Re: [dpdk-dev] [PATCH dpdk-dev] rte_random: fix crash when random init

["Mattias Rönnblom" <mattias.ronnblom@ericsson.com>]

On 2020-04-14 15:35, David Marchand wrote:
> On Tue, Apr 14, 2020 at 3:20 PM Mattias Rönnblom
> <mattias.ronnblom@ericsson.com> wrote:
>> On 2020-04-14 06:43, Tonghao Zhang wrote:
>>> On Tue, Apr 14, 2020 at 12:07 PM Stephen Hemminger
>>> <stephen@networkplumber.org> wrote:
>>>> On Sun, 12 Apr 2020 16:27:53 +0800
>>>> xiangxia.m.yue@gmail.com wrote:
>>>>
>>>>> From: Tonghao Zhang <xiangxia.m.yue@gmail.com>
>>>>>
>>>>> When rte_rand_init is invoked, and the kernel
>>>>> (kernel version < 3.17) running dpdk does't support
>>>>> *getentropy, at the same time, the cpu does't support
>>>>> rdseed, the rte_rand_init will invoke rte_get_timer_cycles
>>>>> which function will invoke rte_get_hpet_cycles
>>>>> (RTE_LIBEAL_USE_HPET was enabled) while *eal_hpet is not
>>>>> allocated.
>>>>>
>>>>> Fixes: faf8fd252785 ("eal: improve entropy for initial PRNG seed")
>>>>> Fixes: 3f002f069612 ("eal: replace libc-based random generation with LFSR")
>>>>>
>>>>> Cc: stable@dpdk.org
>>>>>
>>>>> Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
>>>> Are you sure this patch won't change current default to use HPET (which is slower)?
>>> In rte_eal_timer_init (linux/eal_timer.c)， it will set
>>> eal_timer_source = EAL_TIMER_TSC too.
>>> So after rte_eal_init, eal_timer_source == EAL_TIMER_TSC which is the
>>> default timer source actually.
>>> Then this patch will affect RTE_INIT function which invoke
>>> rte_get_timer_cycles. but hpet is not available yet.
>>
>> Would using rte_rdtsc() directly be an option?
> s/rte_rdtsc/rte_get_tsc_cycles/
>
> This could work, but I am a bit surprised to see an initialisation in
> a constructor.
> The commitlog that moved rte_srand() from rte_eal_init does not
> explain why it was moved.
>
>
The initialization (i.e. automatic seeding) grew in complexity somewhat, 
and with the new rte_random.c file, it felt like it would have a good, 
new home.


That said, maybe it would have been better to add an initialization 
function to the rte_random.h API, and have it called from 
rte_eal_init(), to avoid the ordering issues with constructors. Yet 
another alternative would be to just move the seeding logic.