1 // SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0-or-later
3 * Copyright 2008 - 2016 Freescale Semiconductor Inc.
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/init.h>
10 #include <linux/module.h>
11 #include <linux/of_platform.h>
12 #include <linux/of_mdio.h>
13 #include <linux/of_net.h>
15 #include <linux/if_arp.h>
16 #include <linux/if_vlan.h>
17 #include <linux/icmp.h>
19 #include <linux/ipv6.h>
20 #include <linux/udp.h>
21 #include <linux/tcp.h>
22 #include <linux/net.h>
23 #include <linux/skbuff.h>
24 #include <linux/etherdevice.h>
25 #include <linux/if_ether.h>
26 #include <linux/highmem.h>
27 #include <linux/percpu.h>
28 #include <linux/dma-mapping.h>
29 #include <linux/sort.h>
30 #include <linux/phy_fixed.h>
31 #include <linux/bpf.h>
32 #include <linux/bpf_trace.h>
33 #include <soc/fsl/bman.h>
34 #include <soc/fsl/qman.h>
36 #include "fman_port.h"
40 /* CREATE_TRACE_POINTS only needs to be defined once. Other dpaa files
41 * using trace events only need to #include <trace/events/sched.h>
43 #define CREATE_TRACE_POINTS
44 #include "dpaa_eth_trace.h"
46 static int debug = -1;
47 module_param(debug, int, 0444);
48 MODULE_PARM_DESC(debug, "Module/Driver verbosity level (0=none,...,16=all)");
50 static u16 tx_timeout = 1000;
51 module_param(tx_timeout, ushort, 0444);
52 MODULE_PARM_DESC(tx_timeout, "The Tx timeout in ms");
54 #define FM_FD_STAT_RX_ERRORS \
55 (FM_FD_ERR_DMA | FM_FD_ERR_PHYSICAL | \
56 FM_FD_ERR_SIZE | FM_FD_ERR_CLS_DISCARD | \
57 FM_FD_ERR_EXTRACTION | FM_FD_ERR_NO_SCHEME | \
58 FM_FD_ERR_PRS_TIMEOUT | FM_FD_ERR_PRS_ILL_INSTRUCT | \
59 FM_FD_ERR_PRS_HDR_ERR)
61 #define FM_FD_STAT_TX_ERRORS \
62 (FM_FD_ERR_UNSUPPORTED_FORMAT | \
63 FM_FD_ERR_LENGTH | FM_FD_ERR_DMA)
65 #define DPAA_MSG_DEFAULT (NETIF_MSG_DRV | NETIF_MSG_PROBE | \
66 NETIF_MSG_LINK | NETIF_MSG_IFUP | \
67 NETIF_MSG_IFDOWN | NETIF_MSG_HW)
69 #define DPAA_INGRESS_CS_THRESHOLD 0x10000000
70 /* Ingress congestion threshold on FMan ports
71 * The size in bytes of the ingress tail-drop threshold on FMan ports.
72 * Traffic piling up above this value will be rejected by QMan and discarded
76 /* Size in bytes of the FQ taildrop threshold */
77 #define DPAA_FQ_TD 0x200000
79 #define DPAA_CS_THRESHOLD_1G 0x06000000
80 /* Egress congestion threshold on 1G ports, range 0x1000 .. 0x10000000
81 * The size in bytes of the egress Congestion State notification threshold on
82 * 1G ports. The 1G dTSECs can quite easily be flooded by cores doing Tx in a
83 * tight loop (e.g. by sending UDP datagrams at "while(1) speed"),
84 * and the larger the frame size, the more acute the problem.
85 * So we have to find a balance between these factors:
86 * - avoiding the device staying congested for a prolonged time (risking
87 * the netdev watchdog to fire - see also the tx_timeout module param);
88 * - affecting performance of protocols such as TCP, which otherwise
89 * behave well under the congestion notification mechanism;
90 * - preventing the Tx cores from tightly-looping (as if the congestion
91 * threshold was too low to be effective);
92 * - running out of memory if the CS threshold is set too high.
