The general IT knowledge of security has come along way in the last 20 years. Even more dramatically when considering the last 10 years.
People are generally aware that unless due care is taken, their computer could be injected with a virus, have personal information stolen from it or even be used to facilitate crime. Major OS Vendors have picked up their game and now are putting in a better attempt to prevent compromises from the OS level. Sure, you still hear the odd story about the latest privilege escalation, but compared to what it use to be…
Network level security has been given most of the attention (and IT budget funding) and is *generally* fairly secure these days. Application level is where most of the major hacks are happening now, but unfortunately, corporate uptake on securing their systems at the Application level hasn’t been as good as it was with the Networks.
Let’s be honesty and not undersell ourselves, securing complex applications is no mean feat. It takes knowledge, planning, lots of time & patience and sometimes out-of-the-box thinking. Thankfully, most modern programming languages and Database Management Systems do the heavy lifting for us. From the security features built into C# and Java to the vastly improved safety net found in SQL engines with fine-grained access control and in-built functions for preventing SQL injection, a lot of the basics have been solved.
This is where the U2 family has a few gaps to be filled. UniBasic needs some inbuilt functions for sanitisation, UniObjects needs some form of access control built around it and UniQuery/RetrieVe prepared statements/stored procedures would be nice.
With the increase push in integrating U2 servers as databases for modern front-ends such as web applications, data sanitisation is going to become a prevalent topic in the community. Built-in functions for UniQuery/RetrieVe, SQL and HTML sanitisation/encoding would be welcome additions to the UniBasic command family. Even better would be some form of prepared statements for the query languages. This make it simpler and easier to obtain better program security.
UniObjects is touted as a standard method of connecting GUI application front-ends to a U2 back-end. However, due to the limited access control supported by UniObjects, it is a dangerous hole in your system to have the required port open for anything other than back-end servers. Take into considering user ‘X’. User ‘X’ has appropriate login credentials for the old green screen system. IT brings out a new Windows GUI application, lets say for reporting, that runs on the user’s machine and uses UniObjects to connect to U2. In the old green screen system, User ‘X’ was limited to set menus and programs to run and could not get access to ECL/TCL. With enough knowledge (and malice), User ‘X’ can now freely use his green screen login credentials to log into the U2 system via UniObjects read/write records directly and even execute raw ECL/TCL commands.
So what exactly is the problem with UniObjects? Quite simply put, it has no fine-grained server-side control of what actions can be done, or commands issued via UniObjects. As long as you can log in, you can get a free pass to the back-end’s data. Let’s take MsSQL as a counter example. You can create views, stored procedures, grant or deny users a suite of privileges to tables and commands. Essentially, UniData needs to be able to have some access control scheme for UniQuery that allow you to define whether the users and read/write records in certain files. Ideally, all read/writes would be done through U2 UniBasic subroutines, with RPC daemon having the ability to have a command ‘white-list’ setup. That way, all data access can be moderated with UniBasic code and the RPC daemon having a white-list that only allows access to calling those subroutines.
All this highlights an issue we need to overcome as a community. The lack of U2 specific security literature. Where is the UniData/UniVerse security manual? Where is the “Top 10 common security mistakes” for U2? Sadly, security does seem to be an afterthought. Sometimes even a ‘neverthought’.
A few years ago I read an interesting article titled Denial of Service via Algorithmic Complexity Attacks. When I started working with UniData, It never crossed my mind that U2 had the same class of vulnerabilities, but it does.
If you develop for a U2 system where you cannot afford for malicious internal/external entities to adversely affect system performance, then I highly suggest you read the above linked paper.
I’ll divide this into 3 sections.
The first place I’ll draw your attention to is the humble hash file at the core of UniData and UniVerse. As you probably know, each record is placed in a group dependant on the hash value of its record ID, along with the modulo and hashing algorithm of the file. Now, there are 2 hashing algorithms that a hashed file can use. Type 0 or ‘GENERAL’ is the default, general use hashing algorithm, whereas Type 1 or ‘SEQ.NUM’ is an alternative you can specify and is designed to handle sequential keys. The hash file is basically a hash table with chaining.
Let’s assume we’re working at the HackMe Ltd company that has made a public website to integrate with their existing backend system, which is UniData driven. It is decided that people can pick their own usernames when signing up. Since these usernames are unique, they have been used as the record ID.
Ever since he was not hired after interviewing at HackMe Ltd, Harry has wanted to show them up. Knowing that they used UniData on the backend from his Interview (and their job ads), he installed UniData and makes some initial guesses at the modulo for their ‘users’ tables and calculates a few usernames sets for different modulus.
Now, by going to their website and taking timings for the “Check username availability” feature, Harry was able to become reasonably sure of the modulo for the file. Setting up his computer to run all night generating keys that hashed to a single group. Setting up his email server to automatically do a wget on the confirmation URL on received emails (hence getting around the “Confirm email address” emails).
