* [PATCH net-next V2 1/4] net/mlx5e: Move async ICOSQ lock into ICOSQ struct
2026-01-14 7:46 [PATCH net-next V2 0/4] net/mlx5e: Save per-channel async ICOSQ in default Tariq Toukan
@ 2026-01-14 7:46 ` Tariq Toukan
2026-01-14 7:46 ` [PATCH net-next V2 2/4] net/mlx5e: Use regular ICOSQ for triggering NAPI Tariq Toukan
` (3 subsequent siblings)
4 siblings, 0 replies; 6+ messages in thread
From: Tariq Toukan @ 2026-01-14 7:46 UTC (permalink / raw)
To: Eric Dumazet, Jakub Kicinski, Paolo Abeni, Andrew Lunn,
David S. Miller
Cc: Saeed Mahameed, Leon Romanovsky, Tariq Toukan, Mark Bloch, netdev,
linux-rdma, linux-kernel, Gal Pressman, William Tu, toke
From: William Tu <witu@nvidia.com>
Move the async_icosq spinlock from the mlx5e_channel structure into
the mlx5e_icosq structure itself for better encapsulation and for
later patch to also use it for other icosq use cases.
Changes:
- Add spinlock_t lock field to struct mlx5e_icosq
- Remove async_icosq_lock field from struct mlx5e_channel
- Initialize the new lock in mlx5e_open_icosq()
- Update all lock usage in ktls_rx.c and en_main.c to use sq->lock
instead of c->async_icosq_lock
Signed-off-by: William Tu <witu@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
drivers/net/ethernet/mellanox/mlx5/core/en.h | 4 ++--
.../mellanox/mlx5/core/en_accel/ktls_rx.c | 18 +++++++++---------
.../net/ethernet/mellanox/mlx5/core/en_main.c | 12 +++++++-----
3 files changed, 18 insertions(+), 16 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en.h b/drivers/net/ethernet/mellanox/mlx5/core/en.h
index 262dc032e276..ebd3b90e17fd 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en.h
@@ -545,6 +545,8 @@ struct mlx5e_icosq {
u32 sqn;
u16 reserved_room;
unsigned long state;
+ /* icosq can be accessed from any CPU - the spinlock protects it. */
+ spinlock_t lock;
struct mlx5e_ktls_resync_resp *ktls_resync;
/* control path */
@@ -777,8 +779,6 @@ struct mlx5e_channel {
/* Async ICOSQ */
struct mlx5e_icosq async_icosq;
- /* async_icosq can be accessed from any CPU - the spinlock protects it. */
- spinlock_t async_icosq_lock;
/* data path - accessed per napi poll */
const struct cpumask *aff_mask;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_rx.c b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_rx.c
index da2d1eb52c13..8bc8231f521f 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_rx.c
@@ -203,7 +203,7 @@ static int post_rx_param_wqes(struct mlx5e_channel *c,
err = 0;
sq = &c->async_icosq;
- spin_lock_bh(&c->async_icosq_lock);
+ spin_lock_bh(&sq->lock);
cseg = post_static_params(sq, priv_rx);
if (IS_ERR(cseg))
@@ -214,7 +214,7 @@ static int post_rx_param_wqes(struct mlx5e_channel *c,
mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, cseg);
unlock:
- spin_unlock_bh(&c->async_icosq_lock);
+ spin_unlock_bh(&sq->lock);
return err;
@@ -277,10 +277,10 @@ resync_post_get_progress_params(struct mlx5e_icosq *sq,
buf->priv_rx = priv_rx;
- spin_lock_bh(&sq->channel->async_icosq_lock);
+ spin_lock_bh(&sq->lock);
if (unlikely(!mlx5e_icosq_can_post_wqe(sq, MLX5E_KTLS_GET_PROGRESS_WQEBBS))) {
- spin_unlock_bh(&sq->channel->async_icosq_lock);
+ spin_unlock_bh(&sq->lock);
err = -ENOSPC;
goto err_dma_unmap;
}
@@ -311,7 +311,7 @@ resync_post_get_progress_params(struct mlx5e_icosq *sq,
icosq_fill_wi(sq, pi, &wi);
sq->pc++;
mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, cseg);
- spin_unlock_bh(&sq->channel->async_icosq_lock);
+ spin_unlock_bh(&sq->lock);
return 0;
@@ -413,9 +413,9 @@ static void resync_handle_seq_match(struct mlx5e_ktls_offload_context_rx *priv_r
return;
if (!napi_if_scheduled_mark_missed(&c->napi)) {
- spin_lock_bh(&c->async_icosq_lock);
+ spin_lock_bh(&sq->lock);
mlx5e_trigger_irq(sq);
- spin_unlock_bh(&c->async_icosq_lock);
+ spin_unlock_bh(&sq->lock);
}
}
@@ -772,7 +772,7 @@ bool mlx5e_ktls_rx_handle_resync_list(struct mlx5e_channel *c, int budget)
clear_bit(MLX5E_SQ_STATE_PENDING_TLS_RX_RESYNC, &sq->state);
spin_unlock(&ktls_resync->lock);
- spin_lock(&c->async_icosq_lock);
+ spin_lock(&sq->lock);
for (j = 0; j < i; j++) {
struct mlx5_wqe_ctrl_seg *cseg;
@@ -791,7 +791,7 @@ bool mlx5e_ktls_rx_handle_resync_list(struct mlx5e_channel *c, int budget)
}
if (db_cseg)
mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, db_cseg);
- spin_unlock(&c->async_icosq_lock);
+ spin_unlock(&sq->lock);
priv_rx->rq_stats->tls_resync_res_ok += j;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
index 3ac47df83ac8..e666d9cc1817 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
@@ -2075,6 +2075,8 @@ static int mlx5e_open_icosq(struct mlx5e_channel *c, struct mlx5e_params *params
