Skip to contentSkip to conversation

Allocate memmap from hotadded memory (per device)

PATCH Superseded Oscar Salvador 2 participants 17 messages Mar 9, 2021 – Mar 18, 2021

Subsystems (10)

v4 of 10 Superseded Latest is v10 →

lore
Share & tools
lore

Oscar Salvador

author maintainer
#0 lore raw

v4 cover Allocate memmap from hotadded memory (per device)

Hi,

here is v4.

Changes from v3 -> v4:
- Addressed feedback from David
- Wrap memmap_on_memory module thingy with #ifdef on MHP_MEMMAP_ON_MEMORY
- Move "depend on MEMORY_HOTPLUG" to MHP_MEMMAP_ON_MEMORY in generic mm/Kconfig
- Collect David's Reviewed-bys

Changes from v2 -> v3:
- Addressed feedback from David
- Squash former patch#4 and and patch#5 into patch#1
- Fix config dependency CONFIR_SPARSE_VMEMMAP vs CONFIG_SPARSE_VMEMMAP_ENABLE
- Simplify module parameter functions

Changes from v1 -> v2
- Addressed feedback from David
- Fence off the feature in case struct page size is not multiple of PMD size or pageblock alignment cannot be guaranted
- Tested on x86_64 small and large memory_blocks
- Tested on arm64 4KB and 64KB page sizes (for some reason I cannot boot my VM with 16KB page size).

Arm64 with 4KB page size behaves like x86_64 after [1], which made section size smaller.
With 64KB, the feature gets fenced off due to pageblock alignment.

Changes from RFCv3 -> v1:
- Addressed feedback from David
- Re-order patches

Changes from v2 -> v3 (RFC):
- Re-order patches (Michal)
- Fold "mm,memory_hotplug: Introduce MHP_MEMMAP_ON_MEMORY" in patch#1
- Add kernel boot option to enable this feature (Michal)

Changes from v1 -> v2 (RFC):
- Addressed feedback provided by David
- Add a arch_support_memmap_on_memory to be called from mhp_supports_memmap_on_memory, as atm, only ARM, powerpc and x86_64 have altmat support.

[1] https://lore.kernel.org/lkml/[email protected]/

Original cover letter:

----

The primary goal of this patchset is to reduce memory overhead of the hot-added memory (at least for SPARSEMEM_VMEMMAP memory model).
The current way we use to populate memmap (struct page array) has two main drawbacks:

a) it consumes an additional memory until the hotadded memory itself is onlined and
b) memmap might end up on a different numa node which is especially true for movable_node configuration.
c) due to fragmentation we might end up populating memmap with base pages

One way to mitigate all these issues is to simply allocate memmap array (which is the largest memory footprint of the physical memory hotplug) from the hot-added memory itself. SPARSEMEM_VMEMMAP memory model allows us to map any pfn range so the memory doesn't need to be online to be usable for the array. See patch 3 for more details.
This feature is only usable when CONFIG_SPARSEMEM_VMEMMAP is set.

[Overall design]:

Implementation wise we reuse vmem_altmap infrastructure to override the default allocator used by vmemap_populate. memory_block structure gained a new field called nr_vmemmap_pages. This plays well for two reasons:

1) {offline/online}_pages know the difference between start_pfn and buddy_start_pfn, which is start_pfn + nr_vmemmap_pages. In this way all isolation/migration operations are done to within the right range of memory without vmemmap pages. This allows us for a much cleaner handling.

2) In try_remove_memory, we construct a new vmemap_altmap struct with the right information based on memory_block->nr_vmemap_pages, so we end up calling vmem_altmap_free instead of free_pagetable when removing the memory.

Oscar Salvador (5):
mm,memory_hotplug: Allocate memmap from the added memory range acpi,memhotplug: Enable MHP_MEMMAP_ON_MEMORY when supported mm,memory_hotplug: Add kernel boot option to enable memmap_on_memory x86/Kconfig: Introduce ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE arm64/Kconfig: Introduce ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE

 Documentation/admin-guide/kernel-parameters.txt |  16 +++
 arch/arm64/Kconfig                              |   3 +
 arch/x86/Kconfig                                |   3 +
 drivers/acpi/acpi_memhotplug.c                  |   5 +-
 drivers/base/memory.c                           |  20 +--
 include/linux/memory.h                          |   8 +-
 include/linux/memory_hotplug.h                  |  21 +++-
 include/linux/memremap.h                        |   2 +-
 include/linux/mmzone.h                          |   5 +
 mm/Kconfig                                      |   5 +
 mm/Makefile                                     |   5 +-
 mm/memory_hotplug.c                             | 157 ++++++++++++++++++++----
 mm/page_alloc.c                                 |   4 +-
 13 files changed, 216 insertions(+), 38 deletions(-)
Signature · 1 line
2.16.3
author maintainer
#1 lore raw

PATCH v4 1/5 mm,memory_hotplug: Allocate memmap from the added memory range

Physical memory hotadd has to allocate a memmap (struct page array) for the newly added memory section. Currently, alloc_pages_node() is used for those allocations.

This has some disadvantages:
a) an existing memory is consumed for that purpose

    (eg: ~2MB per 128MB memory section on x86_64)

b) if the whole node is movable then we have off-node struct pages which has performance drawbacks.
c) It might be there are no PMD_ALIGNED chunks so memmap array gets populated with base pages.

This can be improved when CONFIG_SPARSEMEM_VMEMMAP is enabled.

Vmemap page tables can map arbitrary memory.
That means that we can simply use the beginning of each memory section and map struct pages there.
struct pages which back the allocated space then just need to be treated carefully.

Implementation wise we will reuse vmem_altmap infrastructure to override the default allocator used by __populate_section_memmap.
Part of the implementation also relies on memory_block structure gaining a new field which specifies the number of vmemmap_pages at the beginning. This comes in handy as in {online,offline}_pages, all the isolation and migration is being done on (buddy_start_pfn, end_pfn] range, being buddy_start_pfn = start_pfn + nr_vmemmap_pages.

In this way, we have:

[start_pfn, buddy_start_pfn - 1] = Initialized and PageReserved
[buddy_start_pfn, end_pfn - 1] = Initialized and sent to buddy

Hot-remove:

We need to be careful when removing memory, as adding and removing memory needs to be done with the same granularity. To check that this assumption is not violated, we check the memory range we want to remove and if a) any memory block has vmemmap pages and b) the range spans more than a single memory block, we scream out loud and refuse to proceed.

If all is good and the range was using memmap on memory (aka vmemmap pages), we construct an altmap structure so free_hugepage_table does the right thing and calls vmem_altmap_free instead of free_pagetable.

