diff mbox series

[v8,06/23] mm/shmem: Handle uffd-wp special pte in page fault handler

Message ID 20220405014844.14239-1-peterx@redhat.com (mailing list archive)
State New
Headers show
Series userfaultfd-wp: Support shmem and hugetlbfs | expand

Commit Message

Peter Xu April 5, 2022, 1:48 a.m. UTC
File-backed memories are prone to unmap/swap so the ptes are always unstable,
because they can be easily faulted back later using the page cache.  This could
lead to uffd-wp getting lost when unmapping or swapping out such memory.  One
example is shmem.  PTE markers are needed to store those information.

This patch prepares it by handling uffd-wp pte markers first it is applied
elsewhere, so that the page fault handler can recognize uffd-wp pte markers.

The handling of uffd-wp pte markers is similar to missing fault, it's just that
we'll handle this "missing fault" when we see the pte markers, meanwhile we
need to make sure the marker information is kept during processing the fault.

This is a slow path of uffd-wp handling, because zapping of wr-protected shmem
ptes should be rare.  So far it should only trigger in two conditions:

  (1) When trying to punch holes in shmem_fallocate(), there is an optimization
      to zap the pgtables before evicting the page.

  (2) When swapping out shmem pages.

Because of this, the page fault handling is simplifed too by not sending the
wr-protect message in the 1st page fault, instead the page will be installed
read-only, so the uffd-wp message will be generated in the next fault, which
will trigger the do_wp_page() path of general uffd-wp handling.

Disable fault-around for all uffd-wp registered ranges for extra safety just
like uffd-minor fault, and clean the code up.

Signed-off-by: Peter Xu <peterx@redhat.com>
---
 include/linux/userfaultfd_k.h | 17 +++++++++
 mm/memory.c                   | 67 ++++++++++++++++++++++++++++++-----
 2 files changed, 75 insertions(+), 9 deletions(-)

Comments

David Hildenbrand May 11, 2022, 4:30 p.m. UTC | #1
> +/*
> + * This is actually a page-missing access, but with uffd-wp special pte
> + * installed.  It means this pte was wr-protected before being unmapped.
> + */
> +static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
> +{
> +	/*
> +	 * Just in case there're leftover special ptes even after the region
> +	 * got unregistered - we can simply clear them.  We can also do that
> +	 * proactively when e.g. when we do UFFDIO_UNREGISTER upon some uffd-wp
> +	 * ranges, but it should be more efficient to be done lazily here.
> +	 */
> +	if (unlikely(!userfaultfd_wp(vmf->vma) || vma_is_anonymous(vmf->vma)))
> +		return pte_marker_clear(vmf);

What would happen if we do a unregister followed by a register? IMHO we
should start with a clean uffd-wp slate then. Your comment makes ma
assume that we could receive stale WP events, which would be wrong?
Peter Xu May 12, 2022, 4:34 p.m. UTC | #2
On Wed, May 11, 2022 at 06:30:59PM +0200, David Hildenbrand wrote:
> > +/*
> > + * This is actually a page-missing access, but with uffd-wp special pte
> > + * installed.  It means this pte was wr-protected before being unmapped.
> > + */
> > +static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
> > +{
> > +	/*
> > +	 * Just in case there're leftover special ptes even after the region
> > +	 * got unregistered - we can simply clear them.  We can also do that
> > +	 * proactively when e.g. when we do UFFDIO_UNREGISTER upon some uffd-wp
> > +	 * ranges, but it should be more efficient to be done lazily here.
> > +	 */
> > +	if (unlikely(!userfaultfd_wp(vmf->vma) || vma_is_anonymous(vmf->vma)))
> > +		return pte_marker_clear(vmf);
> 
> What would happen if we do a unregister followed by a register? IMHO we
> should start with a clean uffd-wp slate then. Your comment makes ma
> assume that we could receive stale WP events, which would be wrong?

I'd say it's not wrong, but it's true and actually expected.

Firstly, userfaultfd (by design) always allows false positives (getting
same message multiple times) but no tolerance on false negatives (missing
event, which is data corrupt).

The latter should be obvious.  For the former, the simplest example is when
two threads access the same missing page the same time, two same messages
will be generated.  Same applies to wr-protect faults.  And it'll be
non-trivial (or say, impossible.. IMHO) to avoid those.

In this specific case, it's about when to drop the uffd-wp bits when
unregister.  Two obvious options: (1) during unregister, or (2) lazy.

Here I chose the lazy way because unregister could be slowed down by this,
and that's when program quits.  In short with current approach we quit
fast.  We could have leftovers, but we'll take care of them when needed.

One important thing is leftover ptes should not be the major way uffd-wp
should be used by the normal register -> wr-protect -> unprotect ->
unregister sequence.  Normally the process won't unregister probably until
it quits, so the leftover does no harm to anyone.

Meanwhile, any user who wants to avoid the lazy way can simply do a
whole-round unprotect before unregister.  So we leave more choice for the
user and by default we make sure no syscall will be easily slowed down.

