| Message ID | 20201030202727.1053534-1-cezarsa@gmail.com (mailing list archive) |
|---|---|
| State | RFC |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [RFC] ipvs: add genetlink cmd to dump all services and destinations | expand |
Hello, On Fri, 30 Oct 2020, Cezar Sa Espinola wrote: > A common operation for userspace applications managing ipvs is to dump > all services and all destinations and then sort out what needs to be > done. Previously this could only be accomplished by issuing 1 netlink > IPVS_CMD_GET_SERVICE dump command followed by N IPVS_CMD_GET_DEST dump > commands. For a dynamic system with a very large number of services this > could be cause a performance impact. > > This patch introduces a new way of dumping all services and destinations > with the new IPVS_CMD_GET_SERVICE_DEST command. A dump call for this > command will send the services as IPVS_CMD_NEW_SERVICE messages > imediatelly followed by its destinations as multiple IPVS_CMD_NEW_DEST > messages. It's also possible to dump a single service and its > destinations by sending a IPVS_CMD_ATTR_SERVICE argument to the dump > command. > > Signed-off-by: Cezar Sa Espinola <cezarsa@gmail.com> > --- > > To ensure that this patch improves performance I decided to also patch > ipvsadm in order to run some benchmarks comparing 'ipvsadm -Sn' with the > unpatched version. The ipvsadm patch is available on github in [1] for > now but I intend to submit it if this RFC goes forward. > > The benchmarks look nice and detailed results and some scripts to allow > reproducing then are available in another github repository [2]. The > summary of the benchmarks is: > > svcs | dsts | run time compared to unpatched > ----- | ---- | ------------------------------ > 1000 | 4 | -60.63% > 2000 | 2 | -71.10% > 8000 | 2 | -52.83% > 16000 | 1 | -54.13% > 100 | 100 | -4.76% > > [1] - https://github.com/cezarsa/ipvsadm/compare/master...dump-svc-ds > [2] - https://github.com/cezarsa/ipvsadm-validate#benchmark-results > > include/uapi/linux/ip_vs.h | 2 + > net/netfilter/ipvs/ip_vs_ctl.c | 109 +++++++++++++++++++++++++++++++++ > 2 files changed, 111 insertions(+) > > diff --git a/include/uapi/linux/ip_vs.h b/include/uapi/linux/ip_vs.h > index 4102ddcb4e14..353548cb7b81 100644 > --- a/include/uapi/linux/ip_vs.h > +++ b/include/uapi/linux/ip_vs.h > @@ -331,6 +331,8 @@ enum { > IPVS_CMD_ZERO, /* zero all counters and stats */ > IPVS_CMD_FLUSH, /* flush services and dests */ > > + IPVS_CMD_GET_SERVICE_DEST, /* get service and destination info */ > + > __IPVS_CMD_MAX, > }; > > diff --git a/net/netfilter/ipvs/ip_vs_ctl.c b/net/netfilter/ipvs/ip_vs_ctl.c > index e279ded4e306..09a7dd823dc0 100644 > --- a/net/netfilter/ipvs/ip_vs_ctl.c > +++ b/net/netfilter/ipvs/ip_vs_ctl.c > @@ -3396,6 +3396,109 @@ static int ip_vs_genl_dump_dests(struct sk_buff *skb, > return skb->len; > } > > +struct dump_services_dests_ctx { > + struct ip_vs_service *last_svc; > + int idx_svc; > + int idx_dest; > + int start_svc; > + int start_dest; > +}; > + > +static int ip_vs_genl_dump_service_destinations(struct sk_buff *skb, > + struct netlink_callback *cb, > + struct ip_vs_service *svc, > + struct dump_services_dests_ctx *ctx) > +{ > + struct ip_vs_dest *dest; > + > + if (++ctx->idx_svc < ctx->start_svc) > + return 0; > + > + if (ctx->idx_svc == ctx->start_svc && ctx->last_svc != svc) > + return 0; > + > + if (ctx->idx_svc > ctx->start_svc) { > + if (ip_vs_genl_dump_service(skb, svc, cb) < 0) { > + ctx->idx_svc--; > + return -EMSGSIZE; > + } > + ctx->last_svc = svc; > + ctx->start_dest = 0; > + } > + > + ctx->idx_dest = 0; > + list_for_each_entry(dest, &svc->destinations, n_list) { > + if (++ctx->idx_dest <= ctx->start_dest) > + continue; > + if (ip_vs_genl_dump_dest(skb, dest, cb) < 0) { > + ctx->idx_dest--; At this point