---
 drivers/base/memory.c          |  20 +++---
 include/linux/memory.h         |   8 ++-
 include/linux/memory_hotplug.h |  21 ++++--
 include/linux/memremap.h       |   2 +-
 include/linux/mmzone.h         |   5 ++
 mm/Kconfig                     |   5 ++
 mm/memory_hotplug.c            | 149 +++++++++++++++++++++++++++++++++++------
 mm/page_alloc.c                |   4 +-
 8 files changed, 178 insertions(+), 36 deletions(-)
drivers/base/memory.c +12 -8
index f35298425575..5ea2b3fbce02 100644
@@ -175,7 +175,7 @@ int memory_notify(unsigned long val, void *v)
  */
 static int
 memory_block_action(unsigned long start_section_nr, unsigned long action,
-		    int online_type, int nid)
+		    int online_type, int nid, unsigned long nr_vmemmap_pages)
 {
 	unsigned long start_pfn;
 	unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
@@ -185,10 +185,11 @@ memory_block_action(unsigned long start_section_nr, unsigned long action,
 
 	switch (action) {
 	case MEM_ONLINE:
-		ret = online_pages(start_pfn, nr_pages, online_type, nid);
+		ret = online_pages(start_pfn, nr_pages, nr_vmemmap_pages,
+				   online_type, nid);
 		break;
 	case MEM_OFFLINE:
-		ret = offline_pages(start_pfn, nr_pages);
+		ret = offline_pages(start_pfn, nr_pages, nr_vmemmap_pages);
 		break;
 	default:
 		WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
@@ -211,7 +212,7 @@ static int memory_block_change_state(struct memory_block *mem,
 		mem->state = MEM_GOING_OFFLINE;
 
 	ret = memory_block_action(mem->start_section_nr, to_state,
-				  mem->online_type, mem->nid);
+				  mem->online_type, mem->nid, mem->nr_vmemmap_pages);
 
 	mem->state = ret ? from_state_req : to_state;
 
@@ -567,7 +568,8 @@ int register_memory(struct memory_block *memory)
 	return ret;
 }
 
-static int init_memory_block(unsigned long block_id, unsigned long state)
+static int init_memory_block(unsigned long block_id, unsigned long state,
+			     unsigned long nr_vmemmap_pages)
 {
 	struct memory_block *mem;
 	int ret = 0;
@@ -584,6 +586,7 @@ static int init_memory_block(unsigned long block_id, unsigned long state)
 	mem->start_section_nr = block_id * sections_per_block;
 	mem->state = state;
 	mem->nid = NUMA_NO_NODE;
+	mem->nr_vmemmap_pages = nr_vmemmap_pages;
 
 	ret = register_memory(mem);
 
@@ -603,7 +606,7 @@ static int add_memory_block(unsigned long base_section_nr)
 	if (section_count == 0)
 		return 0;
 	return init_memory_block(memory_block_id(base_section_nr),
-				 MEM_ONLINE);
+				 MEM_ONLINE, 0);
 }
 
 static void unregister_memory(struct memory_block *memory)
@@ -625,7 +628,8 @@ static void unregister_memory(struct memory_block *memory)
  *
  * Called under device_hotplug_lock.
  */
-int create_memory_block_devices(unsigned long start, unsigned long size)
+int create_memory_block_devices(unsigned long start, unsigned long size,
+				unsigned long vmemmap_pages)
 {
 	const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
 	unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
@@ -638,7 +642,7 @@ int create_memory_block_devices(unsigned long start, unsigned long size)
 		return -EINVAL;
 
 	for (block_id = start_block_id; block_id != end_block_id; block_id++) {
-		ret = init_memory_block(block_id, MEM_OFFLINE);
+		ret = init_memory_block(block_id, MEM_OFFLINE, vmemmap_pages);
 		if (ret)
 			break;
 	}
include/linux/memory.h +7 -1
index 4da95e684e20..97e92e8b556a 100644
@@ -29,6 +29,11 @@ struct memory_block {
 	int online_type;		/* for passing data to online routine */
 	int nid;			/* NID for this memory block */
 	struct device dev;
+	/*
+	 * Number of vmemmap pages. These pages
+	 * lay at the beginning of the memory block.
+	 */
+	unsigned long nr_vmemmap_pages;
 };
 
 int arch_get_memory_phys_device(unsigned long start_pfn);
@@ -80,7 +85,8 @@ static inline int memory_notify(unsigned long val, void *v)
 #else
 extern int register_memory_notifier(struct notifier_block *nb);
 extern void unregister_memory_notifier(struct notifier_block *nb);
-int create_memory_block_devices(unsigned long start, unsigned long size);
+int create_memory_block_devices(unsigned long start, unsigned long size,
+				unsigned long vmemmap_pages);
 void remove_memory_block_devices(unsigned long start, unsigned long size);
 extern void memory_dev_init(void);
 extern int memory_notify(unsigned long val, void *v);
include/linux/memory_hotplug.h +17 -4
index 7288aa5ef73b..a85d4b7d15c2 100644
@@ -55,6 +55,14 @@ typedef int __bitwise mhp_t;
  */
 #define MHP_MERGE_RESOURCE	((__force mhp_t)BIT(0))
 
+/*
+ * We want memmap (struct page array) to be self contained.
+ * To do so, we will use the beginning of the hot-added range to build
+ * the page tables for the memmap array that describes the entire range.
+ * Only selected architectures support it with SPARSE_VMEMMAP.
+ */
+#define MHP_MEMMAP_ON_MEMORY   ((__force mhp_t)BIT(1))
+
 /*
  * Extended parameters for memory hotplug:
  * altmap: alternative allocator for memmap array (optional)
@@ -101,11 +109,13 @@ extern int zone_grow_waitqueues(struct zone *zone, unsigned long nr_pages);
 extern int add_one_highpage(struct page *page, int pfn, int bad_ppro);
 /* VM interface that may be used by firmware interface */
 extern int online_pages(unsigned long pfn, unsigned long nr_pages,
-			int online_type, int nid);
+			unsigned long nr_vmemmap_pages, int online_type,
+			int nid);
 extern struct zone *test_pages_in_a_zone(unsigned long start_pfn,
 					 unsigned long end_pfn);
 extern void __offline_isolated_pages(unsigned long start_pfn,
-				     unsigned long end_pfn);
+				     unsigned long end_pfn,
+				     unsigned long buddy_start_pfn);
 
 typedef void (*online_page_callback_t)(struct page *page, unsigned int order);
 
@@ -307,7 +317,8 @@ static inline void pgdat_resize_init(struct pglist_data *pgdat) {}
 #ifdef CONFIG_MEMORY_HOTREMOVE
 
 extern void try_offline_node(int nid);
-extern int offline_pages(unsigned long start_pfn, unsigned long nr_pages);
+extern int offline_pages(unsigned long start_pfn, unsigned long nr_pages,
+			 unsigned long nr_vmemmap_pages);
 extern int remove_memory(int nid, u64 start, u64 size);
 extern void __remove_memory(int nid, u64 start, u64 size);
 extern int offline_and_remove_memory(int nid, u64 start, u64 size);
@@ -315,7 +326,8 @@ extern int offline_and_remove_memory(int nid, u64 start, u64 size);
 #else
 static inline void try_offline_node(int nid) {}
 
-static inline int offline_pages(unsigned long start_pfn, unsigned long nr_pages)
+static inline int offline_pages(unsigned long start_pfn, unsigned long nr_pages,
+				unsigned long nr_vmemmap_pages)
 {
 	return -EINVAL;
 }
@@ -359,6 +371,7 @@ extern struct zone *zone_for_pfn_range(int online_type, int nid, unsigned start_
 extern int arch_create_linear_mapping(int nid, u64 start, u64 size,
 				      struct mhp_params *params);
 void arch_remove_linear_mapping(u64 start, u64 size);
+extern bool mhp_supports_memmap_on_memory(unsigned long size);
 #endif /* CONFIG_MEMORY_HOTPLUG */
 
