idpf: disable DIM work before freeing q_vectors
Subsystems (3)
PATCH 1/1 idpf: disable DIM work before freeing q_vectors
idpf never drains the Tx/Rx DIM works before freeing the memory they live in. tx_dim and rx_dim are embedded in struct idpf_q_vector, they are queued from the NAPI poll via net_dim(), and idpf_vport_intr_rel() ends with kfree(rsrc->q_vectors). Nothing in the driver cancels them.
idpf_tx_dim_work() and idpf_rx_dim_work() then run on freed memory:idpf_vport_intr_write_itr() writes the ITR register through q_vector->intr_reg.tx_itr / rx_itr, void __iomem pointers loaded out of the freed q_vector. No configuration is needed to get there -- IDPF_ITR_IS_DYNAMIC() is defined as (itr_mode) and idpf_vport_alloc() initialises both modes to IDPF_ITR_DYNAMIC.
Draining after idpf_vport_intr_napi_dis_all() is not enough on its own. idpf_net_dim() is called from inside the
"if (napi_complete_done(napi, work_done))" branch of the poll, and napi_complete_done() has already cleared NAPIF_STATE_SCHED by then. napi_disable_locked() waits only while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)), so napi_disable() can return while the poll tail is still queueing the work, and a plain cancel_work_sync() would be re-armed behind the drain.
Use disable_work_sync(): schedule_work() on a work with a non-zero disable count is dropped by clear_pending_if_disabled() before __queue_work() is reached.
Move idpf_init_dim() to idpf_vport_intr_alloc() so the works are initialised on every path that can reach the drain -- the three "goto intr_deinit" sites between idpf_vport_intr_init() and idpf_vport_intr_ena() get there without the enable side having run. Nothing re-enables them: rsrc->q_vectors is freed on every exit from idpf_vport_open() and on every idpf_vport_stop(), so the count dies with the object.
It is a race, not a deterministic failure -- net_dim() only schedules once DIM_NEVENTS events have accumulated and the profile index changes. A KASAN ifup/ifdown loop under load is the way to see it.
---
Found by static analysis; not tested on hardware. Testing by someone at Intel with an idpf device would be very welcome.
disable_work_sync() landed in v6.10, which is newer than the commits this fixes, hence the annotation on the stable tag. A backport to older trees needs the work held down another way rather than a cherry-pick; happy to send one if the bot reports a failure.
libwx does the same thing in the same order -- napi_disable() then disable_work_sync() on the dim works -- in wx_napi_disable_all().
Based on v7.2-rc2.
drivers/net/ethernet/intel/idpf/idpf_txrx.c | 24 +++++++++++++++++++++- 1 file changed, 23 insertions(+), 1 deletion(-)
drivers/net/ethernet/intel/idpf/idpf_txrx.c +23 -1
@@ -4148,6 +4148,26 @@ static void idpf_vport_intr_ena_irq_all(struct idpf_vport *vport,writel (rsrc->noirq_dyn_ctl_ena, rsrc->noirq_dyn_ctl); } +/** +* idpf_vport_intr_dis_dim_all - Disable DIM work for all q_vectors +* @rsrc: pointer to queue and vector resources +* +* The DIM works are embedded in the q_vector array that +* idpf_vport_intr_rel() frees, and the poll arms them after +* napi_complete_done() has already cleared NAPI_STATE_SCHED. Disable +* rather than just cancel, so that a poll tail still running past +* napi_disable() cannot queue them again behind the drain. +*/ +static void idpf_vport_intr_dis_dim_all (struct idpf_q_vec_rsrc *rsrc) +{ +for (u16 v_idx =0 ; v_idx < rsrc->num_q_vectors; v_idx++) { +struct idpf_q_vector *q_vector = &rsrc->q_vectors[v_idx]; + +disable_work_sync (&q_vector->tx_dim.work); +disable_work_sync (&q_vector->rx_dim.work); + } +} +/** * idpf_vport_intr_deinit - Release all vector associations for the vport * @vport: main vport structure
@@ -4158,6 +4178,7 @@ void idpf_vport_intr_deinit(struct idpf_vport *vport, {idpf_vport_intr_dis_irq_all (rsrc);idpf_vport_intr_napi_dis_all (rsrc); +idpf_vport_intr_dis_dim_all (rsrc);idpf_vport_intr_napi_del_all (rsrc);idpf_vport_intr_rel_irq (vport, rsrc); }
