| Message ID | 20240603184826.1087245-1-kuba@kernel.org (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | 5b4b62a169e10401cca34a6e7ac39161986f5605 |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [net,v3] rtnetlink: make the "split" NLM_DONE handling generic | expand |
On Mon, Jun 3, 2024 at 8:48 PM Jakub Kicinski <kuba@kernel.org> wrote: > > Jaroslav reports Dell's OMSA Systems Management Data Engine > expects NLM_DONE in a separate recvmsg(), both for rtnl_dump_ifinfo() > and inet_dump_ifaddr(). We already added a similar fix previously in > commit 460b0d33cf10 ("inet: bring NLM_DONE out to a separate recv() again") > > Instead of modifying all the dump handlers, and making them look > different than modern for_each_netdev_dump()-based dump handlers - > put the workaround in rtnetlink code. This will also help us move > the custom rtnl-locking from af_netlink in the future (in net-next). > > Note that this change is not touching rtnl_dump_all(). rtnl_dump_all() > is different kettle of fish and a potential problem. We now mix families > in a single recvmsg(), but NLM_DONE is not coalesced. > > Tested: > > ./cli.py --dbg-small-recv 4096 --spec netlink/specs/rt_addr.yaml \ > --dump getaddr --json '{"ifa-family": 2}' > > ./cli.py --dbg-small-recv 4096 --spec netlink/specs/rt_route.yaml \ > --dump getroute --json '{"rtm-family": 2}' > > ./cli.py --dbg-small-recv 4096 --spec netlink/specs/rt_link.yaml \ > --dump getlink > > Fixes: 3e41af90767d ("rtnetlink: use xarray iterator to implement rtnl_dump_ifinfo()") > Fixes: cdb2f80f1c10 ("inet: use xa_array iterator to implement inet_dump_ifaddr()") > Reported-by: Jaroslav Pulchart <jaroslav.pulchart@gooddata.com> > Link: https://lore.kernel.org/all/CAK8fFZ7MKoFSEzMBDAOjoUt+vTZRRQgLDNXEOfdCCXSoXXKE0g@mail.gmail.com > Signed-off-by: Jakub Kicinski <kuba@kernel.org> Very nice, thanks a lot for taking care of this Jakub. Reviewed-by: Eric Dumazet <edumazet@google.com>
On 6/3/24 12:48 PM, Jakub Kicinski wrote: > @@ -6694,7 +6734,7 @@ void __init rtnetlink_init(void) > register_netdevice_notifier(&rtnetlink_dev_notifier); > > rtnl_register(PF_UNSPEC, RTM_GETLINK, rtnl_getlink, > - rtnl_dump_ifinfo, 0); > + rtnl_dump_ifinfo, RTNL_FLAG_DUMP_SPLIT_NLM_DONE); > rtnl_register(PF_UNSPEC, RTM_SETLINK, rtnl_setlink, NULL, 0); > rtnl_register(PF_UNSPEC, RTM_NEWLINK, rtnl_newlink, NULL, 0); > rtnl_register(PF_UNSPEC, RTM_DELLINK, rtnl_dellink, NULL, 0); > diff --git a/net/ipv4/devinet.c b/net/ipv4/devinet.c > index f3892ee9dfb3..d09f557eaa77 100644 > --- a/net/ipv4/devinet.c > +++ b/net/ipv4/devinet.c > @@ -2805,7 +2805,7 @@ void __init devinet_init(void) > rtnl_register(PF_INET, RTM_NEWADDR, inet_rtm_newaddr, NULL, 0); > rtnl_register(PF_INET, RTM_DELADDR, inet_rtm_deladdr, NULL, 0); > rtnl_register(PF_INET, RTM_GETADDR, NULL, inet_dump_ifaddr, > - RTNL_FLAG_DUMP_UNLOCKED); > + RTNL_FLAG_DUMP_UNLOCKED | RTNL_FLAG_DUMP_SPLIT_NLM_DONE); > rtnl_register(PF_INET, RTM_GETNETCONF, inet_netconf_get_devconf, > inet_netconf_dump_devconf, > RTNL_FLAG_DOIT_UNLOCKED | RTNL_FLAG_DUMP_UNLOCKED); > diff --git a/net/ipv4/fib_frontend.c b/net/ipv4/fib_frontend.c > index c484b1c0fc00..7ad2cafb9276 100644 > --- a/net/ipv4/fib_frontend.c > +++ b/net/ipv4/fib_frontend.c > @@ -1050,11 +1050,6 @@ static int inet_dump_fib(struct sk_buff *skb, struct netlink_callback *cb) > e++; > } > } > - > - /* Don't let NLM_DONE coalesce into a message, even if it could. > - * Some user space expects NLM_DONE in a separate recv(). > - */ > - err = skb->len; > out: > > cb->args[1] = e; > @@ -1665,5 +1660,5 @@ void __init ip_fib_init(void) > rtnl_register(PF_INET, RTM_NEWROUTE, inet_rtm_newroute, NULL, 0); > rtnl_register(PF_INET, RTM_DELROUTE, inet_rtm_delroute, NULL, 0); > rtnl_register(PF_INET, RTM_GETROUTE, NULL, inet_dump_fib, > - RTNL_FLAG_DUMP_UNLOCKED); > + RTNL_FLAG_DUMP_UNLOCKED | RTNL_FLAG_DUMP_SPLIT_NLM_DONE); > } You are using the legacy way for route, link and address dumps with the option to quickly add the others as needed (meaning when a regression is reported). If those 3 need the workaround, I think there are high odds all of the other rtnl_register users will need it. So, if there is not going to be a per-app opt-in to a new way, you might as well make this the default for all rtnl families - sadly.
