net/mlx5e: Add support for HW-GRO to PSP
Subsystems (8)
Tariq Toukan
maintainerv4 cover net/mlx5e: Add support for HW-GRO to PSP
Hi,
Ingress PSP packets cannot be merged by the HW-GRO HW state machine because they are not decapsulated and the current HW-GRO state machine does not understand PSP.
This series by Cosmin decapsulates PSP packets in steering, which allows the now-decapsulated PSP packets (== TCP) to go through HW-GRO and be aggregated. The SPI and PSP version from the PSP header are handed off to the driver in the CQE metadata fields. They are used to terminate the HW GRO session on mismatch, and are required to construct the skb extension which is used higher up in the stack.
Some preparatory work needed to happen to allow that:
- All accel protocol markers were moved away from ft_metadata into flow_tag
- Mutual exclusion between TC and accel protocols was added. Trying to configure both IPsec and TC at the same time for example is now detected at config time instead of doing weird things at runtime.
kperf tests on a pair of CX7 NICs with 200Gbps link speed:
Streams Gbps no HW-GRO Gbps HW-GRO Speedup
------- -------------- ----------- -------
1 28 58 2.07x
2 67 102 1.52x
4 136 180 1.32x
8 175 183 1.05xRegards,
Tariq
Some internal Sashiko findings, plus Cosmin's comments:
Review
Quoted 2 lines
(quoted diff) +2
-0+ cfg.wait_hw_stats_settle() + after = cfg.netnl.qstats_get({"ifindex": cfg.ifindex}, dump=True)[0]
Could this lead to flaky test failures on multiqueue NICs?
By hardcoding the netlink dump array access to index [0], the test statically verifies statistics for the first queue only. Since the test sends traffic over an ephemeral random port, Receive Side Scaling (RSS) will hash this 4-tuple and could deliver the flow to any active RX queue. If the flow lands on a queue other than queue 0, will the test fail to observe the hardware GRO packet increments?
[CR] The premise is wrong. [0] is not queue 0, it's the result for the requested dev.
Review macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, · line 1761
Quoted 10 lines
(quoted diff) +1
-2@@ -1757,11 +1755,10 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, rx_tables = &rx_fs->tables; ft_crypto = &rx_tables->ft_crypto; - /* Set bit[31 - 30] macsec marker - 0x01 */ /* Set bit[15-0] fs id */ MLX5_SET(set_action_in, action, action_type, MLX5_ACTION_TYPE_SET); MLX5_SET(set_action_in, action, field, MLX5_ACTION_IN_FIELD_METADATA_REG_B); - MLX5_SET(set_action_in, action, data, macsec_fs_set_rx_fs_id(fs_id)); + MLX5_SET(set_action_in, action, data, fs_id);
Does removing the BIT(30) marker from the MACsec fs_id metadata cause a regression by exposing RoCE MACsec to an aliasing issue?
By removing BIT(30) from the fs_id in this patch, the fs_id written to METADATA_REG_B becomes a small integer (e.g., 1, 2).
Later in mlx5_macsec_fs_add_roce_rule_rx(), the hardware rule ft_macsec_op_check enforces MACsec on RoCE traffic by matching metadata_reg_c_5 exactly against this same fs_id value.
Could any unencrypted packet that has its REG_B set to this small integer by another NIC feature (such as TC rules setting packet marks, or PSP IPsec decryption) falsely pass the MACsec check and bypass encryption enforcement?
[CR] Preexisting, TC rules could set the fsid to the the previous value even before. 'PSP IPsec' decryption doesn't make sense.
Review macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, · line 1761
Quoted 10 lines
(quoted diff) +1
-2@@ -1757,11 +1755,10 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, rx_tables = &rx_fs->tables; ft_crypto = &rx_tables->ft_crypto; - /* Set bit[31 - 30] macsec marker - 0x01 */ /* Set bit[15-0] fs id */ MLX5_SET(set_action_in, action, action_type, MLX5_ACTION_TYPE_SET); MLX5_SET(set_action_in, action, field, MLX5_ACTION_IN_FIELD_METADATA_REG_B); - MLX5_SET(set_action_in, action, data, macsec_fs_set_rx_fs_id(fs_id)); + MLX5_SET(set_action_in, action, data, fs_id);
Does removing the BIT(30) marker allow metadata collisions in the hardware RoCE datapath?
If an administrator configures a benign TC rule that sets a mark equal to the MACsec fs_id (which is just a 16-bit integer), could an attacker send unauthenticated RoCE traffic that matches the TC mark, falsely setting the metadata and bypassing the MACsec hardware authentication checks?
[CR] This series makes TC and MACsec mutually exclusive.
Review static int mlx5_macsec_fs_add_roce_rule_rx(struct mlx5_macsec_fs *macsec_fs, u32 · line 2171
Quoted 8 lines
(quoted diff) +2
-2@@ -2160,8 +2167,8 @@ static int mlx5_macsec_fs_add_roce_rule_rx(struct mlx5_macsec_fs *macsec_fs, u32 spec->match_criteria_enable |= MLX5_MATCH_MISC_PARAMETERS_2; MLX5_SET_TO_ONES(fte_match_param, spec->match_criteria, misc_parameters_2.metadata_reg_c_5); - MLX5_SET(fte_match_param, spec->match_value, misc_parameters_2.metadata_reg_c_5, - macsec_fs_set_rx_fs_id(fs_id)); + MLX5_SET(fte_match_param, spec->match_value, + misc_parameters_2.metadata_reg_c_5, fs_id);
By matching on the bare fs_id without the BIT(30) marker in metadata_reg_c_5, is this RoCE datapath rule now susceptible to the TC mark collision issue?
[CR] See above.
Review macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, · line 1781
Quoted 8 lines
(quoted diff) +4
-0@@ -1778,6 +1775,11 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, /* Rx crypto table with SCI rule */ macsec_fs_rx_setup_fte(spec, &flow_act, attrs, true); + spec->flow_context.flags |= FLOW_CONTEXT_HAS_TAG; + spec->flow_context.flow_tag = + FIELD_PREP(MLX5E_ACCEL_FLOW_TAG_PROTO_MASK, + MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC);
Can adding FLOW_CONTEXT_HAS_TAG to the shared MACsec crypto table corrupt RoCE Queue Pair Numbers (QPN)?
This unconditionally adds a flow tag to the shared ft_crypto table, which processes both Ethernet and RoCE MACsec traffic. When FLOW_CONTEXT_HAS_TAG is set, the hardware writes the flow tag into the sop_drop_qpn field of the CQE.
Will this cause the mlx5_ib driver to read the MACsec tag (value 2) instead of the actual destination QPN for RDMA traffic, breaking RoCE delivery?
[CR] The flow_tag is only given for Eth CQEs. For RoCE/RDMA, that CQE field is the QPN, and therefore mlx5_ib is unaffected.
Review macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, · line 1811
Quoted 8 lines
(quoted diff) +4
-0@@ -1803,6 +1805,11 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, macsec_fs_rx_setup_fte(spec, &flow_act, attrs, false); + spec->flow_context.flags |= FLOW_CONTEXT_HAS_TAG; + spec->flow_context.flow_tag = + FIELD_PREP(MLX5E_ACCEL_FLOW_TAG_PROTO_MASK, + MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC);
Is it safe to set the flow tag on the shared crypto table here as well? This risks corrupting the destination QPN for RoCE traffic as mentioned above.
[CR] Same as above, QPN isn't corrupted, those CQEs don't carry flow_tag.
Review void mlx5e_macsec_offload_handle_rx_skb(struct net_device *netdev, · line 1723
Quoted 9 lines
(quoted diff) +1
-1@@ -1717,7 +1717,7 @@ void mlx5e_macsec_offload_handle_rx_skb(struct net_device *netdev, if (!macsec) return; - fs_id = MLX5_MACSEC_RX_METADAT_HANDLE(macsec_meta_data); + fs_id = MLX5_MACSEC_RX_METADATA_HANDLE(macsec_meta_data); rcu_read_lock(); sc_xarray_element = xa_load(&macsec->sc_xarray, fs_id);
This is a pre-existing issue, but can this xa_load() race with an immediate kfree() during deletion, causing a use-after-free?
In the deletion path, mlx5e_macsec_del_rxsc() calls xa_erase() and then immediately frees the element without an RCU grace period:
mlx5e_macsec_del_rxsc() {
...
xa_erase(&macsec->sc_xarray, rx_sc->sc_xarray_element->fs_id);
dst_release(&rx_sc->md_dst->dst);
kfree(rx_sc->sc_xarray_element);
...
}Since mlx5e_macsec_offload_handle_rx_skb() retrieves sc_xarray_element locklessly under rcu_read_lock() via xa_load(), would a concurrent deletion result in dereferencing freed memory?
[CR] Preexisting.
Review
Quoted 9 lines
(quoted diff) +9
-0+static inline int mlx5e_accel_block_tc_offload(struct mlx5_core_dev *mdev) +{ + int ret = 0; + + mutex_lock(&mdev->offload_block.lock); + if (mdev->offload_block.num_accel) + ret = -EBUSY; + else + mdev->offload_block.num_tc++;
This isn't a bug, but does the new naming here swap the semantics of the variables?
In the original logic, num_block_tc counted the number of blockers on TC (which are IPsec/accel rules). By renaming it to num_tc, mlx5e_accel_block_tc_offload() now increments num_tc when an accel rule is added.
Would it be clearer to keep "block" in the names, e.g., num_block_tc, so it doesn't look like num_tc is tracking the number of TC rules?
[CR] Intentional, "offload_block.num_{tc, accel}" sounds better.
Review
Quoted 9 lines
(quoted diff) +9
-0+static inline int mlx5e_accel_block_tc_offload(struct mlx5_core_dev *mdev) +{ + int ret = 0; + + mutex_lock(&mdev->offload_block.lock); + if (mdev->offload_block.num_accel) + ret = -EBUSY; + else + mdev->offload_block.num_tc++;
Are the semantic meanings of the num_tc and num_accel variables inverted? When an Accel rule is added here in mlx5e_accel_block_tc_offload(), it increments num_tc.
[CR] Same thing, it's an intentional rename.
Quoted 1 line
(quoted diff) +0
-0@@ -384,6 +429,215 @@ static void fill_transportlayer(void *buf, int seq_offset, int ack_offset,
[ ... ]
Quoted 1 line
+static char psp_scratch[L2_HLEN_MAX +
IP_MAXPACKET+PSP_ENCAP_LEN];
[ ... ]
Quoted 5 lines
(quoted diff) +5
-0+/* Encapsulates & encrypts @pkt with PSP transport mode into psp_scratch. + * Returns the scratch buffer and updates *@lenp. + */ +static char *psp_encapsulate(const char *pkt, int *lenp) +{
[ ... ]
Review
Quoted 4 lines
(quoted diff) +4
-0+ memcpy(psp_scratch, pkt, len); + + if (proto == PF_INET) { + struct iphdr *iph = (struct iphdr *)(psp_scratch + ETH_HLEN);
Does this code violate strict aliasing rules?
Since psp_scratch is declared as a character array, casting it to an incompatible structure pointer like struct iphdr * (and later struct ipv6hdr * and struct udphdr *) violates C11 strict aliasing rules.
Because the tools/ directory assumes standard -fstrict-aliasing optimizations are active, could this cause the compiler to incorrectly reorder or optimize away memory writes to these headers, potentially resulting in malformed packets and spurious test failures?
[CR] Maybe, but there are already 20+ similar things in the file.
V4:
- Removed dead mlx5_esw_lock() (Daniel).
- Rebased on top of f225a7317c18 ("selftests: drv-net: split gro.py into one test per coalescing mode") (Daniel).
- Dropped the gro_lib.py extraction, made obsolete by above change.
- Fixed some typos (Aleksandr).
- Added CONFIG_INET_PSP=y to the test config (Daniel).
V3:
https://lore.kernel.org/netdev/[email protected]/
- Cleared fs->decap_enabled on config down (Daniel).
- Made decap support optional (don't fail device reconfig on errors).
- Used bitfield ops for accel protos & psp ver (Daniel).
- Renamed psp_responder off -> len (Jakub).
- Dedicated HW GRO PSP test (Jakub).
- Extended HW GRO test coverage (Jakub).
V2:
https://lore.kernel.org/netdev/[email protected]/
- Use XFail in patch 13 (Jakub).
V1:
https://lore.kernel.org/all/[email protected]/
Cosmin Ratiu (14):
net/mlx5e: Generalize TC <-> IPsec mutual exclusion net/mlx5e: ipsec: Block TC offload when IPsec is enabled net/mlx5e: psp: Block TC offload when PSP is enabled net/mlx5e: macsec: Block TC offload when MACsec is enabled net/mlx5e: psp: Move RX marker from ft_metadata to flow_tag net/mlx5e: ipsec: Move RX marker from ft_metadata to flow_tag net/mlx5e: macsec: Move RX marker from ft_metadata to flow_tag net/mlx5e: psp: Handle HW-decapsulated RX PSP packets net/mlx5e: psp: Add an rx_decap steering table net/mlx5e: shampo: Flush session on PSP mismatch net/mlx5e: psp: Dynamically reconfigure based on SHAMPO mode
selftests: drv-net: psp: Extract shared helpers into psp_lib.py
selftests: net: gro: Add PSP encapsulation and encryption
selftests: drv-net: Add PSP HW GRO conformance tests
.../net/ethernet/mellanox/mlx5/core/en/fs.h | 1 + .../mellanox/mlx5/core/en_accel/en_accel.h | 27 ++ .../mellanox/mlx5/core/en_accel/flow_tag.h | 49 +++ .../mellanox/mlx5/core/en_accel/ipsec_fs.c | 72 ++-- .../mellanox/mlx5/core/en_accel/ipsec_rxtx.h | 8 +- .../mellanox/mlx5/core/en_accel/macsec.c | 34 +- .../mellanox/mlx5/core/en_accel/macsec.h | 4 +- .../mellanox/mlx5/core/en_accel/psp.c | 316 ++++++++++++++-- .../mellanox/mlx5/core/en_accel/psp.h | 2 + .../mellanox/mlx5/core/en_accel/psp_rxtx.c | 21 +- .../mellanox/mlx5/core/en_accel/psp_rxtx.h | 44 ++- .../net/ethernet/mellanox/mlx5/core/en_main.c | 10 +- .../net/ethernet/mellanox/mlx5/core/en_rx.c | 32 +- .../net/ethernet/mellanox/mlx5/core/en_tc.c | 46 ++- .../net/ethernet/mellanox/mlx5/core/en_tc.h | 7 +- .../net/ethernet/mellanox/mlx5/core/eswitch.c | 12 - .../net/ethernet/mellanox/mlx5/core/eswitch.h | 1 - .../mellanox/mlx5/core/lib/macsec_fs.c | 21 +- .../mellanox/mlx5/core/lib/macsec_fs.h | 9 +- .../net/ethernet/mellanox/mlx5/core/main.c | 3 + include/linux/mlx5/driver.h | 7 +- tools/testing/selftests/drivers/net/Makefile | 1 + .../testing/selftests/drivers/net/gro_lib.py | 15 +- .../testing/selftests/drivers/net/hw/Makefile | 17 + tools/testing/selftests/drivers/net/hw/config | 1 + .../selftests/drivers/net/hw/psp_gro.py | 157 ++++++++ tools/testing/selftests/drivers/net/psp.py | 77 ++-- .../testing/selftests/drivers/net/psp_lib.py | 57 +++ tools/testing/selftests/net/lib/Makefile | 16 + tools/testing/selftests/net/lib/gro.c | 337 +++++++++++++++++- 30 files changed, 1198 insertions(+), 206 deletions(-) create mode 100644 drivers/net/ethernet/mellanox/mlx5/core/en_accel/flow_tag.h create mode 100755 tools/testing/selftests/drivers/net/hw/psp_gro.py create mode 100644 tools/testing/selftests/drivers/net/psp_lib.py
base-commit: bc6fe9d301d59fe5cb236d646abe792bffae5fa6
Signature · 1 line
2.44.0
Tariq Toukan
maintainerPATCH v4 1/14 net/mlx5e: Generalize TC <-> IPsec mutual exclusion
From: Cosmin Ratiu <[email protected]>
There is a mechanism to mutually exclude TC offload and IPsec offload from the same interface (commit [1]) due to ordering issues between SW and HW paths.
TC offload makes use of flow_tag to carry the tc mark (even when 0). Upcoming changes to accel protocols will make use of flow_tag to carry the protocol marker. As the flow_tag cannot be partially modified by a steering rule, the last rule setting the flow_tag will overwrite any previous ones.
This means that TC offload cannot be active at the same time with any of the currently implemented accel protocols (IPsec, MACsec, PSP).
Generalize the TC <-> IPsec mutual exclusion mechanism to be usable by more accel protocols:
- move the existing mlx5e_ipsec_{,un}block_tc_offload functions to en_accel.h, rename them to mlx5e_accel_{,un}block_tc_offload.
- rename the mdev counter from num_block_ipsec to num_accel.
- rename is_tc_ipsec_order_check_needed -> is_tc_accel_check_needed and make it not bail out when IPsec isn't configured.
- replace the conditional use of the esw write lock as a protection for incrementing the counter with a new mutex. The esw might not be available, and it wasn't used correctly on the decrement path anyway, allowing races to happen. Using a dedicated mutex for these counters makes it clear and avoids races.
- add underflow warnings for the two counters to catch future miscounting bugs.
- move counters and the new lock into a dedicated struct in mlx5_core_dev named 'offload_block'. Now offload_block.num_tc means the number of tc blocks due to accel rules and ofload_block.num_accel means the number of accel blocks due to tc rules.
[1] commit c8e350e62fc5 ("net/mlx5e: Make TC and IPsec offloads mutually exclusive on a netdev")
--- .../mellanox/mlx5/core/en_accel/en_accel.h | 22 ++++++++ .../mellanox/mlx5/core/en_accel/ipsec_fs.c | 54 +++---------------- .../net/ethernet/mellanox/mlx5/core/en_tc.c | 46 +++++++++------- .../net/ethernet/mellanox/mlx5/core/eswitch.c | 12 ----- .../net/ethernet/mellanox/mlx5/core/eswitch.h | 1 - .../net/ethernet/mellanox/mlx5/core/main.c | 3 ++ include/linux/mlx5/driver.h | 7 ++- 7 files changed, 64 insertions(+), 81 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/en_accel/en_accel.h +22 -0
@@ -218,6 +218,28 @@ static inline void mlx5e_accel_tx_finish(struct mlx5e_txqsq *sq,#endif } +static inline int mlx5e_accel_block_tc_offload (struct mlx5_core_dev *mdev) +{ +int ret =0 ; + +mutex_lock (&mdev->offload_block.lock); +if (mdev->offload_block.num_accel) + ret = -EBUSY ; +else + mdev->offload_block.num_tc++; +mutex_unlock (&mdev->offload_block.lock); + +return ret; +} + +static inline void mlx5e_accel_unblock_tc_offload (struct mlx5_core_dev *mdev) +{ +mutex_lock (&mdev->offload_block.lock); +if (!WARN_ON_ONCE (!mdev->offload_block.num_tc)) + mdev->offload_block.num_tc--; +mutex_unlock (&mdev->offload_block.lock); +} +static inline int mlx5e_accel_init_rx (struct mlx5e_priv *priv) {return mlx5e_ktls_init_rx (priv);
drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c +7 -47
101 lines not on this page · raw message
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c +27 -19
89 lines not on this page · raw message
drivers/net/ethernet/mellanox/mlx5/core/eswitch.c +0 -12
@@ -3022,18 +3022,6 @@ int mlx5_esw_try_lock(struct mlx5_eswitch *esw)return esw->mode; } -int mlx5_esw_lock (struct mlx5_eswitch *esw) -{ -down_write (&esw->mode_lock); - -if (esw->eswitch_operation_in_progress) { -up_write (&esw->mode_lock); -return -EBUSY ; - } - -return 0 ; -} -/** * mlx5_esw_unlock() - Release write lock on esw mode lock * @esw: eswitch device.
drivers/net/ethernet/mellanox/mlx5/core/eswitch.h +0 -1
@@ -946,7 +946,6 @@ void mlx5_esw_release(struct mlx5_core_dev *dev);void mlx5_esw_get (struct mlx5_core_dev *dev);void mlx5_esw_put (struct mlx5_core_dev *dev);int mlx5_esw_try_lock (struct mlx5_eswitch *esw); -int mlx5_esw_lock (struct mlx5_eswitch *esw);void mlx5_esw_unlock (struct mlx5_eswitch *esw);void esw_vport_change_handle_locked (struct mlx5_vport *vport);
drivers/net/ethernet/mellanox/mlx5/core/main.c +3 -0
@@ -1810,6 +1810,7 @@ int mlx5_mdev_init(struct mlx5_core_dev *dev, int profile_idx)lockdep_register_key (&dev->lock_key);mutex_init (&dev->intf_state_mutex);lockdep_set_class (&dev->intf_state_mutex, &dev->lock_key); +mutex_init (&dev->offload_block.lock);mutex_init (&dev->mlx5e_res.uplink_netdev_lock);mutex_init (&dev->wc_state_lock);
@@ -1901,6 +1902,7 @@ int mlx5_mdev_init(struct mlx5_core_dev *dev, int profile_idx)mutex_destroy (&priv->alloc_mutex);mutex_destroy (&priv->bfregs.wc_head.lock);mutex_destroy (&priv->bfregs.reg_head.lock); +mutex_destroy (&dev->offload_block.lock);mutex_destroy (&dev->intf_state_mutex);lockdep_unregister_key (&dev->lock_key);return err;
@@ -1928,6 +1930,7 @@ void mlx5_mdev_uninit(struct mlx5_core_dev *dev)mutex_destroy (&priv->bfregs.reg_head.lock);mutex_destroy (&dev->wc_state_lock);mutex_destroy (&dev->mlx5e_res.uplink_netdev_lock); +mutex_destroy (&dev->offload_block.lock);mutex_destroy (&dev->intf_state_mutex);lockdep_unregister_key (&dev->lock_key); }
include/linux/mlx5/driver.h +5 -2
@@ -788,8 +788,11 @@ struct mlx5_core_dev {u32 vsc_addr;struct mlx5_hv_vhca *hv_vhca;struct mlx5_hwmon *hwmon; -u64 num_block_tc; -u64 num_block_ipsec; +struct { +struct mutex lock; +u64 num_tc; +u64 num_accel; + } offload_block;#ifdef CONFIG_MLX5_MACSEC struct mlx5_macsec_fs *macsec_fs;/* MACsec notifier chain to sync MACsec core and IB database */
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 01/14] net/mlx5e: Generalize TC <-> IPsec mutual exclusion
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 15 lines
From: Cosmin Ratiu <[email protected]>
There is a mechanism to mutually exclude TC offload and IPsec offload from the same interface (commit [1]) due to ordering issues between SW and HW paths.
TC offload makes use of
flow_tagto carry the tc mark (even when 0). Upcoming changes to accel protocols will make use offlow_tagto carry the protocol marker. As theflow_tagcannot be partially modified by a steering rule, the last rule setting theflow_tagwill overwrite any previous ones.This means that TC offload cannot be active at the same time with any of the currently implemented accel protocols (IPsec, MACsec, PSP).
Quoted 6 lines
[ ... ]
Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Carolina Jubran <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
[email protected]
AI reviewRe: [PATCH net-next V4 01/14] net/mlx5e: Generalize TC <-> IPsec mutual exclusion
In reply to Tariq ToukanThank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Critical: 0 · High: 0 (1) · Medium: 1 (1) · Low: 1
- [Medium] Silent removal of the eswitch-operation gate on the accel (IPsec packet-offload) rule-installation path.
- [Low] The patch fixes a real, pre-existing race on the TC/accel mutual- exclusion counters (both decrement paths were previously completely…
Pre-existing issues:
- [High] offload_block.num_accel is leaked on the driver's own flow- teardown path, permanently blocking all accel offload on the device.
