Re: svn commit: r235777 - head/sys/kern
On Tue, 22 May 2012, Hartmut Brandt wrote: On Tue, 22 May 2012, Bruce Evans wrote: BE>On Tue, 22 May 2012, Hartmut Brandt wrote: BE> BE>> Log: BE>> Make dumptid non-static. It is used by libkvm to detect whether BE>> this is a VNET-kernel or not. gcc used to put the static symbol into BE>> the symbol table, clang does not. This fixes the 'netstat: no namelist' BE>> error seen on clang+VNET systems. BE>> BE>> Modified: BE>> head/sys/kern/kern_shutdown.c BE> BE>That would be a bug in clang if it were done for static symbols generally, BE>but here the bug seems to be that the symbol is not declared as __used. I don't get this. Why should a symbol declared static be in the symbol table (except for debugging purposes) ? It must be there for debugging purposes and historical compatibility (mainly other debugging uses, including kvm). Static symbols are not really special here. The C standard doesn't even require a symbol table, and it is only implementation details and debugging and historical compatibility that require putting global symbols in symbol tables. It has internal linkage and so has a meaning only in the given file. What is the linker supposed to do with several static symbols with the same name from several object files? Same as it always did. The names are in per-object-file namespaces for the linker, so they don't conflict for linking, but they mess up primitive debuggers starting with nm. If several files declared static dumptids, which one would kldsym be supposed to return? libkvm and kldsym would be broken. I don't know of any even non-primitive debuggers that can handle this. In gdb the only way that I know of to print a non-unique variable is to display the source file that declares the variable using something like `l' on a function in that file; the variable scope is then that of the selected file. The address "file.c:foo" only works for displaying functions. The latter probably depends on a full symbol table. When there is no symbol table, `l' of course doesn't work; `disass foo" works. "disass file.c:foo" of course doesn't work. "disass file.o:foo" should work, but doesn't, even when there is a full symbol table. When the "file.*:" address doesn't work, this is is initially because it misparsed as the symbol "file". Actually, I know of the primitive way which works even using ddb with all symbols broken in ddb, and have had to use this several times: you find the address of things using nm or something less primitive on another system, and guess which address to use if several variables have the same name, and type it in. BTW, kernels have lots of conflicting symbols which mess up debugging using primitive debuggers like ddb. /usr/src/tools/tools/kernxref/kernxref.sh was supposed to be used to find and fix these as well as finding and fixing public variables that should be static, but it is rarely used. uniq reports removing 382 out of 28443 (non-unique) symbols in a fairly current kernel with not many features configured. Bruce ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"
Re: svn commit: r235777 - head/sys/kern
On 22. May 2012, at 07:23 , Hartmut Brandt wrote: > Author: harti > Date: Tue May 22 07:23:41 2012 > New Revision: 235777 > URL: http://svn.freebsd.org/changeset/base/235777 > > Log: > Make dumptid non-static. It is used by libkvm to detect whether > this is a VNET-kernel or not. Just for clarifications - it's used to detect whether we are operating on a crash dump or not in the vnet case. > gcc used to put the static symbol into > the symbol table, clang does not. This fixes the 'netstat: no namelist' > error seen on clang+VNET systems. > > Modified: > head/sys/kern/kern_shutdown.c > > Modified: head/sys/kern/kern_shutdown.c > == > --- head/sys/kern/kern_shutdown.c Tue May 22 07:04:23 2012 > (r235776) > +++ head/sys/kern/kern_shutdown.c Tue May 22 07:23:41 2012 > (r235777) > @@ -151,7 +151,7 @@ static struct dumperinfo dumper; /* our > > /* Context information for dump-debuggers. */ > static struct pcb dumppcb;/* Registers. */ > -static lwpid_t dumptid; /* Thread ID. */ > +lwpid_t dumptid; /* Thread ID. */ > > static void poweroff_wait(void *, int); > static void shutdown_halt(void *junk, int howto); -- Bjoern A. Zeeb You have to have visions! It does not matter how good you are. It matters what good you do! ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"
Re: svn commit: r235777 - head/sys/kern
