Re: [PATCH v3 00/20] Draft, incomplete series introducing state migration

[Stefano Brivio <sbrivio@redhat.com>]

On Fri, 31 Jan 2025 20:39:33 +0100
Stefano Brivio <sbrivio@redhat.com> wrote:

> What clearly needs changes:

Maybe something less elegant but actually functional like:

---
static int migrate_data_v1(int fd, struct ctx *c, bool target)
{
	struct iovec context[] { /* All these need to be network endian */
		c->hash_secret,		sizeof(c->hash_secret),
		c->ipv4.addr_seen,	sizeof(c->ipv4.addr_seen),
		c->ipv6.addr_seen,	sizeof(c->ipv6.addr_seen),
		c->ipv6.addr_ll_seen,	sizeof(c->ipv6.addr_ll_seen),
		/* Leave NDP's next_ra alone, start from 0 */
		&log_written /* Drop static */ sizeof(log_written),
	};
	int rc;

	if (target)
		rc = read_remainder(fd, context, ARRAY_SIZE(context), 0);
	else
		rc = write_remainder(fd, context, ARRAY_SIZE(context), 0);

	if (rc)
		return errno;

	if (target)
		return flow_migrate_data_target_v1(fd, c);
	return flow_migrate_data_source_v1(fd, c);
}
---

and:

---
struct tcp_flow_transfer_v1 {
	struct flow_common f;

	uint8_t retrans;

	uint16_t ws_from_tap; /* All BE */
	uint16_t ws_to_tap;
	uint16_t tap_mss;

	...
} __attribute__((packed));

union flow_transfer_v1 {
	struct tcp_flow_transfer_v1 tcp;
	...
};

int flow_migrate_data_source_v1(int fd, struct ctx *c)
{
	for_each_active_flow(flowtab) {
		union flow_transfer_v1 d;

		switch (flow->f.type) {
		case FLOW_TCP:
			d.tcp.f = flow->f;
			d.tcp.retrans = flow->tcp.retrans;
			d.tcp.ws_from_tap = htons(d.tcp.ws_from_tap);
			...;

			/* Fetch window stuff, socket must be in repair mode */

		send(...);
	}

int flow_migrate_data_target_v1(int fd, struct ctx *c)
{
	...
}

And I'm not sure about "[PATCH 0/2] Fancier version handling for
migration": perhaps we could switch to something radically easier
from the beginning. I mean, as we have to drop the declarative
approach altogether, at least let's minimise LoCs...

-- 
Stefano