Conversation with _wmd on Hacker News

Source

First Comment

I admire experimentation, so below comments should preferably be considered constructive..

I want the binary modules any user of Olin would upload today to be still working, untouched, in 5 years, assuming its dependencies outside of the module still work

There are only 3 program formats that enjoyed anywhere like that kind of longevity – PE, Linux ELF, and Jar, all with significant money behind them. Of these, only PE has managed anything like ABI and dependency compatibility across more than 10 years, and that's because of explicitly obsessive compatibility concerns baked into the culture of the company behind it.

The nearest best example to PE is Linux Intel ELF, where despite explicit efforts, and where glibc which nearly-almost-if-you-squint manages long term ABI stability, and the kernel which nearly-almost-if-you-squint manages long term compatibility with glibc, you can run a terminal program compiled at the start of the millennium if you boot with the right kernel command line options, so long as you also run it with an ancient buggy libc, and that program has no other system dependencies.

Coming from Python land, I can say categorically that any attempt to maintain feverishly accurate compatibility for a rich set of interfaces across a community is almost a pointless effort, the options always degrade to running ancient buggy deps, or updated deps with broken interfaces.

The Linux kernel barely manages it despite a strict policy, and they break it occasionally due to security concerns. Downstream community projects follow the whim of whoever is contributing code on that particular day, and their motivations are rarely if ever “compatibility”.

That is not to say a community effort couldn't manage >10 year compatibility, but that such concerns must be a foremost priority (rather than a cute goal) if there is even a dying chance of the attempt succeeding in the long term. It seems better that the inevitability of incompatibility is honestly accepted and designed for rather than hoping things will just work out.

The core idea is that everything is a file, to the point that the file descriptor and file handle array are the only real bits of persistent state for the process

Again an admirable goal, but this is Stockholm syndrome in its purest form, and learning completely the wrong lessons from UNIX land. Files are a shit abstraction, but ignoring those, streams even moreso. Very few elements of a modern application even at the lowest tiers lend themselves well to being expressed as a serialized bytestream.

The minimum primitive (and happy to share many references to support this) should be something like messages and ports. Streams can be expressed as messages, but the converse is not true without application-specific protocol scaffolding, of which there are a thousand examples on every modern UNIX system as a consequence of UNIX picking streams.

When a dagger process is opened, the following files are open:

Sequentially numbered transparent handles are a mistake, as are predefined resources, they allow code to make incorrect assumptions that are difficult to manage over the life of a system. You say that glibc should be built on these APIs, yet one of the principal restrictions glibc suffers from is its inability to open long-lived file handles that might conflict with program assumptions — assumptions often make implicitly without any conscious thought on the part of the programmer. Incremental numbering SUCKS.

A few years back Linux even considered adding a brand new fd namespace just so glibc could safely open files, all because of predictable numbered handles with assumed semantics that did not exist.

Try to learn from NT and elsewhere here — the user receives an opaque word, no predefined words exist, and an interface is provided to grant access to any standard resources without introducing global state.

open, close, read, write, sync

But I thought everything is a file.. why do we need these? Why can't we express these as files too?

Due to incorrect lessons above, in a real world system this list will inevitably grow to multiple times its original size. What does sync mean on a socket? How do I eject a DVD drive? How does an application read the system time? By opening some “time://” and reading a serialized time representation? Good luck with nanosecond resolution when microseconds are already lost on scaffolding and maybe parsing some ASCII format timestamp, and so a “one time only” specialized time() call appears.

10 years later and you have hundreds of entries just like /usr/share/man/man2 on a typical UNIX, probably alongside something like ioctl(). But of course this time will be different!

the open call (defined later), a file URL is specified instead of a file name. This allows for Dagger to natively offer programs using it quick access to common services like HTTP, logging or pretty much anything else.

The open call prototype as defined is totally unusable for any kind of network application. There is no room for setting any kind of meaningful options outside a single bitfield, and so already if I wanted to make any kind of custom HTTP request, the supplied interface is useless and I must link a module with a real client built on top of the tcp:// handler.

There seems to be no long term sense whatsoever in baking a handful of protocols into the “filesystem” interface, especially not with these APIs.

I’d like to add the following handlers in the future:

I see they have thought of time:// already! By discarding one of the few remaining uniformities of the UNIX file API – the namespace it exports. It already looks like we're coping with a bad base abstraction by shoving everything we need into magic strings.

