On Thu, 2006-09-21 at 23:54 +0300, emin ak wrote: Hi Emin: Yep, I edit "static inline" to "static_inline" and then define static_inline as static for KGDB kernels. In include/linux/compiler-gcc3.h and include/linux/compiler-gcc4.h I added: ------------------------------------------------------------------ #if defined(CONFIG_KGDB) || defined(CONFIG_KEXEC) # define static_inline static __attribute__ ((__unused__)) # define static__inline__ static __attribute__ ((__unused__)) # define INLINE __attribute__ ((__unused__)) # define __INLINE__ __attribute__ ((__unused__)) #else # define static_inline static inline # define static__inline__ static __inline__ # define INLINE inline # define __INLINE__ __inline__ #endif ---------------------------------------------------------------------- I'm using it today to understand the device mapping and encryption code. It's great! Inline's make skipping over code with the gdb 'next' instruction impossible and you can't see the local variables. I like having a large stack, compiling -O0 and without inlines can increase the stack size. I think I notices more stability by adding this to include/asm-i386/thread_info.h: ---------------------------------------------------------------------- #if defined(CONFIG_DEBUG_PREEMPT_AUDIT) || defined(CONFIG_KGDB) || defined(CONFIG_KEXEC) #define THREAD_SIZE (8192 * 2) #else #ifdef CONFIG_4KSTACKS #define THREAD_SIZE (4096) #else #define THREAD_SIZE (8192) #endif #endif ----------------------------------------------------------------------- Without OPTIMIZATION I found the MMU code needs a tweak in../linux-4/mm/memory.c: ------------------------------------------------------------------------ #if !defined(__PAGETABLE_PUD_FOLDED) || defined(CONFIG_KGDB) || defined(CONFIG_KEXEC) /* * Allocate page upper directory. * * We've already handled the fast-path in-line, and we own the * page table lock. */ pud_t fastcall *__pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) { . . . } #if !defined(__PAGETABLE_PMD_FOLDED) || defined(CONFIG_KGDB) || defined(CONFIG_KEXEC) /* * Allocate page middle directory. * * We've already handled the fast-path in-line, and we own the * page table lock. */ pmd_t fastcall *__pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) { . . } --------------------------------------------------------------------------- Maybe I should have used #if !defined(__OPTIMIZE__) in ../linux-4/mm/memory.c. Another change is I needed to define a few network byte swapping functions. I currently define them in ../linux-4/net/core/sock.c but I'm not resistant to putting it in a better place: ---------------------------------------------------------------------------- /* * If compiling -O0 we need to define * these functions somewhere. */ #if !defined(__OPTIMIZE__) #define ___htonl(x) __cpu_to_be32(x) #define ___htons(x) __cpu_to_be16(x) #define ___ntohl(x) __be32_to_cpu(x) #define ___ntohs(x) __be16_to_cpu(x) __u32 htonl(__be32 x) { return(___htonl(x)); } __u32 ntohl(__be32 x) { return(___ntohl(x)); } __be16 htons(__u16 x) { return(___htons(x)); } __u16 ntohs(__be16 x) { return(___ntohs(x)); } EXPORT_SYMBOL(htonl); EXPORT_SYMBOL(ntohl); EXPORT_SYMBOL(htons); EXPORT_SYMBOL(ntohs); #endif ------------------------------------------------------------------------------ Sometimes I only want to compile the tcp/ip code -O0, so I modified the networking Makefiles and added: ------------------------------------------------------------------------------- ifdef CONFIG_KGDB CFLAGS += -gdwarf-2 -O0 else ifdef CONFIG_KEXEC CFLAGS += -gdwarf-2 -O0 endif endif ------------------------------------------------------------------------------ In the top level kernel Makefile I have: ------------------------------------------------------------------------------ ifdef CONFIG_FRAME_POINTER CFLAGS += -fno-omit-frame-pointer else CFLAGS += -fomit-frame-pointer endif # # Compiling the complete kernel without optimization (-O0) for