| Message ID | 20240404075740.30682-1-petr@tesarici.cz (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | 38a15d0a50e0a43778561a5861403851f0b0194c |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [net] u64_stats: fix u64_stats_init() for lockdep when used repeatedly in one file | expand |
On Thu, Apr 04, 2024 at 09:57:40AM +0200, Petr Tesarik wrote: > Fix bogus lockdep warnings if multiple u64_stats_sync variables are > initialized in the same file. > > With CONFIG_LOCKDEP, seqcount_init() is a macro which declares: > > static struct lock_class_key __key; > > Since u64_stats_init() is a function (albeit an inline one), all calls > within the same file end up using the same instance, effectively treating > them all as a single lock-class. > > Fixes: 9464ca650008 ("net: make u64_stats_init() a function") > Closes: https://lore.kernel.org/netdev/ea1567d9-ce66-45e6-8168-ac40a47d1821@roeck-us.net/ > Signed-off-by: Petr Tesarik <petr@tesarici.cz> Reviewed-by: Simon Horman <horms@kernel.org> Interesting bug. I'm wondering if you also looked over other users of u64_stats_init() to see if any of them can result in unexpected aliasing of lock keys too. ...
On Thu, Apr 4, 2024 at 9:58 AM Petr Tesarik <petr@tesarici.cz> wrote: > > Fix bogus lockdep warnings if multiple u64_stats_sync variables are > initialized in the same file. > > With CONFIG_LOCKDEP, seqcount_init() is a macro which declares: > > static struct lock_class_key __key; > > Since u64_stats_init() is a function (albeit an inline one), all calls > within the same file end up using the same instance, effectively treating > them all as a single lock-class. > > Fixes: 9464ca650008 ("net: make u64_stats_init() a function") > Closes: https://lore.kernel.org/netdev/ea1567d9-ce66-45e6-8168-ac40a47d1821@roeck-us.net/ > Signed-off-by: Petr Tesarik <petr@tesarici.cz> I thought I gave a Reviewed-by: tag already... Reviewed-by: Eric Dumazet <edumazet@google.com>
Hello: This patch was applied to netdev/net.git (main) by Jakub Kicinski <kuba@kernel.org>: On Thu, 4 Apr 2024 09:57:40 +0200 you wrote: > Fix bogus lockdep warnings if multiple u64_stats_sync variables are > initialized in the same file. > > With CONFIG_LOCKDEP, seqcount_init() is a macro which declares: > > static struct lock_class_key __key; > > [...] Here is the summary with links: - [net] u64_stats: fix u64_stats_init() for lockdep when used repeatedly in one file https://git.kernel.org/netdev/net/c/38a15d0a50e0 You are awesome, thank you!
On Fri, 5 Apr 2024 23:08:58 +0200 Eric Dumazet <edumazet@google.com> wrote: > On Thu, Apr 4, 2024 at 9:58 AM Petr Tesarik <petr@tesarici.cz> wrote: > > > > Fix bogus lockdep warnings if multiple u64_stats_sync variables are > > initialized in the same file. > > > > With CONFIG_LOCKDEP, seqcount_init() is a macro which declares: > > > > static struct lock_class_key __key; > > > > Since u64_stats_init() is a function (albeit an inline one), all calls > > within the same file end up using the same instance, effectively treating > > them all as a single lock-class. > > > > Fixes: 9464ca650008 ("net: make u64_stats_init() a function") > > Closes: https://lore.kernel.org/netdev/ea1567d9-ce66-45e6-8168-ac40a47d1821@roeck-us.net/ > > Signed-off-by: Petr Tesarik <petr@tesarici.cz> > > I thought I gave a Reviewed-by: tag already... You did, but at that time I didn't know which tree should be used to merge the patch, so it appeared as a stranded patch on LKML. Once we agreed it should go through net, I sent it again, but the resend didn't have it. Thanks for your assistance! Petr T > Reviewed-by: Eric Dumazet <edumazet@google.com>
On Fri, 5 Apr 2024 20:51:17 +0100 Simon Horman <horms@kernel.org> wrote: > On Thu, Apr 04, 2024 at 09:57:40AM +0200, Petr Tesarik wrote: > > Fix bogus lockdep warnings if multiple u64_stats_sync variables are > > initialized in the same file. > > > > With CONFIG_LOCKDEP, seqcount_init() is a macro which declares: > > > > static struct lock_class_key __key; > > > > Since u64_stats_init() is a function (albeit an inline one), all calls > > within the same file end up using the same instance, effectively treating > > them all as a single lock-class. > > > > Fixes: 9464ca650008 ("net: make u64_stats_init() a function") > > Closes: https://lore.kernel.org/netdev/ea1567d9-ce66-45e6-8168-ac40a47d1821@roeck-us.net/ > > Signed-off-by: Petr Tesarik <petr@tesarici.cz> > > Reviewed-by: Simon Horman <horms@kernel.org> > > Interesting bug. I'm wondering if you also looked over other users of > u64_stats_init() to see if any of them can result in unexpected aliasing of > lock keys too. I didn't. I have now run `git grep u64_stats_init | sed -e 's/:.