Re: [PATCH] mm/vmalloc: do not warn on -ENOMEM from va_clip() in pcpu_get_vm_areas()

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 25 Sep 2026 21:54:49 +0100 Palla Raghunath <raghunathpalla.0209@gmail.com> wrote:

> When pcpu_get_vm_areas() has to split a free vmap_area in the middle
> (NE_FIT_TYPE), va_clip() needs an extra vmap_area object. It takes the
> per-cpu ne_fit_preload_node if one is there, and otherwise falls back
> to kmem_cache_alloc(GFP_NOWAIT), which may fail and return -ENOMEM.
> pcpu_get_vm_areas() never preloads, and a single call can do more than
> one such split: on a NUMA system it places one area per node group, so
> the first split consumes the preloaded object and the next one depends
> on the GFP_NOWAIT allocation.
> 
> That failure is expected and already handled: the recovery path returns
> the areas clipped so far to the free tree, purges lazily freed areas and
> retries. But the error is checked with WARN_ON_ONCE(), so a transient
> allocation failure under memory pressure or fault injection triggers a
> kernel warning, and a panic with panic_on_warn. syzbot hit this on a
> two-node VM while creating a per-cpu BPF array map.
> 
> Keep the WARN_ON_ONCE() for errors other than -ENOMEM, which do indicate
> a bug, and take the recovery path either way. This matches what commit
> b9183788a2de ("mm/vmalloc: do not warn on -ENOMEM from va_alloc()") did
> for the other va_clip() caller.
> 
> Fixes: 1b23ff80b399 ("mm/vmalloc: invoke classify_va_fit_type() in adjust_va_to_fit_type()")
> Reported-by: syzbot+442828bb356b10813a47@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=442828bb356b10813a47

I assume sysbot hit this via fault injection.    The report doesn't make this
info easily available.  Or maybe it wasn't fault injection.

> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -5107,9 +5107,14 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
>  
>  		ret = va_clip(&free_vmap_area_root,
>  			&free_vmap_area_list, va, start, size);
> -		if (WARN_ON_ONCE(unlikely(ret)))
> -			/* It is a BUG(), but trigger recovery instead. */
> +		if (unlikely(ret)) {
> +			/*
> +			 * -ENOMEM from the GFP_NOWAIT fallback is expected.
> +			 * Anything else is a BUG(), but trigger recovery instead.
> +			 */
> +			WARN_ON_ONCE(ret != -ENOMEM);

Look good.

The WARN_ON_ONCE() is potentially redundant.  We could make va_clip()
remove __GFP_NOWARN from its kmem_cache_alloc() and let the page
allocator do the warn for us.  Doesn't matter.

>  			goto recovery;
> +		}
>  
>  		/* Allocated area. */
>  		va = vas[area];
> -- 
> 2.34.1