[Crash-utility] Re: crash 4.0-8.9 w/ 2.6.30-rc6
Dave Anderson
anderson at redhat.com
Thu May 28 12:59:01 UTC 2009
----- "Daisuke Nishimura" <nishimura at mxp.nes.nec.co.jp> wrote:
> Hi, Dave.
>
> It's a very and very dirty hack, but some of my colleagues say that
> a dump file of 2.6.30-rcX can be analyzed by this patch.
>
> I hope it would be some help for you.
>
> Thanks,
> Daisuke Nishimura.
Actually it's not so dirty at all -- it's pretty well done!
I was also tinkering with a few new schemes yesterday, and did
make it to the "crash> " prompt as well. But I was allowing the
kt->__per_cpu_offset[] array to be initialized in kernel_init()
(line 195) as is done for x86, althougyh I think now that I prefer
the way that you guys have done it.
I also was trying to come up with a clean way to deal with
the x86_64 "per_cpu__xxx" symbols being contained in the
in the crash kernel symbol list. With my test crash -- and
I presume the same with your patch -- the symbol list
look like this:
crash> sym -l
0 (D) __per_cpu_start
0 (D) per_cpu__irq_stack_union
4000 (D) per_cpu__gdt_page
5000 (d) per_cpu__exception_stacks
b000 (D) per_cpu__current_task
b008 (d) per_cpu__is_idle
b010 (D) per_cpu__old_rsp
b018 (D) per_cpu__irq_regs
b020 (D) per_cpu__vector_irq
b420 (d) per_cpu__cpu_devices
b478 (D) per_cpu__cyc2ns
b480 (d) per_cpu__cpuid4_info
b488 (d) per_cpu__cache_kobject
b490 (d) per_cpu__index_kobject
b4a0 (D) per_cpu__irq_stack_ptr
b4a8 (D) per_cpu__kernel_stack
b4b0 (D) per_cpu__irq_count
b4c0 (D) per_cpu__orig_ist
... [ snip ] ...
10540 (D) per_cpu__irq_stat
10580 (D) per_cpu__cpu_info
10680 (D) per_cpu__cpu_tlbstate
106c0 (d) per_cpu__runqueues
10fc0 (d) per_cpu__sched_clock_data
10fe0 (D) __per_cpu_end
ffffffff80200000 (T) _text
ffffffff80200000 (T) startup_64
ffffffff802000b7 (t) ident_complete
...
The problem with having symbols below "_text" in this case
is that commands like "rd -s" or other memory dumps that
do symbolic translations of numeric values run into the
per_cpu symbols. So for example, a page of zeroes would
look like this:
crash> rd empty_zero_page 10
ffffffff80833000: 0000000000000000 0000000000000000 ................
ffffffff80833010: 0000000000000000 0000000000000000 ................
ffffffff80833020: 0000000000000000 0000000000000000 ................
ffffffff80833030: 0000000000000000 0000000000000000 ................
ffffffff80833040: 0000000000000000 0000000000000000 ................
crash> rd -s empty_zero_page 10
ffffffff80833000: per_cpu__irq_stack_union per_cpu__irq_stack_union
ffffffff80833010: per_cpu__irq_stack_union per_cpu__irq_stack_union
ffffffff80833020: per_cpu__irq_stack_union per_cpu__irq_stack_union
ffffffff80833030: per_cpu__irq_stack_union per_cpu__irq_stack_union
ffffffff80833040: per_cpu__irq_stack_union per_cpu__irq_stack_union
crash>
In fact, *any* value would translate to a symbol:
crash> rd ffff88007e4ee820 10
ffff88007e4ee820: 0000000000400000 0000000000409000 .. at .......@.....
ffff88007e4ee830: ffff88007e4eead8 0000000000000025 ..N~....%.......
ffff88007e4ee840: 0000000008001875 ffff88007e4eeb09 u.........N~....
ffff88007e4ee850: 0000000000000000 0000000000000000 ................
ffff88007e4ee860: ffff88007e4ee860 ffff88007e4ee860 `.N~....`.N~....
crash> rd -s ffff88007e4ee820 10
ffff88007e4ee820: __per_cpu_end+4124704 __per_cpu_end+4161568
ffff88007e4ee830: __per_cpu_end+-131939276301576 per_cpu__irq_stack_union+37
ffff88007e4ee840: __per_cpu_end+134154389 __per_cpu_end+-131939276301527
ffff88007e4ee850: per_cpu__irq_stack_union per_cpu__irq_stack_union
ffff88007e4ee860: __per_cpu_end+-131939276302208 __per_cpu_end+-131939276302208
crash>
So clearly there will have to be special handling for those
symbols, because sometimes a caller will *want* to receive
information re: those symbols, and other times will not.
(i.e. in symbol_exists(), value_search(), etc.) It was
never a problem in the past because the per_cpu__xxx
symbols were not "real" symbol values, but they were
very large numberic values in the kernel symbol list.
So in any case, we're on the same page, and I do like
what you've done. Thank you very much for having already
looked into this!