if (err)
goto err_free_icosq;
+ spin_lock_init(&sq->lock);
+
if (param->is_tls) {
sq->ktls_resync = mlx5e_ktls_rx_resync_create_resp_list();
if (IS_ERR(sq->ktls_resync)) {
@@ -2630,8 +2632,6 @@ static int mlx5e_open_queues(struct mlx5e_channel *c,
if (err)
goto err_close_rx_cq;
- spin_lock_init(&c->async_icosq_lock);
-
err = mlx5e_open_icosq(c, params, &cparam->async_icosq, &c->async_icosq,
mlx5e_async_icosq_err_cqe_work);
if (err)
@@ -2750,9 +2750,11 @@ static int mlx5e_channel_stats_alloc(struct mlx5e_priv *priv, int ix, int cpu)
void mlx5e_trigger_napi_icosq(struct mlx5e_channel *c)
{
- spin_lock_bh(&c->async_icosq_lock);
- mlx5e_trigger_irq(&c->async_icosq);
- spin_unlock_bh(&c->async_icosq_lock);
+ struct mlx5e_icosq *async_icosq = &c->async_icosq;
+
+ spin_lock_bh(&async_icosq->lock);
+ mlx5e_trigger_irq(async_icosq);
+ spin_unlock_bh(&async_icosq->lock);
}
void mlx5e_trigger_napi_sched(struct napi_struct *napi)
--
2.31.1
^ permalink raw reply related [flat|nested] 6+ messages in thread* [PATCH net-next V2 2/4] net/mlx5e: Use regular ICOSQ for triggering NAPI
2026-01-14 7:46 [PATCH net-next V2 0/4] net/mlx5e: Save per-channel async ICOSQ in default Tariq Toukan
2026-01-14 7:46 ` [PATCH net-next V2 1/4] net/mlx5e: Move async ICOSQ lock into ICOSQ struct Tariq Toukan
@ 2026-01-14 7:46 ` Tariq Toukan
2026-01-14 7:46 ` [PATCH net-next V2 3/4] net/mlx5e: Move async ICOSQ to dynamic allocation Tariq Toukan
` (2 subsequent siblings)
4 siblings, 0 replies; 6+ messages in thread
From: Tariq Toukan @ 2026-01-14 7:46 UTC (permalink / raw)
To: Eric Dumazet, Jakub Kicinski, Paolo Abeni, Andrew Lunn,
David S. Miller
Cc: Saeed Mahameed, Leon Romanovsky, Tariq Toukan, Mark Bloch, netdev,
linux-rdma, linux-kernel, Gal Pressman, William Tu, toke
From: William Tu <witu@nvidia.com>
Before the cited commit, ICOSQ is used to post NOP WQE to trigger
hardware interrupt and start NAPI, but this mechanism suffers from
a race condition: mlx5e_alloc_rx_mpwqe may post UMR WQEs to ICOSQ
_before_ NOP WQE is posted. The cited commit fixes the issue by
replacing ICOSQ with async ICOSQ, as a new way to post the NOP WQE
to trigger the hardware interrupt and NAPI.
The patch changes it back by replacing async ICOSQ with regular
ICOSQ, for the purpose of saving memory in later patches, and solves
the issue by adding a new SQ state, MLX5E_SQ_STATE_LOCK_NEEDED
for syncing the start of NAPI.
What it does:
- Switch trigger path from async ICOSQ to regular ICOSQ to reduce
need for async SQ.
- Introduce MLX5E_SQ_STATE_LOCK_NEEDED and mlx5e_icosq_sync_lock(),
unlock() to prevent the race where UMR WQEs could be posted before
the NOP WQE used to trigger NAPI.
- Use synchronize_net() once per trigger cycle to quiesce in-flight
softirqs before serializing the NOP WQE and any UMR postings via
the ICOSQ lock.
- Wrap ICOSQ UMR posting in en_rx.c and xsk/rx.c with the new
conditional lock.
The conditional locking approach is critical for performance: always
locking would impose unnecessary overhead. Synchronization is not needed
between regular NAPI cycles once the channel is activated and running.
The lock is only required to protect against the race during channel
activation—specifically, when the very first NOP WQE is posted to trigger
NAPI. After that initial trigger, normal NAPI polling handles subsequent
work without contention. The MLX5E_SQ_STATE_LOCK_NEEDED flag ensures we
pay the synchronization cost only when necessary.
Signed-off-by: William Tu <witu@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
drivers/net/ethernet/mellanox/mlx5/core/en.h | 21 ++++++++++++++++++-
.../mellanox/mlx5/core/en/reporter_tx.c | 1 +
.../ethernet/mellanox/mlx5/core/en/xsk/rx.c | 3 +++
.../net/ethernet/mellanox/mlx5/core/en_main.c | 13 ++++++++----
.../net/ethernet/mellanox/mlx5/core/en_rx.c | 4 ++++
5 files changed, 37 insertions(+), 5 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en.h b/drivers/net/ethernet/mellanox/mlx5/core/en.h
index ebd3b90e17fd..83cfa3983855 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en.h
@@ -388,6 +388,7 @@ enum {
MLX5E_SQ_STATE_DIM,
MLX5E_SQ_STATE_PENDING_XSK_TX,
MLX5E_SQ_STATE_PENDING_TLS_RX_RESYNC,
+ MLX5E_SQ_STATE_LOCK_NEEDED,
MLX5E_NUM_SQ_STATES, /* Must be kept last */
};
@@ -545,7 +546,10 @@ struct mlx5e_icosq {
u32 sqn;
u16 reserved_room;
unsigned long state;
- /* icosq can be accessed from any CPU - the spinlock protects it. */
+ /* icosq can be accessed from any CPU and from different contexts
+ * (NAPI softirq or process/workqueue). Always use spin_lock_bh for
+ * simplicity and correctness across all contexts.
