From: Brett Creeley brett.creeley@intel.com
commit e6ba5273d4ede03d075d7a116b8edad1f6115f4d upstream.
[I had to fix the cherry-pick manually as the patch added a line around some context that was missing.]
The VF can be configured via the PF's ndo ops at the same time the PF is receiving/handling virtchnl messages. This has many issues, with one of them being the ndo op could be actively resetting a VF (i.e. resetting it to the default state and deleting/re-adding the VF's VSI) while a virtchnl message is being handled. The following error was seen because a VF ndo op was used to change a VF's trust setting while the VIRTCHNL_OP_CONFIG_VSI_QUEUES was ongoing:
[35274.192484] ice 0000:88:00.0: Failed to set LAN Tx queue context, error: ICE_ERR_PARAM [35274.193074] ice 0000:88:00.0: VF 0 failed opcode 6, retval: -5 [35274.193640] iavf 0000:88:01.0: PF returned error -5 (IAVF_ERR_PARAM) to our request 6
Fix this by making sure the virtchnl handling and VF ndo ops that trigger VF resets cannot run concurrently. This is done by adding a struct mutex cfg_lock to each VF structure. For VF ndo ops, the mutex will be locked around the critical operations and VFR. Since the ndo ops will trigger a VFR, the virtchnl thread will use mutex_trylock(). This is done because if any other thread (i.e. VF ndo op) has the mutex, then that means the current VF message being handled is no longer valid, so just ignore it.
This issue can be seen using the following commands:
for i in {0..50}; do rmmod ice modprobe ice
sleep 1
echo 1 > /sys/class/net/ens785f0/device/sriov_numvfs echo 1 > /sys/class/net/ens785f1/device/sriov_numvfs
ip link set ens785f1 vf 0 trust on ip link set ens785f0 vf 0 trust on
sleep 2
echo 0 > /sys/class/net/ens785f0/device/sriov_numvfs echo 0 > /sys/class/net/ens785f1/device/sriov_numvfs sleep 1 echo 1 > /sys/class/net/ens785f0/device/sriov_numvfs echo 1 > /sys/class/net/ens785f1/device/sriov_numvfs
ip link set ens785f1 vf 0 trust on ip link set ens785f0 vf 0 trust on done
Fixes: 7c710869d64e ("ice: Add handlers for VF netdevice operations") Cc: stable@vger.kernel.org # 5.8.x Signed-off-by: Brett Creeley brett.creeley@intel.com Tested-by: Konrad Jankowski konrad0.jankowski@intel.com Signed-off-by: Tony Nguyen anthony.l.nguyen@intel.com Signed-off-by: Jacob Keller jacob.e.keller@intel.com --- This is for stable trees 5.8 through 5.12. I sent patches for 5.13 and 5.14 separately since they have slightly different context
.../net/ethernet/intel/ice/ice_virtchnl_pf.c | 25 +++++++++++++++++++ .../net/ethernet/intel/ice/ice_virtchnl_pf.h | 5 ++++ 2 files changed, 30 insertions(+)
diff --git a/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c b/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c index 48dee9c5d534..66da8f540454 100644 --- a/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c +++ b/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c @@ -375,6 +375,8 @@ void ice_free_vfs(struct ice_pf *pf) set_bit(ICE_VF_STATE_DIS, pf->vf[i].vf_states); ice_free_vf_res(&pf->vf[i]); } + + mutex_destroy(&pf->vf[i].cfg_lock); }
if (ice_sriov_free_msix_res(pf)) @@ -1556,6 +1558,8 @@ static void ice_set_dflt_settings_vfs(struct ice_pf *pf) set_bit(ICE_VIRTCHNL_VF_CAP_L2, &vf->vf_caps); vf->spoofchk = true; vf->num_vf_qs = pf->num_qps_per_vf; + + mutex_init(&vf->cfg_lock); } }
@@ -3389,6 +3393,8 @@ ice_set_vf_port_vlan(struct net_device *netdev, int vf_id, u16 vlan_id, u8 qos, return 0; }