 #endif /* __LINUX_MEMORY_HOTPLUG_H */
include/linux/memremap.h +1 -1
index f5b464daeeca..45a79da89c5f 100644
@@ -17,7 +17,7 @@ struct device;
  * @alloc: track pages consumed, private to vmemmap_populate()
  */
 struct vmem_altmap {
-	const unsigned long base_pfn;
+	unsigned long base_pfn;
 	const unsigned long end_pfn;
 	const unsigned long reserve;
 	unsigned long free;
include/linux/mmzone.h +5 -0
index 47946cec7584..747e1c21d8e2 100644
@@ -427,6 +427,11 @@ enum zone_type {
 	 *    techniques might use alloc_contig_range() to hide previously
 	 *    exposed pages from the buddy again (e.g., to implement some sort
 	 *    of memory unplug in virtio-mem).
+	 * 6. Memory-hotplug: when using memmap_on_memory and onlining the memory
+	 *    to the MOVABLE zone, the vmemmap pages are also placed in such
+	 *    zone. Such pages cannot be really moved around as they are
+	 *    self-stored in the range, but they are treated as movable when
+	 *    the range they describe is about to be offlined.
 	 *
 	 * In general, no unmovable allocations that degrade memory offlining
 	 * should end up in ZONE_MOVABLE. Allocators (like alloc_contig_range())
mm/Kconfig +5 -0
index 24c045b24b95..febf805000f8 100644
@@ -183,6 +183,11 @@ config MEMORY_HOTREMOVE
 	depends on MEMORY_HOTPLUG && ARCH_ENABLE_MEMORY_HOTREMOVE
 	depends on MIGRATION
 
+config MHP_MEMMAP_ON_MEMORY
+	def_bool y
+	depends on MEMORY_HOTPLUG && SPARSEMEM_VMEMMAP
+	depends on ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE
+
 # Heavily threaded applications may benefit from splitting the mm-wide
 # page_table_lock, so that faults on different parts of the user address
 # space can be handled with less contention: split it at this NR_CPUS.
mm/memory_hotplug.c +128 -21

310 lines not on this page · raw message

mm/page_alloc.c +3 -1
index 3e4b29ee2b1e..85c478e374d7 100644
@@ -8830,7 +8830,8 @@ void zone_pcp_reset(struct zone *zone)
  * All pages in the range must be in a single zone, must not contain holes,
  * must span full sections, and must be isolated before calling this function.
  */
-void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
+void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn,
+			      unsigned long buddy_start_pfn)
 {
 	unsigned long pfn = start_pfn;
 	struct page *page;
@@ -8841,6 +8842,7 @@ void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
 	offline_mem_sections(pfn, end_pfn);
 	zone = page_zone(pfn_to_page(pfn));
 	spin_lock_irqsave(&zone->lock, flags);
+	pfn = buddy_start_pfn;
 	while (pfn < end_pfn) {
 		page = pfn_to_page(pfn);
 		/*
Signature · 1 line
2.16.3
#2 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

This looks essentially good to me, except some parts in mhp_supports_memmap_on_memory()

mm/memory_hotplug.c static int online_memory_block(struct memory_block *mem, void *arg) · line 1093

Quoted 5 lines
mm/memory_hotplug.c +4 -0
+bool mhp_supports_memmap_on_memory(unsigned long size)
+{
+	unsigned long pageblock_size = PFN_PHYS(pageblock_nr_pages);
+	unsigned long remaining_mem = size - PMD_SIZE;

This looks weird. I think what you want to test is that

a) "nr_vmemmap_pages * sizeof(struct page)" spans complete PMDs (IOW, we won't map too much via the altmap when populating a PMD in the vmemmap)

b) "remaining = size - nr_vmemmap_pages * sizeof(struct page)" spans complete pageblock.

right?

mm/memory_hotplug.c static int online_memory_block(struct memory_block *mem, void *arg) · line 1099

Quoted 6 lines
mm/memory_hotplug.c +6 -0
+
+	/*
+	 * Besides having CONFIG_MHP_MEMMAP_ON_MEMORY, we need a few more
+	 * assumptions to hold true:
+	 * - we are working on a single memory block granularity
+	 * - the size of struct page is multiple of PMD.

That's not what you are checking. (and the way it is phrase, I don;t think it makes sense)

Quoted 10 lines
mm/memory_hotplug.c +10 -0
+	 * - the remaining memory after having used part of the range
+	 *   for vmemmap pages is pageblock aligned.
+	 *
+	 * The reason for the struct page to be multiple of PMD is because
+	 * otherwise two sections would intersect a PMD. This would go against
+	 * memmap-on-memory premise, as memory would stay pinned until both
+	 * sections were removed.
+	 *
+	 * And the reason for the remaining memory to be pageblock aligned is
+	 * because mm core functions work on pageblock granularity in order to

mm/memory_hotplug.c static int online_memory_block(struct memory_block *mem, void *arg) · line 1119

Quoted 10 lines
mm/memory_hotplug.c +10 -0
+	 * change page's migratetype.
+	 */
+	return memmap_on_memory &&
+	       size == memory_block_size_bytes() &&
+	       IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
+	       !(PMD_SIZE % sizeof(struct page)) &&
+	       remaining_mem &&
+	       IS_ALIGNED(remaining_mem, pageblock_size);
+}
+

I suggest a restructuring, compressing the information like:

"
Besides having arch support and the feature enabled at runtime, we need a few more assumptions to hold true:

a) We span a single memory block: memory onlining/offlining happens in memory block granularity. We don't want the vmemmap of online memory blocks to reside on offline memory blocks. In the future, we might want to support variable-sized memory blocks to make the feature more versatile.

b) The vmemmap pages span complete PMDs: We don't want vmemmap code to populate memory from the altmap for unrelated parts (i.e., other memory blocks).

c) The vmemmap pages (and thereby the pages that will be exposed to the buddy) have to cover full pageblocks: memory onlining/offlining code requires applicable ranges to be page-aligned, for example, to set the migratetypes properly. "

Do we have to special case / protect against the vmemmap optimization for hugetlb pages? Or is that already blocked somehow and I missed it?

Signature · 3 lines
Thanks,

David / dhildenb
author maintainer
#3 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 7 lines · David Hildenbrand

This looks essentially good to me, except some parts in mhp_supports_memmap_on_memory()

mm/memory_hotplug.c static int online_memory_block(struct memory_block *mem, void *arg) · line 1093

Quoted 4 lines
mm/memory_hotplug.c +4 -0
+bool mhp_supports_memmap_on_memory(unsigned long size)
+{
+	unsigned long pageblock_size = PFN_PHYS(pageblock_nr_pages);
+	unsigned long remaining_mem = size - PMD_SIZE;

Hi David, thanks for the review!

Quoted 8 lines

This looks weird. I think what you want to test is that

a) "nr_vmemmap_pages * sizeof(struct page)" spans complete PMDs (IOW, we won't map too much via the altmap when populating a PMD in the vmemmap)

b) "remaining = size - nr_vmemmap_pages * sizeof(struct page)" spans complete pageblock.

We do not know the nr_vmemmap_pages at this point in time, although it is easy to pre-compute.

For every section we populate, we use PMD_SIZE. So, PMD_SIZE/PAGE_SIZE lays the nr_vmemmap_pages that are used for populating a single section.

But let me explain the reasoning I use in the current code:

I will enumarate the assumptions that must hold true in order to support the feature together with their check:

- We span a single memory block

size == memory_block_size_bytes()

- The vmemmap pages span a complete PMD and no more than a PMD.

!(PMD_SIZE % sizeof(struct page))

- The vmemmap pages and the pages exposed to the buddy have to cover full pageblocks

  remaining_mem = size - PMD_SIZE;
  IS_ALIGNED(remaining_mem, pageblock_size)

Although this check only covers the range without the vmemmap pages, one could argue that since we use only a PMD_SIZE at a time, we know that PMD_SIZE is pageblock aligned, so the vmemmap range is PMD_SIZE as well.