The next day he runs a script to sign-up all the usernames gradually over the day. After they have all been signed up, Harry now simply scripts a few “Check username availability” calls for his last username generated to start his Denial of Service attack. Essentially, he has taken the non-matching lookup performance of the hash file from O(1 + k/n) to O(k) (where k is the number of keys and n is the modulo). Even worse than that, because of how level 1 overflows work, it now requires multiple disk reads as well (UniData only I believe). Continual random access to that file that is heavily weighted in one group is O(k^2)
Now, to give you a visual example, I have run a test on my home machine and produced 2 graphs.
CPU: Core Duo T7250 (2.0GHZ)
OS: Vista SP2 (32-bit)
DB: UniData 7.2 PE (Built 3771)
Hash File: Modulo 4013 – Type 0
Pre-generate 2 sets of numbers. One is of sequential keys, the other is of keys chosen because they all hash to a single group. Timings are recorded for the total time in milliseconds for:
- Write null records for all the keys and
- read in all the records.
Separate timings for sequential and chosen keys are taken. The test is repeated for different key counts from 1000 to 59000 in 1000 increments.
First Graph – Sequential key timings by themselves:
Second Graph – Chosen key alongside sequential key timings:
Naturally, timings are rough, but they are accurate enough to paint the picture.
Actually, now that I’ve mentioned painting…
Have you heard of Schlemiel the Painter?
Schlemiel gets a job as a street painter, painting the dotted lines down the middle of the road. On the first day he takes a can of paint out to the road and finishes 300 yards of the road. “That’s pretty good!” says his boss, “you’re a fast worker!” and pays him a kopeck.
The next day Schlemiel only gets 150 yards done. “Well, that’s not nearly as good as yesterday, but you’re still a fast worker. 150 yards is respectable,” and pays him a kopeck.
The next day Schlemiel paints 30 yards of the road. “Only 30!” shouts his boss. “That’s unacceptable! On the first day you did ten times that much work! What’s going on?”
“I can’t help it,” says Schlemiel. “Every day I get farther and farther away from the paint can!”
(Credit: Joel Spolsky, 2001)
When looking at Dynamic arrays in U2, you should see how they can be exactly like a computerised version of Schlemiel the Painter. In fact, a public article on PickWiki pointed this out quite some time ago. UniData is affect more so than UniVerse, in that UniVerse has an internal hint mechanism for attributes. The problem with this is, if an uncontrolled (eg, external) entity has control over the number of items in the dynamic array, you could be vulnerable to a Denial of Service attack. It could even be unintentional.
So, let’s see what all fuss is about. Firstly, a quick recap on the issue with dynamic arrays.
Essentially when doing an operation like “CRT STRING” it has to scan the string character by character counting attribute, multi-value and sub-value marks as it does. If you increment Y or Z (or X in UniData’s case) and do the same operation, it has to re-scan the string from the start all over again. As the number of elements increases, the more noticeable the flaw in this method becomes. In fact, cycling through each element in this manner is an O(k^2) algorithm.
I’ve seen this issue bite and bite hard. It involved 2 slightly broken programs finding just the right (or wrong) timing.
The first program was a record lock monitoring program. It used GETREADU() UniBasic command, after which it looped over every entry and generated a report on all locks over 10 minutes old. This process was automatically scheduled to run at regular intervals. It had been operating for months without issues.
The second program was a once off update program. Basically, it read each record in a large file, locked it then if certain complex conditions were met, it updated an attribute and moved on to the next record. See the problem? It didn’t release a record if it didn’t need updating. The processing was estimated to take about 30 minutes and as it turns out, not many records met the complex conditions.
See the bigger problem now? Yup, that’s right, the dynamic array returned by GETREADU() was astronomical! This resulting in the monitoring program saturating a CPU core. The same core the update program was running on. Uh oh! System performance issues ensured until the culprit was found and dealt with.
So, what do we do about these issues? You want a stable system right? One that is less easy to bring to its knees by malicious users and unfortunate timings of buggy code?
DO NOT use external input as record keys! Place it in attribute 1, build a D-type dictionary and index it if you need, but not use it as the @ID!
A further option would be to have Hash files and their hashing algorithms updated to be able to deal with this type of malicious edge case. Other languages have (take Perl for example) updated their hash tables now to use hashing algorithms to be seeded at run-time. These means you cannot prepare ‘attack’ keys ahead of time and cannot replicate how the hashing works on another computer, since the hash algorithm will be seeded differently. Obviously, this cannot be done exactly the same with Hash files, is they are a persistent data store. It could however be done on each CREATE.FILE. That way, even if a malicious party can determine the modulo of a file, they be able to duplicate it on their system as each file will be seeded differently. Doing this would bring UniData and UniVerse inline with the security improvements made in other modern stacks.
This one is simple. Use REMOVE, don’t use simple FOR loops. Think through your data and were it is being sourced from. Is it from external entities? Is it from internal entities whose behaviour cannot be guaranteed to remain within safe bounds? If the answer to either of those questions is even a ‘Maybe’, stay safe and use REMOVE.