| Message ID | 1361210344-14907-3-git-send-email-nikolay@redhat.com |
|---|---|
| State | Superseded, archived |
| Delegated to: | David Miller |
| Headers | show |
Nikolay Aleksandrov <nikolay@redhat.com> wrote: >This patch fixes the following inconsistencies in bond_release_all: >- IFF_BONDING flag is not stripped from slaves >- MTU is not restored >- no netdev notifiers are sent >Instead of trying to keep bond_release and bond_release_all in sync >I think we can re-use bond_release as the environment for calling it >is correct (RTNL is held). I have been running tests for the past >week and they came out successful. The only way for bond_release to fail >is for the slave to be attached in a different bond or to not be a slave >but that cannot happen as RTNL is held and no slave manipulations can be >achieved. It might be worthwhile to add an "all" argument to bond_release that skips some things that don't make sense if all slaves are being released. I'm thinking in particular of this block: if (oldcurrent == slave) { /* * Note that we hold RTNL over this sequence, so there * is no concern that another slave add/remove event * will interfere. */ write_unlock_bh(&bond->lock); read_lock(&bond->lock); write_lock_bh(&bond->curr_slave_lock); bond_select_active_slave(bond); write_unlock_bh(&bond->curr_slave_lock); read_unlock(&bond->lock); write_lock_bh(&bond->lock); } as it's written now, for the release all case, the code may go to the trouble of assigning a new active slave each time one slave is removed (including various log messages, maybe sending IGMPs, etc). If all slaves are being removed, that's pointless. This could be something like: if (release_all) { bond->curr_active_slave = NULL; } else if (oldcurrent == slave) { [ the current block of stuff ] } it's safe here to unconditionally set curr_active_slave to NULL because we hold bond->lock for write. The lock dance stuff for the bond_select_active_slave() call is to satisfy its locking requirements. -J >Signed-off-by: Nikolay Aleksandrov <nikolay@redhat.com> >--- > drivers/net/bonding/bond_main.c | 106 ++-------------------------------------- > 1 file changed, 5 insertions(+), 101 deletions(-) > >diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c >index 94c1534..fcfc880 100644 >--- a/drivers/net/bonding/bond_main.c >+++ b/drivers/net/bonding/bond_main.c >@@ -2140,113 +2140,17 @@ static int bond_release_and_destroy(struct net_device *bond_dev, > /* > * This function releases all slaves. > */ >-static int bond_release_all(struct net_device *bond_dev) >+static void bond_release_all(struct net_device *bond_dev) > { > struct bonding *bond = netdev_priv(bond_dev); >- struct slave *slave; >- struct net_device *slave_dev; >- struct sockaddr addr; >- >- write_lock_bh(&bond->lock); >- >- netif_carrier_off(bond_dev); > > if (bond->slave_cnt == 0) >- goto out; >- >- bond->current_arp_slave = NULL; >- bond->primary_slave = NULL; >- bond_change_active_slave(bond, NULL); >- >- while ((slave = bond->first_slave) != NULL) { >- /* Inform AD package of unbinding of slave >- * before slave is detached from the list. >- */ >- if (bond->params.mode == BOND_MODE_8023AD) >- bond_3ad_unbind_slave(slave); >- >- slave_dev = slave->dev; >- bond_detach_slave(bond, slave); >- >- /* now that the slave is detached, unlock and perform >- * all the undo steps that should not be called from >- * within a lock. >- */ >- write_unlock_bh(&bond->lock); >- >- /* unregister rx_handler early so bond_handle_frame wouldn't >- * be called for this slave anymore. >- */ >- netdev_rx_handler_unregister(slave_dev); >- synchronize_net(); >- >- if (bond_is_lb(bond)) { >- /* must be called only after the slave >- * has been detached from the list >- */ >- bond_alb_deinit_slave(bond, slave); >- } >- >- bond_destroy_slave_symlinks(bond_dev, slave_dev); >- bond_del_vlans_from_slave(bond, slave_dev); >- >- /* If the mode USES_PRIMARY, then we should only remove its >- * promisc and mc settings if it was the curr_active_slave, but that was >- * already taken care of above when we detached the slave >- */ >- if (!USES_PRIMARY(bond->params.mode)) { >- /* unset promiscuity level from slave */ >- if (bond_dev->flags & IFF_PROMISC) >- dev_set_promiscuity(slave_dev, -1); >- >- /* unset allmulti level from slave */ >- if (bond_dev->flags & IFF_ALLMULTI) >- dev_set_allmulti(slave_dev, -1); >- >- /* flush master's mc_list from slave */ >- netif_addr_lock_bh(bond_dev); >- bond_mc_list_flush(bond_dev, slave_dev); >- netif_addr_unlock_bh(bond_dev); >- } >- >- bond_upper_dev_unlink(bond_dev, slave_dev); >- >- slave_disable_netpoll(slave); >- >- /* close slave before restoring its mac address */ >- dev_close(slave_dev); >- >- if (!bond->params.fail_over_mac) { >- /* restore original ("permanent") mac address*/ >- memcpy(addr.sa_data, slave->perm_hwaddr, ETH_ALEN); >- addr.sa_family = slave_dev->type; >- dev_set_mac_address(slave_dev, &addr); >- } >- >- kfree(slave); >- >- /* re-acquire the lock before getting the next slave */ >- write_lock_bh(&bond->lock); >- } >- >- eth_hw_addr_random(bond_dev); >- bond->dev_addr_from_first = true; >- >- if (bond_vlan_used(bond)) { >- pr_warning("%s: Warning: clearing HW address of %s while it still has VLANs.