Now, I see how this might be confusing and rather incomplete. So I guess a better and more clear way to write it would be:

bool mhp_supports_memmap_on_memory(unsigned long size)

 {
         unsigned long nr_vmemmap_pages = PMD_SIZE / PAGE_SIZE;
         unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
         unsigned long remaining_size = size - vmemmap_size;

         return memmap_on_memory &&
                IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
                size == memory_block_size_bytes() &&
                !(PMD_SIZE % vmemmap_size) &&
                IS_ALIGNED(vmemmap_size, pageblock_size) &&
                remaining_size &&
                IS_ALIGNED(remaining_size, pageblock_size);

}

Note that above check is only for a single section, but if assumptions hold true for a single section, it will for many as well.
We could be orthodox and do:

bool mhp_supports_memmap_on_memory(unsigned long size)

 {
         unsigned long nr_sections = (1ULL << SECTION_SHIFT) / memory_block_size_bytes;
         unsigned long nr_vmemmap_pages = (PMD_SIZE / PAGE_SIZE) * nr_sections;
         unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
         unsigned long remaining_size = size - vmemmap_size;

         return memmap_on_memory &&
                IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
                size == memory_block_size_bytes() &&
                !(PMD_SIZE % vmemmap_size) &&
                IS_ALIGNED(vmemmap_size, pageblock_size) &&
                remaining_size &&
                IS_ALIGNED(remaining_size, pageblock_size);

}

to check for all sections, but I do not think it is necessary.

What do you think?

Quoted 15 lines

I suggest a restructuring, compressing the information like:

"
Besides having arch support and the feature enabled at runtime, we need a few more assumptions to hold true:

a) We span a single memory block: memory onlining/offlining happens in memory block granularity. We don't want the vmemmap of online memory blocks to reside on offline memory blocks. In the future, we might want to support variable-sized memory blocks to make the feature more versatile.

b) The vmemmap pages span complete PMDs: We don't want vmemmap code to populate memory from the altmap for unrelated parts (i.e., other memory blocks).

Quoted 5 lines

c) The vmemmap pages (and thereby the pages that will be exposed to the buddy) have to cover full pageblocks: memory onlining/offlining code requires applicable ranges to be page-aligned, for example, to set the migratetypes properly.
"

I am fine with the above, I already added it, thanks.

Quoted 2 lines

Do we have to special case / protect against the vmemmap optimization for hugetlb pages? Or is that already blocked somehow and I missed it?

Yes, hugetlb-vmemmap feature disables vmemmap on PMD population [1]

[1] https://patchwork.kernel.org/project/linux-mm/patch/[email protected]/

Signature · 2 lines
Oscar Salvador
SUSE L3
#4 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 17 lines · Oscar Salvador
Quoted 7 lines · David Hildenbrand

This looks essentially good to me, except some parts in mhp_supports_memmap_on_memory()

mm/memory_hotplug.c static int online_memory_block(struct memory_block *mem, void *arg) · line 1093

Quoted 4 lines
mm/memory_hotplug.c +4 -0
+bool mhp_supports_memmap_on_memory(unsigned long size)
+{
+	unsigned long pageblock_size = PFN_PHYS(pageblock_nr_pages);
+	unsigned long remaining_mem = size - PMD_SIZE;

Hi David, thanks for the review!

Quoted 6 lines

This looks weird. I think what you want to test is that

a) "nr_vmemmap_pages * sizeof(struct page)" spans complete PMDs (IOW, we won't map too much via the altmap when populating a PMD in the vmemmap)

Quoted 8 lines
Quoted 2 lines

b) "remaining = size - nr_vmemmap_pages * sizeof(struct page)" spans complete pageblock.

We do not know the nr_vmemmap_pages at this point in time, although it is easy to pre-compute.

For every section we populate, we use PMD_SIZE. So, PMD_SIZE/PAGE_SIZE lays the nr_vmemmap_pages that are used for populating a single section.

I find that cross reference to vmemmap code a little hard to digest. I would have assume that we don't have to care about PMDs in this code here at all. The vmemmap population code should handle that.

I think I already mentioned that somewhere, I think it should be like this:

a) vmemmap code should *never* populate more memory than requested for a single memory section when we are populating from the altmap. If that cannot be guaranteed for PMDs, then we have to fallback to populating base pages. Populating PMDs from an altmap with sizeof(struct page) == 64 is highly dangerous.

Assume we have sizeof(struct page) == 56. A 128 MiB section spans 32768 pages - we need 32768 * sizeof(struct page) space for the vmemmap.

With 64k pages we *can* use exactly one PMD. With 56k pages we need 448 individual (full!) pages for the vmemmap.

IOW, we don't care how vmemmap code will do the mapping. vmemmap code has to get it right. IMHO, asserting it in this code is wrong.

b) In this code, we really should only care about what memory onlining/offlining code can or can't do. We really only care that

1) size == memory_block_size_bytes()
2) remaining_size
3) IS_ALIGNED(remaining_size, pageblock_size);

I think a) is a logical consequence of b) for x86-64, whereby the pageblock size corresponds to PMD, so we might not have to care about a) right now.

See below for my take.

Quoted 40 lines

But let me explain the reasoning I use in the current code:

I will enumarate the assumptions that must hold true in order to support the feature together with their check:

- We span a single memory block

   size == memory_block_size_bytes()

- The vmemmap pages span a complete PMD and no more than a PMD.

!(PMD_SIZE % sizeof(struct page))

- The vmemmap pages and the pages exposed to the buddy have to cover full pageblocks

   remaining_mem = size - PMD_SIZE;
   IS_ALIGNED(remaining_mem, pageblock_size)

Although this check only covers the range without the vmemmap pages, one could argue that since we use only a PMD_SIZE at a time, we know that PMD_SIZE is pageblock aligned, so the vmemmap range is PMD_SIZE as well.

Now, I see how this might be confusing and rather incomplete. So I guess a better and more clear way to write it would be:

  bool mhp_supports_memmap_on_memory(unsigned long size)
  {
          unsigned long nr_vmemmap_pages = PMD_SIZE / PAGE_SIZE;
          unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
          unsigned long remaining_size = size - vmemmap_size;
          return memmap_on_memory &&
                 IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
                 size == memory_block_size_bytes() &&
                 !(PMD_SIZE % vmemmap_size) &&
                 IS_ALIGNED(vmemmap_size, pageblock_size) &&
                 remaining_size &&
                 IS_ALIGNED(remaining_size, pageblock_size);
Quoted 4 lines

}

Note that above check is only for a single section, but if assumptions hold true for a single section, it will for many as well.

Okay, please document the statement about single sections, that's important to understand what's happening.

My take would be

bool mhp_supports_memmap_on_memory(unsigned long size)

{
	/*
	 * Note: We calculate for a single memory section. The calculation
	 * implicitly covers memory blocks that span multiple sections.
	 */
	unsigned long nr_vmemmap_pages = SECTION_SIZE / PAGE_SIZE;
	unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
	unsigned long remaining_size = size - vmemmap_size;

	/*
	 * Note: vmemmap code has to make sure to not populate more memory
	 * via the altmap than absolutely necessary for a single section.
	 * This always holds when we allocate pageblock-sized chunks from
	 * the altmap, as we require pageblock alignment here.
	 *
	 * TODO, other doc
	 */
	return memmap_on_memory &&
               IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
	       size == memory_block_size_bytes() &&
	       remaining_size &&
	       IS_ALIGNED(remaining_size, pageblock_size);
}

For arm64 with 64k base pages it might not hold and we might have to fix vmemmap code to not over-populate PMDs (512 MB). I did not have a loom at the code, though.

Quoted 22 lines
Quoted 16 lines

I suggest a restructuring, compressing the information like:

"
Besides having arch support and the feature enabled at runtime, we need a few more assumptions to hold true:

a) We span a single memory block: memory onlining/offlining happens in memory block granularity. We don't want the vmemmap of online memory blocks to reside on offline memory blocks. In the future, we might want to support variable-sized memory blocks to make the feature more versatile.

b) The vmemmap pages span complete PMDs: We don't want vmemmap code to populate memory from the altmap for unrelated parts (i.e., other memory blocks).

