hugetlbfs: revert use i_mmap_rwsem for more pmd sharing synchronization

Commit c0d0381ade79 ("hugetlbfs: use i_mmap_rwsem for more pmd sharing
synchronization") added code to take i_mmap_rwsem in read mode for the
duration of fault processing.  However, this has been shown to cause
performance/scaling issues.  Revert the code and go back to only taking
the semaphore in huge_pmd_share during the fault path.

Keep the code that takes i_mmap_rwsem in write mode before calling
try_to_unmap as this is required if huge_pmd_unshare is called.

NOTE: Reverting this code does expose the following race condition.

Faulting thread                                 Unsharing thread
...                                                  ...
ptep = huge_pte_offset()
      or
ptep = huge_pte_alloc()
...
                                                i_mmap_lock_write
                                                lock page table
ptep invalid   <------------------------        huge_pmd_unshare()
Could be in a previously                        unlock_page_table
sharing process or worse                        i_mmap_unlock_write
...
ptl = huge_pte_lock(ptep)
get/update pte
set_pte_at(pte, ptep)

It is unknown if the above race was ever experienced by a user.  It was
discovered via code inspection when initially addressed.

In subsequent patches, a new synchronization mechanism will be added to
coordinate pmd sharing and eliminate this race.

Link: https://lkml.kernel.org/r/20220914221810.95771-3-mike.kravetz@oracle.com
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: James Houghton <jthoughton@google.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mina Almasry <almasrymina@google.com>
Cc: Muchun Song <songmuchun@bytedance.com>
Cc: Naoya Horiguchi <naoya.horiguchi@linux.dev>
Cc: Pasha Tatashin <pasha.tatashin@soleen.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Prakash Sangappa <prakash.sangappa@oracle.com>
Cc: Sven Schnelle <svens@linux.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

1 files changed, 12 insertions, 65 deletions

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 14afb5b67dd4..8283706bd81d 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -4770,7 +4770,6 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 	struct hstate *h = hstate_vma(src_vma);
 	unsigned long sz = huge_page_size(h);
 	unsigned long npages = pages_per_huge_page(h);
-	struct address_space *mapping = src_vma->vm_file->f_mapping;
 	struct mmu_notifier_range range;
 	unsigned long last_addr_mask;
 	int ret = 0;
@@ -4782,14 +4781,6 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 		mmu_notifier_invalidate_range_start(&range);
 		mmap_assert_write_locked(src);
 		raw_write_seqcount_begin(&src->write_protect_seq);
-	} else {
-		/*
-		 * For shared mappings i_mmap_rwsem must be held to call
-		 * huge_pte_alloc, otherwise the returned ptep could go
-		 * away if part of a shared pmd and another thread calls
-		 * huge_pmd_unshare.
-		 */
-		i_mmap_lock_read(mapping);
 	}
 
 	last_addr_mask = hugetlb_mask_last_page(h);
@@ -4936,8 +4927,6 @@ again:
 	if (cow) {
 		raw_write_seqcount_end(&src->write_protect_seq);
 		mmu_notifier_invalidate_range_end(&range);
-	} else {
-		i_mmap_unlock_read(mapping);
 	}
 
 	return ret;
@@ -5346,29 +5335,8 @@ retry_avoidcopy:
 		 * may get SIGKILLed if it later faults.
 		 */
 		if (outside_reserve) {
-			struct address_space *mapping = vma->vm_file->f_mapping;
-			pgoff_t idx;
-			u32 hash;
-
 			put_page(old_page);
-			/*
-			 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
-			 * unmapping.  unmapping needs to hold i_mmap_rwsem
-			 * in write mode.  Dropping i_mmap_rwsem in read mode
-			 * here is OK as COW mappings do not interact with
-			 * PMD sharing.
-			 *
-			 * Reacquire both after unmap operation.
-			 */
-			idx = vma_hugecache_offset(h, vma, haddr);
-			hash = hugetlb_fault_mutex_hash(mapping, idx);
-			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
-			i_mmap_unlock_read(mapping);
-
 			unmap_ref_private(mm, vma, old_page, haddr);
-
-			i_mmap_lock_read(mapping);
-			mutex_lock(&hugetlb_fault_mutex_table[hash]);
 			spin_lock(ptl);
 			ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
 			if (likely(ptep &&
@@ -5523,9 +5491,7 @@ static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
 	 */
 	hash = hugetlb_fault_mutex_hash(mapping, idx);
 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
-	i_mmap_unlock_read(mapping);
 	ret = handle_userfault(&vmf, reason);
-	i_mmap_lock_read(mapping);
 	mutex_lock(&hugetlb_fault_mutex_table[hash]);
 
