big project — but perhaps it’s best to let things sink in for a while and take time to think. The OpenJDK development process is such that the relevant stewards of the platform, people like Alan and Sean, make the decisions, the committers do the development, and relevant (and, preferably succinct) feedback from users is appreciated. It’s important to remember that while all feedback is taken seriously, not all of it is accepted, for the simple reason that it is ultimately the stewards’ job to consider the interests of the entire ecosystem — to the best of their abilities — and not just that of those who give feedback. To see what ideas are currently being considered and which proposals are planned — the stewards will present them to the OpenJDK community at the appropriate time — see the Future Work section of the JEP: https://openjdk.java.net/jeps/411#Future-Work — Ron
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
Maybe we need some criteria, that defines what's not easily instrumented? On 4/08/2021 10:19 am, Peter Firmstone wrote: Excellent, Ron, that's exactly what I'm after. I need to be able to implement an authorization layer on top of the JDK, but reach down into the JDK to use authorization to control access. Can we find out how many such checks that OpenJDK is prepared to support, then we will pick the most important? Don't worry about ClassLoader, I can instrument that (thank you Erik), and maybe I can instrument Properties, and System.exit. So basically things we can't easily instrument with agents, that everybody is most likely to want. * Network access * File System access * User Credentials Maybe we should have a mailing list dedicated to this where we can discuss and potentially collaborate? Regards, Peter. On 3/08/2021 10:15 pm, Ron Pressler wrote: On 3 Aug 2021, at 12:50, Peter Firmstone wrote: Can you think of any workable alternative compromises? If you mean a compromise between no access checks in the JDK and all access checks in the JDK, then yes, which is possibly some callbacks for a small subset of operations that perform access checks today, say, System.exit and opening a file or socket. I am not saying this is what should be done, but that the effort involved is such that I can conceivably see those whose responsibility this would be agreeing to consider it, as the value in such a mechanism might end up being worthy of that amount of effort. But I’m guessing that the more such hooks are requested, the less likely it is that the cost remains acceptable. — Ron -- Regards, Peter Firmstone 0498 286 363 Zeus Project Services Pty Ltd.
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
Excellent, Ron, that's exactly what I'm after. I need to be able to implement an authorization layer on top of the JDK, but reach down into the JDK to use authorization to control access. Can we find out how many such checks that OpenJDK is prepared to support, then we will pick the most important? Don't worry about ClassLoader, I can instrument that (thank you Erik), and maybe I can instrument Properties, and System.exit. So basically things we can't easily instrument with agents, that everybody is most likely to want. * Network access * File System access * User Credentials Maybe we should have a mailing list dedicated to this where we can discuss and potentially collaborate? Regards, Peter. On 3/08/2021 10:15 pm, Ron Pressler wrote: On 3 Aug 2021, at 12:50, Peter Firmstone wrote: Can you think of any workable alternative compromises? If you mean a compromise between no access checks in the JDK and all access checks in the JDK, then yes, which is possibly some callbacks for a small subset of operations that perform access checks today, say, System.exit and opening a file or socket. I am not saying this is what should be done, but that the effort involved is such that I can conceivably see those whose responsibility this would be agreeing to consider it, as the value in such a mechanism might end up being worthy of that amount of effort. But I’m guessing that the more such hooks are requested, the less likely it is that the cost remains acceptable. — Ron
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
> On 3 Aug 2021, at 12:50, Peter Firmstone wrote: > > Can you think of any workable alternative compromises? If you mean a compromise between no access checks in the JDK and all access checks in the JDK, then yes, which is possibly some callbacks for a small subset of operations that perform access checks today, say, System.exit and opening a file or socket. I am not saying this is what should be done, but that the effort involved is such that I can conceivably see those whose responsibility this would be agreeing to consider it, as the value in such a mechanism might end up being worthy of that amount of effort. But I’m guessing that the more such hooks are requested, the less likely it is that the cost remains acceptable. — Ron
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
> On 3 Aug 2021, at 06:48, Peter Firmstone wrote: > > > We can still use these without an SM, Policy or Permissions for authorization > decisions, as mentioned previously I'd replace the inherited thread context > with an unprivileged context, and also allow the stack walk to be disabled > for people only using Subject. > I think what you mean is that you can envision using the same API points for a different, but reasonably similar role to the one they have. But that would mean changing the behaviour of existing classes, possibly making some final classes non-final, in non-trivial ways. > > Just performed a search for java.security.AccessController on GitHub, got > 1,398,418 results for Java: > The plan is to degrade these into no-ops until such time as most of those usages disappear, not to imbue those lines of code with new meaning. The actual removal of the API elements might be a long way off, but, becoming no-ops before then, the JDK and libraries will be free to remove those usages.