95 #define DPAA_CS_THRESHOLD_10G 0x10000000
96 /* The size in bytes of the egress Congestion State notification threshold on
97 * 10G ports, range 0x1000 .. 0x10000000
100 /* Largest value that the FQD's OAL field can hold */
101 #define FSL_QMAN_MAX_OAL 127
103 /* Default alignment for start of data in an Rx FD */
104 #ifdef CONFIG_DPAA_ERRATUM_A050385
105 /* aligning data start to 64 avoids DMA transaction splits, unless the buffer
106 * is crossing a 4k page boundary
108 #define DPAA_FD_DATA_ALIGNMENT (fman_has_errata_a050385() ? 64 : 16)
109 /* aligning to 256 avoids DMA transaction splits caused by 4k page boundary
110 * crossings; also, all SG fragments except the last must have a size multiple
111 * of 256 to avoid DMA transaction splits
113 #define DPAA_A050385_ALIGN 256
114 #define DPAA_FD_RX_DATA_ALIGNMENT (fman_has_errata_a050385() ? \
115 DPAA_A050385_ALIGN : 16)
117 #define DPAA_FD_DATA_ALIGNMENT 16
118 #define DPAA_FD_RX_DATA_ALIGNMENT DPAA_FD_DATA_ALIGNMENT
121 /* The DPAA requires 256 bytes reserved and mapped for the SGT */
122 #define DPAA_SGT_SIZE 256
124 /* Values for the L3R field of the FM Parse Results
126 /* L3 Type field: First IP Present IPv4 */
127 #define FM_L3_PARSE_RESULT_IPV4 0x8000
128 /* L3 Type field: First IP Present IPv6 */
129 #define FM_L3_PARSE_RESULT_IPV6 0x4000
130 /* Values for the L4R field of the FM Parse Results */
131 /* L4 Type field: UDP */
132 #define FM_L4_PARSE_RESULT_UDP 0x40
133 /* L4 Type field: TCP */
134 #define FM_L4_PARSE_RESULT_TCP 0x20
136 /* FD status field indicating whether the FM Parser has attempted to validate
137 * the L4 csum of the frame.
138 * Note that having this bit set doesn't necessarily imply that the checksum
139 * is valid. One would have to check the parse results to find that out.
141 #define FM_FD_STAT_L4CV 0x00000004
143 #define DPAA_SGT_MAX_ENTRIES 16 /* maximum number of entries in SG Table */
144 #define DPAA_BUFF_RELEASE_MAX 8 /* maximum number of buffers released at once */
146 #define FSL_DPAA_BPID_INV 0xff
147 #define FSL_DPAA_ETH_MAX_BUF_COUNT 128
148 #define FSL_DPAA_ETH_REFILL_THRESHOLD 80
150 #define DPAA_TX_PRIV_DATA_SIZE 16
151 #define DPAA_PARSE_RESULTS_SIZE sizeof(struct fman_prs_result)
152 #define DPAA_TIME_STAMP_SIZE 8
153 #define DPAA_HASH_RESULTS_SIZE 8
154 #define DPAA_HWA_SIZE (DPAA_PARSE_RESULTS_SIZE + DPAA_TIME_STAMP_SIZE \
155 + DPAA_HASH_RESULTS_SIZE)
156 #define DPAA_RX_PRIV_DATA_DEFAULT_SIZE (DPAA_TX_PRIV_DATA_SIZE + \
157 XDP_PACKET_HEADROOM - DPAA_HWA_SIZE)
158 #ifdef CONFIG_DPAA_ERRATUM_A050385
159 #define DPAA_RX_PRIV_DATA_A050385_SIZE (DPAA_A050385_ALIGN - DPAA_HWA_SIZE)
160 #define DPAA_RX_PRIV_DATA_SIZE (fman_has_errata_a050385() ? \
161 DPAA_RX_PRIV_DATA_A050385_SIZE : \
162 DPAA_RX_PRIV_DATA_DEFAULT_SIZE)
164 #define DPAA_RX_PRIV_DATA_SIZE DPAA_RX_PRIV_DATA_DEFAULT_SIZE
167 #define DPAA_ETH_PCD_RXQ_NUM 128
169 #define DPAA_ENQUEUE_RETRIES 100000
171 enum port_type {RX, TX};
174 struct dpaa_fq *tx_defq;
175 struct dpaa_fq *tx_errq;
176 struct dpaa_fq *rx_defq;
177 struct dpaa_fq *rx_errq;
178 struct dpaa_fq *rx_pcdq;
181 /* All the dpa bps in use at any moment */
182 static struct dpaa_bp *dpaa_bp_array[BM_MAX_NUM_OF_POOLS];
184 #define DPAA_BP_RAW_SIZE 4096
186 #ifdef CONFIG_DPAA_ERRATUM_A050385
187 #define dpaa_bp_size(raw_size) (SKB_WITH_OVERHEAD(raw_size) & \
