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

[David Marchand]

On Wed, Apr 15, 2020 at 8:01 AM Mattias Rönnblom
<mattias.ronnblom at ericsson.com> wrote:
>
> On 2020-04-15 01:42, Tonghao Zhang wrote:
> > On Tue, Apr 14, 2020 at 11:37 PM Mattias Rönnblom
> > <mattias.ronnblom at ericsson.com> wrote:
> >> On 2020-04-14 15:35, David Marchand wrote:
> >>> On Tue, Apr 14, 2020 at 3:20 PM Mattias Rönnblom
> >>> <mattias.ronnblom at ericsson.com> wrote:
> >>>> On 2020-04-14 06:43, Tonghao Zhang wrote:
> >>>>> On Tue, Apr 14, 2020 at 12:07 PM Stephen Hemminger
> >>>>> <stephen at networkplumber.org> wrote:
> >>>>>> On Sun, 12 Apr 2020 16:27:53 +0800
> >>>>>> xiangxia.m.yue at gmail.com wrote:
> >>>>>>
> >>>>>>> From: Tonghao Zhang <xiangxia.m.yue at 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 at dpdk.org
> >>>>>>>
> >>>>>>> Signed-off-by: Tonghao Zhang <xiangxia.m.yue at 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.
> > If we use that solution  we can register a callback which invoked when
> > almost resources are  available:
> > http://patches.dpdk.org/patch/68313/
>
>
> We might then instead have a new ordering issue, if we wait too long
> with rte_random initialization, since other initialization code might
> well use rte_rand(). It seems like the memory heap expansion code does
> currently. rte_rand() is a pretty basic function, that should be safe to
> call early during initialization, I think.
>
>
> I would suggest first switching to rte_get_tsc_cycles() and later
> potentially change to explicit (i.e. non-constructor), early, rte_random
> initialization.

Switching to rte_get_tsc_cycles seems fine to me.
Who will send the fix?


-- 
David Marchand