Slashdot Mirror
Torvalds on the Microkernel Debate
diegocgteleline.es writes "Linus Torvalds has chimed in on the recently flamed-up (again) micro vs monolithic kernel, but this time with an interesting and unexpected point of view. From the article: 'The real issue, and it's really fundamental, is the issue of sharing address spaces. Nothing else really matters. Everything else ends up flowing from that fundamental question: do you share the address space with the caller or put in slightly different terms: can the callee look at and change the callers state as if it were its own (and the other way around)?'"
This my favorite Linus quote from that whole thread:
"In the UNIX world, we're very used to the notion of having
many small programs that do one thing, and do it well. And
then connecting those programs with pipes, and solving
often quite complicated problems with simple and independent
building blocks. And this is considered good programming.
That's the microkernel approach. It's undeniably a really
good approach, and it makes it easy to do some complex
things using a few basic building blocks. I'm not arguing
against it at all."
He basically continues his previous argument that monolithic kernels are more efficient and easier to implement. Microkernels may seem simpler, but they have complexity in implementing all but the simple tasks. Microkernels have a more marketable name. "Microkernel" just sounds more advanced than "monolithic". He finishes off with the observation that the term "hybrid kernel" is a trick to grab marketing buzz from the microkernel side of things.
My other first post is car post.
pfff, Linus, what would he know?
Philosophy.
Linus FTFA:
"The fundamental result of access space separation is that you can't share data structures. That means that you can't share locking, it means that you must copy any shared data, and that in turn means that you have a much harder time handling coherency. All your algorithms basically end up being distributed algorithms.
And anybody who tells you that distributed algorithms are "simpler" is just so full of sh*t that it's not even funny.
Microkernels are much harder to write and maintain exactly because of this issue. You can do simple things easily - and in particular, you can do things where the information only passes in one direction quite easily, but anythign else is much much harder, because there is no "shared state" (by design). And in the absense of shared state, you have a hell of a lot of problems trying to make any decision that spans more than one entity in the system.
And I'm not just saying that. This is a fact. It's a fact that has been shown in practice over and over again, not just in kernels. But it's been shown in operating systems too - and not just once. The whole "microkernels are simpler" argument is just bull, and it is clearly shown to be bull by the fact that whenever you compare the speed of development of a microkernel and a traditional kernel, the traditional kernel wins. By a huge amount, too.
The whole argument that microkernels are somehow "more secure" or "more stable" is also total crap. The fact that each individual piece is simple and secure does not make the aggregate either simple or secure."
The whole discussion of micro-kernel vs monolithic kernel is totally pointless. All popular OS kernels are monolithic. We can get back to the debate when we have a working fast microkernel in the market that is actually competitive.
Linus is a pragmatist. He didn't write Linux for academic purpose. He wanted it to work.
But you can always prove him wrong by showing him the code, and I bet he'd be glad to accept he was wrong.
Quick slashdoteffect there, that forum is already down. Anyhow.. mirror: http://www.mirrordot.org/stories/3f6b22ec7a7cffcf2 847b92cd5dec7e7/index.html
http://pastebin.ca/54695
"The whole "microkernels are simpler" argument is just bull, and it is clearly shown to be bull by the fact that whenever you compare the speed of development of a microkernel and a traditional kernel,the traditional kernel wins. By a huge amount, too. He goes on to say, "It's ludicrous how microkernel proponents claim that their system is "simpler" than a traditional kernel. It's not. It's much much more complicated, exactly because of the barriers that it has raised between data structures." He states that the most fundamental issue is the sharing of address spaces. "Nothing else really matters. Everything else ends up flowing from that fundamental question: do you share the address space with the caller, or put in slightly different terms: can the callee look at and change the callers state as if it were its own?"
Insinct is stronger than Upbringing - Irish Proverb
I think Linus hit the spot by pointing out that the future of home computing is going to to focus on parallel processing - it's 2006 and all my computers, including my LAPTOP, are dual-processor systems.
By 2010 I suspect at least desktops are 4-CPU systems and as the numbers of cores increase one of the large drawbacks of microkernels raises it's ugly head: microkernels turn simple locking algorithms into distributed computing-style algorithms.