On Mon, 3 Jun 2024 20:33:50 -0600 David Ahern wrote: > You are using the legacy way for route, link and address dumps with the > option to quickly add the others as needed (meaning when a regression is > reported). If those 3 need the workaround, I think there are high odds > all of the other rtnl_register users will need it. So, if there is not > going to be a per-app opt-in to a new way, you might as well make this > the default for all rtnl families - sadly. I prefer the default (for new code) to be "modern".
Hello: This patch was applied to netdev/net.git (main) by David S. Miller <davem@davemloft.net>: On Mon, 3 Jun 2024 11:48:26 -0700 you wrote: > Jaroslav reports Dell's OMSA Systems Management Data Engine > expects NLM_DONE in a separate recvmsg(), both for rtnl_dump_ifinfo() > and inet_dump_ifaddr(). We already added a similar fix previously in > commit 460b0d33cf10 ("inet: bring NLM_DONE out to a separate recv() again") > > Instead of modifying all the dump handlers, and making them look > different than modern for_each_netdev_dump()-based dump handlers - > put the workaround in rtnetlink code. This will also help us move > the custom rtnl-locking from af_netlink in the future (in net-next). > > [...] Here is the summary with links: - [net,v3] rtnetlink: make the "split" NLM_DONE handling generic https://git.kernel.org/netdev/net/c/5b4b62a169e1 You are awesome, thank you!
diff --git a/include/net/rtnetlink.h b/include/net/rtnetlink.h index 3bfb80bad173..b45d57b5968a 100644 --- a/include/net/rtnetlink.h +++ b/include/net/rtnetlink.h @@ -13,6 +13,7 @@ enum rtnl_link_flags { RTNL_FLAG_DOIT_UNLOCKED = BIT(0), RTNL_FLAG_BULK_DEL_SUPPORTED = BIT(1), RTNL_FLAG_DUMP_UNLOCKED = BIT(2), + RTNL_FLAG_DUMP_SPLIT_NLM_DONE = BIT(3), /* legacy behavior */ }; enum rtnl_kinds { diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c index b86b0a87367d..4668d6718040 100644 --- a/net/core/rtnetlink.c +++ b/net/core/rtnetlink.c @@ -6484,6 +6484,46 @@ static int rtnl_mdb_del(struct sk_buff *skb, struct nlmsghdr *nlh, /* Process one rtnetlink message. */ +static int rtnl_dumpit(struct sk_buff *skb, struct netlink_callback *cb) +{ + rtnl_dumpit_func dumpit = cb->data; + int err; + + /* Previous iteration have already finished, avoid calling->dumpit() + * again, it may not expect to be called after it reached the end. + */ + if (!dumpit) + return 0; + + err = dumpit(skb, cb); + + /* Old dump handlers used to send NLM_DONE as in a separate recvmsg(). + * Some applications which parse netlink manually depend on this. + */ + if (cb->flags & RTNL_FLAG_DUMP_SPLIT_NLM_DONE) { + if (err < 0 && err != -EMSGSIZE) + return err; + if (!err) + cb->data = NULL; + + return skb->len; + } + return err; +} + +static int rtnetlink_dump_start(struct sock *ssk, struct sk_buff *skb, + const struct nlmsghdr *nlh, + struct netlink_dump_control *control) +{ + if (control->flags & RTNL_FLAG_DUMP_SPLIT_NLM_DONE) { + WARN_ON(control->data); + control->data = control->dump; + control->dump = rtnl_dumpit; + } + + return netlink_dump_start(ssk, skb, nlh, control); +} + static int rtnetlink_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh, struct netlink_ext_ack *extack) { @@ -6548,7 +6588,7 @@ static int rtnetlink_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh, .module = owner, .flags = flags, }; - err = netlink_dump_start(rtnl, skb, nlh, &c); + err = rtnetlink_dump_start(rtnl, skb, nlh, &c); /* netlink_dump_start() will keep a reference on * module if dump is still in progress. */ @@ -6694,7 +6734,7 @@ void __init