188 ~(DPAA_A050385_ALIGN - 1))
190 #define dpaa_bp_size(raw_size) SKB_WITH_OVERHEAD(raw_size)
193 static int dpaa_max_frm;
195 static int dpaa_rx_extra_headroom;
197 #define dpaa_get_max_mtu() \
198 (dpaa_max_frm - (VLAN_ETH_HLEN + ETH_FCS_LEN))
200 static void dpaa_eth_cgr_set_speed(struct mac_device *mac_dev, int speed);
202 static int dpaa_netdev_init(struct net_device *net_dev,
203 const struct net_device_ops *dpaa_ops,
206 struct dpaa_priv *priv = netdev_priv(net_dev);
207 struct device *dev = net_dev->dev.parent;
208 struct mac_device *mac_dev = priv->mac_dev;
209 struct dpaa_percpu_priv *percpu_priv;
213 /* Although we access another CPU's private data here
214 * we do it at initialization so it is safe
216 for_each_possible_cpu(i) {
217 percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
218 percpu_priv->net_dev = net_dev;
221 net_dev->netdev_ops = dpaa_ops;
222 mac_addr = mac_dev->addr;
224 net_dev->mem_start = (unsigned long)priv->mac_dev->res->start;
225 net_dev->mem_end = (unsigned long)priv->mac_dev->res->end;
227 net_dev->min_mtu = ETH_MIN_MTU;
228 net_dev->max_mtu = dpaa_get_max_mtu();
230 net_dev->hw_features |= (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
231 NETIF_F_LLTX | NETIF_F_RXHASH);
233 net_dev->hw_features |= NETIF_F_SG | NETIF_F_HIGHDMA;
234 /* The kernels enables GSO automatically, if we declare NETIF_F_SG.
235 * For conformity, we'll still declare GSO explicitly.
237 net_dev->features |= NETIF_F_GSO;
238 net_dev->features |= NETIF_F_RXCSUM;
240 net_dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
241 /* we do not want shared skbs on TX */
242 net_dev->priv_flags &= ~IFF_TX_SKB_SHARING;
244 net_dev->features |= net_dev->hw_features;
245 net_dev->vlan_features = net_dev->features;
247 if (is_valid_ether_addr(mac_addr)) {
248 memcpy(net_dev->perm_addr, mac_addr, net_dev->addr_len);
249 eth_hw_addr_set(net_dev, mac_addr);
251 eth_hw_addr_random(net_dev);
252 err = mac_dev->change_addr(mac_dev->fman_mac,
253 (const enet_addr_t *)net_dev->dev_addr);
255 dev_err(dev, "Failed to set random MAC address\n");
258 dev_info(dev, "Using random MAC address: %pM\n",
262 net_dev->ethtool_ops = &dpaa_ethtool_ops;
264 net_dev->needed_headroom = priv->tx_headroom;
265 net_dev->watchdog_timeo = msecs_to_jiffies(tx_timeout);
267 /* The rest of the config is filled in by the mac device already */
268 mac_dev->phylink_config.dev = &net_dev->dev;
269 mac_dev->phylink_config.type = PHYLINK_NETDEV;
270 mac_dev->update_speed = dpaa_eth_cgr_set_speed;
271 mac_dev->phylink = phylink_create(&mac_dev->phylink_config,
272 dev_fwnode(mac_dev->dev),
274 mac_dev->phylink_ops);
275 if (IS_ERR(mac_dev->phylink)) {
276 err = PTR_ERR(mac_dev->phylink);
277 dev_err_probe(dev, err, "Could not create phylink\n");
281 /* start without the RUNNING flag, phylib controls it later */
282 netif_carrier_off(net_dev);
284 err = register_netdev(net_dev);
286 dev_err(dev, "register_netdev() = %d\n", err);
287 phylink_destroy(mac_dev->phylink);
294 static int dpaa_stop(struct net_device *net_dev)
296 struct mac_device *mac_dev;
297 struct dpaa_priv *priv;
300 priv = netdev_priv(net_dev);
301 mac_dev = priv->mac_dev;
303 netif_tx_stop_all_queues(net_dev);
304 /* Allow the Fman (Tx) port to process in-flight frames before we
305 * try switching it off.
