Re: [PATCH] garbage values in file /proc/net/sockstat

[Eric Dumazet <dada1@cosmosbay.com>]

Andi Kleen a écrit :
> On Monday 23 January 2006 12:21, pravin shelar wrote:
>> 	In 2.6.16-rc1-mm1, (for x86_64 arch) cpu_possible_map is not same 
>> as NR_CPUS (prefill_possible_map()). Therefore per cpu areas are allocated 
>> for cpu_possible cpus only (setup_per_cpu_areas()). This causes sockstat 
>> to return garbage value on x84_64 arch.
>>
>> So these per_cpu accesses are geting relocated (RELOC_HIDE) using
>> boot_cpu_pda[]->data_offset which is not initialized.
>>
>> There are other instances of same bug where per_cpu() macro is used
>> without cpu_possible() check. e.g. net/core/utils.c :: 
>> net_random_reseed(), net/core/dev.c :: net_dev_init(), etc.
>>
>> This patch fixes these bugs.
> 
> Thanks. Patches Look good.  Dave, can you push them for 2.6.16 still please?
> 

Shouldnt we force a page fault for not possible cpus in cpu_data
to catch all access to per_cpu(some_object, some_not_possible_cpu) ?

We can use a red zone big enough to hold the whole per_cpu data.

Something like :

file include/asm-x86_64/pgtable.h

#define CPUDATA_RED_ZONE 0xffff808000000000UL  /* start of percpu catcher */

file arch/x86_64/kernel/setup64.c

setup_per_cpu_areas(void)
{
...
     for (i = 0 ; i < NR_CPUS ; i++) {
         if (!cpu_possible(cpu))
             cpu_pda(i)->data_offset = CPUDATA_RED_ZONE - __per_cpu_start ;
     }
}

Eric