rtnetlink_init(void) register_netdevice_notifier(&rtnetlink_dev_notifier); rtnl_register(PF_UNSPEC, RTM_GETLINK, rtnl_getlink, - rtnl_dump_ifinfo, 0); + rtnl_dump_ifinfo, RTNL_FLAG_DUMP_SPLIT_NLM_DONE); rtnl_register(PF_UNSPEC, RTM_SETLINK, rtnl_setlink, NULL, 0); rtnl_register(PF_UNSPEC, RTM_NEWLINK, rtnl_newlink, NULL, 0); rtnl_register(PF_UNSPEC, RTM_DELLINK, rtnl_dellink, NULL, 0); diff --git a/net/ipv4/devinet.c b/net/ipv4/devinet.c index f3892ee9dfb3..d09f557eaa77 100644 --- a/net/ipv4/devinet.c +++ b/net/ipv4/devinet.c @@ -2805,7 +2805,7 @@ void __init devinet_init(void) rtnl_register(PF_INET, RTM_NEWADDR, inet_rtm_newaddr, NULL, 0); rtnl_register(PF_INET, RTM_DELADDR, inet_rtm_deladdr, NULL, 0); rtnl_register(PF_INET, RTM_GETADDR, NULL, inet_dump_ifaddr, - RTNL_FLAG_DUMP_UNLOCKED); + RTNL_FLAG_DUMP_UNLOCKED | RTNL_FLAG_DUMP_SPLIT_NLM_DONE); rtnl_register(PF_INET, RTM_GETNETCONF, inet_netconf_get_devconf, inet_netconf_dump_devconf, RTNL_FLAG_DOIT_UNLOCKED | RTNL_FLAG_DUMP_UNLOCKED); diff --git a/net/ipv4/fib_frontend.c b/net/ipv4/fib_frontend.c index c484b1c0fc00..7ad2cafb9276 100644 --- a/net/ipv4/fib_frontend.c +++ b/net/ipv4/fib_frontend.c @@ -1050,11 +1050,6 @@ static int inet_dump_fib(struct sk_buff *skb, struct netlink_callback *cb) e++; } } - - /* Don't let NLM_DONE coalesce into a message, even if it could. - * Some user space expects NLM_DONE in a separate recv(). - */ - err = skb->len; out: cb->args[1] = e; @@ -1665,5 +1660,5 @@ void __init ip_fib_init(void) rtnl_register(PF_INET, RTM_NEWROUTE, inet_rtm_newroute, NULL, 0); rtnl_register(PF_INET, RTM_DELROUTE, inet_rtm_delroute, NULL, 0); rtnl_register(PF_INET, RTM_GETROUTE, NULL, inet_dump_fib, - RTNL_FLAG_DUMP_UNLOCKED); + RTNL_FLAG_DUMP_UNLOCKED | RTNL_FLAG_DUMP_SPLIT_NLM_DONE); }
Jaroslav reports Dell's OMSA Systems Management Data Engine expects NLM_DONE in a separate recvmsg(), both for rtnl_dump_ifinfo() and inet_dump_ifaddr(). We already added a similar fix previously in commit 460b0d33cf10 ("inet: bring NLM_DONE out to a separate recv() again") Instead of modifying all the dump handlers, and making them look different than modern for_each_netdev_dump()-based dump handlers - put the workaround in rtnetlink code. This will also help us move the custom rtnl-locking from af_netlink in the future (in net-next). Note that this change is not touching rtnl_dump_all(). rtnl_dump_all() is different kettle of fish and a potential problem. We now mix families in a single recvmsg(), but NLM_DONE is not coalesced. Tested: ./cli.py --dbg-small-recv 4096 --spec netlink/specs/rt_addr.yaml \ --dump getaddr --json '{"ifa-family": 2}' ./cli.py --dbg-small-recv 4096 --spec netlink/specs/rt_route.yaml \ --dump getroute --json '{"rtm-family": 2}' ./cli.py --dbg-small-recv 4096 --spec netlink/specs/rt_link.yaml \ --dump getlink Fixes: 3e41af90767d ("rtnetlink: use xarray iterator to implement rtnl_dump_ifinfo()") Fixes: cdb2f80f1c10 ("inet: use xa_array iterator to implement inet_dump_ifaddr()") Reported-by: Jaroslav Pulchart <jaroslav.pulchart@gooddata.com> Link: https://lore.kernel.org/all/CAK8fFZ7MKoFSEzMBDAOjoUt+vTZRRQgLDNXEOfdCCXSoXXKE0g@mail.gmail.com Signed-off-by: Jakub Kicinski <kuba@kernel.org> --- v3: - change the approach to centralized handling + flag v2: - adjust the comment in the kdoc - add getlink CC: dsahern@kernel.org --- include/net/rtnetlink.h | 1 + net/core/rtnetlink.c | 44 +++++++++++++++++++++++++++++++++++++++-- net/ipv4/devinet.c | 2 +- net/ipv4/fib_frontend.c | 7 +------ 4 files changed, 45 insertions(+), 9 deletions(-)