309 phylink_stop(mac_dev->phylink);
310 mac_dev->disable(mac_dev->fman_mac);
312 for (i = 0; i < ARRAY_SIZE(mac_dev->port); i++) {
313 error = fman_port_disable(mac_dev->port[i]);
318 phylink_disconnect_phy(mac_dev->phylink);
319 net_dev->phydev = NULL;
326 static void dpaa_tx_timeout(struct net_device *net_dev, unsigned int txqueue)
328 struct dpaa_percpu_priv *percpu_priv;
329 const struct dpaa_priv *priv;
331 priv = netdev_priv(net_dev);
332 percpu_priv = this_cpu_ptr(priv->percpu_priv);
334 netif_crit(priv, timer, net_dev, "Transmit timeout latency: %u ms\n",
335 jiffies_to_msecs(jiffies - dev_trans_start(net_dev)));
337 percpu_priv->stats.tx_errors++;
340 /* Calculates the statistics for the given device by adding the statistics
341 * collected by each CPU.
343 static void dpaa_get_stats64(struct net_device *net_dev,
344 struct rtnl_link_stats64 *s)
346 int numstats = sizeof(struct rtnl_link_stats64) / sizeof(u64);
347 struct dpaa_priv *priv = netdev_priv(net_dev);
348 struct dpaa_percpu_priv *percpu_priv;
349 u64 *netstats = (u64 *)s;
353 for_each_possible_cpu(i) {
354 percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
356 cpustats = (u64 *)&percpu_priv->stats;
358 /* add stats from all CPUs */
359 for (j = 0; j < numstats; j++)
360 netstats[j] += cpustats[j];
364 static int dpaa_setup_tc(struct net_device *net_dev, enum tc_setup_type type,
367 struct dpaa_priv *priv = netdev_priv(net_dev);
368 struct tc_mqprio_qopt *mqprio = type_data;
372 if (type != TC_SETUP_QDISC_MQPRIO)
375 mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
376 num_tc = mqprio->num_tc;
378 if (num_tc == priv->num_tc)
382 netdev_reset_tc(net_dev);
386 if (num_tc > DPAA_TC_NUM) {
387 netdev_err(net_dev, "Too many traffic classes: max %d supported.\n",
392 netdev_set_num_tc(net_dev, num_tc);
394 for (i = 0; i < num_tc; i++)
395 netdev_set_tc_queue(net_dev, i, DPAA_TC_TXQ_NUM,
396 i * DPAA_TC_TXQ_NUM);
399 priv->num_tc = num_tc ? : 1;
400 netif_set_real_num_tx_queues(net_dev, priv->num_tc * DPAA_TC_TXQ_NUM);
404 static struct mac_device *dpaa_mac_dev_get(struct platform_device *pdev)
406 struct dpaa_eth_data *eth_data;
407 struct device *dpaa_dev;
408 struct mac_device *mac_dev;
410 dpaa_dev = &pdev->dev;
411 eth_data = dpaa_dev->platform_data;
413 dev_err(dpaa_dev, "eth_data missing\n");
414 return ERR_PTR(-ENODEV);
416 mac_dev = eth_data->mac_dev;
418 dev_err(dpaa_dev, "mac_dev missing\n");
419 return ERR_PTR(-EINVAL);
425 static int dpaa_set_mac_address(struct net_device *net_dev, void *addr)
427 const struct dpaa_priv *priv;
428 struct mac_device *mac_dev;
429 struct sockaddr old_addr;
432 priv = netdev_priv(net_dev);