+ mutex_lock(&vf->cfg_lock); + vf->port_vlan_info = vlanprio;
if (vf->port_vlan_info) @@ -3398,6 +3404,7 @@ ice_set_vf_port_vlan(struct net_device *netdev, int vf_id, u16 vlan_id, u8 qos, dev_info(dev, "Clearing port VLAN on VF %d\n", vf_id);
ice_vc_reset_vf(vf); + mutex_unlock(&vf->cfg_lock);
return 0; } @@ -3763,6 +3770,15 @@ void ice_vc_process_vf_msg(struct ice_pf *pf, struct ice_rq_event_info *event) return; }
+ /* VF is being configured in another context that triggers a VFR, so no + * need to process this message + */ + if (!mutex_trylock(&vf->cfg_lock)) { + dev_info(dev, "VF %u is being configured in another context that will trigger a VFR, so there is no need to handle this message\n", + vf->vf_id); + return; + } + switch (v_opcode) { case VIRTCHNL_OP_VERSION: err = ice_vc_get_ver_msg(vf, msg); @@ -3839,6 +3855,8 @@ void ice_vc_process_vf_msg(struct ice_pf *pf, struct ice_rq_event_info *event) dev_info(dev, "PF failed to honor VF %d, opcode %d, error %d\n", vf_id, v_opcode, err); } + + mutex_unlock(&vf->cfg_lock); }
/** @@ -3953,6 +3971,8 @@ int ice_set_vf_mac(struct net_device *netdev, int vf_id, u8 *mac) return -EINVAL; }
+ mutex_lock(&vf->cfg_lock); + /* VF is notified of its new MAC via the PF's response to the * VIRTCHNL_OP_GET_VF_RESOURCES message after the VF has been reset */ @@ -3970,6 +3990,7 @@ int ice_set_vf_mac(struct net_device *netdev, int vf_id, u8 *mac) }
ice_vc_reset_vf(vf); + mutex_unlock(&vf->cfg_lock); return 0; }
@@ -3999,11 +4020,15 @@ int ice_set_vf_trust(struct net_device *netdev, int vf_id, bool trusted) if (trusted == vf->trusted) return 0;
+ mutex_lock(&vf->cfg_lock); + vf->trusted = trusted; ice_vc_reset_vf(vf); dev_info(ice_pf_to_dev(pf), "VF %u is now %strusted\n", vf_id, trusted ? "" : "un");
+ mutex_unlock(&vf->cfg_lock); + return 0; }
diff --git a/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.h b/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.h index 0f519fba3770..59e5b4f16e96 100644 --- a/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.h +++ b/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.h @@ -68,6 +68,11 @@ struct ice_mdd_vf_events { struct ice_vf { struct ice_pf *pf;
+ /* Used during virtchnl message handling and NDO ops against the VF + * that will trigger a VFR + */ + struct mutex cfg_lock; + u16 vf_id; /* VF ID in the PF space */ u16 lan_vsi_idx; /* index into PF struct */ /* first vector index of this VF in the PF space */
From: Brett Creeley brett.creeley@intel.com
commit b385cca47363316c6d9a74ae9db407bbc281f815 upstream.
[ This commit is already in stable/linux-5.10.y as d83832d4a838 ("ice: Fix not stopping Tx queues for VFs"). It's missing from v5.8, v5.9, v5.11, and v5.12 ]
When a VF is removed and/or reset its Tx queues need to be stopped from the PF. This is done by calling the ice_dis_vf_qs() function, which calls ice_vsi_stop_lan_tx_rings(). Currently ice_dis_vf_qs() is protected by the VF state bit ICE_VF_STATE_QS_ENA. Unfortunately, this is causing the Tx queues to not be disabled in some cases and when the VF tries to re-enable/reconfigure its Tx queues over virtchnl the op is failing. This is because a VF can be reset and/or removed before the ICE_VF_STATE_QS_ENA bit is set, but the Tx queues were already configured via ice_vsi_cfg_single_txq() in the VIRTCHNL_OP_CONFIG_VSI_QUEUES op. However, the ICE_VF_STATE_QS_ENA bit is set on a successful VIRTCHNL_OP_ENABLE_QUEUES, which will always happen after the VIRTCHNL_OP_CONFIG_VSI_QUEUES op.