+ */
spinlock_t lock;
struct mlx5e_ktls_resync_resp *ktls_resync;
@@ -801,6 +805,21 @@ struct mlx5e_channel {
struct dim_cq_moder tx_cq_moder;
};
+static inline bool mlx5e_icosq_sync_lock(struct mlx5e_icosq *sq)
+{
+ if (likely(!test_bit(MLX5E_SQ_STATE_LOCK_NEEDED, &sq->state)))
+ return false;
+
+ spin_lock_bh(&sq->lock);
+ return true;
+}
+
+static inline void mlx5e_icosq_sync_unlock(struct mlx5e_icosq *sq, bool locked)
+{
+ if (unlikely(locked))
+ spin_unlock_bh(&sq->lock);
+}
+
struct mlx5e_ptp;
struct mlx5e_channels {
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/reporter_tx.c b/drivers/net/ethernet/mellanox/mlx5/core/en/reporter_tx.c
index 9e2cf191ed30..4adc1adf9897 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/reporter_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/reporter_tx.c
@@ -15,6 +15,7 @@ static const char * const sq_sw_state_type_name[] = {
[MLX5E_SQ_STATE_DIM] = "dim",
[MLX5E_SQ_STATE_PENDING_XSK_TX] = "pending_xsk_tx",
[MLX5E_SQ_STATE_PENDING_TLS_RX_RESYNC] = "pending_tls_rx_resync",
+ [MLX5E_SQ_STATE_LOCK_NEEDED] = "lock_needed",
};
static int mlx5e_wait_for_sq_flush(struct mlx5e_txqsq *sq)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
index 2b05536d564a..4f984f6a2cb9 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
@@ -23,6 +23,7 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
struct mlx5_wq_cyc *wq = &icosq->wq;
struct mlx5e_umr_wqe *umr_wqe;
struct xdp_buff **xsk_buffs;
+ bool sync_locked;
int batch, i;
u32 offset; /* 17-bit value with MTT. */
u16 pi;
@@ -47,6 +48,7 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
goto err_reuse_batch;
}
+ sync_locked = mlx5e_icosq_sync_lock(icosq);
pi = mlx5e_icosq_get_next_pi(icosq, rq->mpwqe.umr_wqebbs);
umr_wqe = mlx5_wq_cyc_get_wqe(wq, pi);
memcpy(umr_wqe, &rq->mpwqe.umr_wqe, sizeof(struct mlx5e_umr_wqe));
@@ -143,6 +145,7 @@ int mlx5e_xsk_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
};
icosq->pc += rq->mpwqe.umr_wqebbs;
+ mlx5e_icosq_sync_unlock(icosq, sync_locked);
icosq->doorbell_cseg = &umr_wqe->hdr.ctrl;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
index e666d9cc1817..fdbcc22b6c61 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
@@ -2750,11 +2750,16 @@ static int mlx5e_channel_stats_alloc(struct mlx5e_priv *priv, int ix, int cpu)
void mlx5e_trigger_napi_icosq(struct mlx5e_channel *c)
{
- struct mlx5e_icosq *async_icosq = &c->async_icosq;
+ bool locked;
- spin_lock_bh(&async_icosq->lock);
- mlx5e_trigger_irq(async_icosq);
- spin_unlock_bh(&async_icosq->lock);
+ if (!test_and_set_bit(MLX5E_SQ_STATE_LOCK_NEEDED, &c->icosq.state))
+ synchronize_net();
+
+ locked = mlx5e_icosq_sync_lock(&c->icosq);
+ mlx5e_trigger_irq(&c->icosq);
+ mlx5e_icosq_sync_unlock(&c->icosq, locked);
+
+ clear_bit(MLX5E_SQ_STATE_LOCK_NEEDED, &c->icosq.state);
}
void mlx5e_trigger_napi_sched(struct napi_struct *napi)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c b/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c
index 1f6930c77437..1fc3720d2201 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c
@@ -778,6 +778,7 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
struct mlx5_wq_cyc *wq = &sq->wq;
struct mlx5e_umr_wqe *umr_wqe;
u32 offset; /* 17-bit value with MTT. */
+ bool sync_locked;
u16 pi;
int err;
int i;
@@ -788,6 +789,7 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
goto err;
}
+ sync_locked = mlx5e_icosq_sync_lock(sq);
pi = mlx5e_icosq_get_next_pi(sq, rq->mpwqe.umr_wqebbs);
umr_wqe = mlx5_wq_cyc_get_wqe(wq, pi);
memcpy(umr_wqe, &rq->mpwqe.umr_wqe, sizeof(struct mlx5e_umr_wqe));
@@ -835,12 +837,14 @@ static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
};
sq->pc += rq->mpwqe.umr_wqebbs;
+ mlx5e_icosq_sync_unlock(sq, sync_locked);
sq->doorbell_cseg = &umr_wqe->hdr.ctrl;
return 0;
err_unmap:
+ mlx5e_icosq_sync_unlock(sq, sync_locked);
while (--i >= 0) {
frag_page--;
mlx5e_page_release_fragmented(rq->page_pool, frag_page);
--
2.31.1
^ permalink raw reply related [flat|nested] 6+ messages in thread* [PATCH net-next V2 3/4] net/mlx5e: Move async ICOSQ to dynamic allocation