@@ -4238,7 +4259,6 @@ static void idpf_vport_intr_napi_ena_all(struct idpf_q_vec_rsrc *rsrc)for (u16 q_idx =0 ; q_idx < rsrc->num_q_vectors; q_idx++) {struct idpf_q_vector *q_vector = &rsrc->q_vectors[q_idx]; -idpf_init_dim (q_vector);napi_enable (&q_vector->napi); } }
@@ -4581,6 +4601,8 @@ int idpf_vport_intr_alloc(struct idpf_vport *vport, q_coal = &user_config->q_coalesce[v_idx]; q_vector->vport = vport; +idpf_init_dim (q_vector); + q_vector->tx_itr_value = q_coal->tx_coalesce_usecs; q_vector->tx_intr_mode = q_coal->tx_intr_mode; q_vector->tx_itr_idx =VIRTCHNL2_ITR_IDX_1 ;
--
2.43.0
Quoted 69 lines
-----Original Message-----
From: Intel-wired-lan <[email protected]> On Behalf Of Myeonghun Pak
Sent: Monday, July 20, 2026 6:59 AM
To: Nguyen, Anthony L <[email protected]>; Kitszel, Przemyslaw <[email protected]>; [email protected]
Cc: Andrew Lunn <[email protected]>; David S . Miller <[email protected]>; Eric Dumazet <[email protected]>; Jakub Kicinski <[email protected]>; Paolo Abeni <[email protected]>; [email protected]; [email protected]; Ijae Kim <[email protected]>
Subject: [Intel-wired-lan] [PATCH net] idpf: disable DIM work before freeingq_vectorsidpf never drains the Tx/Rx DIM works before freeing the memory they live in.
tx_dimandrx_dimare embedded in structidpf_q_vector, they are queued from the NAPI poll vianet_dim(), andidpf_vport_intr_rel()ends withkfree(rsrc->q_vectors). Nothing in the driver cancels them.
idpf_tx_dim_work()andidpf_rx_dim_work()then run on freed memory:idpf_vport_intr_write_itr()writes the ITR register through q_vector-Quoted 1 line
intr_reg.tx_itr /
rx_itr, void__iomempointers loaded out of the freed
q_vector. No configuration is needed to get there --IDPF_ITR_IS_DYNAMIC()is defined as (itr_mode) andidpf_vport_alloc()initialises both modes toIDPF_ITR_DYNAMIC.Draining after
idpf_vport_intr_napi_dis_all()is not enough on its own.idpf_net_dim()is called from inside the "if (napi_complete_done(napi, work_done))" branch of the poll, andnapi_complete_done()has already clearedNAPIF_STATE_SCHEDby then.napi_disable_locked()waits only while (val & (NAPIF_STATE_SCHED|NAPIF_STATE_NPSVC)), sonapi_disable()can return while the poll tail is still queueing the work, and a plaincancel_work_sync()would be re-armed behind the drain.Use
disable_work_sync():schedule_work()on a work with a non-zero disable count is dropped byclear_pending_if_disabled()before__queue_work()is reached.Move
idpf_init_dim()toidpf_vport_intr_alloc()so the works are initialised on every path that can reach the drain -- the three "gotointr_deinit" sites betweenidpf_vport_intr_init()andidpf_vport_intr_ena()get there without the enable side having run. Nothing re-enables them:rsrc->q_vectorsis freed on every exit fromidpf_vport_open()and on everyidpf_vport_stop(), so the count dies with the object.It is a race, not a deterministic failure --
net_dim()only schedules onceDIM_NEVENTSevents have accumulated and the profile index changes. A KASAN ifup/ifdown loop under load is the way to see it.Fixes c2d548cad150 ("idpf: add TX splitq napi poll support")Fixes 3a8845af66ed ("idpf: add RX splitq napi poll support")Cc <[email protected]> # see patch description, needs adjustments for<= 6.9
Co-developed-by Ijae Kim <[email protected]>Signed-off-by Ijae Kim <[email protected]>Signed-off-by Myeonghun Pak <[email protected]>---
Found by static analysis; not tested on hardware. Testing by someone at Intel with an idpf device would be very welcome.
disable_work_sync()landed in v6.10, which is newer than the commits this fixes, hence the annotation on the stable tag. A backport to older trees needs the work held down another way rather than a cherry-pick; happy to send one if the bot reports a failure.
Quoted 5 lines
libwx does the same thing in the same order --
napi_disable()thendisable_work_sync()on the dim works -- inwx_napi_disable_all().Based on v7.2-rc2.
2.43.0