That’s my guess, but there was a conversation on the mailing list a few months ago that wasn’t just immediately shut down, even by other prolific developers
Ts’o seems skeptical, but is at least asking whether c++ has improved
Take a look at what even the proposer is saying wouldn’t be allowed in:
(1) newand delete. There's no way to pass GFP_* flags in.
(2) Constructors and destructors. Nests of implicit code makes the code less
obvious, and the replacement ofstatic initialisation with constructor
calls would make the code size larger.
(3) Exceptions and RTTI. RTTI would bulk the kernel up too much and
exception handling is limited without it, and since destructors are not
allowed, you still have to manually clean up after an error.
(4) Operator overloading (except in special cases).
(5) Function overloading (except in special inline cases).
(6) STL (though some type trait bits are needed to replace __builtins that
don't exist in g++).
(7) 'class', 'private', 'namespace'.
(8) 'virtual'. Don't want virtual base classes, though virtual function
tables might make operations tables more efficient.
C++ without class, constructors, destructors, most overloading and the STL? Wow.
I’ve only worked on a few embedded systems where C++ was even an option, but they allowed 2, 4, 5, and 7. Though, for the most part most classes were simple interfaces to some sort of SPI/I2C/CAN/EtherCAT device, most of which were singletons.
In that post, his critiques were around the problems with the STL and everyone using Boost. The STL has improved significantly since then, and it would be a limited subset of c++ if it was ever allowed
There have been mailing list conversations earlier this year, citing that clang/gcc now allowing c++ in their own code might mean they’ve taken care of the issues that made it unusable for kernel code
I’m not saying it will happen, but it’s not being shot down as an absolute insanity anymore, and I wouldn’t have expected Rust to be allowed in the kernel, either
Oh interesting. I didn’t realize boost was the main issue. Most people I’ve talked to were complaining about VTables introducing a bunch of indirection and people blindly using associative containers.
Vtable equivalents are used extensively in the kernel
You’ll find structs all over the place setting them up, e.g. every driver sets up a .probe function that the core will call, since it doesn’t know what driver it’s loading
Right the issue was more because they’re so easy to throw in without thinking about it so people overuse them. That may just be older devs complaining about newbies though.
For an example from the other poster’s explanation:
https://lwn.net/Articles/249460/
This was pre c++11 - not sure if he’s changed his mind at all with more modern c++
He absolutely has not.
That’s my guess, but there was a conversation on the mailing list a few months ago that wasn’t just immediately shut down, even by other prolific developers
Ts’o seems skeptical, but is at least asking whether c++ has improved
https://lore.kernel.org/lkml/20240110175755.GC1006537@mit.edu/
Take a look at what even the proposer is saying wouldn’t be allowed in:
(1) new and delete. There's no way to pass GFP_* flags in. (2) Constructors and destructors. Nests of implicit code makes the code less obvious, and the replacement of static initialisation with constructor calls would make the code size larger. (3) Exceptions and RTTI. RTTI would bulk the kernel up too much and exception handling is limited without it, and since destructors are not allowed, you still have to manually clean up after an error. (4) Operator overloading (except in special cases). (5) Function overloading (except in special inline cases). (6) STL (though some type trait bits are needed to replace __builtins that don't exist in g++). (7) 'class', 'private', 'namespace'. (8) 'virtual'. Don't want virtual base classes, though virtual function tables might make operations tables more efficient.
C++ without
class
, constructors, destructors, most overloading and the STL? Wow.That doesn’t really surprise me, as most of those are the same requirements from any embedded development use case using c++ that I’ve worked on
4 and 5 are the only ones stricter than I’m used to
I’ve only worked on a few embedded systems where C++ was even an option, but they allowed 2, 4, 5, and 7. Though, for the most part most classes were simple interfaces to some sort of SPI/I2C/CAN/EtherCAT device, most of which were singletons.
time to go pedantic and use parts of the c++stdlib that weren’t included in the stl!
I don’t think its the ergonomics of the language he has an issue with. If anything C++1x probably just made the original critiques of bloat worse.
In that post, his critiques were around the problems with the STL and everyone using Boost. The STL has improved significantly since then, and it would be a limited subset of c++ if it was ever allowed
There have been mailing list conversations earlier this year, citing that clang/gcc now allowing c++ in their own code might mean they’ve taken care of the issues that made it unusable for kernel code
https://lore.kernel.org/lkml/e5949a27-999d-4b6e-8c49-3dbed32a00bc@zytor.com/
I’m not saying it will happen, but it’s not being shot down as an absolute insanity anymore, and I wouldn’t have expected Rust to be allowed in the kernel, either
Oh interesting. I didn’t realize boost was the main issue. Most people I’ve talked to were complaining about VTables introducing a bunch of indirection and people blindly using associative containers.
Vtable equivalents are used extensively in the kernel
You’ll find structs all over the place setting them up, e.g. every driver sets up a .probe function that the core will call, since it doesn’t know what driver it’s loading
Right the issue was more because they’re so easy to throw in without thinking about it so people overuse them. That may just be older devs complaining about newbies though.