Slashdot Mirror
Secretly Monopolizing the CPU Without Being Root
An anonymous reader writes "This year's
Usenix security symposium
includes a
paper
that implements a "cheat" utility, which allows any non-privileged user to
run his/her program, e.g., like so 'cheat 99% program'
thereby insuring that the programs would get 99% of the CPU
cycles, regardless of the presence of any other applications in the
system, and in some cases (like Linux), in a way that keeps the program
invisible from CPU monitoring tools (like 'top'). The utility exclusively
uses standard interfaces and can be trivially implemented by any
beginner non-privileged programmer. Recent efforts to improve the
support for multimedia applications make systems more susceptible to
the attack.
All prevalent operating systems but Mac OS X are vulnerable, though by
this kerneltrap story,
it appears that the new CFS Linux scheduler attempts to address the
problem that were raised by the paper."
I run several websites off of a single host. If I need to login to do maintenance during peak hours, I'm slowed by Apache and MySQL. This would be a nice utility if it were wrapped into SUDO.
I'd rather you do it wrong, than for me to have to do it at all.
For those harboring poisonous grudges against PDFs, the Googlerised HTML version is here.
The gnome desktop for years has been hiding processes that h0rk the cpu.
Using up 99% of the CPU's easy!
#include
int main(int argc, char *argv[])
{
while (1) {}
return 0;
}
Summation 2
Not quite sure what justifies a paper out of this.
If you check the linux kernel mailing list for Vassili Karpov, you should find test cases that demonstrate this behavior and tools for monitoring actual CPU usage for a variety of platforms, though I notice no mention of any of that in the paper.
See http://www.boblycat.org/~malc/apc/ for the tool and 'invisible CPU hog' test case.
Sanity is a sandbox. I prefer the swings.
Back in my day we called it renice.
Yes, I'm kidding. Please don't post a long reply explaining how renice differs from this cheat thing. It isn't necessary.
If you reply, do so only to what I explicitly wrote. If I didn't write it, don't assume or infer it.
That isn't likely to happen without a change in attitude due to both starting furthur behind and progressing more slowly. The current malware situation looks like bad SF and a morality tale of what happens when you allow really stupid things to happen (eg. letting arbitrary code embedded in images run - hopefully that person was dismissed from Microsoft).
If you reply, do so only to what I explicitly wrote. If I didn't write it, don't assume or infer it.
You gun-toting marxist redneck zealot astroturfers make me sick!
I wasn't aware the schedulers for those systems were so deficient !
In my days (yes, I'm an old fart) - the schedulers had basic principles :
- Voluntary yielding led you to get accounted for the time you spent running.
- You could stay in the interactive queue for only a certain amount of time. After some amount of time had passed (a few secs) you were either bumped to non-interactive if you were running (with longer time slices but lower priority) or removed off the scheduler list for good (if the time spent there was idling). They had a special 'idle but interactive' (not eligible for dispatching) queue for that.
- Scheduling a new task restarted a new time slice
That particular scheduler even had a 3 queue system so that if you got accidentally bumped into the non-interactive queue or if your process was semi-interactive you had a better chance of gaining interactive status again. And they had a 'really' not interactive queue for those CPU hogging processes.
Of course this requires the hardware to have a precise timing feature (something with a granularity that is finer than the process interleaving time slice time and ideally in the magnitude of instruction execution). And this scheduler wasn't using time sampling and time quantums.. (but something more like the OSX timer on demand paradigm).
--Ivan
I don't know. I think retractions would screw with everything else. If you make a boneheaded statement (and I've done it more than once myself), it should stand. Otherwise, everyone who responds to correct your misstatement will look insane, and it'd be hard to metamod, because the comments wouldn't necessarily fit the context anymore, etc.
ad logicam Claiming a proposition is false because it was presented as the conclusion of a fallacious argument.
This is accomplished by sleeping for a fixed amount in between OS clock ticks. The timeline looks like this:
The most rabid believers in American Exceptionalism are the exact same people whose policies are destroying it.
We had a user who insisted on abusing the "nice" command, to run his jobs at a higher priority. Pleading and cajoling didn't work, so we decided to get creative.
We changed nice so that whenever this particular user ran it, it lowered his priority by exactly as much as he was attempting to raise it.
He stopped coming to work soon after that. I suppose he had the last laugh though -- NYIT continued to pay him for another six months.
Thad
I love Mondays. On a Monday, anything is possible.
