blob: 052f40cd937a0cecb17906c331b02e61e21ba337 [file] [log] [blame]
developer7f77d022022-06-16 14:15:52 +08001From git@z Thu Jan 1 00:00:00 1970
2Subject: [PATCH v2] napi: fix race inside napi_enable
3From: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
4Date: Sat, 18 Sep 2021 16:52:32 +0800
5Message-Id: <20210918085232.71436-1-xuanzhuo@linux.alibaba.com>
6To: netdev@vger.kernel.org, linyunsheng@huawei.com
7Cc: "David S. Miller" <davem@davemloft.net>, Jakub Kicinski <kuba@kernel.org>, Eric Dumazet <edumazet@google.com>, Daniel Borkmann <daniel@iogearbox.net>, Antoine Tenart <atenart@kernel.org>, Alexander Lobakin <alobakin@pm.me>, Wei Wang <weiwan@google.com>, Taehee Yoo <ap420073@gmail.com>,Björn Töpel <bjorn@kernel.org>, Arnd Bergmann <arnd@arndb.de>, Kumar Kartikeya Dwivedi <memxor@gmail.com>, Neil Horman <nhorman@redhat.com>, Dust Li <dust.li@linux.alibaba.com>
8List-Id: <netdev.vger.kernel.org>
9MIME-Version: 1.0
10Content-Type: text/plain; charset="utf-8"
11Content-Transfer-Encoding: 7bit
12
13The process will cause napi.state to contain NAPI_STATE_SCHED and
14not in the poll_list, which will cause napi_disable() to get stuck.
15
16The prefix "NAPI_STATE_" is removed in the figure below, and
17NAPI_STATE_HASHED is ignored in napi.state.
18
19 CPU0 | CPU1 | napi.state
20===============================================================================
21napi_disable() | | SCHED | NPSVC
22napi_enable() | |
23{ | |
24 smp_mb__before_atomic(); | |
25 clear_bit(SCHED, &n->state); | | NPSVC
26 | napi_schedule_prep() | SCHED | NPSVC
27 | napi_poll() |
28 | napi_complete_done() |
29 | { |
30 | if (n->state & (NPSVC | | (1)
31 | _BUSY_POLL))) |
32 | return false; |
33 | ................ |
34 | } | SCHED | NPSVC
35 | |
36 clear_bit(NPSVC, &n->state); | | SCHED
37} | |
38 | |
39napi_schedule_prep() | | SCHED | MISSED (2)
40
41(1) Here return direct. Because of NAPI_STATE_NPSVC exists.
42(2) NAPI_STATE_SCHED exists. So not add napi.poll_list to sd->poll_list
43
44Since NAPI_STATE_SCHED already exists and napi is not in the
45sd->poll_list queue, NAPI_STATE_SCHED cannot be cleared and will always
46exist.
47
481. This will cause this queue to no longer receive packets.
492. If you encounter napi_disable under the protection of rtnl_lock, it
50 will cause the entire rtnl_lock to be locked, affecting the overall
51 system.
52
53This patch uses cmpxchg to implement napi_enable(), which ensures that
54there will be no race due to the separation of clear two bits.
55
56Fixes: 2d8bff12699abc ("netpoll: Close race condition between poll_one_napi and napi_disable")
57Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
58Reviewed-by: Dust Li <dust.li@linux.alibaba.com>
59---
60 net/core/dev.c | 16 ++++++++++------
61 1 file changed, 10 insertions(+), 6 deletions(-)
62
63diff --git a/net/core/dev.c b/net/core/dev.c
64index 74fd402d26dd..7ee9fecd3aff 100644
65--- a/net/core/dev.c
66+++ b/net/core/dev.c
67@@ -6923,12 +6923,16 @@ EXPORT_SYMBOL(napi_disable);
68 */
69 void napi_enable(struct napi_struct *n)
70 {
71- BUG_ON(!test_bit(NAPI_STATE_SCHED, &n->state));
72- smp_mb__before_atomic();
73- clear_bit(NAPI_STATE_SCHED, &n->state);
74- clear_bit(NAPI_STATE_NPSVC, &n->state);
75- if (n->dev->threaded && n->thread)
76- set_bit(NAPI_STATE_THREADED, &n->state);
77+ unsigned long val, new;
78+
79+ do {
80+ val = READ_ONCE(n->state);
81+ BUG_ON(!test_bit(NAPI_STATE_SCHED, &val));
82+
83+ new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC);
84+ if (n->dev->threaded && n->thread)
85+ new |= NAPIF_STATE_THREADED;
86+ } while (cmpxchg(&n->state, val, new) != val);
87 }
88 EXPORT_SYMBOL(napi_enable);
89
90
91--
922.31.0
93
94