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
> On 3 Aug 2021, at 09:39, Peter Firmstone wrote: > > > > Can you elaborate? A Runnable, similar to registering a shutdown hook? > Maybe, but there have been no specifics beyond saying that something like that, and which requires that level of effort, might, after the dust settles and no one is rushing to decisions, possibly be something that we could contemplate thinking about considering as possibly not being obviously completely unreasonable, and might, perhaps, be worthy of another look.
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
What you are proposing is: 1. Add an SPI for deep permission checks in the JDK. 2. Do so by changing the behaviour of existing classes in a non-trivial way. 3. Keep access checks in the JDK. I do not speak for those who work on the security mechanisms, but I believe that none one of these, even in isolation, would be reasonable. For one, even without the principle against changing the behaviour of existing classes in drastic and non-trivial ways, without the use of doPrivileged in the JDK, AccessController would make a bad API. For another, the cost of SM isn’t in the lines of code that make up the AccessController implementation but in the fact that the JDK does access checks and privileged operations in the first place, as those impose a tax on everything, and also take resources away from what we believe is more robust and effective security. The starting point isn’t removing the policy file implementation, but removing all access checks and privileged operations from the JDK, and everything else follows from there (it will take a long time until API elements are actually removed, anyway). Even if someone might choose to contemplate adding callbacks to one or two I/O operations, it is not the intent for the JDK to participate in a general access checking mechanism deep in its guts in addition to strong encapsulation; quite the opposite. — Ron > On 3 Aug 2021, at 10:44, Peter Firmstone wrote: > > Thanks Ron, reply inline. > > On 3/08/2021 6:48 pm, Ron Pressler wrote: >> >>> On 3 Aug 2021, at 06:48, Peter Firmstone >>> wrote: >>> >>> >>> We can still use these without an SM, Policy or Permissions for >>> authorization decisions, as mentioned previously I'd replace the inherited >>> thread context with an unprivileged context, and also allow the stack walk >>> to be disabled for people only using Subject. >>> >>> >> I think what you mean is that you can envision using the same API points for >> a different, but reasonably similar >> role to the one they have. But that would mean changing the behaviour of >> existing classes, possibly making some >> final classes non-final, in non-trivial ways. >> > I'd limit changes to: > > • Make the stack walk optional (via command line argument to disable > it). > • Remove Thread's inherited context, replace it with an unprivileged > context. > This would allow us to use the API for virtual threads, eg to obtain Subject > credentials to authenticate TLS connections. > > It also means that for someone implementing guard checks, that these only > need check the thread stack back to the last doPrivileged call, or the start > of the thread, in the latter case it will have no privileges. It fixes the > viral permission check problem, usually doPrivileged calls are short and > sweet. > > It may require the addition of doPrivileged calls where they're currently > missing (and should have been used), where they've been responsible for viral > permission checks. > >> >>> Just performed a search for java.security.AccessController on GitHub, got >>> 1,398,418 results for Java: >>> >>> >> The plan is to degrade these into no-ops until such time as most of those >> usages disappear, not to imbue >> those lines of code with new meaning. The actual removal of the API elements >> might be a long way off, >> but, becoming no-ops before then, the JDK and libraries will be free to >> remove those usages. >> > No new meanings, the same as they have now is sufficient, just we leave the > granularity of the checks to the developers of guards and provide a means by > which guards can be registered for common check points that developers > request (perhaps via a poll), rather than all existing permission check > points. Keeping in mind that we are not trying to isolate code, but perform > authorization access checks, as well as provide credentials for > authentication. > > For example, if someone is only concerned with stopping the JVM from exiting, > then they only implement a guard for that particular hook, the actual code > that needs to call System::exit, then calls a doPrivileged method before > doing so. The guard need only check the domain on the stack is the one it > expects, which could be based on Principal, CodeSource, Module or ClassLoader > etc, they may also chose to implement something more complex. > > Someone else may only be concerned with network connections, so they only > implement and register a guard for that. > > So basically we don't dictate how to implement guards or policy, we just > leave enough in place, to ensure that a minimalist authorization access > control api is common among all implementations on all Java versions. > > It is suitable, for Subject's only or code and Subject's. > > The doPrivileged call simply indicates the code is requesting to do something > that might be privileged, or needs to provide credentials for authentication, > as it does now, but
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
> On 3 Aug 2021, at 02:30, Peter Firmstone wrote: > > > I am still hopeful that OpenJDK might create some hooks for us and keep the > AccessController, AccessControlContext and Subject.doAs methods to make > supporting all Java versions easier. When I mentioned “hooks” I was referring to something along the lines of a callback when a file is accessed. > > Just curious, when using Agents, what are the recommendations for line > numbers in code, for exceptions etc, how are these affected when > instrumenting? When you transform a class, you basically replace its content with new content, so the short answer is that what happens to line numbers is whatever you want. If you’re using ASM for your transformation and ignore line numbers, which ASM represents as a kind of no-op instructions, then what would usually happen is that if original instruction X is mapped to line N and when transforming the stream you replace X with YXZ, then all of YXZ will remain mapped to line N.
