use-after-free" with v5.10.y caused by backport of a298232ee6b9 ("io_uring: fix link timeout refs")

Hi Pavel,

We had been running syzkaller on v5.10.y and a "use after free" is being reported for v5.10.43+ kernels.

The syzkaller report is at: https://elisa-builder-00.iol.unh.edu/syzkaller/file?name=crashes%2fb23bc4ad436bbe4afc620d9503730ddd78c382c0%2freport19

The trace is:

refcount_t: underflow; use-after-free.

WARNING: CPU: 1 PID: 8769 at lib/refcount.c:28 refcount_warn_saturate+0x103/0x1f0 lib/refcount.c:28
Modules linked in:
CPU: 1 PID: 8769 Comm: syz-executor.6 Not tainted 5.10.52 #1
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014
RIP: 0010:refcount_warn_saturate+0x103/0x1f0 lib/refcount.c:28
Code: 1d d2 63 54 03 31 ff 89 de e8 c9 22 51 ff 84 db 75 a3 e8 90 29 51 ff 48 c7 c7 20 38 3b 84 c6 05 b2 63 54 03 01 e8 cc 0c c9 01 <0f> 0b eb 87 e8 74 29 51 ff 0f b6 1d 9b 63 54 03 31 ff 89 de e8 94
RSP: 0018:ffff88804ec5f9f8 EFLAGS: 00010286
RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000040000 RSI: ffffffff81293053 RDI: ffffed1009d8bf31
RBP: ffff888048ceb41c R08: 0000000000000001 R09: ffff88806cf1ff9b
R10: 0000000000000000 R11: 0000000000000001 R12: ffff888048ceb41c
R13: 0000000000000000 R14: ffff888048ceb558 R15: ffff88800c857180
FS:  00007f0e798e9700(0000) GS:ffff88806cf00000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000558bbbc7d898 CR3: 0000000048fb6000 CR4: 0000000000350ee0
Call Trace:
 __refcount_sub_and_test include/linux/refcount.h:283 [inline]
 __refcount_dec_and_test include/linux/refcount.h:315 [inline]
 refcount_dec_and_test include/linux/refcount.h:333 [inline]
 io_put_req+0xc6/0x100 fs/io_uring.c:2220
 __io_queue_sqe+0x2b1/0xd00 fs/io_uring.c:6358
 io_queue_sqe+0x5bc/0x1020 fs/io_uring.c:6403
 io_queue_link_head fs/io_uring.c:6414 [inline]
 io_submit_sqe fs/io_uring.c:6455 [inline]
 io_submit_sqes+0x17b5/0x2310 fs/io_uring.c:6700
 __do_sys_io_uring_enter+0x1092/0x1910 fs/io_uring.c:9092
 do_syscall_64+0x33/0x40 arch/x86/entry/common.c:46
 entry_SYSCALL_64_after_hwframe+0x44/0xa9
RIP: 0033:0x466609
Code: ff ff c3 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 40 00 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 bc ff ff ff f7 d8 64 89 01 48
RSP: 002b:00007f0e798e9188 EFLAGS: 00000246 ORIG_RAX: 00000000000001aa
RAX: ffffffffffffffda RBX: 000000000056bf80 RCX: 0000000000466609
RDX: 0000000000000000 RSI: 00000000000058ab RDI: 0000000000000004
RBP: 00000000004bfcb9 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000246 R12: 000000000056bf80
R13: 00007ffdbe4f1c5f R14: 00007f0e798e9300 R15: 0000000000022000
irq event stamp: 1473
hardirqs last  enabled at (1483): [<ffffffff8128f5d9>] console_unlock+0x929/0xb30 kernel/printk/printk.c:2552
hardirqs last disabled at (1494): [<ffffffff83c43a7b>] sysvec_apic_timer_interrupt+0xb/0xa0 arch/x86/kernel/apic/apic.c:1095
softirqs last  enabled at (1072): [<ffffffff83e00f92>] asm_call_irq_on_stack+0x12/0x20
softirqs last disabled at (1067): [<ffffffff83e00f92>] asm_call_irq_on_stack+0x12/0x20

I have done a bisect and the commit causing this is: 0b2a990e5d2f ("io_uring: fix link timeout refs")

