OSDN Git Service

Revert "net: phy: Correctly process PHY_HALTED in phy_stop_machine()"
[sagit-ice-cold/kernel_xiaomi_msm8998.git] / mm / memory.c
index c387430..9ac5517 100644 (file)
@@ -797,6 +797,46 @@ out:
        return pfn_to_page(pfn);
 }
 
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
+                               pmd_t pmd)
+{
+       unsigned long pfn = pmd_pfn(pmd);
+
+       /*
+        * There is no pmd_special() but there may be special pmds, e.g.
+        * in a direct-access (dax) mapping, so let's just replicate the
+        * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
+        */
+       if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
+               if (vma->vm_flags & VM_MIXEDMAP) {
+                       if (!pfn_valid(pfn))
+                               return NULL;
+                       goto out;
+               } else {
+                       unsigned long off;
+                       off = (addr - vma->vm_start) >> PAGE_SHIFT;
+                       if (pfn == vma->vm_pgoff + off)
+                               return NULL;
+                       if (!is_cow_mapping(vma->vm_flags))
+                               return NULL;
+               }
+       }
+
+       if (is_zero_pfn(pfn))
+               return NULL;
+       if (unlikely(pfn > highest_memmap_pfn))
+               return NULL;
+
+       /*
+        * NOTE! We still have PageReserved() pages in the page tables.
+        * eg. VDSO mappings can cause them to exist.
+        */
+out:
+       return pfn_to_page(pfn);
+}
+#endif
+
 /*
  * copy one vm_area from one task to the other. Assumes the page tables
  * already present in the new task to be cleared in the whole range
@@ -1087,6 +1127,7 @@ again:
        init_rss_vec(rss);
        start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
        pte = start_pte;
+       flush_tlb_batched_pending(mm);
        arch_enter_lazy_mmu_mode();
        do {
                pte_t ptent = *pte;
@@ -2622,40 +2663,6 @@ out_release:
 }
 
 /*
- * This is like a special single-page "expand_{down|up}wards()",
- * except we must first make sure that 'address{-|+}PAGE_SIZE'
- * doesn't hit another vma.
- */
-static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
-{
-       address &= PAGE_MASK;
-       if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
-               struct vm_area_struct *prev = vma->vm_prev;
-
-               /*
-                * Is there a mapping abutting this one below?
-                *
-                * That's only ok if it's the same stack mapping
-                * that has gotten split..
-                */
-               if (prev && prev->vm_end == address)
-                       return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
-
-               return expand_downwards(vma, address - PAGE_SIZE);
-       }
-       if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
-               struct vm_area_struct *next = vma->vm_next;
-
-               /* As VM_GROWSDOWN but s/below/above/ */
-               if (next && next->vm_start == address + PAGE_SIZE)
-                       return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
-
-               return expand_upwards(vma, address + PAGE_SIZE);
-       }
-       return 0;
-}
-
-/*
  * We enter with non-exclusive mmap_sem (to exclude vma changes,
  * but allow concurrent faults), and pte mapped but not yet locked.
  * We return with mmap_sem still held, but pte unmapped and unlocked.
@@ -2675,10 +2682,6 @@ static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
        if (vma->vm_flags & VM_SHARED)
                return VM_FAULT_SIGBUS;
 
-       /* Check if we need to add a guard page to the stack */
-       if (check_stack_guard_page(vma, address) < 0)
-               return VM_FAULT_SIGSEGV;
-
        /* Use the zero-page for reads */
        if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
                entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
@@ -3399,8 +3402,18 @@ static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        if (unlikely(pmd_none(*pmd)) &&
            unlikely(__pte_alloc(mm, vma, pmd, address)))
                return VM_FAULT_OOM;
-       /* if an huge pmd materialized from under us just retry later */
-       if (unlikely(pmd_trans_huge(*pmd)))
+       /*
+        * If a huge pmd materialized under us just retry later.  Use
+        * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
+        * didn't become pmd_trans_huge under us and then back to pmd_none, as
+        * a result of MADV_DONTNEED running immediately after a huge pmd fault
+        * in a different thread of this mm, in turn leading to a misleading
+        * pmd_trans_huge() retval.  All we have to ensure is that it is a
+        * regular pmd that we can walk with pte_offset_map() and we can do that
+        * through an atomic read in C, which is what pmd_trans_unstable()
+        * provides.
+        */
+       if (unlikely(pmd_trans_unstable(pmd)))
                return 0;
        /*
         * A regular pmd is established and it can't morph into a huge pmd