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
Our path to making Java more secure is, indeed, similar to what you’re proposing. The general direction is that the application, rather than the libraries, will have the final say on security-sensitive capabilities. Just as strong encapsulation is now on by default, other things, such as native code, some dynamic Agent capabilities, and perhaps others, will likely require explicit permission by the application on the command line. I don’t know the details about management, but I believe that remote management requires explicit permission already. I am not aware, however, of plans to restrict class loading. The idea is that code is trusted to do what it says it does, and what’s restricted is the ability to manipulate it (without permission) to do something else through transformation or circumvention of its declared API. The default security policy today is already more restrictive than it was a couple of versions ago, and we expect it to become even more restrictive. — Ron > On 2 Aug 2021, at 11:33, Peter Firmstone wrote: > > Hello Andrew, > > I think you may be misinterpreting my comment, let me clarify: > > I'm assuming that during the process of removal of security manager, any > external ports or process hooks that we can only turn off now by not granting > a permission will be replaced by a command line property or something > similar? Eg, Agents, Management, etc. If this is the case, it would be nice > if they were set to off by default, such that they needed to be enabled from > the command line. It's a suggestion. So configuration is secure by > default. Eg, similar to how OpenBSD is configured secure by default, is that > something RedHat does? I used RedHat back in the 90's, it was my first > Linux distro, on sparc. :) > > For example, the permissions granted in this policy file: > > https://urldefense.com/v3/__https://github.com/pfirmstone/JGDMS/blob/trunk/qa/harness/policy/defaultsecurephoenix.policy.new__;!!ACWV5N9M2RV99hQ!cVJkwzysH02ZFZWT3rnO_o5viWUCBTm9nnM50QaJ5G9hsg9B3p2naDVYknIrqSNURQ$ > > Ensure that anything that isn't specifically granted, is switched off, eg > agents, management, etc. > > This file is generated by a policy writing tool, the comments in the policy > file, indicate the tests that were run to generate the policy. It's > important to note that we aren't using sandboxing, or isolation, we are > simply running with privileges we don't require switched off. All network > entry points to the JVM require authentication and classes loaded, or data > parsed requires authentication first. > > In my software without SM, authentication isn't required, and there are no > defenses against parsing un-trusted data or downloading and loading malicious > classes. Although you've made it clear, you think this is a /special case/ > and /special loss/ you needn't concern yourself with and that's fine, you've > made your point. > > In future it will not be possible to switch these features off using policy > files, so I imagine that OpenJDK must be considering alternative methods to > switch off features that may be security sensitive? > > We already know de-serialization can be switched off from the command line > using: -Djdk.serialFilter=!*, > > However if you switch off java de-serialization, you best be sure to > initialize the serialization framework when starting your JVM, as SM won't be > there to stop it being turned back on, if it's not used, it's not > initialized, then the property can be changed and is potentially a link in a > gadget chain attack where serialization can be re-enabled. :) Just saying. > > Not sure what canard you're referring to, is the JDK config secure by > default? If so, then it won't be possible for a malicious java attachment in > an email (or an image file) to be run inadvertently and use the attach API to > steal secrets. Although it's processor could be subject to speculative > execution attacks. > > A good reason to use OpenBSD (or Solaris perhaps) on Sparc for secure > systems? I know RedHat is into supporting military application in a big > way, but I'm surprised if they're not concerned about speculative execution > vulnerabilities. > > Speculative execution attacks will eventually be addressed in processor > architectures, they are trying, we shouldn't assume that we don't need to > program defensively because the underlying architecture has speculative > execution bugs. If we do that, then were do we stop? > > Anyway, if possible I would like to get back to discussing how to instrument > the Java API using Agents, and whether we can coordinate removal of > finalizers with SM removal. > > Of course if you'd like to keep SM and cancel JEP 411, as Uwe suggested, I'd > support that. I'm also grateful that Uwe spoke up, and I hope others do as > well, so this topic receives further productive discussion. It would be nice > to sort the use cases prior to SM
Re: [External] : RE: JEP 411, removal of finalizers, a path forward.