\n", >- bond_dev->name, bond_dev->name); >- pr_warning("%s: When re-adding slaves, make sure the bond's HW address matches its VLANs'.\n", >- bond_dev->name); >- } >- >+ return; >+ while (bond->first_slave != NULL) >+ bond_release(bond_dev, bond->first_slave->dev); > pr_info("%s: released all slaves\n", bond_dev->name); > >-out: >- write_unlock_bh(&bond->lock); >- >- bond_compute_features(bond); >- >- return 0; >+ return; > } > > /* >-- >1.7.11.7 --- -Jay Vosburgh, IBM Linux Technology Center, fubar@us.ibm.com -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 18/02/13 22:56, Jay Vosburgh wrote: > Nikolay Aleksandrov <nikolay@redhat.com> wrote: > >> This patch fixes the following inconsistencies in bond_release_all: >> - IFF_BONDING flag is not stripped from slaves >> - MTU is not restored >> - no netdev notifiers are sent >> Instead of trying to keep bond_release and bond_release_all in sync >> I think we can re-use bond_release as the environment for calling it >> is correct (RTNL is held). I have been running tests for the past >> week and they came out successful. The only way for bond_release to fail >> is for the slave to be attached in a different bond or to not be a slave >> but that cannot happen as RTNL is held and no slave manipulations can be >> achieved. > > It might be worthwhile to add an "all" argument to bond_release > that skips some things that don't make sense if all slaves are being > released. I'm thinking in particular of this block: > > if (oldcurrent == slave) { > /* > * Note that we hold RTNL over this sequence, so there > * is no concern that another slave add/remove event > * will interfere. > */ > write_unlock_bh(&bond->lock); > read_lock(&bond->lock); > write_lock_bh(&bond->curr_slave_lock); > > bond_select_active_slave(bond); > > write_unlock_bh(&bond->curr_slave_lock); > read_unlock(&bond->lock); > write_lock_bh(&bond->lock); > } > > as it's written now, for the release all case, the code may go > to the trouble of assigning a new active slave each time one slave is > removed (including various log messages, maybe sending IGMPs, etc). If > all slaves are being removed, that's pointless. This could be something > like: > > if (release_all) { > bond->curr_active_slave = NULL; > } else if (oldcurrent == slave) { > [ the current block of stuff ] > } > > it's safe here to unconditionally set curr_active_slave to NULL > because we hold bond->lock for write. The lock dance stuff for the > bond_select_active_slave() call is to satisfy its locking requirements. > > -J I see your point and I agree. I will prepare another version that incorporates it, although I can't add it as an argument since bond_release is used as ndo_del_slave. I'll have to make it a global variable. Nik -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Nikolay Aleksandrov <nikolay@redhat.com> wrote: >On 18/02/13 22:56, Jay Vosburgh wrote: >> Nikolay Aleksandrov <nikolay@redhat.com> wrote: >> >>> This patch fixes the following inconsistencies in bond_release_all: >>> - IFF_BONDING flag is not stripped from slaves >>> - MTU is not restored >>> - no netdev notifiers are sent >>> Instead of trying to keep bond_release and bond_release_all in sync >>> I think we can re-use bond_release as the environment for calling it >>> is correct (RTNL is held). I have been running tests for the past >>> week and they came out successful. The only way for bond_release to fail >>> is for the slave to be attached in a different bond or to not be a slave >>> but that cannot happen as RTNL is held and no slave manipulations can be >>> achieved. >> >> It might be worthwhile to add an "all" argument to bond_release >> that skips some things that don't make sense if all slaves are being >> released. I'm thinking in particular of this block: >> >> if (oldcurrent == slave) { >> /* >> * Note that we hold RTNL over this sequence, so there >> * is no concern that another slave add/remove event >> * will interfere. >> */ >> write_unlock_bh(&bond->lock); >> read_lock(&bond->lock); >> write_lock_bh(&bond->curr_slave_lock); >> >> bond_select_active_slave(bond); >> >> write_unlock_bh(&bond->curr_slave_lock); >> read_unlock(&bond->lock); >> write_lock_bh(&bond->lock); >> } >> >> as it's written now, for the release all case, the code may go >> to the trouble of assigning a new active slave each time one slave is >> removed (including various log messages, maybe sending IGMPs, etc). If >> all slaves are being removed, that's pointless. This could be something >> like: >> >> if (release_all) { >> bond->curr_active_slave = NULL; >> } else if (oldcurrent == slave) { >> [ the current block of stuff ] >> } >> >> it's safe here to unconditionally set curr_active_slave to NULL >> because