The git bisect log is:
# bad: [71046eac2db9aeccf10763d034a1a123911c9a81] Linux 5.10.53
# good: [2c85ebc57b3e1817b6ce1a6b703928e113a90442] Linux 5.10
git bisect start 'v5.10.53' 'v5.10'
# good: [d29c38dd926d5aba65d177c0b99381ea632ff0a0] staging: rtl8192e: Change state information from u16 to u8 git bisect good d29c38dd926d5aba65d177c0b99381ea632ff0a0
# good: [b67c3d74adc3f7f832f57c170234bbe1fc69c87c] Revert "net: fujitsu: fix a potential NULL pointer dereference" git bisect good b67c3d74adc3f7f832f57c170234bbe1fc69c87c
# bad: [f1f30b3373df2e5ab96dd3781df5c02e5366f845] mmc: usdhi6rol0: fix error return code in usdhi6_probe() git bisect bad f1f30b3373df2e5ab96dd3781df5c02e5366f845
# bad: [3d60457d74d9cc7b36f78f9cb74f29bc6182c1e8] cxgb4: fix endianness when flashing boot image git bisect bad 3d60457d74d9cc7b36f78f9cb74f29bc6182c1e8
# bad: [3a6b69221f96f87c680bbc9fba01db3415b18f27] drm/amdgpu: make sure we unpin the UVD BO
git bisect bad 3a6b69221f96f87c680bbc9fba01db3415b18f27
# good: [65859eca4dff1af0db5e36d1cfbac15b834c6a65] Linux 5.10.42
git bisect good 65859eca4dff1af0db5e36d1cfbac15b834c6a65
# good: [a1bf16616d8351a2e79400d6d19608befb2ce1dd] ixgbe: add correct exception tracing for XDP git bisect good a1bf16616d8351a2e79400d6d19608befb2ce1dd
# bad: [c5155c741a484e036e7997420559431a951f2106] wireguard: allowedips: allocate nodes in kmem_cache git bisect bad c5155c741a484e036e7997420559431a951f2106
# good: [3c23e23c7ad9844a645f4e2bd8ec34a0a2ee5514] riscv: vdso: fix and clean-up Makefile
git bisect good 3c23e23c7ad9844a645f4e2bd8ec34a0a2ee5514
# bad: [74caf718cc7422a957aac381c73d798c0a999a65] Bluetooth: use correct lock to prevent UAF of hdev object git bisect bad 74caf718cc7422a957aac381c73d798c0a999a65
# bad: [58f4d45d8d4d391f60b6f0db6308df1994a265b3] drm/amdgpu/vcn3: add cancel_delayed_work_sync before power gate git bisect bad 58f4d45d8d4d391f60b6f0db6308df1994a265b3
# bad: [ec72cb50c1db39816eae7296686449bba8ca0b2e] io_uring: use better types for cflags
git bisect bad ec72cb50c1db39816eae7296686449bba8ca0b2e
# bad: [0b2a990e5d2f76d020cb840c456e6ec5f0c27530] io_uring: fix link timeout refs
git bisect bad 0b2a990e5d2f76d020cb840c456e6ec5f0c27530

# first bad commit: [0b2a990e5d2f76d020cb840c456e6ec5f0c27530] io_uring: fix link timeout refs

The mainline commit for the bad LTS commit is:
a298232ee6b9 ("io_uring: fix link timeout refs") and I have tested the reproducer on mainline with 'a298232ee6b9' as HEAD and the issue is not reproduced. I think we are missing some change in v5.10.y kernel which was missed while the mainline fix was backported to LTS.

I can reproduce the crash using syzkaller and will be happy to test any patch for this.

--
Regards
Sudip

"... # 5.12+", weird, but perhaps due to dependencies. Thanks for letting know.

Greg, Sasha, should be same as reported for 5.12

https://www.spinics.net/lists/stable/msg485116.html

Can you try to apply it to 5.10 or should I resend?

Pavel Begunkov

I just tried applying those patches and they did not work. So can you provide some new backports?

thanks,

greg k-h

Hi Greg,

Here is the backport for v5.10.y. I have also tested these with the syzkaller repro and the issue is fixed.