Hi. JEP 411 says that there is a feature whose value in fulfilling its purpose no longer justifies its high cost, and so it should be removed. What you’re saying is, that’s true, but you should keep it, anyway. Why? Because it can be put to some other, more modest, accidental uses. Now, the "disaster" of removing any feature is directly felt by those who use it, but the "disaster" of *not* removing it is less visible, even though it may be the bigger disaster, as we believe is the case here. Sure, we must consider those applications that use it and how they use it, but we must also consider all those that do not. By not acknowledging those who lose from keeping that mechanism, you're not appreciating the problem that the JEP tries to address. For example, you say you care about non-security uses of SM, but maintaining it comes at the cost of other security measures, and we believe that removing it would improve Java's security overall. While some of the features you say you want could well be addressed by cheaper mechanisms that could reasonably be added, it's possible that some cannot (or not fully). Thankfully, the JDK is flexible and hackable enough — through things like Java Agents — that the cost of some capabilities could be borne by those who need them. It's perfectly reasonable to explain what functionality would be helpful to you. It is not reasonable to demand keeping some functionality regardless of its cost. The reality is that to adapt to changing circumstances, the JDK must both add new capabilities as well as drop old ones, or its complexity, and, with it, its maintenance cost would grow without bound. Of course, any help in finding the right balance of cost and benefit in consideration of the needs of the *entire* ecosystem would be appreciated. P.S. Loom has nothing to do with JEP 411. I might have mentioned it when pointing out that virtually all relevant Java features developed in the past decade -- like parallel streams, CompletableFuture, and Loom -- have chosen not to integrate with SM. — Ron > On 1 Aug 2021, at 15:28, Uwe Schindler wrote: > > Hi Andrew, > >>> I'm working on the assumption that OpenJDK will close any external holes >>> currently defended by permission checks. It would be good if the JDK >>> was secure by default, with properties required to be set for allowing >>> such things as agents, management, parsing xml and serialization. >> >> You need to stop repeating this canard. There is no absolute need for >> OpenJDK to retain a security mechanism to deal with problems that for >> almost every use case are better solved by using non-OpenJDK >> alternatives (such as OS security measures). Indeed, it's the other way >> round: there is an imperative for the project to spend precious >> resources on alternative capabilities (not necessarily security related). > > Sorry, as another open source project affected by the stupid JEP 411 desaster > I would like to fully confirm to EVERYTHING that Peter said. It is not a > canard, it is the reality and I am really disappointed what happened. > > What I figured out: You intend to remove SecurityManager because it does not > fit your latest ideas how Java threads should behave. I know the main problem > is not "SecurityManager is too complex / too slow / wrongly used /..." -- the > main problem of some OpenJDK people around the Loom project is that it won't > work correctly with those new type of threads. You are now always arguing > against use cases of SecurityManager for the purpose of secuirty because you > just want to hide that the new "light" threads (aka fibers) in project Loom > are incompatible to the stack-based access control provided by > AccessController and SecurityManager. So the only canard is Project Loom - > sorry! > > So I will do the same: repeat the same story like Peter's over an over. Two > large projects affected by your changes are Apache Solr and Elasticsearch. > Both use security manager not only for security purposes, but also to control > access to resources and classloading! Speed is also no issue if you use > security manager correctly. In Apache Lucene/Solr/Elasticsearch there is NO > slowdown visible as the hooks into security manager are only called for > uncritical places (like opening a file before mmapping it, open a listening > socket, loading a class...). > > The problem is: you deprecate for removal without offering any API to replace > the main pain points: > - preventing exit of JDK > - an API to control class loading and prevent code from calling > Method/Field/Constructor#setAccessible(true) for specific cases (e.g., you > want a scripting language like Groovy to not allow caling > sun.misc.Unsafe,...). Those things can never be done with code outside of the > JVM! > - Allow to hook into the I/O system. Unfortunately the java.nio FileSystem > API is not enough: it does not cover java.io.File (why is this the case?) nor > does the FileSystem API
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
Can you think of any workable alternative compromises?
On 3/08/2021 9:21 pm, Ron Pressler wrote:
What you are proposing is:
1. Add an SPI for deep permission checks in the JDK.
2. Do so by changing the behaviour of existing classes in a non-trivial way.
3. Keep access checks in the JDK.
I do not speak for those who work on the security mechanisms, but I believe
that none one of these,
even in isolation, would be reasonable.
For one, even without the principle against changing the behaviour of existing
classes in drastic
and non-trivial ways, without the use of doPrivileged in the JDK,
AccessController would make
a bad API.
That's true, it would better if those doPrivileged calls to remain. An
alternative I thought about if they were removed, was to treat a thread
as privileged if all domains on the thread stack were resolvable from
the bootstrap ClassLoader.
