I summarized my understanding of Linux systems(github.com) |
I summarized my understanding of Linux systems(github.com) |
¹ https://books.google.de/books/about/The_Design_of_the_UNIX_O...
I've been a primary linux user for a couple decades now but I'm not too keen on digging into kernel hacking but love details like the OPs post.
procfs is what's used by pretty much all tools that do something with processes, like ps, top etc.
The everything is a file philosophy becomes much more clear then. Even low level syscalls and their structs are offered by the kernel as file paths so you can interact with them by parsing those files as a struct, without actually needing to use kernel headers for compilation.
eBPF and its bytecode VM are a little over the top, but are essential to known about in the upcoming years cause a lot of tools is moving towards using their own bpf modules.
To be honest, everything is a file is kind of a lie in unix. /proc and /sys are pretty much plan9 inspiration.
The main difference is that plan9 uses read and write for everything, whereas Linux and BSD uses ioctls on file descriptors for everything.
For example, you can't set the baudrate of a serial port by writing it to some /proc node.
It feels like an elaborate mechanism to draw something wrong in the hopes people will correct it.
To me, a block diagram might show [CPU Scheduler], [Virtual Device Manager], [VFS Manager], [Memory Manager], [Interrupt Handlers], etc...
Of course, my knowledge of Linux internals is limited and perhaps it has a separation of the concept of PID and process where there is a literal dependency.
The only use I've had of this kind of documentation is that the process of writing it, made me understand it better. Basically write-only documentation.
I would call myself a Linux expert, and while I can kinda see what you mean with this diagram, it would not have been useful to me back before I was an expert.
If you're curious about the details in this case, ScotXW confuses EGL and OpenGL, the arrows aren't quite right, and the labels aren't quite right either (DRM is labeled "hardware-specific" but KMS isn't? The label for "hardware specific Userspace interface to hardware specific direct rendering manager" is quite weird), and some important boxes are flat out missing. It's nitpicking for sure, but when the diagram goes out of its way to add extremely weird details, it demands nitpicking.
Nobody in the Linux graphics community would draw the diagram like this.
1. User space
2. Kernel
3. Hardware/network
The kernel protects users from hardware, and hardware from users.
Description from the book's website:
"The Linux Programming Interface (TLPI) describes system programming on Linux and UNIX.
TLPI is both a guide and reference book for system programming:
If you are new to system programming, you can read TLPI linearly as an introductory guide: each chapter builds on concepts presented in earlier chapters, with forward references kept to a minimum. Most chapters conclude with a set of exercises intended to consolidate the reader's understanding of the topics covered in the chapter.
If you are an experienced system programmer, TLPI provides a comprehensive reference that you can consult for details of nearly the entire Linux and UNIX (i.e., POSIX) system programming interface. To support this use, the book is thoroughly cross referenced and has an extensive index."
open() read() write() close()
Such a theoretical linux-like or unix-like OS would assume quite literally that "everything is a file" -- including the ability to perform all other syscall/API calls/functions via special system files, probably located in /proc and/or /sys and/or other special directories, as other posters have previously alluded to...
Also, these 4 syscalls could theoretically be combined into a single syscall -- something like (I'll use a Pascal-like syntax here because it will read easier):
FileHandleOrResult = OpenOrReadOrWriteOrClose(FileHandle: Integer; Mode: Integer; pData: Pointer; pDataLen: Integer);
if Mode = 1 then open();
if Mode = 2 then read();
if Mode = 3 then write();
if Mode = 4 then close();
FileHandle is the handle for the file IF we have one; that's for read() write() and close() -- for open() it could be -1, or whatever...
Mode is the mode, as previously enumerated.
pData is a pointer to a pre-allocated data buffer, the data to read or write, or the full filename when opening...
(And of course, the OS could overwrite it with text strings of any error messages that occur... if errors should occur...)
pDataLen is the size of that buffer in bytes.
When the Mode is open(), pData contains a string of the path and file to open.
