Re: crypto: xts - Fix atomic sleep when walking skcipher
From: Eric Biggers
Date: Sun Apr 14 2019 - 12:35:23 EST
Hi Herbert,
On Sun, Apr 14, 2019 at 08:03:38PM +0800, Herbert Xu wrote:
> On Sun, Apr 14, 2019 at 04:26:04AM -0700, syzbot wrote:
> > Hello,
> >
> > syzbot found the following crash on:
> >
> > HEAD commit: e0a092eb Merge branch 'smc-next'
> > git tree: net-next
> > console output: https://syzkaller.appspot.com/x/log.txt?x=139878f3200000
> > kernel config: https://syzkaller.appspot.com/x/.config?x=94c460c3e4cd188a
> > dashboard link: https://syzkaller.appspot.com/bug?extid=6f72c20560060c98b566
> > compiler: gcc (GCC) 9.0.0 20181231 (experimental)
> >
> > Unfortunately, I don't have any reproducer for this crash yet.
> >
> > IMPORTANT: if you fix the bug, please add the following tag to the commit:
> > Reported-by: syzbot+6f72c20560060c98b566@xxxxxxxxxxxxxxxxxxxxxxxxx
> >
> > BUG: sleeping function called from invalid context at crypto/skcipher.c:477
> > in_atomic(): 1, irqs_disabled(): 0, pid: 12, name: kworker/0:1
> > 2 locks held by kworker/0:1/12:
> > #0: 000000000864d9ff ((wq_completion)crypto){+.+.}, at: __write_once_size
> > include/linux/compiler.h:220 [inline]
> > #0: 000000000864d9ff ((wq_completion)crypto){+.+.}, at: arch_atomic64_set
> > arch/x86/include/asm/atomic64_64.h:34 [inline]
> > #0: 000000000864d9ff ((wq_completion)crypto){+.+.}, at: atomic64_set
> > include/asm-generic/atomic-instrumented.h:855 [inline]
> > #0: 000000000864d9ff ((wq_completion)crypto){+.+.}, at: atomic_long_set
> > include/asm-generic/atomic-long.h:40 [inline]
> > #0: 000000000864d9ff ((wq_completion)crypto){+.+.}, at: set_work_data
> > kernel/workqueue.c:619 [inline]
> > #0: 000000000864d9ff ((wq_completion)crypto){+.+.}, at:
> > set_work_pool_and_clear_pending kernel/workqueue.c:646 [inline]
> > #0: 000000000864d9ff ((wq_completion)crypto){+.+.}, at:
> > process_one_work+0x87e/0x1790 kernel/workqueue.c:2240
> > #1: 000000008b3d6218 ((work_completion)(&cpu_queue->work)){+.+.}, at:
> > process_one_work+0x8b4/0x1790 kernel/workqueue.c:2244
> > Preemption disabled at:
> > [<ffffffff830d4130>] local_bh_disable include/linux/bottom_half.h:19
> > [inline]
> > [<ffffffff830d4130>] cryptd_skcipher_complete+0x90/0x170 crypto/cryptd.c:471
> > CPU: 0 PID: 12 Comm: kworker/0:1 Not tainted 5.1.0-rc4+ #135
> > Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS
> > Google 01/01/2011
> > Workqueue: crypto cryptd_queue_worker
> > Call Trace:
> > __dump_stack lib/dump_stack.c:77 [inline]
> > dump_stack+0x172/0x1f0 lib/dump_stack.c:113
> > ___might_sleep.cold+0x1bd/0x1f6 kernel/sched/core.c:6190
> > __might_sleep+0x95/0x190 kernel/sched/core.c:6143
> > skcipher_walk_virt+0x11e/0x150 crypto/skcipher.c:477
> > xor_tweak+0x146/0x350 crypto/xts.c:105
> > xor_tweak_post crypto/xts.c:133 [inline]
> > crypt_done+0x87/0xa0 crypto/xts.c:141
> > cryptd_skcipher_complete+0xbf/0x170 crypto/cryptd.c:472
> > cryptd_skcipher_decrypt+0x2f7/0x420 crypto/cryptd.c:532
> > cryptd_queue_worker+0x126/0x1f0 crypto/cryptd.c:193
> > process_one_work+0x98e/0x1790 kernel/workqueue.c:2269
> > worker_thread+0x98/0xe40 kernel/workqueue.c:2415
> > kthread+0x357/0x430 kernel/kthread.c:253
> > ret_from_fork+0x3a/0x50 arch/x86/entry/entry_64.S:352
>
> Thanks for the report. Please try this patch:
>
> ---8<---
> When we perform a walk in the completion function, we need to ensure
> that it is atomic.