- [Medium] Both unblock paths drop the counter before the corresponding hardware rules are actually gone, so the exclusion the mechanism…
--
net/mlx5e: Generalize TC <-> IPsec mutual exclusion
This generalizes the existing TC <-> IPsec mutual exclusion so other accel protocols can use it: the helpers move to en_accel.h, the two mdev counters move into a new offload_block struct, and the conditional eswitch write lock is replaced by a dedicated mutex. Underflow warnings are added on both decrement paths.
A few questions below.
The commit message says:
- replace the conditional use of the esw write lock as a protection for incrementing the counter with a new mutex. The esw might not be available, and it wasn't used correctly on the decrement path anyway, allowing races to happen. Using a dedicated mutex for these counters makes it clear and avoids races.
[Severity: Low]
This paragraph describes an actual race fix, but the patch carries no
Fixes: tag; c8e350e62fc5 is only cited as background.
Before this patch both decrement sites were completely unserialized:
static void mlx5e_ipsec_unblock_tc_offload(struct mlx5_core_dev *mdev)
{
mdev->num_block_tc--;
}and
priv->mdev->num_block_ipsec--;
the latter reached from mlx5e_delete_flower(), which takes no esw lock at all. Two concurrent unblocks on the same mlx5_core_dev (two xfrm SA deletions, or an SA delete concurrent with tc filter del) can lose an update and leave the counters wrong.
Would a Fixes: tag pointing at c8e350e62fc5 be appropriate here? As it stands the race fix is entangled with the renames, the move to en_accel.h and the new anonymous struct in include/linux/mlx5/driver.h, so it cannot be picked up for stable as-is. Would splitting the locking fix from the rename/move make that possible?
Quoted 33 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/en_accel.h +22
-0@@ -218,6 +218,28 @@ static inline void mlx5e_accel_tx_finish(struct mlx5e_txqsq *sq,#endif } +static inline int mlx5e_accel_block_tc_offload (struct mlx5_core_dev *mdev) +{ +int ret =0 ; + +mutex_lock (&mdev->offload_block.lock); +if (mdev->offload_block.num_accel) + ret = -EBUSY ; +else + mdev->offload_block.num_tc++; +mutex_unlock (&mdev->offload_block.lock); + +return ret; +} + +static inline void mlx5e_accel_unblock_tc_offload (struct mlx5_core_dev *mdev) +{ +mutex_lock (&mdev->offload_block.lock); +if (!WARN_ON_ONCE (!mdev->offload_block.num_tc)) + mdev->offload_block.num_tc--; +mutex_unlock (&mdev->offload_block.lock); +} +static inline int mlx5e_accel_init_rx (struct mlx5e_priv *priv) {return mlx5e_ktls_init_rx (priv);
[ ... ]
Quoted 34 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c +0
-26@@ -2574,53 +2575,12 @@ void mlx5e_accel_ipsec_fs_read_stats(struct mlx5e_priv *priv, void *ipsec_stats) } } -#ifdef CONFIG_MLX5_ESWITCH -static int mlx5e_ipsec_block_tc_offload (struct mlx5_core_dev *mdev) -{ -struct mlx5_eswitch *esw = mdev->priv.eswitch; -int err =0 ; - -if (esw) { - err =mlx5_esw_lock (esw); -if (err) -return err; - } - -if (mdev->num_block_ipsec) { - err = -EBUSY ; - goto unlock; - } - - mdev->num_block_tc++; - -unlock: -if (esw) -mlx5_esw_unlock (esw); - -return err; -} -#else
[ ... ]
Quoted 10 lines
int
mlx5e_accel_ipsec_fs_add_rule(struct mlx5e_ipsec_sa_entry *sa_entry){drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c +1
-1int err;if (sa_entry->attrs.type ==XFRM_DEV_OFFLOAD_PACKET ) { - err =mlx5e_ipsec_block_tc_offload (sa_entry->ipsec->mdev); + err =mlx5e_accel_block_tc_offload (sa_entry->ipsec->mdev);if (err)return err; }
[ ... ]
Quoted 9 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c +1
-1@@ -2652,7 +2612,7 @@ void mlx5e_accel_ipsec_fs_del_rule(struct mlx5e_ipsec_sa_entry *sa_entry)mlx5_packet_reformat_dealloc (mdev, ipsec_rule->pkt_reformat);if (sa_entry->attrs.type ==XFRM_DEV_OFFLOAD_PACKET ) -mlx5e_ipsec_unblock_tc_offload (mdev); +mlx5e_accel_unblock_tc_offload (mdev);if (sa_entry->attrs.dir ==XFRM_DEV_OFFLOAD_OUT ) {tx_ft_put (sa_entry->ipsec, sa_entry->attrs.type);
[ ... ]
Quoted 19 lines
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c +4
-2@@ -4830,27 +4830,35 @@ static bool is_tc_ipsec_order_check_needed(struct net_device *filter, struct mlxreturn true; } -static int mlx5e_tc_block_ipsec_offload (struct net_device *filter,struct mlx5e_priv *priv) +static int mlx5e_tc_block_accel_offload (struct net_device *filter, +struct mlx5e_priv *priv) {struct mlx5_core_dev *mdev = priv->mdev; +int ret =0 ; -if (!is_tc_ipsec_order_check_needed (filter, priv)) +if (!is_tc_accel_check_needed (filter, priv))return 0 ;
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c static bool is_tc_ipsec_order_check_needed(struct net_device *filter, struct mlx · line 4847
Quoted 10 lines
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c +6
-4-if (mdev->num_block_tc) -return -EBUSY ; - - mdev->num_block_ipsec++; +mutex_lock (&mdev->offload_block.lock); +if (mdev->offload_block.num_tc) + ret = -EBUSY ; +else + mdev->offload_block.num_accel++; +mutex_unlock (&mdev->offload_block.lock);
[Severity: High]
This isn't a bug introduced by this patch, since the same asymmetry existed with num_block_ipsec, but can offload_block.num_accel be leaked when the driver tears down its own flows?
The increment here is only undone by mlx5e_tc_unblock_accel_offload() from mlx5e_delete_flower() or the mlx5e_configure_flower() error label. Flows destroyed by the driver go through the bulk destructor instead:
static void _mlx5e_tc_del_flow(void *ptr, void *arg)
{
struct mlx5e_tc_flow *flow = ptr;
struct mlx5e_priv *priv = flow->priv;
mlx5e_tc_del_flow(priv, flow);
kfree(flow);
}which is passed to rhashtable_free_and_destroy() from mlx5e_tc_nic_cleanup() and mlx5e_tc_ht_cleanup(), and never decrements the counter.
mlx5e_tc_nic_cleanup() is reachable without unregistering the netdev:
_mlx5e_suspend() -> mlx5e_detach_netdev() -> mlx5e_cleanup_nic_rx() -> mlx5e_tc_nic_cleanup() -> rhashtable_free_and_destroy(&tc->ht,
_mlx5e_tc_del_flow, NULL)
i.e. devlink dev reload, PCI suspend and FW-reset/health recovery. On those paths the cls_flower filters are still installed and still in tc->ht, so they were never removed via mlx5e_delete_flower().
Since offload_block lives on mlx5_core_dev and nothing else in the tree resets num_tc/num_accel, wouldn't num_accel stay above zero for the lifetime of the driver instance, making every later mlx5e_accel_block_tc_offload() return -EBUSY even though no TC rule exists? That would refuse IPsec packet offload, and after the rest of this series also MACsec and PSP enablement, until the module is reloaded.
The new WARN_ON_ONCE() checks only catch the opposite direction.
Would calling mlx5e_tc_unblock_accel_offload() (or an mdev-only variant of it) from _mlx5e_tc_del_flow(), or resetting the counters on profile detach, be the right place to close this?
Quoted 19 lines
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c +8
-4-return 0 ; +return ret; } -static void mlx5e_tc_unblock_ipsec_offload (struct net_device *filter,struct mlx5e_priv *priv) +static void mlx5e_tc_unblock_accel_offload (struct net_device *filter, +struct mlx5e_priv *priv) { -if (!is_tc_ipsec_order_check_needed (filter, priv)) +if (!is_tc_accel_check_needed (filter, priv))return ; - priv->mdev->num_block_ipsec--; +mutex_lock (&priv->mdev->offload_block.lock); +if (!WARN_ON_ONCE (!priv->mdev->offload_block.num_accel)) + priv->mdev->offload_block.num_accel--; +mutex_unlock (&priv->mdev->offload_block.lock); }
[ ... ]
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c int mlx5e_delete_flower(struct net_device *dev, struct mlx5e_priv *priv, · line 4968
Quoted 6 lines
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c +1
-1@@ -4957,7 +4965,7 @@ int mlx5e_delete_flower(struct net_device *dev, struct mlx5e_priv *priv,trace_mlx5e_delete_flower (f);mlx5e_flow_put (priv, flow); -mlx5e_tc_unblock_ipsec_offload (dev, priv); +mlx5e_tc_unblock_accel_offload (dev, priv);
[Severity: Medium]
This is a pre-existing ordering, unchanged by the patch (the old mlx5e_tc_unblock_ipsec_offload() call sat in the same place), but is the counter dropped before the hardware rules are actually gone?
mlx5e_flow_put() only removes the rules when the last reference goes away:
void mlx5e_flow_put(struct mlx5e_priv *priv, struct mlx5e_tc_flow *flow)
{
if (refcount_dec_and_test(&flow->refcnt)) {
mlx5e_tc_del_flow(priv, flow);
kfree_rcu(flow, rcu_head);
}
}Other contexts legitimately hold references via mlx5e_flow_get() in en/tc_tun_encap.c (encap/neigh update work) and in mlx5e_stats_flower(). With refcnt still above one, the TC rules remain in hardware while mlx5e_tc_unblock_accel_offload() sets num_accel back to 0, so a concurrent
ip xfrm state add ... offload packet dev <uplink>
can pass mlx5e_accel_block_tc_offload() and install its flow_tag rule next to the still-live TC flow_tag rule, which is the overwrite the commit message says must not happen.
The same shape exists in mlx5e_accel_ipsec_fs_del_rule(), where mlx5e_accel_unblock_tc_offload() runs while the trailer/auth/sa_sel/replay rules and the RX/TX steering tables are still installed.
Quoted 13 lines
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c +0
-0mlx5_esw_put (priv->mdev);return 0 ;drivers/net/ethernet/mellanox/mlx5/core/eswitch.c +0
-2@@ -3022,18 +3022,6 @@ int mlx5_esw_try_lock(struct mlx5_eswitch *esw)return esw->mode; } -int mlx5_esw_lock (struct mlx5_eswitch *esw) -{
drivers/net/ethernet/mellanox/mlx5/core/eswitch.c int mlx5_esw_try_lock(struct mlx5_eswitch *esw) · line 3036
Quoted 10 lines
drivers/net/ethernet/mellanox/mlx5/core/eswitch.c +0
-10-down_write (&esw->mode_lock); - -if (esw->eswitch_operation_in_progress) { -up_write (&esw->mode_lock); -return -EBUSY ; - } - -return 0 ; -} -
[Severity: Medium]
Does removing this also remove the eswitch-transition gate on the accel side, and not just a counter lock?
mlx5_esw_lock() returned -EBUSY whenever esw->eswitch_operation_in_progress was set, which aborted mlx5e_accel_ipsec_fs_add_rule() and mlx5e_accel_ipsec_fs_add_pol() for the whole duration of a devlink eswitch mode / inline-mode / encap-mode transition. That check was added on purpose by commit baac8351f74c5 ("net/mlx5e: Reduce eswitch mode_lock protection context"), which carried Fixes: c8e350e62fc5, and it mirrored the check added to mlx5_esw_hold() for the TC path.
The replacement only takes the new private mutex:
mutex_lock(&mdev->offload_block.lock);
if (mdev->offload_block.num_accel)
ret = -EBUSY;
else
mdev->offload_block.num_tc++;
mutex_unlock(&mdev->offload_block.lock);which the eswitch code knows nothing about. So
ip xfrm state add ... offload packet dev <uplink>
can now proceed into rx_add_rule()/tx_add_rule() and create tables and rules in the FDB namespace (ipsec_fs_get_ns() returns MLX5_FLOW_NAMESPACE_FDB for XFRM_DEV_OFFLOAD_PACKET on an uplink rep) while mlx5_devlink_eswitch_encap_mode_set() runs esw_destroy_offloads_fdb_tables()/esw_create_offloads_fdb_tables() with mode_lock dropped.
The commit message describes the esw lock only as "protection for incrementing the counter" and doesn't mention the eswitch_operation_in_progress -> -EBUSY behaviour going away. Was that intentional? If so, could the commit message spell out why the gate is no longer needed?
Quoted 3 lines
drivers/net/ethernet/mellanox/mlx5/core/eswitch.c +0
-0/** * mlx5_esw_unlock() - Release write lock on esw mode lock * @esw: eswitch device.
[ ... ]
Signature · 1 line
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260908065244.3799142-1-tariqt%40nvidia.com
Jakub Kicinski
maintainerRe: [PATCH net-next V4 01/14] net/mlx5e: Generalize TC <-> IPsec mutual exclusion
Quoted 3 lines · On Thu, 10 Sep 2026 21:54:25 +0000 [email protected] wrote:
Pre-existing issues:
- [High] offload_block.num_accel is leaked on the driver's own flow- teardown path, permanently blocking all accel offload on the device.
Does this one need to be fixed in net first ?
Tariq Toukan
maintainerPATCH v4 2/14 net/mlx5e: ipsec: Block TC offload when IPsec is enabled
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
Currently the mutual exclusion mechanism is only used in packet offload mode. But with the upcoming changes to flow_tag, all RX IPsec flows need to block TC offload, so do that with the help of a small helper.
--- .../ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c | 13 ++++++++++--- 1 file changed, 10 insertions(+), 3 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c +10 -3
@@ -2575,11 +2575,18 @@ void mlx5e_accel_ipsec_fs_read_stats(struct mlx5e_priv *priv, void *ipsec_stats) } } +static bool accel_ipsec_should_block_tc (struct mlx5e_ipsec_sa_entry *sa_entry) +{ +return sa_entry->attrs.type ==XFRM_DEV_OFFLOAD_PACKET || + sa_entry->attrs.dir ==XFRM_DEV_OFFLOAD_IN ; +} +int mlx5e_accel_ipsec_fs_add_rule (struct mlx5e_ipsec_sa_entry *sa_entry) { +bool block_tc =accel_ipsec_should_block_tc (sa_entry);int err; -if (sa_entry->attrs.type ==XFRM_DEV_OFFLOAD_PACKET ) { +if (block_tc) { err =mlx5e_accel_block_tc_offload (sa_entry->ipsec->mdev);if (err)return err;
@@ -2596,7 +2603,7 @@ int mlx5e_accel_ipsec_fs_add_rule(struct mlx5e_ipsec_sa_entry *sa_entry)return 0 ; err_out: -if (sa_entry->attrs.type ==XFRM_DEV_OFFLOAD_PACKET ) +if (block_tc)mlx5e_accel_unblock_tc_offload (sa_entry->ipsec->mdev);return err; }
@@ -2611,7 +2618,7 @@ void mlx5e_accel_ipsec_fs_del_rule(struct mlx5e_ipsec_sa_entry *sa_entry)if (ipsec_rule->pkt_reformat)mlx5_packet_reformat_dealloc (mdev, ipsec_rule->pkt_reformat); -if (sa_entry->attrs.type ==XFRM_DEV_OFFLOAD_PACKET ) +if (accel_ipsec_should_block_tc (sa_entry))mlx5e_accel_unblock_tc_offload (mdev);if (sa_entry->attrs.dir ==XFRM_DEV_OFFLOAD_OUT ) {
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 02/14] net/mlx5e: ipsec: Block TC offload when IPsec is enabled
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 10 lines
From: Cosmin Ratiu <[email protected]>
Currently the mutual exclusion mechanism is only used in packet offload mode. But with the upcoming changes to
flow_tag, all RX IPsec flows need to block TC offload, so do that with the help of a small helper.Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Carolina Jubran <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2
Tariq Toukan
maintainerPATCH v4 3/14 net/mlx5e: psp: Block TC offload when PSP is enabled
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
Make use of the mutual exclusion mechanism from the previous patch to make sure only PSP or TC offload is active on a device. Later in the series, the PSP protocol marker will move to the flow_tag, which is also used by TC offload.
--- .../net/ethernet/mellanox/mlx5/core/en_accel/psp.c | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-)
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +13 -1
@@ -6,6 +6,7 @@#include "mlx5_core.h" #include "psp.h" #include "lib/crypto.h" +#include "en_accel/en_accel.h" #include "en_accel/psp.h" #include "fs_core.h"
@@ -526,6 +527,7 @@ accel_psp_fs_rx_decrypt_ft_create(struct mlx5e_psp_fs *fs,static void accel_psp_fs_rx_destroy (struct mlx5e_psp_fs *fs) {struct mlx5_ttc_table *ttc =mlx5e_fs_get_ttc (fs->fs, false); +bool tc_blocked = fs->rx.ft;int i;/* disconnect */
@@ -535,6 +537,8 @@ static void accel_psp_fs_rx_destroy(struct mlx5e_psp_fs *fs) }accel_psp_fs_rx_check_ft_destroy (&fs->check);accel_psp_fs_rx_ft_destroy (&fs->rx); +if (tc_blocked) +mlx5e_accel_unblock_tc_offload (fs->mdev); } static int accel_psp_fs_rx_create(struct mlx5e_psp_fs *fs,
@@ -543,10 +547,16 @@ static int accel_psp_fs_rx_create(struct mlx5e_psp_fs *fs,struct mlx5_ttc_table *ttc =mlx5e_fs_get_ttc (fs->fs, false);int i, err; + err =mlx5e_accel_block_tc_offload (fs->mdev); +if (err) { +NL_SET_ERR_MSG (extack,"TC offload active, cannot enable PSP" ); +return err; + } + err =accel_psp_fs_rx_ft_create (fs, &fs->rx);if (err) {NL_SET_ERR_MSG (extack,"Failed creating RX steering table" ); -return err; + goto err_unblock_tc; } err =accel_psp_fs_rx_check_ft_create (fs, &fs->check);
@@ -583,6 +593,8 @@ static int accel_psp_fs_rx_create(struct mlx5e_psp_fs *fs,accel_psp_fs_rx_check_ft_destroy (&fs->check); err_ft:accel_psp_fs_rx_ft_destroy (&fs->rx); +err_unblock_tc: +mlx5e_accel_unblock_tc_offload (fs->mdev);return err; }
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 03/14] net/mlx5e: psp: Block TC offload when PSP is enabled
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 11 lines
From: Cosmin Ratiu <[email protected]>
Make use of the mutual exclusion mechanism from the previous patch to make sure only PSP or TC offload is active on a device. Later in the series, the PSP protocol marker will move to the
flow_tag, which is also used by TC offload.Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Carolina Jubran <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=3
Tariq Toukan
maintainerPATCH v4 4/14 net/mlx5e: macsec: Block TC offload when MACsec is enabled
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
The MACsec protocol marker will soon move to flow_tag and that will make TC offload unusable at the same time as MACsec on the same device.
This patch makes use of the mutual exclusion mechanism to make sure that TC and MACsec cannot be both active at the same time.
One extra bit of logic is in macsec_upd_secy_hw_address(), where existing macsec rules are drained then readded. During the two loops it's possible a mistimed TC filter add to throw a wrench into things and prevent the 2nd loop from adding anything, since accel rules are now blocked. Fix that by keeping a best-effort TC block across the entire operation.
--- .../mellanox/mlx5/core/en_accel/macsec.c | 28 +++++++++++++++++-- 1 file changed, 26 insertions(+), 2 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.c +26 -2
@@ -10,6 +10,7 @@#include "en.h" #include "lib/aso.h" #include "lib/crypto.h" +#include "en_accel/en_accel.h" #include "en_accel/macsec.h" #define MLX5_MACSEC_EPN_SCOPE_MID 0x80000000L
@@ -324,6 +325,8 @@ static void mlx5e_macsec_cleanup_sa_fs(struct mlx5e_macsec *macsec,mlx5_macsec_fs_del_rule (macsec->mdev->macsec_fs, sa->macsec_rule, action, netdev, fs_id); sa->macsec_rule =NULL ; +if (!is_tx) +mlx5e_accel_unblock_tc_offload (macsec->mdev); } static void mlx5e_macsec_cleanup_sa(struct mlx5e_macsec *macsec,
@@ -343,6 +346,7 @@ static int mlx5e_macsec_init_sa_fs(struct macsec_context *ctx,const struct macsec_tx_sc *tx_sc = &ctx->secy->tx_sc;struct mlx5_macsec_rule_attrs rule_attrs;union mlx5_macsec_rule *macsec_rule; +int err =0 ;if (is_tx && tx_sc->encoding_sa != sa->assoc_num)return 0 ;
@@ -353,13 +357,26 @@ static int mlx5e_macsec_init_sa_fs(struct macsec_context *ctx, rule_attrs.action = (is_tx) ?MLX5_ACCEL_MACSEC_ACTION_ENCRYPT :MLX5_ACCEL_MACSEC_ACTION_DECRYPT ; +if (!is_tx) { + err =mlx5e_accel_block_tc_offload (priv->mdev); +if (err) +return err; + } + macsec_rule =mlx5_macsec_fs_add_rule (macsec_fs, ctx, &rule_attrs, fs_id); -if (!macsec_rule) -return -ENOMEM ; +if (!macsec_rule) { + err = -ENOMEM ; + goto out_unblock_tc; + } sa->macsec_rule = macsec_rule;return 0 ; + +out_unblock_tc: +if (!is_tx) +mlx5e_accel_unblock_tc_offload (priv->mdev); +return err; } static int mlx5e_macsec_init_sa(struct macsec_context *ctx,
@@ -1137,7 +1154,12 @@ static int macsec_upd_secy_hw_address(struct macsec_context *ctx,struct mlx5e_macsec_sa *rx_sa;struct list_head *list;int i, err =0 ; +bool block_tc; +/* Best-effort TC block across the operation, to prevent a mistimed TC +* filter add from preventing the 2nd loop from happening. +*/ + block_tc =mlx5e_accel_block_tc_offload (priv->mdev) ==0 ; list = &macsec_device->macsec_rx_sc_list_head;list_for_each_entry_safe (rx_sc, tmp, list, rx_sc_list_element) {
@@ -1168,6 +1190,8 @@ static int macsec_upd_secy_hw_address(struct macsec_context *ctx,memcpy (macsec_device->dev_addr, dev->dev_addr, dev->addr_len); out: +if (block_tc) +mlx5e_accel_unblock_tc_offload (priv->mdev);return err; }
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 04/14] net/mlx5e: macsec: Block TC offload when MACsec is enabled
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 15 lines
From: Cosmin Ratiu <[email protected]>
The MACsec protocol marker will soon move to
flow_tagand that will make TC offload unusable at the same time as MACsec on the same device.This patch makes use of the mutual exclusion mechanism to make sure that TC and MACsec cannot be both active at the same time.
One extra bit of logic is in
macsec_upd_secy_hw_address(), where existing macsec rules are drained then readded. During the two loops it's possible a mistimed TC filter add to throw a wrench into things and prevent the 2nd loop from adding anything, since accel rules are now blocked. Fix that by keeping a best-effort TC block across the entire operation.
Quoted 4 lines
Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Carolina Jubran <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=4
[email protected]
AI reviewRe: [PATCH net-next V4 04/14] net/mlx5e: macsec: Block TC offload when MACsec is enabled
In reply to Tariq ToukanThank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Critical: 0 · High: 1 · Medium: 0 · Low: 1
- [High] Unbalanced TC-block reference across the MACsec teardown path:
mlx5e_macsec_init_sa_fs() now takes one mdev->offload_block.num_tc…
- [Low] Commit-message scope vs. implementation: the message states the patch makes "sure that TC and MACsec cannot be both active at the same…
--
net/mlx5e: macsec: Block TC offload when MACsec is enabled
The patch takes a reference on the mlx5 TC/accel mutual exclusion counter whenever a MACsec RX (decrypt) steering rule is installed, and drops it when that rule is removed. It also holds a best-effort block across macsec_upd_secy_hw_address() so the drain/re-add loops are not disturbed by a concurrent TC filter add.
