[PATCH RFC v3 0/2] hugetlbfs: use i_mmap_rwsem for more synchronization

[PATCH RFC v3 0/2] hugetlbfs: use i_mmap_rwsem for more synchronization

[Mike Kravetz]

I'm kicking this back to RFC status as there may be a couple other
options we want to explore.  This expanded use of i_mmap_rwsem in
hugetlbfs was recently pushed upstream and reverted due to locking
issues.
http://lkml.kernel.org/r/20181218223557.5202-3-mike.kravetz@oracle.com
http://lkml.kernel.org/r/20181222223013.22193-3-mike.kravetz@oracle.com

The biggest problem with those patches was lock ordering.  Recall, the
two issues we are trying to address:

1) For shared pmds, huge PTE pointers returned by huge_pte_alloc can become
   invalid via a call to huge_pmd_unshare by another thread.
2) hugetlbfs page faults can race with truncation causing invalid global
   reserve counts and state.

To effectively use i_mmap_rwsem to address these issues it needs to
be held (in read mode) during page fault processing.  However, during
fault processing we need to lock the page we will be adding.  Lock
ordering requires we take page lock before i_mmap_rwsem.  Waiting until
after taking the page lock is too late in the fault process for the
synchronization we want to do.

To address this lock ordering issue, the following patches change the
lock ordering for hugetlb pages.  This is not difficult as hugetlbfs
processing is done separate from core mm in many places.  However, I
don't really like this idea.  Much ugliness is contained in the new
routine hugetlb_page_mapping_lock_write() of patch 1.

If this approach of extending the usage of i_mmap_rwsem is not acceptable,
there are two other options I can think of.
- Add a new rw-semaphore that lives in the hugetlbfs specific inode
  extension.  The good thing is that this semaphore would be hugetlbfs
  specific and not directly exposed to the rest of the mm code.  Therefore,
  interaction with other locking schemes is minimized.
- Don't really add any new synchronization, but notice and catch all the
  races.  After catching the race, cleanup, backout, retry ... etc, as
  needed.  This can get really ugly, especially for huge page reservations.
  At one time, I started writing some of the reservation backout code for
  page faults and it got so ugly and complicated I went down the path of
  adding synchronization to avoid the races.

Suggestions on how to proceed would be appreciated.  If you think the
following patches are not too ugly, comments on those would also be
welcome.

Mike Kravetz (2):
  hugetlbfs: use i_mmap_rwsem for more pmd sharing synchronization
  hugetlbfs: Use i_mmap_rwsem to fix page fault/truncate race

 fs/hugetlbfs/inode.c    |  34 +++++---
 include/linux/fs.h      |   5 ++
 include/linux/hugetlb.h |   8 ++
 mm/hugetlb.c            | 186 ++++++++++++++++++++++++++++++++++------
 mm/memory-failure.c     |  29 ++++++-
 mm/migrate.c            |  24 +++++-
 mm/rmap.c               |  17 +++-
 mm/userfaultfd.c        |  11 ++-
 8 files changed, 270 insertions(+), 44 deletions(-)

-- 
2.17.2

[PATCH RFC v3 1/2] hugetlbfs: use i_mmap_rwsem for more pmd sharing synchronization

[Mike Kravetz]

While looking at BUGs associated with invalid huge page map counts,
it was discovered and observed that a huge pte pointer could become
'invalid' and point to another task's page table.  Consider the
following:

A task takes a page fault on a shared hugetlbfs file and calls
huge_pte_alloc to get a ptep.  Suppose the returned ptep points to a
shared pmd.

Now, another task truncates the hugetlbfs file.  As part of truncation,
it unmaps everyone who has the file mapped.  If the range being
truncated is covered by a shared pmd, huge_pmd_unshare will be called.
For all but the last user of the shared pmd, huge_pmd_unshare will
clear the pud pointing to the pmd.  If the task in the middle of the
page fault is not the last user, the ptep returned by huge_pte_alloc
now points to another task's page table or worse.  This leads to bad
things such as incorrect page map/reference counts or invalid memory
references.