Every game developer tells us how difficult it is to write multi-threaded code for even our monolithic operating systems (Windows, Linux, OSX). In microkernels you constantly have to worry how to share data with other threads as you can't trust them to give even correct pointers! If you would explicitly trust them, then a single failure at any driver or module would bring down the whole system - just like in monolithic kernels but with a performance penalty that scales nicely with the number of cores. What's even worse is that at a multi-core environment you'll have to be very, very careful when designing and implementing the distribution algorithms or a simple user-space program could easily crash the system or gain superuser privileges.
Capitalization is the difference between "Helping your uncle jack off a horse" and "Helping your uncle Jack off a horse"
That would be monolithic+marketing.
MacOS X is no microkernel system. It does have Mach, sure. Mach is arguably not a microkernel by today's standards, and in any case MacOS X has a full BSD kernel bolted onto the Mach kernel. Mach and BSD are sharing address space. In other words, it's not a microkernel.
NT is the same way.
I don't know all that much about NetWare, but I'd never before heard anyone claim it to be a microkernel. It's not terribly popular anyway. (it was, but back then I'm sure it wasn't a microkernel system) ReactOS isn't much yet. BeOS died for unrelated reasons, so we really can't judge.
Monolithic kernels can be very modular. Microkernels can get really convoluted as the developers struggle with the stupid restrictions.
Individual pieces aren't really any simpler either. In fact, if you want your kernel to scale, to work well with lots of processes, you are going to run into a simple problem: multitasking.
Consider a filesystem driver in a monolithic kernel. If a dozen or so processes are all doing filesystem calls, then, assuming proper locking and in-kernel pre-emption, there's no problem - each process that executes the call enters kernel mode and starts executing the relevant kernel code immediately. If you have a multiprocessor machine, they could even be executing the calls simultaneously. If the processes have different priorities, those priorities will affect the CPU time they get when processing the call too, just as they should.
Now consider a microkernel. The filesystem driver is a separate server process. Executing a system call means sending a message to that server and waiting for an answer. Now, what happens if the server is already executing another call ? The calling process blocks, possibly for a long time if there's lots of other requests queued up. This is an especially fun situation if the calling process has a higher priority than some CPU-consuming process, which in turn has a higher priority than the filesystem server. But, even if there are no other queued requests, and the server is ready and waiting, there's no guarantee that it will be scheduled for execution next, so latencies will be higher on average than on a monolithic kernel even in the best case.
Sure, there are ways around this. The server could be multi-threaded, for example. But how many threads should it spawn ? And how much system resources are they going to waste ? A monolithic kernel has none of these problems.
I don't know if a microkernel is better than monolithic kernel, but it sure isn't simpler - not if you want performance or scalability from it, but if you don't, then a monolithic kernel can be made pretty simple too...
Forget magic. Any technology distinguishable from divine power is insufficiently advanced.
By today's standards, Mach is not much of a microkernel. Mach has been disowned by microkernel proponents because it was so big and nasty.
MacOS has that sharing address space with the monolithic BSD kernel. So a semi-microkernel and a monolithic kernel are firmly bolted together. That's only a microkernel if your degree is in marketing.
"You can do simple things easily - and in particular, you can do things where the information only passes in one direction quite easily, but anythign else is much much harder, because there is no "shared state" (by design). And in the absense of shared state, you have a hell of a lot of problems trying to make any decision that spans more than one entity in the system."
I think you're looking at this the wrong way around.
There has been a lot of research into this over the past 40 years, ever since Dijkstra first talked about coordination on a really big scale in the THE operating system. Any decent CS program has a class on distributed programming. Any decent SW architect can break down these different parts of the OS into weakly-connected pieces that communicate via a message passing interface (check out this comment by a guy talking about how Dragonfly BSD does this).
It's obvious that breaking something like your process dispatcher into a set of processes or threads is silly, but that can be easily separated from the core context switcher. Most device driver bottom halves live fine as a userland process (each with a message-passing interface to their top-halves).
If you're compiling for an embedded system, I'm sure you could even entirely remove the interface via some #define magic; only debug designs could actually have things in separate address spaces.
