[PATCH] mm, kasan: don't poison boot memory

[George Kennedy]

On 2/18/2021 3:55 AM, David Hildenbrand wrote:
> On 17.02.21 21:56, Andrey Konovalov wrote:
>> During boot, all non-reserved memblock memory is exposed to the buddy
>> allocator. Poisoning all that memory with KASAN lengthens boot time,
>> especially on systems with large amount of RAM. This patch makes
>> page_alloc to not call kasan_free_pages() on all new memory.
>>
>> __free_pages_core() is used when exposing fresh memory during system
>> boot and when onlining memory during hotplug. This patch adds a new
>> FPI_SKIP_KASAN_POISON flag and passes it to __free_pages_ok() through
>> free_pages_prepare() from __free_pages_core().
>>
>> This has little impact on KASAN memory tracking.
>>
>> Assuming that there are no references to newly exposed pages before they
>> are ever allocated, there won't be any intended (but buggy) accesses to
>> that memory that KASAN would normally detect.
>>
>> However, with this patch, KASAN stops detecting wild and large
>> out-of-bounds accesses that happen to land on a fresh memory page that
>> was never allocated. This is taken as an acceptable trade-off.
>>
>> All memory allocated normally when the boot is over keeps getting
>> poisoned as usual.
>>
>> Signed-off-by: Andrey Konovalov <andreyknvl at google.com>
>> Change-Id: Iae6b1e4bb8216955ffc14af255a7eaaa6f35324d
>
> Not sure this is the right thing to do, see
>
> https://lkml.kernel.org/r/bcf8925d-0949-3fe1-baa8-cc536c529860@oracle.com
>
> Reversing the order in which memory gets allocated + used during boot 
> (in a patch by me) might have revealed an invalid memory access during 
> boot.
>
> I suspect that that issue would no longer get detected with your 
> patch, as the invalid memory access would simply not get detected. 
> Now, I cannot prove that :)

Since David's patch we're having trouble with the iBFT ACPI table, which 
is mapped in via kmap() - see acpi_map() in "drivers/acpi/osl.c". KASAN 
detects that it is being used after free when ibft_init() accesses the 
iBFT table, but as of yet we can't find where it get's freed (we've 
instrumented calls to kunmap()).

Thank you,
George