Blog Original diff → Current Codebase Refactoring Summary

Original post: Linux Kernel Hacking for Non-Majors — Let’s Tame Jitter

This article summarizes how the patch from that post was refactored into a cleaner architecture: separate struct extraction, abstract accessor helper, and improved selection logic.

1. Core Structural Change: Direct dst_entry Extension → struct dst_power Separation

Before

Fields were tacked directly onto the end of struct dst_entry.

/* include/net/dst.h */
struct dst_entry {
    ...
    struct lwtunnel_state *lwtstate;
#endif
    /* USER ADDED */
    u64 ema_load;
    u64 ema_time_delta;
    u64 last_update_jiffies;
    unsigned int ema_k_factor;
    unsigned int power_cost_weight;
};

Problems

• sizeof(struct dst_entry) grows, risking cache-line / page-boundary crossing.
• Every path structure that embeds dst_entry (rtable, rt6_info, xfrm_dst, etc.) implicitly shares the same offset, creating ABI fragility.

After

struct dst_power is extracted into a standalone structure and embedded only where needed (IPv4/IPv6 route structures).

/* include/net/dst.h */
struct dst_power {
    u64 ema_load;
    u64 ema_time_delta;
    u64 last_update_jiffies;
    unsigned int ema_k_factor;
    unsigned int power_cost_weight;
};

struct dst_power *dst_power_ptr(struct dst_entry *dst);

/* include/net/route.h */
struct rtable {
    struct dst_entry    dst;
    ...
    struct dst_power    power;   /* ← embedded only for IPv4 routes */
};

/* include/net/ip6_fib.h */
struct rt6_info {
    struct dst_entry    dst;
    ...
    struct dst_power    power;   /* ← embedded only for IPv6 routes */
};

Benefits

• dst_entry size/layout stays intact → cache-line and ABI safety.
• dst-derived structures that do not need power metrics remain unaffected.
• Additional protocols (SCTP, XFRM, etc.) can opt-in by embedding struct dst_power.

2. Access Abstraction: Introducing dst_power_ptr()

Before

Every call site dereferenced struct dst_entry directly: dst->ema_load.

After

A family-type dispatch helper is centralized in net/core/dst.c.

/* net/core/dst.c */
struct dst_power *dst_power_ptr(struct dst_entry *dst)
{
    if (!dst || !dst->ops)
        return NULL;

#if IS_ENABLED(CONFIG_IPV6)
    if (dst->ops->family == AF_INET6) {
        struct rt6_info *rt = (struct rt6_info *)dst;
        return &rt->power;
    }
#endif
    if (dst->ops->family == AF_INET) {
        struct rtable *rt = (struct rtable *)dst;
        return &rt->power;
    }
    return NULL;
}
EXPORT_SYMBOL(dst_power_ptr);

Unified call sites

• net/core/dev.c — update_dst_ems_metrics()
• net/ipv4/fib_semantics.c — calculate_lowpower_weight()
• net/ipv6/route.c — calculate_lowpower_weight(), rt6_score_route()
• net/ipv4/route.c — rt_dst_alloc() initialization
• net/ipv6/route.c — ip6_dst_alloc() initialization
• net/xfrm/xfrm_policy.c — xfrm_bundle_create() initialization
• net/sctp/outqueue.c — calculate_lowpower_weight()

All locations now follow the uniform pattern: dst_power_ptr(dst) → NULL check → field access.

3. EMA Update Logic

The delta is small. Apart from obtaining the pointer via dst_power_ptr(), update_dst_ems_metrics() in net/core/dev.c (called on successful dev_hard_start_xmit()) is nearly identical to the original blog version.

static void update_dst_ems_metrics(struct dst_entry *dst, unsigned int tx_bytes)
{
    u64 cur_jiffies = get_jiffies_64();
    struct dst_power *p = dst_power_ptr(dst);
    u64 cur_load_rate;
    u64 delta_t;

    if (!p)
        return;

    delta_t = cur_jiffies - READ_ONCE(p->last_update_jiffies);
    if (!delta_t)
        return;

    cur_load_rate = tx_bytes / delta_t;

    WRITE_ONCE(p->ema_load,
               EMA_UPDATE(READ_ONCE(p->ema_k_factor),
                          READ_ONCE(p->ema_load), cur_load_rate));

    u64 diff = (cur_load_rate > READ_ONCE(p->ema_load))
        ? cur_load_rate - READ_ONCE(p->ema_load)
        : READ_ONCE(p->ema_load) - cur_load_rate;

    WRITE_ONCE(p->ema_time_delta,
               EMA_UPDATE(READ_ONCE(p->ema_k_factor),
                          READ_ONCE(p->ema_time_delta), diff));
    WRITE_ONCE(p->last_update_jiffies, cur_jiffies);
}

4. Multipath Selection Improvements (fib_select_multipath)

Before

Two-phase loop structure:

1. 1st loop: Iterate all nexthops, evaluate calculate_lowpower_weight(), pick the single nexthop with max_ema_weight.
2. 2nd loop: If all weights are zero, fall back to legacy fib_nh_upper_bound + hash-based ECMP.

