Re: [RFC PATCH v4 3/3] KVM: guest_memfd: Enforce NUMA mempolicy using shared policy

[David Hildenbrand]

On 10.02.25 07:32, Shivank Garg wrote:
> Previously, guest-memfd allocations were following local NUMA node id
> in absence of process mempolicy, resulting in random memory allocation.
> Moreover, mbind() couldn't be used since memory wasn't mapped to userspace
> in VMM.
>
> Enable NUMA policy support by implementing vm_ops for guest-memfd mmap
> operation. This allows VMM to map the memory and use mbind() to set the
> desired NUMA policy. The policy is then retrieved via
> mpol_shared_policy_lookup() and passed to filemap_grab_folio_mpol() to
> ensure that allocations follow the specified memory policy.
>
> This enables VMM to control guest memory NUMA placement by calling mbind()
> on the mapped memory regions, providing fine-grained control over guest
> memory allocation across NUMA nodes.

Yes, I think that is the right direction, especially with upcoming 
in-place conversion of shared<->private in mind.

> Suggested-by: David Hildenbrand <david@xxxxxxxxxx>
> Signed-off-by: Shivank Garg <shivankg@xxxxxxx>
> ---
>   virt/kvm/guest_memfd.c | 84 +++++++++++++++++++++++++++++++++++++++---
>   1 file changed, 78 insertions(+), 6 deletions(-)
>
> diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
> index b2aa6bf24d3a..e1ea8cb292fa 100644
> --- a/virt/kvm/guest_memfd.c
> +++ b/virt/kvm/guest_memfd.c
> @@ -2,6 +2,7 @@
>   #include <linux/backing-dev.h>
>   #include <linux/falloc.h>
>   #include <linux/kvm_host.h>
> +#include <linux/mempolicy.h>
>   #include <linux/pagemap.h>
>   #include <linux/anon_inodes.h>
>   
> @@ -11,8 +12,13 @@ struct kvm_gmem {
>   	struct kvm *kvm;
>   	struct xarray bindings;
>   	struct list_head entry;
> +	struct shared_policy policy;
>   };
>   
> +static struct mempolicy *kvm_gmem_get_pgoff_policy(struct kvm_gmem *gmem,
> +						   pgoff_t index,
> +						   pgoff_t *ilx);
> +
>   /**
>    * folio_file_pfn - like folio_file_page, but return a pfn.
>    * @folio: The folio which contains this index.
> @@ -96,10 +102,20 @@ static int kvm_gmem_prepare_folio(struct kvm *kvm, struct kvm_memory_slot *slot,
>    * Ignore accessed, referenced, and dirty flags.  The memory is
>    * unevictable and there is no storage to write back to.
>    */
> -static struct folio *kvm_gmem_get_folio(struct inode *inode, pgoff_t index)
> +static struct folio *kvm_gmem_get_folio(struct file *file, pgoff_t index)

I'd probably do that change in a separate prep-patch; would remove some 
of the unrelated noise in this patch.

>   {
>   	/* TODO: Support huge pages. */
> -	return filemap_grab_folio(inode->i_mapping, index);
> +	struct folio *folio = NULL;

No need to init folio.

> +	struct inode *inode = file_inode(file);
> +	struct kvm_gmem *gmem = file->private_data;

Prefer reverse christmas-tree (longest line first) as possible.

> +	struct mempolicy *policy;
> +	pgoff_t ilx;

Why do you return the ilx from kvm_gmem_get_pgoff_policy() if it is 
completely unused?

> +
> +	policy = kvm_gmem_get_pgoff_policy(gmem, index, &ilx);
> +	folio =  filemap_grab_folio_mpol(inode->i_mapping, index, policy);
> +	mpol_cond_put(policy);

The downside is that we always have to lookup the policy, even if we 
don't have to allocate anything because the pagecache already contains a 
folio.

Would there be a way to lookup if there is something already allcoated 
(fast-path) and fallback to the slow-path (lookup policy+call 
filemap_grab_folio_mpol) only if that failed?

Note that shmem.c does exactly that: shmem_alloc_folio() is only called 
after filemap_get_entry() told us that there is nothing.

> +
> +	return folio;
>   }
>   

[...]

> +#ifdef CONFIG_NUMA
> +static int kvm_gmem_set_policy(struct vm_area_struct *vma, struct mempolicy *new)
> +{
> +	struct file *file = vma->vm_file;
> +	struct kvm_gmem *gmem = file->private_data;
> +
> +	return mpol_set_shared_policy(&gmem->policy, vma, new);
> +}
> +
> +static struct mempolicy *kvm_gmem_get_policy(struct vm_area_struct *vma,
> +		unsigned long addr, pgoff_t *pgoff)
> +{
> +	struct file *file = vma->vm_file;
> +	struct kvm_gmem *gmem = file->private_data;
> +
> +	*pgoff = vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT);
> +	return mpol_shared_policy_lookup(&gmem->policy, *pgoff);
> +}
> +
> +static struct mempolicy *kvm_gmem_get_pgoff_policy(struct kvm_gmem *gmem,
> +						   pgoff_t index,
> +						   pgoff_t *ilx)
> +{
> +	struct mempolicy *mpol;
> +
> +	*ilx = NO_INTERLEAVE_INDEX;
> +	mpol = mpol_shared_policy_lookup(&gmem->policy, index);
> +	return mpol ? mpol : get_task_policy(current);
> +}
> +
> +static const struct vm_operations_struct kvm_gmem_vm_ops = {
> +	.get_policy	= kvm_gmem_get_policy,
> +	.set_policy	= kvm_gmem_set_policy,
> +};
> +
> +static int kvm_gmem_mmap(struct file *file, struct vm_area_struct *vma)
> +{
> +	file_accessed(file);
> +	vma->vm_ops = &kvm_gmem_vm_ops;
> +	return 0;
> +}
> +#else
> +static struct mempolicy *kvm_gmem_get_pgoff_policy(struct kvm_gmem *gmem,
> +						   pgoff_t index,
> +						   pgoff_t *ilx)
> +{
> +	*ilx = 0;
> +	return NULL;
> +}
> +#endif /* CONFIG_NUMA */
>   
>   static struct file_operations kvm_gmem_fops = {
> +#ifdef CONFIG_NUMA
> +	.mmap		= kvm_gmem_mmap,
> +#endif

With Fuad's work, this will be unconditional, and you'd only set the 
kvm_gmem_vm_ops conditionally -- just like shmem.c. Maybe best to 
prepare for that already: allow unconditional mmap (Fuad will implement 
the faulting logic of shared pages, until then all accesses would SIGBUS 
I assume, did you try that?) and only mess with get_policy/set_policy.

--
Cheers,

David / dhildenb