Shachar Shemesh <wine-patches(a)shemesh.biz> writes:
After much talk, the patch seems ready for commit. Since there are many systems with slightly older glibc that have the header but do not have the implementation in glibc, or that have a stub implementation that is guaranteed to fail, we are calling the kernel functions directly. This also allows compiling on older systems than the code actually ships to.
Accordingly, the autoconf checks for epoll_create in glibc were dropped.
That's wrong; if the libc function exists you should use it, direct system calls should only be used as a fallback. And when using system calls you must not depend on the libc header, you have to define the structures yourself. -- Alexandre Julliard julliard(a)winehq.org
Alexandre Julliard wrote:
Shachar Shemesh <wine-patches(a)shemesh.biz> writes:
After much talk, the patch seems ready for commit. Since there are many systems with slightly older glibc that have the header but do not have the implementation in glibc, or that have a stub implementation that is guaranteed to fail, we are calling the kernel functions directly. This also allows compiling on older systems than the code actually ships to.
Accordingly, the autoconf checks for epoll_create in glibc were dropped.
That's wrong; if the libc function exists you should use it, direct system calls should only be used as a fallback. And when using system calls you must not depend on the libc header, you have to define the structures yourself.
The problem is that on many platforms, the functions are defined in glibc, but stubbed out to always fail. It seems a waste to lose these platforms. Shachar -- Shachar Shemesh Lingnu Open Source Consulting ltd. http://www.lingnu.com/
Alexandre Julliard wrote:
Shachar Shemesh <wine-patches(a)shemesh.biz> writes:
After much talk, the patch seems ready for commit. Since there are many systems with slightly older glibc that have the header but do not have the implementation in glibc, or that have a stub implementation that is guaranteed to fail, we are calling the kernel functions directly. This also allows compiling on older systems than the code actually ships to.
Accordingly, the autoconf checks for epoll_create in glibc were dropped.
That's wrong; if the libc function exists you should use it, direct system calls should only be used as a fallback.
But what to do if the library functions exist during compilation, and do not during runtime? Merely linking with them will cause wineserver not to load, claiming that it requires "GLIBC version 3.2.3". Some wine uses (crossover office being one of them) require you to compile once and run almost everywhere, which run counter to this behavior.
And when using system calls you must not depend on the libc header, you have to define the structures yourself.
Just yank the definitions from the header? I can do that, the license is compatible. However, I think it's pointless. Most systems will have that header, and for those who don't, we will have a fallback. What's wrong with using the header if it's there? Shachar -- Shachar Shemesh Lingnu Open Source Consulting ltd. http://www.lingnu.com/
Shachar Shemesh <wine-devel(a)shemesh.biz> writes:
But what to do if the library functions exist during compilation, and do not during runtime? Merely linking with them will cause wineserver not to load, claiming that it requires "GLIBC version 3.2.3". Some wine uses (crossover office being one of them) require you to compile once and run almost everywhere, which run counter to this behavior.
Building with a recent glibc and running on an older one is not supported by glibc, and it won't work for many reasons having nothing to do with epoll. Crossover runs everywhere because it is built against a very old glibc.
Just yank the definitions from the header? I can do that, the license is compatible. However, I think it's pointless. Most systems will have that header, and for those who don't, we will have a fallback. What's wrong with using the header if it's there?
The libc definitions must be used with the libc function, and the kernel definitions with the system call. There is no guarantee at all that they will remain identical (check the 64-bit file support for an example of how that can happen). -- Alexandre Julliard julliard(a)winehq.org
Building with a recent glibc and running on an older one is not supported by glibc, and it won't work for many reasons having nothing to do with epoll. Crossover runs everywhere because it is built against a very old glibc.
To clarify, it is possible to build against a new glibc and have it run on older libcs, but the process is far from intuitive and requires special tools, header overrides etc. We do this in autopackage, and it works, but it's quite tricky to get right. So I guess Shachar could use dlopen, or possibly the epoll functions are weak symbols - quite a lot of glibc functions are these days. In that case you can just mark them with the weak attribute and that's no problem.
Mike Hearn wrote:
Building with a recent glibc and running on an older one is not supported by glibc, and it won't work for many reasons having nothing to do with epoll. Crossover runs everywhere because it is built against a very old glibc.
To clarify, it is possible to build against a new glibc and have it run on older libcs, but the process is far from intuitive and requires special tools, header overrides etc. We do this in autopackage, and it works, but it's quite tricky to get right.
So I guess Shachar could use dlopen, or possibly the epoll functions are weak symbols - quite a lot of glibc functions are these days. In that case you can just mark them with the weak attribute and that's no problem.
Details, please. How do I do that? Shachar
Details, please. How do I do that?
readelf -aW /lib/libc.so.6 | grep epoll if the symbols are declared WEAK then you can do this (assuming AJ agrees). #pragma weak epoll_whatever should do the trick. In the resulting binary "readelf -aW mybinary | grep epoll" should should you the weak linkage. At runtime: if (epoll_foobar) { ... it's safe to use }
On Tue, Aug 31, 2004 at 11:48:16AM +0300, Shachar Shemesh wrote:
Mike Hearn wrote:
Building with a recent glibc and running on an older one is not supported by glibc, and it won't work for many reasons having nothing to do with epoll. Crossover runs everywhere because it is built against a very old glibc.
To clarify, it is possible to build against a new glibc and have it run on older libcs, but the process is far from intuitive and requires special tools, header overrides etc. We do this in autopackage, and it works, but it's quite tricky to get right.
So I guess Shachar could use dlopen, or possibly the epoll functions are weak symbols - quite a lot of glibc functions are these days. In that case you can just mark them with the weak attribute and that's no problem.
Details, please. How do I do that?
Guys, pretty pretty please, just use the glibc function, add a regular configure check and use it if it does not return -ENOSYS. All this confusing handling of dlopening etc etc is way overblown. Ciao, Marcus
participants (5)
-
Alexandre Julliard -
Marcus Meissner -
Mike Hearn -
Shachar Shemesh -
Shachar Shemesh