Btrfs: fix invalid page accesses in extent_same (dedup) ioctl

In the extent_same ioctl we are getting the pages for the source and
target ranges and unlocking them immediately after, which is incorrect
because later we attempt to map them (with kmap_atomic) and access their
contents at btrfs_cmp_data(). When we do such access the pages might have
been relocated or removed from memory, which leads to an invalid memory
access. This issue is detected on a kernel with CONFIG_DEBUG_PAGEALLOC=y
which produces a trace like the following:

186736.677437] general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
[186736.680382] Modules linked in: btrfs dm_flakey dm_mod ppdev xor raid6_pq sha256_generic hmac drbg ansi_cprng acpi_cpufreq evdev sg aesni_intel aes_x86_64
parport_pc ablk_helper tpm_tis psmouse parport i2c_piix4 tpm cryptd i2c_core lrw processor button serio_raw pcspkr gf128mul glue_helper loop autofs4 ext4
crc16 mbcache jbd2 sd_mod sr_mod cdrom ata_generic virtio_scsi ata_piix libata virtio_pci virtio_ring crc32c_intel scsi_mod e1000 virtio floppy [last
unloaded: btrfs]
[186736.681319] CPU: 13 PID: 10222 Comm: duperemove Tainted: G        W       4.4.0-rc6-btrfs-next-18+ #1
[186736.681319] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS by qemu-project.org 04/01/2014
[186736.681319] task: ffff880132600400 ti: ffff880362284000 task.ti: ffff880362284000
[186736.681319] RIP: 0010:[<ffffffff81264d00>]  [<ffffffff81264d00>] memcmp+0xb/0x22
[186736.681319] RSP: 0018:ffff880362287d70  EFLAGS: 00010287
[186736.681319] RAX: 000002c002468acf RBX: 0000000012345678 RCX: 0000000000000000
[186736.681319] RDX: 0000000000001000 RSI: 0005d129c5cf9000 RDI: 0005d129c5cf9000
[186736.681319] RBP: ffff880362287d70 R08: 0000000000000000 R09: 0000000000001000
[186736.681319] R10: ffff880000000000 R11: 0000000000000476 R12: 0000000000001000
[186736.681319] R13: ffff8802f91d4c88 R14: ffff8801f2a77830 R15: ffff880352e83e40
[186736.681319] FS:  00007f27b37fe700(0000) GS:ffff88043dda0000(0000) knlGS:0000000000000000
[186736.681319] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[186736.681319] CR2: 00007f27a406a000 CR3: 0000000217421000 CR4: 00000000001406e0
[186736.681319] Stack:
[186736.681319]  ffff880362287ea0 ffffffffa048d0bd 000000000009f000 0000000000001000
[186736.681319]  0100000000000000 ffff8801f2a77850 ffff8802f91d49b0 ffff880132600400
[186736.681319]  00000000000004f8 ffff8801c1efbe41 0000000000000000 0000000000000038
[186736.681319] Call Trace:
[186736.681319]  [<ffffffffa048d0bd>] btrfs_ioctl+0x24cb/0x2731 [btrfs]
[186736.681319]  [<ffffffff8108a8b0>] ? arch_local_irq_save+0x9/0xc
[186736.681319]  [<ffffffff8118b3d4>] ? rcu_read_unlock+0x3e/0x5d
[186736.681319]  [<ffffffff811822f8>] do_vfs_ioctl+0x42b/0x4ea
[186736.681319]  [<ffffffff8118b4f3>] ? __fget_light+0x62/0x71
[186736.681319]  [<ffffffff8118240e>] SyS_ioctl+0x57/0x79
[186736.681319]  [<ffffffff814872d7>] entry_SYSCALL_64_fastpath+0x12/0x6f
[186736.681319] Code: 0a 3c 6e 74 0d 3c 79 74 04 3c 59 75 0c c6 06 01 eb 03 c6 06 00 31 c0 eb 05 b8 ea ff ff ff 5d c3 55 31 c9 48 89 e5 48 39 d1 74 13 <0f> b6
04 0f 44 0f b6 04 0e 48 ff c1 44 29 c0 74 ea eb 02 31 c0