434 memcpy(old_addr.sa_data, net_dev->dev_addr, ETH_ALEN);
436 err = eth_mac_addr(net_dev, addr);
438 netif_err(priv, drv, net_dev, "eth_mac_addr() = %d\n", err);
442 mac_dev = priv->mac_dev;
444 err = mac_dev->change_addr(mac_dev->fman_mac,
445 (const enet_addr_t *)net_dev->dev_addr);
447 netif_err(priv, drv, net_dev, "mac_dev->change_addr() = %d\n",
449 /* reverting to previous address */
450 eth_mac_addr(net_dev, &old_addr);
458 static void dpaa_set_rx_mode(struct net_device *net_dev)
460 const struct dpaa_priv *priv;
463 priv = netdev_priv(net_dev);
465 if (!!(net_dev->flags & IFF_PROMISC) != priv->mac_dev->promisc) {
466 priv->mac_dev->promisc = !priv->mac_dev->promisc;
467 err = priv->mac_dev->set_promisc(priv->mac_dev->fman_mac,
468 priv->mac_dev->promisc);
470 netif_err(priv, drv, net_dev,
471 "mac_dev->set_promisc() = %d\n",
475 if (!!(net_dev->flags & IFF_ALLMULTI) != priv->mac_dev->allmulti) {
476 priv->mac_dev->allmulti = !priv->mac_dev->allmulti;
477 err = priv->mac_dev->set_allmulti(priv->mac_dev->fman_mac,
478 priv->mac_dev->allmulti);
480 netif_err(priv, drv, net_dev,
481 "mac_dev->set_allmulti() = %d\n",
485 err = priv->mac_dev->set_multi(net_dev, priv->mac_dev);
487 netif_err(priv, drv, net_dev, "mac_dev->set_multi() = %d\n",
491 static struct dpaa_bp *dpaa_bpid2pool(int bpid)
493 if (WARN_ON(bpid < 0 || bpid >= BM_MAX_NUM_OF_POOLS))
496 return dpaa_bp_array[bpid];
499 /* checks if this bpool is already allocated */
500 static bool dpaa_bpid2pool_use(int bpid)
502 if (dpaa_bpid2pool(bpid)) {
503 refcount_inc(&dpaa_bp_array[bpid]->refs);
510 /* called only once per bpid by dpaa_bp_alloc_pool() */
511 static void dpaa_bpid2pool_map(int bpid, struct dpaa_bp *dpaa_bp)
513 dpaa_bp_array[bpid] = dpaa_bp;
514 refcount_set(&dpaa_bp->refs, 1);
517 static int dpaa_bp_alloc_pool(struct dpaa_bp *dpaa_bp)
521 if (dpaa_bp->size == 0 || dpaa_bp->config_count == 0) {
522 pr_err("%s: Buffer pool is not properly initialized! Missing size or initial number of buffers\n",
527 /* If the pool is already specified, we only create one per bpid */
528 if (dpaa_bp->bpid != FSL_DPAA_BPID_INV &&
529 dpaa_bpid2pool_use(dpaa_bp->bpid))
532 if (dpaa_bp->bpid == FSL_DPAA_BPID_INV) {
533 dpaa_bp->pool = bman_new_pool();
534 if (!dpaa_bp->pool) {
535 pr_err("%s: bman_new_pool() failed\n",
540 dpaa_bp->bpid = (u8)bman_get_bpid(dpaa_bp->pool);
543 if (dpaa_bp->seed_cb) {
544 err = dpaa_bp->seed_cb(dpaa_bp);
546 goto pool_seed_failed;
549 dpaa_bpid2pool_map(dpaa_bp->bpid, dpaa_bp);
554 pr_err("%s: pool seeding failed\n", __func__);
555 bman_free_pool(dpaa_bp->pool);
560 /* remove and free all the buffers from the given buffer pool */