For another, the cost of SM isn’t in the lines of code that make up the
AccessController implementation
but in the fact that the JDK does access checks and privileged operations in
the first place, as those
impose a tax on everything, and also take resources away from what we believe
is more robust and effective
security.
Yes, I realise this, however I'm not looking to replace all permission
checks, just the most common. Perhaps then a static method added to a
class, without implementation would be better, in cases where it's not
practical to instrument the JVM. Such that it does nothing by default
and would be optimised away by hotspot.
eg:
static instrumentGuard(String [] args) throws SecurityException{}
Thanks,
Peter.
The starting point isn’t removing the policy file implementation, but removing
all access checks
and privileged operations from the JDK, and everything else follows from there
(it will take a long
time until API elements are actually removed, anyway). Even if someone might
choose to contemplate adding
callbacks to one or two I/O operations, it is not the intent for the JDK to
participate in a general access
checking mechanism deep in its guts in addition to strong encapsulation; quite
the opposite.
— Ron
On 3 Aug 2021, at 10:44, Peter Firmstone wrote:
Thanks Ron, reply inline.
On 3/08/2021 6:48 pm, Ron Pressler wrote:
On 3 Aug 2021, at 06:48, Peter Firmstone
wrote:
We can still use these without an SM, Policy or Permissions for authorization
decisions, as mentioned previously I'd replace the inherited thread context
with an unprivileged context, and also allow the stack walk to be disabled for
people only using Subject.
I think what you mean is that you can envision using the same API points for a
different, but reasonably similar
role to the one they have. But that would mean changing the behaviour of
existing classes, possibly making some
final classes non-final, in non-trivial ways.
I'd limit changes to:
• Make the stack walk optional (via command line argument to disable
it).
• Remove Thread's inherited context, replace it with an unprivileged
context.
This would allow us to use the API for virtual threads, eg to obtain Subject
credentials to authenticate TLS connections.
It also means that for someone implementing guard checks, that these only need
check the thread stack back to the last doPrivileged call, or the start of the
thread, in the latter case it will have no privileges. It fixes the viral
permission check problem, usually doPrivileged calls are short and sweet.
It may require the addition of doPrivileged calls where they're currently
missing (and should have been used), where they've been responsible for viral
permission checks.
Just performed a search for java.security.AccessController on GitHub, got
1,398,418 results for Java:
The plan is to degrade these into no-ops until such time as most of those
usages disappear, not to imbue
those lines of code with new meaning. The actual removal of the API elements
might be a long way off,
but, becoming no-ops before then, the JDK and libraries will be free to remove
those usages.
No new meanings, the same as they have now is sufficient, just we leave the
granularity of the checks to the developers of guards and provide a means by
which guards can be registered for common check points that developers request
(perhaps via a poll), rather than all existing permission check points.
Keeping in mind that we are not trying to isolate code, but perform
authorization access checks, as well as provide credentials for authentication.
For example, if someone is only concerned with stopping the JVM from exiting,
then they only implement a guard for that particular hook, the actual code that
needs to call System::exit, then calls a doPrivileged method before doing so.
The guard need only check the domain on the stack is the one it expects, which
could be based on Principal, CodeSource, Module or ClassLoader etc, they may
also chose to implement something more complex.
Someone
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
Wow, thanks Andrew,
There's certainly some complexity there, I appreciate your guidance.
Regards,
Peter.
On 3/08/2021 7:11 pm, Andrew Dinn wrote:
On 03/08/2021 02:30, Peter Firmstone wrote:
Just curious, when using Agents, what are the recommendations for
line numbers in code, for exceptions etc, how are these affected when
instrumenting?
For Byteman I use ASM to inject bytecode sequences into methods
without any (significant) modification to existing bytecodes. ASM
successfully adjusts line number info to allow for the injected code
without the need for my agent to take any action.
I also add extra exception handling around injected regions and
propagate exceptions through enclosing regions to an outer handler.
ASM updates the exception table according to where the method visitor
notifies new try and catch region boundaries while retaining the
regions determined by existing notifications.
