mmap is not a C feature, but POSIX. There are C platforms that don't provide mmap, and on those that do you can use mmap from other languages (there's mmap module in the Python's standard library, for example).
> However, in other most languages, you have to read() in tiny chunks, parse, process, serialize and finally write() back to the disk. This works, but is verbose and needlessly limited
C has those too and am glad that they do. This is what allows one to do other things while the buffer gets filled, without the need for multithreading.
Yes easier standardized portable async interfaces would have been nice, not sure how well supported they are.
I think C# standard library is better. You can do same unsafe code as in C, SafeBuffer.AcquirePointer method then directly access the memory. Or you can do safer and slightly slower by calling Read or Write methods of MemoryMappedViewAccessor.
All these methods are in the standard library, i.e. they work on all platforms. The C code is specific to POSIX; Windows supports memory mapped files too but the APIs are quite different.
Using mmap means that you need to be able to handle memory access exceptions when a disk read or write fails. Examples of disk access that fails includes reading from a file on a Wifi network drive, a USB device with a cable that suddenly loses its connection when the cable is jiggled, or even a removable USB drive where all disk reads fail after it sees one bad sector. If you're not prepared to handle a memory access exception when you access the mapped file, don't use mmap.
Do these really ever result in access failures instead of just hangs? How are they surfaced to processes?
In my experience, all these things just cause whatever process is memory mapping to freeze up horribly and make me regret ever using a network file system or external hard drive.
Aside from what https://news.ycombinator.com/item?id=47210893 said, mmap() is a low-level design that makes it easier to work with files that don't fit in memory and fundamentally represent a single homogeneous array of some structure. But it turns out that files commonly do fit in memory (nowadays you commonly have on the order of ~100x as much disk as memory, but millions of files); and you very often want to read them in order, because that's the easiest way to make sense of them (and tape is not at all the only storage medium historically that had a much easier time with linear access than random access); and you need to parse them because they don't represent any such array.
When I was first taught C formally, they definitely walked us through all the standard FILE* manipulators and didn't mention mmap() at all. And when I first heard about mmap() I couldn't imagine personally having a reason to use it.
C's API does not include mmap, nor does it contain any API to deal with file paths, nor does it contain any support for opening up a file picker. This paired with C's bad string support results in one of it being one of the worst file APIs.
Also using mmap is not as simple as the article lays out. For example what happens when another process modifies the file and now your processes' mapped memory consists of parts of 2 different versions of the file at the same time. You also need to build a way to know how to grow the mapping if you run out room. You also want to be able to handle failures to read or write. This means you pretty much will need to reimplement a fread and fwrite going back to the approach the author didn't like: "This works, but is verbose and needlessly limited to sequential access." So it turns out "It ends up being just a nicer way to call read() and write()" is only true if you ignore the edge cases.
No it doesn't. If you have a file that's 2^36 bytes and your address space is only 2^32, it won't work.
On a related digression, I've seen so many cases of programs that could've handled infinitely long input in constant space instead implemented as some form of "read the whole input into memory", which unnecessarily puts a limit on the input length.
I'm not sure what the author really wants to say. mmap is available in many languages (e.g. Python) on Linux (and many other *nix I suppose). C provides you with raw memory access, so using mmap is sort-of-convenient for this use case.
But if you use Python then, yes, you'll need a bytearray, because Python doesn't give you raw access to such memory - and I'm not sure you'd want to mmap a PyObject anyway?
Then, writing and reading this kind of raw memory can be kind of dangerous and non-portable - I'm not really sure that the pickle analogy even makes sense. I very much suppose (I've never tried) that if you mmap-read malicious data in C, a vulnerability would be _quite_ easy to exploit.
It has the best API for the author, that's for sure. One size does not fit all: believe it or not, different files have different uses. One does not mmap a pipe or /dev/urandom.
mmap is nice. But, I find sqlite is a better filesystem API [1]. If you are going to use mmap why not take it further and use LMDB? Both have bindings for most languages.
I guess the author didn't use that many other programming languages or OSes. You can do the same even in garbage collected languages like Java and C# and on Windows too.
I'd be careful though, as they all have quirks due to how tricky it is handling mmap faults. The Java API mentions both unique garbage collection behavior and throwing unspecified exceptions at unspecified times.
technically yes, because there's a failure path for every single failure that an OS knows about. And most others aren't so resilient. However, mmap bypasses a lot of that....
mmap is not a language feature. it is also full of its own pitfalls that you need to be aware of. recommended reading: https://db.cs.cmu.edu/mmap-cidr2022/
When a developer that usually consumes the language starts critiquing the language.
I could go on as to why it's a bad signal, psychologically, but let's just say that empirically it usually doesn't come from a good place, it's more like a developer raising the stakes of their application failing and blaming the language.
Sure one out of a thousand devs might be the next Linus Torvalds and develop the next Rust. But the odds aren't great.
mmap() is useful for some narrow use-cases, I think, but error-handling is a huge pain. I don't want to have to deal with SIGBUS in my application.
I agree that the model of mmap() is amazing, though: being able to treat a file as just a range of bytes in memory, while letting the OS handle the fussy details, is incredibly useful. (It's just that the OS doesn't handle all of the fussy details, and that's a pain.)
I can’t entirely tell what the article’s point is. It seems to be trying to say that many languages can mmap bytes, but:
> (as far as I'm aware) C is the only language that lets you specify a binary format and just use it.