When Mode is read(), pData is read to, that is, overwritten.
When Mode is write(), pData is used to write from.
All in all, pretty simple, pretty straightforward...
A "one syscall Linux or Unix (or Linux-like or Unix-like) operating system", if you will... for simplicity and understanding!
(Andrew Tannenbaum would be pleased!)
Related: "One-instruction set computer" (OISC): https://en.wikipedia.org/wiki/One-instruction_set_computer
SUN RPC for NFS, yellow pages,...
Userspace assembly runs directly on the CPU* executing in the unprivileged ring. When the userspace program makes a system call by calling a kernel entry function which is mapped into the process's address space by the dynamic loader, then part of that entry into kernelspace is to put the CPU into the privileged ring and kernel assembly then runs on the CPU.
The process scheduler can stop execution to kick a task off the CPU and switch to another one, depending on OS and kernel some things can be kicked off the CPU and some cannot.
Userspace memory allocations are serviced by virtual memory where the page tables track the translation of virtual memory pages into physical memory pages using the MMU.
The kernel is involved during allocation and page fault, but iiuc a regular successful virtual memory access is a hardware operation only.
I don't have a diagram of how this works. Neither processes nor memory are my usual area of kernel.
You'd be better to read the x86 version of the XV6 book to learn how this stuff really works. It's really well written and implements enough to be tangibly useful. Reading the code is optional when just learning concepts. Reading the XV6 code will hopefully help you understand the O'Reilly Linux books better, which will hopefully help you understand the actual Linux kernel better.
(*yes I'm aware CPUs don't directly execute assembly anymore, but the microcode guarantees the observable CPU state at any Instruction Pointer matches the expectation as if you were running assembly on a PDP or C64, or close enough for 99.999% purposes and definitely enough for debugging your program in gdb)
I am currently reading "Asynchronous Programming in Rust: Learn asynchronous programming by building working examples of futures, green threads, and runtimes" and there is vague talk about how cpus process things, but it really would like to know more. I am even curious about what is actually happening in hardware. It seems hard to determine where to start.
I don't have formal education in this field unfortunately.
I always thought that assembly was a type of programming language.
VMs (well not emulated vms, but if you're doing an x86 vm on x86 or an arm vm on arm) do something similar - an inaccurate (but useful for the concept) way to think of it is that the cpu does 2 layers of MMU for user processes in a vm.
The kernel code isn't running directly when your program accesses memory, but it sets up the cpu so that the kernel still has control over your memory, and you only have access to what the kernel allows - its mediated by the kernel.
Even with full emulation, I'd say memory is being accessed directly, unless you really go out of your way to make it weird.
There are some differences which may interest you: https://9fans.github.io/plan9port/man/man9/intro.html
I think you may find that some of the additional complexity of 9p is necessary, but perhaps not all of it.
A process has access to a set of capabilities (if it does not have any capabilities, then it is automatically terminated (unless a debugger is attached), since there is nothing for the program to do).
A "message" consists of a sequence of bytes and/or capabilities. (The message format will be system-independent (e.g. the endianness is always the same) so that it works with emulation and network transparency, described below.)
A process can send messages to capabilities it has access to, receive messages from capabilities it has access to, create new capabilities (called "proxy capabilities"), discard capabilities, and wait for capabilities.
Terminating the process is equivalent to a mandatory blocking wait for an empty set of capabilities; discarding all capabilities also terminates the process. A non-blocking wait for an empty set of capabilities means that you wish to yield, so that other processes get a chance to run, before this process continues.
Some further options may be needed to handle multiple locking and transactions, and to avoid some kinds of race conditions, but mostly that is just it.
This is useful for many things, including sandboxing, emulation, network transparency (this can be done by one program keeping track of which capabilities need to be sent across the network link and assign an index number to each one, and then the other end will create a proxy capability for each index number and use that number when it wants to send back), security with user accouts, etc; the kernel does not need to know about all of these things since they can be implemented in user code.