ASM provides mechanism for more complex cases. If you are rewriting
existing bytecodes or exception regions then you can use the ASM
visitor pattern to process and then renotify both line numbers and try
catch boundaries as you visit the bytecode stream. Of course, the
logic of how to do that can be as complex as (or more so than) the
rewrite logic.
n.b. Remapping exception regions at load time is tricky for two
reasons. Firstly, you cannot afford to load exception classes to
investigate the exception class hierarchy without losing the
opportunity to transform the exception class's bytecode. So, you have
to identify exception shadowing by parsing the bytecode as a resource
to establish exception super relationships. Secondly, if you tinker
with or inject extra try regions you need to a) ensure they are
properly nested inside containing try regions b) make sure that new
exception exit paths out of synchronized regions include monitorexits.
regards,
Andrew Dinn
---
Red Hat Distinguished Engineer
Red Hat UK Ltd
Registered in England and Wales under Company Registration No. 03798903
Directors: Michael Cunningham, Michael ("Mike") O'Neill
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
Thanks Ron, reply inline. On 3/08/2021 6:48 pm, Ron Pressler wrote: On 3 Aug 2021, at 06:48, Peter Firmstone wrote: We can still use these without an SM, Policy or Permissions for authorization decisions, as mentioned previously I'd replace the inherited thread context with an unprivileged context, and also allow the stack walk to be disabled for people only using Subject. I think what you mean is that you can envision using the same API points for a different, but reasonably similar role to the one they have. But that would mean changing the behaviour of existing classes, possibly making some final classes non-final, in non-trivial ways. I'd limit changes to: 1. Make the stack walk optional (via command line argument to disable it). 2. Remove Thread's inherited context, replace it with an unprivileged context. This would allow us to use the API for virtual threads, eg to obtain Subject credentials to authenticate TLS connections. It also means that for someone implementing guard checks, that these only need check the thread stack back to the last doPrivileged call, or the start of the thread, in the latter case it will have no privileges. It fixes the viral permission check problem, usually doPrivileged calls are short and sweet. It may require the addition of doPrivileged calls where they're currently missing (and should have been used), where they've been responsible for viral permission checks. Just performed a search for java.security.AccessController on GitHub, got 1,398,418 results for Java: The plan is to degrade these into no-ops until such time as most of those usages disappear, not to imbue those lines of code with new meaning. The actual removal of the API elements might be a long way off, but, becoming no-ops before then, the JDK and libraries will be free to remove those usages. No new meanings, the same as they have now is sufficient, just we leave the granularity of the checks to the developers of guards and provide a means by which guards can be registered for common check points that developers request (perhaps via a poll), rather than all existing permission check points. Keeping in mind that we are not trying to isolate code, but perform authorization access checks, as well as provide credentials for authentication. For example, if someone is only concerned with stopping the JVM from exiting, then they only implement a guard for that particular hook, the actual code that needs to call System::exit, then calls a doPrivileged method before doing so. The guard need only check the domain on the stack is the one it expects, which could be based on Principal, CodeSource, Module or ClassLoader etc, they may also chose to implement something more complex. Someone else may only be concerned with network connections, so they only implement and register a guard for that. So basically we don't dictate how to implement guards or policy, we just leave enough in place, to ensure that a minimalist authorization access control api is common among all implementations on all Java versions. It is suitable, for Subject's only or code and Subject's. The doPrivileged call simply indicates the code is requesting to do something that might be privileged, or needs to provide credentials for authentication, as it does now, but it's the light version of the stack walk, if doPrivileged is not called, then the context will have an unprivileged domain on the stack (that initialized when Thread was created.). It's also possible to register guards that do fine grained permission checks, similar to the way Java does now. Then there's the use case, or registering no guards at all, and disabling the stack walk, and only using the api to obtain and preserve Subject credentials for authentication. You can trust me on this one, I'm experienced with the current API and have pushed it to all sorts of limits. Cheers, Peter.
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
On 03/08/2021 02:30, Peter Firmstone wrote:
Just curious, when using Agents, what are the recommendations for line
numbers in code, for exceptions etc, how are these affected when
instrumenting?
For Byteman I use ASM to inject bytecode sequences into methods without
any (significant) modification to existing bytecodes. ASM successfully
adjusts line number info to allow for the injected code without the need
for my agent to take any action.
I also add extra exception handling around injected regions and
propagate exceptions through enclosing regions to an outer handler. ASM
updates the exception table according to where the method visitor
notifies new try and catch region boundaries while retaining the regions
determined by existing notifications.
ASM provides mechanism for more complex cases. If you are rewriting
existing bytecodes or exception regions then you can use the ASM visitor
pattern to process and then renotify both line numbers and try catch
boundaries as you visit the bytecode stream. Of course, the logic of how
to do that can be as complex as (or more so than) the rewrite logic.
n.b. Remapping exception regions at load time is tricky for two reasons.