enhanced debugging # with kgdb/kdump requires ./mm/memory.c to have: # # if !defined(__PAGETABLE_PUD_FOLDED) || defined(CONFIG_KGDB) || defined(CONFIG_KEXEC) # and # if !defined(__PAGETABLE_PMD_FOLDED) || defined(CONFIG_KGDB) || defined(CONFIG_KEXEC) # # A less invasive procedure is to use -O1 and only use -O0 for networking code. # The networking Makefiles have been setup to support this. So just change # -O0 to -O1 below and back out the kgdb change in ./mm/memory.c for a # less invasive change. Compiling -O0 also required increasing ROUNDUP_WAIT in # linux/kernel/kgdb.c; value in 2.6.12 patch was way to low and value in 2.6.16 # is marginal and frequently causes lead CPU to times out prematurely waiting for # other CPU's to stop. # ifdef CONFIG_DEBUG_INFO ifdef CONFIG_KGDB CFLAGS += -gdwarf-2 -O0 else ifdef CONFIG_KEXEC CFLAGS += -gdwarf-2 -O1 else CFLAGS += -g endif endif endif ------------------------------------------------------------------------------------ Yep, I don't understand why everyone else doesn't. It's also like using debug printf's, I like being able to trace the code to get the big picture and then a -O0 to look at details with kgdb. Some believe doing this kind of stuff is blasphemy. The Bible says I should be killed for working on Sunday; I happen to disagree. I offered to post them to Amit back on Sept 06(2:45 PM) but I don't think I ever heard back. I'd prefer to see the -O0 and KGDB_DEBUG code for tracing the kgdb stub assimilated. If they would be accepted I could make a patch to Tom's git repository... I think it's an obsession with performance. As long as I/we can map "static inline" to "static" it's not a big deal. I'm using it every day; works great. I also recommend by SOCK_DEBUG, SKB_DEBUG, and TCP_DEBUG macros to trace the TCP code. I also indent the trace to make it easy to read. function1() { function2() { function3() { function4(); } } } The brackets make it easy to see the scope of the trace with vi. I like tracing with 'C' syntax since it what the reader is use to. If folks are interested I could also add that to the git diff, but I think that likely belongs else where and isn't likely the current dogma. See snippet from attached network trace. I gave a talk at a UNENIX conference back in the 1980 recommending a common UNIX tracing paradigm and a few liked it. The director of Siemens, Struck Zimmerman, didn't; you can't please everyone, so I just do what I think is best and live with the world not being as I'd expect it to be. For TCP I'm using the attached sock.h fragment which has a backward compatible SOCK_DEBUG() macro. I used the same paradigm in skbuff.h; see attachment. Likewise I'm doing the same in kgdb.h; also attached. In printk I added: -------------------------------------------------------------------- for (tp = tbuf; tp < tbuf + tlen; tp++) emit_log_char(*tp); printed_len += tlen - 3; #ifdef CONFIG_PRINTK_INDENT if (!in_interrupt()) { int depth = stack_depth(); int i; if ((depth > 0) && (depth < 120)) { for(i = 0; i < depth; i++) { emit_log_char(' '); printed_len++; } } } #endif ----------------------------------------------------------- and I added stack_depth() function to ../linux-4/arch/i386/kernel/traps.c ----------------------------------------------------------- int stack_depth(void) { struct thread_info *tinfo; unsigned long ebp; int depth = 0; #ifdef CONFIG_FRAME_POINTER asm("andl %%esp,%0; ":"=r" (tinfo) : "0" (~(THREAD_SIZE - 1))); asm ("movl %%ebp, %0" : "=r" (ebp) : ); while (valid_stack_ptr(tinfo, (void *)ebp)) { ebp = *(unsigned long *)ebp; if (depth++ > 100) { break; } } #endif return(depth); } --------------------------------------------------------------------------- Let me know if you you have any questions. Sounds like your on the right track; IMHO. -piet -- Piet Delaney Phone: (408) 200-5256 Blue Lane Technologies Fax: (408) 200-5299 10450 Bubb Rd. Cupertino, Ca. 95014 Email: piet(a)bluelane.com