This was causing the following error message when loading the ice driver, creating VFs, and modifying VF trust in an endless loop:
[35274.192484] ice 0000:88:00.0: Failed to set LAN Tx queue context, error: ICE_ERR_PARAM [35274.193074] ice 0000:88:00.0: VF 0 failed opcode 6, retval: -5 [35274.193640] iavf 0000:88:01.0: PF returned error -5 (IAVF_ERR_PARAM) to our request 6
Fix this by always calling ice_dis_vf_qs() and silencing the error message in ice_vsi_stop_tx_ring() since the calling code ignores the return anyway. Also, all other places that call ice_vsi_stop_tx_ring() catch the error, so this doesn't affect those flows since there was no change to the values the function returns.
Other solutions were considered (i.e. tracking which VF queues had been "started/configured" in VIRTCHNL_OP_CONFIG_VSI_QUEUES, but it seemed more complicated than it was worth. This solution also brings in the chance for other unexpected conditions due to invalid state bit checks. So, the proposed solution seemed like the best option since there is no harm in failing to stop Tx queues that were never started.
This issue can be seen using the following commands:
for i in {0..50}; do rmmod ice modprobe ice
sleep 1
echo 1 > /sys/class/net/ens785f0/device/sriov_numvfs echo 1 > /sys/class/net/ens785f1/device/sriov_numvfs
ip link set ens785f1 vf 0 trust on ip link set ens785f0 vf 0 trust on
sleep 2
echo 0 > /sys/class/net/ens785f0/device/sriov_numvfs echo 0 > /sys/class/net/ens785f1/device/sriov_numvfs sleep 1 echo 1 > /sys/class/net/ens785f0/device/sriov_numvfs echo 1 > /sys/class/net/ens785f1/device/sriov_numvfs
ip link set ens785f1 vf 0 trust on ip link set ens785f0 vf 0 trust on done
Fixes: 77ca27c41705 ("ice: add support for virtchnl_queue_select.[tx|rx]_queues bitmap") Cc: stable@vger.kernel.org # 5.8.x Signed-off-by: Brett Creeley brett.creeley@intel.com Tested-by: Konrad Jankowski konrad0.jankowski@intel.com Signed-off-by: Tony Nguyen anthony.l.nguyen@intel.com --- This is for stable trees 5.8 through 5.12, excepting 5.10 which already has it. I sent a series with these fixes for 5.13 and 5.14 separately.
drivers/net/ethernet/intel/ice/ice_base.c | 2 +- drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c | 6 ++---- 2 files changed, 3 insertions(+), 5 deletions(-)
diff --git a/drivers/net/ethernet/intel/ice/ice_base.c b/drivers/net/ethernet/intel/ice/ice_base.c index 1148d768f8ed..203a3b21adc8 100644 --- a/drivers/net/ethernet/intel/ice/ice_base.c +++ b/drivers/net/ethernet/intel/ice/ice_base.c @@ -859,7 +859,7 @@ ice_vsi_stop_tx_ring(struct ice_vsi *vsi, enum ice_disq_rst_src rst_src, } else if (status == ICE_ERR_DOES_NOT_EXIST) { dev_dbg(ice_pf_to_dev(vsi->back), "LAN Tx queues do not exist, nothing to disable\n"); } else if (status) { - dev_err(ice_pf_to_dev(vsi->back), "Failed to disable LAN Tx queues, error: %s\n", + dev_dbg(ice_pf_to_dev(vsi->back), "Failed to disable LAN Tx queues, error: %s\n", ice_stat_str(status)); return -ENODEV; } diff --git a/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c b/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c index 66da8f540454..8fa941231500 100644 --- a/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c +++ b/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c @@ -362,8 +362,7 @@ void ice_free_vfs(struct ice_pf *pf)
/* Avoid wait time by stopping all VFs at the same time */ ice_for_each_vf(pf, i) - if (test_bit(ICE_VF_STATE_QS_ENA, pf->vf[i].vf_states)) - ice_dis_vf_qs(&pf->vf[i]); + ice_dis_vf_qs(&pf->vf[i]);
tmp = pf->num_alloc_vfs; pf->num_qps_per_vf = 0; @@ -1329,8 +1328,7 @@ bool ice_reset_vf(struct ice_vf *vf, bool is_vflr)
vsi = pf->vsi[vf->lan_vsi_idx];
- if (test_bit(ICE_VF_STATE_QS_ENA, vf->vf_states)) - ice_dis_vf_qs(vf); + ice_dis_vf_qs(vf);
/* Call Disable LAN Tx queue AQ whether or not queues are * enabled. This is needed for successful completion of VFR.