Other things (outside of the kernel) can also be implemented in terms of proxy capabilities, and I had ideas about those other parts of the operating system too, for example it has a hypertext file system (with no file names, but files can contain multiple numbered streams, which can include both bytes and links to other files (which can be either to the current version or to a fixed version; if to a fixed version then copy-on-write will be used if the file is modified)), and the "foreign links table", and a common (binar) data format, and a command shell with some similarities than Nushell (but also many differences), and the system uses the "Extended TRON Code" character set, and details about the working of the package manager and IME and window manager, etc.
The FreeBSD kernel/world is almost exquisite in it's engineering simplicity. Compared to the sometimes chaotic world of GNU/Linux (in my experience).
I love wikipedia and I read to procrastinate by reading the articles for everything I am interested in so I have found quite a lot of factually incorrect information or statements of fact which are really philosophical opinions. However usually these problems coexist with a certain change in writing style (loss of formatting, grammatical errors, random capitalizations, change of tone, etc.). I find I haven't found many problems with content written in the usual "wikipedia" style, so I assume the hardcore wk editors who enforce this style care a lot about factual accuracy. However I don't read enough outside of wikipedia so I wouldn't know if everything is correct...
(actually now that I think about, maybe I am more likely to agree with things in the wikipedia style. But I think the style errors and factual errors are at least a bit correlated.)
Everything being a file, and everything being read/write calls are different things. There's pros and cons.
You could argue that nothing is direct unless you're running on a bare metal system, no MMU, no page tables. How do you define "direct"?
https://www.youtube.com/watch?v=oO8_2JJV0B4
In short, assembly for a given CPU is very nearly one-to-one with the machine language for that CPU. It's not correct to conflate the two, but close enough when speaking informally.
It's even blurrier when you consider that most modern assembler dialects have plenty of high level functionality (structs, macros, labels, etc) that do not correlate to machine instructions.
My instructions are executed directly on the cpu. My file reads and writes directly translated from a stream of bytes to block instructions by code in the kernel.
Memory is a wierd in-between place, or maybe a 3rd option, since the kernel has to run a bunch of code on my behalf for me to use memory, sort of like the filesystem thing, but I'm using the direct hardware units afterward, sort of like instructions.
With modern computers, AFAIK even OS does not have absolutely full control over CPU.
Forget virtualisation. Compile a userspace program which just adds numbers into a stack variable. That program is running directly on the CPU in the unprivileged ring.
A userspace program in a VT-x virtual machine is exactly the same.
If those programs attempt privileged access then that access will fail and a trap is raised. That's what the CPU ring controls.
Why wouldn't it? Several features are simply not available in ring 3. Several features are configured for you in a way you cannot change. Several instructions will simply fault your program.
> which just adds numbers into a stack variable
Yes.. and when you eventually overflow that stack, what happens? How did the stack segment selector get created? Can you change that selector or it's attributes? Can you set the stack pointer to any valid memory address you like?
> A userspace program in a VT-x virtual machine is exactly the same.
What does an IOMMU do?
> If those programs attempt privileged access then that access will fail and a trap is raised.
Right.. so you are not directly using the CPU. You're not even in control of what timeslices are afforded to you by the OS. You are in an exceptionally limited environment most of which you cannot control or alter and much of which you cannot even observe.
The fact that instructions get dispatched according to the system ABI when you run a program is not material to this problem, and in particular, is not at all correctly represented by this diagram.
And at that point, the whole "everything is" turns into nonsense, because yes, everything is a pointer to something, so what.
The difference to plan9 is not the files, but the way plan9 uses text protocols with read/write to ctl files. To open a TCP connection - if memory serves me right - you first have to write to a ctl file, which creates a new file for the connection. Then, you write the dial command to the ctl file of that connection, and after which you can open the connection file. On Linux, a syscall creates an anonymous file, and then everything after is operations on this anonymous file.