561 static void dpaa_bp_drain(struct dpaa_bp *bp)
567 struct bm_buffer bmb[8];
570 ret = bman_acquire(bp->pool, bmb, num);
573 /* we have less than 8 buffers left;
574 * drain them one by one
580 /* Pool is fully drained */
586 for (i = 0; i < num; i++)
587 bp->free_buf_cb(bp, &bmb[i]);
591 static void dpaa_bp_free(struct dpaa_bp *dpaa_bp)
593 struct dpaa_bp *bp = dpaa_bpid2pool(dpaa_bp->bpid);
595 /* the mapping between bpid and dpaa_bp is done very late in the
596 * allocation procedure; if something failed before the mapping, the bp
597 * was not configured, therefore we don't need the below instructions
602 if (!refcount_dec_and_test(&bp->refs))
608 dpaa_bp_array[bp->bpid] = NULL;
609 bman_free_pool(bp->pool);
612 static void dpaa_bps_free(struct dpaa_priv *priv)
614 dpaa_bp_free(priv->dpaa_bp);
617 /* Use multiple WQs for FQ assignment:
618 * - Tx Confirmation queues go to WQ1.
619 * - Rx Error and Tx Error queues go to WQ5 (giving them a better chance
620 * to be scheduled, in case there are many more FQs in WQ6).
621 * - Rx Default goes to WQ6.
622 * - Tx queues go to different WQs depending on their priority. Equal
623 * chunks of NR_CPUS queues go to WQ6 (lowest priority), WQ2, WQ1 and
624 * WQ0 (highest priority).
625 * This ensures that Tx-confirmed buffers are timely released. In particular,
626 * it avoids congestion on the Tx Confirm FQs, which can pile up PFDRs if they
627 * are greatly outnumbered by other FQs in the system, while
628 * dequeue scheduling is round-robin.
630 static inline void dpaa_assign_wq(struct dpaa_fq *fq, int idx)
632 switch (fq->fq_type) {
633 case FQ_TYPE_TX_CONFIRM:
634 case FQ_TYPE_TX_CONF_MQ:
637 case FQ_TYPE_RX_ERROR:
638 case FQ_TYPE_TX_ERROR:
641 case FQ_TYPE_RX_DEFAULT:
646 switch (idx / DPAA_TC_TXQ_NUM) {
648 /* Low priority (best effort) */
652 /* Medium priority */
660 /* Very high priority */
664 WARN(1, "Too many TX FQs: more than %d!\n",
669 WARN(1, "Invalid FQ type %d for FQID %d!\n",
670 fq->fq_type, fq->fqid);
674 static struct dpaa_fq *dpaa_fq_alloc(struct device *dev,
675 u32 start, u32 count,
676 struct list_head *list,
677 enum dpaa_fq_type fq_type)
679 struct dpaa_fq *dpaa_fq;
682 dpaa_fq = devm_kcalloc(dev, count, sizeof(*dpaa_fq),
687 for (i = 0; i < count; i++) {
688 dpaa_fq[i].fq_type = fq_type;
689 dpaa_fq[i].fqid = start ? start + i : 0;
690 list_add_tail(&dpaa_fq[i].list, list);
693 for (i = 0; i < count; i++)
694 dpaa_assign_wq(dpaa_fq + i, i);
699 static int dpaa_alloc_all_fqs(struct device *dev, struct list_head *list,
700 struct fm_port_fqs *port_fqs)
702 struct dpaa_fq *dpaa_fq;
703 u32 fq_base, fq_base_aligned, i;
705 dpaa_fq = dpaa_fq_alloc(dev, 0, 1, list, FQ_TYPE_RX_ERROR);
