On 10/17/18 10:25 AM, Sébastien Boisvert wrote:
>> @@ -10,14 +11,18 @@ struct align_footer {
>> unsigned int offset;
>> };
>>
>> -void *fio_memalign(size_t alignment, size_t size)
>> +void *fio_memalign(size_t alignment, size_t size, bool shared)
>> {
>> struct align_footer *f;
>> void *ptr, *ret = NULL;
>>
>> assert(!(alignment & (alignment - 1)));
>>
>> - ptr = malloc(size + alignment + sizeof(*f) - 1);
>> + if (shared)
>> + ptr = smalloc(size + alignment + sizeof(*f) - 1);
>> + else
>> + ptr = malloc(size + alignment + sizeof(*f) - 1);
>
> Hello Vincent,
>
> In your patch serives version 2, you removed the memset() call because
> smalloc() is already zeroing memory, according to Jens.
>
> From the hunk above, the memory pointed by ptr is not zeroed, when
> being allocated by malloc.
>
> I don't think that malloc is guaranteed to return zeroed memory.
>
> Sure, if the malloc from the glibc just called brk() or mmap(), then,
> maybe those pages will be zeroed by the kernel, and the glibc malloc
> arena from which the chunk is drawn may be zeroed.
>
> But otherwise, if it's a malloc that stays in user-space (existing
> arenas can fulfill the call), then, it is not zeroed, I think.
malloc() isn't zeroed, the code was using calloc() before (which is fine.
--
Jens Axboe