 	return ret;
@@ -5760,11 +5726,6 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 
 	ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
 	if (ptep) {
-		/*
-		 * Since we hold no locks, ptep could be stale.  That is
-		 * OK as we are only making decisions based on content and
-		 * not actually modifying content here.
-		 */
 		entry = huge_ptep_get(ptep);
 		if (unlikely(is_hugetlb_entry_migration(entry))) {
 			migration_entry_wait_huge(vma, ptep);
@@ -5772,31 +5733,20 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 		} else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
 			return VM_FAULT_HWPOISON_LARGE |
 				VM_FAULT_SET_HINDEX(hstate_index(h));
+	} else {
+		ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
+		if (!ptep)
+			return VM_FAULT_OOM;
 	}
 
-	/*
-	 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
-	 * until finished with ptep.  This prevents huge_pmd_unshare from
-	 * being called elsewhere and making the ptep no longer valid.
-	 *
-	 * ptep could have already be assigned via huge_pte_offset.  That
-	 * is OK, as huge_pte_alloc will return the same value unless
-	 * something has changed.
-	 */
 	mapping = vma->vm_file->f_mapping;
-	i_mmap_lock_read(mapping);
-	ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
-	if (!ptep) {
-		i_mmap_unlock_read(mapping);
-		return VM_FAULT_OOM;
-	}
+	idx = vma_hugecache_offset(h, vma, haddr);
 
 	/*
 	 * Serialize hugepage allocation and instantiation, so that we don't
 	 * get spurious allocation failures if two CPUs race to instantiate
 	 * the same page in the page cache.
 	 */
-	idx = vma_hugecache_offset(h, vma, haddr);
 	hash = hugetlb_fault_mutex_hash(mapping, idx);
 	mutex_lock(&hugetlb_fault_mutex_table[hash]);
 
@@ -5861,7 +5811,6 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 			put_page(pagecache_page);
 		}
 		mutex_unlock(&hugetlb_fault_mutex_table[hash]);
-		i_mmap_unlock_read(mapping);
 		return handle_userfault(&vmf, VM_UFFD_WP);
 	}
 
@@ -5905,7 +5854,6 @@ out_ptl:
 	}
 out_mutex:
 	mutex_unlock(&hugetlb_fault_mutex_table[hash]);
-	i_mmap_unlock_read(mapping);
 	/*
 	 * Generally it's safe to hold refcount during waiting page lock. But
 	 * here we just wait to defer the next page fault to avoid busy loop and
@@ -6745,12 +6693,10 @@ void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
  * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
  * and returns the corresponding pte. While this is not necessary for the
  * !shared pmd case because we can allocate the pmd later as well, it makes the
- * code much cleaner.
- *
- * This routine must be called with i_mmap_rwsem held in at least read mode if
- * sharing is possible.  For hugetlbfs, this prevents removal of any page
- * table entries associated with the address space.  This is important as we
- * are setting up sharing based on existing page table entries (mappings).
+ * code much cleaner. pmd allocation is essential for the shared case because
+ * pud has to be populated inside the same i_mmap_rwsem section - otherwise
+ * racing tasks could either miss the sharing (see huge_pte_offset) or select a
+ * bad pmd for sharing.
  */
 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
 		      unsigned long addr, pud_t *pud)
@@ -6764,7 +6710,7 @@ pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
 	pte_t *pte;
 	spinlock_t *ptl;
 
-	i_mmap_assert_locked(mapping);
+	i_mmap_lock_read(mapping);
 	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
 		if (svma == vma)
 			continue;
@@ -6794,6 +6740,7 @@ pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
 	spin_unlock(ptl);
 out:
 	pte = (pte_t *)pmd_alloc(mm, pud, addr);
+	i_mmap_unlock_read(mapping);
 	return pte;
 }
 
@@ -6804,7 +6751,7 @@ out:
  * indicated by page_count > 1, unmap is achieved by clearing pud and
  * decrementing the ref count. If count == 1, the pte page is not shared.
  *
- * Called with page table lock held and i_mmap_rwsem held in write mode.
+ * Called with page table lock held.
  *
  * returns: 1 successfully unmapped a shared pte page
  *	    0 the underlying pte page is not shared, or it is the last user