The point I'm trying to make is: yes, you can access these fancy data structures inside the same address space, but you still have to serialize the access, otherwise your kernel could get into a strange state. If you mapped out the state diagram of your kernel, you'd want the transistions to be explicit and synchronized.
Once you introduce the abstraction that does this, how much harder is it to make that work between processes as well as between threads in the kernel? How much of a benefit do you gain by not having random poorly-written chunks pissing over memory?
How about security benefits from state-machine breakdowns being controlled and sectioned off from the rest of the machine? A buffer overflow is just a clever way of breaking a state diagram and adding your own state where you have control over the IP; by being in a separate address space, that poorly written module can't interact with the rest of the system to give elevated privileges for the attacker (unless, of course, they find flaws in more of the state machines and can chain them all together, which is highly unlikely!).
Clearly there is a security benefit as much as there is a consistency benefit. Provably correct systems will always be better.
--
Internet Explorer (n): Another bug -- that is, a feature that can't be turned off -- in Windows.
You are forgiven for being wrong, but not for spouting off nonsense despite knowing that you don't know what you're talking about, apparently applying the principal "if my argument involves M$ doing the wrong thing, it must be right".
While neither NT nor Mac OS X are true microkernels, the architecture of both is strongly inspired by microkernel ideas. Like Linus, the developers of these kernels recognized the practical difficulties involved in making full-on microkernels work, but unlike Linus, instead of throwing in the towel completely and doing full-on monolithic kernels, they created cleanly seperated layers interacting via well-defined interfaces whenever they practically could.
If you talk to kernel programmers, most will express a high degree of respect for the NT kernel, which is based on the DEC VMS kernel. It mostly the poor design of systems that sit on top of the kernel that has earned Windows its reputation.
Hi folks,
I worked two years for a society that was developing its own micro-kernel system, for embedded targets. I was involved in system programing and adaptation of the whole compiler tools, based on GCC chain.
Linus is right: basic problem is address space sharing, and if you want to implement memory protection, you rapidly falls into address space fragmentation problem.
The main advantage of the system I worked on wasn't really its micro-kernel architecture, but the fact that its design allowed to suppress most of glue code that is needed between a C++ program and a more classic system.
In my opinion, micro-kernel architecture has the same advantage and drawbacks that so-called "object-oriented" programing scheme : it is somewhat intellectually seducive for presentations but it is just a tool.
It would certainly be intersting for Linux to provide the dynamic link management specificities of a micro-kernel system, for instance to allow someone to quickly modify IP stack for its own purpose, but should the whole system being design that way ? I am not sure.
If you want to have an idea of the problem encountered with programing for these systems, one can look at the history of the AmigaOS, which have a design very close to a micro-kernel one.
abstraction and state isolation considered harmful
That's a pretty big assumption. Or rather, you have basically taken all the hard parts of doing shared code and said "Let's hope someone else already solved this for us".
Sooooo, it's easy to have someone else handle the multi-process bits in a monolithic design. But when it comes to writing services for microkernels suddenly everyone is an idiot?
Besides, as Linus pointed out, when data is going one way microkernels are easy. And in the case of file systems that is really the case. Sure multiple processes can access it at once, but the time scale on handling the incoming signals is extremely fast compared to waiting for data from disk. Only a really, *really* incompetent idiot would write such a server which blocked until the read was finished.
Name: Linus Torvalds (torvalds AT osdl.org) 5/9/06
___________________
_Arthur (Arthur_ AT sympatico.ca) on 5/9/06 wrote:
I found that distinction between microkernels and "monolithic" kernels useful: With microkernels, when you call a system service, a "message" is generated to be handled by the kernel *task*, to be dispatched to the proper handler (task). There is likely to be at least 2 levels of task-switching (and ring-level switching) in a microkernel call.
___________________
I don't think you should focus on implementation details.
For example, the task-switching could be basically hidden by hardware, and a "ukernel task switch" is not necessarily the same as a traditional task switch, because you may have things - hardware or software conventions - that basically might turn it into something that acts more like a normal subroutine call.