Problem: Loops are duplicated, and lowpower policy operates independently from existing hash/saddr/upper_bound policy.

After

Single-loop integration. Introduces nh_score priority + tie-breaker concept.

void fib_select_multipath(struct fib_result *res, int hash,
                          const struct flowi4 *fl4)
{
    ...
    int score = -1;
    int lowpower_nh_index = -1;
    long max_ema_weight = -1;

    change_nexthops(fi) {
        int nh_score = 0;
        ...

        /* 1) Score existing policies */
        if (saddr && nexthop_nh->nh_saddr == saddr)
            nh_score += 2;
        if (hash <= nh_upper_bound)
            nh_score++;

        /* 2) lowpower weight acts as a tie-breaker */
        dst = get_dst_entry_from_nhc(&nexthop_nh->nh_common);
        current_weight = calculate_lowpower_weight(dst);

        if (nh_score > score ||
            (nh_score == score && current_weight > max_ema_weight)) {
            score = nh_score;
            max_ema_weight = current_weight;
            lowpower_nh_index = nhsel;

            /* 3) Early return on perfect match */
            if (nh_score == 3 || (!saddr && nh_score == 1)) {
                res->nh_sel = nhsel;
                res->nhc = &nexthop_nh->nh_common;
                return;
            }
        }
    } endfor_nexthops(fi);

    if (lowpower_nh_index != -1) {
        res->nh_sel = lowpower_nh_index;
        res->nhc = fib_info_nhc(fi, lowpower_nh_index);
        return;
    }
    ...
}

Summary of improvements

In short, the legacy ECMP policies (saddr affinity, hash bound) and lowpower metrics now coexist hierarchically without collision.

5. Unified Initialization Pattern

Previously, rtable, rt6_info, and xfrm_dst creators each poked fields directly. Now they all follow the same dst_power_ptr() template:

/* IPv4 example (net/ipv4/route.c) */
struct dst_power *p = dst_power_ptr(&rt->dst);
if (p) {
    WRITE_ONCE(p->ema_k_factor, ...);
    WRITE_ONCE(p->power_cost_weight, ...);
    WRITE_ONCE(p->ema_load, 0);
    WRITE_ONCE(p->ema_time_delta, 0);
    p->last_update_jiffies = 0;
}

The same template is reused in xfrm_policy.c and net/ipv6/route.c.

6. Removed vs. Preserved Parts

Removed changes

• netif_addr_lock_bh() simplification: The blog originally replaced local_bh_disable() + spin_lock_nested() with spin_lock_bh(). This change is not reflected in the current codebase (likely split into a separate cleanup commit or excluded from the experimental branch).

Preserved as-is

• EMA_UPDATE(K, Old, New) macro (top of net/core/dev.c)
• sysctl interfaces (lowpower_ema_k_factor, lowpower_power_cost_weight)
• Defaults (512, 100)
• IPv6 rt6_score_route() adding weight via m += weight
• Lowpower field initialization and weight function definitions in SCTP/XFRM

7. Overall Flow Diagram

Packet transmit (dev_hard_start_xmit)
    └── update_dst_ems_metrics(dst, skb->len)
            └── dst_power_ptr(dst) → struct dst_power *
            └── EMA_UPDATE(ema_load, ema_time_delta)

Multipath selection (fib_select_multipath / rt6_score_route)
    └── get_dst_entry_from_nhc / get_dst_entry_from_fib6_nh
    └── calculate_lowpower_weight(dst)
            └── dst_power_ptr(dst)
            └── (ema_load + ema_time_delta) * power_cost_weight
    └── Hierarchical combination with legacy policy (saddr, hash bound)

8. Conclusion

If the original patch was simply “glue fields onto the end of dst_entry and use them,” the current codebase achieves the following engineering-quality improvements:

1. Encapsulation: struct dst_power separation + dst_power_ptr() abstraction
2. ABI / cache safety: dst_entry size unchanged; embedded only where needed
3. Policy integration: fib_select_multipath improved to a single-loop, score-based design so legacy ECMP and lowpower metrics combine naturally
4. Templatization: Initialization / access code is unified around dst_power_ptr(), improving maintainability