Quoted 4 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.c +0
-0
[ ... ]
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.c static void mlx5e_macsec_cleanup_sa_fs(struct mlx5e_macsec *macsec, · line 330
Quoted 7 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.c +2
-0@@ -324,6 +325,8 @@ static void mlx5e_macsec_cleanup_sa_fs(struct mlx5e_macsec *macsec,mlx5_macsec_fs_del_rule (macsec->mdev->macsec_fs, sa->macsec_rule, action, netdev, fs_id); sa->macsec_rule =NULL ; +if (!is_tx) +mlx5e_accel_unblock_tc_offload (macsec->mdev); }
[Severity: High]
This is the only place that gives back the count taken in mlx5e_macsec_init_sa_fs(). Is mlx5e_macsec_cleanup() a second teardown path that destroys the same RX rules without passing through here?
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.c:mlx5e_macsec_cleanup() {
mlx5_notifier_unregister(mdev, &macsec->nb);
mlx5_macsec_fs_cleanup(mdev->macsec_fs);
destroy_workqueue(macsec->wq);
mlx5e_macsec_aso_cleanup(&macsec->aso, mdev);
mutex_destroy(&macsec->lock);
kfree(macsec);
}It never walks macsec_device_list_head / rx_sc->rx_sa[] and never calls mlx5e_accel_unblock_tc_offload(), so wouldn't every RX SA rule still installed at that moment leave mdev->offload_block.num_tc permanently incremented?
The path looks reachable without any mdo_del_rxsa/mdo_del_secy first:
mlx5_unload_one()/mlx5e_suspend() -> _mlx5e_suspend() -> mlx5e_detach_netdev() -> mlx5e_nic_profile.disable == mlx5e_nic_disable() -> mlx5e_macsec_cleanup()
In those flows (system suspend, devlink reload, health recovery) the lower netdev is only netif_device_detach()ed, not unregistered, so the stacked MACsec devices and their offloaded RX SAs are still live. mdev->offload_block is initialized once in mlx5_mdev_init() and is not reset by unload/load.
After such an event, does every later TC flower add on that device fail forever?
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c:mlx5e_tc_block_accel_offload() {
mutex_lock(&mdev->offload_block.lock);
if (mdev->offload_block.num_tc)
ret = -EBUSY;
...
}The WARN_ON_ONCE() in mlx5e_accel_unblock_tc_offload() would not fire here because the leak is a missing decrement, so there is no diagnostic either. Would it be more robust to drop the reference in mlx5e_macsec_cleanup(), or to hold it per MACsec device in add_secy/del_secy instead of per RX rule?
Quoted 11 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.c +1
-0static void mlx5e_macsec_cleanup_sa(struct mlx5e_macsec *macsec,@@ -343,6 +346,7 @@ static int mlx5e_macsec_init_sa_fs(struct macsec_context *ctx,const struct macsec_tx_sc *tx_sc = &ctx->secy->tx_sc;struct mlx5_macsec_rule_attrs rule_attrs;union mlx5_macsec_rule *macsec_rule; +int err =0 ;if (is_tx && tx_sc->encoding_sa != sa->assoc_num)return 0 ;
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.c static int mlx5e_macsec_init_sa_fs(struct macsec_context *ctx, · line 365
Quoted 9 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.c +6
-0@@ -353,13 +357,26 @@ static int mlx5e_macsec_init_sa_fs(struct macsec_context *ctx, rule_attrs.action = (is_tx) ?MLX5_ACCEL_MACSEC_ACTION_ENCRYPT :MLX5_ACCEL_MACSEC_ACTION_DECRYPT ; +if (!is_tx) { + err =mlx5e_accel_block_tc_offload (priv->mdev); +if (err) +return err; + } +
[Severity: Low]
This isn't a bug, but the commit message says the patch makes
"sure that TC and MACsec cannot be both active at the same time"
while the block here is taken only for RX rules (!is_tx), and only for as long as an RX rule exists. A device configured with just an active TX SA goes through mlx5e_macsec_add_txsa() -> mlx5e_macsec_init_sa() -> mlx5e_macsec_init_sa_fs() with is_tx=true and never touches mdev->offload_block.num_tc, so TC flower filters are still accepted while MACsec offload is active.
That scope looks functionally sufficient, since the follow-up patch "net/mlx5e: macsec: Move RX marker from ft_metadata to flow_tag" only sets FLOW_CONTEXT_HAS_TAG/flow_tag on the RX crypto rules in macsec_fs_rx_add_rule(), the TX path keeps using MLX5_MACSEC_TX_METADATA, and tc_act_can_offload_mark() limits TC marks to 16 bits so a TC filter cannot forge the proto bits. Could the commit message state the RX-only scope explicitly, the way the sibling IPsec patch does?
Related to the same narrow scope: because the block is per RX rule, a TC filter can be installed during any window with no RX rule present (all RX SAs inactive, or after a partially failed macsec_upd_secy_hw_address()). The next RX SA activation then returns a bare -EBUSY from mlx5e_macsec_init_sa_fs() with no netdev_err() or extack, unlike the other error paths in this file. Would adding a message here help users understand the failure?
[ ... ]
Signature · 1 line
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260908065244.3799142-1-tariqt%40nvidia.com
Tariq Toukan
maintainerPATCH v4 5/14 net/mlx5e: psp: Move RX marker from ft_metadata to flow_tag
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
Move PSP RX marker from ft_metadata (set via modify_hdr action) to flow_tag (set via flow_context). This frees ft_metadata for storing SPI for future decapsulated PSP packets.
All mlx5e accel protos have to move to avoid misinterpreting packets with high bits of ft_metadata as IPsec or Macsec. This patch is the first step, defining a new header, bit layout and macros.
--- .../mellanox/mlx5/core/en_accel/flow_tag.h | 41 +++++++++++++++++++ .../mellanox/mlx5/core/en_accel/psp.c | 36 ++++------------ .../mellanox/mlx5/core/en_accel/psp_rxtx.h | 10 ++--- 3 files changed, 55 insertions(+), 32 deletions(-) create mode 100644 drivers/net/ethernet/mellanox/mlx5/core/en_accel/flow_tag.h
drivers/net/ethernet/mellanox/mlx5/core/en_accel/flow_tag.h new file +41 -0
@@ -0,0 +1,41 @@ +/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */ +/* Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved. */ + +#ifndef __MLX5E_FLOW_TAG_H__ +#define __MLX5E_FLOW_TAG_H__ + +#include <linux/bitfield.h> +#include <linux/bits.h> +#include <linux/kernel.h> +#include <linux/types.h> +#include <linux/mlx5/device.h> + +/* Unified accel flow_tag layout in CQE sop_drop_qpn [23:0]: +* +* [23:21] = protocol ID (3 bits): +* 0 = none (default) +* 3 = PSP (HW decrypted, PSP header present) +* 1,2,4-7 = reserved +* [20:16] = reserved +* [15:0] = used by other subsystems (e.g. TC). +*/ +#define MLX5E_ACCEL_FLOW_TAG_MASK GENMASK(23, 0) +#define MLX5E_ACCEL_FLOW_TAG_PROTO_MASK GENMASK(23, 21) + +enum { +MLX5E_ACCEL_FLOW_TAG_PROTO_NONE =0 , +MLX5E_ACCEL_FLOW_TAG_PROTO_PSP =3 , +}; + +static inline u32 mlx5e_accel_flow_tag (struct mlx5_cqe64 *cqe) +{ +return be32_to_cpu (cqe->sop_drop_qpn) &MLX5E_ACCEL_FLOW_TAG_MASK ; +} + +static inline u32 mlx5e_accel_flow_tag_proto (struct mlx5_cqe64 *cqe) +{ +return FIELD_GET (MLX5E_ACCEL_FLOW_TAG_PROTO_MASK , +mlx5e_accel_flow_tag (cqe)); +} + +#endif /* __MLX5E_FLOW_TAG_H__ */
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +9 -27
@@ -8,6 +8,7 @@#include "lib/crypto.h" #include "en_accel/en_accel.h" #include "en_accel/psp.h" +#include "en_accel/psp_rxtx.h" #include "fs_core.h" enum accel_fs_psp_type {
@@ -42,7 +43,6 @@ struct mlx5e_psp_rx_decrypt_table {struct mlx5_flow_table *ft;struct mlx5_flow_group *miss_group;struct mlx5_flow_handle *miss_rule; -struct mlx5_modify_hdr *rx_modify_hdr;struct mlx5_flow_handle *rule; };
@@ -408,10 +408,6 @@ accel_psp_fs_rx_decrypt_ft_destroy(struct mlx5e_psp_fs *fs, struct mlx5e_psp_rx_decrypt_table *decrypt) {accel_psp_fs_del_flow_rule (&decrypt->rule); -if (decrypt->rx_modify_hdr) { -mlx5_modify_header_dealloc (fs->mdev, decrypt->rx_modify_hdr); - decrypt->rx_modify_hdr =NULL ; - }accel_psp_fs_del_flow_rule (&decrypt->miss_rule);accel_psp_fs_destroy_flow_group (&decrypt->miss_group);accel_psp_fs_destroy_ft (&decrypt->ft);
@@ -431,8 +427,6 @@ accel_psp_fs_rx_decrypt_ft_create(struct mlx5e_psp_fs *fs,struct mlx5e_psp_rx_decrypt_table *decrypt, struct mlx5_flow_destination *default_dest) { - u8 action[MLX5_UN_SZ_BYTES(set_add_copy_action_in_auto)] = {}; -struct mlx5_modify_hdr *modify_hdr =NULL ; struct mlx5_flow_table_attr ft_attr = {};struct mlx5_flow_destination dest = {};struct mlx5_core_dev *mdev = fs->mdev;
@@ -482,28 +476,16 @@ accel_psp_fs_rx_decrypt_ft_create(struct mlx5e_psp_fs *fs,/* Add PSP RX decrypt rule */ setup_fte_udp_psp (spec,PSP_DEFAULT_UDP_PORT ); - flow_act.crypto.type =MLX5_FLOW_CONTEXT_ENCRYPT_DECRYPT_TYPE_PSP ; -/* Set bit[31, 30] PSP marker */ -#define MLX5E_PSP_MARKER_BIT (BIT(30) | BIT(31)) -MLX5_SET (set_action_in, action, action_type,MLX5_ACTION_TYPE_SET ); -MLX5_SET (set_action_in, action, field,MLX5_ACTION_IN_FIELD_METADATA_REG_B ); -MLX5_SET (set_action_in, action, data,MLX5E_PSP_MARKER_BIT ); -MLX5_SET (set_action_in, action, offset,0 ); -MLX5_SET (set_action_in, action, length,32 ); - - modify_hdr =mlx5_modify_header_alloc (mdev,MLX5_FLOW_NAMESPACE_KERNEL ,1 , action); -if (IS_ERR (modify_hdr)) { - err =PTR_ERR (modify_hdr); -mlx5_core_err (mdev,"fail to alloc psp set modify_header_id err=%d\n" , err); - modify_hdr =NULL ; - goto out_err; - } - decrypt->rx_modify_hdr = modify_hdr; +/* Set PSP marker via flow_tag */ + spec->flow_context.flags =FLOW_CONTEXT_HAS_TAG ; + spec->flow_context.flow_tag = +FIELD_PREP (MLX5E_ACCEL_FLOW_TAG_PROTO_MASK , +MLX5E_ACCEL_FLOW_TAG_PROTO_PSP ); + + flow_act.crypto.type =MLX5_FLOW_CONTEXT_ENCRYPT_DECRYPT_TYPE_PSP ; flow_act.action =MLX5_FLOW_CONTEXT_ACTION_FWD_DEST | -MLX5_FLOW_CONTEXT_ACTION_CRYPTO_DECRYPT | -MLX5_FLOW_CONTEXT_ACTION_MOD_HDR ; - flow_act.modify_hdr = modify_hdr; +MLX5_FLOW_CONTEXT_ACTION_CRYPTO_DECRYPT ; dest.type =MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE ; dest.ft = fs->check.ft; rule =mlx5_add_flow_rules (decrypt->ft, spec, &flow_act, &dest,1 );
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp_rxtx.h +5 -5
@@ -9,10 +9,7 @@#include <net/psp.h> #include "en.h" #include "en/txrx.h" - -/* Bit30: PSP marker, Bit22-0: PSP obj id */ -#define MLX5_PSP_METADATA_MARKER (metadata) ((((metadata) >> 30) & 0x3) == 0x3) -#define MLX5_PSP_METADATA_HANDLE (metadata) ((metadata) & GENMASK(22, 0)) +#include "en_accel/flow_tag.h" struct mlx5e_accel_tx_psp_state {u32 tailen;
@@ -82,7 +79,10 @@ static inline unsigned int mlx5e_psp_tx_ids_len(struct mlx5e_accel_tx_psp_statestatic inline bool mlx5e_psp_is_rx_flow (struct mlx5_cqe64 *cqe) { -return MLX5_PSP_METADATA_MARKER (be32_to_cpu (cqe->ft_metadata)); +u32 proto =mlx5e_accel_flow_tag_proto (cqe); + +return proto ==MLX5E_ACCEL_FLOW_TAG_PROTO_PSP ; + } bool mlx5e_psp_offload_handle_rx_skb(struct net_device *netdev, struct sk_buff *skb,
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 05/14] net/mlx5e: psp: Move RX marker from ft_metadata to flow_tag
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 6 lines
From: Cosmin Ratiu <[email protected]>
Move PSP RX marker from
ft_metadata(set viamodify_hdraction) toflow_tag(set viaflow_context). This freesft_metadatafor storing SPI for future decapsulated PSP packets.
Quoted 7 lines
All mlx5e accel protos have to move to avoid misinterpreting packets with high bits of
ft_metadataas IPsec or Macsec. This patch is the first step, defining a new header, bit layout and macros.Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=5
Tariq Toukan
maintainerPATCH v4 6/14 net/mlx5e: ipsec: Move RX marker from ft_metadata to flow_tag
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
IPsec used BIT(31) in cqe.ft_metadata as a marker for ingress IPsec-decrypted packets. This will conflict with PSP which stores SPI in ft_metadata - an SPI with bit 31 set would falsely trigger mlx5_ipsec_is_rx_flow().
Define a new marker for IPsec in flow_tag and use it in steering rules. The obj_id stays in ft_metadata bits[23:0].
Remove the unused MARKER and SYNDROM (sic) macros.
This is only done for IPsec NIC RX flows. The esw path uses a different mechanism and is unaffected: RX handling in mlx5e_rep_tc_receive() makes use of a few bits in ft_metadata to detect and hand off IPsec packets to mlx5e_ipsec_offload_handle_rx_skb().
An additional complication is that mlx5e_cqe_regb_chain() used the high order bits of ft_metadata to differentiate between TC chains and protocol markers. With the IPSec marker moving to flow_tag, this would falsely trigger mlx5e_cqe_regb_chain() to believe the chain is set, when in fact, ft_metadata is something completely different. Solve that by requiring that the new flow_tag proto is NONE, since only those packets can carry a chain id in ft_metadata.
--- .../ethernet/mellanox/mlx5/core/en_accel/flow_tag.h | 4 +++- .../ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c | 11 +++++++++-- .../ethernet/mellanox/mlx5/core/en_accel/ipsec_rxtx.h | 8 ++++---- drivers/net/ethernet/mellanox/mlx5/core/en_tc.h | 7 +++++-- 4 files changed, 21 insertions(+), 9 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/en_accel/flow_tag.h +3 -1
@@ -14,8 +14,9 @@* * [23:21] = protocol ID (3 bits): * 0 = none (default) +* 1 = IPsec * 3 = PSP (HW decrypted, PSP header present) -* 1,2,4-7 = reserved +* 2,4-7 = reserved * [20:16] = reserved * [15:0] = used by other subsystems (e.g. TC). */
@@ -24,6 +25,7 @@enum {MLX5E_ACCEL_FLOW_TAG_PROTO_NONE =0 , +MLX5E_ACCEL_FLOW_TAG_PROTO_IPSEC =1 ,MLX5E_ACCEL_FLOW_TAG_PROTO_PSP =3 , };
drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c +9 -2
@@ -5,6 +5,7 @@#include "en.h" #include "en/fs.h" #include "en_accel/en_accel.h" +#include "en_accel/flow_tag.h" #include "eswitch.h" #include "ipsec.h" #include "fs_core.h"
@@ -190,7 +191,7 @@ static void ipsec_rx_rule_add_match_obj(struct mlx5e_ipsec_sa_entry *sa_entry, misc_parameters_2.metadata_reg_c_2);MLX5_SET (fte_match_param, spec->match_value, misc_parameters_2.metadata_reg_c_2, - sa_entry->ipsec_obj_id |BIT (31 )); + sa_entry->ipsec_obj_id); spec->match_criteria_enable |=MLX5_MATCH_MISC_PARAMETERS_2 ; }
@@ -2066,7 +2067,7 @@ static int rx_add_rule(struct mlx5e_ipsec_sa_entry *sa_entry)if (!attrs->drop) {if (rx != ipsec->rx_esw) err =setup_modify_header (ipsec, attrs->type, - sa_entry->ipsec_obj_id |BIT (31 ), + sa_entry->ipsec_obj_id,XFRM_DEV_OFFLOAD_IN , &flow_act);else err =mlx5_esw_ipsec_rx_setup_modify_header (sa_entry, &flow_act);
@@ -2099,6 +2100,12 @@ static int rx_add_rule(struct mlx5e_ipsec_sa_entry *sa_entry) flow_act.action |=MLX5_FLOW_CONTEXT_ACTION_DROP ; else flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST; +if (!attrs->drop && rx != ipsec->rx_esw) { + spec->flow_context.flags |=FLOW_CONTEXT_HAS_TAG ; + spec->flow_context.flow_tag = +FIELD_PREP (MLX5E_ACCEL_FLOW_TAG_PROTO_MASK , +MLX5E_ACCEL_FLOW_TAG_PROTO_IPSEC ); + } dest[0 ].type =MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE ; dest[0 ].ft = rx->ft.status; dest[1 ].type =MLX5_FLOW_DESTINATION_TYPE_COUNTER ;
drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_rxtx.h +4 -4
@@ -38,10 +38,9 @@#include <net/xfrm.h> #include "en.h" #include "en/txrx.h" +#include "en_accel/flow_tag.h" -/* Bit31: IPsec marker, Bit30: reserved, Bit29-24: IPsec syndrome, Bit23-0: IPsec obj id */ -#define MLX5_IPSEC_METADATA_MARKER (metadata) ((((metadata) >> 30) & 0x3) == 0x2) -#define MLX5_IPSEC_METADATA_SYNDROM (metadata) (((metadata) >> 24) & GENMASK(5, 0)) +/* IPsec obj id in ft_metadata bits[23:0] */ #define MLX5_IPSEC_METADATA_HANDLE (metadata) ((metadata) & GENMASK(23, 0))struct mlx5e_accel_tx_ipsec_state {
@@ -74,7 +73,8 @@ static inline unsigned int mlx5e_ipsec_tx_ids_len(struct mlx5e_accel_tx_ipsec_ststatic inline bool mlx5_ipsec_is_rx_flow (struct mlx5_cqe64 *cqe) { -return MLX5_IPSEC_METADATA_MARKER (be32_to_cpu (cqe->ft_metadata)); +return mlx5e_accel_flow_tag_proto (cqe) == +MLX5E_ACCEL_FLOW_TAG_PROTO_IPSEC ; }static inline bool mlx5e_ipsec_eseg_meta (struct mlx5_wqe_eth_seg *eseg)
drivers/net/ethernet/mellanox/mlx5/core/en_tc.h +5 -2
@@ -36,6 +36,7 @@#include <net/pkt_cls.h> #include "en.h" #include "eswitch.h" +#include "en_accel/flow_tag.h" #include "en/tc_ct.h" #include "en/tc_tun.h" #include "en/tc/int_port.h"
@@ -370,11 +371,13 @@ struct mlx5e_tc_table *mlx5e_tc_table_alloc(void);void mlx5e_tc_table_free (struct mlx5e_tc_table *tc);static inline bool mlx5e_cqe_regb_chain (struct mlx5_cqe64 *cqe) { -u32 chain, reg_b; +u32 flow_tag, reg_b, chain; + flow_tag =mlx5e_accel_flow_tag_proto (cqe); reg_b =be32_to_cpu (cqe->ft_metadata); -if (reg_b >> (MLX5E_TC_TABLE_CHAIN_TAG_BITS +ESW_ZONE_ID_BITS )) +if (flow_tag !=MLX5E_ACCEL_FLOW_TAG_PROTO_NONE || + (reg_b >> (MLX5E_TC_TABLE_CHAIN_TAG_BITS +ESW_ZONE_ID_BITS )))return false; chain = reg_b &MLX5E_TC_TABLE_CHAIN_TAG_MASK ;
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 06/14] net/mlx5e: ipsec: Move RX marker from ft_metadata to flow_tag
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 24 lines
From: Cosmin Ratiu <[email protected]>
IPsec used BIT(31) in cqe.ft_metadata as a marker for ingress IPsec-decrypted packets. This will conflict with PSP which stores SPI in
ft_metadata- an SPI with bit 31 set would falsely triggermlx5_ipsec_is_rx_flow().Define a new marker for IPsec in
flow_tagand use it in steering rules. Theobj_idstays inft_metadatabits[23:0].
Remove the unused MARKER and SYNDROM (sic) macros.This is only done for IPsec NIC RX flows. The esw path uses a different mechanism and is unaffected: RX handling in
mlx5e_rep_tc_receive()makes use of a few bits inft_metadatato detect and hand off IPsec packets tomlx5e_ipsec_offload_handle_rx_skb().An additional complication is that
mlx5e_cqe_regb_chain()used the high order bits offt_metadatato differentiate between TC chains and protocol markers. With the IPSec marker moving toflow_tag, this would falsely triggermlx5e_cqe_regb_chain()to believe the chain is set, when in fact,ft_metadatais something completely different. Solve that by requiring that the newflow_tagproto is NONE, since only those packets can carry a chain id inft_metadata.
Quoted 4 lines
Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Carolina Jubran <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=6
Tariq Toukan
maintainerPATCH v4 7/14 net/mlx5e: macsec: Move RX marker from ft_metadata to flow_tag
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
MACsec uses BIT(30) in cqe.ft_metadata as a marker to identify MACsec-decrypted packets. This will conflict with PSP SPI values with bit 30 set, which would falsely trigger mlx5e_macsec_is_rx_flow().
Define a new marker for MACsec in flow_tag and use it in steering rules. The fs_id stays in ft_metadata bits[15:0].
Correct typos of MLX5_MACSEC_RX_METADAT_HANDLE and MLX5_MACEC_RX_FS_ID_MAX while touching this.
Delete the METADATA_MARKER macro.
Set flow_tag on both with-SCI and without-SCI crypto rules.
