[PATCH net-next v7 00/17] net: macb: implement context swapping

MACB has a pretty primitive approach to buffer management. They are all
stored in `struct macb *bp`. On operations that require buffer realloc
(set_ringparam & change_mtu at the moment), the only option is to close
the interface, change our global state and re-open the interface.

Two issues:
- It doesn't fly on memory pressured systems; we free our precious
  buffers and don't manage to reallocate fully, meaning our machine
  just lost its network access.
- Anecdotally, it is pretty slow because it implies a full PHY reinit.

Instead, we shall:
 - allocate a new context (including buffers) first
 - if it fails, early return without any impact to the interface
 - stop interface
 - update global state (bp, netdev, etc)
 - pass newly allocated buffer pointers to the hardware
 - start interface
 - free old context

This is what we implement here. Both .set_ringparam() and
.ndo_change_mtu() are covered by this series. In the future,
at least .set_channels() [0], XDP [1] and XSK [2] would benefit.

---
What your state of mind on this series? Past two iterations were about
fixing Sashiko edge-cases.

The most important fix was netconsole related and only lightly related
to this series. For the rest, it is nice that they are fixed but
overall I couldn't find a way to reproduce any of them, seeing how
unlikely they all are.

Can we get this in net-next? Goal being to work on XDP [1] & XSK [2]
during the next kernel cycle, bringing those long awaited features
into MACB.

Thanks,
Have a nice day,
Théo

[0]: https://lore.kernel.org/netdev/[email protected]/
[1]: https://lore.kernel.org/netdev/[email protected]/
[2]: https://lore.kernel.org/netdev/[email protected]/
[3]: https://lore.kernel.org/all/[email protected]/

---
Changes in v7:
- queue->tx_{head,tail} got renamed: fixup queue->tx_ptr_lock field
  comment and netdev_dbg() call in macb_start_xmit().
- Ensure bp->configured_{rx,tx}_ring_size have correct values in the
  EMAC case so that get_ringparam doesn't lie.
- Refuse set_ringparam on EMAC, which is invalid (sizes are hardcoded).
- Fix whitespace issue in "change function signatures to take contexts".
- Fix rebase mistake: macb_tx_unmap() signature change moved from
  patch "change function signatures to take contexts" to "change caps
  helpers signatures".
- Patch "move printk() calls out of bp->lock critical section":
  - Undrop netdev_vdbg() in macb_tx_error_task(); let's ignore the race
    for debug printk calls.
  - Move netdev_err() from macb_interrupt_misc() into macb_interrupt()
    so it can live outside the bp->lock critical section.
  - Add details to commit message.
- Patch "use context swapping in .set_ringparam()":
  - Re-schedule LPI after swap with delay=1s. Pick 1s so that if
    link-up -> swap instantly, we still respect the IEEE spec.
  - Add bp->link_up set under bp->mac_cfg_lock. netif_carrier_ok() bit
    is unsafe because not set inside a critical section we can enter.
  - Drop mac_cfg_lock from macb_hresp_error_task(); made useless by
    above.
  - Move macb_halt_tx() out of bp->lock to avoid 14ms latency spike
    possibility. Safe to do so.
- Patch "use context swapping in .ndo_change_mtu()":
  - Clarify commit message about EMAC.
- Rebase onto latest net-next/main (69963a0678a3), nothing to report.
- Link to v6: https://patch.msgid.link/[email protected]

Changes in v6:
- Non Sashiko based:
  - Take 3x Reviewed-By: Nicolai.
  - Silence checkpatch warnings in [02/15] about printk(..., __func__).
  - Rebase to latest net-next/main (a5c6ae8de11e), nothing to report.
- Fix netconsole/netpoll deadlock coming from macb_tx_error_task()
  calling printk() inside the bp->lock critical section.
- Move at91ether synchronize_irq() on close to protect more kfree().
- Clarify commit message for [11/15]: functions operating on the
  interface have no risk of using the wrong interface.
- On swap end, only re-enable the interface if !netif_carrier_ok().
- macb_hresp_error_task() grabs bp->mac_cfg_lock as it toggles the
  carrier state, to avoid race with context swap.
- On swap end, move netif_tx_start_all_queues() into bp->mac_cfg_lock.
- Link to v5: https://patch.msgid.link/[email protected]