Firstly, you cannot afford to load exception classes to investigate the
exception class hierarchy without losing the opportunity to transform
the exception class's bytecode. So, you have to identify exception
shadowing by parsing the bytecode as a resource to establish exception
super relationships. Secondly, if you tinker with or inject extra try
regions you need to a) ensure they are properly nested inside containing
try regions b) make sure that new exception exit paths out of
synchronized regions include monitorexits.
regards,
Andrew Dinn
---
Red Hat Distinguished Engineer
Red Hat UK Ltd
Registered in England and Wales under Company Registration No. 03798903
Directors: Michael Cunningham, Michael ("Mike") O'Neill
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
On 3/08/2021 6:31 pm, Ron Pressler wrote: On 3 Aug 2021, at 02:30, Peter Firmstone wrote: I am still hopeful that OpenJDK might create some hooks for us and keep the AccessController, AccessControlContext and Subject.doAs methods to make supporting all Java versions easier. When I mentioned “hooks” I was referring to something along the lines of a callback when a file is accessed. Can you elaborate? A Runnable, similar to registering a shutdown hook? Just curious, when using Agents, what are the recommendations for line numbers in code, for exceptions etc, how are these affected when instrumenting? When you transform a class, you basically replace its content with new content, so the short answer is that what happens to line numbers is whatever you want. If you’re using ASM for your transformation and ignore line numbers, which ASM represents as a kind of no-op instructions, then what would usually happen is that if original instruction X is mapped to line N and when transforming the stream you replace X with YXZ, then all of YXZ will remain mapped to line N. Ok, Currently my experience with ASM is finding dependencies, and unpacking the pack200 format. I hadn't stopped to consider what it might do to line numbers. Cheers, Peter.
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
Just thought I'd share some thoughts around a couple of statements in
JEP 411:
*|java.security.{AccessController, AccessControlContext,
AccessControlException, DomainCombiner}|* — The primary APIs for the
access controller, which is the default implementation to which the
Security Manager delegates permission checks. These APIs do not have
value without the Security Manager, since certain operations will
not work without both a policy implementation and access-control
context support in the VM.
*|javax.security.auth.SubjectDomainCombiner|* and
*|javax.security.auth.Subject::{doAsPrivileged, getSubject}|* — APIs
for user-based authorization that are dependent on Security Manager
APIs such as |AccessControlContext| and |DomainCombiner|. We plan to
provide a replacement API
for
|Subject::getSubject| since it is commonly used for use cases that
do not require the Security Manager, and to continue to support use
cases involving |Subject::doAs| (see below
).
We can still use these without an SM, Policy or Permissions for
authorization decisions, as mentioned previously I'd replace the
inherited thread context with an unprivileged context, and also allow
the stack walk to be disabled for people only using Subject.
The advantage of retaining these methods, is that where these are used
to preserve context for establishing TLS and Kerberos connections, they
continue to function, even without SM or policy.
They are also a standard API for an authorization layer, for those
wishing to re-implement it.
Just performed a search for java.security.AccessController on GitHub,
got 1,398,418 results for Java:
https://github.com/search?l=Java=desc=3=java.security.AccessController+extension%3A.java+language%3AJava+language%3AJava+language%3AJava+language%3AJava=indexed=Code
An example of how DomainCombiner provides access to the underlying
ProtectionDomain's on the stack. We don't need to call
ProtectionDomain::Implies(Permission) to make authorization decisions.
https://github.com/pfirmstone/JGDMS/blob/trunk/JGDMS/jgdms-platform/src/main/java/org/apache/river/api/security/CombinerSecurityManager.java
--
Regards,
Peter
On 3/08/2021 11:30 am, Peter Firmstone wrote:
Thanks Ron,
What we do now is dynamic, so we need to figure out how to replicate
that post SM. Things we don't grant dynamically are good candidates
for command line argument options.
We basically authenticate, then authorize class loading dynamically at
runtime, along with other things, such as deserialization (not to be
confused with Java Serialization) and network connections. The dynamic
grants are removed by garbage collection. Dynamic grants are probably
very rare for most developers, although OSGi also supports dynamic
grants also, I had a discussion with Peter Kriens around OSGi dynamic
grants some years ago.
I am still hopeful that OpenJDK might create some hooks for us and
keep the AccessController, AccessControlContext and Subject.doAs
methods to make supporting all Java versions easier. Also replacing
thread inherited context, with an unprivileged context would be good
too. Then remove all the permission implementations and we could use
the hooks we need. The hooks could simply be an empty static method
that's not subject to change, so we can instrument it with our own
implementation and it gets optimized out at runtime if not used by
hotspot. Clearly if there are alternative mechanisms available from
the command line that we don't need to dynamically enable, then they
are preferable to adding clutter to policy files. The existing
policy provider can be removed, so we can strip it down to just the
essential minimum components, which removes much of the testing
required by OpenJDK.
Just curious, when using Agents, what are the recommendations for line
numbers in code, for exceptions etc, how are these affected when
instrumenting?