To make a stupid analogy: a function call is certainly "more expensive" than a straight jump (because the function call implies the setup for returning, and the return itself). But you can optimize certain function calls into plain jumps - and it's such a common optimization that it has a name of its own ("tailcall conversion").
In a similar manner, those task switches for the system call have very specific semantics, so it's possible to do them as less than "real" task-switches.
So I wouldn't focus on them, since they aren't necessarily even the biggest performance problem of an ukernel.
The real issue, and it's really fundamental, is the issue of sharing address spaces. Nothing else really matters. Everything else ends up flowing from that fundamental question: do you share the address space with the caller, or put in slightly different terms: can the callee look at and change the callers state as if it were its own (and the other way around)?
Even for a monolithic kernel, the answer is a very emphatic no when you cross from user space into kernel space. Obviously the user space program cannot change kernel state, but it is equally true that the kernel cannot just consider user space to be equivalent to its own data structures (it might use the exact same physical instructions, but it cannot trust the user pointers, which means that in practice, they are totally different things from kernel pointers).
That's another example of where "implementation" doesn't much matter, this time in the reverse sense. When a kernel accesses user space, the actual implementation of that - depending on hw concepts and implementation - may be exactly the same as when it accesses its own data structures: a normal "load" or "store". But despite that identical low-level implementation, there are high-level issues that radically differ.
And that separation of "access space" is a really big deal. I say "access space", because it really is something conceptually different from "address space". The two parts may even "share" the address space (in a monolithic kernel they normally do), and that has huge advantages (no TLB issues etc), but there are issues that means that you end up having protection differences or simply semantic differences between the accesses.
(Where one common example of "semantic" difference might be that one "access space" might take a page fault, while another one is guaranteed to be pinned down - this has some really huge issues for locking around the access, and for dead-lock avoidance etc etc).
So in a traditional kernel, you usually would share the address space, but you'd have protection issues and some semantic differences that mean that the kernel and user space can't access each other freely. And that makes for some really big issues, but a traditional kernel very much tries to minimize them. And most importantly, a traditional kernel shares the access space across all the basic system calls, so that user/kernel difference is the only access space boundary.
Now, the real problem with split acce
Here is some good readign amterial, maybe people should read and _understand_ it before posting on the subject..
i tem.5d61c1d591162e4b0ef1bd108bcd45f3/index.jsp?&pN ame=computer_level1_article&TheCat=1005&path=compu ter/homepage/0506&file=cover1.xml&xsl=article.xsl&
0 ,1144,0131429388,00.html
This does not mean you have to agree with the guy.
http://www.computer.org/portal/site/computer/menu
http://vig.prenhall.com/catalog/academic/product/
Because it's based on a huge, monstrous monolithic microkernel.
Anagram("United States of America") == "Dine out, taste a Mac, fries"
Distributed algorithms are of course difficult to implement with a f***ed up language like C.
Here, it seems, the means justify the ends. Linus basically says "I won't take any challenges".
Linus tells me that we can never write a proper scalable OS for a NUMA machine, or a modular system that can serve well to parallel I/O systems and the like. I highly disagree.
Because these things are not pipe dreams, they have been done. IBM guys have made amazingly abstract and modular OS stuff and they've been using them for years, so I think it's rather pathetic to say that there is only one true path to OS implementation. Why not admit that it is the only path that you have any experience in?
--exa--
The analogy of centralisation vs. local autonomy is not totally accurate either. Both the monolithic and the microkernel are centralized, except that in the first case there a large beaurocratic structure and in the second case it just a dictator and a couple of "advisors". If the dictator or the king is chosen well, the system will be more predictable and will work much better. If case of the large beaurocratic system, if some of its members get corrupted [and they will because there are so many of them] the whole system will fail. It is like saying that a small bug in the mouse driver will freeze and crash the system with a monolithic kernel. Good thing if the system was only running Doom at the time and not controlling a reactor, or administering a drug. If the same happens in the microkernel system, the kernel will reload the driver, raise an alarm, or in general -- be able to take the system to a predictable predetermined state. Going back to the analogy is it is like having the dictator execute a corrupted staff member and replace him immediately.