To fix, expand the use of i_mmap_rwsem as follows:
- i_mmap_rwsem is held in read mode whenever huge_pmd_share is called.
  huge_pmd_share is only called via huge_pte_alloc, so callers of
  huge_pte_alloc take i_mmap_rwsem before calling.  In addition, callers
  of huge_pte_alloc continue to hold the semaphore until finished with
  the ptep.
- i_mmap_rwsem is held in write mode whenever huge_pmd_unshare is called.

One problem with this scheme is that it requires taking i_mmap_rwsem
before taking the page lock during page faults.  This is not the order
specified in the rest of mm code.  Handling of hugetlbfs pages is mostly
isolated today.  Therefore, we use this alternative locking order for
PageHuge() pages.
         mapping->i_mmap_rwsem
           hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
             page->flags PG_locked (lock_page)

To help with lock ordering issues, hugetlb_page_mapping_lock_write()
is introduced to write lock the i_mmap_rwsem associated with a page.

In most cases it is easy to get address_space via vma->vm_file->f_mapping.
However, in the case of migration or memory errors for anon pages we do
not have an associated vma.  A new routine _get_hugetlb_page_mapping()
will use anon_vma to get address_space in these cases.

Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
---
 fs/hugetlbfs/inode.c    |   2 +
 include/linux/fs.h      |   5 ++
 include/linux/hugetlb.h |   8 ++
 mm/hugetlb.c            | 167 ++++++++++++++++++++++++++++++++++++----
 mm/memory-failure.c     |  29 ++++++-
 mm/migrate.c            |  24 +++++-
 mm/rmap.c               |  17 +++-
 mm/userfaultfd.c        |  11 ++-
 8 files changed, 239 insertions(+), 24 deletions(-)

diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c
index fb6de1db8806..5280fe3aad2f 100644
--- a/fs/hugetlbfs/inode.c
+++ b/fs/hugetlbfs/inode.c
@@ -443,7 +443,9 @@ static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 			if (unlikely(page_mapped(page))) {
 				BUG_ON(truncate_op);
 
+				mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 				i_mmap_lock_write(mapping);
+				mutex_lock(&hugetlb_fault_mutex_table[hash]);
 				hugetlb_vmdelete_list(&mapping->i_mmap,
 					index * pages_per_huge_page(h),
 					(index + 1) * pages_per_huge_page(h));
diff --git a/include/linux/fs.h b/include/linux/fs.h
index c95c0807471f..2369ae58a24d 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -501,6 +501,11 @@ static inline void i_mmap_lock_write(struct address_space *mapping)
 	down_write(&mapping->i_mmap_rwsem);
 }
 
+static inline int i_mmap_trylock_write(struct address_space *mapping)
+{
+	return down_write_trylock(&mapping->i_mmap_rwsem);
+}
+
 static inline void i_mmap_unlock_write(struct address_space *mapping)
 {
 	up_write(&mapping->i_mmap_rwsem);
diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
index 087fd5f48c91..c6959eb23344 100644
--- a/include/linux/hugetlb.h
+++ b/include/linux/hugetlb.h
@@ -130,6 +130,8 @@ u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
 
 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud);
 
+struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage);
+
 extern int sysctl_hugetlb_shm_group;
 extern struct list_head huge_boot_pages;
 
@@ -172,6 +174,12 @@ static inline unsigned long hugetlb_total_pages(void)
 	return 0;
 }
 
+static inline struct address_space *hugetlb_page_mapping_lock_write(
+							struct page *hpage)
+{
+	return NULL;
+}
+
 static inline int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr,
 					pte_t *ptep)
 {
diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index a80832487981..9198b00d25e6 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -1356,6 +1356,106 @@ int PageHeadHuge(struct page *page_head)
 	return get_compound_page_dtor(page_head) == free_huge_page;
 }
 