we hold bond->lock for write. The lock dance stuff for the >> bond_select_active_slave() call is to satisfy its locking requirements. >> >> -J >I see your point and I agree. I will prepare another version that >incorporates it, although I can't add it as an argument since >bond_release is used as ndo_del_slave. I'll have to make it a global >variable. No, just rename the current bond_release to __bond_release_one, add the extra argument, and create a new bond_release .ndo_del_slave that calls __bond_release_one with "all=0". Then, bond_release_all calls __bond_release_one with all=1. Also, there's only one caller of bond_release_all, and since the new & improved bond_release_all is trivial, it could be open coded into bond_uninit, eliminating bond_release_all as a function. -J --- -Jay Vosburgh, IBM Linux Technology Center, fubar@us.ibm.com -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c index 94c1534..fcfc880 100644 --- a/drivers/net/bonding/bond_main.c +++ b/drivers/net/bonding/bond_main.c @@ -2140,113 +2140,17 @@ static int bond_release_and_destroy(struct net_device *bond_dev, /* * This function releases all slaves. */ -static int bond_release_all(struct net_device *bond_dev) +static void bond_release_all(struct net_device *bond_dev) { struct bonding *bond = netdev_priv(bond_dev); - struct slave *slave; - struct net_device *slave_dev; - struct sockaddr addr; - - write_lock_bh(&bond->lock); - - netif_carrier_off(bond_dev); if (bond->slave_cnt == 0) - goto out; - - bond->current_arp_slave = NULL; - bond->primary_slave = NULL; - bond_change_active_slave(bond, NULL); - - while ((slave = bond->first_slave) != NULL) { - /* Inform AD package of unbinding of slave - * before slave is detached from the list. - */ - if (bond->params.mode == BOND_MODE_8023AD) - bond_3ad_unbind_slave(slave); - - slave_dev = slave->dev; - bond_detach_slave(bond, slave); - - /* now that the slave is detached, unlock and perform - * all the undo steps that should not be called from - * within a lock. - */ - write_unlock_bh(&bond->lock); - - /* unregister rx_handler early so bond_handle_frame wouldn't - * be called for this slave anymore. - */ - netdev_rx_handler_unregister(slave_dev); - synchronize_net(); - - if (bond_is_lb(bond)) { - /* must be called only after the slave - * has been detached from the list - */ - bond_alb_deinit_slave(bond, slave); - } - - bond_destroy_slave_symlinks(bond_dev, slave_dev); - bond_del_vlans_from_slave(bond, slave_dev); - - /* If the mode USES_PRIMARY, then we should only remove its - * promisc and mc settings if it was the curr_active_slave, but that was - * already taken care of above when we detached the slave - */ - if (!USES_PRIMARY(bond->params.mode)) { - /* unset promiscuity level from slave */ - if (bond_dev->flags & IFF_PROMISC) - dev_set_promiscuity(slave_dev, -1); - - /* unset allmulti level from slave */ - if (bond_dev->flags & IFF_ALLMULTI) - dev_set_allmulti(slave_dev, -1); - - /* flush master's mc_list from slave */ - netif_addr_lock_bh(bond_dev); - bond_mc_list_flush(bond_dev, slave_dev); - netif_addr_unlock_bh(bond_dev); - } - - bond_upper_dev_unlink(bond_dev, slave_dev); - - slave_disable_netpoll(slave); - - /* close slave before restoring its mac address */ - dev_close(slave_dev); - - if (!bond->params.fail_over_mac) { - /* restore original ("permanent") mac address*/ - memcpy(addr.sa_data, slave->perm_hwaddr, ETH_ALEN); - addr.sa_family = slave_dev->type; - dev_set_mac_address(slave_dev, &addr); - } - - kfree(slave); - - /* re-acquire the lock before getting the next slave */ - write_lock_bh(&bond->lock); - } - - eth_hw_addr_random(bond_dev); - bond->dev_addr_from_first = true; - - if (bond_vlan_used(bond)) { - pr_warning("%s: Warning: clearing HW address of %s while it still has VLANs.\n", - bond_dev->name, bond_dev->name); - pr_warning("%s: When re-adding slaves, make sure the bond's HW address matches its VLANs'.\n", - bond_dev->name); - } - + return; + while (bond->first_slave != NULL) + bond_release(bond_dev, bond->first_slave->dev); pr_info("%s: released all slaves\n", bond_dev->name); -out: - write_unlock_bh(&bond->lock); - - bond_compute_features(bond); - - return 0; + return; } /*
This patch fixes the following inconsistencies in bond_release_all: - IFF_BONDING flag is not stripped from slaves - MTU is not restored - no netdev notifiers are sent Instead of trying to keep bond_release and bond_release_all in sync I think we can re-use bond_release as the environment for calling it is correct (RTNL is held). I have been running tests for the past week and they came out successful. The only way for bond_release to fail is for the slave to be attached in a different bond or to not be a slave but that cannot happen as RTNL is held and no slave manipulations can be achieved. Signed-off-by: Nikolay Aleksandrov <nikolay@redhat.com> --- drivers/net/bonding/bond_main.c | 106 ++-------------------------------------- 1 file changed, 5 insertions(+), 101 deletions(-)