There's some ideological benefits, but plan9 creates a mess of implicit text protocols, ugly string handling, syscall storms and memory inefficiencies. Their design is pretty much solely a limitation caused by the idea that all filesystems should exist through the 9p protocol, which as a networked protocol cannot share data (fds, structs), only copy (payloads to/from read, write). With the idea that all functionality had to be replaceable and mountable from remote machines, the only possible API became read/write for everything.
I'd argue that fd-using syscalls and ioctls - basically per-file syscalls - is a superior approach to implement everything-as-a-file.
If I've got a Plan9 system mounted over, say, NFS, would this all mean that (ignoring permissions) I could effectively open a TCP connection from that remote machine by writing appropriate information to a file on the NFS share? It would be pretty inefficient I suspect, tunnelling TCP over NFS, but it seems like there could be an incredible amount of cool hacks that might Just Work as a side-effect of them going all-in on "everything is a file".
There was perhaps a time where the differences between having everything in binary ioctls and bound to specifically one device was a necessary component in order to reach reasonable performance, but I don't believe that is the case anymore. Anecdotally these days everything on Plan 9 feels snappier. We have some benchmarks that show that 9front outperforms linux with naive pipe io and context switches. What Plan 9 misses in micro optimizations it makes up for by having a incredibly consistent and versatile base.
I want to reiterate the benefits of the network transparency by talking about how drawterm works. Drawterm can be thought of the plan 9 equivalent of windows RDP. How it works is that internally drawterm creates routines to expose a /dev/draw, /dev/mouse and /dev/keyboard through whichever native way there is on the target system (macos, windows, linux, etc). It then attaches to the remote system and overlays these files over a namespace. Programs like our window manager rio can then be run completely transparently, forwarding not compressed images, but individual draw RPC messages. There is no need for any special code on the plan 9 host side in order to accommodate drawterm, again it is something that just falls out of the core design of the system.
It's not clear to me that 9p itself could not be extended to allow for shared memory. With low-level control over the operating system and rebuilding of existing binaries, distributed shared memory becomes a possibility. (I.e. the existing VM system ought to be enough to implement whatever cache coherence is needed for shared memory over the network.)
Whats magical about the segmet(3)[1] device? The '#' devices are kernel file servers. There's no magic.
It's even more amusing with Type 2 hypervisors.
On the other hand, linux ioctl and syscalls have infinite binary structs you need to know (and cannot let the compiler reorder fields in), which then doesn't make cross-platform development any easier.
Re-ordering of structs is always forbidden with the binary format being strictly specified, so there's nothing to worry about there. Can't exactly shuffle bytes in a text format either, and plan9 control strings tend to have positional arguments.
The current structs do leave something to be desired though.
It doesn’t matter which order the client sends {“a”: 1, “b”: 2 }, it’s one object with “a” and “b” regardless
https://git.cloudef.pw/uinputd.git/tree/common/packet.h#n34 https://git.cloudef.pw/uinputd.git/tree/server/uinputd.c#n16
(Excuse to write code to use my PS Vita as gamepad :D)
User namespaces are still a hell of a lot clunkier than "each process inherits its parents' namespace".
Position having meaning is a common part of most data representations, not something specific to structs. If you need to, you can also engineer order independence in structs with arrays of type and value unions, but there's no need to as the order has never been a problem.
I imagine the issue you experienced might have been either that you hardcoded the numeric size instead of using sizeof and then building on a different architecture, or that you redeclared the struct for one architecture, ignoring that you were building for another architecture.
Not to undermine the annoyance or surprise-factor of your bug, I am tempted to say that it falls into generic logic bugs rather than the fault of any system. Changing to an encoded with dynamic length for example would just introduce different classes of logic bugs, like the whole series of possible issues within JSON parsing/serialization and requiring dynamic buffers.
A struct on the other hand only requires referencing the right struct declaration - the only amount of care you need is to include and use them. In my opinion, this is the maximum possible convenience for such system.