+/*
+ * Find address_space associated with hugetlbfs page.
+ * Upon entry page is locked and page 'was' mapped although mapped state
+ * could change.  If necessary, use anon_vma to find vma and associated
+ * address space.  The returned mapping may be stale, but it can not be
+ * invalid as page lock (which is held) is required to destroy mapping.
+ */
+static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
+{
+	struct anon_vma *anon_vma;
+	pgoff_t pgoff_start, pgoff_end;
+	struct anon_vma_chain *avc;
+	struct address_space *mapping = page_mapping(hpage);
+
+	/* Simple file based mapping */
+	if (mapping)
+		return mapping;
+
+	/*
+	 * Even anonymous hugetlbfs mappings are associated with an
+	 * underlying hugetlbfs file (see hugetlb_file_setup in mmap
+	 * code).  Find a vma associated with the anonymous vma, and
+	 * use the file pointer to get address_space.
+	 */
+	anon_vma = page_lock_anon_vma_read(hpage);
+	if (!anon_vma)
+		return mapping;  /* NULL */
+
+	/* Use first found vma */
+	pgoff_start = page_to_pgoff(hpage);
+	pgoff_end = pgoff_start + hpage_nr_pages(hpage) - 1;
+	anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
+					pgoff_start, pgoff_end) {
+		struct vm_area_struct *vma = avc->vma;
+
+		mapping = vma->vm_file->f_mapping;
+		break;
+	}
+
+	anon_vma_unlock_read(anon_vma);
+	return mapping;
+}
+
+/*
+ * Find and lock address space (mapping) in write mode.
+ *
+ * Upon entry, the page is locked which allows us to find the mapping
+ * even in the case of an anon page.  However, locking order dictates
+ * the i_mmap_rwsem be acquired BEFORE the page lock.  This is hugetlbfs
+ * specific.  So, we first try to lock the sema while still holding the
+ * page lock.  If this works, great!  If not, then we need to drop the
+ * page lock and then acquire i_mmap_rwsem and reacquire page lock.  Of
+ * course, need to revalidate state along the way.
+ */
+struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
+{
+	struct address_space *mapping, *mapping2;
+
+	mapping = _get_hugetlb_page_mapping(hpage);
+retry:
+	if (!mapping)
+		return mapping;
+
+	/*
+	 * If no contention, take lock and return
+	 */
+	if (i_mmap_trylock_write(mapping))
+		return mapping;
+
+	/*
+	 * Must drop page lock and wait on mapping sema.
+	 * Note:  Once page lock is dropped, mapping could become invalid.
+	 * As a hack, increase map count until we lock page again.
+	 */
+	atomic_inc(&hpage->_mapcount);
+	unlock_page(hpage);
+	i_mmap_lock_write(mapping);
+	lock_page(hpage);
+	atomic_add_negative(-1, &hpage->_mapcount);
+
+	/* verify page is still mapped */
+	if (!page_mapped(hpage)) {
+		i_mmap_unlock_write(mapping);
+		return NULL;
+	}
+
+	/*
+	 * Get address space again and verify it is the same one
+	 * we locked.  If not, drop lock and retry.
+	 */
+	mapping2 = _get_hugetlb_page_mapping(hpage);
+	if (mapping2 != mapping) {
+		i_mmap_unlock_write(mapping);
+		mapping = mapping2;
+		goto retry;
+	}
+
+	return mapping;
+}
+
 pgoff_t __basepage_index(struct page *page)
 {
 	struct page *page_head = compound_head(page);
@@ -3240,6 +3340,7 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 	int cow;
 	struct hstate *h = hstate_vma(vma);
 	unsigned long sz = huge_page_size(h);
+	struct address_space *mapping = vma->vm_file->f_mapping;
 	unsigned long mmun_start;	/* For mmu_notifiers */
 	unsigned long mmun_end;		/* For mmu_notifiers */
 	int ret = 0;
@@ -3248,14 +3349,25 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 
 	mmun_start = vma->vm_start;
 	mmun_end = vma->vm_end;
-	if (cow)
+	if (cow) {
 		mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
+	} 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);
+	}
 
 	for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
 		spinlock_t *src_ptl, *dst_ptl;
+
 		src_pte = huge_pte_offset(src, addr, sz);
 		if (!src_pte)
 			continue;
+
 		dst_pte = huge_pte_alloc(dst, addr, sz);
 		if (!dst_pte) {
 			ret = -ENOMEM;
@@ -3326,6 +3438,8 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 
 	if (cow)
 		mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
+	else
+		i_mmap_unlock_read(mapping);
 
 	return ret;
 }
@@ -3773,14 +3887,18 @@ static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
 			};
 