Quoted 5 lines

c) The vmemmap pages (and thereby the pages that will be exposed to the buddy) have to cover full pageblocks: memory onlining/offlining code requires applicable ranges to be page-aligned, for example, to set the migratetypes properly.
"

Quoted 8 lines

I am fine with the above, I already added it, thanks.

Quoted 2 lines

Do we have to special case / protect against the vmemmap optimization for hugetlb pages? Or is that already blocked somehow and I missed it?

Yes, hugetlb-vmemmap feature disables vmemmap on PMD population [1]

[1] https://patchwork.kernel.org/project/linux-mm/patch/[email protected]/

Sorry, I might be missing something important. How does that make both features run mutually exclusive?

hugetlb-vmemmap only instructs to populate base pages on !altmap. I assume if hugetlb-vmemmap code stumbles over a PMD it will skip optimization.

But what if your machine does not have boot_cpu_has(X86_FEATURE_PSE) -- IOW, you always populate base pages?

Either hugetlb-vmemmap code would have to identify that these pages come from an altmap (how? PageReserved() is also used for boot memory), or both features have to somehow run mutually exclusive.

Signature · 3 lines
Thanks,

David / dhildenb
#5 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 132 lines · David Hildenbrand
Quoted 17 lines · Oscar Salvador
Quoted 7 lines · David Hildenbrand

This looks essentially good to me, except some parts in mhp_supports_memmap_on_memory()

mm/memory_hotplug.c static int online_memory_block(struct memory_block *mem, void *arg) · line 1093

Quoted 4 lines
mm/memory_hotplug.c +4 -0
+bool mhp_supports_memmap_on_memory(unsigned long size)
+{
+	unsigned long pageblock_size = PFN_PHYS(pageblock_nr_pages);
+	unsigned long remaining_mem = size - PMD_SIZE;

Hi David, thanks for the review!

Quoted 6 lines

This looks weird. I think what you want to test is that

a) "nr_vmemmap_pages * sizeof(struct page)" spans complete PMDs (IOW, we won't map too much via the altmap when populating a PMD in the vmemmap)

Quoted 8 lines
Quoted 2 lines

b) "remaining = size - nr_vmemmap_pages * sizeof(struct page)" spans complete pageblock.

We do not know the nr_vmemmap_pages at this point in time, although it is easy to pre-compute.

For every section we populate, we use PMD_SIZE. So, PMD_SIZE/PAGE_SIZE lays the nr_vmemmap_pages that are used for populating a single section.

I find that cross reference to vmemmap code a little hard to digest. I would have assume that we don't have to care about PMDs in this code here at all. The vmemmap population code should handle that.

I think I already mentioned that somewhere, I think it should be like this:

a) vmemmap code should *never* populate more memory than requested for a single memory section when we are populating from the altmap. If that cannot be guaranteed for PMDs, then we have to fallback to populating base pages. Populating PMDs from an altmap with sizeof(struct page) == 64 is highly dangerous.

Assume we have sizeof(struct page) == 56. A 128 MiB section spans 32768 pages - we need 32768 * sizeof(struct page) space for the vmemmap.

With 64k pages we *can* use exactly one PMD. With 56k pages we need 448 individual (full!) pages for the vmemmap.

IOW, we don't care how vmemmap code will do the mapping. vmemmap code has to get it right. IMHO, asserting it in this code is wrong.

b) In this code, we really should only care about what memory onlining/offlining code can or can't do. We really only care that

1) size == memory_block_size_bytes()
2) remaining_size
3) IS_ALIGNED(remaining_size, pageblock_size);

I think a) is a logical consequence of b) for x86-64, whereby the pageblock size corresponds to PMD, so we might not have to care about a) right now.

See below for my take.

Quoted 42 lines

But let me explain the reasoning I use in the current code:

I will enumarate the assumptions that must hold true in order to support the feature together with their check:

- We span a single memory block

    size == memory_block_size_bytes()

- The vmemmap pages span a complete PMD and no more than a PMD.

!(PMD_SIZE % sizeof(struct page))

- The vmemmap pages and the pages exposed to the buddy have to cover full pageblocks

remaining_mem = size - PMD_SIZE;
IS_ALIGNED(remaining_mem, pageblock_size)

Although this check only covers the range without the vmemmap pages, one could argue that since we use only a PMD_SIZE at a time, we know that PMD_SIZE is pageblock aligned, so the vmemmap range is PMD_SIZE as well.

Now, I see how this might be confusing and rather incomplete. So I guess a better and more clear way to write it would be:

   bool mhp_supports_memmap_on_memory(unsigned long size)
   {
           unsigned long nr_vmemmap_pages = PMD_SIZE / PAGE_SIZE;
           unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
           unsigned long remaining_size = size - vmemmap_size;
           return memmap_on_memory &&
                  IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
                  size == memory_block_size_bytes() &&
                  !(PMD_SIZE % vmemmap_size) &&
                  IS_ALIGNED(vmemmap_size, pageblock_size) &&
                  remaining_size &&
                  IS_ALIGNED(remaining_size, pageblock_size);
    }
Quoted 2 lines

Note that above check is only for a single section, but if assumptions hold true for a single section, it will for many as well.

Okay, please document the statement about single sections, that's important to understand what's happening.

My take would be

bool mhp_supports_memmap_on_memory(unsigned long size)

{
	/*
	 * Note: We calculate for a single memory section. The calculation
	 * implicitly covers memory blocks that span multiple sections.
	 */
	unsigned long nr_vmemmap_pages = SECTION_SIZE / PAGE_SIZE;
	unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
	unsigned long remaining_size = size - vmemmap_size;

	/*
	 * Note: vmemmap code has to make sure to not populate more memory
	 * via the altmap than absolutely necessary for a single section.
	 * This always holds when we allocate pageblock-sized chunks from
	 * the altmap, as we require pageblock alignment here.
Quoted 8 lines
	 *
	 * TODO, other doc
	 */
	return memmap_on_memory &&
                 IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
	       size == memory_block_size_bytes() &&
	       remaining_size &&
	       IS_ALIGNED(remaining_size, pageblock_size);

Oh, I forgot, we can also drop the check for "remaining_size" in that code. If we ever have PAGE_SIZE == sizeof(struct page) we'll be in bigger trouble :D

Signature · 3 lines
Thanks,

David / dhildenb
author maintainer
#6 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 11 lines · David Hildenbrand
Quoted 11 lines

I find that cross reference to vmemmap code a little hard to digest. I would have assume that we don't have to care about PMDs in this code here at all. The vmemmap population code should handle that.

I think I already mentioned that somewhere, I think it should be like this:

a) vmemmap code should *never* populate more memory than requested for a single memory section when we are populating from the altmap. If that cannot be guaranteed for PMDs, then we have to fallback to populating base pages. Populating PMDs from an altmap with sizeof(struct page) == 64 is highly dangerous.

I guess you meant sizeof(struct page) != 64

But other usecases of using altmap (ZONE_DEVICE stuff) might not care whether they have sub-populated PMDs when populating sections from altmap?

Current vmemmap code populates PMD with PMD_SIZE if empty, and with basepages if there are still holes.

Quoted 11 lines
Quoted 11 lines

Assume we have sizeof(struct page) == 56. A 128 MiB section spans 32768 pages - we need 32768 * sizeof(struct page) space for the vmemmap.
With 64k pages we *can* use exactly one PMD. With 56k pages we need 448 individual (full!) pages for the vmemmap.

IOW, we don't care how vmemmap code will do the mapping. vmemmap code has to get it right. IMHO, asserting it in this code is wrong.