Why not leave the post but allow a "retracted" tickbox? Thus at least the owner of the comment can effectively say "I was wrong, boneheaded, whatever" without having to post another comment and wait two minutes to do it? and all that shows up it a one-liner under the comment:
This comment has been retracted by its poster
-nB
whois gawk date unzip strip find touch finger mount join nice man top fsck grep eject more yes exit umount sleep dump
According to the paper, the reason Mac OS X is not vulnerable is that it uses one-shot timers scheduled for exactly when the next event needs to occur, rather than periodic "ticks" with a fixed interval between them. The "tickless idle" feature introduced in Linux 2.6.21 (currently only on x86, I believe) takes the same approach, and very possibly makes Linux immune too.
(Ironically, immediately after discussing OSX's ticklessness, the paper mentions that "the Linux 2.6.16 kernel source tree contains 8,997 occurrences of the tick frequency HZ macro, spanning 3,199 files", to illustrate how difficult it is to take a tick-based kernel and make it tickless. But those kernel hackers went and did it anyway.)
The tickless feature isn't yet implemented on all architectures that Linux supports, though. I think AMD64 support for it is supposed to come in 2.6.23, along with the new CFS scheduler.
It works by sleeping at the right point in time. You really hack up the timeslices and decrease the overall efficiency (more context switches), so it's only good if you want to steal cycles where you are not really allowed to.
That'd be fine, or even cool. It'd deflect the inevitable storm of 500 people saying, "Wrong n00b!" and not reading down far enough to see that you admitted it already, and let the whole discussion move on to more productive things.
ad logicam Claiming a proposition is false because it was presented as the conclusion of a fallacious argument.
They took too long to publish this. Linux 2.6.21 (released in April) added support for using one-shot timers instead of a periodic tick, so it avoids the problem like OS X does. In addition to resolving this issue, tickless is important for saving power (because the processor can stay in a low-power state for long enough to get substantial benefits compared to the power cost of starting and stopping) and for virtual hosting (where the combined load of the guest OS scheduler ticks is significant on a system with a large number of idle guests). As a side effect, while the accounting didn't change at that point, the pattern a task has to use to fool the accounting became impossible to guess.
The CFS additionally removes the interactivity boost in favor of giving interactive tasks no extra time but rather just quick access to their available time, which is what they really benefit from.
My mother is a gun-toting marxist redneck zealot astroturfer, you insensitive clod!
The creator of this post (Jacob Smith) hereby releases it, and all of his other posts, into the public domain.
This is an outrage. You cannot 'sue' without lawyerd! What about the required functionality of 'sue --counter' and 'appeal'?!
The only change I can believe in is what I find in my couch cushions.
(reply to self after RTFA)
What 'saved' the Mac OS was its different use of timing triggers. "All" other OS'es use one common steadily ticking clock as a dealer of time slots. This allows the cheat to "skip to the start of the line (queue)" every time it's had its turn.
OTOH, the Mac uses a stack of alarms set to specific points in the future, and polled in order as they occur. So the difference on Mac OS is that there's no skipping the queue, it's rather "there is no queue, we'll call you when it's your turn".
I don't know the details of the OpenBSD scheduler, but it's very likely the same (clock tick) method as used by the rest of the susceptible OS'es.
"Good news, everyone!"
The paper is quite long, so here's a summary (take this with a grain of salt, who wants accurate information should still RTFP):
Most OSes (Linux, Solaris, Windows but not Mac OS X) are tick-based. This means that the kernel is called from hardware periodically (this is the "HZ" value you set in the Linux kernel). Some of them (Linux) simply check which process is running at each tick and compute statistics based on that ("sample-based statistics"). This means that the process running when the tick happens is billed for the entire period of the tick.
Since ticks are typically "long" (typically 1-10 ms on Linux) more than one process may run during this period. In other words, using this approach leads to inaccuracies in the process billing. If all programs "play by the rules" this works quite well on average though.
Next thing: the classic schedulers typically maintain some sort of "priority" value for each process, which decreases whenever the process is running and increases when it's not. This means that a process runs for some time, its priority decreases, and then another process (which hasn't been running for some time) takes over.
You can exploit that by always sleeping when a tick happens and running only in-between ticks. This makes the kernel thinks that your process is never running and give it a high priority. So, when your process wakes up just after a tick happened, it will have a higher priority than most other processes and be given the CPU. If it goes to sleep again just before the next tick, its priority will not be decreased. Your process will (almost) always run when it wants to and the kernel will think that it's (almost) never running and keep its priority high. You win!