 			/*
-			 * hugetlb_fault_mutex must be dropped before
-			 * handling userfault.  Reacquire after handling
-			 * fault to make calling code simpler.
+			 * hugetlb_fault_mutex and i_mmap_rwsem must be
+			 * dropped before handling userfault.  Reacquire
+			 * after handling fault to make calling code simpler.
 			 */
 			hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
 							idx, haddr);
 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
+			i_mmap_unlock_read(mapping);
+
 			ret = handle_userfault(&vmf, VM_UFFD_MISSING);
+
+			i_mmap_lock_read(mapping);
 			mutex_lock(&hugetlb_fault_mutex_table[hash]);
 			goto out;
 		}
@@ -3928,6 +4046,11 @@ 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, mm, ptep);
@@ -3935,20 +4058,31 @@ 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, 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 changed.
+	 */
 	mapping = vma->vm_file->f_mapping;
-	idx = vma_hugecache_offset(h, vma, haddr);
+	i_mmap_lock_read(mapping);
+	ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
+	if (!ptep) {
+		i_mmap_unlock_read(mapping);
+		return VM_FAULT_OOM;
+	}
 
 	/*
 	 * 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(h, mm, vma, mapping, idx, haddr);
 	mutex_lock(&hugetlb_fault_mutex_table[hash]);
 
@@ -4036,6 +4170,7 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 	}
 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
@@ -4640,10 +4775,12 @@ 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. 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.
+ * code much cleaner.
+ *
+ * This routine must be called with i_mmap_rwsem held in at least read mode.
+ * 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).
  */
 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 {
@@ -4660,7 +4797,6 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	if (!vma_shareable(vma, addr))
 		return (pte_t *)pmd_alloc(mm, pud, addr);
 
-	i_mmap_lock_write(mapping);
 	vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
 		if (svma == vma)
 			continue;
@@ -4690,7 +4826,6 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	spin_unlock(ptl);
 out:
 	pte = (pte_t *)pmd_alloc(mm, pud, addr);
-	i_mmap_unlock_write(mapping);
 	return pte;
 }
 
@@ -4701,7 +4836,7 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
  * 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.
+ * Called with page table lock held and i_mmap_rwsem held in write mode.
  *
  * returns: 1 successfully unmapped a shared pte page
  *	    0 the underlying pte page is not shared, or it is the last user
diff --git a/mm/memory-failure.c b/mm/memory-failure.c
index 7c72f2a95785..ecca1eb2827e 100644
--- a/mm/memory-failure.c
+++ b/mm/memory-failure.c
@@ -966,7 +966,7 @@ static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
 	enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
 	struct address_space *mapping;
 	LIST_HEAD(tokill);
-	bool unmap_success;
+	bool unmap_success = true;
 	int kill = 1, forcekill;
 	struct page *hpage = *hpagep;
 	bool mlocked = PageMlocked(hpage);
@@ -1028,7 +1028,32 @@ static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
 	if (kill)
 		collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
 
-	unmap_success = try_to_unmap(hpage, ttu);
+	if (!PageHuge(hpage)) {
+		unmap_success = try_to_unmap(hpage, ttu);
+	} else {
+		/*
+		 * For hugetlb pages, try_to_unmap could potentially call
+		 * huge_pmd_unshare.  Because of this, take semaphore in
+		 * write mode here and set TTU_RMAP_LOCKED to indicate we
+		 * have taken the lock at this higer level.
+		 *
+		 * Note that the call to hugetlb_page_mapping_lock_write
+		 * is necessary even if mapping is already set.  It handles
+		 * ugliness of potentially having to drop page lock to obtain
+		 * i_mmap_rwsem.
+		 */
+		mapping = hugetlb_page_mapping_lock_write(hpage);
+
+		if (mapping) {
+			unmap_success = try_to_unmap(hpage,
+						     ttu|TTU_RMAP_LOCKED);
+			i_mmap_unlock_write(mapping);
+		} else {
+			pr_info("Memory failure: %#lx: could not find mapping for mapped huge page\n",
+				pfn);
+			unmap_success = false;
+		}
+	}
 	if (!unmap_success)
 		pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
 		       pfn, page_mapcount(hpage));
diff --git a/mm/migrate.c b/mm/migrate.c
index 0d9708803553..4952d54138b4 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1266,6 +1266,7 @@ static int unmap_and_move_huge_page(new_page_t get_new_page,
 	int page_was_mapped = 0;
 	struct page *new_hpage;
 	struct anon_vma *anon_vma = NULL;
+	struct address_space *mapping = NULL;
 
