On 12.10.19 00:16, Andrew Morton wrote:
On Thu, 10 Oct 2019 16:12:00 +0200 David Hildenbrand <david@xxxxxxxxxx> wrote:Uninitialized memmaps contain garbage and in the worst case trigger kernel BUGs, especially with CONFIG_PAGE_POISONING. They should not get touched. Right now, when trying to soft-offline a PFN that resides on a memory block that was never onlined, one gets a misleading error with CONFIG_PAGE_POISONING: :/# echo 5637144576 > /sys/devices/system/memory/soft_offline_page [ 23.097167] soft offline: 0x150000 page already poisoned But the actual result depends on the garbage in the memmap. soft_offline_page() can only work with online pages, it returns -EIO in case of ZONE_DEVICE. Make sure to only forward pages that are online (iow, managed by the buddy) and, therefore, have an initialized memmap. Add a check against pfn_to_online_page() and similarly return -EIO. Fixes: f1dd2cd13c4b ("mm, memory_hotplug: do not associate hotadded memory to zones until online") # visible after d0dc12e86b319Should this be cc:stable?
I think yes, more on that below.
What is the relationship between this and some similar fixes in the series "mm/memory_hotplug: Shrink zones before removing memory", v6?
In general, they all have the same root cause. With f1dd2cd13c4b, we started to initialize the memmap when onlining memory, however, we at least zeroed it out when adding the memory. With d0dc12e86b319 we removed the zeroing, and added conditional poisoning instead.
All these BUGs can be reproduced by adding memory and keeping some memory blocks offline. Most distributions either online memory directly in the kernel when added or userspace onlines it via udev rules. s390x is special, because there we don't online memory blocks as default in user space. So on !s390x systems, these BUGs are quite hard to reproduce.
With "mm/memory_hotplug: Shrink zones before removing memory" these BUGs get easier to reproduce, because it is now sufficient to offline a memory block that was already onlined.
Also, devmem with "driver reserved memory" (for which part we don't initialize the memmap) is able to trigger these BUGs, but that feature is more recent AFAIK.
So, cc:stable, I am not sure if it applies to all patches. Some really only trigger when page poisoning is active, but don't result in any damage (as so far observed). We really produce damage in case we de-reference the NID/zone via the garbage memmap (and probably when doing a page_to_pfn(pfn_to_page(gargabe_page))).
But here, it is quite hard to tell what could happen, so I guess, if in doubt, better add cc:stable?
Should any of the patches in "mm/memory_hotplug: Shrink zones before removing memory", v6 be cc:stable?
I'll go over all patches and reply to the relevant ones. So for this patch, please add: Cc: stable@xxxxxxxxxxxxxxx # v4.13+ -- Thanks, David / dhildenb