(gdb) list *(btrfs_ioctl+0x24cb)
0x5e0e1 is in btrfs_ioctl (fs/btrfs/ioctl.c:2972).
2967                    dst_addr = kmap_atomic(dst_page);
2968
2969                    flush_dcache_page(src_page);
2970                    flush_dcache_page(dst_page);
2971
2972                    if (memcmp(addr, dst_addr, cmp_len))
2973                            ret = BTRFS_SAME_DATA_DIFFERS;
2974
2975                    kunmap_atomic(addr);
2976                    kunmap_atomic(dst_addr);

So fix this by making sure we keep the pages locked and respect the same
locking order as everywhere else: get and lock the pages first and then
lock the range in the inode's io tree (like for example at
__btrfs_buffered_write() and extent_readpages()). If an ordered extent
is found after locking the range in the io tree, unlock the range,
unlock the pages, wait for the ordered extent to complete and repeat the
entire locking process until no overlapping ordered extents are found.

Cc: stable@vger.kernel.org   # 4.2+
Signed-off-by: Filipe Manana <fdmanana@suse.com>

1 files changed, 76 insertions, 14 deletions

diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c
index 83c9ad3f2621..1d6767c4c092 100644
--- a/fs/btrfs/ioctl.c
+++ b/fs/btrfs/ioctl.c
@@ -2811,7 +2811,6 @@ static struct page *extent_same_get_page(struct inode *inode, pgoff_t index)
 			return NULL;
 		}
 	}
-	unlock_page(page);
 
 	return page;
 }
@@ -2830,10 +2829,17 @@ static int gather_extent_pages(struct inode *inode, struct page **pages,
 	return 0;
 }
 
-static inline void lock_extent_range(struct inode *inode, u64 off, u64 len)
+static int lock_extent_range(struct inode *inode, u64 off, u64 len,
+			     bool retry_range_locking)
 {
-	/* do any pending delalloc/csum calc on src, one way or
-	   another, and lock file content */
+	/*
+	 * Do any pending delalloc/csum calculations on inode, one way or
+	 * another, and lock file content.
+	 * The locking order is:
+	 *
+	 *   1) pages
+	 *   2) range in the inode's io tree
+	 */
 	while (1) {
 		struct btrfs_ordered_extent *ordered;
 		lock_extent(&BTRFS_I(inode)->io_tree, off, off + len - 1);
@@ -2851,8 +2857,11 @@ static inline void lock_extent_range(struct inode *inode, u64 off, u64 len)
 		unlock_extent(&BTRFS_I(inode)->io_tree, off, off + len - 1);
 		if (ordered)
 			btrfs_put_ordered_extent(ordered);
+		if (!retry_range_locking)
+			return -EAGAIN;
 		btrfs_wait_ordered_range(inode, off, len);
 	}
+	return 0;
 }
 
 static void btrfs_double_inode_unlock(struct inode *inode1, struct inode *inode2)
@@ -2877,15 +2886,24 @@ static void btrfs_double_extent_unlock(struct inode *inode1, u64 loff1,
 	unlock_extent(&BTRFS_I(inode2)->io_tree, loff2, loff2 + len - 1);
 }
 
-static void btrfs_double_extent_lock(struct inode *inode1, u64 loff1,
-				     struct inode *inode2, u64 loff2, u64 len)
+static int btrfs_double_extent_lock(struct inode *inode1, u64 loff1,
+				    struct inode *inode2, u64 loff2, u64 len,
+				    bool retry_range_locking)
 {
+	int ret;
+
 	if (inode1 < inode2) {
 		swap(inode1, inode2);
 		swap(loff1, loff2);
 	}
-	lock_extent_range(inode1, loff1, len);
-	lock_extent_range(inode2, loff2, len);
+	ret = lock_extent_range(inode1, loff1, len, retry_range_locking);
+	if (ret)
+		return ret;
+	ret = lock_extent_range(inode2, loff2, len, retry_range_locking);
+	if (ret)
+		unlock_extent(&BTRFS_I(inode1)->io_tree, loff1,
+			      loff1 + len - 1);
+	return ret;
 }
 