 	/*
 	 * Movability of hugepages depends on architectures and hugepage size.
@@ -1303,17 +1304,34 @@ static int unmap_and_move_huge_page(new_page_t get_new_page,
 		goto put_anon;
 
 	if (page_mapped(hpage)) {
+		/*
+		 * try_to_unmap could potentially call huge_pmd_unshare.
+		 * Because of this, take semaphore in write mode here and
+		 * set TTU_RMAP_LOCKED to let lower levels know we have
+		 * taken the lock.
+		 */
+		mapping = hugetlb_page_mapping_lock_write(hpage);
+		if (unlikely(!mapping))
+			goto put_anon;
+
 		try_to_unmap(hpage,
-			TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS);
+			TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS|
+			TTU_RMAP_LOCKED);
 		page_was_mapped = 1;
+		/*
+		 * Leave mapping locked until after subsequent call to
+		 * remove_migration_ptes()
+		 */
 	}
 
 	if (!page_mapped(hpage))
 		rc = move_to_new_page(new_hpage, hpage, mode);
 
-	if (page_was_mapped)
+	if (page_was_mapped) {
 		remove_migration_ptes(hpage,
-			rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, false);
+			rc == MIGRATEPAGE_SUCCESS ? new_hpage : hpage, true);
+		i_mmap_unlock_write(mapping);
+	}
 
 	unlock_page(new_hpage);
 
diff --git a/mm/rmap.c b/mm/rmap.c
index 85b7f9423352..c87b7277f1b0 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -22,9 +22,10 @@
  *
  * inode->i_mutex	(while writing or truncating, not reading or faulting)
  *   mm->mmap_sem
- *     page->flags PG_locked (lock_page)
+ *     page->flags PG_locked (lock_page)   * (see huegtlbfs below)
  *       hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share)
  *         mapping->i_mmap_rwsem
+ *           hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
  *           anon_vma->rwsem
  *             mm->page_table_lock or pte_lock
  *               zone_lru_lock (in mark_page_accessed, isolate_lru_page)
@@ -43,6 +44,11 @@
  * anon_vma->rwsem,mapping->i_mutex      (memory_failure, collect_procs_anon)
  *   ->tasklist_lock
  *     pte map lock
+ *
+ * * hugetlbfs PageHuge() pages take locks in this order:
+ *         mapping->i_mmap_rwsem
+ *           hugetlb_fault_mutex (hugetlbfs specific page fault mutex)
+ *             page->flags PG_locked (lock_page)
  */
 
 #include <linux/mm.h>
@@ -1374,6 +1380,9 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 		/*
 		 * If sharing is possible, start and end will be adjusted
 		 * accordingly.
+		 *
+		 * If called for a huge page, caller must hold i_mmap_rwsem
+		 * in write mode as it is possible to call huge_pmd_unshare.
 		 */
 		adjust_range_if_pmd_sharing_possible(vma, &start, &end);
 	}
@@ -1420,6 +1429,12 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 		address = pvmw.address;
 
 		if (PageHuge(page)) {
+			/*
+			 * To call huge_pmd_unshare, i_mmap_rwsem must be
+			 * held in write mode.  Caller needs to explicitly
+			 * do this outside rmap routines.
+			 */
+			VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
 			if (huge_pmd_unshare(mm, &address, pvmw.pte)) {
 				/*
 				 * huge_pmd_unshare unmapped an entire PMD
diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
index 458acda96f20..48368589f519 100644
--- a/mm/userfaultfd.c
+++ b/mm/userfaultfd.c
@@ -267,10 +267,14 @@ static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 		VM_BUG_ON(dst_addr & ~huge_page_mask(h));
 
