Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
Merge branch 'net-mana-enforce-tx-sge-limit-and-fix-error-cleanup'
Aditya Garg says:
====================
net: mana: Enforce TX SGE limit and fix error cleanup
Add pre-transmission checks to block SKBs that exceed the hardware's SGE
limit. Force software segmentation for GSO traffic and linearize non-GSO
packets as needed.
Update TX error handling to drop failed SKBs and unmap resources
immediately.
====================
Link: https://patch.msgid.link/1763464269-10431-1-git-send-email-gargaditya@linux.microsoft.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
+53
-12
+1
-5
drivers/net/ethernet/microsoft/mana/gdma_main.c
+1
-5
drivers/net/ethernet/microsoft/mana/gdma_main.c
···
1300
1300
struct gdma_posted_wqe_info *wqe_info)
1301
1301
{
1302
1302
u32 client_oob_size = wqe_req->inline_oob_size;
1303
-
struct gdma_context *gc;
1304
1303
u32 sgl_data_size;
1305
1304
u32 max_wqe_size;
1306
1305
u32 wqe_size;
···
1329
1330
if (wqe_size > max_wqe_size)
1330
1331
return -EINVAL;
1331
1332
1332
-
if (wq->monitor_avl_buf && wqe_size > mana_gd_wq_avail_space(wq)) {
1333
-
gc = wq->gdma_dev->gdma_context;
1334
-
dev_err(gc->dev, "unsuccessful flow control!\n");
1333
+
if (wq->monitor_avl_buf && wqe_size > mana_gd_wq_avail_space(wq))
1335
1334
return -ENOSPC;
1336
-
}
1337
1335
1338
1336
if (wqe_info)
1339
1337
wqe_info->wqe_size_in_bu = wqe_size / GDMA_WQE_BU_SIZE;
+41
-6
drivers/net/ethernet/microsoft/mana/mana_en.c
+41
-6
drivers/net/ethernet/microsoft/mana/mana_en.c
···
11
11
#include <linux/mm.h>
12
12
#include <linux/pci.h>
13
13
#include <linux/export.h>
14
+
#include <linux/skbuff.h>
14
15
15
16
#include <net/checksum.h>
16
17
#include <net/ip6_checksum.h>
···
330
329
cq = &apc->tx_qp[txq_idx].tx_cq;
331
330
tx_stats = &txq->stats;
332
331
332
+
BUILD_BUG_ON(MAX_TX_WQE_SGL_ENTRIES != MANA_MAX_TX_WQE_SGL_ENTRIES);
333
+
if (MAX_SKB_FRAGS + 2 > MAX_TX_WQE_SGL_ENTRIES &&
334
+
skb_shinfo(skb)->nr_frags + 2 > MAX_TX_WQE_SGL_ENTRIES) {
335
+
/* GSO skb with Hardware SGE limit exceeded is not expected here
336
+
* as they are handled in mana_features_check() callback
337
+
*/
338
+
if (skb_linearize(skb)) {
339
+
netdev_warn_once(ndev, "Failed to linearize skb with nr_frags=%d and is_gso=%d\n",
340
+
skb_shinfo(skb)->nr_frags,
341
+
skb_is_gso(skb));
342
+
goto tx_drop_count;
343
+
}
344
+
apc->eth_stats.tx_linear_pkt_cnt++;
345
+
}
346
+
333
347
pkg.tx_oob.s_oob.vcq_num = cq->gdma_id;
334
348
pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame;
335
349
···
458
442
}
459
443
}
460
444
461
-
WARN_ON_ONCE(pkg.wqe_req.num_sge > MAX_TX_WQE_SGL_ENTRIES);
462
-
463
445
if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) {
464
446
pkg.wqe_req.sgl = pkg.sgl_array;
465
447
} else {
···
492
478
493
479
if (err) {
494
480
(void)skb_dequeue_tail(&txq->pending_skbs);
481
+
mana_unmap_skb(skb, apc);
495
482
netdev_warn(ndev, "Failed to post TX OOB: %d\n", err);
496
-
err = NETDEV_TX_BUSY;
497
-
goto tx_busy;
483
+
goto free_sgl_ptr;
498
484
}
499
485
500
486
err = NETDEV_TX_OK;
···
514
500
tx_stats->bytes += len + ((num_gso_seg - 1) * gso_hs);
515
501
u64_stats_update_end(&tx_stats->syncp);
516
502
517
-
tx_busy:
518
503
if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) {
519
504
netif_tx_wake_queue(net_txq);
520
505
apc->eth_stats.wake_queue++;
···
530
517
dev_kfree_skb_any(skb);
531
518
return NETDEV_TX_OK;
532
519
}
520
+
521
+
#if (MAX_SKB_FRAGS + 2 > MANA_MAX_TX_WQE_SGL_ENTRIES)
522
+
static netdev_features_t mana_features_check(struct sk_buff *skb,
523
+
struct net_device *ndev,
524
+
netdev_features_t features)
525
+
{
526
+
if (skb_shinfo(skb)->nr_frags + 2 > MAX_TX_WQE_SGL_ENTRIES) {
527
+
/* Exceeds HW SGE limit.