2026-01-14 7:46 [PATCH net-next V2 0/4] net/mlx5e: Save per-channel async ICOSQ in default Tariq Toukan
2026-01-14 7:46 ` [PATCH net-next V2 1/4] net/mlx5e: Move async ICOSQ lock into ICOSQ struct Tariq Toukan
2026-01-14 7:46 ` [PATCH net-next V2 2/4] net/mlx5e: Use regular ICOSQ for triggering NAPI Tariq Toukan
@ 2026-01-14 7:46 ` Tariq Toukan
2026-01-14 7:46 ` [PATCH net-next V2 4/4] net/mlx5e: Conditionally create async ICOSQ Tariq Toukan
2026-01-19 21:00 ` [PATCH net-next V2 0/4] net/mlx5e: Save per-channel async ICOSQ in default patchwork-bot+netdevbpf
4 siblings, 0 replies; 6+ messages in thread
From: Tariq Toukan @ 2026-01-14 7:46 UTC (permalink / raw)
To: Eric Dumazet, Jakub Kicinski, Paolo Abeni, Andrew Lunn,
David S. Miller
Cc: Saeed Mahameed, Leon Romanovsky, Tariq Toukan, Mark Bloch, netdev,
linux-rdma, linux-kernel, Gal Pressman, William Tu, toke
From: William Tu <witu@nvidia.com>
Dynamically allocate async ICOSQ. ICO (Internal Communication
Operations) is for driver to communicate with the HW, and it's
not used for traffic. Currently mlx5 driver has sync and async
ICO send queues. The async ICOSQ means that it's not necessarily
under NAPI context protection. The patch is in preparation for
the later patch to detect its usage and enable it when necessary.
Signed-off-by: William Tu <witu@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
drivers/net/ethernet/mellanox/mlx5/core/en.h | 2 +-
.../ethernet/mellanox/mlx5/core/en/xsk/tx.c | 6 +-
.../mellanox/mlx5/core/en_accel/ktls_rx.c | 8 +--
.../mellanox/mlx5/core/en_accel/ktls_txrx.h | 3 +-
.../net/ethernet/mellanox/mlx5/core/en_main.c | 67 ++++++++++++++-----
.../net/ethernet/mellanox/mlx5/core/en_txrx.c | 7 +-
6 files changed, 65 insertions(+), 28 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en.h b/drivers/net/ethernet/mellanox/mlx5/core/en.h
index 83cfa3983855..a7076b26fd5c 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en.h
@@ -782,7 +782,7 @@ struct mlx5e_channel {
struct mlx5e_xdpsq xsksq;
/* Async ICOSQ */
- struct mlx5e_icosq async_icosq;
+ struct mlx5e_icosq *async_icosq;
/* data path - accessed per napi poll */
const struct cpumask *aff_mask;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.c
index a59199ed590d..9e33156fac8a 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.c
@@ -26,10 +26,12 @@ int mlx5e_xsk_wakeup(struct net_device *dev, u32 qid, u32 flags)
* active and not polled by NAPI. Return 0, because the upcoming
* activate will trigger the IRQ for us.
*/
- if (unlikely(!test_bit(MLX5E_SQ_STATE_ENABLED, &c->async_icosq.state)))
+ if (unlikely(!test_bit(MLX5E_SQ_STATE_ENABLED,
+ &c->async_icosq->state)))
return 0;
- if (test_and_set_bit(MLX5E_SQ_STATE_PENDING_XSK_TX, &c->async_icosq.state))
+ if (test_and_set_bit(MLX5E_SQ_STATE_PENDING_XSK_TX,
+ &c->async_icosq->state))
return 0;
mlx5e_trigger_napi_icosq(c);
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_rx.c b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_rx.c
index 8bc8231f521f..5d8fe252799e 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_rx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_rx.c
@@ -202,7 +202,7 @@ static int post_rx_param_wqes(struct mlx5e_channel *c,
int err;
err = 0;
- sq = &c->async_icosq;
+ sq = c->async_icosq;
spin_lock_bh(&sq->lock);
cseg = post_static_params(sq, priv_rx);
@@ -344,7 +344,7 @@ static void resync_handle_work(struct work_struct *work)
}
c = resync->priv->channels.c[priv_rx->rxq];
- sq = &c->async_icosq;
+ sq = c->async_icosq;
if (resync_post_get_progress_params(sq, priv_rx)) {
priv_rx->rq_stats->tls_resync_req_skip++;
@@ -371,7 +371,7 @@ static void resync_handle_seq_match(struct mlx5e_ktls_offload_context_rx *priv_r
struct mlx5e_icosq *sq;
bool trigger_poll;
- sq = &c->async_icosq;
+ sq = c->async_icosq;
ktls_resync = sq->ktls_resync;
trigger_poll = false;
@@ -753,7 +753,7 @@ bool mlx5e_ktls_rx_handle_resync_list(struct mlx5e_channel *c, int budget)
LIST_HEAD(local_list);
int i, j;
- sq = &c->async_icosq;