idx_svc is incremented and we stop at the middle of dest list, so we need ctx->idx_svc-- too. And now what happens if all dests can not fit in a packet? We should start next packet with the same svc? And then user space should merge the dests when multiple packets start with same service? The main points are: - the virtual services are in hash table, their order is not important, user space can sort them - order of dests in a service is important for the schedulers - every packet should contain info for svc, so that we can properly add dests to the right svc > + return -EMSGSIZE; > + } > + } > + > + return 0; > +} > + > +static int ip_vs_genl_dump_services_destinations(struct sk_buff *skb, > + struct netlink_callback *cb) > +{ > + /* Besides usual index based counters, saving a pointer to the last > + * dumped service is useful to ensure we only dump destinations that > + * belong to it, even when services are removed while the dump is still > + * running causing indexes to shift. > + */ > + struct dump_services_dests_ctx ctx = { > + .idx_svc = 0, > + .idx_dest = 0, > + .start_svc = cb->args[0], > + .start_dest = cb->args[1], > + .last_svc = (struct ip_vs_service *)(cb->args[2]), > + }; > + struct net *net = sock_net(skb->sk); > + struct netns_ipvs *ipvs = net_ipvs(net); > + struct ip_vs_service *svc = NULL; > + struct nlattr *attrs[IPVS_CMD_ATTR_MAX + 1]; > + int i; > + > + mutex_lock(&__ip_vs_mutex); > + > + if (nlmsg_parse_deprecated(cb->nlh, GENL_HDRLEN, attrs, IPVS_CMD_ATTR_MAX, > + ip_vs_cmd_policy, cb->extack) == 0) { > + svc = ip_vs_genl_find_service(ipvs, attrs[IPVS_CMD_ATTR_SERVICE]); > + > + if (!IS_ERR_OR_NULL(svc)) { > + ip_vs_genl_dump_service_destinations(skb, cb, svc, &ctx); > + goto nla_put_failure; > + } > + } > + > + for (i = 0; i < IP_VS_SVC_TAB_SIZE; i++) { > + hlist_for_each_entry(svc, &ip_vs_svc_table[i], s_list) { > + if (svc->ipvs != ipvs) > + continue; > + if (ip_vs_genl_dump_service_destinations(skb, cb, svc, &ctx) < 0) > + goto nla_put_failure; > + } > + } > + > + for (i = 0; i < IP_VS_SVC_TAB_SIZE; i++) { > + hlist_for_each_entry(svc, &ip_vs_svc_fwm_table[i], s_list) { > + if (svc->ipvs != ipvs) > + continue; > + if (ip_vs_genl_dump_service_destinations(skb, cb, svc, &ctx) < 0) > + goto nla_put_failure; > + } > + } > + > +nla_put_failure: > + mutex_unlock(&__ip_vs_mutex); > + cb->args[0] = ctx.idx_svc; > + cb->args[1] = ctx.idx_dest; > + cb->args[2] = (long)ctx.last_svc; last_svc is used out of __ip_vs_mutex region, so it is not safe. We can get a reference count but this is bad if user space blocks. But even if we use just indexes it should be ok. If multiple agents are used in parallel it is not our problem. What can happen is that we can send duplicates or to skip entries (both svcs and dests). It is impossible to keep any kind of references to current entries or even keys to lookup them if another agent can remove them. > + > + return skb->len; > +} > + > static int ip_vs_genl_parse_dest(struct ip_vs_dest_user_kern *udest, > struct nlattr *nla, bool full_entry) > { > @@ -3991,6 +4094,12 @@ static const struct genl_small_ops ip_vs_genl_ops[] = { > .flags = GENL_ADMIN_PERM, > .doit = ip_vs_genl_set_cmd, > }, > + { > + .cmd = IPVS_CMD_GET_SERVICE_DEST, > + .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP, > + .flags = GENL_ADMIN_PERM, > + .dumpit = ip_vs_genl_dump_services_destinations, > + }, > }; > > static struct genl_family ip_vs_genl_family __ro_after_init = { > -- Regards -- Julian Anastasov <ja@ssi.bg>
