bes2600: Patch C — direct-deliver RX SKBs from sdio_rx_work #3

Closed

claude-noether wants to merge 1 commits from bes2600/sdio-rx-direct-deliver into cleanups

Author	SHA1	Message	Date
test0r	c4797b1dbf	bes2600: deliver RX SKBs directly into wsm_handle_rx from sdio_rx_work Patch C — collapse the sdio_rx_work → rx_queue → bh_work relay into a direct call from bes2600_sdio_extract_packets into a new helper bes2600_bh_handle_rx_skb that performs the per-SKB bookkeeping previously done inside bes2600_bh_rx_helper after pipe_read. What this saves per RX frame: - two spinlock acquires on self->rx_queue->lock (skb_queue_tail in extract_packets + skb_dequeue in pipe_read) - one bh wait-queue wake-up per IRQ batch (sdio_rx_work no longer calls self->irq_handler) Pre-patch baseline on ohm (4 MB/s sender, ~5 min, srcversion 1B3B3ED0): - 387,532 RX packets, 578 MB, 1.36 MB/s observed receive - sdio_rx_work dispatched 34,994 times = 86.4/s = 90.3 per 1000 RX pkts - sdio_tx_work dispatched 111,770 times = 276.1/s - bes2600_bh_work redispatches: 0 (single long-lived work item, refutes earlier review claim of "9 events per frame") API contract for wsm_handle_rx (cited in bh.c block comment): - declared wsm.h:2108, defined wsm.c:2211, EXPORT_SYMBOL wsm.c:2463 - process context, sleepable - caller MUST hold no bes2600 spinlock; SDIO mutex is released at bes2600_sdio.c before extract_packets is called - SKB ownership: zeroes *skb_p if consumed, caller frees otherwise. bes2600_bh_handle_rx_skb frees on every path (success + error) After patch, bh thread retains responsibility for TX work. TX-confirm packets that release a TX buffer wake the bh thread via bes2600_bh_wakeup() inside bes2600_bh_handle_rx_skb (mirrors the in-bh tx=1 signaling the old bh_rx_helper used to do). Early-boot IRQs before fw_started are still served via the GPIO IRQ handler's fallback self->irq_handler path; that branch is unchanged. Minimum-diff scope: - struct sbus_priv->rx_queue + ->rx_queue_lock are still defined and initialised (allocated but never used after this patch). Removed in a follow-up hygiene patch, not bundled here. - bes2600_sdio_pipe_read still exists; returns NULL after this patch (rx_queue always empty). Removed in same follow-up. - bes2600_bh_rx_helper still exists and still calls pipe_read; it now always returns 0. Wasted 1 spinlock + skb_dequeue per bh wake; quantified in Phase 7 to decide whether worth a follow-up. Awaiting Phase 7 verification on ohm. Per Phase 4 plan §4.5 predicted delta: - rx_queue->lock acquire rate: 1914/s -> 0 (high confidence) - _raw_spin_unlock_irqrestore CPU%: 20% -> 12-15% (medium) - observed RX KB/s @ 4 MB/s: floor lifts toward >= 1 MB/s sustained (medium; rep 3's 75 KB/s death has multiple causes) Phase 7 will N=3 ramp 1 -> 2 -> 4 -> 8 MB/s with identical instrumentation pre/post. See notes/patch-c-phase4-plan-2026-05-07.md in marfrit/besser for the plan; this commit is the Phase 6 deliverable.	2026-05-07 19:49:57 +02:00

Author

SHA1

Message

Date

test0r

c4797b1dbf

bes2600: deliver RX SKBs directly into wsm_handle_rx from sdio_rx_work

Patch C — collapse the sdio_rx_work → rx_queue → bh_work relay into a
direct call from bes2600_sdio_extract_packets into a new helper
bes2600_bh_handle_rx_skb that performs the per-SKB bookkeeping
previously done inside bes2600_bh_rx_helper after pipe_read.

What this saves per RX frame:
  - two spinlock acquires on self->rx_queue->lock
    (skb_queue_tail in extract_packets + skb_dequeue in pipe_read)
  - one bh wait-queue wake-up per IRQ batch
    (sdio_rx_work no longer calls self->irq_handler)

Pre-patch baseline on ohm (4 MB/s sender, ~5 min, srcversion 1B3B3ED0):
  - 387,532 RX packets, 578 MB, 1.36 MB/s observed receive
  - sdio_rx_work dispatched 34,994 times = 86.4/s = 90.3 per 1000 RX pkts
  - sdio_tx_work dispatched 111,770 times = 276.1/s
  - bes2600_bh_work redispatches: 0  (single long-lived work item,
    refutes earlier review claim of "9 events per frame")

API contract for wsm_handle_rx (cited in bh.c block comment):
  - declared wsm.h:2108, defined wsm.c:2211, EXPORT_SYMBOL wsm.c:2463
  - process context, sleepable
  - caller MUST hold no bes2600 spinlock; SDIO mutex is released at
    bes2600_sdio.c before extract_packets is called
  - SKB ownership: zeroes *skb_p if consumed, caller frees otherwise.
    bes2600_bh_handle_rx_skb frees on every path (success + error)

After patch, bh thread retains responsibility for TX work.  TX-confirm
packets that release a TX buffer wake the bh thread via
bes2600_bh_wakeup() inside bes2600_bh_handle_rx_skb (mirrors the in-bh
tx=1 signaling the old bh_rx_helper used to do).  Early-boot IRQs
before fw_started are still served via the GPIO IRQ handler's fallback
self->irq_handler path; that branch is unchanged.

Minimum-diff scope:
  - struct sbus_priv->rx_queue + ->rx_queue_lock are still defined and
    initialised (allocated but never used after this patch).  Removed
    in a follow-up hygiene patch, not bundled here.
  - bes2600_sdio_pipe_read still exists; returns NULL after this patch
    (rx_queue always empty).  Removed in same follow-up.
  - bes2600_bh_rx_helper still exists and still calls pipe_read; it
    now always returns 0.  Wasted 1 spinlock + skb_dequeue per bh wake;
    quantified in Phase 7 to decide whether worth a follow-up.

Awaiting Phase 7 verification on ohm.  Per Phase 4 plan §4.5 predicted
delta:
  - rx_queue->lock acquire rate: 1914/s -> 0  (high confidence)
  - _raw_spin_unlock_irqrestore CPU%: 20% -> 12-15%  (medium)
  - observed RX KB/s @ 4 MB/s: floor lifts toward >= 1 MB/s sustained
    (medium; rep 3's 75 KB/s death has multiple causes)

Phase 7 will N=3 ramp 1 -> 2 -> 4 -> 8 MB/s with identical
instrumentation pre/post.  See notes/patch-c-phase4-plan-2026-05-07.md
in marfrit/besser for the plan; this commit is the Phase 6 deliverable.

2026-05-07 19:49:57 +02:00

bes2600: Patch C — direct-deliver RX SKBs from sdio_rx_work #3

1 Commits