Quoted 7 lines
Quoted 7 lines

b) In this code, we really should only care about what memory onlining/offlining code can or can't do. We really only care that

1) size == memory_block_size_bytes()
2) remaining_size
3) IS_ALIGNED(remaining_size, pageblock_size);

I agree with the above, but see below:

Quoted 5 lines
Quoted 5 lines

Okay, please document the statement about single sections, that's important to understand what's happening.

My take would be

Quoted 8 lines
Quoted 8 lines

bool mhp_supports_memmap_on_memory(unsigned long size)

{
	/*
	 * Note: We calculate for a single memory section. The calculation
	 */
	unsigned long nr_vmemmap_pages = SECTION_SIZE / PAGE_SIZE;
	unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
	unsigned long remaining_size = size - vmemmap_size;

While it might be true that we need to back off from populating with altmap in case PMDs are not going to be fully populated because of the size of the struct page (I am not still not sure though as I said above, other usecases might not care at all), I would go __for now__ with placing vmemmap_size == PMD_SIZE in the check below as well.

If the check comes true, we know that we fully populate PMDs when populating sections, so the feature can be used.

Then I commit to have a look whether we need to back off in vmemmap-populating code in case altmap && !NOT_FULLY_POPULATED_PMDS.

What do you think?

Signature · 2 lines
Oscar Salvador
SUSE L3
#7 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 7 lines · Oscar Salvador
Quoted 6 lines · David Hildenbrand
Quoted 6 lines

I find that cross reference to vmemmap code a little hard to digest. I would have assume that we don't have to care about PMDs in this code here at all. The vmemmap population code should handle that.

I think I already mentioned that somewhere, I think it should be like this:

Quoted 8 lines
Quoted 5 lines
Quoted 5 lines

a) vmemmap code should *never* populate more memory than requested for a single memory section when we are populating from the altmap. If that cannot be guaranteed for PMDs, then we have to fallback to populating base pages. Populating PMDs from an altmap with sizeof(struct page) == 64 is highly dangerous.

I guess you meant sizeof(struct page) != 64

Yes!

Quoted 2 lines

But other usecases of using altmap (ZONE_DEVICE stuff) might not care whether they have sub-populated PMDs when populating sections from altmap?

Just assume you have two ranges

[ ZONE_DEVICE 0 ][ ZONE_DEVICE 1]

If the vmemmap of ZONE_DEVICE 1 could be taken from the altmap of ZONE_DEVICE 0, we could be in trouble, as both parts can be removed/repurposed independently ...

Quoted 45 lines

Current vmemmap code populates PMD with PMD_SIZE if empty, and with basepages if there are still holes.

Quoted 11 lines
Quoted 11 lines

Assume we have sizeof(struct page) == 56. A 128 MiB section spans 32768 pages - we need 32768 * sizeof(struct page) space for the vmemmap.
With 64k pages we *can* use exactly one PMD. With 56k pages we need 448 individual (full!) pages for the vmemmap.

IOW, we don't care how vmemmap code will do the mapping. vmemmap code has to get it right. IMHO, asserting it in this code is wrong.

Quoted 7 lines
Quoted 7 lines

b) In this code, we really should only care about what memory onlining/offlining code can or can't do. We really only care that

1) size == memory_block_size_bytes()
2) remaining_size
3) IS_ALIGNED(remaining_size, pageblock_size);

I agree with the above, but see below:

Quoted 5 lines
Quoted 5 lines

Okay, please document the statement about single sections, that's important to understand what's happening.

My take would be

Quoted 8 lines
Quoted 8 lines

bool mhp_supports_memmap_on_memory(unsigned long size)

{
	/*
	 * Note: We calculate for a single memory section. The calculation
	 */
	unsigned long nr_vmemmap_pages = SECTION_SIZE / PAGE_SIZE;
	unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page);
	unsigned long remaining_size = size - vmemmap_size;

While it might be true that we need to back off from populating with altmap in case PMDs are not going to be fully populated because of the size of the struct page (I am not still not sure though as I said above, other usecases might not care at all), I would go __for now__ with placing vmemmap_size == PMD_SIZE in the check below as well.

Quoted 7 lines

If the check comes true, we know that we fully populate PMDs when populating sections, so the feature can be used.

Then I commit to have a look whether we need to back off in vmemmap-populating code in case altmap && !NOT_FULLY_POPULATED_PMDS.

What do you think?

If we check for

IS_ALIGNED(nr_vmemmap_pages, PMD_SIZE), please add a proper TODO comment that this is most probably the wrong place to take care of this.

Signature · 3 lines
Thanks,

David / dhildenb
author maintainer
#8 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 7 lines · David Hildenbrand

Just assume you have two ranges

[ ZONE_DEVICE 0 ][ ZONE_DEVICE 1]

If the vmemmap of ZONE_DEVICE 1 could be taken from the altmap of ZONE_DEVICE 0, we could be in trouble, as both parts can be removed/repurposed independently ...

I have to say my knowledge about ZONE_DEVICE and its intrinsencs tend to 0, that is why I thought it might not matter, but I agree that this is only asking for trouble.

Quoted 4 lines

If we check for

IS_ALIGNED(nr_vmemmap_pages, PMD_SIZE), please add a proper TODO comment that this is most probably the wrong place to take care of this.

Sure, I will stuff the check in there and place a big TODO comment so we do not forget about addressing this issue the right way.

I will prepare a v5 (hopefully the last one) and do some more testing before sending it out.

Thanks David!

Signature · 2 lines
Oscar Salvador
SUSE L3
author maintainer
#9 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 7 lines · Oscar Salvador
Quoted 4 lines

If we check for

IS_ALIGNED(nr_vmemmap_pages, PMD_SIZE), please add a proper TODO comment that this is most probably the wrong place to take care of this.

Sure, I will stuff the check in there and place a big TODO comment so we do not forget about addressing this issue the right way.

Ok, I realized something while working on v5.

Here is what I have right now:

bool mhp_supports_memmap_on_memory(unsigned long size)
{
/*
* Note: We calculate for a single memory section. The calculation
* implicitly covers memory blocks that span multiple sections. *
* Not all archs define SECTION_SIZE, but MIN_MEMORY_BLOCK_SIZE always
* equals SECTION_SIZE, so use that instead.
*/
unsigned long nr_vmemmap_pages = MIN_MEMORY_BLOCK_SIZE / PAGE_SIZE; unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page); unsigned long remaining_size = size - vmemmap_size;

/*
* Besides having arch support and the feature enabled at runtime, we
* need a few more assumptions to hold true:
*
* a) We span a single memory block: memory onlining/offlinin;g happens
* in memory block granularity. We don't want the vmemmap of online
* memory blocks to reside on offline memory blocks. In the future,
* we might want to support variable-sized memory blocks to make the
* feature more versatile.
*
* b) The vmemmap pages span complete PMDs: We don't want vmemmap code
* to populate memory from the altmap for unrelated parts (i.e.,
* other memory blocks)
*
* c) The vmemmap pages (and thereby the pages that will be exposed to
* the buddy) have to cover full pageblocks: memory onlining/offlining
* code requires applicable ranges to be page-aligned, for example, to
* set the migratetypes properly.
*
* TODO: Although we have a check here to make sure that vmemmap pages
* fully populate a PMD, it is not the right place to check for
* this. A much better solution involves improving vmemmap code
* to fallback to base pages when trying to populate vmemmap using
* altmap as an alternative source of memory, and we do not exactly
* populate a single PMD.
*/
return memmap_on_memory &&
IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
size == memory_block_size_bytes() &&
remaining_size &&
IS_ALIGNED(remaining_size, pageblock_size) && IS_ALIGNED(vmemmap_size, PMD_SIZE);
}

Assume we are on x86_64 to simplify the case.