>
> Reported-by: syzbot+6f72c20560060c98b566@xxxxxxxxxxxxxxxxxxxxxxxxx
> Fixes: 78105c7e769b ("crypto: xts - Drop use of auxiliary buffer")
> Cc: <stable@xxxxxxxxxxxxxxx>
> Signed-off-by: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>
>
> diff --git a/crypto/xts.c b/crypto/xts.c
> index 847f54f76789..c915a45711f5 100644
> --- a/crypto/xts.c
> +++ b/crypto/xts.c
> @@ -88,7 +88,8 @@ static int setkey(struct crypto_skcipher *parent, const u8 *key,
> * mutliple calls to the 'ecb(..)' instance, which usually would be slower than
> * just doing the gf128mul_x_ble() calls again.
> */
> -static int xor_tweak(struct skcipher_request *req, bool second_pass)
> +static int xor_tweak(struct skcipher_request *req, bool second_pass,
> + bool atomic)
> {
> struct rctx *rctx = skcipher_request_ctx(req);
> struct crypto_skcipher *tfm = crypto_skcipher_reqtfm(req);
> @@ -102,7 +103,7 @@ static int xor_tweak(struct skcipher_request *req, bool second_pass)
> /* set to our TFM to enforce correct alignment: */
> skcipher_request_set_tfm(req, tfm);
> }
> - err = skcipher_walk_virt(&w, req, false);
> + err = skcipher_walk_virt(&w, req, atomic);
>
> while (w.nbytes) {
> unsigned int avail = w.nbytes;
> @@ -125,12 +126,12 @@ static int xor_tweak(struct skcipher_request *req, bool second_pass)
>
> static int xor_tweak_pre(struct skcipher_request *req)
> {
> - return xor_tweak(req, false);
> + return xor_tweak(req, false, false);
> }
>
> -static int xor_tweak_post(struct skcipher_request *req)
> +static int xor_tweak_post(struct skcipher_request *req, bool atomic)
> {
> - return xor_tweak(req, true);
> + return xor_tweak(req, true, atomic);
> }
>
> static void crypt_done(struct crypto_async_request *areq, int err)
> @@ -138,7 +139,7 @@ static void crypt_done(struct crypto_async_request *areq, int err)
> struct skcipher_request *req = areq->data;
>
> if (!err)
> - err = xor_tweak_post(req);
> + err = xor_tweak_post(req, true);
>
> skcipher_request_complete(req, err);
> }
> @@ -166,7 +167,7 @@ static int encrypt(struct skcipher_request *req)
> init_crypt(req);
> return xor_tweak_pre(req) ?:
> crypto_skcipher_encrypt(subreq) ?:
> - xor_tweak_post(req);
> + xor_tweak_post(req, false);
> }
>
> static int decrypt(struct skcipher_request *req)
> @@ -177,7 +178,7 @@ static int decrypt(struct skcipher_request *req)
> init_crypt(req);
> return xor_tweak_pre(req) ?:
> crypto_skcipher_decrypt(subreq) ?:
> - xor_tweak_post(req);
> + xor_tweak_post(req, false);
> }
>
> static int init_tfm(struct crypto_skcipher *tfm)
> --
This isn't correct because skcipher_walk_virt() can still sleep when
atomic=true, since it may need to copy the IV. That's why I added might_sleep()
to it in commit bb648291fc04, and that's what's triggering here.
This is the correct fix:
diff --git a/crypto/xts.c b/crypto/xts.c
index aed11e63ca315..359ef15e1673e 100644
--- a/crypto/xts.c
+++ b/crypto/xts.c
@@ -101,6 +101,7 @@ static int xor_tweak(struct skcipher_request *req, bool second_pass)
req = &rctx->subreq;
/* set to our TFM to enforce correct alignment: */
skcipher_request_set_tfm(req, tfm);
+ req->base.flags &= ~CRYPTO_TFM_REQ_MAY_SLEEP;
}
err = skcipher_walk_virt(&w, req, false);
Likewise for LRW.
With the crypto self-tests and CONFIG_DEBUG_ATOMIC_SLEEP enabled, this bug can
be reproduced by trying to allocate the skcipher algorithm
"xts(cryptd(ecb(aes-generic)))", e.g. by running in Python:
import socket
fd = socket.socket(socket.AF_ALG, 5, 0)
fd.bind(("skcipher", "xts(cryptd(ecb(aes-generic)))"))
Likewise, use "lrw(cryptd(ecb(aes-generic)))" for LRW.
- Eric