Hope that answers, thanks!
diff mbox series

Patch

diff --git a/include/linux/userfaultfd_k.h b/include/linux/userfaultfd_k.h
index bd09c3c89b59..827e38b7be65 100644
--- a/include/linux/userfaultfd_k.h
+++ b/include/linux/userfaultfd_k.h
@@ -96,6 +96,18 @@  static inline bool uffd_disable_huge_pmd_share(struct vm_area_struct *vma)
 	return vma->vm_flags & (VM_UFFD_WP | VM_UFFD_MINOR);
 }
 
+/*
+ * Don't do fault around for either WP or MINOR registered uffd range.  For
+ * MINOR registered range, fault around will be a total disaster and ptes can
+ * be installed without notifications; for WP it should mostly be fine as long
+ * as the fault around checks for pte_none() before the installation, however
+ * to be super safe we just forbid it.
+ */
+static inline bool uffd_disable_fault_around(struct vm_area_struct *vma)
+{
+	return vma->vm_flags & (VM_UFFD_WP | VM_UFFD_MINOR);
+}
+
 static inline bool userfaultfd_missing(struct vm_area_struct *vma)
 {
 	return vma->vm_flags & VM_UFFD_MISSING;
@@ -236,6 +248,11 @@  static inline void userfaultfd_unmap_complete(struct mm_struct *mm,
 {
 }
 
+static inline bool uffd_disable_fault_around(struct vm_area_struct *vma)
+{
+	return false;
+}
+
 #endif /* CONFIG_USERFAULTFD */
 
 static inline bool pte_marker_entry_uffd_wp(swp_entry_t entry)
diff --git a/mm/memory.c b/mm/memory.c
index b1af996b09ca..21abb8a30553 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3559,6 +3559,39 @@  static inline bool should_try_to_free_swap(struct page *page,
 		page_count(page) == 2;
 }
 
+static vm_fault_t pte_marker_clear(struct vm_fault *vmf)
+{
+	vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd,
+				       vmf->address, &vmf->ptl);
+	/*
+	 * Be careful so that we will only recover a special uffd-wp pte into a
+	 * none pte.  Otherwise it means the pte could have changed, so retry.
+	 */
+	if (is_pte_marker(*vmf->pte))
+		pte_clear(vmf->vma->vm_mm, vmf->address, vmf->pte);
+	pte_unmap_unlock(vmf->pte, vmf->ptl);
+	return 0;
+}
+
+/*
+ * This is actually a page-missing access, but with uffd-wp special pte
+ * installed.  It means this pte was wr-protected before being unmapped.
+ */
+static vm_fault_t pte_marker_handle_uffd_wp(struct vm_fault *vmf)
+{
+	/*
+	 * Just in case there're leftover special ptes even after the region
+	 * got unregistered - we can simply clear them.  We can also do that
+	 * proactively when e.g. when we do UFFDIO_UNREGISTER upon some uffd-wp
+	 * ranges, but it should be more efficient to be done lazily here.
+	 */
+	if (unlikely(!userfaultfd_wp(vmf->vma) || vma_is_anonymous(vmf->vma)))
+		return pte_marker_clear(vmf);
+
+	/* do_fault() can handle pte markers too like none pte */
+	return do_fault(vmf);
+}
+
 static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
 {
 	swp_entry_t entry = pte_to_swp_entry(vmf->orig_pte);
@@ -3572,8 +3605,11 @@  static vm_fault_t handle_pte_marker(struct vm_fault *vmf)
 	if (WARN_ON_ONCE(vma_is_anonymous(vmf->vma) || !marker))
 		return VM_FAULT_SIGBUS;
 
-	/* TODO: handle pte markers */
-	return 0;
+	if (pte_marker_entry_uffd_wp(entry))
+		return pte_marker_handle_uffd_wp(vmf);
+
+	/* This is an unknown pte marker */
+	return VM_FAULT_SIGBUS;
 }
 
 /*
@@ -4157,6 +4193,7 @@  vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
 {
 	struct vm_area_struct *vma = vmf->vma;
+	bool uffd_wp = pte_marker_uffd_wp(vmf->orig_pte);
 	bool write = vmf->flags & FAULT_FLAG_WRITE;
 	bool prefault = vmf->address != addr;
 	pte_t entry;
@@ -4171,6 +4208,8 @@  void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
 
 	if (write)
 		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
+	if (unlikely(uffd_wp))
+		entry = pte_mkuffd_wp(pte_wrprotect(entry));
 	/* copy-on-write page */
 	if (write && !(vma->vm_flags & VM_SHARED)) {
 		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
@@ -4344,9 +4383,21 @@  static vm_fault_t do_fault_around(struct vm_fault *vmf)
 	return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
 }
 
+/* Return true if we should do read fault-around, false otherwise */
+static inline bool should_fault_around(struct vm_fault *vmf)
+{
+	/* No ->map_pages?  No way to fault around... */
+	if (!vmf->vma->vm_ops->map_pages)
+		return false;
+
+	if (uffd_disable_fault_around(vmf->vma))
+		return false;
+
+	return fault_around_bytes >> PAGE_SHIFT > 1;
+}
+
 static vm_fault_t do_read_fault(struct vm_fault *vmf)
 {
-	struct vm_area_struct *vma = vmf->vma;
 	vm_fault_t ret = 0;
 
 	/*
@@ -4354,12 +4405,10 @@  static vm_fault_t do_read_fault(struct vm_fault *vmf)
 	 * if page by the offset is not ready to be mapped (cold cache or
 	 * something).
 	 */
-	if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
-		if (likely(!userfaultfd_minor(vmf->vma))) {
-			ret = do_fault_around(vmf);
-			if (ret)
-				return ret;
-		}
+	if (should_fault_around(vmf)) {
+		ret = do_fault_around(vmf);
+		if (ret)
+			return ret;
 	}
 
 	ret = __do_fault(vmf);