| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Date: Fri, 11 Feb 2022 09:25:00 +0100
From: Dominik Brodowski <linux@...inikbrodowski.net>
To: Sebastian Andrzej Siewior <bigeasy@...utronix.de>
Cc: "Jason A. Donenfeld" <Jason@...c4.com>,
linux-kernel@...r.kernel.org, linux-crypto@...r.kernel.org,
Thomas Gleixner <tglx@...utronix.de>,
Peter Zijlstra <peterz@...radead.org>,
Theodore Ts'o <tytso@....edu>,
Sultan Alsawaf <sultan@...neltoast.com>,
Jonathan Neuschäfer <j.neuschaefer@....net>,
Eric Biggers <ebiggers@...nel.org>
Subject: Re: [PATCH v4 2/2] random: defer fast pool mixing to worker
Am Thu, Feb 10, 2022 at 07:04:20PM +0100 schrieb Sebastian Andrzej Siewior:
> > @@ -999,9 +1016,10 @@ void add_interrupt_randomness(int irq)
> >
> > fast_mix(fast_pool);
> > add_interrupt_bench(cycles);
> > + new_count = ++fast_pool->count;
> >
> > if (unlikely(crng_init == 0)) {
> > - if ((fast_pool->count >= 64) &&
> > + if (new_count >= 64 &&
> > crng_fast_load((u8 *)fast_pool->pool, sizeof(fast_pool->pool)) > 0) {
>
> crng_fast_load() does spin_trylock_irqsave() in hardirq context. It does
> not produce any warning on RT but is still wrong IMHO:
> - lockdep will see a random task and I remember in the past it produced
> strange lock chains based on this.
>
> - Should another task attempt to acquire this lock then it will PI-boost the
> wrong task.
>
> If we just could move this, too.
>
> I don't know how timing critical this is but the first backtrace from
> crng_fast_load() came (to my surprise) from hwrng_fillfn() (a kthread)
> and added 64bytes in one go.
That's a hw rng (such as a tpm chip or the virtio-rng driver) providing some
entropy; if it's 64 bytes of input, crng_init progresses to 1, and
crng_fast_load() should never be called again.[*] I'm a bit suprised that the
hw_rng input occurred so early (it's only at device_initcall() level), and
earlier than 64 interrupts. But that may differ from system to system.
Note that crng_fast_load() will also never be called from
add_interrupt_randomness() if
EFI, DT or kexec provides bootloader entropy of at least 64 bytes,
and CONFIG_RANDOM_TRUST_BOOTLOADER is set
and/or CONFIG_RANDOM_TRUST_CPU is set and the RDRAND/RDSEED instructions do
not fail.
If neither of these three conditions (hw_rng is run early, bootloader or CPU
randomness) are met, the initial and early seeding of the base_crng depends
on add_interrupt_randomness(), and should happen rather quickly.
> I did move that crng_fast_load() into the worker and did made some
> numbers:
> <idle>-0 [000] d..h1.. 2.069924: add_interrupt_randomness: Tick
>
> first interrupt
> …
> swapper/0-1 [000] d..h.11 2.341938: add_interrupt_randomness: Tick
> swapper/0-1 [000] d..h.11 2.341938: add_interrupt_randomness: work
>
> the 64th interrupt, scheduling the worker.
>
> swapper/0-1 [000] d..h.11 2.345937: add_interrupt_randomness: Tick
> swapper/0-1 [000] d..h111 2.349938: add_interrupt_randomness: Tick
> swapper/0-1 [000] d..h.11 2.353939: add_interrupt_randomness: Tick
> swapper/0-1 [000] d..h.11 2.357940: add_interrupt_randomness: Tick
> swapper/0-1 [000] d..h111 2.361939: add_interrupt_randomness: Tick
> swapper/0-1 [000] d..h111 2.365939: add_interrupt_randomness: Tick
> swapper/0-1 [000] d..h.11 2.369941: add_interrupt_randomness: Tick
> kworker/0:0H-6 [000] ....... 2.384714: mix_interrupt_randomness: load
> kworker/0:0H-6 [000] ....... 2.384715: crng_fast_load: 16
> <idle>-0 [001] dn.h1.. 3.205766: add_interrupt_randomness: Tick
> <idle>-0 [019] dn.h1.. 6.771047: add_interrupt_randomness: Tick
>
> 7 interrupts got lost before the worker could run & load first 16 bytes.
> The workqueue core gets initialized at that point and spawns first
> worker.
So the reason for the longer delay here is that the workqueue core had not
been initialized beforehand?
> After that the interrupts took a break.
> And then the work-to-load delay was quite low:
>
> <idle>-0 [019] dn.h1.. 7.586234: add_interrupt_randomness: Tick
> <idle>-0 [019] dn.h1.. 7.586234: add_interrupt_randomness: work
> kworker/19:0H-175 [019] ....... 7.586504: mix_interrupt_randomness: load
> kworker/19:0H-175 [019] ....... 7.586507: crng_fast_load: 16
> <idle>-0 [020] dn.h1.. 7.614649: add_interrupt_randomness: Tick
> <idle>-0 [020] dn.h1.. 7.614651: add_interrupt_randomness: work
> <idle>-0 [020] dn.h1.. 7.614736: add_interrupt_randomness: Tick
> kworker/20:0H-183 [020] dn.h... 7.614859: add_interrupt_randomness: Tick
> kworker/20:0H-183 [020] ....... 7.614871: mix_interrupt_randomness: load
> kworker/20:0H-183 [020] ....... 7.614872: crng_fast_load: 16
> <idle>-0 [018] dn.h1.. 8.352423: add_interrupt_randomness: Tick
> <idle>-0 [018] dn.h1.. 8.352423: add_interrupt_randomness: work
> kworker/18:0H-167 [018] dn.h1.. 8.352438: add_interrupt_randomness: Tick
> kworker/18:0H-167 [018] dn.h1.. 8.352448: add_interrupt_randomness: Tick
> kworker/18:0H-167 [018] dn.h1.. 8.352459: add_interrupt_randomness: Tick
> kworker/18:0H-167 [018] dn.h1.. 8.352491: add_interrupt_randomness: Tick
> kworker/18:0H-167 [018] ....... 8.352505: mix_interrupt_randomness: load
> kworker/18:0H-167 [018] ....... 8.352506: crng_fast_load: 16
>
> In total we lost 13 ticks.
Was this still way before the initramfs was up and running?
> I did the same test on PREEMPT_VOLUNTARY and lost 2 ticks only.
Thanks,
Dominik
[*] Actually, there's some contradiciton going on: If we do not trust the
hw_rng device (that is, its quality setting is 0), crng_fast_load() will be
called nonetheless, and the hw_rng-provided input will be used to increment
crng_init to 1. If !CONFIG_RANDOM_TRUST_BOOTLOADER, only crng_slow_load() is
called, and crng_init will remain at 0. Similar for
!CONFIG_RANDOM_TRUST_CPU.
Powered by blists - more mailing lists