+ sq = c->async_icosq;
if (unlikely(!test_bit(MLX5E_SQ_STATE_ENABLED, &sq->state)))
return false;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_txrx.h b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_txrx.h
index cb08799769ee..4022c7e78a2e 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_txrx.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_txrx.h
@@ -50,7 +50,8 @@ bool mlx5e_ktls_rx_handle_resync_list(struct mlx5e_channel *c, int budget);
static inline bool
mlx5e_ktls_rx_pending_resync_list(struct mlx5e_channel *c, int budget)
{
- return budget && test_bit(MLX5E_SQ_STATE_PENDING_TLS_RX_RESYNC, &c->async_icosq.state);
+ return budget && test_bit(MLX5E_SQ_STATE_PENDING_TLS_RX_RESYNC,
+ &c->async_icosq->state);
}
static inline void
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
index fdbcc22b6c61..aa4ff3963b86 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
@@ -2589,6 +2589,47 @@ static int mlx5e_open_rxq_rq(struct mlx5e_channel *c, struct mlx5e_params *param
return mlx5e_open_rq(params, rq_params, NULL, cpu_to_node(c->cpu), q_counter, &c->rq);
}
+static struct mlx5e_icosq *
+mlx5e_open_async_icosq(struct mlx5e_channel *c,
+ struct mlx5e_params *params,
+ struct mlx5e_channel_param *cparam,
+ struct mlx5e_create_cq_param *ccp)
+{
+ struct dim_cq_moder icocq_moder = {0, 0};
+ struct mlx5e_icosq *async_icosq;
+ int err;
+
+ async_icosq = kvzalloc_node(sizeof(*async_icosq), GFP_KERNEL,
+ cpu_to_node(c->cpu));
+ if (!async_icosq)
+ return ERR_PTR(-ENOMEM);
+
+ err = mlx5e_open_cq(c->mdev, icocq_moder, &cparam->async_icosq.cqp, ccp,
+ &async_icosq->cq);
+ if (err)
+ goto err_free_async_icosq;
+
+ err = mlx5e_open_icosq(c, params, &cparam->async_icosq, async_icosq,
+ mlx5e_async_icosq_err_cqe_work);
+ if (err)
+ goto err_close_async_icosq_cq;
+
+ return async_icosq;
+
+err_close_async_icosq_cq:
+ mlx5e_close_cq(&async_icosq->cq);
+err_free_async_icosq:
+ kvfree(async_icosq);
+ return ERR_PTR(err);
+}
+
+static void mlx5e_close_async_icosq(struct mlx5e_icosq *async_icosq)
+{
+ mlx5e_close_icosq(async_icosq);
+ mlx5e_close_cq(&async_icosq->cq);
+ kvfree(async_icosq);
+}
+
static int mlx5e_open_queues(struct mlx5e_channel *c,
struct mlx5e_params *params,
struct mlx5e_channel_param *cparam)
@@ -2600,15 +2641,10 @@ static int mlx5e_open_queues(struct mlx5e_channel *c,
mlx5e_build_create_cq_param(&ccp, c);
- err = mlx5e_open_cq(c->mdev, icocq_moder, &cparam->async_icosq.cqp, &ccp,
- &c->async_icosq.cq);
- if (err)
- return err;
-
err = mlx5e_open_cq(c->mdev, icocq_moder, &cparam->icosq.cqp, &ccp,
&c->icosq.cq);
if (err)
- goto err_close_async_icosq_cq;
+ return err;
err = mlx5e_open_tx_cqs(c, params, &ccp, cparam);
if (err)
@@ -2632,10 +2668,11 @@ static int mlx5e_open_queues(struct mlx5e_channel *c,
if (err)
goto err_close_rx_cq;
- err = mlx5e_open_icosq(c, params, &cparam->async_icosq, &c->async_icosq,
- mlx5e_async_icosq_err_cqe_work);
- if (err)
+ c->async_icosq = mlx5e_open_async_icosq(c, params, cparam, &ccp);
+ if (IS_ERR(c->async_icosq)) {
+ err = PTR_ERR(c->async_icosq);
goto err_close_rq_xdpsq_cq;
+ }
mutex_init(&c->icosq_recovery_lock);
@@ -2671,7 +2708,7 @@ static int mlx5e_open_queues(struct mlx5e_channel *c,
mlx5e_close_icosq(&c->icosq);
err_close_async_icosq:
- mlx5e_close_icosq(&c->async_icosq);
+ mlx5e_close_async_icosq(c->async_icosq);
err_close_rq_xdpsq_cq:
if (c->xdp)
@@ -2690,9 +2727,6 @@ static int mlx5e_open_queues(struct mlx5e_channel *c,
err_close_icosq_cq:
mlx5e_close_cq(&c->icosq.cq);
-err_close_async_icosq_cq:
- mlx5e_close_cq(&c->async_icosq.cq);
-
return err;
}
@@ -2706,7 +2740,7 @@ static void mlx5e_close_queues(struct mlx5e_channel *c)
mlx5e_close_sqs(c);
mlx5e_close_icosq(&c->icosq);
mutex_destroy(&c->icosq_recovery_lock);
- mlx5e_close_icosq(&c->async_icosq);
+ mlx5e_close_async_icosq(c->async_icosq);
if (c->xdp)
mlx5e_close_cq(&c->rq_xdpsq.cq);
mlx5e_close_cq(&c->rq.cq);
@@ -2714,7 +2748,6 @@ static void mlx5e_close_queues(struct mlx5e_channel *c)
mlx5e_close_xdpredirect_sq(c->xdpsq);
mlx5e_close_tx_cqs(c);
