hugetlbfs: emit hugetlbfs_free_inode from ->destroy_inode()
Subsystems (3)
PATCH 1/1 hugetlbfs: emit hugetlbfs_free_inode from ->destroy_inode()
hugetlbfs_free_inode() calls trace_hugetlbfs_free_inode(), and the hugetlbfs__inode event class records inode->i_sb->s_dev. ->free_inode() runs from an RCU callback (destroy_inode() queues i_callback() with call_rcu()), and Documentation/filesystems/porting.rst says that this callback "is *NOT* ordered wrt superblock destruction; as the matter of fact, the superblock and all associated structures might be already gone". put_super() frees the superblock with call_rcu() followed by a workqueue item (destroy_super_rcu() -> destroy_super_work()). syzbot hit this with the event enabled:
BUG: KASAN: slab-use-after-free in trace_event_raw_event_hugetlbfs__inode+0x45b/0x4c0 include/trace/events/hugetlbfs.h:36 Read of size 4 at addr ffff888038850010 by task syz-executor952/6134 hugetlbfs_free_inode+0x15e/0x1f0 fs/hugetlbfs/inode.c:1182 i_callback+0x46/0x80 fs/inode.c:317 rcu_do_batch kernel/rcu/tree.c:2650 [inline] rcu_core+0x5f0/0x1220 kernel/rcu/tree.c:2919 Allocated by task 6134: alloc_super fs/super.c:326 [inline] sget_fc+0x1ba/0x1f60 fs/super.c:874 hugetlbfs_get_tree+0x1e7/0x2b0 fs/hugetlbfs/inode.c:1431 Freed by task 855: kfree+0x22b/0x6d0 mm/slub.c:6923 process_one_work+0xac7/0x1b10 kernel/workqueue.c:3495
Emit the event from hugetlbfs_destroy_inode() instead. destroy_inode() calls ->destroy_inode() synchronously, before it queues the RCU callback, while the superblock is still alive; hugetlbfs_destroy_inode() already dereferences inode->i_sb there. ->free_inode() is left with only the kmem_cache_free().
---
Tested on net/main 94cea342e891 (arm64, QEMU, generic KASAN, tracing enabled) with the syzbot C reproducer, which enables events/hugetlbfs/hugetlbfs_free_inode and races hugetlbfs mounts and lazy unmounts with open + unlink. Without this patch KASAN reports the slab-use-after-free in trace_event_raw_event_hugetlbfs__inode() from hugetlbfs_free_inode() <- i_callback() during the first 10 second run, on 3 of 3 boots. With it, 12 runs over 2 boots are clean, and the trace buffer still holds the hugetlbfs_free_inode events (about 46k per boot) with valid dev numbers. fs/hugetlbfs/inode.o builds cleanly with W=1.
The event now fires when the inode is destroyed instead of when its memory goes back to the slab cache, one RCU grace period earlier; the recorded fields are the same.
fs/hugetlbfs/inode.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
fs/hugetlbfs/inode.c +5 -1
@@ -1179,12 +1179,16 @@ static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)static void hugetlbfs_free_inode (struct inode *inode) { -trace_hugetlbfs_free_inode (inode);kmem_cache_free (hugetlbfs_inode_cachep,HUGETLBFS_I (inode)); }static void hugetlbfs_destroy_inode (struct inode *inode) { +/* +* Not in ->free_inode(): that runs from an RCU callback, when the +* superblock the event reads may already be freed. +*/ +trace_hugetlbfs_free_inode (inode);hugetlbfs_inc_free_inodes (HUGETLBFS_SB (inode->i_sb)); }
Signature · 1 line
2.39.5
David Hildenbrand (Arm)
reviewerQuoted 66 lines · Joas Antonio dos Santos
