Re: [PATCH v3 5/5] x86/sgx: Add a basic NUMA allocation scheme to sgx_alloc_epc_page()

[Dave Hansen <dave.hansen@intel.com>]

What changed from the last patch?

On 3/3/21 7:03 AM, Jarkko Sakkinen wrote:
> Background
> ==========
> 
> EPC section is covered by one or more SRAT entries that are associated with
> one and only one PXM (NUMA node). The motivation behind this patch is to
> provide basic elements of building allocation scheme based on this premise.

Just like normal RAM, enclave memory (EPC) should be covered by entries
in the ACPI SRAT table.  These entries allow each EPC section to be
associated with a NUMA node.

Use this information to implement a simple NUMA-aware allocator for
enclave memory.

> Use phys_to_target_node() to associate each NUMA node with the EPC
> sections contained within its range. In sgx_alloc_epc_page(), first try
> to allocate from the NUMA node, where the CPU is executing. If that
> fails, fallback to the legacy allocation.

By "legacy", you mean the one from the last patch? :)

> Link: https://lore.kernel.org/lkml/158188326978.894464.217282995221175417.stgit@dwillia2-desk3.amr.corp.intel.com/
> Signed-off-by: Jarkko Sakkinen <jarkko@kernel.org>
> ---
>  arch/x86/Kconfig               |  1 +
>  arch/x86/kernel/cpu/sgx/main.c | 84 ++++++++++++++++++++++++++++++++++
>  arch/x86/kernel/cpu/sgx/sgx.h  |  9 ++++
>  3 files changed, 94 insertions(+)
> 
> diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> index a5f6a3013138..7eb1e96cfe8a 100644
> --- a/arch/x86/Kconfig
> +++ b/arch/x86/Kconfig
> @@ -1940,6 +1940,7 @@ config X86_SGX
>  	depends on CRYPTO_SHA256=y
>  	select SRCU
>  	select MMU_NOTIFIER
> +	select NUMA_KEEP_MEMINFO if NUMA

This dependency is worth mentioning somewhere.  Why do we suddenly need
NUMA_KEEP_MEMINFO?

> +/* Nodes with one or more EPC sections. */
> +static nodemask_t sgx_numa_mask;
> +
> +/*
> + * Array with one list_head for each possible NUMA node.  Each
> + * list contains all the sgx_epc_section's which are on that

					   ^ no "'", please

> + * node.
> + */
> +static struct sgx_numa_node *sgx_numa_nodes;
> +
> +/*
> + * sgx_free_epc_page() uses this to find out the correct struct sgx_numa_node,
> + * to put the page in.
> + */
> +static int sgx_section_to_numa_node_id[SGX_MAX_EPC_SECTIONS];

If this is per-section, why not put it in struct sgx_epc_section?

>  /*
> @@ -434,6 +451,36 @@ static bool __init sgx_page_reclaimer_init(struct list_head *laundry)
>  	return true;
>  }
>  
> +static struct sgx_epc_page *__sgx_alloc_epc_page_from_node(int nid)
> +{
> +	struct sgx_epc_page *page = NULL;
> +	struct sgx_numa_node *sgx_node;
> +
> +	if (WARN_ON_ONCE(nid < 0 || nid >= num_possible_nodes()))
> +		return NULL;

This has exactly one call-site which plumbs numa_node_id() in here
pretty directly.  Is this check worthwhile?

> +	if (!node_isset(nid, sgx_numa_mask))
> +		return NULL;
> +
> +	sgx_node = &sgx_numa_nodes[nid];
> +
> +	spin_lock(&sgx_free_page_list_lock);

The glocal lock protecting a per-node structure is a bit unsightly.

> +	if (list_empty(&sgx_node->free_page_list)) {
> +		spin_unlock(&sgx_free_page_list_lock);
> +		return NULL;
> +	}
> +
> +	page = list_first_entry(&sgx_node->free_page_list, struct sgx_epc_page, numa_list);
> +	list_del_init(&page->numa_list);
> +	list_del_init(&page->list);
> +	sgx_nr_free_pages--;
> +
> +	spin_unlock(&sgx_free_page_list_lock);
> +
> +	return page;
> +}
> +
>  /**
>   * __sgx_alloc_epc_page() - Allocate an EPC page
>   *
> @@ -446,8 +493,14 @@ static bool __init sgx_page_reclaimer_init(struct list_head *laundry)
>   */
>  struct sgx_epc_page *__sgx_alloc_epc_page(void)
>  {
> +	int current_nid = numa_node_id();
>  	struct sgx_epc_page *page;
>  
> +	/* Try to allocate EPC from the current node, first: */
> +	page = __sgx_alloc_epc_page_from_node(current_nid);
> +	if (page)
> +		return page;
> +
>  	spin_lock(&sgx_free_page_list_lock);
>  
>  	if (list_empty(&sgx_free_page_list)) {
> @@ -456,6 +509,7 @@ struct sgx_epc_page *__sgx_alloc_epc_page(void)
>  	}
>  
>  	page = list_first_entry(&sgx_free_page_list, struct sgx_epc_page, list);
> +	list_del_init(&page->numa_list);
>  	list_del_init(&page->list);
>  	sgx_nr_free_pages--;

I would much rather prefer that this does what the real page allocator
does: kep the page on a single list.  That list is maintained
per-NUMA-node.  Allocations try local NUMA node structures, then fall
back to other structures (hopefully in a locality-aware fashion).

