|
Simon Horman a écrit :
> On Thu, Apr 23, 2009 at 06:23:32PM +0200, Fabien Duchêne wrote:
>> Hello Joe,
>>
>> Thanks for your reply!
>>
>> Okay for the -SH scheduler, but we have a script that dynamically change
>> the weight of the servers (because the are very different -hardware
>> speaking there-) and i think it wouldn't solve the problem (you'll see why).
>> It was very simple to test, a openned the LDAP port on a webserver with
>> a perl script (a server that just write "hello"), and then connected to
>> this port with a telnet client just after connecting to the web service
>> with my browser.
>> Result: i'm connected to the webserver (via the LDAP mark...).
>>
>> In fact, i looked into the code, and i think that LVS can't handle
>> multiple fwmark + persistence services (maybe we found a bug?).
>>
>> If you look in ip_vs.h (in the headers):
>>
>> static inline void ip_vs_addr_copy(int af, union nf_inet_addr *dst,
>> const union nf_inet_addr *src)
>> {
>> #ifdef CONFIG_IP_VS_IPV6
>> if (af == AF_INET6)
>> ipv6_addr_copy(&dst->in6, &src->in6);
>> else
>> #endif
>> dst->ip = src->ip;
>> }
>>
>> static inline int ip_vs_addr_equal(int af, const union nf_inet_addr *a,
>> const union nf_inet_addr *b)
>> {
>> #ifdef CONFIG_IP_VS_IPV6
>> if (af == AF_INET6)
>> return ipv6_addr_equal(&a->in6, &b->in6);
>> #endif
>>
>> return a->ip == b->ip;
>> }
>>
>> These functions are used the check if a template already exist.
>> In the fwmarked template, ->ip is always 0.0.0.0 and the ->all[3] (where
>> the fwmark is written) isn't tested (and not copied as you can see!).
>> So, the first template created by a "fwmark persistent service" will
>> match every fwmark persistent service (ip = 0.0.0.0, it's the same for
>> all!).
>>
>> Correct me if i'm wrong?
>>
>> If it's a bug, I hope a Dev' could fix this..
>
> Hi Fabien,
>
> that looks like a bug in the IPv4 to me too. If so, it would have been
> introduced when the IPv6 code was added to LVS, which was fairly recently.
>
> It seems to me that it should be easy enough to fix by changing
> fwmark in ip_vs_sched_persist() from:
>
> union nf_inet_addr fwmark = {
> .all = { 0, 0, 0, htonl(svc->fwmark) }
> };
>
> to:
>
> union nf_inet_addr fwmark = {
> .all = { htonl(svc->fwmark), 0, 0, 0 }
> };
>
> Assuming that this would result in fwmark->ip being set to
> htonl(svc->fwmark), which is relevant if svc->af is AF_INET - that is,
> for IPv4. IPv6 should work fine in both cases, as all 4 octets of
> all will be tested - which would have presumably been the focus
> of the IPv6 work that changed this code.
>
> This assums that the ip element of union nf_inet_addr always corresponds
> to the first octet of all.
>
> An alternate idea would be to change the af value used for fwmarks,
> but this seems to be even less clean than the current (slightly broken)
> technique of using nf_inet_addr for IPv4 or IPv6 addresses, or fwmarks.
>
> _______________________________________________
> Please read the documentation before posting - it's available at:
> http://www.linuxvirtualserver.org/
>
> LinuxVirtualServer.org mailing list - lvs-users@xxxxxxxxxxxxxxxxxxxxxx
> Send requests to lvs-users-request@xxxxxxxxxxxxxxxxxxxxxx
> or go to http://lists.graemef.net/mailman/listinfo/lvs-users
Hello Simon,
I made the same fix and tested it for 4 days, it seem to be stable.
union nf_inet_addr fwmark = {
.ip = htonl(svc->fwmark)
};
bwt, it's the same.
Any Dev' here can confirm that it's a bug and that it will be fixed?
Thanks!
Fabien
_______________________________________________
Please read the documentation before posting - it's available at:
http://www.linuxvirtualserver.org/
LinuxVirtualServer.org mailing list - lvs-users@xxxxxxxxxxxxxxxxxxxxxx
Send requests to lvs-users-request@xxxxxxxxxxxxxxxxxxxxxx
or go to http://lists.graemef.net/mailman/listinfo/lvs-users
|