Another aspect is that modern kernels (at least Linux and Windows) distinguish between "interactive" (e.g. media players) and "non-interactive" processes. They do so by looking how many times a process goes to sleep voluntarily. An interactive program (such as a media player) will have many voluntary sleeps (e.g. inbetween displaying frames) while a non-interactive program (e.g. a compiler or some number crunching program) will likely never go to sleep voluntarily. The scheduler gives the interactive programs an additional priority boost.
Since the cheating programs go to sleep very often (at every tick) the kernel thinks they're "very interactive", which makes the situation worse.
Some of the analyzed OSes - even if tick-based - do not use sample-based statistics in the kernel but they do use sample-based statistics for scheduling decisions. So the kernel sees that a process is taking more CPU than it should but it will still keep on scheduling it.
Mac OS X is not affected because it has a tickless kernel (e.g. without periodic interrupts). Because of that sample-based statistics don't work and it has to use accurate statistics, which make it unaffected by the bug.
This bug can be exploited to (at least)
- get more CPU than you're supposed to
- hinder other programs in their normal work
- hide malicious programs (such as rootkits) which do work in the background
Here's a list with the OSes (this USED TO BE a nicely formatted table, but the darned Slashdot "lameness filter" forced me to remove much of the nice lines and the "ecode" tag collapses whitespace).
OS, Process statistics, Scheduler decisions, Interactive/non-interactive decision, Affected
Linux, sample, sample, yes, yes
Solaris, accurate, sample, ?, yes
FreeBSD 4BSD, ?, sample, no?, yes
FreeBSD ULE, ?, sample, yes, yes
Windows, accurate, sample, yes, yes
Mac OS X, accurate, accurate, not needed?, yes
I guess that Mac OS X doesn't need a interactive/non-interactive distinction because of its different (tickless) approach. I assume that interactive applications can (implicitly or explicitly) can be recognized as such in a different way. Does anyone have more information on that?
How does tickless Linux compare? What abo
A very simple patch is to issue RDTSC instructions at process restart and blocking syscall to count the cycles actually used. That way the extensive tick-code doesn't need to be modified.
Besides the syntax comment the other poster said, it could've also been that the school implemented per-user process limits on the machine. Linux has had this capability for years and years; most people just don't bother setting it, but universities hosting machines for programming students pretty much have to set it for exactly this sort of thing, whether it be accidental or malicious.
If it's for-profit but free, you're not the customer -- you're the product (e.g., the Slashdot Beta's "audience").
What a scary, scary thought...
Chris Torek gave a presentation at UseNIX about how a constant quantum could result in a process having its CPU usage unaccounted.
His solution was to use a randomized quantum. Not unique per process, but randomized when the kernel starts running each process. That gave you a better accounting of the CPU time (statistics, doncha know :)), but also made this kind of attach much, much harder.
I'm somewhat disappointed that I did not see Chris and Steven's paper referenced in this one. (I believe that the title of that paper was "Randomized Sampling Clock for CPU Utilization Estimation and Code Profiling," for those who care to find it.)
I'm assuming that we're saying that this application can get 99% of the time-slices on an otherwise occupied system, starving other tasks for resources.
We already have that. They're called McAfee Automatic Update and Windows Automatic Update.
God dammit, I hate those things. I turn on my office computer in the morning, and just let it sit for ten minutes because it's otherwise useless. (I turned-off Windows Automatic Update, but McAfee was more than happy to fill its shoes in needless resource hogging.)
Does it make you happy you're so strange?
You're missing the point here. Because the CPU accounting is off it's possible to do a QoS attack on a box rather than a DoS, that's virtually impossible to detect as the end user. From his or her standpoint, the system will be sluggish, but because of the way the attack works various random processes will seem to be taking up all that extra slack so that most likely no one process will appear to be hogging the CPU.
There's also the possibility when combined with a worm or rootkit, as well as a bot net to setup a difficult to detect distributed computing environment to perform massive computations in short amounts of time.
Like any concealment based vulnerability this is just a tool to be combined with others for a complete attack, but a serious issue nonetheless.
Curiosity was framed, Ignorance killed the cat.
Do you think this might be related to that incident where thousands of English teachers all burst into flames moments after the first SMS-enabled phone was released?
The second one is obviously the better one. I think this is basically what the CFS does. (the following is my understanding. It may be wrong) For processes it figures out what amount of time each process should have (based on the the number of processes. It tracks how much time each process is owed (in the case of 5 processes each deserves 1/5 of the total processor time). It subtracts the time used on each scheduler event (clock tick or voluntary yield.) Each clock tick the scheduler transfers control to the process owed the most time (but there is a minimum number of clock ticks before mandatory switching to prevent cache thrashing.) I presume voluntary yielding has some form of impact on the time owed amount, or else idling processes would always stay at the top of the list. Obviously there is more complications, such as nice levels, and everything.