I wrote you the loop that I want to see this implement in an earlier
review.  This, basically:

	page = NULL;
	nid = numa_node_id();
	while (true) {
		page = __sgx_alloc_epc_page_from_node(nid);	
		if (page)
			break;

		nid = // ... some search here, next_node_in()...
		// check if we wrapped around:
		if (nid == numa_node_id())
			break;
	}

There's no global list.  You just walk around nodes trying to find one
with space.  If you wrap around, you stop.

Please implement this.  If you think it's a bad idea, or can't, let's
talk about it in advance.  Right now, it appears that my review comments
aren't being incorporated into newer versions.

>  void sgx_free_epc_page(struct sgx_epc_page *page)
>  {
> +	int nid = sgx_section_to_numa_node_id[page->section];
> +	struct sgx_numa_node *sgx_node = &sgx_numa_nodes[nid];
>  	int ret;
>  
>  	WARN_ON_ONCE(page->flags & SGX_EPC_PAGE_RECLAIMER_TRACKED);
> @@ -575,7 +631,15 @@ void sgx_free_epc_page(struct sgx_epc_page *page)
>  		return;
>  
>  	spin_lock(&sgx_free_page_list_lock);
> +
> +	/*   Enable NUMA local allocation in sgx_alloc_epc_page(). */
> +	if (!node_isset(nid, sgx_numa_mask)) {
> +		INIT_LIST_HEAD(&sgx_node->free_page_list);
> +		node_set(nid, sgx_numa_mask);
> +	}
> +
>  	list_add_tail(&page->list, &sgx_free_page_list);
> +	list_add_tail(&page->numa_list, &sgx_node->free_page_list);
>  	sgx_nr_free_pages++;
>  	spin_unlock(&sgx_free_page_list_lock);
>  }
> @@ -626,8 +690,28 @@ static bool __init sgx_page_cache_init(struct list_head *laundry)
>  {
>  	u32 eax, ebx, ecx, edx, type;
>  	u64 pa, size;
> +	int nid;
>  	int i;
>  
> +	nodes_clear(sgx_numa_mask);

Is this really required for a variable allocated in .bss?

> +	sgx_numa_nodes = kmalloc_array(num_possible_nodes(), sizeof(*sgx_numa_nodes), GFP_KERNEL);

This is what I was looking for here, thanks!

> +	/*
> +	 * Create NUMA node lookup table for sgx_free_epc_page() as the very
> +	 * first step, as it is used to populate the free list's during the
> +	 * initialization.
> +	 */
> +	for (i = 0; i < ARRAY_SIZE(sgx_epc_sections); i++) {
> +		nid = numa_map_to_online_node(phys_to_target_node(pa));
> +		if (nid == NUMA_NO_NODE) {
> +			/* The physical address is already printed above. */
> +			pr_warn(FW_BUG "Unable to map EPC section to online node. Fallback to the NUMA node 0.\n");
> +			nid = 0;
> +		}
> +
> +		sgx_section_to_numa_node_id[i] = nid;
> +	}
> +
>  	for (i = 0; i < ARRAY_SIZE(sgx_epc_sections); i++) {
>  		cpuid_count(SGX_CPUID, i + SGX_CPUID_EPC, &eax, &ebx, &ecx, &edx);
>  
> diff --git a/arch/x86/kernel/cpu/sgx/sgx.h b/arch/x86/kernel/cpu/sgx/sgx.h
> index 41ca045a574a..3a3c07fc0c8e 100644
> --- a/arch/x86/kernel/cpu/sgx/sgx.h
> +++ b/arch/x86/kernel/cpu/sgx/sgx.h
> @@ -27,6 +27,7 @@ struct sgx_epc_page {
>  	unsigned int flags;
>  	struct sgx_encl_page *owner;
>  	struct list_head list;
> +	struct list_head numa_list;
>  };

I'll say it again, explicitly: Each sgx_epc_page should be on one and
only one free list: a per-NUMA-node list.

>  /*
> @@ -43,6 +44,14 @@ struct sgx_epc_section {
>  
>  extern struct sgx_epc_section sgx_epc_sections[SGX_MAX_EPC_SECTIONS];
>  
> +/*
> + * Contains the tracking data for NUMA nodes having EPC pages. Most importantly,
> + * the free page list local to the node is stored here.
> + */
> +struct sgx_numa_node {
> +	struct list_head free_page_list;
> +};

I think it's unconscionable to leave this protected by a global lock.
Please at least give us a per-node spinlock proteting this list.