Quite happy to contribute time to adding hooks and removing other
components.
Regards,
Peter.
On 2/08/2021 9:00 pm, Ron Pressler wrote:
Our path to making Java more secure is, indeed, similar to what
you’re proposing.
The general direction is that the application, rather than the
libraries, will have the final say on security-sensitive capabilities.
Just as strong encapsulation is now on by default, other things, such
as native code, some dynamic Agent capabilities, and
perhaps others, will likely require explicit permission by the
application on the command line. I don’t know the details about
management, but I believe that remote management requires explicit
permission already.
I am not aware, however, of plans to restrict class loading. The idea
is that code is trusted to do what it says it does, and
what’s restricted is the ability to manipulate it (without
permission) to do something else
Re: [External] : Re: JEP 411, removal of finalizers, a path forward.
Thanks Ron, What we do now is dynamic, so we need to figure out how to replicate that post SM. Things we don't grant dynamically are good candidates for command line argument options. We basically authenticate, then authorize class loading dynamically at runtime, along with other things, such as deserialization (not to be confused with Java Serialization) and network connections. The dynamic grants are removed by garbage collection. Dynamic grants are probably very rare for most developers, although OSGi also supports dynamic grants also, I had a discussion with Peter Kriens around OSGi dynamic grants some years ago. I am still hopeful that OpenJDK might create some hooks for us and keep the AccessController, AccessControlContext and Subject.doAs methods to make supporting all Java versions easier. Also replacing thread inherited context, with an unprivileged context would be good too. Then remove all the permission implementations and we could use the hooks we need. The hooks could simply be an empty static method that's not subject to change, so we can instrument it with our own implementation and it gets optimized out at runtime if not used by hotspot. Clearly if there are alternative mechanisms available from the command line that we don't need to dynamically enable, then they are preferable to adding clutter to policy files. The existing policy provider can be removed, so we can strip it down to just the essential minimum components, which removes much of the testing required by OpenJDK. Just curious, when using Agents, what are the recommendations for line numbers in code, for exceptions etc, how are these affected when instrumenting? Quite happy to contribute time to adding hooks and removing other components. Regards, Peter. On 2/08/2021 9:00 pm, Ron Pressler wrote: Our path to making Java more secure is, indeed, similar to what you’re proposing. The general direction is that the application, rather than the libraries, will have the final say on security-sensitive capabilities. Just as strong encapsulation is now on by default, other things, such as native code, some dynamic Agent capabilities, and perhaps others, will likely require explicit permission by the application on the command line. I don’t know the details about management, but I believe that remote management requires explicit permission already. I am not aware, however, of plans to restrict class loading. The idea is that code is trusted to do what it says it does, and what’s restricted is the ability to manipulate it (without permission) to do something else through transformation or circumvention of its declared API. The default security policy today is already more restrictive than it was a couple of versions ago, and we expect it to become even more restrictive. — Ron On 2 Aug 2021, at 11:33, Peter Firmstone wrote: Hello Andrew, I think you may be misinterpreting my comment, let me clarify: I'm assuming that during the process of removal of security manager, any external ports or process hooks that we can only turn off now by not granting a permission will be replaced by a command line property or something similar? Eg, Agents, Management, etc. If this is the case, it would be nice if they were set to off by default, such that they needed to be enabled from the command line. It's a suggestion. So configuration is secure by default. Eg, similar to how OpenBSD is configured secure by default, is that something RedHat does? I used RedHat back in the 90's, it was my first Linux distro, on sparc. :) For example, the permissions granted in this policy file: https://urldefense.com/v3/__https://github.com/pfirmstone/JGDMS/blob/trunk/qa/harness/policy/defaultsecurephoenix.policy.new__;!!ACWV5N9M2RV99hQ!cVJkwzysH02ZFZWT3rnO_o5viWUCBTm9nnM50QaJ5G9hsg9B3p2naDVYknIrqSNURQ$ Ensure that anything that isn't specifically granted, is switched off, eg agents, management, etc. This file is generated by a policy writing tool, the comments in the policy file, indicate the tests that were run to generate the policy. It's important to note that we aren't using sandboxing, or isolation, we are simply running with privileges we don't require switched off. All network entry points to the JVM require authentication and classes loaded, or data parsed requires authentication first. In my software without SM, authentication isn't required, and there are no defenses against parsing un-trusted data or downloading and loading malicious classes. Although you've made it clear, you think this is a /special case/ and /special loss/ you needn't concern yourself with and that's fine, you've made your point. In future it will not be possible to switch these features off using policy files, so I imagine that OpenJDK must be considering alternative methods to switch off features that may be security sensitive? We already know de-serialization can be switched off from the