Andy likes microkernels because they force you to do that. Time spent on design leads to insight, which may well point to better and cleaner ways to do the task you originally set out to acomplish.
Linus hates microkernels because they force you to do that. Time spent on design is time lost getting working code out the door, and working code will give you experience that will point to better and cleaner ways to do the task you originally set out to acomplish.
Individual pieces aren't really any simpler either. In fact, if you want your kernel to scale, to work well with lots of processes, you are going to run into a simple problem: multitasking.
This is very true.
Consider a filesystem driver in a monolithic kernel. If a dozen or so processes are all doing filesystem calls, then, assuming proper locking and in-kernel pre-emption, there's no problem - each process that executes the call enters kernel mode and starts executing the relevant kernel code immediately.
OK, here's where things start getting a little tricky. The whole locking setup in a monolithic kernel is pretty tricky. Early multi-processor kernels often took the course of "one big lock" at the top of the call stack - essentially only one process could be executing in the kernel. Why? Because all that "proper locking" is tricky. Took years to get this working right. Of course it's done now in Linux so you can take advantage of it, but it wasn't easy.
Now consider a microkernel. The filesystem driver is a separate server process. Executing a system call means sending a message to that server and waiting for an answer.
OK, now here, you're kind of running off the rails. What is a "message"? There is no magical processor construct called a "message" - it's something that the OS provides. How messages are implemented can vary quite a bit. What you're thinking of is a messaging system ala sockets - that is the message would be placed onto a queue and then a process switch would happen sometime and the server on the other end would read messages out of the queue and do something. That's how microkernels are usually presented conceptually so it tends to get stuck in peoples' heads.
However, messages can be implemented in other ways. For example, you could make a message be more like a procedure call - you create a new stack, swap your address table around, and then jump into the function in the "server". No need to instantiate threads in the "server" anymore than there is a need to instantiate threads within a monolithic kernel. The server would essentially share the thread of the caller. I've worked on microkernel architectures that were implemented just this way.
If the number of data structures that you can directly access is smaller, the amount of locking that you have to take into account is smaller. Modularity and protection makes most people's tasks easier.
Many of the arguments made for monolithic kernels are similar to the arguments you used to hear from Mac programmers who didn't want to admit that protected memory and multi-tasking were good things. Mac programmers liked to (as I used to say) "look in each other's underware". Programs rummaged about through system data structures and other apps data structures sometimes, changing things where they felt like it. This can be pretty fun sometimes and you can do some really spiffy things. However, set one byte the wrong way and the whole system comes crashing down.
The only way the monolithic vs microkernel debate will go away is if CPUs provide a better way of sharing resources between modules.
One solution to the problem is to use memory maps. Right now each process has its own address space, and that creates lots of problems. It would have been much better if each module had its own memory map, ala virtual memory, so as that the degree of sharing was defined by the O/S. Two modules could then see each other as if they belong to the same address space, but other modules would be inaccessible. In other words, each module should have its own unique view of the memory.
Of course the above is hard to implement, so there is another solution: the ring protection scheme of 80x86 should move down to paging level. Each page shall have its own ring number for read, write, and execute access. Code in page A could access code/data in page B only if the ring number of A is less than or equal to the ring number of B. That's a very easy to implement solution that would greatly enhance modularity of operating systems.
A third solution is to provide implicit segmentation. Right now 80x86 has an explicit segmentation model that forces inter-segment addresses to be 48 bits wide on 32-bit machines (32 bits for the target address and 16 bits for the segment id). The implicit segmentation model is to use a 32-bit flat addressing mode but load the segment from a table indexed by the destination address, as it is done with virtual memory. Each segment shall have a base address and a limit, as it is right now. If a 32-bit address falls within the current segment, then the instruction is executed, otherwise a new segment is loaded from the address and a security check is performed. This is also a very easy to implement solution that would provide better modularization of code without the problems associated with monolithic kernels.
There are various technical solutions that can be supported at CPU level that are not very complex and do not impose a big performance hit. These solutions must be adopted by CPU manufacturers if software is to be improved.