Hi, > > + if (ctx->idx_svc == ctx->start_svc && ctx->last_svc != svc) > > + return 0; > > + > > + if (ctx->idx_svc > ctx->start_svc) { > > + if (ip_vs_genl_dump_service(skb, svc, cb) < 0) { > > + ctx->idx_svc--; > > + return -EMSGSIZE; > > + } > > + ctx->last_svc = svc; > > + ctx->start_dest = 0; > > + } > > + > > + ctx->idx_dest = 0; > > + list_for_each_entry(dest, &svc->destinations, n_list) { > > + if (++ctx->idx_dest <= ctx->start_dest) > > + continue; > > + if (ip_vs_genl_dump_dest(skb, dest, cb) < 0) { > > + ctx->idx_dest--; > > At this point idx_svc is incremented and we > stop at the middle of dest list, so we need ctx->idx_svc-- too. > > And now what happens if all dests can not fit in a packet? > We should start next packet with the same svc? And then > user space should merge the dests when multiple packets > start with same service? My (maybe not so great) idea was to avoid repeating the svc on each packet. It's possible for a packet to start with a destination and user space must consider then as belonging to the last svc received on the previous packet. The comparison "ctx->last_svc != svc" was intended to ensure that a packet only starts with destinations if the current service is the same as the svc we sent on the previous packet. > > The main points are: > > - the virtual services are in hash table, their order is > not important, user space can sort them > > - order of dests in a service is important for the schedulers > > - every packet should contain info for svc, so that we can > properly add dests to the right svc Thanks, I will rework the patch with these points in mind. It does sound safer to ensure every packet starts with service information. > > +nla_put_failure: > > + mutex_unlock(&__ip_vs_mutex); > > + cb->args[0] = ctx.idx_svc; > > + cb->args[1] = ctx.idx_dest; > > + cb->args[2] = (long)ctx.last_svc; > > last_svc is used out of __ip_vs_mutex region, > so it is not safe. We can get a reference count but this > is bad if user space blocks. I thought it would be relatively safe to store a pointer to the last svc since I would only use it for pointer comparison and never dereferencing it. But in retrospect it does look unsafe and fragile and could probably lead to errors especially if services are modified during a dump causing the stored pointer to point to a different service. > But even if we use just indexes it should be ok. > If multiple agents are used in parallel it is not our > problem. What can happen is that we can send duplicates > or to skip entries (both svcs and dests). It is impossible > to keep any kind of references to current entries or even > keys to lookup them if another agent can remove them. Got it. I noticed this behavior while writing this patch and even created a few crude validation scripts running parallel agents and checking the diff in [1]. [1] - https://github.com/cezarsa/ipvsadm-validate/blob/37ebd39785b1e835c6d4b5c58aaca7be60d5e194/test.sh#L86-L87 Thanks a lot for the review, -- Cezar Sá Espinola
Hello, On Tue, 3 Nov 2020, Cezar Sá Espinola wrote: > > And now what happens if all dests can not fit in a packet? > > We should start next packet with the same svc? And then > > user space should merge the dests when multiple packets > > start with same service? > > My (maybe not so great) idea was to avoid repeating the svc on each > packet. It's possible for a packet to start with a destination and > user space must consider then as belonging to the last svc received on > the previous packet. The comparison "ctx->last_svc != svc" was > intended to ensure that a packet only starts with destinations if the > current service is the same as the svc we sent on the previous packet. You can also consider the idea of having 3 coordinates for start svc: idx_svc_tab (0 or 1), idx_svc_row (0..IP_VS_SVC_TAB_SIZE-1) and idx_svc for index in row's chain. On new packet this will indicate the htable and its row and we have to skip svcs in this row to find our starting svc. I think, this will still fit in the netlink_callback's args area. If not, we can always kmalloc our context in