What happens if I only read 2 bytes from time://? Do I get pieces of the last sampled timestamp because some internal buffering now exists, or do i get half the old timestamp half the new one, or do I continually get the first 2 bytes of unrelated timestamps, or perhaps we simply return an error because the buffer wasn't big enough? It's almost like even a timestamp doesn't really fit the stream-of-bytes file abstraction.

Very happy to see all these new WebAssembly efforts, but this one's present design fails to learn any lesson from our past suffering.

My Reply

below comments should preferably be considered constructive

Author of the post here. Don't worry, I am taking them as such.

It seems better that the inevitability of incompatibility is honestly accepted and designed for rather than hoping things will just work out.

How would this look like? I'm interested by what you mean by this. I'm not sure how incompatibility with things like reading out sequential memory is going to be tolerable though.

Again an admirable goal, but this is Stockholm syndrome in its purest form, and learning completely the wrong lessons from UNIX land. Files are a shit abstraction, but ignoring those, streams even moreso. Very few elements of a modern application even at the lowest tiers lend themselves well to being expressed as a serialized bytestream.

The minimum primitive (and happy to share many references to support this) should be something like messages and ports. Streams can be expressed as messages, but the converse is not true without application-specific protocol scaffolding, of which there are a thousand examples on every modern UNIX system as a consequence of UNIX picking streams.

Yes, I eventually want to settle on messages. The reason I have picked file I/O in the meantime is so I can hack something together that will “just work”. From a generic level, I don't know what memory addresses are safe to write to. Using the read() function makes the program choose for itself.

Streams are a shitty abstraction, but most of the world is currently built on them, so I'm using streams in Dagger if only to make Dagger have to be thrown away in the long run. Streams may be bad in the long term, but for now we can do HTTP via the filesystem calls: https://github.com/Xe/olin/blob/master/internal/abi/dagger/t...

file descriptors are a mistake

Yeah, I'm betting that they are gonna be a mistake in the long term. I just don't know what abstraction to use that won't be yet. I guess I sort of am using file streams as messages. I'm gonna see what it would look like to make messages the primitive.

But I thought everything is a file.. why do we need these? Why can't we express these as files too?

File descriptors are adjectives, system calls are verbs. The verb is the action you want to do, such as read. The adjective is the source, ie the semantically standard input of the process.

Due to incorrect lessons above, in a real world system this list will inevitably grow to multiple times its original size.

Yep. I'm expecting this design to be wrong in the long term. It really started as an experiment to see how far minimalism could go.

The open call prototype as defined is totally unusable for any kind of network application. There is no room for setting any kind of meaningful options outside a single bitfield, and so already if I wanted to make any kind of custom HTTP request, the supplied interface is useless and I must link a module with a real client built on top of the tcp:// handler.

https://AzureDiamond:hunter2@bash.org/244321 is what I usually do with URL's and hard-defined authentication tokens. Works for me.

I see they have thought of time:// already! By discarding one of the few remaining uniformities of the UNIX file API – the namespace it exports. It already looks like we're coping with a bad base abstraction by shoving everything we need into magic strings.

Yeah, I'm starting to see that Dagger is WAY TOO MINIMAL. I'm gonna go poke around POSIX and other minimal OS' to see what syscalls they expose. For what it's worth, the Go runtime has its own time() call that I expose here: https://github.com/Xe/olin/blob/master/internal/abi/wasmgo/abi.go#L97

Very happy to see all these new WebAssembly efforts, but this one's present design fails to learn any lesson from our past suffering.

The first design always has to be thrown out :D

Their reply

Picking just the most fun aspect to reply to, because this is too long already :)

would this look? I'm interested by what you mean by this

One area UNIX fails at is a single global flat system interface, it makes evolutionary change difficult without introducing permanent “variant” system calls (e.g. openat()) that must be supported in tandem with the original interface, in addition to an automatic penalty for any innovation that might wish to add more, because it must be supported forever.