1. From AST (I'd assume you know who he is since you are interested in Linus/microkernel debate): http://www.cs.vu.nl/~ast/brown/followup/ Read the section "Microkernels Revisited":
I can't resist saying a few words about microkernels. A microkernel is a very small kernel. If the file system runs inside the kernel, it is NOT a microkernel. The microkernel should handle low-level process management, scheduling, interprocess communication, interrupt handling, and the basics of memory management and little else. ... Microsoft claimed that Windows NT 3.51 was a microkernel. It wasn't. It wasn't even close. Even they dropped the claim with NT 4.0.
2. From Windows Internals, the 4th edition, published by Microsoft Press. Page 36: Windows is similar to most Unix systems in that it's a monolithic operating system in the sense that the bulk of the operating system and device driver code shares the same kernel-mode protected memory space. Can we stop claiming Windows has a microkernel now?
That's interesting because those are exactly my thoughts every time I hear the arguments people use to defend microkernels: Java is to microkernels as C/C++ is to monolithic kernels.
Linus Torvalds summed it well when he mentioned that microkernels are simpler only when data flow goes in one direction only. It's very hard to get a function to fill a complicated data structure for you if you cannot work with pointers. Passing a reference will do only for simple structures, it will not work if there are structures within structures, it is very hard to do if the called function must itself pass some subset of that structure to another function. And for operating systems where one must contend with multiple access and locking, it's almost impossible to do without a performance penalty.
Let's face it: pointer manipulation is necessary because there are real life problems that are more complex than textbook examples. If there weren't, inventing the C language wouldn't have been necessary, we could have stuck with Fortran all along.
Well maybe that's how *you* would design *your* Microkernel. And yes, it would suck.
The way I would design the filesystem driver, would be to accept a request, add it to a queue of pending requests to serve. If there are no initiated requests, find the request that can most efficiently be served based upon your preferred policy (closest seek time, for example, or first come first serve, your choice), and initiate that request. Add some smarts for multiple devices, so multiple requests can be initiated at the same time to different devices. When data comes back, answer the requesting process with their data. Rather than sitting around blocking on a request, go grab more requests from other processes and queue them up. No need to block. When an initiated request comes back, send back the data to the requesting process, and everyone's happy. Just because things are separated out into different processes, doesn't mean that they can't do some asynchronous juggling to be efficient. Add multi-threading, and the coding becomes a bit easier; but multi-threading isn't necessary to rely upon to have this work well.
I'm pretty sure the monolithic kernels do things somewhat similarly; build a request queue, service that queue. They could also block until they're done other requests, but that would be bad design. Don't assume a Microkernel Filesystem server has to suffer from similarly bad design.
Love many, trust a few, do harm to none.
Only a really, *really* incompetent idiot would write such a server which blocked until the read was finished.
This sounds like a veiled reference to something; would you care to name it?
"Sometimes the truth is stupid." - Lawrence, creator of Prime Intellect
While neither NT nor Mac OS X are true microkernels, the architecture of both is strongly inspired by microkernel ideas.
What exactly does "inspired" mean in this case? I am "inspired" by John Holmes but that doesn't mean I have a 12" cock does it?
If you talk to kernel programmers, most will express a high degree of respect for the NT kernel, which is based on the DEC VMS kernel. It mostly the poor design of systems that sit on top of the kernel that has earned Windows its reputation.
So, did VMS have a graphics subsystem in the kernel as well? Also can you provide some examples of kernel experts praising the NT kernel for its microkernel properties? Thanks in advance.
I see you are also a fan of monolithic posts.
This micro-post shows a division into seperable units.
Using message passing, I can efficiently communicate this to you.
Note that other readers may be reading different sections of my post while you read this one.
This section of my post never has to access internal structures of the other sections. In fact, I could have written each section in any order. Feel free to reorder them yourself.
Intron: the portion of DNA which expresses nothing useful.
I worked on a commercial microkernel OS.
The learning curve was very steep. New developers took at least half a year to be productive. A number of people never became productive and had to be fired.
Linux is really clean and tidy compared to that. Even BSD is clean and tidy compared to that microkernel OS.
Separated components tend to get complex interactions. Sharing data can be very awkward, even if you are co-located.