mlx5e_close_cq(&c->icosq.cq);
- mlx5e_close_cq(&c->async_icosq.cq);
}
static u8 mlx5e_enumerate_lag_port(struct mlx5_core_dev *mdev, int ix)
@@ -2879,7 +2912,7 @@ static void mlx5e_activate_channel(struct mlx5e_channel *c)
for (tc = 0; tc < c->num_tc; tc++)
mlx5e_activate_txqsq(&c->sq[tc]);
mlx5e_activate_icosq(&c->icosq);
- mlx5e_activate_icosq(&c->async_icosq);
+ mlx5e_activate_icosq(c->async_icosq);
if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
mlx5e_activate_xsk(c);
@@ -2900,7 +2933,7 @@ static void mlx5e_deactivate_channel(struct mlx5e_channel *c)
else
mlx5e_deactivate_rq(&c->rq);
- mlx5e_deactivate_icosq(&c->async_icosq);
+ mlx5e_deactivate_icosq(c->async_icosq);
mlx5e_deactivate_icosq(&c->icosq);
for (tc = 0; tc < c->num_tc; tc++)
mlx5e_deactivate_txqsq(&c->sq[tc]);
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c b/drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c
index 76108299ea57..57c54265dbda 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c
@@ -180,11 +180,12 @@ int mlx5e_napi_poll(struct napi_struct *napi, int budget)
busy |= work_done == budget;
mlx5e_poll_ico_cq(&c->icosq.cq);
- if (mlx5e_poll_ico_cq(&c->async_icosq.cq))
+ if (mlx5e_poll_ico_cq(&c->async_icosq->cq))
/* Don't clear the flag if nothing was polled to prevent
* queueing more WQEs and overflowing the async ICOSQ.
*/
- clear_bit(MLX5E_SQ_STATE_PENDING_XSK_TX, &c->async_icosq.state);
+ clear_bit(MLX5E_SQ_STATE_PENDING_XSK_TX,
+ &c->async_icosq->state);
/* Keep after async ICOSQ CQ poll */
if (unlikely(mlx5e_ktls_rx_pending_resync_list(c, budget)))
@@ -236,7 +237,7 @@ int mlx5e_napi_poll(struct napi_struct *napi, int budget)
mlx5e_cq_arm(&rq->cq);
mlx5e_cq_arm(&c->icosq.cq);
- mlx5e_cq_arm(&c->async_icosq.cq);
+ mlx5e_cq_arm(&c->async_icosq->cq);
if (c->xdpsq)
mlx5e_cq_arm(&c->xdpsq->cq);
--
2.31.1
^ permalink raw reply related [flat|nested] 6+ messages in thread* [PATCH net-next V2 4/4] net/mlx5e: Conditionally create async ICOSQ
2026-01-14 7:46 [PATCH net-next V2 0/4] net/mlx5e: Save per-channel async ICOSQ in default Tariq Toukan
` (2 preceding siblings ...)
2026-01-14 7:46 ` [PATCH net-next V2 3/4] net/mlx5e: Move async ICOSQ to dynamic allocation Tariq Toukan
@ 2026-01-14 7:46 ` Tariq Toukan
2026-01-19 21:00 ` [PATCH net-next V2 0/4] net/mlx5e: Save per-channel async ICOSQ in default patchwork-bot+netdevbpf
4 siblings, 0 replies; 6+ messages in thread
From: Tariq Toukan @ 2026-01-14 7:46 UTC (permalink / raw)
To: Eric Dumazet, Jakub Kicinski, Paolo Abeni, Andrew Lunn,
David S. Miller
Cc: Saeed Mahameed, Leon Romanovsky, Tariq Toukan, Mark Bloch, netdev,
linux-rdma, linux-kernel, Gal Pressman, William Tu, toke
From: William Tu <witu@nvidia.com>
The async ICOSQ is only required by TLS RX (for re-sync flow) and XSK
TX. Create it only when these features are enabled instead of always
allocating it. This reduces per-channel memory usage, saves hardware
resources, improves latency, and decreases the default number of SQs
(from 3 to 2) and CQs (from 4 to 3). It also speeds up channel
open/close operations for a netdev when async ICOSQ is not needed.
Currently when TLS RX is enabled, there is no channel reset triggered.
As a result, async ICOSQ allocation is not triggered, causing a NULL
pointer crash. One solution is to do channel reset every time when
toggling TLS RX. However, it's not straightforward as the offload
state matters only on connection creation, and can go on beyond the
channels reset.
Instead, introduce a new field 'ktls_rx_was_enabled': if TLS RX is
enabled for the first time: reset channels, create async ICOSQ, set
the field. From that point on, no need to reset channels for any TLS
RX enable/disable. Async ICOSQ will always be needed.
For XSK TX, async ICOSQ is used in wakeup control and is guaranteed
to have async ICOSQ allocated.
This improves the latency of interface up/down operations when it
applies.
Perf numbers:
NIC: Connect-X7.
Test: Latency of interface up + down operations.
Measured 20% speedup.
Saving ~0.36 sec for 248 channels (~1.45 msec per channel).