--- .../mellanox/mlx5/core/en_accel/flow_tag.h | 4 +++- .../mellanox/mlx5/core/en_accel/macsec.c | 6 +++--- .../mellanox/mlx5/core/en_accel/macsec.h | 4 +++- .../mellanox/mlx5/core/lib/macsec_fs.c | 21 ++++++++++++------- .../mellanox/mlx5/core/lib/macsec_fs.h | 9 ++++---- 5 files changed, 27 insertions(+), 17 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/en_accel/flow_tag.h +3 -1
@@ -15,8 +15,9 @@* [23:21] = protocol ID (3 bits): * 0 = none (default) * 1 = IPsec +* 2 = MACsec * 3 = PSP (HW decrypted, PSP header present) -* 2,4-7 = reserved +* 4-7 = reserved * [20:16] = reserved * [15:0] = used by other subsystems (e.g. TC). */
@@ -26,6 +27,7 @@enum {MLX5E_ACCEL_FLOW_TAG_PROTO_NONE =0 ,MLX5E_ACCEL_FLOW_TAG_PROTO_IPSEC =1 , +MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC =2 ,MLX5E_ACCEL_FLOW_TAG_PROTO_PSP =3 , };
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.c +3 -3
@@ -752,12 +752,12 @@ static int mlx5e_macsec_add_rxsc(struct macsec_context *ctx)* a non-NULL md_dst with uninitialised contents. */ err =xa_alloc (&macsec->sc_xarray, &sc_xarray_element->fs_id, sc_xarray_element, -XA_LIMIT (1 ,MLX5_MACEC_RX_FS_ID_MAX ),GFP_KERNEL ); +XA_LIMIT (1 ,MLX5_MACSEC_RX_FS_ID_MAX ),GFP_KERNEL );if (err) {if (err == -EBUSY )netdev_err (ctx->netdev,"MACsec offload: unable to create entry for RX SC (%d Rx SCs already allocated)\n" , -MLX5_MACEC_RX_FS_ID_MAX ); +MLX5_MACSEC_RX_FS_ID_MAX ); goto destroy_md_dst; }
@@ -1724,7 +1724,7 @@ void mlx5e_macsec_offload_handle_rx_skb(struct net_device *netdev,if (!macsec)return ; - fs_id =MLX5_MACSEC_RX_METADAT_HANDLE (macsec_meta_data); + fs_id =MLX5_MACSEC_RX_METADATA_HANDLE (macsec_meta_data);rcu_read_lock (); sc_xarray_element =xa_load (&macsec->sc_xarray, fs_id);
drivers/net/ethernet/mellanox/mlx5/core/en_accel/macsec.h +3 -1
@@ -9,6 +9,7 @@#include <linux/mlx5/driver.h> #include <net/macsec.h> #include <net/dst_metadata.h> +#include "en_accel/flow_tag.h" #include "lib/macsec_fs.h" struct mlx5e_priv ;
@@ -31,7 +32,8 @@ static inline bool mlx5e_macsec_skb_is_offload(struct sk_buff *skb)static inline bool mlx5e_macsec_is_rx_flow (struct mlx5_cqe64 *cqe) { -return MLX5_MACSEC_METADATA_MARKER (be32_to_cpu (cqe->ft_metadata)); +return mlx5e_accel_flow_tag_proto (cqe) == +MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC ; } void mlx5e_macsec_offload_handle_rx_skb(struct net_device *netdev, struct sk_buff *skb,
drivers/net/ethernet/mellanox/mlx5/core/lib/macsec_fs.c +14 -7
@@ -8,6 +8,7 @@#include <linux/mlx5/macsec.h> #include "fs_core.h" #include "lib/macsec_fs.h" +#include "en_accel/flow_tag.h" #include "mlx5_core.h" /* MACsec TX flow steering */
@@ -45,9 +46,6 @@#define MLX5_SECTAG_HEADER_SIZE_WITHOUT_SCI 0x8#define MLX5_SECTAG_HEADER_SIZE_WITH_SCI (MLX5_SECTAG_HEADER_SIZE_WITHOUT_SCI + MACSEC_SCI_LEN) -/* MACsec fs_id handling for steering */ -#define macsec_fs_set_rx_fs_id (fs_id) ((fs_id) | BIT(30)) -struct mlx5_sectag_header {__be16 ethertype;u8 tci_an;
@@ -1757,11 +1755,10 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, rx_tables = &rx_fs->tables; ft_crypto = &rx_tables->ft_crypto; -/* Set bit[31 - 30] macsec marker - 0x01 */ /* Set bit[15-0] fs id */ MLX5_SET (set_action_in, action, action_type,MLX5_ACTION_TYPE_SET );MLX5_SET (set_action_in, action, field,MLX5_ACTION_IN_FIELD_METADATA_REG_B ); -MLX5_SET (set_action_in, action, data,macsec_fs_set_rx_fs_id (fs_id)); +MLX5_SET (set_action_in, action, data, fs_id);MLX5_SET (set_action_in, action, offset,0 );MLX5_SET (set_action_in, action, length,32 );
@@ -1778,6 +1775,11 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs,/* Rx crypto table with SCI rule */ macsec_fs_rx_setup_fte (spec, &flow_act, attrs, true); + spec->flow_context.flags |=FLOW_CONTEXT_HAS_TAG ; + spec->flow_context.flow_tag = +FIELD_PREP (MLX5E_ACCEL_FLOW_TAG_PROTO_MASK , +MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC ); + flow_act.modify_hdr = modify_hdr; flow_act.action =MLX5_FLOW_CONTEXT_ACTION_FWD_DEST | MLX5_FLOW_CONTEXT_ACTION_CRYPTO_DECRYPT |
@@ -1803,6 +1805,11 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs,macsec_fs_rx_setup_fte (spec, &flow_act, attrs, false); + spec->flow_context.flags |=FLOW_CONTEXT_HAS_TAG ; + spec->flow_context.flow_tag = +FIELD_PREP (MLX5E_ACCEL_FLOW_TAG_PROTO_MASK , +MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC ); + flow_act.modify_hdr = modify_hdr; flow_act.action =MLX5_FLOW_CONTEXT_ACTION_FWD_DEST | MLX5_FLOW_CONTEXT_ACTION_CRYPTO_DECRYPT |
@@ -2160,8 +2167,8 @@ static int mlx5_macsec_fs_add_roce_rule_rx(struct mlx5_macsec_fs *macsec_fs, u32 spec->match_criteria_enable |=MLX5_MATCH_MISC_PARAMETERS_2 ;MLX5_SET_TO_ONES (fte_match_param, spec->match_criteria, misc_parameters_2.metadata_reg_c_5); -MLX5_SET (fte_match_param, spec->match_value, misc_parameters_2.metadata_reg_c_5, -macsec_fs_set_rx_fs_id (fs_id)); +MLX5_SET (fte_match_param, spec->match_value, + misc_parameters_2.metadata_reg_c_5, fs_id); flow_act.action =MLX5_FLOW_CONTEXT_ACTION_ALLOW ; new_rule =mlx5_add_flow_rules (rx_fs->roce.ft_macsec_op_check, spec, &flow_act,NULL ,0 );
drivers/net/ethernet/mellanox/mlx5/core/lib/macsec_fs.h +4 -5
@@ -6,11 +6,10 @@#ifdef CONFIG_MLX5_MACSEC -/* Bit31 - 30: MACsec marker, Bit15-0: MACsec id */ -#define MLX5_MACEC_RX_FS_ID_MAX USHRT_MAX/* Must be power of two */ -#define MLX5_MACSEC_RX_FS_ID_MASK MLX5_MACEC_RX_FS_ID_MAX -#define MLX5_MACSEC_METADATA_MARKER (metadata) ((((metadata) >> 30) & 0x3) == 0x1) -#define MLX5_MACSEC_RX_METADAT_HANDLE (metadata) ((metadata) & MLX5_MACSEC_RX_FS_ID_MASK) +/* MACsec fs_id in ft_metadata bits[15:0] */ +#define MLX5_MACSEC_RX_FS_ID_MAX USHRT_MAX/* Must be power of two */ +#define MLX5_MACSEC_RX_FS_ID_MASK MLX5_MACSEC_RX_FS_ID_MAX +#define MLX5_MACSEC_RX_METADATA_HANDLE (metadata) ((metadata) & MLX5_MACSEC_RX_FS_ID_MASK)/* MACsec TX flow steering */ #define MLX5_ETH_WQE_FT_META_MACSEC_MASK \
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 07/14] net/mlx5e: macsec: Move RX marker from ft_metadata to flow_tag
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Pre-existing issues:
- [High] Deterministic NULL pointer dereference in macsec_fs_id_del during error recovery.
- [High] The function macsec_fs_id_del frees an RCU-protected struct mlx5_fs_id synchronously with kfree() instead of deferring it via RCU, risking a use-after-free. --
commit b66a5cce46cf15521232a2fb7c33b033c4391c99
Author: Cosmin Ratiu <[email protected]>
net/mlx5e: macsec: Move RX marker from ft_metadata to flow_tag
This commit changes the MACsec marker from bit 30 of ft_metadata to a dedicated flow_tag protocol ID. It updates steering rules and helper macros accordingly to prevent conflicts with PSP SPI values.
Quoted 6 lines
drivers/net/ethernet/mellanox/mlx5/core/lib/macsec_fs.c +0
-0@@ -1803,6 +1805,11 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs,
drivers/net/ethernet/mellanox/mlx5/core/lib/macsec_fs.c macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, · line 1815
Quoted 10 lines
drivers/net/ethernet/mellanox/mlx5/core/lib/macsec_fs.c +5
-0macsec_fs_rx_setup_fte (spec, &flow_act, attrs, false); + spec->flow_context.flags |=FLOW_CONTEXT_HAS_TAG ; + spec->flow_context.flow_tag = +FIELD_PREP (MLX5E_ACCEL_FLOW_TAG_PROTO_MASK , +MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC ); + flow_act.modify_hdr = modify_hdr; flow_act.action =MLX5_FLOW_CONTEXT_ACTION_FWD_DEST | MLX5_FLOW_CONTEXT_ACTION_CRYPTO_DECRYPT |
[Severity: High]
This isn't a bug introduced by this patch, but later in this function during error recovery, there is a deterministic NULL pointer dereference:
macsec_fs_rx_add_rule() {
...
err = macsec_fs_id_add(&macsec_fs->macsec_devices_list, fs_id,
macsec_ctx->secy->netdev,
&macsec_fs->fs_id_hash, attrs->sci, false);
if (err) {
mlx5_core_err(mdev, "Failed to save fs_id, err=%d\n", err);
goto err;
}
...
err:
macsec_fs_rx_del_rule(macsec_fs, rx_rule, macsec_ctx->secy->netdev,
fs_id);
macsec_rule = NULL;
}When macsec_fs_id_add() fails (e.g., due to memory allocation failure), it cleans up by freeing the newly allocated macsec_device and removing it from the list. However, the err block then unconditionally calls macsec_fs_rx_del_rule(), which in turn calls macsec_fs_id_del():
macsec_fs_id_del() {
...
fs_id_xa = (is_tx) ? &macsec_device->tx_id_xa :
&macsec_device->rx_id_xa;
xa_lock(fs_id_xa);
...
}Since macsec_device was never successfully added, the lookup fails and returns NULL. Computing fs_id_xa from a NULL macsec_device pointer and calling xa_lock() on it causes a direct NULL pointer dereference. Could the error path be updated to avoid calling macsec_fs_id_del() when the device was not added?
[Severity: High]
This is a pre-existing issue, but macsec_fs_id_del() appears to free an RCU-protected structure synchronously instead of deferring it via RCU:
macsec_fs_id_del() {
...
rhashtable_remove_fast(fs_id_hash, &fs_id_found->rhash_node,
macsec_fs_id_rhash_params);
...
kfree(fs_id_found);
}Because it does not use kfree_rcu() or wait for an RCU grace period, concurrent lockless readers running under rcu_read_lock() can access the freed memory:
mlx5_macsec_fs_get_fs_id_from_hashtable() {
...
mlx5_fs_id = rhashtable_lookup(&macsec_fs->sci_hash, sci, rhash_sci);
if (mlx5_fs_id)
fs_id = mlx5_fs_id->id;
}If a reader obtains a pointer just before removal, it will dereference it after the writer has freed it. Should this memory be freed using kfree_rcu() to prevent use-after-free during concurrent datapath packet processing?
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=7
Tariq Toukan
maintainerPATCH v4 8/14 net/mlx5e: psp: Handle HW-decapsulated RX PSP packets
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
mlx5e_psp_offload_handle_rx_skb() handles RX PSP packets.
Add an additional flow tag marker and helpers for decapsulated PSP packets and extend the handler to construct the PSP skb extension from the CQE fields (SPI in ft_metadata and PSP version in flow_tag).
--- .../mellanox/mlx5/core/en_accel/flow_tag.h | 8 +++-- .../mellanox/mlx5/core/en_accel/psp_rxtx.c | 21 +++++++++-- .../mellanox/mlx5/core/en_accel/psp_rxtx.h | 36 ++++++++++++++++++- 3 files changed, 59 insertions(+), 6 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/en_accel/flow_tag.h +6 -2
@@ -17,18 +17,22 @@* 1 = IPsec * 2 = MACsec * 3 = PSP (HW decrypted, PSP header present) -* 4-7 = reserved -* [20:16] = reserved +* 4 = PSP decap (HW decrypted & decapsulated) +* 5-7 = reserved +* [20:17] = PSP version (4 bits, valid when proto == PSP decap) +* 16 = reserved * [15:0] = used by other subsystems (e.g. TC). */ #define MLX5E_ACCEL_FLOW_TAG_MASK GENMASK(23, 0)#define MLX5E_ACCEL_FLOW_TAG_PROTO_MASK GENMASK(23, 21) +#define MLX5E_ACCEL_FLOW_TAG_PSP_VER_MASK GENMASK(20, 17)enum {MLX5E_ACCEL_FLOW_TAG_PROTO_NONE =0 ,MLX5E_ACCEL_FLOW_TAG_PROTO_IPSEC =1 ,MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC =2 ,MLX5E_ACCEL_FLOW_TAG_PROTO_PSP =3 , +MLX5E_ACCEL_FLOW_TAG_PROTO_PSP_DECAP =4 , };static inline u32 mlx5e_accel_flow_tag (struct mlx5_cqe64 *cqe)
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp_rxtx.c +18 -3
@@ -118,10 +118,25 @@ bool mlx5e_psp_offload_handle_rx_skb(struct net_device *netdev, struct sk_buff * {struct mlx5e_priv *priv =netdev_priv (netdev);u16 dev_id = priv->psp->psd->id; -bool strip_icv = true; -u8 generation =0 ; +struct psp_skb_ext *pse; + +if (mlx5e_psp_is_decap (cqe)) { +/* UDP + PSP headers and PSP trailer removed by HW. +* Construct the PSP extension from CQE metadata. +*/ + pse =skb_ext_add (skb,SKB_EXT_PSP ); +if (unlikely (!pse)) + goto drop; + + pse->spi =mlx5e_psp_get_spi (cqe); + pse->version =mlx5e_psp_get_version (cqe); + pse->dev_id = dev_id; + pse->generation =0 ; + skb->decrypted =1 ; +return false; + } -if (psp_dev_rcv (skb, dev_id, generation, strip_icv)) +if (psp_dev_rcv (skb, dev_id,0 , true)) goto drop; skb->decrypted =1 ;
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp_rxtx.h +35 -1
@@ -81,10 +81,29 @@ static inline bool mlx5e_psp_is_rx_flow(struct mlx5_cqe64 *cqe) {u32 proto =mlx5e_accel_flow_tag_proto (cqe); -return proto ==MLX5E_ACCEL_FLOW_TAG_PROTO_PSP ; +return proto ==MLX5E_ACCEL_FLOW_TAG_PROTO_PSP || + proto ==MLX5E_ACCEL_FLOW_TAG_PROTO_PSP_DECAP ; } +static inline bool mlx5e_psp_is_decap (struct mlx5_cqe64 *cqe) +{ +u32 proto =mlx5e_accel_flow_tag_proto (cqe); + +return proto ==MLX5E_ACCEL_FLOW_TAG_PROTO_PSP_DECAP ; +} + +static inline u8 mlx5e_psp_get_version (struct mlx5_cqe64 *cqe) +{ +return FIELD_GET (MLX5E_ACCEL_FLOW_TAG_PSP_VER_MASK , +mlx5e_accel_flow_tag (cqe)); +} + +static inline __be32 mlx5e_psp_get_spi (struct mlx5_cqe64 *cqe) +{ +return cqe->ft_metadata; +} +bool mlx5e_psp_offload_handle_rx_skb (struct net_device *netdev,struct sk_buff *skb,struct mlx5_cqe64 *cqe);#else
@@ -110,6 +129,21 @@ static inline bool mlx5e_psp_is_rx_flow(struct mlx5_cqe64 *cqe)return false; } +static inline bool mlx5e_psp_is_decap (struct mlx5_cqe64 *cqe) +{ +return false; +} + +static inline u8 mlx5e_psp_get_version (struct mlx5_cqe64 *cqe) +{ +return 0 ; +} + +static inline __be32 mlx5e_psp_get_spi (struct mlx5_cqe64 *cqe) +{ +return 0 ; +} +static inline bool mlx5e_psp_offload_handle_rx_skb (struct net_device *netdev,struct sk_buff *skb,struct mlx5_cqe64 *cqe)
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 08/14] net/mlx5e: psp: Handle HW-decapsulated RX PSP packets
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 11 lines
From: Cosmin Ratiu <[email protected]>
mlx5e_psp_offload_handle_rx_skb()handles RX PSP packets.Add an additional flow tag marker and helpers for decapsulated PSP packets and extend the handler to construct the PSP skb extension from the CQE fields (SPI in
ft_metadataand PSP version inflow_tag).Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=8
Tariq Toukan
maintainerPATCH v4 9/14 net/mlx5e: psp: Add an rx_decap steering table
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
Introduce an additional steering table for PSP transport mode decapsulation, containing:
- one rule per supported PSP version which does:
- transport mode decap (removes UDP+PSP headers and PSP trailer)
- recomputes iph->tot_len
- recomputes IP checksum
- reparses packet headers
- copy SPI into reg_b (which ends up as cqe.ft_metadata)
- set a decap marker and the PSP version in the flow_tag, so the RX handler can make sense of the packet
- default drop rule for unsupported PSP versions (per PSP spec).
Packets are forwarded to the previously added rx table, where:
- one rule forwards UDP traffic to the UDP default destination.
- default rule forwards traffic to the TTC table.
The reason is to avoid steering loops. If packets were to be injected into the TTC directly after rx_decap, it may be possible to create a steering loop with RX packets of the form IP|UDP|PSP|UDP|PSP... The rx flow table guarantees that packets go through PSP steering at most once.
The steering mode is saved in a new field 'fs.decap_enabled'. Updating the mode is done through accel_psp_fs_rx_reconfigure(), which creates the decap steering table if needed and possible. It then uses an atomic rule update to redirect traffic to the new table. This is now invoked with decap_wanted == false.
The intention is for failures creating the new table to not block feature reconfig/PSP toggling. Decap table creation and traffic redirection failures therefore result in a warning that PSP HW GRO will not work.
Nothing happens on older HW/FW without the ability to decapsulate PSP transport.
An upcoming patch will add dynamic reconfiguration of PSP steering based on HW GRO.
--- .../net/ethernet/mellanox/mlx5/core/en/fs.h | 1 + .../mellanox/mlx5/core/en_accel/psp.c | 249 +++++++++++++++++- 2 files changed, 248 insertions(+), 2 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/en/fs.h +1 -0
@@ -98,6 +98,7 @@ enum {#if defined(CONFIG_MLX5_EN_PSP )MLX5E_ACCEL_FS_PSP_FT_LEVEL =MLX5E_INNER_TTC_FT_LEVEL +1 ,MLX5E_ACCEL_FS_PSP_ERR_FT_LEVEL , +MLX5E_ACCEL_FS_PSP_DECAP_FT_LEVEL , MLX5E_ACCEL_FS_PSP_RX_FT_LEVEL, #endif };
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +247 -2
333 lines not on this page · raw message
Signature · 1 line
2.44.0
Re: [PATCH net-next V4 09/14] net/mlx5e: psp: Add an rx_decap steering table
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c accel_psp_fs_rx_decrypt_ft_create(struct mlx5e_psp_fs *fs, · line 720
Quoted 8 lines · Tariq Toukan
From: Cosmin Ratiu <[email protected]>
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +6
-0+static void accel_psp_fs_rx_reconfigure (struct mlx5e_psp_fs *fs, +bool decap_wanted) +{ +bool decap_supported = +MLX5_CAP_FLOWTABLE (fs->mdev, + flow_table_properties_nic_receive.reformat_del_psp_transport);
not for this series, but this makes me think having some way of letting the user know whether or not psp with hw gro is enabled would be nice, so that they don't wonder why stuff isn't coalescing when they have old fw like I do :) I could see a similar thing being useful for gso features if the psp device can't segment something that it otherwise can for non-psp.
Quoted 29 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +29
-0+bool decap_enable = decap_wanted && decap_supported; +struct mlx5_flow_destination dest = {}; +int err; + +if (decap_enable == fs->decap_enabled) +return ; + +/* Create the decap table if needed. */ +if (decap_enable && !fs->decap.ft) { + err =accel_psp_fs_rx_decap_ft_create (fs, &fs->decap); +if (err) + goto out_err; + } + +/* Redirect traffic to the correct table. */ + dest.type =MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE ; + dest.ft = decap_enable ? fs->decap.ft : fs->rx.ft; + err =mlx5_modify_rule_destination (fs->check.rule, &dest,NULL ); +if (err) + goto out_err; + + fs->decap_enabled = decap_enable; +return ; + +out_err: +mlx5_core_warn (fs->mdev, +"Failed to create/modify PSP decapsulation rules (err %d), HW GRO for PSP unavailable\n" , + err); +}
[email protected]
AI reviewRe: [PATCH net-next V4 09/14] net/mlx5e: psp: Add an rx_decap steering table
In reply to Tariq ToukanSashiko has reviewed this patch and found no issues. It looks great!
Quoted 14 lines
From: Cosmin Ratiu <[email protected]>
Introduce an additional steering table for PSP transport mode decapsulation, containing:
- one rule per supported PSP version which does:
- transport mode decap (removes UDP+PSP headers and PSP trailer)
- recomputesiph->tot_len
- recomputes IP checksum
- reparses packet headers
- copy SPI intoreg_b(which ends up as cqe.ft_metadata)
- set a decap marker and the PSP version in theflow_tag, so the RX handler can make sense of the packet
- default drop rule for unsupported PSP versions (per PSP spec).
Quoted 7 lines
Packets are forwarded to the previously added rx table, where:
[ ... ]
on HW GRO.Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=9
[email protected]
AI reviewRe: [PATCH net-next V4 09/14] net/mlx5e: psp: Add an rx_decap steering table
In reply to Tariq ToukanThank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
Critical: 0 · High: 1 · Medium: 1 · Low: 0
- [High] accel_psp_fs_rx_decap_ft_create() (drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c) creates the…
- [Medium] accel_psp_fs_rx_decrypt_ft_create() now unconditionally allocates a modify-header context that copies…
--
net/mlx5e: psp: Add an rx_decap steering table
This adds a new PSP RX decapsulation steering table with one rule per supported PSP version (transport-mode decap, SPI copy into reg_b, flow tag marker) plus a drop rule for unsupported versions. The mode is tracked in fs.decap_enabled and switched through accel_psp_fs_rx_reconfigure(), which is called with decap_wanted == false in this commit.
A couple of questions below.
Quoted 4 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +0
-0
[ ... ]
Quoted 8 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +4
-0@@ -403,11 +434,164 @@ int accel_psp_fs_rx_check_ft_create(struct mlx5e_psp_fs *fs,return err; } +static +int accel_psp_fs_rx_decap_ft_create (struct mlx5e_psp_fs *fs, +struct mlx5e_psp_rx_decap_table *decap) +{
[ ... ]
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c int accel_psp_fs_rx_check_ft_create(struct mlx5e_psp_fs *fs, · line 494
Quoted 6 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +6
-0+/* Create FT */ + ft_attr.max_fte =1 +MLX5E_PSP_NUM_SUPPORTED_VERSIONS ; + ft_attr.level =MLX5E_ACCEL_FS_PSP_DECAP_FT_LEVEL ; + ft_attr.prio =MLX5E_NIC_PRIO ; + ft_attr.autogroup.num_reserved_entries =1 ; + err =accel_psp_fs_create_ft (fs, &ft_attr, &decap->ft);
[Severity: High]
Should ft_attr.flags here include MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT?