 		/*
-		 * Serialize via hugetlb_fault_mutex
+		 * Serialize via i_mmap_rwsem and hugetlb_fault_mutex.
+		 * i_mmap_rwsem ensures the dst_pte remains valid even
+		 * in the case of shared pmds.  fault mutex prevents
+		 * races with other faulting threads.
 		 */
-		idx = linear_page_index(dst_vma, dst_addr);
 		mapping = dst_vma->vm_file->f_mapping;
+		i_mmap_lock_read(mapping);
+		idx = linear_page_index(dst_vma, dst_addr);
 		hash = hugetlb_fault_mutex_hash(h, dst_mm, dst_vma, mapping,
 								idx, dst_addr);
 		mutex_lock(&hugetlb_fault_mutex_table[hash]);
@@ -279,6 +283,7 @@ static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 		dst_pte = huge_pte_alloc(dst_mm, dst_addr, huge_page_size(h));
 		if (!dst_pte) {
 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
+			i_mmap_unlock_read(mapping);
 			goto out_unlock;
 		}
 
@@ -286,6 +291,7 @@ static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 		dst_pteval = huge_ptep_get(dst_pte);
 		if (!huge_pte_none(dst_pteval)) {
 			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
+			i_mmap_unlock_read(mapping);
 			goto out_unlock;
 		}
 
@@ -293,6 +299,7 @@ static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 						dst_addr, src_addr, &page);
 
 		mutex_unlock(&hugetlb_fault_mutex_table[hash]);
+		i_mmap_unlock_read(mapping);
 		vm_alloc_shared = vm_shared;
 
 		cond_resched();
-- 
2.17.2

[PATCH RFC v3 2/2] hugetlbfs: Use i_mmap_rwsem to fix page fault/truncate race

[Mike Kravetz]

hugetlbfs page faults can race with truncate and hole punch operations.
Current code in the page fault path attempts to handle this by 'backing
out' operations if we encounter the race.  One obvious omission in the
current code is removing a page newly added to the page cache.  This is
pretty straight forward to address, but there is a more subtle and
difficult issue of backing out hugetlb reservations.  To handle this
correctly, the 'reservation state' before page allocation needs to be
noted so that it can be properly backed out.  There are four distinct
possibilities for reservation state: shared/reserved, shared/no-resv,
private/reserved and private/no-resv.  Backing out a reservation may
require memory allocation which could fail so that needs to be taken
into account as well.

Instead of writing the required complicated code for this rare
occurrence, just eliminate the race.  i_mmap_rwsem is now held in read
mode for the duration of page fault processing.  Hold i_mmap_rwsem in
write mode when modifying i_size.  In this way, truncation can not
proceed when page faults are being processed.  In addition, i_size
will not change during fault processing so a single check can be made
to ensure faults are not beyond (proposed) end of file.  Faults can
still race with hole punch, but that race is handled by existing code
and the use of hugetlb_fault_mutex.

With this modification, checks for races with truncation in the page
fault path can be simplified and removed.  remove_inode_hugepages no
longer needs to take hugetlb_fault_mutex in the case of truncation.
Comments are expanded to explain reasoning behind locking.

Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
---
 fs/hugetlbfs/inode.c | 32 ++++++++++++++++++++++----------
 mm/hugetlb.c         | 23 +++++++++++------------
 2 files changed, 33 insertions(+), 22 deletions(-)

diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c
index 5280fe3aad2f..039566175b48 100644
--- a/fs/hugetlbfs/inode.c
+++ b/fs/hugetlbfs/inode.c
@@ -384,10 +384,9 @@ hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
  *	In this case, we first scan the range and release found pages.
  *	After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
  *	maps and global counts.  Page faults can not race with truncation
- *	in this routine.  hugetlb_no_page() prevents page faults in the
- *	truncated range.  It checks i_size before allocation, and again after
- *	with the page table lock for the page held.  The same lock must be
- *	acquired to unmap a page.
+ *	in this routine.  hugetlb_no_page() holds i_mmap_rwsem and prevents
+ *	page faults in the truncated range by checking i_size.  i_size is
+ *	modified while holding i_mmap_rwsem.
  * hole punch is indicated if end is not LLONG_MAX
  *	In the hole punch case we scan the range and release found pages.
  *	Only when releasing a page is the associated region/reserv map
@@ -425,11 +424,19 @@ static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 			struct page *page = pvec.pages[i];
 			u32 hash;
 
-			index = page->index;
-			hash = hugetlb_fault_mutex_hash(h, current->mm,
+			if (!truncate_op) {
+				/*
+				 * Only need to hold the fault mutex in the
+				 * hole punch case.  This prevents races with
+				 * page faults.  Races are not possible in the
+				 * case of truncation.
+				 */
+				index = page->index;
+				hash = hugetlb_fault_mutex_hash(h, current->mm,
 							&pseudo_vma,
 							mapping, index, 0);
-			mutex_lock(&hugetlb_fault_mutex_table[hash]);
+				mutex_lock(&hugetlb_fault_mutex_table[hash]);
+			}
 
 			/*
 			 * If page is mapped, it was faulted in after being
@@ -472,7 +479,8 @@ static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 			}
 
 			unlock_page(page);
-			mutex_unlock(&hugetlb_fault_mutex_table[hash]);
+			if (!truncate_op)
+				mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 		}
 		huge_pagevec_release(&pvec);
 		cond_resched();
@@ -503,8 +511,8 @@ static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 	BUG_ON(offset & ~huge_page_mask(h));
 	pgoff = offset >> PAGE_SHIFT;
 
-	i_size_write(inode, offset);
 	i_mmap_lock_write(mapping);
+	i_size_write(inode, offset);
 	if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 		hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 	i_mmap_unlock_write(mapping);
@@ -626,7 +634,11 @@ static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 		/* addr is the offset within the file (zero based) */
 		addr = index * hpage_size;
 
-		/* mutex taken here, fault path and hole punch */
+		/*
+		 * fault mutex taken here, protects against fault path
+		 * and hole punch.  inode_lock previously taken protects
+		 * against truncation.
+		 */
 		hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
 						index, addr);
 		mutex_lock(&hugetlb_fault_mutex_table[hash]);
diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 9198b00d25e6..acd3411fd9f0 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -3858,16 +3858,17 @@ static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
 	}
 
 	/*
-	 * Use page lock to guard against racing truncation
-	 * before we get page_table_lock.
+	 * We can not race with truncation due to holding i_mmap_rwsem.
+	 * i_size is modified when holding i_mmap_rwsem, so check here
+	 * once for faults beyond end of file.
 	 */
+	size = i_size_read(mapping->host) >> huge_page_shift(h);
+	if (idx >= size)
+		goto out;
+
 retry:
 	page = find_lock_page(mapping, idx);
 	if (!page) {
-		size = i_size_read(mapping->host) >> huge_page_shift(h);
-		if (idx >= size)
-			goto out;
-
 		/*
 		 * Check for page in userfault range
 		 */
@@ -3957,10 +3958,6 @@ static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
 	}
 
 	ptl = huge_pte_lock(h, mm, ptep);
-	size = i_size_read(mapping->host) >> huge_page_shift(h);
-	if (idx >= size)
-		goto backout;
-
 	ret = 0;
 	if (!huge_pte_none(huge_ptep_get(ptep)))
 		goto backout;
@@ -4062,8 +4059,10 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 
 	/*
 	 * 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.
+	 * until finished with ptep.  This serves two purposes:
+	 * 1) It prevents huge_pmd_unshare from being called elsewhere
+	 *    and making the ptep no longer valid.
+	 * 2) It synchronizes us with i_size modifications during truncation.
 	 *
 	 * ptep could have already be assigned via huge_pte_offset.  That
 	 * is OK, as huge_pte_alloc will return the same value unless
-- 
2.17.2