The big thing is that is (AFAIK) tracking the exact amount of time used by each process. The only proper way to do that is to do it at both each clock tick, and each volentary yield.
One other rant I have is the naming of the so called O(1) scheduler. That scheduler was apparently O(1) but only because there is a limit to the maximum number of processes. In nearly every case it is possible to construct on O(1) algorithm if the maximum number of possibilities is known in advance. Technically the algorithm's timing was some function of the maximum number of processes. Since the maximum number of processes is a compile time constant, the algorithm is constant-time.
Stylish sheet to fix many problems in Slashdot's D3: https://gist.github.com/801524
A DoS attack is an extreme form of QoS. If you perform a QoS attack on someone their performance is reduced, but the system is still usable, where as in a DoS the goal is to make the system totally unusable. In some ways a QoS is even more effective than a DoS because it's more subtle and causes more frustration. If for instance a website gets DoSed the owner is upset and will try to get someone to investigate and shutdown if possible whoever is DoSing them, and the users simply cannot connect to the service and go somewhere else. If, on the other hand you QoS attack a server, the owner will be frustrated because performance is poor, but they will have to spend a good bit of time trying to track down WHY exactly the performance is poor, but, more importantly the users connecting to the service will have a very poor experience, and that hurts the servers owner. A user is willing to cut someone slack if the server goes down, but they're much less forgiving when the servers performance is just poor.
I think a mistake you are making a mistake in understanding, the process isn't invisible or even hard to spot, but rather the resources being used are, a simple ps would still show the process.The process is not invisible, you are correct, but it is hard to spot. If the malicious program is named something innocuous such as srvchost.exe (check your process list in Windows, there's a ton of the suckers), or maybe httpd in linux, and the user attempts to figure out what's causing slow downs on their system, they will be looking at anywhere but these processes because they will be showing 0% CPU utilization. Also, as I said, this is only part of a proper attack, this combined with some other exploit that hides the presence of the process will be even more confusing to the user because this attack actually re-allocates the used timeslices for the process to other random processes, so to the user it looks like the entire system is just using way more CPU time than normal. Of course, if you have a root kit you can perform this re-allocation at the kernel level, but part of the point of this exploit is that it's 100% userland so has a much smaller barrier to entry.
Curiosity was framed, Ignorance killed the cat.
Nothing new here.
I remember seeing this done on the VAX/VMS mainframe back in 1987. In that environment, it simply meant that you kept track of your timeslice and voluntarily gave it up before the scheduler took it away from you. That meant you got put at the top of the run queue, and unless someone else was doing the same thing, you were the next program to run. Voila... 99% CPU for you!
Of course, ordinary users were given a limited amount of CPU time (as well as connect time, disk space, etc), so for the ordinary student, this just meant they used it up in a day or two instead of having a whole month. But then again, for class accounts, they could usually beg for more.
Under unix variants, one could do the same by implementing cpu quotas at the user level. I've seen network packet quotas, and I'm sure someone out there has done cpu quotas along the same lines.
May contain traces of nut.
Made from the freshest electrons.
Absolutely. In fact I think it should go half a step further. In the interest of civility, using this feature should hide the message from casual viewing. But a single click will still bring up the original so that you can't use slashdot to be a complete ass and then censor yourself after the damage is done :)
Of course not. It shows that OS research work is likely to be done on a Unix of some sort where the source code is available for anaylsis
TFA points out that Windows is just as vulnerable to these cheats as BSD, Linux and Solaris. The cheat works by releasing the CPU just before the end of a time tick there by allowing the whole tick to be charged to whatever task gets the rest of the tick. Windows, like Solaris, has accurate job accounting information available, but choses not to use it for scheduling. In addition, like the Linux 2.6 kernel, Windows will actually artificially raise the priority of a cheating task under the misaprehension that the job is interactive.
You can't see ANYTHING from a car, You've got to get out of the goddamned contraption and walk...Edward Abbey
Not exactly. This is a technique that will, in prinicple, work with any scheduler that prioritizes tasks on the basis of time ticks previously used by the task. That turns out to be most of them. The technique does not require being an I/O driver, other special task, or having unusual user priviliges.
So yes, it IS news.
You can't see ANYTHING from a car, You've got to get out of the goddamned contraption and walk...Edward Abbey