*//' | uniq -d` and there indeed appears to be some aliasing in other files. block/blk-cgroup.c: - struct blkg_iostat_set.sync - struct blkg_iostat_set.sync drivers/net/ethernet/amazon/ena/ena_netdev.c: - struct ena_ring.syncp struct ena_adapter.syncp drivers/net/ethernet/aquantia/atlantic/aq_ring.c: - struct aq_ring_stats_tx_s.syncp - struct aq_ring_stats_rx_s.syncp drivers/net/ethernet/emulex/benet/be_main.c: - struct be_tx_stats.sync + aliasing between this struct in - struct be_tx_obj - struct be_rx_obj - struct be_tx_stats.sync_compl drivers/net/ethernet/google/gve/gve_main.c: - struct gve_priv.rx[].statss - struct gve_priv.tx[].statss drivers/net/ethernet/intel/fm10k/fm10k_netdev.c: - tx_ring->syncp - rx_ring->syncp drivers/net/ethernet/intel/i40e/i40e_txrx.c: - tx_ring->syncp - rx_ring->syncp drivers/net/ethernet/intel/igb/igb_main.c: - struct igb_ring.tx_syncp - struct igb_ring.tx_syncp2 - struct igb_ring.rx_syncp drivers/net/ethernet/intel/ixgbe/ixgbe_main.c: - struct ixgbe_adapter.rx_ring[]->syncp - struct ixgbe_adapter.tx_ring[]->syncp - struct ixgbe_adapter.xdp_ring[]->syncp drivers/net/ethernet/intel/ixgbevf/ixgbevf_main.c: - tx_ring->syncp - rx_ring->syncp drivers/net/ethernet/marvell/sky2.c: - struct sky2_port.tx_stats.syncp - struct sky2_port.rx_stats.syncp drivers/net/ethernet/microsoft/mana/mana_en.c: - struct mana_txq.stats.syncp - struct mana_rxq.stats.syncp drivers/net/ethernet/netronome/nfp/nfp_net_dp.c: - struct nfp_net_tx_ring.r_vec->tx_sync - struct nfp_net_rx_ring.r_vec->rx_sync drivers/net/ethernet/nvidia/forcedeth.c: - struct fe_priv.swstats_tx_syncp - struct fe_priv.swstats_rx_syncp drivers/net/ethernet/realtek/8139too.c: - struct rtl8139_private.tx_stats.syncp - struct rtl8139_private.rx_stats.syncp drivers/net/ethernet/socionext/sni_ave.c: - struct ave_private.stats_tx.syncp - struct ave_private.stats_rx.syncp drivers/net/ethernet/stmicro/stmmac/stmmac_main.c: - this was the trigger ;-) drivers/net/ethernet/ti/netcp_core.c: - struct netcp_stats.syncp_tx - struct netcp_stats.syncp_rx drivers/net/ethernet/via/via-rhine.c: - struct rhine_private.tx_stats.syncp - struct rhine_private.rx_stats.syncp drivers/net/ethernet/xilinx/xilinx_axienet_main.c: - struct axienet_local.tx_stat_sync - struct axienet_local.rx_stat_sync drivers/net/hyperv/netvsc.c: - struct netvsc_channel.tx_stats.syncp - struct netvsc_channel.rx_stats.syncp drivers/net/ifb.c: - struct ifb_q_private.tx_stats.sync - struct ifb_q_private.rx_stats.sync drivers/net/mhi_net.c: - struct mhi_net_stats.tx_syncp - struct mhi_net_stats.rx_syncp drivers/net/virtio_net.c: - struct virtnet_info.rq[].stats.syncp - struct virtnet_info.sq[].stats.syncp drivers/net/wwan/mhi_wwan_mbim.c: - struct mhi_mbim_link.tx_syncp - struct mhi_mbim_link.rx_syncp drivers/vdpa/vdpa_sim/vdpa_sim_net.c: - struct vdpasim_net.tx_stats.syncp - struct vdpasim_net.rx_stats.syncp - struct vdpasim_net.cq_stats.syncp HTH Petr T
diff --git a/include/linux/u64_stats_sync.h b/include/linux/u64_stats_sync.h index ffe48e69b3f3..457879938fc1 100644 --- a/include/linux/u64_stats_sync.h +++ b/include/linux/u64_stats_sync.h @@ -135,10 +135,11 @@ static inline void u64_stats_inc(u64_stats_t *p) p->v++; } -static inline void u64_stats_init(struct u64_stats_sync *syncp) -{ - seqcount_init(&syncp->seq); -} +#define u64_stats_init(syncp) \ + do { \ + struct u64_stats_sync *__s = (syncp); \ + seqcount_init(&__s->seq); \ + } while (0) static inline void __u64_stats_update_begin(struct u64_stats_sync *syncp) {
Fix bogus lockdep warnings if multiple u64_stats_sync variables are initialized in the same file. With CONFIG_LOCKDEP, seqcount_init() is a macro which declares: static struct lock_class_key __key; Since u64_stats_init() is a function (albeit an inline one), all calls within the same file end up using the same instance, effectively treating them all as a single lock-class. Fixes: 9464ca650008 ("net: make u64_stats_init() a function") Closes: https://lore.kernel.org/netdev/ea1567d9-ce66-45e6-8168-ac40a47d1821@roeck-us.net/ Signed-off-by: Petr Tesarik <petr@tesarici.cz> --- include/linux/u64_stats_sync.h | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-)