528
+
* GSO case:
529
+
* Disable GSO so the stack will software-segment the skb
530
+
* into smaller skbs that fit the SGE budget.
531
+
* Non-GSO case:
532
+
* The xmit path will attempt skb_linearize() as a fallback.
533
+
*/
534
+
features &= ~NETIF_F_GSO_MASK;
535
+
}
536
+
return features;
537
+
}
538
+
#endif
533
539
534
540
static void mana_get_stats64(struct net_device *ndev,
535
541
struct rtnl_link_stats64 *st)
···
915
883
.ndo_open = mana_open,
916
884
.ndo_stop = mana_close,
917
885
.ndo_select_queue = mana_select_queue,
886
+
#if (MAX_SKB_FRAGS + 2 > MANA_MAX_TX_WQE_SGL_ENTRIES)
887
+
.ndo_features_check = mana_features_check,
888
+
#endif
918
889
.ndo_start_xmit = mana_start_xmit,
919
890
.ndo_validate_addr = eth_validate_addr,
920
891
.ndo_get_stats64 = mana_get_stats64,
···
1686
1651
return 0;
1687
1652
}
1688
1653
1689
-
static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1654
+
void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc)
1690
1655
{
1691
1656
struct mana_skb_head *ash = (struct mana_skb_head *)skb->head;
1692
1657
struct gdma_context *gc = apc->ac->gdma_dev->gdma_context;
+2
drivers/net/ethernet/microsoft/mana/mana_ethtool.c
+2
drivers/net/ethernet/microsoft/mana/mana_ethtool.c
···
18
18
{"tx_cq_err", offsetof(struct mana_ethtool_stats, tx_cqe_err)},
19
19
{"tx_cqe_unknown_type", offsetof(struct mana_ethtool_stats,
20
20
tx_cqe_unknown_type)},
21
+
{"tx_linear_pkt_cnt", offsetof(struct mana_ethtool_stats,
22
+
tx_linear_pkt_cnt)},
21
23
{"rx_coalesced_err", offsetof(struct mana_ethtool_stats,
22
24
rx_coalesced_err)},
23
25
{"rx_cqe_unknown_type", offsetof(struct mana_ethtool_stats,
+7
-1
include/net/mana/gdma.h
+7
-1
include/net/mana/gdma.h
···
486
486
#define INLINE_OOB_SMALL_SIZE 8
487
487
#define INLINE_OOB_LARGE_SIZE 24
488
488
489
+
#define MANA_MAX_TX_WQE_SGL_ENTRIES 30
490
+
489
491
#define MAX_TX_WQE_SIZE 512
490
492
#define MAX_RX_WQE_SIZE 256
491
493
···
594
592
#define GDMA_DRV_CAP_FLAG_1_HANDLE_RECONFIG_EQE BIT(17)
595
593
#define GDMA_DRV_CAP_FLAG_1_HW_VPORT_LINK_AWARE BIT(6)
596
594
595
+
/* Driver supports linearizing the skb when num_sge exceeds hardware limit */
596
+
#define GDMA_DRV_CAP_FLAG_1_SKB_LINEARIZE BIT(20)
597
+
597
598
/* Driver can send HWC periodically to query stats */
598
599
#define GDMA_DRV_CAP_FLAG_1_PERIODIC_STATS_QUERY BIT(21)
599
600
···
610
605
GDMA_DRV_CAP_FLAG_1_SELF_RESET_ON_EQE | \
611
606
GDMA_DRV_CAP_FLAG_1_HANDLE_RECONFIG_EQE | \
612
607
GDMA_DRV_CAP_FLAG_1_HW_VPORT_LINK_AWARE | \
613
-
GDMA_DRV_CAP_FLAG_1_PERIODIC_STATS_QUERY)
608
+
GDMA_DRV_CAP_FLAG_1_PERIODIC_STATS_QUERY | \
609
+
GDMA_DRV_CAP_FLAG_1_SKB_LINEARIZE)
614
610
615
611
#define GDMA_DRV_CAP_FLAGS2 0
616
612
+2
include/net/mana/mana.h
+2
include/net/mana/mana.h
···
377
377
u64 wake_queue;
378
378
u64 tx_cqe_err;
379
379
u64 tx_cqe_unknown_type;
380
+
u64 tx_linear_pkt_cnt;
380
381
u64 rx_coalesced_err;
381
382
u64 rx_cqe_unknown_type;
382
383
};
···
593
592
void mana_query_phy_stats(struct mana_port_context *apc);
594
593
int mana_pre_alloc_rxbufs(struct mana_port_context *apc, int mtu, int num_queues);
595
594
void mana_pre_dealloc_rxbufs(struct mana_port_context *apc);
595
+
void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc);
596
596
597
597
extern const struct ethtool_ops mana_ethtool_ops;
598
598
extern struct dentry *mana_debugfs_root;