> On Aug 26, 2019, at 2:23 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> So only the high mapping is ever executable; the identity map should not
> be. Both should be RO.
>
>> kprobe (with CONFIG_KPROBES_ON_FTRACE) should work on kernel identity
>> mapping.
>
> Please provide more information; kprobes shouldn't be touching either
> mapping. That is, afaict kprobes uses text_poke() which uses a temporary
> mapping (in 'userspace' even) to alias the high text mapping.
kprobe without CONFIG_KPROBES_ON_FTRACE uses text_poke(). But kprobe with
CONFIG_KPROBES_ON_FTRACE uses another path. The split happens with
set_kernel_text_rw() -> ... -> __change_page_attr() -> split_large_page().
The split is introduced by commit 585948f4f695. do_split in
__change_page_attr() becomes true after commit 585948f4f695. This patch
tries to fix/workaround this part.
>
> I'm also not sure how it would then result in any 4k text maps. Yes the
> alias is 4k, but it should not affect the actual high text map in any
> way.
I am confused by the alias logic. set_kernel_text_rw() makes the high map
rw, and split the PMD in the high map.
>
> kprobes also allocates executable slots, but it does that in the module
> range (afaict), so that, again, should not affect the high text mapping.
>
>> We found with 5.2 kernel (no CONFIG_PAGE_TABLE_ISOLATION, w/
>> CONFIG_KPROBES_ON_FTRACE), a single kprobe will split _all_ PMDs in
>> kernel text mapping into pte-mapped pages. This increases iTLB
>> miss rate from about 300 per million instructions to about 700 per
>> million instructions (for the application I test with).
>>
>> Per bisect, we found this behavior happens after commit 585948f4f695
>> ("x86/mm/cpa: Avoid the 4k pages check completely"). That's why I
>> proposed this PATCH to fix/workaround this issue. However, per
>> Peter's comment and my study of the code, this doesn't seem the
>> real problem or the only here.
>>
>> I also tested that the PMD split issue doesn't happen w/o
>> CONFIG_KPROBES_ON_FTRACE.
>
> Right, because then ftrace doesn't flip the whole kernel map writable;
> which it _really_ should stop doing anyway.
>
> But I'm still wondering what causes that first 4k split...
Please see above.
Thanks,
Song