On Tue, 22 May 2012, Bruce Evans wrote: BE>On Tue, 22 May 2012, Hartmut Brandt wrote: BE> BE>> Log: BE>> Make dumptid non-static. It is used by libkvm to detect whether BE>> this is a VNET-kernel or not. gcc used to put the static symbol into BE>> the symbol table, clang does not. This fixes the 'netstat: no namelist' BE>> error seen on clang+VNET systems. BE>> BE>> Modified: BE>> head/sys/kern/kern_shutdown.c BE> BE>That would be a bug in clang if it were done for static symbols generally, BE>but here the bug seems to be that the symbol is not declared as __used. I don't get this. Why should a symbol declared static be in the symbol table (except for debugging purposes) ? It has internal linkage and so has a meaning only in the given file. What is the linker supposed to do with several static symbols with the same name from several object files? If several files declared static dumptids, which one would kldsym be supposed to return? harti BE> BE>gcc does the same for a file containing only "static int x;", but it BE>is apparently confused by dumptid being initialized non-statically, BE>although the initialization has no side effects. If dumptid were a BE>local variable, then clang would probably warn about the variable being BE>unused, but gcc-4.2.1 never detects such unused variables (thus code BE>that compiles with gcc -Wunused -Werror often fails with clang). Here BE>the initialization is to curthread->td_tid, so it isn't clear if the BE>compiler can tell if it has no side effects. curthread() is actually BE>__curthread(). __curthread() is now declared as __pure2, but that BE>never worked for me with older compilers (its result wasn't cached). BE>If the compilers can tell that the expression has no side effects, BE>then it is another bug that they don't warn about it having no effect BE>when it is only assigned to the apparently-unused variable dumptid. BE> BE>> Modified: head/sys/kern/kern_shutdown.c BE>> == BE>> --- head/sys/kern/kern_shutdown.c Tue May 22 07:04:23 2012 BE>> (r235776) BE>> +++ head/sys/kern/kern_shutdown.c Tue May 22 07:23:41 2012 BE>> (r235777) BE>> @@ -151,7 +151,7 @@ static struct dumperinfo dumper; /* our BE>> BE>> /* Context information for dump-debuggers. */ BE>> static struct pcb dumppcb; /* Registers. */ BE>> -static lwpid_t dumptid; /* Thread ID. */ BE>> +lwpid_t dumptid; /* Thread ID. */ BE>> BE>> static void poweroff_wait(void *, int); BE>> static void shutdown_halt(void *junk, int howto); BE> BE>Now there are 3 bugs instead of 1: BE>- the variable is declared (implicit) extern instead of static BE>- the extern declaration is in a section for static declaration BE>- the variable is still not declared as __used. If the compiler did BE> a more extensive usage analysis, that looked at all object files but BE> not at the libkvm API, then it should remove this variable anyway BE> when it is not declared as __used. BE> BE>Bruce BE> BE> ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"
Re: svn commit: r235777 - head/sys/kern
On Tue, 22 May 2012, Hartmut Brandt wrote: Log: Make dumptid non-static. It is used by libkvm to detect whether this is a VNET-kernel or not. gcc used to put the static symbol into the symbol table, clang does not. This fixes the 'netstat: no namelist' error seen on clang+VNET systems. Modified: head/sys/kern/kern_shutdown.c That would be a bug in clang if it were done for static symbols generally, but here the bug seems to be that the symbol is not declared as __used. gcc does the same for a file containing only "static int x;", but it is apparently confused by dumptid being initialized non-statically, although the initialization has no side effects. If dumptid were a local variable, then clang would probably warn about the variable being unused, but gcc-4.2.1 never detects such unused variables (thus code that compiles with gcc -Wunused -Werror often fails with clang). Here the initialization is to curthread->td_tid, so it isn't clear if the compiler can tell if it has no side effects. curthread() is actually __curthread(). __curthread() is now declared as __pure2, but that never worked for me with older compilers (its result wasn't cached). If the compilers can tell that the expression has no side effects, then it is another bug that they don't warn about it having no effect when it is only assigned to the apparently-unused variable dumptid. Modified: head/sys/kern/kern_shutdown.c == --- head/sys/kern/kern_shutdown.c Tue May 22 07:04:23 2012 (r235776) +++ head/sys/kern/kern_shutdown.c Tue May 22 07:23:41 2012 (r235777) @@ -151,7 +151,7 @@ static struct dumperinfo dumper;/* our /* Context information for dump-debuggers. */ static struct pcb dumppcb; /* Registers. */ -static lwpid_t dumptid;/* Thread ID. */ +lwpid_t dumptid; /* Thread ID. */ static void poweroff_wait(void *, int); static void shutdown_halt(void *junk, int howto); Now there are 3 bugs instead of 1: - the variable is declared (implicit) extern instead of static - the extern declaration is in a section for static declaration - the variable is still not declared as __used. If the compiler did a more extensive usage analysis, that looked at all object files but not at the libkvm API, then it should remove this variable anyway when it is not declared as __used. Bruce ___ svn-src-all@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/svn-src-all To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"