Changes in v5:
- Fix build on [PATCH 02/15]; `bp->dev` got renamed to `bp->netdev`.
- Fix one word in commit message of [PATCH 07/15].
- Move bp->ctx->info assignment in at91ether_open() from [PATCH 12/15]
  to [PATCH 09/15].
- In [PATCH 14/15], reset DQL (dynamic queue limits) after tx disable,
  not before.
- Take 7 Reviewed-by: Nicolai (weirdly not all detected by b4).
- Rebase on latest net-next/main (89d8006259b8), nothing to report.
- Link to v4: https://patch.msgid.link/[email protected]

Changes in v4:
- Disable tx_error_task then disable NAPI because error task does a
  napi_disable() which deadlocks if NAPI is already disabled.
- Disable NAPI then disable tx_lpi_work, because NAPI might re-arm the
  latter.
- Last iteration did mask-irqs then disable-and-wait-for-bh then
  disable-hw. This is flawed because BH rearm IRQs once done. We don't
  have a good way to signal to them they shouldn't do so (we don't want
  to lock from NAPI context). So instead we add a bp->ctx_swap flag,
  shielded by bp->lock, to indicate to our IRQ handler to ignore IRQs
  and self-disarm.
- at91ether_close(): synchronize_irq() before freeing context.
- macb_interrupt(): drop double ISR read (outside & inside bp->lock).
- Rebase on latest net-next/main (f6f3b36c15ed). macb_free() since
  commit 27f575836cfe ("net: macb: drop in-flight Tx SKBs on close")
  needs access to queue stats => add them to macb_info.
- Take trailers.
- I did NOT init ctx->rx_ring_size/tx_ring_size/rx_buffer_size from
  at91ether_open(), as recommended by the netdev LLM. They are not the
  first fields in MACB that are present in both instances and that are
  uninitialised in one case. AT91 is almost a different driver. [3]
- Link to v3: https://patch.msgid.link/[email protected]

Changes in v3:
- Use `const struct macb_info *info` instead of bare `u32 caps` as
  helper arguments, for type safety.
- macb_interrupt(): the pre-lock readl(ISR) to detect spurious
  interrupts is only done if CLEAR_ON_WRITE.
- Don't forget allocating context in at91ether_open().
- swap:
  - Refuse swap for EMAC HW; it would crash because codepaths are so
    different.
  - Grab new bp->mac_cfg_lock to serialise with phylink MAC callbacks.
    We cannot rely on phydev->lock because it isn't present in the SFP
    or fixed-link cases. We also want to avoid phylink_stop() which
    triggers a slow PHY retrain.
- swap start:
  - We used to do disable-irqs-and-hw then drain-all-bh-features, but
    then HW might be raced against. Instead we disable-irqs then
    drain-all-bh then disable-hw which means at disable-hw step no BH
    context can be active.
  - Use macb_halt_tx() helper to properly stop HW.
  - Disable BH features before netif_tx_disable() call to avoid queue
    wakeup races.
  - Use macb_queue_isr_clear() helper instead of manual if-then-writel.
- swap end:
  - Grab bp->lock for the hardware reinit sequence composed of DMACFG
    and NCR writes.
  - Drop now useless EMAC check (we refuse EMAC HW before swapping).
- nits:
  - New patch to rename macb_{alloc,free}_consistent() which don't only
    allocate consistent buffers since a long time ago.
  - Fix the start_xmit verbose netdev_vdbg() format string from %hu to %u
    because the queue index type changed.
  - Strong commit reword from "unify `struct macb *` naming convention"
    to "unify variable naming convention in at91ether functions" which
    was underselling the changes.
- Rebase upon latest net-next/main (1c664ec4b9ea).
- Link to v2: https://patch.msgid.link/[email protected]