hugetlbfs_free_inode()callstrace_hugetlbfs_free_inode(), and thehugetlbfs__inodeevent class recordsinode->i_sb->s_dev. ->free_inode() runs from an RCU callback (destroy_inode()queuesi_callback()withcall_rcu()), andDocumentation/filesystems/porting.rstsays that this callback "is *NOT* ordered wrt superblock destruction; as the matter of fact, the superblock and all associated structures might be already gone".put_super()frees the superblock withcall_rcu()followed by a workqueue item (destroy_super_rcu()->destroy_super_work()). syzbot hit this with the event enabled:BUG: KASAN: slab-use-after-free in trace_event_raw_event_hugetlbfs__inode+0x45b/0x4c0 include/trace/events/hugetlbfs.h:36 Read of size 4 at addr ffff888038850010 by task syz-executor952/6134 hugetlbfs_free_inode+0x15e/0x1f0 fs/hugetlbfs/inode.c:1182 i_callback+0x46/0x80 fs/inode.c:317 rcu_do_batch kernel/rcu/tree.c:2650 [inline] rcu_core+0x5f0/0x1220 kernel/rcu/tree.c:2919 Allocated by task 6134: alloc_super fs/super.c:326 [inline] sget_fc+0x1ba/0x1f60 fs/super.c:874 hugetlbfs_get_tree+0x1e7/0x2b0 fs/hugetlbfs/inode.c:1431 Freed by task 855: kfree+0x22b/0x6d0 mm/slub.c:6923 process_one_work+0xac7/0x1b10 kernel/workqueue.c:3495Emit the event from
hugetlbfs_destroy_inode()instead.destroy_inode()calls ->destroy_inode() synchronously, before it queues the RCU callback, while the superblock is still alive;hugetlbfs_destroy_inode()already dereferencesinode->i_sbthere. ->free_inode() is left with only thekmem_cache_free().Fixes 014ad7c42a69 ("hugetlbfs: use tracepoints in hugetlbfs functions.")Reported-by [email protected]Signed-off-by Joas Antonio dos Santos <[email protected]>Assisted-by Yaga Agent HackerSec Claude:claude-opus-5-5---
Tested on net/main94cea342e891(arm64, QEMU, generic KASAN, tracing enabled) with the syzbot C reproducer, which enables events/hugetlbfs/hugetlbfs_free_inode and races hugetlbfs mounts and lazy unmounts with open + unlink. Without this patch KASAN reports the slab-use-after-free intrace_event_raw_event_hugetlbfs__inode()fromhugetlbfs_free_inode()<-i_callback()during the first 10 second run, on 3 of 3 boots. With it, 12 runs over 2 boots are clean, and the trace buffer still holds thehugetlbfs_free_inodeevents (about 46k per boot) with valid dev numbers.fs/hugetlbfs/inode.obuilds cleanly with W=1.The event now fires when the inode is destroyed instead of when its memory goes back to the slab cache, one RCU grace period earlier; the recorded fields are the same.
fs/hugetlbfs/inode.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)fs/hugetlbfs/inode.c +0
-1@@ -1179,12 +1179,16 @@ static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)static void hugetlbfs_free_inode (struct inode *inode) { -trace_hugetlbfs_free_inode (inode);kmem_cache_free (hugetlbfs_inode_cachep,HUGETLBFS_I (inode)); }
fs/hugetlbfs/inode.c static struct inode *hugetlbfs_alloc_inode(struct super_block *sb) · line 1194
Quoted 10 lines
fs/hugetlbfs/inode.c +5
-0static void hugetlbfs_destroy_inode (struct inode *inode) { +/* +* Not in ->free_inode(): that runs from an RCU callback, when the +* superblock the event reads may already be freed. +*/ +trace_hugetlbfs_free_inode (inode);hugetlbfs_inc_free_inodes (HUGETLBFS_SB (inode->i_sb)); }
Should we rename the tracepoint then?
Signature · 3 lines
Cheers, David
Quoted 1 line · David Hildenbrand
Should we rename the tracepoint then?
Fair point, "free_inode" no longer matches where it fires.
I'd keep this patch as is, though, since it is tagged for stable:
the event name is visible to users in tracefs (events/hugetlbfs/hugetlbfs_free_inode), and renaming it in stable kernels would break existing tracing setups there just to fix the use-after-free.
I can send a follow-up on top of it for mm-unstable that renames the event to hugetlbfs_destroy_inode, which also matches the hugetlbfs_evict_inode name next to it. Would that work for you, or would you prefer the rename folded into a v2?
Thanks,
Joas