Dave
>
> Signed-off-by: Daisuke Nishimura <nishimura at mxp.nes.nec.co.jp>
> ---
> defs.h | 1 +
> x86_64.c | 78
> ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++----
> 2 files changed, 74 insertions(+), 5 deletions(-)
>
> diff --git a/defs.h b/defs.h
> index e5e3538..32b2e1e 100755
> --- a/defs.h
> +++ b/defs.h
> @@ -3851,6 +3851,7 @@ struct machine_specific {
> #define PHYS_BASE (0x80)
> #define VM_XEN_RHEL4 (0x100)
> #define VMEMMAP (0x200)
> +#define KSYMS_PERCPU (0x400)
>
> #define VM_FLAGS (VM_ORIG|VM_2_6_11|VM_XEN|VM_XEN_RHEL4)
>
> diff --git a/x86_64.c b/x86_64.c
> index 993827e..4813f96 100755
> --- a/x86_64.c
> +++ b/x86_64.c
> @@ -71,6 +71,7 @@ static int x86_64_is_uvaddr(ulong, struct
> task_context *);
> void x86_64_compiler_warning_stub(void);
> static void x86_64_init_kernel_pgd(void);
> static void x86_64_cpu_pda_init(void);
> +static void x86_64_per_cpu_init(void);
> static void x86_64_ist_init(void);
> static void x86_64_post_init(void);
> static void parse_cmdline_arg(void);
> @@ -290,7 +291,10 @@ x86_64_init(int when)
> MEMBER_OFFSET_INIT(user_regs_struct_ss,
> "user_regs_struct", "ss");
> STRUCT_SIZE_INIT(user_regs_struct, "user_regs_struct");
> - x86_64_cpu_pda_init();
> + if (STRUCT_EXISTS("x8664_pda"))
> + x86_64_cpu_pda_init();
> + else
> + x86_64_per_cpu_init();
> x86_64_ist_init();
> if ((machdep->machspec->irqstack = (char *)
> malloc(machdep->machspec->stkinfo.isize)) == NULL)
> @@ -677,6 +681,51 @@ x86_64_cpu_pda_init(void)
> FREEBUF(cpu_pda_buf);
> }
>
> +static void
> +x86_64_per_cpu_init(void)
> +{
> + int i, cpus, cpunumber;
> + ulong istacksize;
> +
> + if (!(kt->flags & PER_CPU_OFF))
> + return;
> +
> + if (!symbol_exists("per_cpu__cpu_number") || !symbol_exists("per_cpu__irq_stack_ptr"))
> + return;
> +
> + for (i = cpus = 0; i < NR_CPUS; i++) {
> + readmem(symbol_value("per_cpu__cpu_number") + kt->__per_cpu_offset[i],
> + KVADDR, &cpunumber, sizeof(int),
> + "cpu number (per_cpu)", FAULT_ON_ERROR);
> + if (cpunumber != cpus)
> + break;
> + cpus++;
> +
> + readmem(symbol_value("per_cpu__irq_stack_ptr") + kt->__per_cpu_offset[i],
> + KVADDR, &machdep->machspec->stkinfo.ibase[i],
> + sizeof(ulong), "irq stack ptr (per cpu)", FAULT_ON_ERROR);
> + }
> +
> + istacksize = 16384; /* 16K */
> + machdep->machspec->stkinfo.isize = istacksize;
> +
> + /*
> + * Adjust the kernel top-of-stack values down to their base.
> + */
> + for (i = 0; i < NR_CPUS; i++) {
> + if (machdep->machspec->stkinfo.ibase[i])
> + machdep->machspec->stkinfo.ibase[i] -= (istacksize-64);
> + else
> + break;
> + }
> +
> + kt->cpus = cpus;
> + if (kt->cpus > 1)
> + kt->flags |= SMP;
> +
> + verify_spinlock();
> +}
> +
> /*
> * Gather the ist addresses for each CPU.
> */
> @@ -754,7 +803,7 @@ x86_64_ist_init(void)
> */
> sp = value_search(ms->stkinfo.ebase[0][0], &offset);
> if (!sp || offset || !STREQ(sp->name,
> "boot_exception_stacks")) {
> - if (symbol_value("boot_exception_stacks")) {
> + if (symbol_exists("boot_exception_stacks")) {
> error(WARNING,
> "cpu 0 first exception stack: %lx\n
> boot_exception_stacks: %lx\n\n",
> ms->stkinfo.ebase[0][0],
> @@ -1706,8 +1755,12 @@ x86_64_verify_symbol(const char *name, ulong
> value, char type)
> {
> if (STREQ(name, "_text") || STREQ(name, "_stext"))
> machdep->flags |= KSYMS_START;
> + if (STREQ(name, "__per_cpu_start"))
> + machdep->flags |= KSYMS_PERCPU;
> + if (STREQ(name, "__per_cpu_end"))
> + machdep->flags &= ~KSYMS_PERCPU;
>
> - if (!name || !strlen(name) || !(machdep->flags & KSYMS_START))
> + if (!name || !strlen(name) || !(machdep->flags & (KSYMS_START | KSYMS_PERCPU)))
> return FALSE;
> return TRUE;
> }
> @@ -3943,8 +3996,23 @@ x86_64_get_smp_cpus(void)
> char *cpu_pda_buf;
> ulong level4_pgt, cpu_pda_addr;
>
> - if (!VALID_STRUCT(x8664_pda))
> - return 1;
> + if (!VALID_STRUCT(x8664_pda)) {
> + if (!(kt->flags & PER_CPU_OFF))
> + return 1;
> +
> + if (!symbol_exists("per_cpu__cpu_number"))
> + return 1;
> +
> + for (i = cpus = 0; i < NR_CPUS; i++) {
> + readmem(symbol_value("per_cpu__cpu_number") + kt->__per_cpu_offset[i],
> + KVADDR, &cpunumber, sizeof(int),
> + "cpu number (per_cpu)",FAULT_ON_ERROR);
> + if (cpunumber != cpus)
> + break;
> + cpus++;
> + }
> + return cpus;
> + }
>
> cpu_pda_buf = GETBUF(SIZE(x8664_pda));
>
More information about the Crash-utility
mailing list