Skip to contentSkip to conversation

git 9c501935a3cd broke the strace build.

DISCUSSION Rob Landley 2 participants 2 messages Jan 14, 2010

lore
Share & tools
lore
#0 lore raw
Quoted 6 lines

net: Support inclusion of <linux/socket.h> before <sys/socket.h>

   The following user-space program fails to compile:
   
       #include <linux/socket.h>
       #include <sys/socket.h>

Did it ever? Isn't #including both sys/thingy and a linux/thingy in the same program considered bad form?

Quoted 6 lines
       int main() { return 0; }

The reason is that <linux/socket.h> tests __GLIBC__ to decide whether it should define various structures and macros that are now defined for user-space by <sys/socket.h>, but __GLIBC__ is not defined if no libc headers have yet been included.

Specifically, one of the tests was !__GLIBC__, I.E. checking that we _haven't_ included any of glibc's headers yet, which is presumably how you broke the strace 4.5.18 ./configure test checking for the existence of the linux/netlink.h file. (That test #includes stddef.h, linux/socket.h, and linux/netlink.h and nothing else. Since stddef.h is a gcc header and not a glibc header, __GLIBC__ isn't defined for the test code and thus that used to work under 2.5.31. Now under 2.6.32 it dies with:

linux/netlink.h:35: error: expected specifier-qualifier-list before 'sa_family_t'

And later on the strace build breaks because it doesn't #include linux/netlink.h when it needs to. (Why yes, autoconf _is_ brittle and near- useless, thanks for noticing.)

*shrug* I can patch the strace ./configure to include bits/socket.h instead of linux/socket.h, but what I can't understand is why _you_ couldn't. (What use case did this patch actually fix? Would reversing the order of those two headers have helped?)

Rob

Signature · 1 line
Latency is more important than throughput. It's that simple. - Linus Torvalds
#1 lore raw
Quoted 9 lines · Rob Landley
Quoted 6 lines

net: Support inclusion of <linux/socket.h> before <sys/socket.h>

   The following user-space program fails to compile:
   
       #include <linux/socket.h>
       #include <sys/socket.h>

Did it ever? Isn't #including both sys/thingy and a linux/thingy in the same program considered bad form?

I don't see why.

Even if a source file doesn't include both of them directly, they may both be included indirectly.

[...]

Quoted 4 lines

Now under 2.6.32 it dies with:

linux/netlink.h:35: error: expected specifier-qualifier-list before 'sa_family_t'

I did actually ask for this to be reverted when I saw that it causes breakage elsewhere, but David Miller insisted that my change was correct.

Quoted 6 lines

And later on the strace build breaks because it doesn't #include linux/netlink.h when it needs to. (Why yes, autoconf _is_ brittle and near- useless, thanks for noticing.)

*shrug* I can patch the strace ./configure to include bits/socket.h instead of linux/socket.h,

Including libc headers under 'bits' is a bad idea.

Quoted 3 lines

but what I can't understand is why _you_ couldn't. (What use case did this patch actually fix? Would reversing the order of those two headers have helped?)

I think that reversing the order of the two headers should work for kernel headers before and after this change.

Ben.

Signature · 3 lines
Ben Hutchings
The generation of random numbers is too important to be left to chance.
                                                            - Robert Coveyou