Re: [PATCH v3 1/1] kasan: Avoid sleepable page allocation from atomic context

[Alexander Gordeev <agordeev@linux.ibm.com>]

On Tue, May 06, 2025 at 04:55:20PM +0200, Andrey Ryabinin wrote:
> >>> -	if (likely(pte_none(ptep_get(ptep)))) {
> >>> +	if (likely(pte_none(ptep_get(ptep))))
> >>>  		set_pte_at(&init_mm, addr, ptep, pte);
> >>> -		page = 0;
> >>
> >> With this patch, now if the pte is already set, the page is leaked?
> > 
> > Yes. But currently it is leaked for previously allocated pages anyway,
> > so no change in behaviour (unless I misread the code).
> 
> Current code doesn't even allocate page if pte set, and if set pte discovered only after
> taking spinlock, the page will be freed, not leaked.

Oh, right. I rather meant pages that are leaked in case of a failure. My bad.

> Whereas, this patch leaks page for every single !pte_none case. This will build up over time
> as long as vmalloc called.
> 
> > 
> >> Should we set data->pages[PFN_DOWN(addr - data->start)] = NULL 
> >> and free non-null elements later in __kasan_populate_vmalloc()?
> > 
> > Should the allocation fail on boot, the kernel would not fly anyway.
> 
> This is not boot code, it's called from vmalloc() code path.

FWIW, it is called from rest_init() too.

> > If for whatever reason we want to free, that should be a follow-up
> > change, as far as I am concerned.
> > 
> We want to free it, because we don't want unbound memory leak.

Will send v5.

Thanks!