 struct cmp_pages {
@@ -2901,11 +2919,15 @@ static void btrfs_cmp_data_free(struct cmp_pages *cmp)
 
 	for (i = 0; i < cmp->num_pages; i++) {
 		pg = cmp->src_pages[i];
-		if (pg)
+		if (pg) {
+			unlock_page(pg);
 			page_cache_release(pg);
+		}
 		pg = cmp->dst_pages[i];
-		if (pg)
+		if (pg) {
+			unlock_page(pg);
 			page_cache_release(pg);
+		}
 	}
 	kfree(cmp->src_pages);
 	kfree(cmp->dst_pages);
@@ -2966,6 +2988,8 @@ static int btrfs_cmp_data(struct inode *src, u64 loff, struct inode *dst,
 
 		src_page = cmp->src_pages[i];
 		dst_page = cmp->dst_pages[i];
+		ASSERT(PageLocked(src_page));
+		ASSERT(PageLocked(dst_page));
 
 		addr = kmap_atomic(src_page);
 		dst_addr = kmap_atomic(dst_page);
@@ -3078,14 +3102,46 @@ static int btrfs_extent_same(struct inode *src, u64 loff, u64 olen,
 		goto out_unlock;
 	}
 
+again:
 	ret = btrfs_cmp_data_prepare(src, loff, dst, dst_loff, olen, &cmp);
 	if (ret)
 		goto out_unlock;
 
 	if (same_inode)
-		lock_extent_range(src, same_lock_start, same_lock_len);
+		ret = lock_extent_range(src, same_lock_start, same_lock_len,
+					false);
 	else
-		btrfs_double_extent_lock(src, loff, dst, dst_loff, len);
+		ret = btrfs_double_extent_lock(src, loff, dst, dst_loff, len,
+					       false);
+	/*
+	 * If one of the inodes has dirty pages in the respective range or
+	 * ordered extents, we need to flush dellaloc and wait for all ordered
+	 * extents in the range. We must unlock the pages and the ranges in the
+	 * io trees to avoid deadlocks when flushing delalloc (requires locking
+	 * pages) and when waiting for ordered extents to complete (they require
+	 * range locking).
+	 */
+	if (ret == -EAGAIN) {
+		/*
+		 * Ranges in the io trees already unlocked. Now unlock all
+		 * pages before waiting for all IO to complete.
+		 */
+		btrfs_cmp_data_free(&cmp);
+		if (same_inode) {
+			btrfs_wait_ordered_range(src, same_lock_start,
+						 same_lock_len);
+		} else {
+			btrfs_wait_ordered_range(src, loff, len);
+			btrfs_wait_ordered_range(dst, dst_loff, len);
+		}
+		goto again;
+	}
+	ASSERT(ret == 0);
+	if (WARN_ON(ret)) {
+		/* ranges in the io trees already unlocked */
+		btrfs_cmp_data_free(&cmp);
+		return ret;
+	}
 
 	/* pass original length for comparison so we stay within i_size */
 	ret = btrfs_cmp_data(src, loff, dst, dst_loff, olen, &cmp);
@@ -3907,9 +3963,15 @@ static noinline long btrfs_ioctl_clone(struct file *file, unsigned long srcfd,
 		u64 lock_start = min_t(u64, off, destoff);
 		u64 lock_len = max_t(u64, off, destoff) + len - lock_start;
 
-		lock_extent_range(src, lock_start, lock_len);
+		ret = lock_extent_range(src, lock_start, lock_len, true);
 	} else {
-		btrfs_double_extent_lock(src, off, inode, destoff, len);
+		ret = btrfs_double_extent_lock(src, off, inode, destoff, len,
+					       true);
+	}
+	ASSERT(ret == 0);
+	if (WARN_ON(ret)) {
+		/* ranges in the io trees already unlocked */
+		goto out_unlock;
 	}
 
 	ret = btrfs_clone(src, inode, off, olen, len, destoff, 0);