On Mon, Feb 28, 2022 at 12:51:13PM -0800, Jacob Keller wrote:
From: Brett Creeley brett.creeley@intel.com
commit b385cca47363316c6d9a74ae9db407bbc281f815 upstream.
[ This commit is already in stable/linux-5.10.y as d83832d4a838 ("ice: Fix not stopping Tx queues for VFs"). It's missing from v5.8, v5.9, v5.11, and v5.12 ]
That's because all of those branches are end-of-life, can't add new stuff to them, sorry.
greg k-h
commit fadead80fe4c033b5e514fcbadd20b55c4494112 upstream.
[ The original upstream commit fixed c503e63200c6 ("ice: Stop processing VF messages during teardown"). This change is not in the 5.8 through 5.12 stable trees. However, that commit was itself a fix for ddf30f7ff840 ("ice: Add handler to configure SR-IOV"). This fix originally reverted c503e63200c6 and supplied this more complete fix based on properly locking reset and remove. This backport only applies the new fix, rather than applying the broken inbetween fix and then reverting it. ]
Commit c503e63200c6 ("ice: Stop processing VF messages during teardown") introduced a driver state flag, ICE_VF_DEINIT_IN_PROGRESS, which is intended to prevent some issues with concurrently handling messages from VFs while tearing down the VFs.
This change was motivated by crashes caused while tearing down and bringing up VFs in rapid succession.
It turns out that the fix actually introduces issues with the VF driver caused because the PF no longer responds to any messages sent by the VF during its .remove routine. This results in the VF potentially removing its DMA memory before the PF has shut down the device queues.
Additionally, the fix doesn't actually resolve concurrency issues within the ice driver. It is possible for a VF to initiate a reset just prior to the ice driver removing VFs. This can result in the remove task concurrently operating while the VF is being reset. This results in similar memory corruption and panics purportedly fixed by that commit.
Fix this concurrency at its root by protecting both the reset and removal flows using the existing VF cfg_lock. This ensures that we cannot remove the VF while any outstanding critical tasks such as a virtchnl message or a reset are occurring.
This locking change also fixes the root cause originally fixed by commit c503e63200c6 ("ice: Stop processing VF messages during teardown"), so we can simply revert it.
Note that I kept these two changes together because simply reverting the original commit alone would leave the driver vulnerable to worse race conditions.
Fixes: c503e63200c6 ("ice: Stop processing VF messages during teardown") Cc: stable@vger.kernel.org # 5.8.x: e6ba5273d4ed: ice: Fix race conditions between virtchnl handling and VF ndo ops Cc: stable@vger.kernel.org # 5.8.x: b385cca47363: ice: Fix not stopping Tx queues for VF Cc: stable@vger.kernel.org # 5.8.x Signed-off-by: Jacob Keller jacob.e.keller@intel.com Tested-by: Konrad Jankowski konrad0.jankowski@intel.com Signed-off-by: Tony Nguyen anthony.l.nguyen@intel.com --- This is for stable trees 5.8 through 5.12. I sent a series with this fix for 5.13 and 5.14 separately since they have slightly different context.
