netfilter: nfnetlink: place subsys mutexes in distinct lockdep classes
Subsystems (2)
Florian Westphal
maintainerPATCH 1/1 netfilter: nfnetlink: place subsys mutexes in distinct lockdep classes
From time to time there are lockdep reports similar to this one:
WARNING: possible circular locking dependency detected
------------------------------------------------------
000000004f61aa56 (&table[i].mutex){+.+.}, at: nfnl_lock [nfnetlink] but task is already holding lock:
[..] (&net->nft.commit_mutex){+.+.}, at: nf_tables_valid_genid [nf_tables] which lock already depends on the new lock.
the existing dependency chain (in reverse order) is:
-> #1 (&net->nft.commit_mutex){+.+.}:
[..]
nf_tables_valid_genid+0x18/0x60 [nf_tables]
nfnetlink_rcv_batch+0x24c/0x620 [nfnetlink]
nfnetlink_rcv+0x110/0x140 [nfnetlink]
netlink_unicast+0x12c/0x1e0
[..]
sys_sendmsg+0x18/0x40
linux_sparc_syscall+0x34/0x44
-> #0 (&table[i].mutex){+.+.}:
[..]
nfnl_lock+0x24/0x40 [nfnetlink]
ip_set_nfnl_get_byindex+0x19c/0x280 [ip_set]
set_match_v1_checkentry+0x14/0xc0 [xt_set]
xt_check_match+0x238/0x260 [x_tables]
__nft_match_init+0x160/0x180 [nft_compat]
[..]
sys_sendmsg+0x18/0x40
linux_sparc_syscall+0x34/0x44
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0 CPU1
---- ----
lock(&net->nft.commit_mutex);
lock(&table[i].mutex);
lock(&net->nft.commit_mutex);
lock(&table[i].mutex);Lockdep considers this an ABBA deadlock because the different nfnl subsys mutexes reside in the same lockdep class, but this is a false positive.
CPU1 table[i] refers to the nftables subsys mutex, whereas CPU1 locks the ipset subsys mutex.
Yi Che reported a similar lockdep splat, this time between ipset and ctnetlink subsys mutexes.
Time to place them in distinct classes to avoid these warnings.
--- net/netfilter/nfnetlink.c | 19 ++++++++++++++++++- 1 file changed, 18 insertions(+), 1 deletion(-)
net/netfilter/nfnetlink.c +18 -1
@@ -46,6 +46,23 @@ static struct {const struct nfnetlink_subsystem __rcu *subsys; } table[NFNL_SUBSYS_COUNT ]; +static struct lock_class_key nfnl_lockdep_keys[NFNL_SUBSYS_COUNT ]; + +static const char *const nfnl_lockdep_names[NFNL_SUBSYS_COUNT ] = { + [NFNL_SUBSYS_NONE ] ="nfnl_subsys_none" , + [NFNL_SUBSYS_CTNETLINK ] ="nfnl_subsys_ctnetlink" , + [NFNL_SUBSYS_CTNETLINK_EXP ] ="nfnl_subsys_ctnetlink_exp" , + [NFNL_SUBSYS_QUEUE ] ="nfnl_subsys_queue" , + [NFNL_SUBSYS_ULOG ] ="nfnl_subsys_ulog" , + [NFNL_SUBSYS_OSF ] ="nfnl_subsys_osf" , + [NFNL_SUBSYS_IPSET ] ="nfnl_subsys_ipset" , + [NFNL_SUBSYS_ACCT ] ="nfnl_subsys_acct" , + [NFNL_SUBSYS_CTNETLINK_TIMEOUT ] ="nfnl_subsys_cttimeout" , + [NFNL_SUBSYS_CTHELPER ] ="nfnl_subsys_cthelper" , + [NFNL_SUBSYS_NFTABLES ] ="nfnl_subsys_nftables" , + [NFNL_SUBSYS_NFT_COMPAT ] ="nfnl_subsys_nftcompat" , +}; +static const int nfnl_group2type[NFNLGRP_MAX +1 ] = { [NFNLGRP_CONNTRACK_NEW ] =NFNL_SUBSYS_CTNETLINK , [NFNLGRP_CONNTRACK_UPDATE ] =NFNL_SUBSYS_CTNETLINK ,
@@ -632,7 +649,7 @@ static int __init nfnetlink_init(void)BUG_ON (nfnl_group2type[i] ==NFNL_SUBSYS_NONE );for (i=0 ; i<NFNL_SUBSYS_COUNT ; i++) -mutex_init (&table[i].mutex); +__mutex_init (&table[i].mutex, nfnl_lockdep_names[i], &nfnl_lockdep_keys[i]);return register_pernet_subsys (&nfnetlink_net_ops); }
Signature · 1 line
2.26.2
On Fri, Oct 02, 2020 at 01:51:29PM +0200, Florian Westphal wrote:
[...]
Quoted 1 line
Time to place them in distinct classes to avoid these warnings.
Applied, thanks