Re: [PATCH v8 04/10] mm: thp: Support allocation of anonymous multi-size THP

From: David Hildenbrand
Date: Tue Dec 05 2023 - 11:33:19 EST

On 04.12.23 11:20, Ryan Roberts wrote:
Introduce the logic to allow THP to be configured (through the new sysfs
interface we just added) to allocate large folios to back anonymous
memory, which are larger than the base page size but smaller than
PMD-size. We call this new THP extension "multi-size THP" (mTHP).

mTHP continues to be PTE-mapped, but in many cases can still provide
similar benefits to traditional PMD-sized THP: Page faults are
significantly reduced (by a factor of e.g. 4, 8, 16, etc. depending on
the configured order), but latency spikes are much less prominent
because the size of each page isn't as huge as the PMD-sized variant and
there is less memory to clear in each page fault. The number of per-page
operations (e.g. ref counting, rmap management, lru list management) are
also significantly reduced since those ops now become per-folio.

Some architectures also employ TLB compression mechanisms to squeeze
more entries in when a set of PTEs are virtually and physically
contiguous and approporiately aligned. In this case, TLB misses will
occur less often.

The new behaviour is disabled by default, but can be enabled at runtime
by writing to /sys/kernel/mm/transparent_hugepage/hugepage-XXkb/enabled
(see documentation in previous commit). The long term aim is to change
the default to include suitable lower orders, but there are some risks
around internal fragmentation that need to be better understood first.

Signed-off-by: Ryan Roberts <ryan.roberts@xxxxxxx>

In general, looks good to me, some comments/nits. And the usual "let's make sure we don't degrade order-0 and keep that as fast as possible" comment.

---
include/linux/huge_mm.h | 6 ++-
mm/memory.c | 106 ++++++++++++++++++++++++++++++++++++----
2 files changed, 101 insertions(+), 11 deletions(-)

diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
index bd0eadd3befb..91a53b9835a4 100644
--- a/include/linux/huge_mm.h
+++ b/include/linux/huge_mm.h
@@ -68,9 +68,11 @@ extern struct kobj_attribute shmem_enabled_attr;
#define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
/*
- * Mask of all large folio orders supported for anonymous THP.
+ * Mask of all large folio orders supported for anonymous THP; all orders up to
+ * and including PMD_ORDER, except order-0 (which is not "huge") and order-1
+ * (which is a limitation of the THP implementation).
*/
-#define THP_ORDERS_ALL_ANON BIT(PMD_ORDER)
+#define THP_ORDERS_ALL_ANON ((BIT(PMD_ORDER + 1) - 1) & ~(BIT(0) | BIT(1)))
/*
* Mask of all large folio orders supported for file THP.
diff --git a/mm/memory.c b/mm/memory.c
index 3ceeb0f45bf5..bf7e93813018 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -4125,6 +4125,84 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
return ret;
}
+static bool pte_range_none(pte_t *pte, int nr_pages)
+{
+ int i;
+
+ for (i = 0; i < nr_pages; i++) {
+ if (!pte_none(ptep_get_lockless(pte + i)))
+ return false;
+ }
+
+ return true;
+}
+
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+static struct folio *alloc_anon_folio(struct vm_fault *vmf)
+{
+ gfp_t gfp;
+ pte_t *pte;
+ unsigned long addr;
+ struct folio *folio;
+ struct vm_area_struct *vma = vmf->vma;
+ unsigned long orders;
+ int order;

Nit: reverse christmas tree encouraged ;)

+
+ /*
+ * If uffd is active for the vma we need per-page fault fidelity to
+ * maintain the uffd semantics.
+ */
+ if (userfaultfd_armed(vma))

Nit: unlikely()

+ goto fallback;
+
+ /*
+ * Get a list of all the (large) orders below PMD_ORDER that are enabled
+ * for this vma. Then filter out the orders that can't be allocated over
+ * the faulting address and still be fully contained in the vma.
+ */
+ orders = thp_vma_allowable_orders(vma, vma->vm_flags, false, true, true,
+ BIT(PMD_ORDER) - 1);
+ orders = thp_vma_suitable_orders(vma, vmf->address, orders);

Comment: Both will eventually loop over all orders, correct? Could eventually be sped up in the future.

Nit: the orders = ... order = ... looks like this might deserve a helper function that makes this easier to read.

Nit: Why call thp_vma_suitable_orders if the orders are already 0? Again, some helper might be reasonable where that is handled internally.

