[PATCH] mm: drop mmap_sem before calling balance_dirty_pages() in write fault

[Johannes Weiner]

From: Johannes Weiner <jweiner@xxxxxx>

One of our services is observing hanging ps/top/etc under heavy write
IO, and the task states show this is an mmap_sem priority inversion:

A write fault is holding the mmap_sem in read-mode and waiting for
(heavily cgroup-limited) IO in balance_dirty_pages():

[<0>] balance_dirty_pages+0x724/0x905
[<0>] balance_dirty_pages_ratelimited+0x254/0x390
[<0>] fault_dirty_shared_page.isra.96+0x4a/0x90
[<0>] do_wp_page+0x33e/0x400
[<0>] __handle_mm_fault+0x6f0/0xfa0
[<0>] handle_mm_fault+0xe4/0x200
[<0>] __do_page_fault+0x22b/0x4a0
[<0>] page_fault+0x45/0x50
[<0>] 0xffffffffffffffff

Somebody tries to change the address space, contending for the
mmap_sem in write-mode:

[<0>] call_rwsem_down_write_failed_killable+0x13/0x20
[<0>] do_mprotect_pkey+0xa8/0x330
[<0>] SyS_mprotect+0xf/0x20
[<0>] do_syscall_64+0x5b/0x100
[<0>] entry_SYSCALL_64_after_hwframe+0x3d/0xa2
[<0>] 0xffffffffffffffff

The waiting writer locks out all subsequent readers to avoid lock
starvation, and several threads can be seen hanging like this:

[<0>] call_rwsem_down_read_failed+0x14/0x30
[<0>] proc_pid_cmdline_read+0xa0/0x480
[<0>] __vfs_read+0x23/0x140
[<0>] vfs_read+0x87/0x130
[<0>] SyS_read+0x42/0x90
[<0>] do_syscall_64+0x5b/0x100
[<0>] entry_SYSCALL_64_after_hwframe+0x3d/0xa2
[<0>] 0xffffffffffffffff

To fix this, do what we do for cache read faults already: drop the
mmap_sem before calling into anything IO bound, in this case the
balance_dirty_pages() function, and return VM_FAULT_RETRY.

Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
---
 mm/memory.c | 53 ++++++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 40 insertions(+), 13 deletions(-)

diff --git a/mm/memory.c b/mm/memory.c
index 2e796372927f..da5eb1d67447 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -2221,12 +2221,14 @@ static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
  *
  * The function expects the page to be locked and unlocks it.
  */
-static void fault_dirty_shared_page(struct vm_area_struct *vma,
-				    struct page *page)
+static int fault_dirty_shared_page(struct vm_fault *vmf)
 {
+	struct vm_area_struct *vma = vmf->vma;
 	struct address_space *mapping;
+	struct page *page = vmf->page;
 	bool dirtied;
 	bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
+	int ret = 0;
 
 	dirtied = set_page_dirty(page);
 	VM_BUG_ON_PAGE(PageAnon(page), page);
@@ -2239,16 +2241,36 @@ static void fault_dirty_shared_page(struct vm_area_struct *vma,
 	mapping = page_rmapping(page);
 	unlock_page(page);
 
+	if (!page_mkwrite)
+		file_update_time(vma->vm_file);
+
+	/*
+	 * Throttle page dirtying rate down to writeback speed.
+	 *
+	 * mapping may be NULL here because some device drivers do not
+	 * set page.mapping but still dirty their pages
+	 *
+	 * Drop the mmap_sem before waiting on IO, if we can. The file
+	 * is pinning the mapping, as per above.
+	 */
 	if ((dirtied || page_mkwrite) && mapping) {
-		/*
-		 * Some device drivers do not set page.mapping
-		 * but still dirty their pages
-		 */
+		struct file *fpin = NULL;
+
+		if ((vmf->flags &
+		     (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT)) ==
+		    FAULT_FLAG_ALLOW_RETRY) {
+			fpin = get_file(vma->vm_file);
+			up_read(&vma->vm_mm->mmap_sem);
+			ret = VM_FAULT_RETRY;
+		}
+
 		balance_dirty_pages_ratelimited(mapping);
+
+		if (fpin)
+			fput(fpin);
 	}
 
-	if (!page_mkwrite)
-		file_update_time(vma->vm_file);
+	return ret;
 }
 
 /*
@@ -2491,6 +2513,7 @@ static vm_fault_t wp_page_shared(struct vm_fault *vmf)
 	__releases(vmf->ptl)
 {
 	struct vm_area_struct *vma = vmf->vma;
+	int ret = VM_FAULT_WRITE;
 
 	get_page(vmf->page);
 
@@ -2514,10 +2537,10 @@ static vm_fault_t wp_page_shared(struct vm_fault *vmf)
 		wp_page_reuse(vmf);
 		lock_page(vmf->page);
 	}
-	fault_dirty_shared_page(vma, vmf->page);
+	ret |= fault_dirty_shared_page(vmf);
 	put_page(vmf->page);
 
-	return VM_FAULT_WRITE;
+	return ret;
 }
 
 /*
@@ -3561,7 +3584,7 @@ static vm_fault_t do_shared_fault(struct vm_fault *vmf)
 		return ret;
 	}
 
-	fault_dirty_shared_page(vma, vmf->page);
+	ret |= fault_dirty_shared_page(vmf);
 	return ret;
 }
 
@@ -3576,7 +3599,6 @@ static vm_fault_t do_shared_fault(struct vm_fault *vmf)
 static vm_fault_t do_fault(struct vm_fault *vmf)
 {
 	struct vm_area_struct *vma = vmf->vma;
-	struct mm_struct *vm_mm = vma->vm_mm;
 	vm_fault_t ret;
 
 	/*
@@ -3617,7 +3639,12 @@ static vm_fault_t do_fault(struct vm_fault *vmf)
 
 	/* preallocated pagetable is unused: free it */
 	if (vmf->prealloc_pte) {
-		pte_free(vm_mm, vmf->prealloc_pte);
+		/*
+		 * XXX: Accessing vma->vm_mm now is not safe. The page
+		 * fault handler may have dropped the mmap_sem a long
+		 * time ago. Only s390 derefs that parameter.
+		 */
+		pte_free(vma->vm_mm, vmf->prealloc_pte);
 		vmf->prealloc_pte = NULL;
 	}
 	return ret;
-- 
2.23.0