[PATCH] ocfs2: validate bg_list.l_next_free_rec in discontig group descriptor

[ZhengYuan Huang <gality369@gmail.com>]

[BUG]
Running ocfs2 on a corrupted image with a discontiguous block
group whose bg_list.l_next_free_rec is set to an excessively
large value triggers a KASAN use-after-free crash:

BUG: KASAN: use-after-free in ocfs2_bg_discontig_fix_by_rec fs/ocfs2/suballoc.c:1678 [inline]
BUG: KASAN: use-after-free in ocfs2_bg_discontig_fix_result+0x4a4/0x560 fs/ocfs2/suballoc.c:1715
Read of size 4 at addr ffff88801a85f000 by task syz.0.115/552

Call Trace:
 <TASK>
 ...
 __asan_report_load4_noabort+0x14/0x30 mm/kasan/report_generic.c:380
 ocfs2_bg_discontig_fix_by_rec fs/ocfs2/suballoc.c:1678 [inline]
 ocfs2_bg_discontig_fix_result+0x4a4/0x560 fs/ocfs2/suballoc.c:1715
 ocfs2_search_one_group fs/ocfs2/suballoc.c:1752 [inline]
 ocfs2_claim_suballoc_bits+0x13c3/0x1cd0 fs/ocfs2/suballoc.c:1984
 ocfs2_claim_new_inode+0x2e7/0x8a0 fs/ocfs2/suballoc.c:2292
 ocfs2_mknod_locked.constprop.0+0x121/0x2a0 fs/ocfs2/namei.c:637
 ocfs2_mknod+0xc71/0x2400 fs/ocfs2/namei.c:384
 ocfs2_create+0x158/0x390 fs/ocfs2/namei.c:676
 lookup_open.isra.0+0x10a1/0x1460 fs/namei.c:3796
 open_last_lookups fs/namei.c:3895 [inline]
 path_openat+0x11fe/0x2ce0 fs/namei.c:4131
 do_filp_open+0x1f6/0x430 fs/namei.c:4161
 do_sys_openat2+0x117/0x1c0 fs/open.c:1437
 do_sys_open fs/open.c:1452 [inline]
 __do_sys_openat fs/open.c:1468 [inline]
 __se_sys_openat fs/open.c:1463 [inline]
 __x64_sys_openat+0x15b/0x220 fs/open.c:1463
 ...

[CAUSE]
ocfs2_bg_discontig_fix_result() iterates over bg->bg_list.l_recs[]
using l_next_free_rec as the upper bound without any sanity check:

  for (i = 0; i < le16_to_cpu(bg->bg_list.l_next_free_rec); i++) {
          rec = &bg->bg_list.l_recs[i];

l_next_free_rec is read directly from the on-disk group descriptor and
is trusted blindly. On a 4 KiB block device, bg_list.l_recs[] can hold
at most 235 entries (ocfs2_extent_recs_per_gd(sb)). A corrupted or
crafted filesystem image can set l_next_free_rec to an arbitrarily
large value, causing the loop to index past the end of the group
descriptor buffer_head data page and into an adjacent freed page.

[FIX]
Fix this by adding a bounds check in ocfs2_validate_gd_self(), which
is called for every group descriptor read via ocfs2_read_group_descriptor().
Use ocfs2_gd_is_discontig() to restrict the check to discontiguous
block groups, and ocfs2_extent_recs_per_gd(sb) as the physical upper
bound (rather than trusting the on-disk l_count, which could also be
corrupted). This follows the same do_error() pattern used by the
existing field sanity checks in ocfs2_validate_gd_self().

Signed-off-by: ZhengYuan Huang <gality369@gmail.com>
---
 fs/ocfs2/suballoc.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/fs/ocfs2/suballoc.c b/fs/ocfs2/suballoc.c
index 6ac4dcd54588..6dcf45fda457 100644
--- a/fs/ocfs2/suballoc.c
+++ b/fs/ocfs2/suballoc.c
@@ -196,6 +196,22 @@ static int ocfs2_validate_gd_self(struct super_block *sb,
 			 8 * le16_to_cpu(gd->bg_size));
 	}
 
+	/*
+	 * For discontiguous block groups, validate that bg_list.l_next_free_rec
+	 * does not exceed the maximum number of extent records that can physically
+	 * fit in a single block.
+	 */
+	if (ocfs2_gd_is_discontig(gd)) {
+		u16 max_recs = ocfs2_extent_recs_per_gd(sb);
+
+		if (le16_to_cpu(gd->bg_list.l_next_free_rec) > max_recs) {
+			do_error("Group descriptor #%llu bad discontig l_next_free_rec %u max %u\n",
+				 (unsigned long long)bh->b_blocknr,
+				 le16_to_cpu(gd->bg_list.l_next_free_rec),
+				 max_recs);
+		}
+	}
+
 	return 0;
 }
 
-- 
2.43.0