I assume they mean:
struct foo { fields; };
foo *data = mmap(…);
And yes, C is one of relatively few languages that let you do this without complaint, because it’s a terrible idea. And C doesn’t even let you specify a binary format — it lets you write a struct that will correspond to a binary format in accordance with the C ABI on your particular system.
If you want to access a file containing a bunch of records using mmap, and you want a well defined format and good performance, then use something actually intended for the purpose. Cap’n Proto and FlatBuffers are fast but often produce rather large output; protobuf and its ilk are more space efficient and very widely supported; Parquet and Feather can have excellent performance and space efficiency if you use them for their intended purposes. And everything needs to deal with the fact that, if you carelessly access mmapped data that is modified while you read it in any C-like language, you get UB.
How often does anyone care about using data on a different system than it was created on?
These days, any C struct you built on amd64 will work identically on arm64. There really aren't any other architectures that matter.
And yes, managing concurrent access to shared resources requires care and cooperation. That has always been true, and has nothing specific to do with mmap.
It may have a tidy mmap api, but Smalltalk has a much better file api through its Streams hierarchy IMHO. You can create a stream on a diskfile, you can create a stream on a byteArray, you can create a stream on standard Unix streams, you can create a stream on anything where "next" makes sense.
43 comments
[ 4.7 ms ] story [ 51.0 ms ] threadC has those too and am glad that they do. This is what allows one to do other things while the buffer gets filled, without the need for multithreading.
Yes easier standardized portable async interfaces would have been nice, not sure how well supported they are.
All these methods are in the standard library, i.e. they work on all platforms. The C code is specific to POSIX; Windows supports memory mapped files too but the APIs are quite different.
In my experience, all these things just cause whatever process is memory mapping to freeze up horribly and make me regret ever using a network file system or external hard drive.
When I was first taught C formally, they definitely walked us through all the standard FILE* manipulators and didn't mention mmap() at all. And when I first heard about mmap() I couldn't imagine personally having a reason to use it.
It also often saves at least one copy operation (page cache to/from an application-level byte array), doesn't it?
> Look inside
> Platform APIs
Ok.
I agree platform APIs are better than most generic language APIs at least. I disagree on mmap being the "best".
Also using mmap is not as simple as the article lays out. For example what happens when another process modifies the file and now your processes' mapped memory consists of parts of 2 different versions of the file at the same time. You also need to build a way to know how to grow the mapping if you run out room. You also want to be able to handle failures to read or write. This means you pretty much will need to reimplement a fread and fwrite going back to the approach the author didn't like: "This works, but is verbose and needlessly limited to sequential access." So it turns out "It ends up being just a nicer way to call read() and write()" is only true if you ignore the edge cases.
No it doesn't. If you have a file that's 2^36 bytes and your address space is only 2^32, it won't work.
On a related digression, I've seen so many cases of programs that could've handled infinitely long input in constant space instead implemented as some form of "read the whole input into memory", which unnecessarily puts a limit on the input length.
I'm not sure what the author really wants to say. mmap is available in many languages (e.g. Python) on Linux (and many other *nix I suppose). C provides you with raw memory access, so using mmap is sort-of-convenient for this use case.
But if you use Python then, yes, you'll need a bytearray, because Python doesn't give you raw access to such memory - and I'm not sure you'd want to mmap a PyObject anyway?
Then, writing and reading this kind of raw memory can be kind of dangerous and non-portable - I'm not really sure that the pickle analogy even makes sense. I very much suppose (I've never tried) that if you mmap-read malicious data in C, a vulnerability would be _quite_ easy to exploit.
[1] - https://sqlite.org/fasterthanfs.html
https://docs.oracle.com/javase/8/docs/api/java/nio/MappedByt...
https://learn.microsoft.com/en-us/dotnet/api/system.io.memor...
https://learn.microsoft.com/en-us/windows/win32/memory/creat...
Memory mapping is very common.
When a developer that usually consumes the language starts critiquing the language.
I could go on as to why it's a bad signal, psychologically, but let's just say that empirically it usually doesn't come from a good place, it's more like a developer raising the stakes of their application failing and blaming the language.
Sure one out of a thousand devs might be the next Linus Torvalds and develop the next Rust. But the odds aren't great.
I agree that the model of mmap() is amazing, though: being able to treat a file as just a range of bytes in memory, while letting the OS handle the fussy details, is incredibly useful. (It's just that the OS doesn't handle all of the fussy details, and that's a pain.)
> (as far as I'm aware) C is the only language that lets you specify a binary format and just use it.
I assume they mean:
And yes, C is one of relatively few languages that let you do this without complaint, because it’s a terrible idea. And C doesn’t even let you specify a binary format — it lets you write a struct that will correspond to a binary format in accordance with the C ABI on your particular system.If you want to access a file containing a bunch of records using mmap, and you want a well defined format and good performance, then use something actually intended for the purpose. Cap’n Proto and FlatBuffers are fast but often produce rather large output; protobuf and its ilk are more space efficient and very widely supported; Parquet and Feather can have excellent performance and space efficiency if you use them for their intended purposes. And everything needs to deal with the fact that, if you carelessly access mmapped data that is modified while you read it in any C-like language, you get UB.
Flatbuffers etc. is cool but they can be very bloaty and clunky.
These days, any C struct you built on amd64 will work identically on arm64. There really aren't any other architectures that matter.
And yes, managing concurrent access to shared resources requires care and cooperation. That has always been true, and has nothing specific to do with mmap.