Re: [PATCH net-next 02/15] net: dsa: sja1105: Remove usage of iterator for list_add() after loop

From: Jakob Koschel
Date: Sun Apr 10 2022 - 14:24:46 EST




> On 10. Apr 2022, at 14:39, Jakob Koschel <jakobkoschel@xxxxxxxxx> wrote:
>
>
>
>> On 10. Apr 2022, at 13:05, Vladimir Oltean <olteanv@xxxxxxxxx> wrote:
>>
>> On Sun, Apr 10, 2022 at 12:51:56PM +0200, Jakob Koschel wrote:
>>> I've just looked at this again in a bit more detail while integrating it into the patch series.
>>>
>>> I realized that this just shifts the 'problem' to using the 'pos' iterator variable after the loop.
>>> If the scope of the list iterator would be lowered to the list traversal loop it would also make sense
>>> to also do it for list_for_each().
>>
>> Yes, but list_for_each() was never formulated as being problematic in
>> the same way as list_for_each_entry(), was it? I guess I'm starting to
>> not understand what is the true purpose of the changes.
>
> Sorry for having caused the confusion. Let me elaborate a bit to give more context.
>
> There are two main benefits of this entire effort.
>
> 1) fix the architectural bugs and avoid any missuse of the list iterator after the loop
> by construction. This only concerns the list_for_each_entry_*() macros and your change
> will allow lowering the scope for all of those. It might be debatable that it would be
> more consistent to lower the scope for list_for_each() as well, but it wouldn't be
> strictly necessary.
>
> 2) fix *possible* speculative bugs. In our project, Kasper [1], we were able to show
> that this can be an issue for the list traversal macros (and this is how the entire
> effort started).
> The reason is that the processor might run an additional loop iteration in speculative
> execution with the iterator variable computed based on the head element. This can
> (and we have verified this) happen if the CPU incorrectly
> assumes !list_entry_is_head(pos, head, member).
>
> If this happens, all memory accesses based on the iterator variable *potentially* open
> the chance for spectre [2] gadgets. The proposed mitigation was setting the iterator variable
> to NULL when the terminating condition is reached (in speculative safe way). Then,
> the additional speculative list iteration would still execute but won't access any
> potential secret data.
>
> And this would also be required for list_for_each() since combined with the list_entry()
> within the loop it basically is semantically identical to list_for_each_entry()
> for the additional speculative iteration.
>
> Now, I have no strong opinion on going all the way and since 2) is not the main motivation
> for this I'm also fine with sticking to your proposed solution, but it would mean that implementing
> a "speculative safe" list_for_each() will be more difficult in the future since it is using
> the iterator of list_for_each() past the loop.
>
> I hope this explains the background a bit better.
>
>>
>>> What do you think about doing it this way:
>>>
>>> diff --git a/drivers/net/dsa/sja1105/sja1105_vl.c b/drivers/net/dsa/sja1105/sja1105_vl.c
>>> index b7e95d60a6e4..f5b0502c1098 100644
>>> --- a/drivers/net/dsa/sja1105/sja1105_vl.c
>>> +++ b/drivers/net/dsa/sja1105/sja1105_vl.c
>>> @@ -28,6 +28,7 @@ static int sja1105_insert_gate_entry(struct sja1105_gating_config *gating_cfg,
>>> list_add(&e->list, &gating_cfg->entries);
>>> } else {
>>> struct sja1105_gate_entry *p;
>>> + struct list_head *pos = NULL;
>>>
>>> list_for_each_entry(p, &gating_cfg->entries, list) {
>>> if (p->interval == e->interval) {
>>> @@ -37,10 +38,14 @@ static int sja1105_insert_gate_entry(struct sja1105_gating_config *gating_cfg,
>>> goto err;
>>> }
>>>
>>> - if (e->interval < p->interval)
>>> + if (e->interval < p->interval) {
>>> + pos = &p->list;
>>> break;
>>> + }
>>> }
>>> - list_add(&e->list, p->list.prev);
>>> + if (!pos)
>>> + pos = &gating_cfg->entries;
>>> + list_add(&e->list, pos->prev);
>>> }
>>>
>>> gating_cfg->num_entries++;
>>> --
>>>
>>>>
>>>> Thanks for the suggestion.
>>>>
>>>>> }
>>>>>
>>>>> gating_cfg->num_entries++;
>>>>> -----------------------------[ cut here ]-----------------------------
>>>>
>>>> [1] https://lore.kernel.org/linux-kernel/20220407102900.3086255-12-jakobkoschel@xxxxxxxxx/
>>>>
>>>> Jakob
>>>
>>> Thanks,
>>> Jakob
>
> Thanks,
> Jakob
>
> [1] https://www.vusec.net/projects/kasper/
> [2] https://spectreattack.com/spectre.pdf


Btw, I just realized that the if (!pos) is not necessary. This should simply do it:

diff --git a/drivers/net/dsa/sja1105/sja1105_vl.c b/drivers/net/dsa/sja1105/sja1105_vl.c
index b7e95d60a6e4..2d59e75a9e3d 100644
--- a/drivers/net/dsa/sja1105/sja1105_vl.c
+++ b/drivers/net/dsa/sja1105/sja1105_vl.c
@@ -28,6 +28,7 @@ static int sja1105_insert_gate_entry(struct sja1105_gating_config *gating_cfg,
list_add(&e->list, &gating_cfg->entries);
} else {
+ struct list_head *pos = &gating_cfg->entries;
struct sja1105_gate_entry *p;

list_for_each_entry(p, &gating_cfg->entries, list) {
if (p->interval == e->interval) {
@@ -37,10 +38,12 @@ static int sja1105_insert_gate_entry(struct sja1105_gating_config *gating_cfg,
goto err;
}

- if (e->interval < p->interval)
+ if (e->interval < p->interval) {
+ pos = &p->list;
break;
+ }
}
- list_add(&e->list, p->list.prev);
+ list_add(&e->list, pos->prev);
}

gating_cfg->num_entries++;
--
2.25.1

Thanks,
Jakob