Comment: For order-0 we'll always perform a function call to both thp_vma_allowable_orders() / thp_vma_suitable_orders(). We should perform some fast and efficient check if any <PMD THP are even enabled in the system / for this VMA, and in that case just fallback before doing more expensive checks.

+
+ if (!orders)
+ goto fallback;
+
+ pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK);
+ if (!pte)
+ return ERR_PTR(-EAGAIN);
+
+ order = first_order(orders);
+ while (orders) {
+ addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
+ vmf->pte = pte + pte_index(addr);
+ if (pte_range_none(vmf->pte, 1 << order))
+ break;

Comment: Likely it would make sense to scan only once and determine the "largest none range" around that address, having the largest suitable order in mind.

+ order = next_order(&orders, order);
+ }
+
+ vmf->pte = NULL;

Nit: Can you elaborate why you are messing with vmf->pte here? A simple helper variable will make this code look less magical. Unless I am missing something important :)

+ pte_unmap(pte);
+
+ gfp = vma_thp_gfp_mask(vma);
+
+ while (orders) {
+ addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
+ folio = vma_alloc_folio(gfp, order, vma, addr, true);
+ if (folio) {
+ clear_huge_page(&folio->page, addr, 1 << order);
+ return folio;
+ }
+ order = next_order(&orders, order);
+ }
+

Queestion: would it make sense to combine both loops? I suspect memory allocations with pte_offset_map()/kmao are problematic.

+fallback:
+ return vma_alloc_zeroed_movable_folio(vma, vmf->address);
+}
+#else
+#define alloc_anon_folio(vmf) \
+ vma_alloc_zeroed_movable_folio((vmf)->vma, (vmf)->address)
+#endif
+
/*
* We enter with non-exclusive mmap_lock (to exclude vma changes,
* but allow concurrent faults), and pte mapped but not yet locked.
@@ -4132,6 +4210,9 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
*/
static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
{
+ int i;
+ int nr_pages = 1;
+ unsigned long addr = vmf->address;
bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
struct vm_area_struct *vma = vmf->vma;
struct folio *folio;

Nit: reverse christmas tree :)

@@ -4176,10 +4257,15 @@ static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
/* Allocate our own private page. */
if (unlikely(anon_vma_prepare(vma)))
goto oom;
- folio = vma_alloc_zeroed_movable_folio(vma, vmf->address);
+ folio = alloc_anon_folio(vmf);
+ if (IS_ERR(folio))
+ return 0;
if (!folio)
goto oom;
+ nr_pages = folio_nr_pages(folio);
+ addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE);
+
if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL))
goto oom_free_page;
folio_throttle_swaprate(folio, GFP_KERNEL);
@@ -4196,12 +4282,13 @@ static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
if (vma->vm_flags & VM_WRITE)
entry = pte_mkwrite(pte_mkdirty(entry), vma);
- vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
- &vmf->ptl);
+ vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
if (!vmf->pte)
goto release;
- if (vmf_pte_changed(vmf)) {
- update_mmu_tlb(vma, vmf->address, vmf->pte);
+ if ((nr_pages == 1 && vmf_pte_changed(vmf)) ||
+ (nr_pages > 1 && !pte_range_none(vmf->pte, nr_pages))) {
+ for (i = 0; i < nr_pages; i++)
+ update_mmu_tlb(vma, addr + PAGE_SIZE * i, vmf->pte + i);

Comment: separating the order-0 case from the other case might make this easier to read.

goto release;
}
@@ -4216,16 +4303,17 @@ static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
return handle_userfault(vmf, VM_UFFD_MISSING);
}
- inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
- folio_add_new_anon_rmap(folio, vma, vmf->address);
+ folio_ref_add(folio, nr_pages - 1);
+ add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages);
+ folio_add_new_anon_rmap(folio, vma, addr);
folio_add_lru_vma(folio, vma);
setpte:
if (uffd_wp)
entry = pte_mkuffd_wp(entry);
- set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
+ set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr_pages);
/* No need to invalidate - it was non-present before */
- update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
+ update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr_pages);
unlock:
if (vmf->pte)
pte_unmap_unlock(vmf->pte, vmf->ptl);

Benchmarking order-0 allocations might be interesting. There will be some added checks + multiple loops/conditionals for order-0 that could be avoided by having two separate code paths. If we can't measure a difference, all good.

--
Cheers,

David / dhildenb