Signed-off-by: William Tu <witu@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
drivers/net/ethernet/mellanox/mlx5/core/en.h | 1 +
.../mellanox/mlx5/core/en_accel/ktls.c | 10 ++++-
.../net/ethernet/mellanox/mlx5/core/en_main.c | 30 ++++++++++-----
.../net/ethernet/mellanox/mlx5/core/en_txrx.c | 38 ++++++++++---------
4 files changed, 50 insertions(+), 29 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en.h b/drivers/net/ethernet/mellanox/mlx5/core/en.h
index a7076b26fd5c..d16bdef95703 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en.h
@@ -939,6 +939,7 @@ struct mlx5e_priv {
u8 max_opened_tc;
bool tx_ptp_opened;
bool rx_ptp_opened;
+ bool ktls_rx_was_enabled;
struct kernel_hwtstamp_config hwtstamp_config;
u16 q_counter[MLX5_SD_MAX_GROUP_SZ];
u16 drop_rq_q_counter;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls.c b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls.c
index e3e57c849436..1c2cc2aad2b0 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls.c
@@ -135,10 +135,15 @@ int mlx5e_ktls_set_feature_rx(struct net_device *netdev, bool enable)
int err = 0;
mutex_lock(&priv->state_lock);
- if (enable)
+ if (enable) {
err = mlx5e_accel_fs_tcp_create(priv->fs);
- else
+ if (!err && !priv->ktls_rx_was_enabled) {
+ priv->ktls_rx_was_enabled = true;
+ mlx5e_safe_reopen_channels(priv);
+ }
+ } else {
mlx5e_accel_fs_tcp_destroy(priv->fs);
+ }
mutex_unlock(&priv->state_lock);
return err;
@@ -161,6 +166,7 @@ int mlx5e_ktls_init_rx(struct mlx5e_priv *priv)
destroy_workqueue(priv->tls->rx_wq);
return err;
}
+ priv->ktls_rx_was_enabled = true;
}
return 0;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
index aa4ff3963b86..d04ba93fe617 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
@@ -2632,7 +2632,8 @@ static void mlx5e_close_async_icosq(struct mlx5e_icosq *async_icosq)
static int mlx5e_open_queues(struct mlx5e_channel *c,
struct mlx5e_params *params,
- struct mlx5e_channel_param *cparam)
+ struct mlx5e_channel_param *cparam,
+ bool async_icosq_needed)
{
const struct net_device_ops *netdev_ops = c->netdev->netdev_ops;
struct dim_cq_moder icocq_moder = {0, 0};
@@ -2668,10 +2669,13 @@ static int mlx5e_open_queues(struct mlx5e_channel *c,
if (err)
goto err_close_rx_cq;
- c->async_icosq = mlx5e_open_async_icosq(c, params, cparam, &ccp);
- if (IS_ERR(c->async_icosq)) {
- err = PTR_ERR(c->async_icosq);
- goto err_close_rq_xdpsq_cq;
+ if (async_icosq_needed) {
+ c->async_icosq = mlx5e_open_async_icosq(c, params, cparam,
+ &ccp);
+ if (IS_ERR(c->async_icosq)) {
+ err = PTR_ERR(c->async_icosq);
+ goto err_close_rq_xdpsq_cq;
+ }
}
mutex_init(&c->icosq_recovery_lock);
@@ -2708,7 +2712,8 @@ static int mlx5e_open_queues(struct mlx5e_channel *c,
mlx5e_close_icosq(&c->icosq);
err_close_async_icosq:
- mlx5e_close_async_icosq(c->async_icosq);
+ if (c->async_icosq)
+ mlx5e_close_async_icosq(c->async_icosq);
err_close_rq_xdpsq_cq:
if (c->xdp)
@@ -2740,7 +2745,8 @@ static void mlx5e_close_queues(struct mlx5e_channel *c)
mlx5e_close_sqs(c);
mlx5e_close_icosq(&c->icosq);
mutex_destroy(&c->icosq_recovery_lock);
- mlx5e_close_async_icosq(c->async_icosq);
+ if (c->async_icosq)
+ mlx5e_close_async_icosq(c->async_icosq);
if (c->xdp)
mlx5e_close_cq(&c->rq_xdpsq.cq);
mlx5e_close_cq(&c->rq.cq);
@@ -2825,6 +2831,7 @@ static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
struct mlx5e_channel_param *cparam;
struct mlx5_core_dev *mdev;
struct mlx5e_xsk_param xsk;
+ bool async_icosq_needed;
struct mlx5e_channel *c;
unsigned int irq;
int vec_ix;
@@ -2874,7 +2881,8 @@ static int mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
netif_napi_add_config_locked(netdev, &c->napi, mlx5e_napi_poll, ix);
netif_napi_set_irq_locked(&c->napi, irq);
- err = mlx5e_open_queues(c, params, cparam);
+ async_icosq_needed = !!xsk_pool || priv->ktls_rx_was_enabled;
+ err = mlx5e_open_queues(c, params, cparam, async_icosq_needed);
if (unlikely(err))
goto err_napi_del;
@@ -2912,7 +2920,8 @@ static void mlx5e_activate_channel(struct mlx5e_channel *c)
for (tc = 0; tc < c->num_tc; tc++)
mlx5e_activate_txqsq(&c->sq[tc]);
mlx5e_activate_icosq(&c->icosq);
- mlx5e_activate_icosq(c->async_icosq);
+ if (c->async_icosq)
+ mlx5e_activate_icosq(c->async_icosq);
if (test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
mlx5e_activate_xsk(c);
@@ -2933,7 +2942,8 @@ static void mlx5e_deactivate_channel(struct mlx5e_channel *c)
else
mlx5e_deactivate_rq(&c->rq);
- mlx5e_deactivate_icosq(c->async_icosq);
+ if (c->async_icosq)
+ mlx5e_deactivate_icosq(c->async_icosq);
mlx5e_deactivate_icosq(&c->icosq);
for (tc = 0; tc < c->num_tc; tc++)
mlx5e_deactivate_txqsq(&c->sq[tc]);
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c b/drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c
index 57c54265dbda..b31f689fe271 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_txrx.c
@@ -125,6 +125,7 @@ int mlx5e_napi_poll(struct napi_struct *napi, int budget)
{
struct mlx5e_channel *c = container_of(napi, struct mlx5e_channel,
napi);
+ struct mlx5e_icosq *aicosq = c->async_icosq;
struct mlx5e_ch_stats *ch_stats = c->stats;
struct mlx5e_xdpsq *xsksq = &c->xsksq;
struct mlx5e_txqsq __rcu **qos_sqs;
@@ -180,16 +181,18 @@ int mlx5e_napi_poll(struct napi_struct *napi, int budget)
busy |= work_done == budget;
mlx5e_poll_ico_cq(&c->icosq.cq);
- if (mlx5e_poll_ico_cq(&c->async_icosq->cq))
- /* Don't clear the flag if nothing was polled to prevent
- * queueing more WQEs and overflowing the async ICOSQ.
