On 19.7.2021 15.56, Qi Zheng wrote:

On 7/18/21 2:19 PM, Mika Penttilä wrote:

quoted

quoted

+
+/*
+ * returns true if the pmd has been populated with PTE page table,
+ * or false for all other cases.
+ */
+bool pte_install_try_get(struct mm_struct *mm, pmd_t *pmd, 
pgtable_t *pte)
+{
+    spinlock_t *ptl;
+    bool retval = true;
+
+retry:
+    ptl = pmd_lock(mm, pmd);
+    if (likely(pmd_none(*pmd))) {
+        __pte_install(mm, pmd, pte);
+    } else if (pmd_leaf(*pmd) || !pmd_present(*pmd)) {
+        retval = false;
+    } else if (!pte_get_unless_zero(pmd)) {
+        spin_unlock(ptl);
+        goto retry;
+    }
+    spin_unlock(ptl);
+    return retval;
+}
+

Can pte_get_unless_zero() return true above? Can the pmd have been by 
populated by others? In that case the ref count is wrongly incremented.

Here we only have mmap_read_lock(mm), so the pmd can be populated with
other PTE page table page after a page fault in a different thread B 
of this mm. In this case, thread B already hold a pte_refcount of the 
PTE page table page populated in the pmd, so pte_get_unless_zero() can
return true above.

Yes but if thread B populates the page table page and pte, then we also 
increase the refcount with pte_get_unless_zero() , but dont decrease it 
when notice !pte_none().
And in the pte_none() case, the refcount is increased again, so double 
accounting. see finish_fault().

Similarly, if THP is enabled, the pmd also can be populated with a THP 
page, we can see more detail in comment in handle_pte_fault(). The
pmd_leaf() above is to detect this situation.