Every rule installed into this table below carries MLX5_FLOW_CONTEXT_ACTION_PACKET_REFORMAT with a MLX5_REFORMAT_TYPE_REMOVE_PSP_TRANSPORT context, but the table flag is the only source of flow_table_context.reformat_en:
fs_cmd.c:mlx5_cmd_create_flow_table() {
int en_encap = !!(ft->flags & MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT);
...
MLX5_SET(create_flow_table_in, in, flow_table_context.reformat_en,
en_encap);
}The SWS and HWS backends derive reformat_en from the same flag, so with flags == 0 the table is created with reformat disabled while its FTEs request a reformat action.
The directly analogous header-removal case in the same driver does set the flag, for the MACsec RX check table that hosts MLX5_REFORMAT_TYPE_DEL_MACSEC:
lib/macsec_fs.c:macsec_fs_rx_create() {
flow_table = macsec_fs_auto_group_table_create(ns,
MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT,
RX_CHECK_TABLE_LEVEL,
RX_CHECK_TABLE_NUM_FTE);
}IPsec does the same on rx->ft.sa / tx->ft.sa in en_accel/ipsec_fs.c, and so do esw/bridge.c, lib/fs_chains.c, esw/vporttbl.c and eswitch_offloads.c for their reformat-bearing tables.
If firmware rejects the reformat FTE, mlx5_add_flow_rules() fails in the version loop, accel_psp_fs_rx_decap_ft_create() unwinds, and accel_psp_fs_rx_reconfigure() only prints "HW GRO for PSP unavailable", so the feature stays inert even on HW/FW advertising reformat_del_psp_transport. If a backend instead accepts the FTE without performing the reformat, packets that still carry UDP+PSP headers get the PSP-decap flow tag and are interpreted as decapsulated by the RX handler added later in the series.
Note there is no software-side check in fs_core.c for this mismatch, so it only shows up as a firmware syndrome at rule installation time. Also, as IPsec shows (mlx5_eswitch_block_encap() before setting the flag), adding the flag alone may not be the whole fix.
At the end of the series (selftests: drv-net: Add PSP HW GRO conformance tests) psp.c still has no ft_attr.flags assignment and no MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT anywhere.
Quoted 5 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +5
-0+if (err) { +mlx5_core_err (mdev,"fail to create psp decap rx ft err=%d\n" , + err); + goto out_spec; + }
[ ... ]
Quoted 11 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +11
-0+/* pkt_reformat: decap PSP transport */ + reformat_params.type =MLX5_REFORMAT_TYPE_REMOVE_PSP_TRANSPORT ; + reformat =mlx5_packet_reformat_alloc (mdev, &reformat_params, +MLX5_FLOW_NAMESPACE_KERNEL ); +if (IS_ERR (reformat)) { + err =PTR_ERR (reformat); + goto out_err; + } + decap->reformat = reformat; + +for (i =0 ; i <MLX5E_PSP_NUM_SUPPORTED_VERSIONS ; i++) {
[ ... ]
Quoted 24 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +24
-0+ flow_act.action =MLX5_FLOW_CONTEXT_ACTION_PACKET_REFORMAT | +MLX5_FLOW_CONTEXT_ACTION_MOD_HDR | +MLX5_FLOW_CONTEXT_ACTION_FWD_DEST ; + flow_act.pkt_reformat = reformat; + flow_act.modify_hdr = modify_hdr; + dest.type =MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE ; + dest.ft = fs->rx.ft; + + rule =mlx5_add_flow_rules (decap->ft, spec, &flow_act, &dest,1 ); +if (IS_ERR (rule)) { + err =PTR_ERR (rule); + goto out_err; + } + decap->rule[i] = rule; + } + goto out_spec; + +out_err: +accel_psp_fs_rx_decap_ft_destroy (fs, decap); +out_spec: +kvfree (spec); +return err; +} +
[ ... ]
Quoted 12 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +8
-0@@ -474,6 +660,24 @@ accel_psp_fs_rx_decrypt_ft_create(struct mlx5e_psp_fs *fs, } decrypt->miss_rule = rule; +/* Create modify_hdr to copy SPI to REG_C_0 */ +MLX5_SET (copy_action_in, action, action_type,MLX5_ACTION_TYPE_COPY ); +MLX5_SET (copy_action_in, action, src_field, +MLX5_ACTION_IN_FIELD_PSP_HEADER_1 ); +MLX5_SET (copy_action_in, action, src_offset,0 ); +MLX5_SET (copy_action_in, action, length,0 );/* 0 = 32 bits */ +MLX5_SET (copy_action_in, action, dst_field, +MLX5_ACTION_IN_FIELD_METADATA_REG_C_0 );
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c accel_psp_fs_rx_decrypt_ft_create(struct mlx5e_psp_fs *fs, · line 680
Quoted 10 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +10
-0+MLX5_SET (copy_action_in, action, dst_offset,0 ); + + modhdr =mlx5_modify_header_alloc (mdev,MLX5_FLOW_NAMESPACE_KERNEL ,1 , + action); +if (IS_ERR (modhdr)) { + err =PTR_ERR (modhdr); + goto out_err; + } + decrypt->modify_hdr = modhdr; +
[Severity: Medium]
Should this SPI copy be gated on the same capability as the decap table?
The only consumer of REG_C_0 here is the optional decap table, which re-copies REG_C_0 into REG_B and is gated in accel_psp_fs_rx_reconfigure() on:
bool decap_supported = MLX5_CAP_FLOWTABLE(fs->mdev, flow_table_properties_nic_receive.reformat_del_psp_transport);
with failures there being warn-only. On the decrypt path, however, the MLX5_ACTION_IN_FIELD_PSP_HEADER_1 copy is unconditional, and a failure is fatal for the whole feature:
accel_psp_fs_rx_create()
accel_psp_fs_rx_decrypt_ft_create() /* -> goto out_err */
NL_SET_ERR_MSG(extack, "Failed creating RX decrypt steering table");
mlx5_cmd_modify_header_alloc() validates only num_actions against max_modify_header_actions, so the source field is validated by firmware alone and there is no capability bit for it. Both MLX5_ACTION_IN_FIELD_PSP_HEADER_1 and the reformat_del_psp_transport bit came in with commit bee40a7d0bd1 ("net/mlx5: Add PSP related fields to the mlx5_ifc"), while PSP crypto offload shipped earlier with commit 04a3134f88a4. On FW that has PSP crypto offload but not the newer PSP header field, does PSP offload stop being enable-able at all rather than just losing decap/GRO?
That would not match the commit message:
The intention is for failures creating the new table to not block feature reconfig/PSP toggling.
...
Nothing happens on older HW/FW without the ability to decapsulate PSP transport.
The change to the pre-existing decrypt rule is not mentioned in the commit message at all; could it be described there?
Two smaller points in the same block. This is the only failure branch of accel_psp_fs_rx_decrypt_ft_create() without an mlx5_core_err(), while the code it replaces did log ("fail to alloc psp set modify_header_id err=%d"). And since accel_psp_fs_rx_create() calls accel_psp_fs_rx_reconfigure(fs, false) in this commit, the extra MOD_HDR action runs for every PSP RX packet with nothing reading REG_C_0 until the follow-up patch enables decap. At the end of the series the decrypt-side copy is still unconditional and still unlogged.
Quoted 16 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +4
-2/* Add PSP RX decrypt rule */ setup_fte_udp_psp (spec,PSP_DEFAULT_UDP_PORT );@@ -484,8 +688,10 @@ accel_psp_fs_rx_decrypt_ft_create(struct mlx5e_psp_fs *fs, MLX5E_ACCEL_FLOW_TAG_PROTO_PSP); flow_act.crypto.type =MLX5_FLOW_CONTEXT_ENCRYPT_DECRYPT_TYPE_PSP ; - flow_act.action =MLX5_FLOW_CONTEXT_ACTION_FWD_DEST | -MLX5_FLOW_CONTEXT_ACTION_CRYPTO_DECRYPT ; + flow_act.action =MLX5_FLOW_CONTEXT_ACTION_CRYPTO_DECRYPT | +MLX5_FLOW_CONTEXT_ACTION_MOD_HDR | +MLX5_FLOW_CONTEXT_ACTION_FWD_DEST ; + flow_act.modify_hdr = modhdr; dest.type =MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE ; dest.ft = fs->check.ft; rule =mlx5_add_flow_rules (decrypt->ft, spec, &flow_act, &dest,1 );
[ ... ]
Signature · 1 line
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260908065244.3799142-1-tariqt%40nvidia.com
Tariq Toukan
maintainerPATCH v4 10/14 net/mlx5e: shampo: Flush session on PSP mismatch
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
Flush SHAMPO session on PSP status change (no PSP -> PSP or vice-versa) or on SPI/version mismatch. HW aggregates by 5-tuple but is unaware of PSP fields, so sessions must be terminated when security parameters change to avoid mixing packets from different PSP associations.
The session is also forcefully flushed if non-decapsulated packets are received. This could happen if, for example, the decap table could not be created so HW GRO is active but there's no decapsulation.
--- .../net/ethernet/mellanox/mlx5/core/en_rx.c | 32 +++++++++++++++++-- 1 file changed, 30 insertions(+), 2 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/en_rx.c +30 -2
@@ -2224,6 +2224,33 @@ static bool mlx5e_hw_gro_skb_has_enough_space(struct sk_buff *skb,return page_size * nr_frags + data_bcnt <=GRO_LEGACY_MAX_SIZE ; } +static bool mlx5e_hw_gro_psp_match (struct sk_buff *skb,struct mlx5_cqe64 *cqe) +{ +#ifdef CONFIG_MLX5_EN_PSP +struct psp_skb_ext *pse =skb_ext_find (skb,SKB_EXT_PSP ); +bool is_psp =mlx5e_psp_is_rx_flow (cqe); + +if (likely (!is_psp && !pse)) +return true; + +/* No match on PSP status change (no crypto -> crypto or vice-versa). */ +if (unlikely (is_psp != !!pse)) +return false; + +/* SPI and version are only available in CQE metadata for decap flows. +* Non-decap PSP cannot be matched here, force a flush. +*/ +if (unlikely (!mlx5e_psp_is_decap (cqe))) +return false; + +/* No match on security parameters change. */ +return pse->spi ==mlx5e_psp_get_spi (cqe) && + pse->version ==mlx5e_psp_get_version (cqe); +#else +return true; +#endif +} +static void mlx5e_handle_rx_cqe_mpwrq_shampo (struct mlx5e_rq *rq,struct mlx5_cqe64 *cqe) {u16 data_bcnt =mpwrq_get_cqe_byte_cnt (cqe) - cqe->shampo.header_size;
@@ -2266,8 +2293,9 @@ static void mlx5e_handle_rx_cqe_mpwrq_shampo(struct mlx5e_rq *rq, struct mlx5_cq }if (*skb && - !(match &&mlx5e_hw_gro_skb_has_enough_space (*skb, data_bcnt, - page_size))) { + !(match && +mlx5e_hw_gro_skb_has_enough_space (*skb, data_bcnt, page_size) && +mlx5e_hw_gro_psp_match (*skb, cqe))) { match = false;mlx5e_shampo_flush_skb (rq, cqe, match); }
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 10/14] net/mlx5e: shampo: Flush session on PSP mismatch
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 7 lines
From: Cosmin Ratiu <[email protected]>
Flush SHAMPO session on PSP status change (no PSP -> PSP or vice-versa) or on SPI/version mismatch. HW aggregates by 5-tuple but is unaware of PSP fields, so sessions must be terminated when security parameters change to avoid mixing packets from different PSP associations.
Quoted 7 lines
The session is also forcefully flushed if non-decapsulated packets are received. This could happen if, for example, the decap table could not be created so HW GRO is active but there's no decapsulation.
Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=10
[email protected]
AI reviewRe: [PATCH net-next V4 10/14] net/mlx5e: shampo: Flush session on PSP mismatch
In reply to Tariq ToukanThank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
Critical: 0 · High: 0 · Medium: 1 · Low: 0
- [Medium] Understated impact / missing Fixes: tag.
--
net/mlx5e: shampo: Flush session on PSP mismatch
This adds mlx5e_hw_gro_psp_match() and wires it into the SHAMPO session match test in mlx5e_handle_rx_cqe_mpwrq_shampo(), so a session is flushed when the PSP status changes, when SPI/version differ, or when the CQE indicates a non-decapsulated PSP packet.
(No Link: tags present in the commit header.)
Quoted 16 lines
drivers/net/ethernet/mellanox/mlx5/core/en_rx.c +8
-0@@ -2224,6 +2224,33 @@ static bool mlx5e_hw_gro_skb_has_enough_space(struct sk_buff *skb,return page_size * nr_frags + data_bcnt <=GRO_LEGACY_MAX_SIZE ; } +static bool mlx5e_hw_gro_psp_match (struct sk_buff *skb,struct mlx5_cqe64 *cqe) +{ +#ifdef CONFIG_MLX5_EN_PSP +struct psp_skb_ext *pse =skb_ext_find (skb,SKB_EXT_PSP ); +bool is_psp =mlx5e_psp_is_rx_flow (cqe); + +if (likely (!is_psp && !pse)) +return true;
drivers/net/ethernet/mellanox/mlx5/core/en_rx.c static bool mlx5e_hw_gro_skb_has_enough_space(struct sk_buff *skb, · line 2244
Quoted 10 lines
drivers/net/ethernet/mellanox/mlx5/core/en_rx.c +10
-0+ +/* No match on PSP status change (no crypto -> crypto or vice-versa). */ +if (unlikely (is_psp != !!pse)) +return false; + +/* SPI and version are only available in CQE metadata for decap flows. +* Non-decap PSP cannot be matched here, force a flush. +*/ +if (unlikely (!mlx5e_psp_is_decap (cqe))) +return false;
[Severity: Medium]
Is the commit message understating what this particular hunk fixes? The message says:
The session is also forcefully flushed if non-decapsulated packets are received. This could happen if, for example, the decap table could not be created so HW GRO is active but there's no decapsulation.
At the baseline commit (bc6fe9d301d5), non-decap looks like the only PSP RX mode rather than an exceptional fallback:mlx5e_psp_offload_handle_rx_skb() in en_accel/psp_rxtx.c only calls psp_dev_rcv(), and mlx5e_psp_update_rx()/shampo_enabled() coupling between PSP and NETIF_F_GRO_HW is only added by a later patch in this series.
If that is right, does the following sequence mis-deliver payload bytes today, without this patch?
mlx5e_handle_rx_cqe_mpwrq_shampo()
-> HW coalesces several PSP datagrams sharing the outer IP/UDP 5-tuple
(mlx5e_shampo_update_hdr() -> mlx5e_shampo_update_ipv4_udp_hdr()
sets SKB_GSO_UDP_L4, so UDP flows are aggregated too)
-> mlx5e_shampo_complete_rx_cqe():
if (NAPI_GRO_CB(skb)->count != 1) return false; if (mlx5e_build_rx_skb(cqe, cqe_bcnt, rq, skb)) return true;
so mlx5e_build_rx_skb() -> mlx5e_psp_offload_handle_rx_skb() -> psp_dev_rcv() runs only for the first record of a session. psp_dev_rcv() in net/psp/psp_main.c strips one 16-byte PSP header plus one 16-byte trailer and adjusts iph->tot_len once, so do the PSP headers and ICVs of records 2..N remain embedded in the payload handed to the socket?
If so, should the commit message say that PSP + HW GRO (ethtool -K <dev> rx-gro-hw on, also implied by tcp-data-split) is broken in the current tree, and should the patch carry a Fixes: tag so stable kernels shipping mlx5e PSP pick it up? A natural candidate looks like 29d7f433fcec ("net/mlx5e: Add Rx data path offload"), even if a backport needs adjustment because mlx5e_psp_is_decap() is new in this series.
Quoted 24 lines
drivers/net/ethernet/mellanox/mlx5/core/en_rx.c +12
-2+ +/* No match on security parameters change. */ +return pse->spi ==mlx5e_psp_get_spi (cqe) && + pse->version ==mlx5e_psp_get_version (cqe); +#else +return true; +#endif +} +static void mlx5e_handle_rx_cqe_mpwrq_shampo (struct mlx5e_rq *rq,struct mlx5_cqe64 *cqe) {u16 data_bcnt =mpwrq_get_cqe_byte_cnt (cqe) - cqe->shampo.header_size;@@ -2266,8 +2293,9 @@ static void mlx5e_handle_rx_cqe_mpwrq_shampo(struct mlx5e_rq *rq, struct mlx5_cq }if (*skb && - !(match &&mlx5e_hw_gro_skb_has_enough_space (*skb, data_bcnt, - page_size))) { + !(match && +mlx5e_hw_gro_skb_has_enough_space (*skb, data_bcnt, page_size) && +mlx5e_hw_gro_psp_match (*skb, cqe))) { match = false;mlx5e_shampo_flush_skb (rq, cqe, match); }
Signature · 1 line
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260908065244.3799142-1-tariqt%40nvidia.com
Tariq Toukan
maintainerPATCH v4 11/14 net/mlx5e: psp: Dynamically reconfigure based on SHAMPO mode
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
Add an mlx5e_update_nic_rx() -> mlx5e_psp_update_rx() hook. This allows PSP steering to reconfigure when netdev features change.
When this signal is received, toggle PSP between standard and decap modes when SHAMPO mode changes, optionally creating the decap steering table.
The toggling of traffic is atomic (a single steering rule update), so no packets should get eaten/lost. It is probable though that the rxhash will change, and decapsulated packets will be received by a different core than before. Established PSP connections receiving traffic might experience reordering/retransmits as a result.
Also enable decap when configuring the device if the mode indicates it.
--- .../mellanox/mlx5/core/en_accel/en_accel.h | 5 ++++ .../mellanox/mlx5/core/en_accel/psp.c | 23 +++++++++++++++++-- .../mellanox/mlx5/core/en_accel/psp.h | 2 ++ .../net/ethernet/mellanox/mlx5/core/en_main.c | 10 +++++++- 4 files changed, 37 insertions(+), 3 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/en_accel/en_accel.h +5 -0
@@ -251,6 +251,11 @@ static inline void mlx5e_accel_cleanup_rx(struct mlx5e_priv *priv)mlx5_accel_psp_fs_cleanup_rx_tables (priv); } +static inline void mlx5e_accel_update_rx (struct mlx5e_priv *priv) +{ +mlx5e_psp_update_rx (priv); +} +static inline int mlx5e_accel_init_tx (struct mlx5e_priv *priv) {return mlx5e_ktls_init_tx (priv);
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +21 -2
@@ -92,6 +92,12 @@ struct mlx5e_psp_fs {struct mlx5e_psp_rx_table rx; }; +static bool shampo_enabled (struct mlx5e_priv *priv) +{ +return priv->channels.params.packet_merge.type == +MLX5E_PACKET_MERGE_SHAMPO ; +} +/* PSP RX flow steering */ static enum mlx5_traffic_types fs_psp2tt (enum accel_fs_psp_type i) {
@@ -767,6 +773,7 @@ static void accel_psp_fs_rx_destroy(struct mlx5e_psp_fs *fs) }static int accel_psp_fs_rx_create (struct mlx5e_psp_fs *fs, +bool decap_enable,struct netlink_ext_ack *extack) {struct mlx5_ttc_table *ttc =mlx5e_fs_get_ttc (fs->fs, false);
@@ -808,7 +815,7 @@ static int accel_psp_fs_rx_create(struct mlx5e_psp_fs *fs,mlx5_ttc_fwd_dest (ttc,fs_psp2tt (i), &dest); } -accel_psp_fs_rx_reconfigure (fs, false); +accel_psp_fs_rx_reconfigure (fs, decap_enable);return 0 ;
@@ -1078,7 +1085,8 @@ static int accel_psp_fs_create(struct mlx5e_priv *priv, {int err; - err =accel_psp_fs_rx_create (priv->psp->fs, extack); + err =accel_psp_fs_rx_create (priv->psp->fs,shampo_enabled (priv), + extack);if (err)return err;
@@ -1355,3 +1363,14 @@ void mlx5e_psp_cleanup(struct mlx5e_priv *priv) priv->psp =NULL ;kfree (psp); } + +void mlx5e_psp_update_rx (struct mlx5e_priv *priv) +{ +struct mlx5e_psp *psp = priv->psp; + +netdev_assert_locked (priv->netdev); +if (!psp || !psp->fs->check.ft) +return ; + +accel_psp_fs_rx_reconfigure (psp->fs,shampo_enabled (priv)); +}
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.h +2 -0
@@ -49,6 +49,7 @@ int mlx5e_psp_register(struct mlx5e_priv *priv);void mlx5e_psp_unregister (struct mlx5e_priv *priv);int mlx5e_psp_init (struct mlx5e_priv *priv);void mlx5e_psp_cleanup (struct mlx5e_priv *priv); +void mlx5e_psp_update_rx (struct mlx5e_priv *priv);#else static inline void mlx5_accel_psp_fs_cleanup_rx_tables (struct mlx5e_priv *priv) { }static inline void mlx5_accel_psp_fs_cleanup_tx_tables (struct mlx5e_priv *priv) { }
@@ -61,5 +62,6 @@ static inline int mlx5e_psp_register(struct mlx5e_priv *priv) { return 0; }static inline void mlx5e_psp_unregister (struct mlx5e_priv *priv) { }static inline int mlx5e_psp_init (struct mlx5e_priv *priv) {return 0 ; }static inline void mlx5e_psp_cleanup (struct mlx5e_priv *priv) { } +static inline void mlx5e_psp_update_rx (struct mlx5e_priv *priv) { }#endif /* CONFIG_MLX5_EN_PSP */ #endif /* __MLX5E_ACCEL_PSP_H__ */
drivers/net/ethernet/mellanox/mlx5/core/en_main.c +9 -1
@@ -6288,7 +6288,15 @@ static void mlx5e_nic_disable(struct mlx5e_priv *priv)static int mlx5e_update_nic_rx (struct mlx5e_priv *priv) { -return mlx5e_refresh_tirs (priv->mdev, false, false); +int err; + + err =mlx5e_refresh_tirs (priv->mdev, false, false); +if (err) +return err; + +mlx5e_accel_update_rx (priv); + +return 0 ; }static const struct mlx5e_profile mlx5e_nic_profile = {
Signature · 1 line
2.44.0
[email protected]
AI reviewRe: [PATCH net-next V4 11/14] net/mlx5e: psp: Dynamically reconfigure based on SHAMPO mode
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 15 lines
From: Cosmin Ratiu <[email protected]>
Add an
mlx5e_update_nic_rx()->mlx5e_psp_update_rx()hook. This allows PSP steering to reconfigure when netdev features change.When this signal is received, toggle PSP between standard and decap modes when SHAMPO mode changes, optionally creating the decap steering table.
The toggling of traffic is atomic (a single steering rule update), so no packets should get eaten/lost. It is probable though that the rxhash will change, and decapsulated packets will be received by a different core than before. Established PSP connections receiving traffic might experience reordering/retransmits as a result.
Quoted 5 lines
Also enable decap when configuring the device if the mode indicates it.
Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=11
[email protected]
AI reviewRe: [PATCH net-next V4 11/14] net/mlx5e: psp: Dynamically reconfigure based on SHAMPO mode
In reply to Tariq ToukanThank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
Critical: 0 · High: 0 · Medium: 1 · Low: 0
- [Medium] In mlx5e_update_nic_rx() (drivers/net/ethernet/mellanox/mlx5/core/en_main.c:6289-6300) the…
--
net/mlx5e: psp: Dynamically reconfigure based on SHAMPO mode
This adds an mlx5e_update_nic_rx() -> mlx5e_accel_update_rx() -> mlx5e_psp_update_rx() hook so PSP steering can switch between standard and decap modes when the SHAMPO mode changes. It also enables decap at fs creation time when shampo_enabled() says so.