--
Regards
Sudip

From bd136e16d2f1b5480c60f4e78f18727d568c7d86 Mon Sep 17 00:00:00 2001
From: Pavel Begunkov <[email protected]>
Date: Thu, 1 Apr 2021 15:43:59 +0100
Subject: [PATCH 1/2] io_uring: put link timeout req consistently

commit df9727affa058f4f18e388b30247650f8ae13cd8 upstream

Don't put linked timeout req in io_async_find_and_cancel() but do it in io_link_timeout_fn(), so we have only one point for that and won't have to do it differently as it's now (put vs put_deferred). Btw, improve a bit io_async_find_and_cancel()'s locking.

---
 fs/io_uring.c | 10 +++-------
 1 file changed, 3 insertions(+), 7 deletions(-)

index 42153106b7bc..a6c9c55ca3a3 100644
@@ -5697,12 +5697,9 @@ static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
 	int ret;
 
 	ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
-	if (ret != -ENOENT) {
-		spin_lock_irqsave(&ctx->completion_lock, flags);
-		goto done;
-	}
-
 	spin_lock_irqsave(&ctx->completion_lock, flags);
+	if (ret != -ENOENT)
+		goto done;
 	ret = io_timeout_cancel(ctx, sqe_addr);
 	if (ret != -ENOENT)
 		goto done;

@@ -5717,7 +5714,6 @@ static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
 
 	if (ret < 0)
 		req_set_fail_links(req);
-	io_put_req(req);
 }
 
 static int io_async_cancel_prep(struct io_kiocb *req,

@@ -6263,8 +6259,8 @@ static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
 		io_put_req_deferred(req, 1);
 	} else {
 		io_cqring_add_event(req, -ETIME, 0);
-		io_put_req_deferred(req, 1);
 	}
+	io_put_req_deferred(req, 1);
 	return HRTIMER_NORESTART;
 }
 

2.30.2

From b901fb09894731c4ad84a359509508eff50a4920 Mon Sep 17 00:00:00 2001
From: Pavel Begunkov <[email protected]>
Date: Fri, 7 May 2021 21:06:38 +0100
Subject: [PATCH 2/2] io_uring: fix link timeout refs

commit a298232ee6b9a1d5d732aa497ff8be0d45b5bd82 upstream

WARNING: CPU: 0 PID: 10242 at lib/refcount.c:28 refcount_warn_saturate+0x15b/0x1a0 lib/refcount.c:28
RIP: 0010:refcount_warn_saturate+0x15b/0x1a0 lib/refcount.c:28
Call Trace:
 __refcount_sub_and_test include/linux/refcount.h:283 [inline]
 __refcount_dec_and_test include/linux/refcount.h:315 [inline]
 refcount_dec_and_test include/linux/refcount.h:333 [inline]
 io_put_req fs/io_uring.c:2140 [inline]
 io_queue_linked_timeout fs/io_uring.c:6300 [inline]
 __io_queue_sqe+0xbef/0xec0 fs/io_uring.c:6354
 io_submit_sqe fs/io_uring.c:6534 [inline]
 io_submit_sqes+0x2bbd/0x7c50 fs/io_uring.c:6660
 __do_sys_io_uring_enter fs/io_uring.c:9240 [inline]
 __se_sys_io_uring_enter+0x256/0x1d60 fs/io_uring.c:9182

io_link_timeout_fn() should put only one reference of the linked timeout request, however in case of racing with the master request's completion first io_req_complete() puts one and then io_put_req_deferred() is called.

---
 fs/io_uring.c | 1 -
 1 file changed, 1 deletion(-)

index a6c9c55ca3a3..1a5951fbc763 100644
@@ -6260,7 +6260,6 @@ static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
 	} else {
 		io_cqring_add_event(req, -ETIME, 0);
 	}
-	io_put_req_deferred(req, 1);
 	return HRTIMER_NORESTART;
 }
 

2.30.2

Thanks trying out, but it should be leaking requests (rarely), because io_cqring_add_event() doesn't put a ref unlike io_req_complete_post() from the original patch.

I'll get to it today double checking refcounting.

Pavel Begunkov