Above, nr_vmemmap_pages would be 32768 and vmemmap_size 2MB (exactly a PMD).

Now, although correct, this nr_vmemmap_pages does not match with the altmap->alloc.

static void * __meminit altmap_alloc_block_buf(unsigned long size,
struct altmap)
{
...
...
nr_pfns = size >> PAGE_SHIFT; //size is PMD_SIZE altmap->alloc += nr_pfns;
}

altmap->alloc will be 512, 512 * 4K pages = 2MB.

Of course, the reason they do not match is because in one case, we are saying a) how many pfns we need to cover a PMD_SIZE, while in the other case we say b) how many pages we need to cover SECTION_SIZE

Then b) multiply for page_size to get the current size of it.

So, I have mixed feeling about this.
Would it be more clear to just do:

bool mhp_supports_memmap_on_memory(unsigned long size)

 {
        /*
         * Note: We calculate for a single memory section. The calculation
         * implicitly covers memory blocks that span multiple sections.
         */
        unsigned long nr_vmemmap_pages = PMD_SIZE / PAGE_SIZE;
        unsigned long vmemmap_size = nr_vmemmap_pages * PAGE_SIZE;
        unsigned long remaining_size = size - vmemmap_size;
	...
	...
Signature · 2 lines
Oscar Salvador
SUSE L3
#10 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 13 lines · Oscar Salvador
Quoted 7 lines · Oscar Salvador
Quoted 4 lines

If we check for

IS_ALIGNED(nr_vmemmap_pages, PMD_SIZE), please add a proper TODO comment that this is most probably the wrong place to take care of this.

Sure, I will stuff the check in there and place a big TODO comment so we do not forget about addressing this issue the right way.

Ok, I realized something while working on v5.

Here is what I have right now:

Quoted 10 lines

bool mhp_supports_memmap_on_memory(unsigned long size)
{
/*
* Note: We calculate for a single memory section. The calculation
* implicitly covers memory blocks that span multiple sections. *
* Not all archs define SECTION_SIZE, but MIN_MEMORY_BLOCK_SIZE always
* equals SECTION_SIZE, so use that instead.
*/
unsigned long nr_vmemmap_pages = MIN_MEMORY_BLOCK_SIZE / PAGE_SIZE;

Even clearer would be just using "size / PAGE_SIZE" here. The you can even drop the comment.

Quoted 54 lines

unsigned long vmemmap_size = nr_vmemmap_pages * sizeof(struct page); unsigned long remaining_size = size - vmemmap_size;

/*
* Besides having arch support and the feature enabled at runtime, we
* need a few more assumptions to hold true:
*
* a) We span a single memory block: memory onlining/offlinin;g happens
* in memory block granularity. We don't want the vmemmap of online
* memory blocks to reside on offline memory blocks. In the future,
* we might want to support variable-sized memory blocks to make the
* feature more versatile.
*
* b) The vmemmap pages span complete PMDs: We don't want vmemmap code
* to populate memory from the altmap for unrelated parts (i.e.,
* other memory blocks)
*
* c) The vmemmap pages (and thereby the pages that will be exposed to
* the buddy) have to cover full pageblocks: memory onlining/offlining
* code requires applicable ranges to be page-aligned, for example, to
* set the migratetypes properly.
*
* TODO: Although we have a check here to make sure that vmemmap pages
* fully populate a PMD, it is not the right place to check for
* this. A much better solution involves improving vmemmap code
* to fallback to base pages when trying to populate vmemmap using
* altmap as an alternative source of memory, and we do not exactly
* populate a single PMD.
*/
return memmap_on_memory &&
IS_ENABLED(CONFIG_MHP_MEMMAP_ON_MEMORY) &&
size == memory_block_size_bytes() &&
remaining_size &&
IS_ALIGNED(remaining_size, pageblock_size) && IS_ALIGNED(vmemmap_size, PMD_SIZE);
}

Assume we are on x86_64 to simplify the case.

Above, nr_vmemmap_pages would be 32768 and vmemmap_size 2MB (exactly a PMD).

Now, although correct, this nr_vmemmap_pages does not match with the altmap->alloc.

  static void * __meminit altmap_alloc_block_buf(unsigned long size,
   struct altmap)
  {
    ...
    ...
    nr_pfns = size >> PAGE_SHIFT; //size is PMD_SIZE
    altmap->alloc += nr_pfns;
  }
Quoted 7 lines

altmap->alloc will be 512, 512 * 4K pages = 2MB.

Of course, the reason they do not match is because in one case, we are saying a) how many pfns we need to cover a PMD_SIZE, while in the other case we say b) how many pages we need to cover SECTION_SIZE

Then b) multiply for page_size to get the current size of it.

I don't follow. 2MB == 2MB. And if there would be difference then we would be in the problem I brought up: vmemmap code allocating too much via the altmap, which can be very bad because might be populating more vmemmap than we actually need.

Quoted 10 lines

So, I have mixed feeling about this.
Would it be more clear to just do:

bool mhp_supports_memmap_on_memory(unsigned long size)
{
/*
* Note: We calculate for a single memory section. The calculation
* implicitly covers memory blocks that span multiple sections. */

Then this comment is wrong

Quoted 1 line
         unsigned long nr_vmemmap_pages = PMD_SIZE / PAGE_SIZE;

And this stuff just gets confusing.

nr_vmemmap_pages = 2MiB / 4 KiB = 512;

Quoted 1 line
         unsigned long vmemmap_size = nr_vmemmap_pages * PAGE_SIZE;

vmemmap_size = 512 * 4KiB = 2 MiB.

That calculation wasn't very useful (/ PAGE_SIZE * PAGE_SIZE)?

Quoted 1 line
         unsigned long remaining_size = size - vmemmap_size;

And here we could get something like

remaining_size = 2 GiB - 2 MiB

?

Which does not make any sense.

Quoted 4 lines
	...
	...
Signature · 3 lines
Thanks,

David / dhildenb
author maintainer
#11 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 4 lines · David Hildenbrand

I don't follow. 2MB == 2MB. And if there would be difference then we would be in the problem I brought up: vmemmap code allocating too much via the altmap, which can be very bad because might be populating more vmemmap than we actually need.

Yes, I meant to say nr_vmemmap_pages won't match, or IOW, won't have the same meaning.
The end result is the same.

Quoted 3 lines

vmemmap_size = 512 * 4KiB = 2 MiB.

That calculation wasn't very useful (/ PAGE_SIZE * PAGE_SIZE)?

Yeah, somewhat redundant.

Quoted 6 lines
Quoted 2 lines
         unsigned long remaining_size = size - vmemmap_size;

And here we could get something like

remaining_size = 2 GiB - 2 MiB

Yes, vmemmap_size would need to scale with nr_sections to be relative to size.

Just wanted to bring it up, because somene might wonder "ok, why do we have altmap->nr_pfns = X, and here nr_vmemmap_pages
is Y"

It was an effort to make it consistent, although I see it would bring more confusion other than anything, so disregard.

Signature · 2 lines
Oscar Salvador
SUSE L3
#12 lore raw

Re: [PATCH v4 1/5] mm,memory_hotplug: Allocate memmap from the added memory range

Quoted 26 lines · Oscar Salvador
Quoted 4 lines · David Hildenbrand

I don't follow. 2MB == 2MB. And if there would be difference then we would be in the problem I brought up: vmemmap code allocating too much via the altmap, which can be very bad because might be populating more vmemmap than we actually need.

Yes, I meant to say nr_vmemmap_pages won't match, or IOW, won't have the same meaning.
The end result is the same.

Quoted 3 lines

vmemmap_size = 512 * 4KiB = 2 MiB.