Quoted 4 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +0
-0
[ ... ]
Quoted 15 lines
drivers/net/ethernet/mellanox/mlx5/core/en_accel/psp.c +11
-0@@ -1355,3 +1363,14 @@ void mlx5e_psp_cleanup(struct mlx5e_priv *priv) priv->psp =NULL ;kfree (psp); } + +void mlx5e_psp_update_rx (struct mlx5e_priv *priv) +{ +struct mlx5e_psp *psp = priv->psp; + +netdev_assert_locked (priv->netdev); +if (!psp || !psp->fs->check.ft) +return ; + +accel_psp_fs_rx_reconfigure (psp->fs,shampo_enabled (priv)); +}
[ ... ]
Quoted 9 lines
drivers/net/ethernet/mellanox/mlx5/core/en_main.c +0
-1@@ -6288,7 +6288,15 @@ static void mlx5e_nic_disable(struct mlx5e_priv *priv)static int mlx5e_update_nic_rx (struct mlx5e_priv *priv) { -return mlx5e_refresh_tirs (priv->mdev, false, false);
drivers/net/ethernet/mellanox/mlx5/core/en_main.c static void mlx5e_nic_disable(struct mlx5e_priv *priv) · line 6300
Quoted 10 lines
drivers/net/ethernet/mellanox/mlx5/core/en_main.c +9
-0+int err; + + err =mlx5e_refresh_tirs (priv->mdev, false, false); +if (err) +return err; + +mlx5e_accel_update_rx (priv); + +return 0 ; }
[Severity: Medium]
Should the new PSP hook be gated on the TIR refresh result? The two operations look unrelated (TIR self-loopback refresh versus the PSP check rule destination), so a failing mlx5e_refresh_tirs() now also skips the PSP steering reconfiguration.
mlx5e_refresh_tirs() does have reachable failure modes on devices where MLX5_CAP_GEN(mdev, tis_tir_td_order) is 0, since it then calls mlx5e_modify_tirs_lb() in en_common.c, which returns -ENOMEM when mlx5e_tir_builder_alloc() fails and otherwise propagates the MODIFY_TIR firmware error:
drivers/net/ethernet/mellanox/mlx5/core/en_common.c:mlx5e_modify_tirs_lb() {
builder = mlx5e_tir_builder_alloc(true);
if (!builder)
return -ENOMEM;
...
err = mlx5e_tir_modify(tir, builder);
if (err) {
...
break;
}The live reconfiguration caller does not look at the return value either, and it has already committed the new channel params by then:
drivers/net/ethernet/mellanox/mlx5/core/en_main.c:mlx5e_switch_priv_channels() {
priv->channels = *new_chs;
...
priv->profile->update_rx(priv);
...
mlx5e_activate_priv_channels(priv);
}In that case the datapath runs with the new packet_merge.type while fs->decap_enabled still reflects the old mode, and nothing re-evaluates it, because shampo_enabled() is only consulted from mlx5e_psp_update_rx() and accel_psp_fs_create(). When SHAMPO has just been enabled the PSP check rule keeps pointing at fs->rx.ft rather than fs->decap.ft, so HW GRO for PSP stays unavailable, and unlike accel_psp_fs_rx_reconfigure()'s own error path there is no mlx5_core_warn() to indicate it.
The RX path itself appears to tolerate the mismatch, since mlx5e_psp_is_decap() keys off the per-CQE flow tag and en_rx.c flushes the SHAMPO session for non-decap PSP packets, and the state self-heals on the next channel switch or open. Would calling mlx5e_accel_update_rx(priv) unconditionally, or before mlx5e_refresh_tirs(), be preferable here?
This condition is still present at the end of the series (a9e41034d482a6cbfe0487f9ed8050c5bf5c3ce1), where mlx5e_update_nic_rx() is unchanged.
Signature · 1 line
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260908065244.3799142-1-tariqt%40nvidia.com
Tariq Toukan
maintainerPATCH v4 12/14 selftests: drv-net: psp: Extract shared helpers into psp_lib.py
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
Put init_psp_dev() and the logic for requiring a psp version into a shared lib, since they will soon be used by a second test file.
No functional change for psp.py.
--- tools/testing/selftests/drivers/net/Makefile | 1 + tools/testing/selftests/drivers/net/psp.py | 77 ++++++------------- .../testing/selftests/drivers/net/psp_lib.py | 57 ++++++++++++++ 3 files changed, 82 insertions(+), 53 deletions(-) create mode 100644 tools/testing/selftests/drivers/net/psp_lib.py
tools/testing/selftests/drivers/net/Makefile +1 -0
@@ -32,6 +32,7 @@ TEST_PROGS := \ TEST_FILES := \ gro_lib.py \ + psp_lib.py \ # end of TEST_FILES # YNL files, must be before "include ..lib.mk"
tools/testing/selftests/drivers/net/psp.py +24 -53
209 lines not on this page · raw message
tools/testing/selftests/drivers/net/psp_lib.py new file +57 -0
57 lines not on this page · raw message
Signature · 1 line
2.44.0
Re: [PATCH net-next V4 12/14] selftests: drv-net: psp: Extract shared helpers into psp_lib.py
Quoted 11 lines · Tariq Toukan
From: Cosmin Ratiu <[email protected]>
Put
init_psp_dev()and the logic for requiring a psp version into a shared lib, since they will soon be used by a second test file.No functional change for
psp.py.Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>---
[email protected]
AI reviewRe: [PATCH net-next V4 12/14] selftests: drv-net: psp: Extract shared helpers into psp_lib.py
In reply to Tariq ToukanSashiko has reviewed this patch and found no issues. It looks great!
Quoted 10 lines
From: Cosmin Ratiu <[email protected]>
Put
init_psp_dev()and the logic for requiring a psp version into a shared lib, since they will soon be used by a second test file.No functional change for
psp.py.Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>
Sashiko has reviewed this patch and found no issues. It looks great!
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=12
Tariq Toukan
maintainerPATCH v4 13/14 selftests: net: gro: Add PSP encapsulation and encryption
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
Add a PSP mode which wraps every transmitted frame in PSP transport mode and encrypts them in SW with the help of OpenSSL.
The receiving NIC decrypts and decapsulates, so the frames seen on the receiver on its AF_PACKET socket are identical to the non-PSP ones and every geometry assertion in this file applies unchanged.
Sender packet construction is unchanged, PSP uses a different buffer to avoid double encryption.
A few PSP-only cases are added: psp_spi_diff, psp_ver_diff, psp_mixed, to cover coalescing scenarios which only exist for PSP.
OpenSSL is optional and scoped to this binary: gro is deployed by several suites, so a missing libcrypto costs only the PSP mode.
--- tools/testing/selftests/net/lib/Makefile | 16 ++ tools/testing/selftests/net/lib/gro.c | 337 ++++++++++++++++++++++- 2 files changed, 351 insertions(+), 2 deletions(-)
tools/testing/selftests/net/lib/Makefile +16 -0
@@ -22,4 +22,20 @@ TEST_INCLUDES := $(wildcard py/*.py sh/*.sh) include ../../lib.mk +# gro optionally needs OpenSSL for PSP. +PKG_CONFIG ?= pkg-config +LIBCRYPTO_CFLAGS := $(shell $(PKG_CONFIG) --cflags libcrypto 2>/dev/null) +LIBCRYPTO_LIBS := $(shell $(PKG_CONFIG) --libs libcrypto 2>/dev/null || \ + echo -lcrypto) +HAVE_OPENSSL := $(shell echo 'int main(void) { return 0; }' | \ + $(CC) -x c - -include openssl/evp.h $(LIBCRYPTO_CFLAGS) \ + $(LIBCRYPTO_LIBS) -o /dev/null >/dev/null 2>&1 && echo 1) + +ifeq ($(HAVE_OPENSSL),1) +$(OUTPUT)/gro: CFLAGS += -DHAVE_OPENSSL $(LIBCRYPTO_CFLAGS) +$(OUTPUT)/gro: LDLIBS += $(LIBCRYPTO_LIBS) +else +$(warning libcrypto not found, gro will be built without PSP support) +endif + include ../bpf.mk
tools/testing/selftests/net/lib/gro.c +335 -2
430 lines not on this page · raw message
Signature · 1 line
2.44.0
Re: [PATCH net-next V4 13/14] selftests: net: gro: Add PSP encapsulation and encryption
Quoted 21 lines · Tariq Toukan
From: Cosmin Ratiu <[email protected]>
Add a PSP mode which wraps every transmitted frame in PSP transport mode and encrypts them in SW with the help of OpenSSL.
The receiving NIC decrypts and decapsulates, so the frames seen on the receiver on itsAF_PACKETsocket are identical to the non-PSP ones and every geometry assertion in this file applies unchanged.Sender packet construction is unchanged, PSP uses a different buffer to avoid double encryption.
A few PSP-only cases are added:
psp_spi_diff,psp_ver_diff,psp_mixed, to cover coalescing scenarios which only exist for PSP.OpenSSL is optional and scoped to this binary: gro is deployed by several suites, so a missing libcrypto costs only the PSP mode.
Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>---
...
Quoted 5 lines
tools/testing/selftests/net/lib/gro.c +5
-0+/* PSP transport mode encapsulation, as built by the sender: +* +* [eth][IP][UDP dport=1000][PSP hdr][encrypted L4][ICV] +* +* Can't use the kernel-only include/net/psp/types.h header, so copy the
tools/testing/selftests/net/lib/gro.c line 136
Quoted 10 lines
tools/testing/selftests/net/lib/gro.c +10
-0+* encoding here. +*/ +struct psphdr { +uint8_t nexthdr; +uint8_t hdrlen; +uint8_t crypt_offset; +uint8_t verfl; +uint32_t spi;/* big endian */ +uint64_t iv;/* big endian */ +} __packed;
My llm is complaining that it thinks __packed is not defined for userspace code and being used as the name of an object being declared, instead of an attribute. Otherwise:
[email protected]
AI reviewRe: [PATCH net-next V4 13/14] selftests: net: gro: Add PSP encapsulation and encryption
In reply to Tariq ToukanThank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] Casting a statically allocated byte array (`psp_scratch`) to struct pointers violates strict…
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Casting a statically allocated byte array (psp_scratch) to struct pointers violates strict aliasing rules and causes potential unaligned memory access traps. --
commit bedf89d5cc02c2e4bc6f000e8888dc6adb4414c6
Author: Cosmin Ratiu <[email protected]>
selftests: net: gro: Add PSP encapsulation and encryption
This commit adds a PSP mode to the gro selftest which wraps transmitted frames in PSP transport mode and encrypts them using OpenSSL. The receiver decapsulates the frames for unchanged geometry assertions.
Quoted 4 lines
tools/testing/selftests/net/lib/gro.c +0
-0
[ ... ]
Quoted 2 lines
tools/testing/selftests/net/lib/gro.c +2
-0+static char *psp_encapsulate (const char *pkt,int *lenp) +{
[ ... ]
Quoted 2 lines
tools/testing/selftests/net/lib/gro.c +2
-0+if (proto ==PF_INET ) { +struct iphdr *iph = (struct iphdr *)(psp_scratch +ETH_HLEN );
[ ... ]
Quoted 2 lines
tools/testing/selftests/net/lib/gro.c +2
-0+ sport = *(__be16 *)(psp_scratch + l4_off); + dport = *(__be16 *)(psp_scratch + l4_off +sizeof (sport));
[ ... ]
tools/testing/selftests/net/lib/gro.c static void fill_transportlayer(void *buf, int seq_offset, int ack_offset, · line 598
Quoted 7 lines
tools/testing/selftests/net/lib/gro.c +7
-0+ udph = (struct udphdr *)(psp_scratch + l4_off); + udph->source = sport ^ dport; + udph->dest =htons (PSP_UDP_PORT ); + udph->len =htons (PSP_UDP_LEN +PSP_HDR_LEN + l4_len +PSP_ICV_LEN ); + udph->check =0 ; + + psph = (struct psphdr *)(udph +1 );
[Severity: Medium]
Does casting a statically allocated byte array directly to structure pointers violate strict aliasing rules?
Looking at psp_encapsulate(), the statically allocated psp_scratch array is directly cast to struct iphdr, __be16, struct udphdr, and struct psphdr pointers.
Because the tools compilation environment assumes standard -fstrict-aliasing, can this cause the compiler to incorrectly optimize away reads or writes?
Additionally, can this lead to fatal unaligned access traps on architectures that enforce strict alignment, since a byte array is not guaranteed to align to the requirements of these structures?
[ ... ]
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=13
[email protected]
AI reviewRe: [PATCH net-next V4 13/14] selftests: net: gro: Add PSP encapsulation and encryption
In reply to Tariq ToukanThank you for your contribution! Sashiko AI review found 7 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 7 potential issue(s) to consider:
Critical: 0 · High: 0 · Medium: 0 · Low: 7
- [Low] struct psphdr { ... } __packed; in tools/testing/selftests/net/lib/gro.c uses the __packed macro, but…
- [Low] The psp_ver_diff sub-test claims to verify that 'different PSP version doesn't coalesce', but it is implemented by the same…
- [Low] The file-level comment block of tools/testing/selftests/net/lib/gro.c enumerates every test case and…
- [Low] Concern (validated as NOT a defect): gro.c gains --psp-assoc and three psp_* sub-tests with no in-tree caller, making the feature…
- [Low] Concern (validated as NOT a defect): psp_encapsulate() hardcodes IPv6 header length and does not walk extension headers or reject…
- [Low] Concern (validated as NOT a defect): psp_encapsulate() recomputes the IPv4 header checksum from scratch, repairing the deliberately…
- [Low] Concern (validated as NOT a defect): each parsed PSP association allocates an EVP_CIPHER_CTX that is never freed, on error paths and at…
--
selftests: net: gro: Add PSP encapsulation and encryption
This adds an optional PSP transport mode to the gro selftest helper: frames are encapsulated and encrypted in software with OpenSSL before being sent, and the receiving NIC is expected to decrypt and decapsulate them. Three PSP-only sub-tests (psp_spi_diff, psp_ver_diff, psp_mixed) and a new --psp-assoc option are added, with libcrypto detection scoped to the gro binary in the Makefile.
A few questions below.
Quoted 4 lines
tools/testing/selftests/net/lib/gro.c +0
-0
[ ... ]
Quoted 9 lines
tools/testing/selftests/net/lib/gro.c +5
-0@@ -110,6 +119,38 @@#define EXIT_OVER_COALESCE 42 +/* PSP transport mode encapsulation, as built by the sender: +* +* [eth][IP][UDP dport=1000][PSP hdr][encrypted L4][ICV] +* +* Can't use the kernel-only include/net/psp/types.h header, so copy the
tools/testing/selftests/net/lib/gro.c line 136
Quoted 10 lines
tools/testing/selftests/net/lib/gro.c +10
-0+* encoding here. +*/ +struct psphdr { +uint8_t nexthdr; +uint8_t hdrlen; +uint8_t crypt_offset; +uint8_t verfl; +uint32_t spi;/* big endian */ +uint64_t iv;/* big endian */ +} __packed;
[Severity: Low]
Where does __packed come from in this translation unit?
The include path for this binary is only:
CFLAGS += -I../../../../../usr/include/ $(KHDR_INCLUDES) CFLAGS += -I../../
scripts/headers_install.sh textually rewrites the macro when exporting uapi headers:
s/(^|[^a-zA-Z0-9])__packed([^a-zA-Z0-9_]|$)/\1__attribute__((packed))\2/g
so the installed headers (including linux/psp.h) never define __packed, and kselftest.h plus net/lib/ksft.h only pull in libc headers. In tools/ the macro is defined by tools/include/linux/compiler.h and tools/include/linux/compiler-gcc.h, neither of which is on this include path.
Doesn't that make the declaration parse as a file-scope tentative definition of an object named __packed of type struct psphdr, leaving the struct unpacked?
The layouts happen to agree today (spi at offset 4, iv at offset 8, sizeof == 16), so PSP_HDR_LEN and the offsetof() used for the nonce are still right. But the unpacked struct carries 8-byte alignment while psph lands at ETH_HLEN + 20 + 8 == 42 inside psp_scratch, i.e. 2 mod 8, so:
psph->spi = htonl(assoc->spi); psph->iv = htobe64(psp_next_iv++);
are stores the compiler is entitled to treat as aligned. Would using __attribute__((packed)) directly, or defining __packed locally as the sgx and bpf selftests do, be preferable here?
Quoted 4 lines
tools/testing/selftests/net/lib/gro.c +4
-0+ +#define PSP_UDP_PORT 1000 +#define PSP_UDP_LEN sizeof(struct udphdr) +#define PSP_HDR_LEN sizeof(struct psphdr)
[ ... ]
Quoted 11 lines
tools/testing/selftests/net/lib/gro.c +7
-0@@ -1111,6 +1374,39 @@ static void send_changed_pppoe_sid(int fd, struct sockaddr_ll *daddr)write_packet (fd, buf, pkt_size, daddr); } +/* PSP packets don't coalesce across SPIs or versions. */ +static void send_psp_assoc_switch (int fd,struct sockaddr_ll *daddr) +{ +static char buf[MAX_HDR_LEN +PAYLOAD_LEN ]; +int pkt_size = total_hdr_len +PAYLOAD_LEN ; + +if (psp_num_assoc <2 )
tools/testing/selftests/net/lib/gro.c static void send_changed_pppoe_sid(int fd, struct sockaddr_ll *daddr) · line 1393
Quoted 10 lines
tools/testing/selftests/net/lib/gro.c +10
-0+error (1 ,0 ,"%s needs two PSP associations" , testname); + +create_packet (buf,0 ,0 ,PAYLOAD_LEN ,0 ); +write_packet (fd, buf, pkt_size, daddr); + +create_packet (buf,PAYLOAD_LEN ,0 ,PAYLOAD_LEN ,0 ); + psp_idx =1 ; +write_packet (fd, buf, pkt_size, daddr); + psp_idx =0 ; +}
[Severity: Low]
Does psp_ver_diff actually exercise the version dimension?
Both psp_spi_diff and psp_ver_diff dispatch to send_psp_assoc_switch(), which only flips psp_idx from association 0 to association 1. Since the two RX associations always carry distinct device-allocated SPIs, the second frame differs in SPI as well as version.
The kernel side compares the whole extension:
include/net/psp/functions.h:__psp_skb_coalesce_diff() {
...
diffs |= memcmp(a, b, sizeof(*a));
...
}over struct psp_skb_ext { __be32 spi; u16 dev_id; u8 generation; u8 version; }, so the differing SPI alone already forces the flush.
Wouldn't psp_ver_diff then keep passing even if the version field were dropped from the software or hardware comparison, making it a duplicate of psp_spi_diff? Since SPIs are allocated by the device and two associations cannot share one, would it make sense to at least assert psp_assocs[0].version != psp_assocs[1].version in the sender, or note in a comment that this case mainly exercises multi-version decap?
Quoted 3 lines
tools/testing/selftests/net/lib/gro.c +0
-0static void bind_packetsocket (int fd) {struct sockaddr_ll daddr = {};
[ ... ]
Quoted 10 lines
tools/testing/selftests/net/lib/gro.c +6
-0@@ -1559,6 +1855,16 @@ static void gro_sender(void)usleep (fin_delay_us);write_packet (txfd, fin_pkt, total_hdr_len, &daddr); +/* PSP sub-tests */ + }else if (strcmp (testname,"psp_spi_diff" ) ==0 || +strcmp (testname,"psp_ver_diff" ) ==0 ) { +send_psp_assoc_switch (txfd, &daddr); +usleep (fin_delay_us); +write_packet (txfd, fin_pkt, total_hdr_len, &daddr);
[ ... ]
Quoted 15 lines
tools/testing/selftests/net/lib/gro.c +11
-0@@ -1782,6 +2088,22 @@ static void gro_receiver(void)printf ("different PPPoE session ID doesn't coalesce: " );check_recv_pkts (rxfd, correct_payload,2 ); +/* PSP sub-tests */ + }else if (strcmp (testname,"psp_spi_diff" ) ==0 ) { + correct_payload[0 ] =PAYLOAD_LEN ; + correct_payload[1 ] =PAYLOAD_LEN ; +printf ("different PSP SPI doesn't coalesce: " ); +check_recv_pkts (rxfd, correct_payload,2 ); + }else if (strcmp (testname,"psp_ver_diff" ) ==0 ) { + correct_payload[0 ] =PAYLOAD_LEN ; + correct_payload[1 ] =PAYLOAD_LEN ; +printf ("different PSP version doesn't coalesce: " ); +check_recv_pkts (rxfd, correct_payload,2 );
[ ... ]
tools/testing/selftests/net/lib/gro.c static void parse_args(int argc, char **argv) · line 2127
Quoted 5 lines
tools/testing/selftests/net/lib/gro.c +1
-0@@ -1802,6 +2124,7 @@ static void parse_args(int argc, char **argv) { "ip6ip6", no_argument, NULL, 'E' }, { "pppoev4", no_argument, NULL, 'p' }, { "pppoev6", no_argument, NULL, 'P' }, + { "psp-assoc", required_argument, NULL, OPT_PSP_ASSOC },
[Severity: Low]
Should the file-level comment block at the top of gro.c be extended too?
It enumerates every sub-test family (data_*, flags_*, tcp_*, ip_*, large_*, single, capacity), but psp_spi_diff, psp_ver_diff, psp_mixed and the new --psp-assoc version,spi,key option are not listed there, and the last patch of the series does not touch gro.c either.
The requirement that both sides be invoked with the same --psp-assoc arguments is also undocumented, even though the receiver's expected sizes depend on it:
static int calc_mss(void)
{
return ASSUMED_MTU - (total_hdr_len - ETH_HLEN) -
(psp_enabled ? PSP_ENCAP_LEN : 0);
}Could that coupling be spelled out next to the other test descriptions?
Signature · 1 line
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260908065244.3799142-1-tariqt%40nvidia.com
Tariq Toukan
maintainerPATCH v4 14/14 selftests: drv-net: Add PSP HW GRO conformance tests
In reply to Tariq ToukanFrom: Cosmin Ratiu <[email protected]>
Add PSP conformance tests using the same gro helper binary as other gro tests , but in --psp mode. SPIs are procured by psp_gro.py from a real PSP device and handed off to the gro sender & receiver. The sender crafts and encrypts packets in software, the receiver relies fully on the HW to decrypt, decapsulate and do HW GRO.
Because the NIC decrypts and decapsulates before the frames reach the receiver AF_PACKET tap, the gro receiver sees the plain frames and the assertions remain exactly the same as other gro tests.
So these tests verify, at length, that the device does PSP HW-GRO the same way as plain TCP.
Additional PSP-specific tests are defined to check that GRO doesn't merge packets across PSP versions, SPIs, encryption-status, etc.
Some gro tests are not included because they don't work:
- ip_csum: the checksum is recomputed by psp_dev_rcv().
- tcp_csum: packets are marked with CHECKSUM_UNNECESSARY.
- ip_frag4/ip_frag6: PSP is incompatible with IP fragmentation.
- IPv6 extension header tests: PSP doesn't deal with IPv6 ext headers.