args[0]. In single table, this should speedup the start svc lookup 128 times in average (we have 256 rows). In setup with 1024 svcs (average 4 in each of the 256 rows) we should skip these 0..3 entries instead of 512 in average. > > last_svc is used out of __ip_vs_mutex region, > > so it is not safe. We can get a reference count but this > > is bad if user space blocks. > > I thought it would be relatively safe to store a pointer to the last > svc since I would only use it for pointer comparison and never > dereferencing it. But in retrospect it does look unsafe and fragile > and could probably lead to errors especially if services are modified > during a dump causing the stored pointer to point to a different > service. Yes, nobody is using such pointers. We should create packets that correctly identify svc for the dests. The drawback is that user space may need more work for merging. We can always create a sorted array of pointers to svcs, so that we can binary search with bsearch() the svc from every received packet. Then we will know if this is a new svc or an old one (with dests in multiple packets). Should we also check for dest duplicates in the svc? The question is how much safe we should play. In user space the max work we can do is to avoid duplicates and to put dests to their correct svc. > > But even if we use just indexes it should be ok. > > If multiple agents are used in parallel it is not our > > problem. What can happen is that we can send duplicates > > or to skip entries (both svcs and dests). It is impossible > > to keep any kind of references to current entries or even > > keys to lookup them if another agent can remove them. > > Got it. I noticed this behavior while writing this patch and even > created a few crude validation scripts running parallel agents and > checking the diff in [1]. Ok, make sure your tests cover cases with multiple dests, so that single service occupies multiple packets, I'm not sure if 100 dests fit in one packet or not. Regards -- Julian Anastasov <ja@ssi.bg>
diff --git a/include/uapi/linux/ip_vs.h b/include/uapi/linux/ip_vs.h index 4102ddcb4e14..353548cb7b81 100644 --- a/include/uapi/linux/ip_vs.h +++ b/include/uapi/linux/ip_vs.h @@ -331,6 +331,8 @@ enum { IPVS_CMD_ZERO, /* zero all counters and stats */ IPVS_CMD_FLUSH, /* flush services and dests */ + IPVS_CMD_GET_SERVICE_DEST, /* get service and destination info */ + __IPVS_CMD_MAX, }; diff --git a/net/netfilter/ipvs/ip_vs_ctl.c b/net/netfilter/ipvs/ip_vs_ctl.c index e279ded4e306..09a7dd823dc0 100644 --- a/net/netfilter/ipvs/ip_vs_ctl.c +++ b/net/netfilter/ipvs/ip_vs_ctl.c @@ -3396,6 +3396,109 @@ static int ip_vs_genl_dump_dests(struct sk_buff *skb, return skb->len; } +struct dump_services_dests_ctx { + struct ip_vs_service *last_svc; + int idx_svc; + int idx_dest; + int start_svc; + int start_dest; +}; + +static int ip_vs_genl_dump_service_destinations(struct sk_buff *skb, + struct netlink_callback *cb, + struct ip_vs_service *svc, + struct dump_services_dests_ctx *ctx) +{ + struct ip_vs_dest *dest; + + if (++ctx->idx_svc < ctx->start_svc) + return 0; + + if (ctx->idx_svc == ctx->start_svc && ctx->last_svc != svc) + return 0; + + if (ctx->idx_svc > ctx->start_svc) { + if (ip_vs_genl_dump_service(skb, svc, cb) < 0) { + ctx->idx_svc--; + return -EMSGSIZE; + } + ctx->last_svc = svc; + ctx->start_dest = 0; + } + + ctx->idx_dest = 0; + list_for_each_entry(dest, &svc->destinations, n_list) { + if (++ctx->idx_dest <= ctx->start_dest) + continue; + if (ip_vs_genl_dump_dest(skb, dest, cb) < 0) { + ctx->idx_dest--; + return -EMSGSIZE; + } + } + + return 0; +} + +static int ip_vs_genl_dump_services_destinations(struct sk_buff *skb, + struct netlink_callback *cb) +{ + /* Besides usual index based counters, saving a pointer to the last + * dumped