|
On Sat, Oct 30, 2010 at 05:55:32PM +0300, Julian Anastasov wrote:
>
> Hello,
>
> On Fri, 29 Oct 2010, Hans Schillstrom wrote:
>
> >PATCH STATUS:
> >- Persistence data is not tested.
> >
> >
> >THANKS
> >The Persistence part is based on Simon Hormans Backup RFC
> >Julian for the comments and the Review of the RFC
> >
> >*v2
> >Simlified fwmark handling patch 1/4
> >New handling of optional data patch 2/4
> >Seconds in timeout
> >Basically all changes is based on Julians and Simons comments on the RFC
> >For details see individual patches.
>
> v2 PATCH 1/4
> OK
>
> v2 PATCH 2/4
> - Do we change the name from "Option" to "Parameter" ?
> Also, 'Option Length" will be part of the DATA, i.e.
> it should be present depending on type, in some cases
> it will be implicit, eg. for IPVS_OPT_SEQ_DATA:
>
> IPVS_OPT_SEQ_DATA:
> - 1 octet IPVS_OPT_SEQ_DATA followed by
> struct ip_vs_sync_conn_options
>
> IPVS_OPT_PE_DATA:
> - 1 octet IPVS_OPT_PE_DATA, 1 or 2 octets
> for Parameter Length (pe_data_len) followed by
> pe_data. If the decision is to support large
> data please use 2 octets and get_unaligned_be16
> to read it. Other PEs may need it, they can decide
> to add many things in PE data, they will not
> allocate other cp fields for their data.
>
> IPVS_OPT_PE_NAME:
> - 1 octet for IPVS_OPT_PE_NAME, 1 octet
> for Parameter Length (pe_name_len) followed by
> pe_name
>
> - struct ip_vs_sync_mesg_v2 with spaces instead of tabs?
>
> - can we add check in backup for timeout not to exceed
> (MAX_SCHEDULE_TIMEOUT / HZ)
>
> For example:
>
> if (timeout) {
> if (timeout > MAX_SCHEDULE_TIMEOUT / HZ)
> timeout = MAX_SCHEDULE_TIMEOUT / HZ;
> cp->timeout = timeout * HZ;
> } else ...
>
> - extra space in protocol:
> ip_vs_proc_conn(¶m, flags, state, s-> protocol, AF_INET,
>
> - lets drop the optional parameters from ip_vs_process_message()?
>
> - ip_vs_process_message: do not use mask: m2->size & SVER_MASK
> because the message size has no version.
>
> - What is the right thing to do in ip_vs_conn_fill_param_sync
> when ip_vs_pe_get() does not find PE? May be to drop
> connection? May be in ip_vs_process_message() we should
> use one pointer for end of message (it was 'p' before now),
> so that we can skip connections, for example, when some
> mandatory parameter as PE is not supported. We should not
> drop the whole sync message. When message is invalid it
> should be dropped but lack of support should not hurt
> other connections. In the case for PE may be
> ip_vs_conn_fill_param_sync() should return 1 if
> PE is unknown (ip_vs_pe_get). Then caller will continue
> with next connection if result > 0 or will drop message
> if result < 0 as in current patch. May be the pe_name
> presence should be the first check in
> ip_vs_conn_fill_param_sync.
Yes, I agree with this.
> Also note that p->pe is
> pointer without reference while called in ip_vs_sched_persist.
> In our case with ip_vs_conn_fill_param_sync we should
> put this reference after calling ip_vs_proc_conn().
> May be ip_vs_find_dest should check that p->pe matches
> svc->pe. ip_vs_try_bind_dest should provide p->pe too
> for this check. But we must somehow ignore new conns
> with PE if they don't have cp->dest (service) because
> it is risky to attach PE that is not held by svc.
> It is bad that the check for p->pe is before
> ip_vs_find_dest.
I have a patch, "IPVS: Add persistence engine to connection entry"
that attaches the pe to the connection rather than the dest.
In this case, if pe is NULL, then the connection doesn't use
pe, which is fine. If its non-NULL, its there so there is no
problem in that case either.
I think this resolves the issue that you raise.
>
> Example for parsing:
>
> - rename msgEnd/p to 'char *endp;' and move it before loop.
> It will have the 'p' role as before: to point to end
> of connection.
>
> before loop:
> endp = buffer + sizeof(struct ip_vs_sync_mesg_v2);
>
> In loop:
>
> + p = endp;
> + s = (union ip_vs_sync_conn *) endp;
> + size = ntohs(s->v4.ver_size) & SVER_MASK;
> + endp = p + size; <--- we are ready for 'continue'
>
> By this way endp should point to next connection in case
> we want to ignore current one at any time. 'p' will walk
> parameters as you do now. After that all checks for p should be
> with endp, I mean use 'if (p > endp)' here:
>
> + if (p > buffer+buflen) {
> + IP_VS_ERR_RL("bogus conn in sync message\n");
> + return;
> + }
>
> - Note that pe_data is leaked when ip_vs_proc_conn() fails
> to create connection. May be PE info for non-templates
> should be ignored?
Yes I think it should be ignored in that case.
> And we need a way to know if
> ip_vs_proc_conn called ip_vs_conn_new at all and that
> it succeeded so that we can safely free pe_data.
Perhaps ip_vs_proc_conn() could just free pe_data if
it is present but ip_vs_conn_new() is not called.
>
> Simon should tell what happens if some PE updates
> ct->pe_data, may be we should replace it too for the
> case when ip_vs_proc_conn works with existing template?
At this stage there is no facility for updating PE data.
But I think replacing it should be fine, so long as
the appropriate locking is in place.
> v2 PATCH 4/4
> May be we should add configuration option if sync is enabled
> to default to version 0 because how we solve the problem if
> backup can not be upgraded?
I think that defaulting to v1 and having an option
to move back to v0 would make more sense. Initially
I would expect people to need to use v0 for the reason
that you describe. But once the backup has been upgraded then,
bugs not withstanding, there shouldn't be a reason not to use v1
moving forward.
> For IPVS_OPT_PE_NAME: because we have length better not to
> add zero terminator in sync message. It complicates the
> checks in backup. Or backup prefers to check with zero
> terminated string directly from message? In this case we
> should check for existing terminator.
I have no preference with regards to this.
--
To unsubscribe from this list: send the line "unsubscribe lvs-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
|