--- .../testing/selftests/drivers/net/gro_lib.py | 15 +- .../testing/selftests/drivers/net/hw/Makefile | 17 ++ tools/testing/selftests/drivers/net/hw/config | 1 + .../selftests/drivers/net/hw/psp_gro.py | 157 ++++++++++++++++++ 4 files changed, 186 insertions(+), 4 deletions(-) create mode 100755 tools/testing/selftests/drivers/net/hw/psp_gro.py
tools/testing/selftests/drivers/net/gro_lib.py +11 -4
@@ -189,7 +189,8 @@ def _setup_queue_count(cfg, num_queues): def _run_gro_bin(cfg, test_name, protocol=None, num_flows=None, - order_check=False, verbose=False, fail=False): + order_check=False, verbose=False, fail=False, + common_args=None): """Run gro binary with given test and return the process result.""" if not hasattr(cfg, "bin_remote"): cfg.bin_local = cfg.net_lib_dir / "gro"
@@ -217,6 +218,8 @@ def _run_gro_bin(cfg, test_name, protocol=None, num_flows=None, base_args.append("--order-check") if verbose: base_args.append("--verbose") + if common_args: + base_args += common_args args = " ".join(base_args)
@@ -341,8 +344,11 @@ def _gro_variants(): yield protocol, test_name -def run_test(cfg, mode, protocol, test_name): - """Run a single GRO test with retries.""" +def run_test(cfg, mode, protocol, test_name, common_args=None): + """Run a single GRO test with retries. + + common_args are extra arguments passed to both gro sender and receiver. + """ ipver = "6" if protocol[-1] == "6" else "4" cfg.require_ipver(ipver)
@@ -356,7 +362,8 @@ def run_test(cfg, mode, protocol, test_name): for attempt in range(max_retries): fail_now = attempt >= max_retries - 1 rx_proc = _run_gro_bin(cfg, test_name, protocol=protocol, - verbose=True, fail=fail_now) + verbose=True, fail=fail_now, + common_args=common_args) if rx_proc.ret == 0: return
tools/testing/selftests/drivers/net/hw/Makefile +17 -0
@@ -13,6 +13,20 @@ else $(warning excluding iouring tests, liburing not installed or too old) endif +# psp_gro.py uses the gro binary in PSP mode, which requires OpenSSL. +PKG_CONFIG ?= pkg-config +HAVE_OPENSSL := $(shell echo 'int main(void) { return 0; }' | \ + $(CC) -x c - -include openssl/evp.h \ + $(shell $(PKG_CONFIG) --cflags --libs libcrypto 2>/dev/null \ + || echo -lcrypto) \ + -o /dev/null >/dev/null 2>&1 && echo 1) + +ifeq ($(HAVE_OPENSSL),1) +COND_PROGS += psp_gro.py +else +$(warning excluding PSP GRO tests, libcrypto not installed) +endif + TEST_GEN_FILES := \ $(COND_GEN_FILES) \ # end of TEST_GEN_FILES
@@ -40,6 +54,7 @@ TEST_PROGS = \ nk_qlease.py \ ntuple.py \ pp_alloc_fail.py \ + $(COND_PROGS) \ rss_api.py \ rss_ctx.py \ rss_drv.py \
@@ -64,6 +79,8 @@ TEST_FILES := \ TEST_INCLUDES := \ $(wildcard lib/py/*.py ../lib/py/*.py) \ + ../gro_lib.py \ + ../psp_lib.py \ ../../../net/lib.sh \ ../../../net/forwarding/ipip_lib.sh \ ../../../net/forwarding/lib.sh \
tools/testing/selftests/drivers/net/hw/config +1 -0
@@ -8,6 +8,7 @@ CONFIG_INET6_ESP=y CONFIG_INET6_ESP_OFFLOAD=y CONFIG_INET_ESP=y CONFIG_INET_ESP_OFFLOAD=y +CONFIG_INET_PSP=y CONFIG_IO_URING=y CONFIG_IPV6=y CONFIG_IPV6_GRE=y
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +157 -0
157 lines not on this page · raw message
Signature · 1 line
2.44.0
Re: [PATCH net-next V4 14/14] selftests: drv-net: Add PSP HW GRO conformance tests
Quoted 12 lines · Tariq Toukan
From: Cosmin Ratiu <[email protected]>
Add PSP conformance tests using the same gro helper binary as other gro tests , but in --psp mode. SPIs are procured by
psp_gro.pyfrom a real PSP device and handed off to the gro sender & receiver. The sender crafts and encrypts packets in software, the receiver relies fully on the HW to decrypt, decapsulate and do HW GRO.Because the NIC decrypts and decapsulates before the frames reach the receiver
AF_PACKETtap, the gro receiver sees the plain frames and the assertions remain exactly the same as other gro tests.
Quoted 8 lines
So these tests verify, at length, that the device does PSP HW-GRO the same way as plain TCP.
Additional PSP-specific tests are defined to check that GRO doesn't merge packets across PSP versions, SPIs, encryption-status, etc.
Some gro tests are not included because they don't work:
-ip_csum: the checksum is recomputed bypsp_dev_rcv().
psp_dev_rcv() isn't on the hw gro path for mlx5, right?
Quoted 1 line
-
tcp_csum: packets are marked withCHECKSUM_UNNECESSARY.
this also seems strange to me as the hw gro engine and its decision to coalesce two packets sits upstream of this skb->ip_summed determination in the driver, right?
Quoted 7 lines
- ip_frag4/ip_frag6: PSP is incompatible with IP fragmentation.
- IPv6 extension header tests: PSP doesn't deal with IPv6 ext headers.Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>---
...
tools/testing/selftests/drivers/net/hw/psp_gro.py line 119
Quoted 6 lines
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +6
-0+ +@ksft_variants(_ip_variants()) +def test_psp_ver_diff(cfg, protocol): + """Frames from two different PSP versions must not coalesce.""" + init_psp_dev(cfg) + if len(cfg.psp_info['psp-versions-cap']) < 2:
nit: what the above checks is slightly different than what the _run() below uses, e.g. dev supports versions 0 and 2.
tools/testing/selftests/drivers/net/hw/psp_gro.py line 123
Quoted 4 lines
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +4
-0+ raise KsftSkipEx("Device supports a single PSP version") + + _run(cfg, "psp_ver_diff", protocol, [0, 1]) +
couple of minor comments:
Re: [PATCH net-next V4 14/14] selftests: drv-net: Add PSP HW GRO conformance tests
Quoted 22 lines · Daniel Zahka
Quoted 21 lines · Tariq Toukan
From: Cosmin Ratiu <[email protected]>
Add PSP conformance tests using the same gro helper binary as other gro
tests , but in --psp mode. SPIs are procured bypsp_gro.pyfrom a real
PSP device and handed off to the gro sender & receiver. The sender crafts and encrypts packets in software, the receiver relies fully on
the HW to decrypt, decapsulate and do HW GRO.Because the NIC decrypts and decapsulates before the frames reach the
receiverAF_PACKETtap, the gro receiver sees the plain frames and the
assertions remain exactly the same as other gro tests.So these tests verify, at length, that the device does PSP HW-GRO the
same way as plain TCP.
Quoted 7 lines
Quoted 5 lines
Additional PSP-specific tests are defined to check that GRO doesn't merge packets across PSP versions, SPIs, encryption-status, etc.
Some gro tests are not included because they don't work:
-ip_csum: the checksum is recomputed bypsp_dev_rcv().
psp_dev_rcv()isn't on the hw gro path for mlx5, right?
You're right, that's a bogus argument, sorry.
The truth was that the encap support in gro.c was recomputing the IP checksum, sabotaging the bad csum test... Will be fixed in the next version.
Quoted 7 lines
Quoted 1 line
-
tcp_csum: packets are marked withCHECKSUM_UNNECESSARY.this also seems strange to me as the hw gro engine and its decision to
coalesce two packets sits upstream of thisskb->ip_summeddetermination
in the driver, right?
Another weak argument from my side, sorry. I will add the test back.
Quoted 10 lines
Quoted 9 lines
- ip_frag4/ip_frag6: PSP is incompatible with IP fragmentation.
- IPv6 extension header tests: PSP doesn't deal with IPv6 ext headers.Signed-off-by Cosmin Ratiu <[email protected]>Reviewed-by Dragos Tatulea <[email protected]>Signed-off-by Tariq Toukan <[email protected]>---
...
Quoted 10 lines
Quoted 7 lines
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +4
-0+ +@ksft_variants(_ip_variants()) +def test_psp_ver_diff(cfg, protocol): + """Frames from two different PSP versions must notcoalesce."""
+init_psp_dev(cfg)
+ iflen(cfg.psp_info['psp-versions-cap'])< 2:nit: what the above checks is slightly different than what the
_run()below uses, e.g. dev supports versions 0 and 2.
Will drop this check.
Quoted 9 lines
tools/testing/selftests/drivers/net/hw/psp_gro.py line 123
Quoted 5 lines
+ raise
KsftSkipEx("Device supports a single PSP version")tools/testing/selftests/drivers/net/hw/psp_gro.py new file +3
-0+ + _run(cfg, "psp_ver_diff", protocol, [0, 1]) +couple of minor comments:
Reviewed-by Daniel Zahka <[email protected]>
Thank you for the review!
Cosmin.
[email protected]
AI reviewRe: [PATCH net-next V4 14/14] selftests: drv-net: Add PSP HW GRO conformance tests
In reply to Tariq ToukanThank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider: - [Medium] Hardcoded PSP versions [0, 1] in test_psp_ver_diff instead of dynamically using supported…
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [Medium] Hardcoded PSP versions [0, 1] in test_psp_ver_diff instead of dynamically using supported capabilities.
- [Medium] Unconditional creation of AF_INET6 socket crashes on IPv6-disabled systems.
--
commit 2fbb01e97add94de0a14266aa0b2565a8246c7ea
Author: Cosmin Ratiu <[email protected]>
selftests: drv-net: Add PSP HW GRO conformance tests
This commit adds PSP conformance tests to kselftests using the gro helper binary in --psp mode. The sender crafts encapsulated and SW-encrypted PSP packets, and the receiver's device decrypts and decapsulates the packets before an AF_PACKET tap gets to analyze them.
[ ... ]
Quoted 5 lines
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +0
-0
[ ... ]
tools/testing/selftests/drivers/net/hw/psp_gro.py line 74
Quoted 3 lines
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +3
-0+def _psp_assoc(cfg, version=0): + # This socket receives no traffic, exists solely to own the rx assoc. + s = socket.socket(socket.AF_INET6, socket.SOCK_STREAM)
[Severity: Medium]
Will unconditionally creating an AF_INET6 socket here crash the test framework with an OSError (EAFNOSUPPORT) on systems where IPv6 is disabled at boot?
If the test is running an IPv4 variant on such a system, should the socket creation match the protocol family, or gracefully skip the test instead of failing?
[ ... ]
tools/testing/selftests/drivers/net/hw/psp_gro.py line 122
Quoted 8 lines
+@ksft_variants(_ip_variants())
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +7
-0+def test_psp_ver_diff(cfg, protocol): + """Frames from two different PSP versions must not coalesce.""" + init_psp_dev(cfg) + if len(cfg.psp_info['psp-versions-cap']) < 2: + raise KsftSkipEx("Device supports a single PSP version") + + _run(cfg, "psp_ver_diff", protocol, [0, 1])
[Severity: Medium]
Could this hardcoded list of versions [0, 1] cause the test to fail on devices that support multiple versions but do not support version 0 (for example, if they only support versions 1 and 2)?
Since the test verifies len(cfg.psp_info['psp-versions-cap']) < 2 to ensure at least two versions are supported, would it be better to dynamically pick the first two versions from that capabilities array instead of hardcoding them?
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=14
[email protected]
AI reviewRe: [PATCH net-next V4 14/14] selftests: drv-net: Add PSP HW GRO conformance tests
In reply to Tariq ToukanThank you for your contribution! Sashiko AI review found 5 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 5 potential issue(s) to consider:
Critical: 0 · High: 0 · Medium: 2 · Low: 3
- [Medium] test_psp_ver_diff() in tools/testing/selftests/drivers/net/hw/psp_gro.py gates on the *count*…
- [Medium] test_psp_after_reconfig() (tools/testing/selftests/drivers/net/hw/psp_gro.py:132-141) disables…
- [Low] The new HAVE_OPENSSL probe in tools/testing/selftests/drivers/net/hw/Makefile (lines 16-28) decides…
- [Low] test_psp_ver_diff() (tools/testing/selftests/drivers/net/hw/psp_gro.py:116-122) claims to…
- [Low] The psp_gro.py module docstring (line 7: "This reuses the gro binary in --psp mode:") and the new Makefile comment ("psp_gro.py uses…
--
selftests: drv-net: Add PSP HW GRO conformance tests
This adds psp_gro.py, which drives the existing gro helper binary with PSP associations obtained from a real PSP device, so the sender encrypts in software and the receiver relies on the NIC for decrypt, decap and HW GRO. It also plumbs a common_args parameter through gro_lib.run_test() and adds a libcrypto probe to the hw Makefile.
No Link: tags in the commit header.
Quoted 2 lines
Add PSP conformance tests using the same gro helper binary as other gro tests , but in --psp mode. SPIs are procured by
psp_gro.pyfrom a real
[Severity: Low]
The commit message, the new Makefile comment ("psp_gro.py uses the gro binary in PSP mode") and the psp_gro.py docstring all describe a "--psp mode", but the gro helper has no such option.
In tools/testing/selftests/net/lib/gro.c the only PSP entry in long_options[] is:
{ "psp-assoc", required_argument, NULL, OPT_PSP_ASSOC },
and psp_enabled is just a side effect of psp_parse_assoc(). _psp_args() in psp_gro.py correctly emits --psp-assoc ver,spi,key. Could the wording be changed to --psp-assoc so nobody tries to pass --psp and gets an unknown-option error from getopt_long()?
Quoted 11 lines
tools/testing/selftests/drivers/net/hw/Makefile +3
-0@@ -13,6 +13,20 @@ else $(warning excluding iouring tests, liburing not installed or too old) endif +# psp_gro.py uses the gro binary in PSP mode, which requires OpenSSL. +PKG_CONFIG ?= pkg-config +HAVE_OPENSSL := $(shell echo 'int main(void) { return 0; }' | \
tools/testing/selftests/drivers/net/hw/Makefile else · line 28
Quoted 10 lines
tools/testing/selftests/drivers/net/hw/Makefile +10
-0+ $(CC) -x c - -include openssl/evp.h \ + $(shell $(PKG_CONFIG) --cflags --libs libcrypto 2>/dev/null \ + || echo -lcrypto) \ + -o /dev/null >/dev/null 2>&1 && echo 1) + +ifeq ($(HAVE_OPENSSL),1) +COND_PROGS += psp_gro.py +else +$(warning excluding PSP GRO tests, libcrypto not installed) +endif
[Severity: Low]
Is the build host the right place to test this condition? The PSP software encryption runs in the gro binary on the remote endpoint, not on the build host:
gro_lib.py:_run_gro_bin()
cfg.bin_local = cfg.net_lib_dir / "gro"
cfg.bin_remote = cfg.remote.deploy(cfg.bin_local)
...
tx_cmd = f"{cfg.bin_remote} {args} --iface {cfg.remote_ifname}"
...
cmd(tx_cmd, host=cfg.remote)gro is dynamically linked against libcrypto:
tools/testing/selftests/net/lib/Makefile
$(OUTPUT)/gro: LDLIBS += $(LIBCRYPTO_LIBS)
and Remote.deploy() only scp's the executable:
tools/testing/selftests/drivers/net/lib/py/remote_ssh.py
cmd(f"scp {what} {self.name}:{file_name}")So on a setup where the build host has libcrypto but the remote endpoint lacks the runtime library, psp_gro.py is installed and every case fails with a dynamic linker error on the sender instead of skipping. Would a runtime check on the remote be preferable here?
[ ... ]
tools/testing/selftests/drivers/net/hw/psp_gro.py line 7
Quoted 13 lines
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +7
-0@@ -0,0 +1,157 @@ +#!/usr/bin/env python3 +# SPDX-License-Identifier: GPL-2.0 + +""" +PSP HW GRO conformance tests. + +This reuses the gro binary in --psp mode:
[Severity: Low]
Same --psp naming question as in the commit message: the implemented option in gro.c is --psp-assoc.
Quoted 6 lines
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +6
-0+The sender crafts encapsulated & SW-encrypted PSP packets with receiver's PSP +rx-assoc, and the receiver's device decrypts and decapsulates the packets +before an AF_PACKET tap gets to analyze them. + +All GRO conformance tests which could run with PSP are included. +"""
[ ... ]
Quoted 19 lines
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +19
-0+def _setup(cfg, version): + """Enables PSP on the device under test.""" + init_psp_dev(cfg) + require_version(cfg, version) + + +def _psp_args(cfg, versions): + """Produces PSP associations as gro binary --psp-assoc arguments.""" + keys = [_psp_assoc(cfg, ver) for ver in versions] + return [f"--psp-assoc {ver},{key['spi']:x},{key['key'].hex()}" + for ver, key in zip(versions, keys)] + + +def _run(cfg, test_name, protocol, versions): + """Sets up, associates and runs one gro test case under PSP + HW GRO.""" + _setup(cfg, max(versions)) + + run_test(cfg, "hw", protocol, test_name, + common_args=_psp_args(cfg, versions))
[ ... ]
tools/testing/selftests/drivers/net/hw/psp_gro.py line 122
Quoted 8 lines
+@ksft_variants(_ip_variants())
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +7
-0+def test_psp_ver_diff(cfg, protocol): + """Frames from two different PSP versions must not coalesce.""" + init_psp_dev(cfg) + if len(cfg.psp_info['psp-versions-cap']) < 2: + raise KsftSkipEx("Device supports a single PSP version") + + _run(cfg, "psp_ver_diff", protocol, [0, 1])
[Severity: Medium]
Does the guard here match what the test then asks for? It counts how many versions the device advertises, but the following call hardcodes versions 0 and 1.
psp-versions-cap is an independent flag bitmask in Documentation/netlink/specs/psp.yaml:
- name: psp-versions-cap
doc: Bitmask of PSP versions supported by the device.
type: u32
enum: version
enum-as-flags: true
so the supported set is not required to be a contiguous prefix starting at 0. On a device advertising, say, hdr0-aes-gcm-128 (0) and hdr0-aes-gmac-128 (2), the length check passes and then:
psp_gro.py:_run()
_setup(cfg, max([0, 1])) /* == 1 */
psp_lib.py:require_version()
name = cfg.pspnl.consts["version"].entries_by_val[version].name if name in cfg.psp_info['psp-versions-cap']:
return
...
raise KsftSkipEx("PSP version not supported", name)
which skips with a reason unrelated to the guard, and the cross-version non-coalescing assertion never runs on a device that could run it.
The existing peer test picks the versions from the actual caps instead:
tools/testing/selftests/drivers/net/psp.py:assoc_version_mismatch()
versions = [cfg.pspnl.consts["version"].entries[v].value for v in versions]
Would deriving the two versions from cfg.psp_info['psp-versions-cap'] be better here? Today mlx5 happens to advertise exactly versions 0 and 1 (psp_supported_versions[] = { GCM_128, GCM_256 }), so this only bites the next driver that advertises a GMAC version.
[Severity: Low]
This isn't a bug, but does psp_ver_diff assert anything that psp_spi_diff does not? The two associations passed here differ in version and in SPI, since each rx_assoc allocates a fresh device SPI (psp_nl_rx_assoc_doit() -> psd->ops->rx_spi_alloc()).
On the gro.c side both cases share the same sender helper, which only flips psp_idx:
send_psp_assoc_switch()
create_packet(buf, PAYLOAD_LEN, 0, PAYLOAD_LEN, 0); psp_idx = 1; write_packet(fd, buf, pkt_size, daddr); psp_idx = 0;
and the receive expectations are byte-for-byte the same as psp_spi_diff (correct_payload[0] = correct_payload[1] = PAYLOAD_LEN, then check_recv_pkts(rxfd, correct_payload, 2)); only the printf text differs. A version-only difference does not look constructible from user space, so this may be the closest achievable approximation - it is just worth noting that the case would pass even if the GRO path ignored the version field.
[ ... ]
tools/testing/selftests/drivers/net/hw/psp_gro.py line 138
Quoted 8 lines
+@ksft_variants(_ip_variants())
tools/testing/selftests/drivers/net/hw/psp_gro.py new file +7
-0+def test_psp_after_reconfig(cfg, protocol): + """Verifies that decap still works after PSP off + on.""" + _setup(cfg, 0) + + cap = cfg.psp_info['psp-versions-cap'] + cfg.pspnl.dev_set({'id': cfg.psp_dev_id, 'psp-versions-ena': []}) + cfg.pspnl.dev_set({'id': cfg.psp_dev_id, 'psp-versions-ena': cap})
[Severity: Medium]
Should the disable here be paired with a defer() for the restore? As written, the only thing that re-enables PSP is the very next dev_set() call, with nothing covering an NlError from it or a kill between the two calls.
init_psp_dev() only registers a rollback when the cached psp_info says the enabled set differs from the caps:
tools/testing/selftests/drivers/net/psp_lib.py:init_psp_dev()
if cap != ena:
cfg.pspnl.dev_set({'id': cfg.psp_dev_id, 'psp-versions-ena': cap})
defer(cfg.pspnl.dev_set, {'id': cfg.psp_dev_id,
'psp-versions-ena': ena})so on a device discovered with PSP already fully enabled (cap == ena) there is no defer at all. If the re-enable does not happen, the NIC is left with PSP offload off while cfg.psp_info still caches the old enabled set, and later init_psp_dev() calls compare against that stale snapshot and conclude nothing needs enabling - subsequent PSP tests then fail on rx-assoc/decap in a way that looks like a driver problem.
Would registering the re-enable via defer() right after the disable, and re-reading the device state rather than trusting the cached psp_info, address that?
[ ... ]
Signature · 1 line
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260908065244.3799142-1-tariqt%40nvidia.com
Jakub Kicinski
maintainerRe: [PATCH net-next V4 14/14] selftests: drv-net: Add PSP HW GRO conformance tests
In reply to Tariq ToukanQuoted 15 lines · On Tue, 8 Sep 2026 09:52:44 +0300 Tariq Toukan wrote:
From: Cosmin Ratiu <[email protected]>
Add PSP conformance tests using the same gro helper binary as other gro tests , but in --psp mode. SPIs are procured by
psp_gro.pyfrom a real PSP device and handed off to the gro sender & receiver. The sender crafts and encrypts packets in software, the receiver relies fully on the HW to decrypt, decapsulate and do HW GRO.Because the NIC decrypts and decapsulates before the frames reach the receiver
AF_PACKETtap, the gro receiver sees the plain frames and the assertions remain exactly the same as other gro tests.So these tests verify, at length, that the device does PSP HW-GRO the same way as plain TCP.
Quoted 8 lines
Additional PSP-specific tests are defined to check that GRO doesn't merge packets across PSP versions, SPIs, encryption-status, etc.
Some gro tests are not included because they don't work:
-ip_csum: the checksum is recomputed bypsp_dev_rcv().
-tcp_csum: packets are marked withCHECKSUM_UNNECESSARY.
- ip_frag4/ip_frag6: PSP is incompatible with IP fragmentation.
- IPv6 extension header tests: PSP doesn't deal with IPv6 ext headers.
Daniel commented on the csum tests already, it'd also be good to understand what the problem with frag tests is. Presumably they get passed thru still encrypted? Or dropped and counted as errors?
Quoted 11 lines
tools/testing/selftests/drivers/net/hw/Makefile +3
-0@@ -13,6 +13,20 @@ else $(warning excluding iouring tests, liburing not installed or too old) endif +# psp_gro.py uses the gro binary in PSP mode, which requires OpenSSL. +PKG_CONFIG ?= pkg-config +HAVE_OPENSSL := $(shell echo 'int main(void) { return 0; }' | \
tools/testing/selftests/drivers/net/hw/Makefile else · line 28
Quoted 10 lines
tools/testing/selftests/drivers/net/hw/Makefile +10
-0+ $(CC) -x c - -include openssl/evp.h \ + $(shell $(PKG_CONFIG) --cflags --libs libcrypto 2>/dev/null \ + || echo -lcrypto) \ + -o /dev/null >/dev/null 2>&1 && echo 1) + +ifeq ($(HAVE_OPENSSL),1) +COND_PROGS += psp_gro.py +else +$(warning excluding PSP GRO tests, libcrypto not installed) +endif
Looks like slop. Why do this? Just run the test and see if it works. Don't hide the python tests because build was lacking. We explicitly want to see all tests run and fail if something is missing in the build setup in NIPA.