Changes in v2:
- Patch "add subset of `struct macb` to `struct macb_context`" was
  messed up. It contained much more than what the name implied. Split
  into three commits (I caused trouble by rebase reordering).
- Fix tieoff; V1 allocated it without initialisation.
- Fix NULL pointer dereference on context in mab_get_regs() and
  macb_get_ringparam() when interface is offline.
- Patch "unify device pointer naming convention":
  - Fix build issue when CONFIG_NETCONSOLE=y.
  - Rename `struct net_device *dev` to `netdev` in macb.h.
  - Rename `struct phy_device *phy` to `phydev` in macb_main.c.
- On swap, call netdev_tx_reset_queue() to reset all DQL counters.
- At end of swap, add missing kfree(old_ctx).
- During HW disabling in swap, grab bp->lock to protect against IRQ
  handler.
- On swap, cancel the three BH features MACB has:
  bp->hresp_err_bh_work, bp->tx_lpi_work and queue->tx_error_task.
- On swap, call macb_configure_dma() which writes buffer size to
  hardware registers. This is important because the change_mtu codepath
  changes the buffer size.
- Rebase onto latest net-next/main (58dd34dbd5b0) & resolve conflicts.
- Link to v1: https://patch.msgid.link/[email protected]

To: Théo Lebrun <[email protected]>
To: Conor Dooley <[email protected]>
To: Andrew Lunn <[email protected]>
To: "David S. Miller" <[email protected]>
To: Eric Dumazet <[email protected]>
To: Jakub Kicinski <[email protected]>
To: Paolo Abeni <[email protected]>
To: Richard Cochran <[email protected]>
To: Russell King <[email protected]>
Cc: [email protected]
Cc: [email protected]
Cc: Nicolas Ferre <[email protected]>
Cc: Claudiu Beznea <[email protected]>
Cc: Paolo Valerio <[email protected]>
Cc: Nicolai Buchwitz <[email protected]>
Cc: Vladimir Kondratiev <[email protected]>
Cc: Gregory CLEMENT <[email protected]>
Cc: Benoît Monin <[email protected]>
Cc: Tawfik Bayouk <[email protected]>
Cc: Thomas Petazzoni <[email protected]>
Cc: Maxime Chevallier <[email protected]>
Signed-off-by: Théo Lebrun <[email protected]>

---
Théo Lebrun (17):
      net: macb: drop "consistent" from alloc/free function names
      net: macb: unify device pointer naming convention
      net: macb: unify variable naming convention in at91ether functions
      net: macb: unify queue index variable naming convention and types
      net: macb: enforce reverse christmas tree (RCT) convention
      net: macb: allocate tieoff descriptor once across device lifetime
      net: macb: refuse set_ringparam on EMAC
      net: macb: introduce macb_context struct for buffer management
      net: macb: avoid macb_init_rx_buffer_size() modifying state
      net: macb: make `struct macb` subset reachable from macb_context struct
      net: macb: change caps helpers signatures
      net: macb: change function signatures to take contexts
      net: macb: introduce macb_context_alloc() helper
      net: macb: move printk() calls out of bp->lock critical section
      net: macb: read ISR inside bp->lock critical section
      net: macb: use context swapping in .set_ringparam()
      net: macb: use context swapping in .ndo_change_mtu()

 drivers/net/ethernet/cadence/macb.h      |  134 +-
 drivers/net/ethernet/cadence/macb_main.c | 1962 ++++++++++++++++++------------
 drivers/net/ethernet/cadence/macb_pci.c  |   46 +-
 drivers/net/ethernet/cadence/macb_ptp.c  |   26 +-
 4 files changed, 1292 insertions(+), 876 deletions(-)
---
base-commit: 492568fab54e8eb8ab0aa7928e6ea44505729854
change-id: 20260401-macb-context-bd0caf20414d

Best regards,
--  
Théo Lebrun <[email protected]>