Re: [PATCH] x86/mm: Handle alloc failure in phys_*_init()

[Ingo Molnar <mingo@kernel.org>]

* Em Sharnoff <sharnoff@neon.tech> wrote:

> tl;dr:
> 
> * When setting up page table mappings for physical addresses after boot,
>   alloc_low_page() uses GFP_ATOMIC, which is allowed to fail.
> * This isn't currently handled, and results in a null pointer
>   dereference when it occurs.
> * This allocation failure can happen during memory hotplug.
> 
> To handle failure, change phys_pud_init() and similar functions to
> return zero if allocation failed (either directly or transitively), and
> convert that to -ENOMEM in arch_add_memory().

> +		/*
> +		 * Bail only after updating pgd/p4d to keep progress from p4d across retries.
> +		 */
> +		if (!paddr_last)
> +			return 0;
> +
>  		pgd_changed = true;

> -	init_memory_mapping(start, start + size, params->pgprot);
> +	if (!init_memory_mapping(start, start + size, params->pgprot))
> +		return -ENOMEM;

I agree that it makes total sense to fix all this (especially since you 
are actively triggering it), but have you tried also changing it away 
from GFP_ATOMIC? There's no real reason why it should be GFP_ATOMIC 
AFAICS, other than some historic inertia that nobody bothered to fix.

Plus, could you please change the return flow from this zero 
special-case over to something like ERR_PTR(-ENOMEM) and IS_ERR()?

*Technically* zero is a valid physical address, although we 
intentionally never use it in the kernel AFAIK and wouldn't ever put a 
page table there either. ERR_PTR()/IS_ERR() is much easier on the eyes 
than the zero special-case.

Finally, could you make this a 2-patch fix series: first one to fix the 
error return path to not crash, and the second one to change it away 
from GFP_ATOMIC?

Thanks,

	Ingo