If you really want you can "probe" or display the support in some --help message from the binary and check at runtime.
Quoted 12 lines
tools/testing/selftests/drivers/net/hw/Makefile +1
-0TEST_GEN_FILES := \ $(COND_GEN_FILES) \ # end of TEST_GEN_FILES@@ -40,6 +54,7 @@ TEST_PROGS = \ nk_qlease.py \ ntuple.py \ pp_alloc_fail.py \ + $(COND_PROGS) \ rss_api.py \ rss_ctx.py \ rss_drv.py \
tools/testing/selftests/drivers/net/hw/Makefile TEST_FILES := \ · line 83
Quoted 5 lines
tools/testing/selftests/drivers/net/hw/Makefile +2
-0@@ -64,6 +79,8 @@ TEST_FILES := \ TEST_INCLUDES := \ $(wildcard lib/py/*.py ../lib/py/*.py) \ + ../gro_lib.py \ + ../psp_lib.py \
ugh, just move the test down to the drv-net dir, and Xfail for netdevsim. Both PSP and GRO live in ../ better to keep this test there as well.
tools/testing/selftests/drivers/net/hw/Makefile TEST_FILES := \
Quoted 10 lines
tools/testing/selftests/drivers/net/hw/Makefile +10
-0+# Tests for both IP versions, a subset of gro tests. Missing: +# - tcp_csum: PSP packets are validated by HW (with the ICV) and marked with +# CHECKSUM_UNNECESSARY. +_COMMON = ["data_same", "data_lrg_sml", "data_sml_lrg", "data_lrg_1byte", + "data_burst", + "ack", + "flags_psh", "flags_syn", "flags_rst", "flags_urg", "flags_cwr", + "tcp_seq", "tcp_ts", "tcp_opt", + "ip_ecn", "ip_tos", + "large_max", "large_rem",
Really not great that we have to duplicate all the cases under psp. Narcisa is adding BIG TCP cases for instance, how are we going to keep this in sync :S
Re: [PATCH net-next V4 14/14] selftests: drv-net: Add PSP HW GRO conformance tests
Quoted 32 lines · Jakub Kicinski
Quoted 24 lines · On Tue, 8 Sep 2026 09:52:44 +0300 Tariq Toukan wrote:
From: Cosmin Ratiu <[email protected]>
Add PSP conformance tests using the same gro helper binary as other gro
tests , but in --psp mode. SPIs are procured bypsp_gro.pyfrom a real
PSP device and handed off to the gro sender & receiver. The sender crafts and encrypts packets in software, the receiver relies fully on
the HW to decrypt, decapsulate and do HW GRO.Because the NIC decrypts and decapsulates before the frames reach the
receiverAF_PACKETtap, the gro receiver sees the plain frames and the
assertions remain exactly the same as other gro tests.So these tests verify, at length, that the device does PSP HW-GRO the
same way as plain TCP.Additional PSP-specific tests are defined to check that GRO doesn't merge packets across PSP versions, SPIs, encryption-status, etc.
Quoted 6 lines
Some gro tests are not included because they don't work:
-ip_csum: the checksum is recomputed bypsp_dev_rcv().
-tcp_csum: packets are marked withCHECKSUM_UNNECESSARY.
- ip_frag4/ip_frag6: PSP is incompatible with IP fragmentation.
- IPv6 extension header tests: PSP doesn't deal with IPv6 ext headers.
Quoted 3 lines
Daniel commented on the csum tests already, it'd also be good to understand what the problem with frag tests is. Presumably they get passed thru still encrypted? Or dropped and counted as errors?
It seems ConnectX HW doesn't parse IP fragments beyond the IP header in steering (presumably there are too many corner cases), so the PSP rule selecting packets with UDP dport == 1000 doesn't decrypt those packets and they get dropped by SW. IPsec behaves similarly, IP fragments are just left to SW, but for PSP there's no such option.
In any case, this is probably HW-specific and there's no reason for this test not to be included, so I will include it in the next version.
Quoted 24 lines
Quoted 24 lines
tools/testing/selftests/drivers/net/hw/Makefile +0
-0@@ -13,6 +13,20 @@ else$(warning excluding iouring tests, liburing not installed or too old)
endif
+#psp_gro.pyuses the gro binary in PSP mode, which requires OpenSSL.tools/testing/selftests/drivers/net/hw/Makefile +4
-0+PKG_CONFIG ?= pkg-config +HAVE_OPENSSL := $(shell echo 'int main(void) { return 0; }' | \ + $(CC) -x c - -include openssl/evp.h \ + $(shell $(PKG_CONFIG) --cflags --libs libcrypto2>/dev/null \
tools/testing/selftests/drivers/net/hw/Makefile +6
-0+ || echo -lcrypto) \ + -o /dev/null >/dev/null 2>&1 && echo 1) + +ifeq ($(HAVE_OPENSSL),1) +COND_PROGS += psp_gro.py +else
Quoted 8 lines
Quoted 2 lines
+$(warning excluding PSP GRO tests, libcrypto not installed) +endif
Looks like slop. Why do this? Just run the test and see if it works. Don't hide the python tests because build was lacking. We explicitly want to see all tests run and fail if something is missing in the build
setup in NIPA.
Will do.
Quoted 18 lines
If you really want you can "probe" or display the support in some -- help
message from the binary and check at runtime.Quoted 13 lines
TEST_GEN_FILES := \
$(COND_GEN_FILES) \# end of TEST_GEN_FILEStools/testing/selftests/drivers/net/hw/Makefile +1
-0@@ -40,6 +54,7 @@ TEST_PROGS = \ nk_qlease.py \ ntuple.py \ pp_alloc_fail.py \ + $(COND_PROGS) \ rss_api.py \ rss_ctx.py \ rss_drv.py \@@ -64,6 +79,8 @@ TEST_FILES := \
Quoted 8 lines
tools/testing/selftests/drivers/net/hw/Makefile TEST_FILES := \ · line 83
Quoted 4 lines
TEST_INCLUDES := \
$(wildcardlib/py/*.py../lib/py/*.py) \tools/testing/selftests/drivers/net/hw/Makefile +2
-0+ ../gro_lib.py \ + ../psp_lib.py \ugh, just move the test down to the drv-net dir, and Xfail for netdevsim. Both PSP and GRO live in ../ better to keep this test there as well.
It was there originally, but I figured since PSP is a HW-only protocol, the test belongs in hw/. But I'll move it back, it makes more sense in this context.
Quoted 10 lines
Quoted 10 lines
tools/testing/selftests/drivers/net/hw/Makefile +2
-0+# Tests for both IP versions, a subset of gro tests. Missing: +# - tcp_csum: PSP packets are validated by HW (with the ICV) andmarked with
tools/testing/selftests/drivers/net/hw/Makefile +2
-0+# CHECKSUM_UNNECESSARY. +_COMMON = ["data_same", "data_lrg_sml", "data_sml_lrg","
data_lrg_1byte",
+ "data_burst",
+ "ack",
+ "flags_psh", "flags_syn", "flags_rst", "flags_urg",
Quoted 9 lines
Quoted 4 lines
"
flags_cwr",
+ "tcp_seq", "tcp_ts", "tcp_opt",
+ "ip_ecn", "ip_tos",
+ "large_max", "large_rem",Really not great that we have to duplicate all the cases under psp. Narcisa is adding BIG TCP cases for instance, how are we going to keep this in sync :S
Will drop the lists and just use the gro_variants() generator, with a filter on top to be able to avoid some tests.
Cosmin.
Quoted 367 lines · Tariq Toukan
Hi,
Ingress PSP packets cannot be merged by the HW-GRO HW state machine because they are not decapsulated and the current HW-GRO state machine
does not understand PSP.This series by Cosmin decapsulates PSP packets in steering, which allows
the now-decapsulated PSP packets (== TCP) to go through HW-GRO and be aggregated. The SPI and PSP version from the PSP header are handed off
to the driver in the CQE metadata fields. They are used to terminate the
HW GRO session on mismatch, and are required to construct the skb extension which is used higher up in the stack.Some preparatory work needed to happen to allow that:
- All accel protocol markers were moved away fromft_metadataintoflow_tag
- Mutual exclusion between TC and accel protocols was added. Trying to
configure both IPsec and TC at the same time for example is now detected at config time instead of doing weird things at runtime.kperf tests on a pair of CX7 NICs with 200Gbps link speed:
Streams Gbps no HW-GRO Gbps HW-GRO Speedup
------- -------------- ----------- -------
1 28 58 2.07x
2 67 102 1.52x
4 136 180 1.32x
8 175 183 1.05x
Regards,
TariqSome internal Sashiko findings, plus Cosmin's comments:
Quoted 3 lines
+
cfg.wait_hw_stats_settle()
+ after =cfg.netnl.qstats_get({"ifindex": cfg.ifindex}, dump=True)[0]Could this lead to flaky test failures on multiqueue NICs? By hardcoding the netlink dump array access to index [0], the test statically
verifies statistics for the first queue only. Since the test sends traffic over
an ephemeral random port, Receive Side Scaling (RSS) will hash this 4-tuple and
could deliver the flow to any active RX queue. If the flow lands on a queue
other than queue 0, will the test fail to observe the hardware GRO packet
increments?[CR] The premise is wrong. [0] is not queue 0, it's the result for the
requested dev.Quoted 12 lines
(quoted diff) +0
-2@@ -1757,11 +1755,10 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, rx_tables = &rx_fs->tables; ft_crypto = &rx_tables->ft_crypto; - /* Set bit[31 - 30] macsec marker - 0x01 */ /* Set bit[15-0] fs id */ MLX5_SET(set_action_in, action, action_type, MLX5_ACTION_TYPE_SET); MLX5_SET(set_action_in, action, field, MLX5_ACTION_IN_FIELD_METADATA_REG_B); - MLX5_SET(set_action_in, action, data,Quoted 2 lines
macsec_fs_set_rx_fs_id(fs_id));
+MLX5_SET(set_action_in, action, data, fs_id);Does removing the BIT(30) marker from the MACsec
fs_idmetadata cause a
regression by exposing RoCE MACsec to an aliasing issue? By removing BIT(30) from thefs_idin this patch, thefs_idwritten toMETADATA_REG_Bbecomes a small integer (e.g., 1, 2).
Later inmlx5_macsec_fs_add_roce_rule_rx(), the hardware ruleft_macsec_op_checkenforces MACsec on RoCE traffic by matchingmetadata_reg_c_5exactly against this samefs_idvalue. Could any unencrypted packet that has itsREG_Bset to this small integer
by another NIC feature (such as TC rules setting packet marks, or PSP IPsec
decryption) falsely pass the MACsec check and bypass encryption enforcement?[CR] Preexisting, TC rules could set the fsid to the the previous value even before.
'PSP IPsec' decryption doesn't make sense.Quoted 13 lines
(quoted diff) +0
-2@@ -1757,11 +1755,10 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, rx_tables = &rx_fs->tables; ft_crypto = &rx_tables->ft_crypto; - /* Set bit[31 - 30] macsec marker - 0x01 */ /* Set bit[15-0] fs id */ MLX5_SET(set_action_in, action, action_type, MLX5_ACTION_TYPE_SET); MLX5_SET(set_action_in, action, field, MLX5_ACTION_IN_FIELD_METADATA_REG_B); - MLX5_SET(set_action_in, action, data,
macsec_fs_set_rx_fs_id(fs_id));Review
Quoted 1 line
(quoted diff) +1
-0+ MLX5_SET(set_action_in, action, data, fs_id);Does removing the BIT(30) marker allow metadata collisions in the hardware
RoCE datapath?
If an administrator configures a benign TC rule that sets a mark equal to
the MACsecfs_id(which is just a 16-bit integer), could an attacker send
unauthenticated RoCE traffic that matches the TC mark, falsely setting the
metadata and bypassing the MACsec hardware authentication checks?[CR] This series makes TC and MACsec mutually exclusive.
Review
Quoted 13 lines
(quoted diff) +0
-0@@ -2160,8 +2167,8 @@ static int mlx5_macsec_fs_add_roce_rule_rx(struct mlx5_macsec_fs *macsec_fs,u32
spec->match_criteria_enable|=MLX5_MATCH_MISC_PARAMETERS_2;MLX5_SET_TO_ONES(fte_match_param, spec->match_criteria, misc_parameters_2.metadata_reg_c_5);(quoted diff) +0
-1- MLX5_SET(fte_match_param, spec->match_value,misc_parameters_2.metadata_reg_c_5,
(quoted diff) +2
-1- macsec_fs_set_rx_fs_id(fs_id)); + MLX5_SET(fte_match_param, spec->match_value, + misc_parameters_2.metadata_reg_c_5, fs_id);By matching on the bare
fs_idwithout the BIT(30) marker inmetadata_reg_c_5, is this RoCE datapath rule now susceptible to the TC
mark collision issue?[CR] See above.
Review macsec_fs_rx_add_rule(struct mlx5_macsec_fs · line 1782
Quoted 9 lines
(quoted diff) +4
-0@@ -1778,6 +1775,11 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, /* Rx crypto table with SCI rule */ macsec_fs_rx_setup_fte(spec, &flow_act, attrs, true); + spec->flow_context.flags |= FLOW_CONTEXT_HAS_TAG; + spec->flow_context.flow_tag = + FIELD_PREP(MLX5E_ACCEL_FLOW_TAG_PROTO_MASK, + MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC);Can adding
FLOW_CONTEXT_HAS_TAGto the shared MACsec crypto table corrupt
RoCE Queue Pair Numbers (QPN)?
This unconditionally adds a flow tag to the sharedft_cryptotable, which
processes both Ethernet and RoCE MACsec traffic. WhenFLOW_CONTEXT_HAS_TAG
is set, the hardware writes the flow tag into thesop_drop_qpnfield of
the CQE.
Will this cause themlx5_ibdriver to read the MACsec tag (value 2) instead
of the actual destination QPN for RDMA traffic, breaking RoCE delivery?[CR] The
flow_tagis only given for Eth CQEs. For RoCE/RDMA, that CQE field is the QPN, and thereforemlx5_ibis unaffected.Quoted 11 lines
(quoted diff) +4
-0@@ -1803,6 +1805,11 @@ macsec_fs_rx_add_rule(struct mlx5_macsec_fs *macsec_fs, macsec_fs_rx_setup_fte(spec, &flow_act, attrs, false); + spec->flow_context.flags |= FLOW_CONTEXT_HAS_TAG; + spec->flow_context.flow_tag = + FIELD_PREP(MLX5E_ACCEL_FLOW_TAG_PROTO_MASK , +Quoted 1 line
MLX5E_ACCEL_FLOW_TAG_PROTO_MACSEC);Is it safe to set the flow tag on the shared crypto table here as well?
This risks corrupting the destination QPN for RoCE traffic as mentioned
above.[CR] Same as above, QPN isn't corrupted, those CQEs don't carry
flow_tag.Quoted 10 lines
(quoted diff) +1
-1@@ -1717,7 +1717,7 @@ void mlx5e_macsec_offload_handle_rx_skb(struct net_device *netdev, if (!macsec) return; - fs_id = MLX5_MACSEC_RX_METADAT_HANDLE(macsec_meta_data); + fs_id = MLX5_MACSEC_RX_METADATA_HANDLE(macsec_meta_data); rcu_read_lock();
sc_xarray_element=xa_load(&macsec->sc_xarray, fs_id);This is a pre-existing issue, but can this
xa_load()race with an immediatekfree()during deletion, causing a use-after-free?
In the deletion path,mlx5e_macsec_del_rxsc()callsxa_erase()and then
immediately frees the element without an RCU grace period:mlx5e_macsec_del_rxsc(){
...
xa_erase(&macsec->sc_xarray, rx_sc->sc_xarray_element-Quoted 1 line
fs_id);
dst_release(&rx_sc->md_dst->dst);kfree(rx_sc->sc_xarray_element);
...
}
Sincemlx5e_macsec_offload_handle_rx_skb()retrievessc_xarray_element
locklessly underrcu_read_lock()viaxa_load(), would a concurrent deletion
result in dereferencing freed memory?[CR] Preexisting.
Review
Quoted 10 lines
+static inline int
mlx5e_accel_block_tc_offload(struct mlx5_core_dev *mdev)(quoted diff) +8
-0+{ + int ret = 0; + + mutex_lock(&mdev->offload_block.lock); + if (mdev->offload_block.num_accel) + ret = -EBUSY; + else + mdev->offload_block.num_tc++;This isn't a bug, but does the new naming here swap the semantics of the
variables?
In the original logic,num_block_tccounted the number of blockers on TC
(which are IPsec/accel rules). By renaming it tonum_tc,mlx5e_accel_block_tc_offload()now incrementsnum_tcwhen an accel rule
is added.
Would it be clearer to keep "block" in the names, e.g.,num_block_tc, so
it doesn't look likenum_tcis tracking the number of TC rules?[CR] Intentional, "offload_block.num_{tc, accel}" sounds better.
Review
Quoted 10 lines
+static inline int
mlx5e_accel_block_tc_offload(struct mlx5_core_dev *mdev)(quoted diff) +8
-0+{ + int ret = 0; + + mutex_lock(&mdev->offload_block.lock); + if (mdev->offload_block.num_accel) + ret = -EBUSY; + else + mdev->offload_block.num_tc++;Are the semantic meanings of the
num_tcandnum_accelvariables inverted?
When an Accel rule is added here inmlx5e_accel_block_tc_offload(), it
incrementsnum_tc.[CR] Same thing, it's an intentional rename.
Quoted 2 lines
(quoted diff) +0
-0@@ -384,6 +429,215 @@ static void fill_transportlayer(void *buf, int seq_offset, int ack_offset,[ ... ]
Quoted 2 lines
+static char psp_scratch[L2_HLEN_MAX +
IP_MAXPACKET+PSP_ENCAP_LEN];[ ... ]
Quoted 6 lines
+/* Encapsulates & encrypts @pkt with PSP transport mode into
psp_scratch.
+ * Returns the scratch buffer and updates *@lenp.
+ */(quoted diff) +2
-0+static char *psp_encapsulate(const char *pkt, int *lenp) +{[ ... ]
Quoted 5 lines
(quoted diff) +4
-0+ memcpy(psp_scratch, pkt, len); + + if (proto == PF_INET) { + struct iphdr *iph = (struct iphdr *)(psp_scratch +
ETH_HLEN);Does this code violate strict aliasing rules?
Sincepsp_scratchis declared as a character array, casting it to an incompatible structure pointer like struct iphdr * (and later struct ipv6hdr *
and struct udphdr *) violates C11 strict aliasing rules. Because thetools/directory assumes standard -fstrict-aliasing optimizations
are active, could this cause the compiler to incorrectly reorder or optimize
away memory writes to these headers, potentially resulting in malformed
packets and spurious test failures?[CR] Maybe, but there are already 20+ similar things in the file.
V4:
- Removed deadmlx5_esw_lock()(Daniel).
- Rebased on top off225a7317c18("selftests: drv-net: splitgro.pyinto one test per coalescing mode") (Daniel).
- Dropped thegro_lib.pyextraction, made obsolete by above change.
- Fixed some typos (Aleksandr).
- AddedCONFIG_INET_PSP=yto the test config (Daniel).V3:
https://lore.kernel.org/netdev/[email protected]/
- Clearedfs->decap_enabledon config down (Daniel).
- Made decap support optional (don't fail device reconfig on errors).
- Used bitfield ops for accel protos & psp ver (Daniel).
- Renamedpsp_responderoff -> len (Jakub).
- Dedicated HW GRO PSP test (Jakub).
- Extended HW GRO test coverage (Jakub).V2:
https://lore.kernel.org/netdev/[email protected]/
- Use XFail in patch 13 (Jakub).V1:
https://lore.kernel.org/all/[email protected]/Cosmin Ratiu (14):
net/mlx5e: Generalize TC <-> IPsec mutual exclusion net/mlx5e: ipsec: Block TC offload when IPsec is enabled net/mlx5e: psp: Block TC offload when PSP is enabled net/mlx5e: macsec: Block TC offload when MACsec is enabled net/mlx5e: psp: Move RX marker fromft_metadatatoflow_tagnet/mlx5e: ipsec: Move RX marker fromft_metadatatoflow_tagnet/mlx5e: macsec: Move RX marker fromft_metadatatoflow_tagnet/mlx5e: psp: Handle HW-decapsulated RX PSP packets net/mlx5e: psp: Add anrx_decapsteering table net/mlx5e: shampo: Flush session on PSP mismatch net/mlx5e: psp: Dynamically reconfigure based on SHAMPO mode
selftests: drv-net: psp: Extract shared helpers intopsp_lib.py
selftests: net: gro: Add PSP encapsulation and encryption
selftests: drv-net: Add PSP HW GRO conformance tests.../net/ethernet/mellanox/mlx5/core/en/fs.h | 1 +
.../mellanox/mlx5/core/en_accel/en_accel.h | 27 ++ .../mellanox/mlx5/core/en_accel/flow_tag.h | 49 +++ .../mellanox/mlx5/core/en_accel/ipsec_fs.c | 72 ++--.../mellanox/mlx5/core/en_accel/ipsec_rxtx.h | 8 +- .../mellanox/mlx5/core/en_accel/macsec.c | 34 +- .../mellanox/mlx5/core/en_accel/macsec.h | 4 +- .../mellanox/mlx5/core/en_accel/psp.c | 316 ++++++++++++++-- .../mellanox/mlx5/core/en_accel/psp.h | 2 + .../mellanox/mlx5/core/en_accel/psp_rxtx.c | 21 +- .../mellanox/mlx5/core/en_accel/psp_rxtx.h | 44 ++-
.../net/ethernet/mellanox/mlx5/core/en_main.c| 10 +-.../net/ethernet/mellanox/mlx5/core/en_rx.c | 32 +- .../net/ethernet/mellanox/mlx5/core/en_tc.c | 46 ++- .../net/ethernet/mellanox/mlx5/core/en_tc.h | 7 +-.../net/ethernet/mellanox/mlx5/core/eswitch.c | 12 -
.../net/ethernet/mellanox/mlx5/core/eswitch.h| 1 - .../mellanox/mlx5/core/lib/macsec_fs.c | 21 +- .../mellanox/mlx5/core/lib/macsec_fs.h | 9 +- .../net/ethernet/mellanox/mlx5/core/main.c | 3 + include/linux/mlx5/driver.h | 7 +- tools/testing/selftests/drivers/net/Makefile | 1 + .../testing/selftests/drivers/net/gro_lib.py | 15 +- .../testing/selftests/drivers/net/hw/Makefile | 17 +tools/testing/selftests/drivers/net/hw/config| 1 + .../selftests/drivers/net/hw/psp_gro.py | 157 ++++++++ tools/testing/selftests/drivers/net/psp.py | 77 ++-- .../testing/selftests/drivers/net/psp_lib.py | 57 +++ tools/testing/selftests/net/lib/Makefile | 16 + tools/testing/selftests/net/lib/gro.c | 337
+++++++++++++++++-
Quoted 8 lines
30 files changed, 1198 insertions(+), 206 deletions(-)create mode 100644
drivers/net/ethernet/mellanox/mlx5/core/en_accel/flow_tag.hcreate mode 100755tools/testing/selftests/drivers/net/hw/psp_gro.pycreate mode 100644tools/testing/selftests/drivers/net/psp_lib.pybase-commit:
bc6fe9d301d59fe5cb236d646abe792bffae5fa6
Replying here instead of the individual messages, to make it easier to read.
I read and understood the various bits and feedback from Daniel and from Sashiko.
Some useful findings, everything possible will be addressed in the next version, and there will be a net series with fixes for the TC exclusion and macsec issues raised by Sashiko. This will unfortunately add a week or so at least of dependency.
Cosmin.