That calculation wasn't very useful (/ PAGE_SIZE * PAGE_SIZE)?

Yeah, somewhat redundant.

Quoted 6 lines
Quoted 2 lines
          unsigned long remaining_size = size - vmemmap_size;

And here we could get something like

remaining_size = 2 GiB - 2 MiB

Yes, vmemmap_size would need to scale with nr_sections to be relative to size.

Quoted 6 lines

Just wanted to bring it up, because somene might wonder "ok, why do we have altmap->nr_pfns = X, and here nr_vmemmap_pages
is Y"

It was an effort to make it consistent, although I see it would bring more confusion other than anything, so disregard.

I am also unhappy that we have to replicate the same computation at two places, but I don't see an easy way to avoid that ... we have to trust on vmemmap code to do the right thing either way :(

Signature · 3 lines
Thanks,

David / dhildenb
author maintainer
#13 lore raw

PATCH v4 2/5 acpi,memhotplug: Enable MHP_MEMMAP_ON_MEMORY when supported

In reply to Oscar Salvador

Let the caller check whether it can pass MHP_MEMMAP_ON_MEMORY by checking mhp_supports_memmap_on_memory(). MHP_MEMMAP_ON_MEMORY can only be set in case ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE is enabled, the architecture supports altmap, and the range to be added spans a single memory block.

---
 drivers/acpi/acpi_memhotplug.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/acpi/acpi_memhotplug.c +4 -1
index b02fd51e5589..8cc195c4c861 100644
@@ -171,6 +171,7 @@ static int acpi_memory_enable_device(struct acpi_memory_device *mem_device)
 	acpi_handle handle = mem_device->device->handle;
 	int result, num_enabled = 0;
 	struct acpi_memory_info *info;
+	mhp_t mhp_flags = MHP_NONE;
 	int node;
 
 	node = acpi_get_node(handle);
@@ -194,8 +195,10 @@ static int acpi_memory_enable_device(struct acpi_memory_device *mem_device)
 		if (node < 0)
 			node = memory_add_physaddr_to_nid(info->start_addr);
 
+		if (mhp_supports_memmap_on_memory(info->length))
+			mhp_flags |= MHP_MEMMAP_ON_MEMORY;
 		result = __add_memory(node, info->start_addr, info->length,
-				      MHP_NONE);
+				      mhp_flags);
 
 		/*
 		 * If the memory block has been used by the kernel, add_memory()
Signature · 1 line
2.16.3
author maintainer
#14 lore raw

PATCH v4 3/5 mm,memory_hotplug: Add kernel boot option to enable memmap_on_memory

In reply to Oscar Salvador

Self stored memmap leads to a sparse memory situation which is unsuitable for workloads that requires large contiguous memory chunks, so make this an opt-in which needs to be explicitly enabled.

To control this, let memory_hotplug have its own memory space, as suggested by David, so we can add memory_hotplug.memmap_on_memory parameter.

---
 Documentation/admin-guide/kernel-parameters.txt | 16 ++++++++++++++++
 mm/Makefile                                     |  5 ++++-
 mm/memory_hotplug.c                             | 10 +++++++++-
 3 files changed, 29 insertions(+), 2 deletions(-)
Documentation/admin-guide/kernel-parameters.txt +16 -0
index 04545725f187..d29b96e50c5c 100644
@@ -2794,6 +2794,22 @@
 			seconds.  Use this parameter to check at some
 			other rate.  0 disables periodic checking.
 
+	memory_hotplug.memmap_on_memory
+			[KNL,X86,ARM] Boolean flag to enable this feature.
+			Format: {on | off (default)}
+			When enabled, memory to build the pages tables for the
+			memmap array describing the hot-added range will be taken
+			from the range itself, so the memmap page tables will be
+			self-hosted.
+			Since only single memory device ranges are supported at
+			the moment, this option is disabled by default because
+			it might have an impact on workloads that needs large
+			contiguous memory chunks.
+			The state of the flag can be read in
+			/sys/module/memory_hotplug/parameters/memmap_on_memory.
+			Note that even when enabled, there are a few cases where
+			the feature is not effective.
+
 	memtest=	[KNL,X86,ARM,PPC] Enable memtest
 			Format: <integer>
 			default : 0 <disable>
mm/Makefile +4 -1
index 72227b24a616..82ae9482f5e3 100644
@@ -58,9 +58,13 @@ obj-y			:= filemap.o mempool.o oom_kill.o fadvise.o \
 page-alloc-y := page_alloc.o
 page-alloc-$(CONFIG_SHUFFLE_PAGE_ALLOCATOR) += shuffle.o
 
+# Give 'memory_hotplug' its own module-parameter namespace
+memory-hotplug-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o
+
 obj-y += page-alloc.o
 obj-y += init-mm.o
 obj-y += memblock.o
+obj-y += $(memory-hotplug-y)
 
 ifdef CONFIG_MMU
 	obj-$(CONFIG_ADVISE_SYSCALLS)	+= madvise.o
@@ -83,7 +87,6 @@ obj-$(CONFIG_SLUB) += slub.o
 obj-$(CONFIG_KASAN)	+= kasan/
 obj-$(CONFIG_KFENCE) += kfence/
 obj-$(CONFIG_FAILSLAB) += failslab.o
-obj-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o
 obj-$(CONFIG_MEMTEST)		+= memtest.o
 obj-$(CONFIG_MIGRATION) += migrate.o
 obj-$(CONFIG_TRANSPARENT_HUGEPAGE) += huge_memory.o khugepaged.o
mm/memory_hotplug.c +9 -1
index 63e5a0e9a6f3..53ed9dcd3824 100644
@@ -42,7 +42,15 @@
 #include "internal.h"
 #include "shuffle.h"
 
-static bool memmap_on_memory;
+
+/*
+ * memory_hotplug.memmap_on_memory parameter
+ */
+static bool memmap_on_memory __ro_after_init;
+#ifdef CONFIG_MHP_MEMMAP_ON_MEMORY
+module_param(memmap_on_memory, bool, 0444);
+MODULE_PARM_DESC(memmap_on_memory, "Enable memmap on memory for memory hotplug");
+#endif
 
 /*
  * online_page_callback contains pointer to current page onlining function.
Signature · 1 line
2.16.3
author maintainer
#15 lore raw

PATCH v4 4/5 x86/Kconfig: Introduce ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE

In reply to Oscar Salvador

Enable x86_64 platform to use the MHP_MEMMAP_ON_MEMORY feature.

---
 arch/x86/Kconfig | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/Kconfig +3 -0
index 2792879d398e..9f0211df1746 100644
@@ -2433,6 +2433,9 @@ config ARCH_ENABLE_MEMORY_HOTREMOVE
 	def_bool y
 	depends on MEMORY_HOTPLUG
 
+config ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE
+	def_bool y
+
 config USE_PERCPU_NUMA_NODE_ID
 	def_bool y
 	depends on NUMA
Signature · 1 line
2.16.3
author maintainer
#16 lore raw

PATCH v4 5/5 arm64/Kconfig: Introduce ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE

In reply to Oscar Salvador

Enable arm64 platform to use the MHP_MEMMAP_ON_MEMORY feature.

---
 arch/arm64/Kconfig | 3 +++
 1 file changed, 3 insertions(+)
arch/arm64/Kconfig +3 -0
index 1f212b47a48a..c4ade8b47cf8 100644
@@ -309,6 +309,9 @@ config ARCH_ENABLE_MEMORY_HOTPLUG
 config ARCH_ENABLE_MEMORY_HOTREMOVE
 	def_bool y
 
+config ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE
+	def_bool y
+
 config SMP
 	def_bool y
 
Signature · 1 line
2.16.3