Re: [PATCH v3 1/2] mm: Abstract THP allocation

From: Dev Jain
Date: Wed Sep 11 2024 - 08:03:31 EST


On 9/11/24 14:59, David Hildenbrand wrote:
On 11.09.24 11:27, David Hildenbrand wrote:
On 11.09.24 08:55, Dev Jain wrote:
In preparation for the second patch, abstract away the THP allocation
logic present in the create_huge_pmd() path, which corresponds to the
faulting case when no page is present.

There should be no functional change as a result of applying
this patch.


Hi,

Signed-off-by: Dev Jain <dev.jain@xxxxxxx>
---
   mm/huge_memory.c | 110 +++++++++++++++++++++++++++++------------------
   1 file changed, 67 insertions(+), 43 deletions(-)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 67c86a5d64a6..b96a1ff2bf40 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -943,47 +943,88 @@ unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
   }
   EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
   -static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
-            struct page *page, gfp_t gfp)
+static struct folio *pmd_thp_fault_alloc(gfp_t gfp, struct vm_area_struct *vma,
+                     unsigned long haddr, unsigned long addr)

I suggest calling this something like "vma_alloc_anon_folio_pmd()"? Then
it's more consistent with vma_alloc_folio().

Also, likely we should just only pass in "addr" and calculate "haddr"
ourselves, it's cheap and reduces the number of function parameters.

   {
-    struct vm_area_struct *vma = vmf->vma;
-    struct folio *folio = page_folio(page);
-    pgtable_t pgtable;
-    unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
-    vm_fault_t ret = 0;
+    const int order = HPAGE_PMD_ORDER;
+    struct folio *folio = vma_alloc_folio(gfp, order, vma, haddr, true);
   -    VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
+    if (unlikely(!folio)) {
+        count_vm_event(THP_FAULT_FALLBACK);
+        count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK);
+        goto out;
+    }
   +    VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
       if (mem_cgroup_charge(folio, vma->vm_mm, gfp)) {
           folio_put(folio);
           count_vm_event(THP_FAULT_FALLBACK);
           count_vm_event(THP_FAULT_FALLBACK_CHARGE);
-        count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK);
-        count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
-        return VM_FAULT_FALLBACK;
+        count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK);
+        count_mthp_stat(order, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
+        goto out;
       }
       folio_throttle_swaprate(folio, gfp);
   -    pgtable = pte_alloc_one(vma->vm_mm);
-    if (unlikely(!pgtable)) {
-        ret = VM_FAULT_OOM;
-        goto release;
-    }
-
-    folio_zero_user(folio, vmf->address);
+    folio_zero_user(folio, addr);
       /*
        * The memory barrier inside __folio_mark_uptodate makes sure that
        * folio_zero_user writes become visible before the set_pmd_at()
        * write.
        */
       __folio_mark_uptodate(folio);
+out:
+    return folio;
+}
+
+static void __pmd_thp_fault_success_stats(struct vm_area_struct *vma)
+{
+    count_vm_event(THP_FAULT_ALLOC);
+    count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_ALLOC);
+    count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
+}

Why isn't that moved into map_pmd_thp()

Note that in this patch you do:

map_pmd_thp(folio, vmf, vma, haddr);
spin_unlock(vmf->ptl);
__pmd_thp_fault_success_stats(vma);

But in patch #2

map_pmd_thp(folio, vmf, vma, haddr);
__pmd_thp_fault_success_stats(vma);
goto unlock;
release:
    folio_put(folio);
unlock:
    spin_unlock(vmf->ptl);

Please make that consistent, meaning:

1) Inline __pmd_thp_fault_success_stats() into map_pmd_thp(). No need to
have the separated out.

2) Either do the PTL unlocking in __pmd_thp_fault_success_stats() or in
     the caller. In the caller is likely easiest. Adjusting the counters
     should be cheap, if not we could revisit this later with real data.

+
+static void map_pmd_thp(struct folio *folio, struct vm_fault *vmf,
+            struct vm_area_struct *vma, unsigned long haddr)
+{
+    pmd_t entry;
+
+    entry = mk_huge_pmd(&folio->page, vma->vm_page_prot);
+    entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
+    folio_add_new_anon_rmap(folio, vma, haddr, RMAP_EXCLUSIVE);
+    folio_add_lru_vma(folio, vma);
+    set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
+    update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);

It's quite weird to see a mixture of haddr and vmf->address, and likely
this mixture is wrong or not not required.

Looking at arc's update_mmu_cache_pmd() implementation, I cannot see how
passing in the unaligned address would do the right thing. But maybe arc
also doesn't trigger that code path ... who knows :)


Staring at some other update_mmu_cache_pmd() users, it's quite
inconsistent. Primarily only do_huge_pmd_numa_page() and
__do_huge_pmd_anonymous_page() use the unaligned address. The others
seem to use the aligned address ... as one would expect when modifying a
PMD.


I suggest to change this function to *not* pass in the vmf, and rename
it to something like:

static void folio_map_anon_pmd(struct folio *folio, struct
vm_area_struct *vma, pmd_t *pmd, unsigned long haddr)

... or better "map_anon_folio_pmd" so it better matches vma_alloc_folio_pmd() suggested above.

I'll vote for this.