- */
- clear_bit(MLX5E_SQ_STATE_PENDING_XSK_TX,
- &c->async_icosq->state);
-
- /* Keep after async ICOSQ CQ poll */
- if (unlikely(mlx5e_ktls_rx_pending_resync_list(c, budget)))
- busy |= mlx5e_ktls_rx_handle_resync_list(c, budget);
+ if (aicosq) {
+ if (mlx5e_poll_ico_cq(&aicosq->cq))
+ /* Don't clear the flag if nothing was polled to prevent
+ * queueing more WQEs and overflowing the async ICOSQ.
+ */
+ clear_bit(MLX5E_SQ_STATE_PENDING_XSK_TX,
+ &aicosq->state);
+
+ /* Keep after async ICOSQ CQ poll */
+ if (unlikely(mlx5e_ktls_rx_pending_resync_list(c, budget)))
+ busy |= mlx5e_ktls_rx_handle_resync_list(c, budget);
+ }
busy |= INDIRECT_CALL_2(rq->post_wqes,
mlx5e_post_rx_mpwqes,
@@ -237,16 +240,17 @@ int mlx5e_napi_poll(struct napi_struct *napi, int budget)
mlx5e_cq_arm(&rq->cq);
mlx5e_cq_arm(&c->icosq.cq);
- mlx5e_cq_arm(&c->async_icosq->cq);
+ if (aicosq) {
+ mlx5e_cq_arm(&aicosq->cq);
+ if (xsk_open) {
+ mlx5e_handle_rx_dim(xskrq);
+ mlx5e_cq_arm(&xsksq->cq);
+ mlx5e_cq_arm(&xskrq->cq);
+ }
+ }
if (c->xdpsq)
mlx5e_cq_arm(&c->xdpsq->cq);
- if (xsk_open) {
- mlx5e_handle_rx_dim(xskrq);
- mlx5e_cq_arm(&xsksq->cq);
- mlx5e_cq_arm(&xskrq->cq);
- }
-
if (unlikely(aff_change && busy_xsk)) {
mlx5e_trigger_irq(&c->icosq);
ch_stats->force_irq++;
--
2.31.1
^ permalink raw reply related [flat|nested] 6+ messages in thread* Re: [PATCH net-next V2 0/4] net/mlx5e: Save per-channel async ICOSQ in default
2026-01-14 7:46 [PATCH net-next V2 0/4] net/mlx5e: Save per-channel async ICOSQ in default Tariq Toukan
` (3 preceding siblings ...)
2026-01-14 7:46 ` [PATCH net-next V2 4/4] net/mlx5e: Conditionally create async ICOSQ Tariq Toukan
@ 2026-01-19 21:00 ` patchwork-bot+netdevbpf
4 siblings, 0 replies; 6+ messages in thread
From: patchwork-bot+netdevbpf @ 2026-01-19 21:00 UTC (permalink / raw)
To: Tariq Toukan
Cc: edumazet, kuba, pabeni, andrew+netdev, davem, saeedm, leon,
mbloch, netdev, linux-rdma, linux-kernel, gal, witu, toke
Hello:
This series was applied to netdev/net-next.git (main)
by Jakub Kicinski <kuba@kernel.org>:
On Wed, 14 Jan 2026 09:46:36 +0200 you wrote:
> Hi,
>
> This is V2, find V1 here:
> https://lore.kernel.org/all/1762939749-1165658-1-git-send-email-tariqt@nvidia.com/
>
> This series by William reduces the default number of SQs in a channel
> from 3 down to 2, by not creating the async ICOSQ (asynchronous
> internal-communication-operations send-queue).
>
> [...]
Here is the summary with links:
- [net-next,V2,1/4] net/mlx5e: Move async ICOSQ lock into ICOSQ struct
https://git.kernel.org/netdev/net-next/c/ea945f4f3991
- [net-next,V2,2/4] net/mlx5e: Use regular ICOSQ for triggering NAPI
https://git.kernel.org/netdev/net-next/c/56aca3e0f730
- [net-next,V2,3/4] net/mlx5e: Move async ICOSQ to dynamic allocation
https://git.kernel.org/netdev/net-next/c/1b080bd74840
- [net-next,V2,4/4] net/mlx5e: Conditionally create async ICOSQ
https://git.kernel.org/netdev/net-next/c/abed42f9cd80
You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html
^ permalink raw reply [flat|nested] 6+ messages in thread