drivers/net/ethernet/intel/ice/ice_main.c | 2 ++ .../net/ethernet/intel/ice/ice_virtchnl_pf.c | 35 +++++++++++++------ 2 files changed, 27 insertions(+), 10 deletions(-)
diff --git a/drivers/net/ethernet/intel/ice/ice_main.c b/drivers/net/ethernet/intel/ice/ice_main.c index b61cd84be97f..171cb9198c8d 100644 --- a/drivers/net/ethernet/intel/ice/ice_main.c +++ b/drivers/net/ethernet/intel/ice/ice_main.c @@ -1601,7 +1601,9 @@ static void ice_handle_mdd_event(struct ice_pf *pf) * reset, so print the event prior to reset. */ ice_print_vf_rx_mdd_event(vf); + mutex_lock(&pf->vf[i].cfg_lock); ice_reset_vf(&pf->vf[i], false); + mutex_unlock(&pf->vf[i].cfg_lock); } } } diff --git a/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c b/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c index 8fa941231500..af1fbb2c1479 100644 --- a/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c +++ b/drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c @@ -360,22 +360,26 @@ void ice_free_vfs(struct ice_pf *pf) else dev_warn(dev, "VFs are assigned - not disabling SR-IOV\n");
- /* Avoid wait time by stopping all VFs at the same time */ - ice_for_each_vf(pf, i) - ice_dis_vf_qs(&pf->vf[i]); - tmp = pf->num_alloc_vfs; pf->num_qps_per_vf = 0; pf->num_alloc_vfs = 0; for (i = 0; i < tmp; i++) { - if (test_bit(ICE_VF_STATE_INIT, pf->vf[i].vf_states)) { + struct ice_vf *vf = &pf->vf[i]; + + mutex_lock(&vf->cfg_lock); + + ice_dis_vf_qs(vf); + + if (test_bit(ICE_VF_STATE_INIT, vf->vf_states)) { /* disable VF qp mappings and set VF disable state */ - ice_dis_vf_mappings(&pf->vf[i]); - set_bit(ICE_VF_STATE_DIS, pf->vf[i].vf_states); - ice_free_vf_res(&pf->vf[i]); + ice_dis_vf_mappings(vf); + set_bit(ICE_VF_STATE_DIS, vf->vf_states); + ice_free_vf_res(vf); }
- mutex_destroy(&pf->vf[i].cfg_lock); + mutex_unlock(&vf->cfg_lock); + + mutex_destroy(&vf->cfg_lock); }
if (ice_sriov_free_msix_res(pf)) @@ -1259,9 +1263,13 @@ bool ice_reset_all_vfs(struct ice_pf *pf, bool is_vflr) ice_for_each_vf(pf, v) { vf = &pf->vf[v];
+ mutex_lock(&vf->cfg_lock); + ice_vf_pre_vsi_rebuild(vf); ice_vf_rebuild_vsi(vf); ice_vf_post_vsi_rebuild(vf); + + mutex_unlock(&vf->cfg_lock); }
ice_flush(hw); @@ -1308,6 +1316,8 @@ bool ice_reset_vf(struct ice_vf *vf, bool is_vflr) u32 reg; int i;
+ lockdep_assert_held(&vf->cfg_lock); + dev = ice_pf_to_dev(pf);
if (test_bit(__ICE_VF_RESETS_DISABLED, pf->state)) { @@ -1765,9 +1775,12 @@ void ice_process_vflr_event(struct ice_pf *pf) bit_idx = (hw->func_caps.vf_base_id + vf_id) % 32; /* read GLGEN_VFLRSTAT register to find out the flr VFs */ reg = rd32(hw, GLGEN_VFLRSTAT(reg_idx)); - if (reg & BIT(bit_idx)) + if (reg & BIT(bit_idx)) { /* GLGEN_VFLRSTAT bit will be cleared in ice_reset_vf */ + mutex_lock(&vf->cfg_lock); ice_reset_vf(vf, true); + mutex_unlock(&vf->cfg_lock); + } } }
@@ -1844,7 +1857,9 @@ ice_vf_lan_overflow_event(struct ice_pf *pf, struct ice_rq_event_info *event) if (!vf) return;
+ mutex_lock(&vf->cfg_lock); ice_vc_reset_vf(vf); + mutex_unlock(&vf->cfg_lock); }
/**
linux-stable-mirror@lists.linaro.org
-
Greg KH -
Jacob Keller