On 26.03.21 15:38, Michal Hocko wrote:

On Fri 26-03-21 09:52:58, David Hildenbrand wrote:
[...]

quoted

Something else to note:


We'll not call the memory notifier (e.g., MEM_ONLINE) for the vmemmap. The
result is that

1. We won't allocate extended struct pages for the range. Don't think this
is really problematic (pages are never allocated/freed, so I guess we don't
care - like ZONE_DEVICE code).

Agreed. I do not think we need them. Future might disagree but let's
handle it when we have a clear demand.

quoted

2. We won't allocate kasan shadow memory. We most probably have to do it
explicitly via kasan_add_zero_shadow()/kasan_remove_zero_shadow(), see
mm/memremap.c:pagemap_range()

I think this is similar to the above. Does kasan has to know about
memory which will never be used for anything?

IIRC, kasan will track read/writes to the vmemmap as well. So it could 
theoretically detect if we read from the vmemmap before writing 
(initializing) it IIUC.

This is also why mm/memremap.c does a kasan_add_zero_shadow() before the 
move_pfn_range_to_zone()->memmap_init_range() for the whole region, 
including altmap space.

Now, I am no expert on KASAN, what would happen in case we have access 
to non-tracked memory.

commit 0207df4fa1a869281ddbf72db6203dbf036b3e1a
Author: Andrey Ryabinin [off-list ref]
Date:   Fri Aug 17 15:47:04 2018 -0700

     kernel/memremap, kasan: make ZONE_DEVICE with work with KASAN

indicates that kasan will crash the system on "non-existent shadow memory"

quoted

Further a locking rework might be necessary. We hold the device hotplug
lock, but not the memory hotplug lock. E.g., for get_online_mems(). Might
have to move that out online_pages.

Could you be more explicit why this locking is needed? What it would
protect from for vmemmap pages?

One example is in mm/kmemleak.c:kmemleak_scan(), where we scan the 
vmemmap for pointers. We don't want the vmemmap to get unmapped while we 
are working on it (-> fault).

-- 
Thanks,

David / dhildenb