shmem: fix sleeping from atomic context

Syzkaller fuzzer managed to trigger this:

    BUG: sleeping function called from invalid context at mm/shmem.c:852
    in_atomic(): 1, irqs_disabled(): 0, pid: 529, name: khugepaged
    3 locks held by khugepaged/529:
     #0:  (shrinker_rwsem){++++..}, at: [<ffffffff818d7ef1>] shrink_slab.part.59+0x121/0xd30 mm/vmscan.c:451
     #1:  (&type->s_umount_key#29){++++..}, at: [<ffffffff81a63630>] trylock_super+0x20/0x100 fs/super.c:392
     #2:  (&(&sbinfo->shrinklist_lock)->rlock){+.+.-.}, at: [<ffffffff818fd83e>] spin_lock include/linux/spinlock.h:302 [inline]
     #2:  (&(&sbinfo->shrinklist_lock)->rlock){+.+.-.}, at: [<ffffffff818fd83e>] shmem_unused_huge_shrink+0x28e/0x1490 mm/shmem.c:427
    CPU: 2 PID: 529 Comm: khugepaged Not tainted 4.10.0-rc5+ torvalds#201
    Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
    Call Trace:
       shmem_undo_range+0xb20/0x2710 mm/shmem.c:852
       shmem_truncate_range+0x27/0xa0 mm/shmem.c:939
       shmem_evict_inode+0x35f/0xca0 mm/shmem.c:1030
       evict+0x46e/0x980 fs/inode.c:553
       iput_final fs/inode.c:1515 [inline]
       iput+0x589/0xb20 fs/inode.c:1542
       shmem_unused_huge_shrink+0xbad/0x1490 mm/shmem.c:446
       shmem_unused_huge_scan+0x10c/0x170 mm/shmem.c:512
       super_cache_scan+0x376/0x450 fs/super.c:106
       do_shrink_slab mm/vmscan.c:378 [inline]
       shrink_slab.part.59+0x543/0xd30 mm/vmscan.c:481
       shrink_slab mm/vmscan.c:2592 [inline]
       shrink_node+0x2c7/0x870 mm/vmscan.c:2592
       shrink_zones mm/vmscan.c:2734 [inline]
       do_try_to_free_pages+0x369/0xc80 mm/vmscan.c:2776
       try_to_free_pages+0x3c6/0x900 mm/vmscan.c:2982
       __perform_reclaim mm/page_alloc.c:3301 [inline]
       __alloc_pages_direct_reclaim mm/page_alloc.c:3322 [inline]
       __alloc_pages_slowpath+0xa24/0x1c30 mm/page_alloc.c:3683
       __alloc_pages_nodemask+0x544/0xae0 mm/page_alloc.c:3848
       __alloc_pages include/linux/gfp.h:426 [inline]
       __alloc_pages_node include/linux/gfp.h:439 [inline]
       khugepaged_alloc_page+0xc2/0x1b0 mm/khugepaged.c:750
       collapse_huge_page+0x182/0x1fe0 mm/khugepaged.c:955
       khugepaged_scan_pmd+0xfdf/0x12a0 mm/khugepaged.c:1208
       khugepaged_scan_mm_slot mm/khugepaged.c:1727 [inline]
       khugepaged_do_scan mm/khugepaged.c:1808 [inline]
       khugepaged+0xe9b/0x1590 mm/khugepaged.c:1853
       kthread+0x326/0x3f0 kernel/kthread.c:227
       ret_from_fork+0x31/0x40 arch/x86/entry/entry_64.S:430

The iput() from atomic context was a bad idea: if after igrab() somebody
else calls iput() and we left with the last inode reference, our iput()
would lead to inode eviction and therefore sleeping.

This patch should fix the situation.

Link: http://lkml.kernel.org/r/20170131093141.GA15899@node.shutemov.name
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/shmem.c

@@ -415,6 +415,7 @@ static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 		struct shrink_control *sc, unsigned long nr_to_split)
 {
 	LIST_HEAD(list), *pos, *next;
+	LIST_HEAD(to_remove);
 	struct inode *inode;
 	struct shmem_inode_info *info;
 	struct page *page;
@@ -441,9 +442,8 @@ static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 		/* Check if there's anything to gain */
 		if (round_up(inode->i_size, PAGE_SIZE) ==
 				round_up(inode->i_size, HPAGE_PMD_SIZE)) {
-			list_del_init(&info->shrinklist);
+			list_move(&info->shrinklist, &to_remove);
 			removed++;
-			iput(inode);
 			goto next;
 		}
 
@@ -454,6 +454,13 @@ static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
 	}
 	spin_unlock(&sbinfo->shrinklist_lock);
 
+	list_for_each_safe(pos, next, &to_remove) {
+		info = list_entry(pos, struct shmem_inode_info, shrinklist);
+		inode = &info->vfs_inode;
+		list_del_init(&info->shrinklist);
+		iput(inode);
+	}
+
 	list_for_each_safe(pos, next, &list) {
 		int ret;