service is useful to ensure we only dump destinations that + * belong to it, even when services are removed while the dump is still + * running causing indexes to shift. + */ + struct dump_services_dests_ctx ctx = { + .idx_svc = 0, + .idx_dest = 0, + .start_svc = cb->args[0], + .start_dest = cb->args[1], + .last_svc = (struct ip_vs_service *)(cb->args[2]), + }; + struct net *net = sock_net(skb->sk); + struct netns_ipvs *ipvs = net_ipvs(net); + struct ip_vs_service *svc = NULL; + struct nlattr *attrs[IPVS_CMD_ATTR_MAX + 1]; + int i; + + mutex_lock(&__ip_vs_mutex); + + if (nlmsg_parse_deprecated(cb->nlh, GENL_HDRLEN, attrs, IPVS_CMD_ATTR_MAX, + ip_vs_cmd_policy, cb->extack) == 0) { + svc = ip_vs_genl_find_service(ipvs, attrs[IPVS_CMD_ATTR_SERVICE]); + + if (!IS_ERR_OR_NULL(svc)) { + ip_vs_genl_dump_service_destinations(skb, cb, svc, &ctx); + goto nla_put_failure; + } + } + + for (i = 0; i < IP_VS_SVC_TAB_SIZE; i++) { + hlist_for_each_entry(svc, &ip_vs_svc_table[i], s_list) { + if (svc->ipvs != ipvs) + continue; + if (ip_vs_genl_dump_service_destinations(skb, cb, svc, &ctx) < 0) + goto nla_put_failure; + } + } + + for (i = 0; i < IP_VS_SVC_TAB_SIZE; i++) { + hlist_for_each_entry(svc, &ip_vs_svc_fwm_table[i], s_list) { + if (svc->ipvs != ipvs) + continue; + if (ip_vs_genl_dump_service_destinations(skb, cb, svc, &ctx) < 0) + goto nla_put_failure; + } + } + +nla_put_failure: + mutex_unlock(&__ip_vs_mutex); + cb->args[0] = ctx.idx_svc; + cb->args[1] = ctx.idx_dest; + cb->args[2] = (long)ctx.last_svc; + + return skb->len; +} + static int ip_vs_genl_parse_dest(struct ip_vs_dest_user_kern *udest, struct nlattr *nla, bool full_entry) { @@ -3991,6 +4094,12 @@ static const struct genl_small_ops ip_vs_genl_ops[] = { .flags = GENL_ADMIN_PERM, .doit = ip_vs_genl_set_cmd, }, + { + .cmd = IPVS_CMD_GET_SERVICE_DEST, + .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP, + .flags = GENL_ADMIN_PERM, + .dumpit = ip_vs_genl_dump_services_destinations, + }, }; static struct genl_family ip_vs_genl_family __ro_after_init = {
A common operation for userspace applications managing ipvs is to dump all services and all destinations and then sort out what needs to be done. Previously this could only be accomplished by issuing 1 netlink IPVS_CMD_GET_SERVICE dump command followed by N IPVS_CMD_GET_DEST dump commands. For a dynamic system with a very large number of services this could be cause a performance impact. This patch introduces a new way of dumping all services and destinations with the new IPVS_CMD_GET_SERVICE_DEST command. A dump call for this command will send the services as IPVS_CMD_NEW_SERVICE messages imediatelly followed by its destinations as multiple IPVS_CMD_NEW_DEST messages. It's also possible to dump a single service and its destinations by sending a IPVS_CMD_ATTR_SERVICE argument to the dump command. Signed-off-by: Cezar Sa Espinola <cezarsa@gmail.com> --- To ensure that this patch improves performance I decided to also patch ipvsadm in order to run some benchmarks comparing 'ipvsadm -Sn' with the unpatched version. The ipvsadm patch is available on github in [1] for now but I intend to submit it if this RFC goes forward. The benchmarks look nice and detailed results and some scripts to allow reproducing then are available in another github repository [2]. The summary of the benchmarks is: svcs | dsts | run time compared to unpatched ----- | ---- | ------------------------------ 1000 | 4 | -60.63% 2000 | 2 | -71.10% 8000 | 2 | -52.83% 16000 | 1 | -54.13% 100 | 100 | -4.76% [1] - https://github.com/cezarsa/ipvsadm/compare/master...dump-svc-ds [2] - https://github.com/cezarsa/ipvsadm-validate#benchmark-results include/uapi/linux/ip_vs.h | 2 + net/netfilter/ipvs/ip_vs_ctl.c | 109 +++++++++++++++++++++++++++++++++ 2 files changed, 111 insertions(+)