From: "Mike Rapoport (IBM)" <rppt@kernel.org>

vmalloc allocations with VM_ALLOW_HUGE_VMAP that do not explictly
specify node ID will use huge pages only if size_per_node is larger than
PMD_SIZE.
Still the actual allocated memory is not distributed between nodes and
there is no advantage in such approach.
On the contrary, BPF allocates PMD_SIZE * num_possible_nodes() for each
new bpf_prog_pack, while it could do with PMD_SIZE'ed packs.

Don't account for number of nodes for VM_ALLOW_HUGE_VMAP with
NUMA_NO_NODE and use huge pages whenever the requested allocation size
is larger than PMD_SIZE.

Signed-off-by: Mike Rapoport (IBM) <rppt@kernel.org>
---
 mm/vmalloc.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 22aa63f4ef63..5fc8b514e457 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -3737,8 +3737,6 @@ void *__vmalloc_node_range(unsigned long size, unsigned long align,
 	}
 
 	if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
-		unsigned long size_per_node;
-
 		/*
 		 * Try huge pages. Only try for PAGE_KERNEL allocations,
 		 * others like modules don't yet expect huge pages in
@@ -3746,13 +3744,10 @@ void *__vmalloc_node_range(unsigned long size, unsigned long align,
 		 * supporting them.
 		 */
 
-		size_per_node = size;
-		if (node == NUMA_NO_NODE)
-			size_per_node /= num_online_nodes();
-		if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
+		if (arch_vmap_pmd_supported(prot) && size >= PMD_SIZE)
 			shift = PMD_SHIFT;
 		else
-			shift = arch_vmap_pte_supported_shift(size_per_node);
+			shift = arch_vmap_pte_supported_shift(size);
 
 		align = max(real_align, 1UL << shift);
 		size = ALIGN(real_size, 1UL << shift);
-- 
2.43.0

Le 11/04/2024 à 18:05, Mike Rapoport a écrit :
> From: "Mike Rapoport (IBM)" <rppt@kernel.org>
> 
> vmalloc allocations with VM_ALLOW_HUGE_VMAP that do not explictly
> specify node ID will use huge pages only if size_per_node is larger than
> PMD_SIZE.
> Still the actual allocated memory is not distributed between nodes and
> there is no advantage in such approach.
> On the contrary, BPF allocates PMD_SIZE * num_possible_nodes() for each
> new bpf_prog_pack, while it could do with PMD_SIZE'ed packs.
> 
> Don't account for number of nodes for VM_ALLOW_HUGE_VMAP with
> NUMA_NO_NODE and use huge pages whenever the requested allocation size
> is larger than PMD_SIZE.

Patch looks ok but message is confusing. We also use huge pages at PTE 
size, for instance 512k pages or 16k pages on powerpc 8xx, while 
PMD_SIZE is 4M.

Christophe

> 
> Signed-off-by: Mike Rapoport (IBM) <rppt@kernel.org>
> ---
>   mm/vmalloc.c | 9 ++-------
>   1 file changed, 2 insertions(+), 7 deletions(-)
> 
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 22aa63f4ef63..5fc8b514e457 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -3737,8 +3737,6 @@ void *__vmalloc_node_range(unsigned long size, unsigned long align,
>   	}
>   
>   	if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
> -		unsigned long size_per_node;
> -
>   		/*
>   		 * Try huge pages. Only try for PAGE_KERNEL allocations,
>   		 * others like modules don't yet expect huge pages in
> @@ -3746,13 +3744,10 @@ void *__vmalloc_node_range(unsigned long size, unsigned long align,
>   		 * supporting them.
>   		 */
>   
> -		size_per_node = size;
> -		if (node == NUMA_NO_NODE)
> -			size_per_node /= num_online_nodes();
> -		if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
> +		if (arch_vmap_pmd_supported(prot) && size >= PMD_SIZE)
>   			shift = PMD_SHIFT;
>   		else
> -			shift = arch_vmap_pte_supported_shift(size_per_node);
> +			shift = arch_vmap_pte_supported_shift(size);
>   
>   		align = max(real_align, 1UL << shift);
>   		size = ALIGN(real_size, 1UL << shift);

On Fri, Apr 12, 2024 at 06:07:19AM +0000, Christophe Leroy wrote:
> 
> 
> Le 11/04/2024 à 18:05, Mike Rapoport a écrit :
> > From: "Mike Rapoport (IBM)" <rppt@kernel.org>
> > 
> > vmalloc allocations with VM_ALLOW_HUGE_VMAP that do not explictly
> > specify node ID will use huge pages only if size_per_node is larger than
> > PMD_SIZE.
> > Still the actual allocated memory is not distributed between nodes and
> > there is no advantage in such approach.
> > On the contrary, BPF allocates PMD_SIZE * num_possible_nodes() for each
> > new bpf_prog_pack, while it could do with PMD_SIZE'ed packs.
> > 
> > Don't account for number of nodes for VM_ALLOW_HUGE_VMAP with
> > NUMA_NO_NODE and use huge pages whenever the requested allocation size
> > is larger than PMD_SIZE.
> 
> Patch looks ok but message is confusing. We also use huge pages at PTE 
> size, for instance 512k pages or 16k pages on powerpc 8xx, while 
> PMD_SIZE is 4M.

Ok, I'll rephrase.
 
> Christophe
> 
> > 
> > Signed-off-by: Mike Rapoport (IBM) <rppt@kernel.org>
> > ---
> >   mm/vmalloc.c | 9 ++-------
> >   1 file changed, 2 insertions(+), 7 deletions(-)
> > 
> > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > index 22aa63f4ef63..5fc8b514e457 100644
> > --- a/mm/vmalloc.c
> > +++ b/mm/vmalloc.c
> > @@ -3737,8 +3737,6 @@ void *__vmalloc_node_range(unsigned long size, unsigned long align,
> >   	}
> >   
> >   	if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
> > -		unsigned long size_per_node;
> > -
> >   		/*
> >   		 * Try huge pages. Only try for PAGE_KERNEL allocations,
> >   		 * others like modules don't yet expect huge pages in
> > @@ -3746,13 +3744,10 @@ void *__vmalloc_node_range(unsigned long size, unsigned long align,
> >   		 * supporting them.
> >   		 */
> >   
> > -		size_per_node = size;
> > -		if (node == NUMA_NO_NODE)
> > -			size_per_node /= num_online_nodes();
> > -		if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
> > +		if (arch_vmap_pmd_supported(prot) && size >= PMD_SIZE)
> >   			shift = PMD_SHIFT;
> >   		else
> > -			shift = arch_vmap_pte_supported_shift(size_per_node);
> > +			shift = arch_vmap_pte_supported_shift(size);
> >   
> >   		align = max(real_align, 1UL << shift);
> >   		size = ALIGN(real_size, 1UL << shift);

-- 
Sincerely yours,
Mike.