To skirt around this, one popular option in UNIX land is instead not to define any interface: just dump whatever magic is required into some ioctl() and claim it is device-specific, and so we end up with what an interface that isn't quite guaranteed and definitely unofficial, but that's okay because it's driver-specific, but then a second driver comes along and adds a similar interface, and now we have 2 options for frobbing a particular device, with no obviously correct choice, and no guarantee frobbing will be supported that way forever. In other words, we got the worst of both worlds by trying to ignore impermanence. (Just a side note, a real-life example of those ioctls might be e.g. how Linux used to support asking the filesystem driver for a mapping of unused sectors, or how you set or query the FAT filesystem's DOS filesystem label)

UNIX exports APIs for all of this, and it's mostly frozen in stone forever:

Just taking the silly 'rewinding tapes' example, planning an API for permanent compatibility with a problem space that 99.9% of people no longer have, rising to 100% at some point eventually, it's easy in hindsight to see this interface should be modularized somehow, and preferably made entirely optional.

How that modularity looks is definitely up for debate.. looking at Windows for one approach, in COM (and ignoring its baroque useability!) handles export only one interface (IUnknown) whose sole method allows enumerating the contracts the handle (object) supports. For example here is DirectShow: https://docs.microsoft.com/en-us/windows/desktop/directshow/... , see how a single input can implement IDvdControl, IAMTVTuner, both, or neither, and if neither are implemented, the handle is still useable via IPin and related interfaces that actually control how data flows from the input.

Mapping that to UNIX, a handle might have protocols like IUnixSecurity, ITapeOperations, IBraindeadLocking, ILinuxSpecificLocking, IStream, all or none etc., where a handle has no guaranteed fixed vocabularity perhaps except for the most fundamental expected anywhere that handle is found (in the case of a file, perhaps that is IStream).

By exporting the impermanence down to consumer code, accessing any particular interface becomes an explicit affair, because you must always request it upfront. You can still implement a media player program by wiring up some filter graph using a handle you received from someplace else, but disable the rewind and eject buttons if querying those interfaces failed — this is literally how some stuff works on Windows:

handle = ask_user_for_media_input()
wire_up_graph(handle)
if(supported(handle, "dvd player")) {
    enable_eject_button();
}

Meanwhile at the top of the stack, outside of absolutely having to support some core interfaces that can never go away, the cost of adding and retiring sundry features is reduced, as the reality of permanence that has always existed has been pushed down the stack, where end users can choose to accept or ignore it.

An ideal interface might resemble a predefined root handle or sentinel value with the ability to construct new handles bound to some protocol. Imagine a language like (but not quite) protocolbuffer services, where those read/write/open/close/sync calls were methods of a particular named interface. The IDL would output wrappers to speak it. Probably already you can see room for splitting 'sync' out into some kind of IBufferedStream optional interface.

Maybe like:

#include <os.h>                             // Invoke(handle, char *interface, void *in, void *out) -> bool
                                            // Close(handle) -> void
#include <os/interface.h>                   // GetInterface(handle, char *interface) -> void *
                                            // Require(handle, char *interface) -> void *
#include <os/services/console.h>            // IF_CONSOLE, Console, Console_write
#include <os/services/filesystem.h>         // IF_FILESYSTEM, Filesystem, Filesystem_open, Filesystem_error
#include <os/services/stream.h>             // IF_STREAM, Stream_read, Stream_write
#include <os/services/buffered_stream.h>    // IF_BUFFERED_STREAM, BufferedStream_sync

int main(void)
{
    Filesystem fs = Require(NULL, IF_FILESYSTEM);
    Console console = Require(NULL, IF_CONSOLE);

    File *fp;
    if((fp = Filesystem_open(fs, "/etc/passwd", O_RDWR)) == NULL) {
        Console_write(console, "open failed: %s", Filesystem_error(fs));
        return 1;
    }

    if(File_write(fp, "foo\n") == -1) {
        Console_write(console, "write failed: %s", Filesystem_error(fs));
        return 1;
    }

    // Optional. Underlying Invoke() will fail if interface absent.
    BufferedStream_sync((BufferedStream) fp);

    return 0;
}

A real killer would be other than simplifying that calling style, somehow unifying it with the calling convention for linked modules, so by replacing a handle you could direct an OS-level protocol to a library, or perhaps better, the act of calling an OS service caused an optional library to be loaded and bound to the handle you were returned, say in the case of an ancient ITapeDrive interface that's long since stopped being a once-upon-a-time core feature.

But the main point from the example is probably the notion of an OS with only two core interfaces – Invoke() and Close(). Another place you could look for inspiration of this sort is Objective C — a similar concept is built into its guts, and its dynamic dispatch msgsend() function is optimized to all hell to match!