707 goto fq_alloc_failed;
709 port_fqs->rx_errq = &dpaa_fq[0];
711 dpaa_fq = dpaa_fq_alloc(dev, 0, 1, list, FQ_TYPE_RX_DEFAULT);
713 goto fq_alloc_failed;
715 port_fqs->rx_defq = &dpaa_fq[0];
717 /* the PCD FQIDs range needs to be aligned for correct operation */
718 if (qman_alloc_fqid_range(&fq_base, 2 * DPAA_ETH_PCD_RXQ_NUM))
719 goto fq_alloc_failed;
721 fq_base_aligned = ALIGN(fq_base, DPAA_ETH_PCD_RXQ_NUM);
723 for (i = fq_base; i < fq_base_aligned; i++)
724 qman_release_fqid(i);
726 for (i = fq_base_aligned + DPAA_ETH_PCD_RXQ_NUM;
727 i < (fq_base + 2 * DPAA_ETH_PCD_RXQ_NUM); i++)
728 qman_release_fqid(i);
730 dpaa_fq = dpaa_fq_alloc(dev, fq_base_aligned, DPAA_ETH_PCD_RXQ_NUM,
731 list, FQ_TYPE_RX_PCD);
733 goto fq_alloc_failed;
735 port_fqs->rx_pcdq = &dpaa_fq[0];
737 if (!dpaa_fq_alloc(dev, 0, DPAA_ETH_TXQ_NUM, list, FQ_TYPE_TX_CONF_MQ))
738 goto fq_alloc_failed;
740 dpaa_fq = dpaa_fq_alloc(dev, 0, 1, list, FQ_TYPE_TX_ERROR);
742 goto fq_alloc_failed;
744 port_fqs->tx_errq = &dpaa_fq[0];
746 dpaa_fq = dpaa_fq_alloc(dev, 0, 1, list, FQ_TYPE_TX_CONFIRM);
748 goto fq_alloc_failed;
750 port_fqs->tx_defq = &dpaa_fq[0];
752 if (!dpaa_fq_alloc(dev, 0, DPAA_ETH_TXQ_NUM, list, FQ_TYPE_TX))
753 goto fq_alloc_failed;
758 dev_err(dev, "dpaa_fq_alloc() failed\n");
762 static u32 rx_pool_channel;
763 static DEFINE_SPINLOCK(rx_pool_channel_init);
765 static int dpaa_get_channel(void)
767 spin_lock(&rx_pool_channel_init);
768 if (!rx_pool_channel) {
772 ret = qman_alloc_pool(&pool);
775 rx_pool_channel = pool;
777 spin_unlock(&rx_pool_channel_init);
778 if (!rx_pool_channel)
780 return rx_pool_channel;
783 static void dpaa_release_channel(void)
785 qman_release_pool(rx_pool_channel);
788 static void dpaa_eth_add_channel(u16 channel, struct device *dev)
790 u32 pool = QM_SDQCR_CHANNELS_POOL_CONV(channel);
791 const cpumask_t *cpus = qman_affine_cpus();
792 struct qman_portal *portal;
795 for_each_cpu_and(cpu, cpus, cpu_online_mask) {
796 portal = qman_get_affine_portal(cpu);
797 qman_p_static_dequeue_add(portal, pool);
798 qman_start_using_portal(portal, dev);
802 /* Congestion group state change notification callback.
803 * Stops the device's egress queues while they are congested and
804 * wakes them upon exiting congested state.
805 * Also updates some CGR-related stats.
807 static void dpaa_eth_cgscn(struct qman_portal *qm, struct qman_cgr *cgr,
810 struct dpaa_priv *priv = (struct dpaa_priv *)container_of(cgr,
811 struct dpaa_priv, cgr_data.cgr);
814 priv->cgr_data.congestion_start_jiffies = jiffies;
815 netif_tx_stop_all_queues(priv->net_dev);
816 priv->cgr_data.cgr_congested_count++;
818 priv->cgr_data.congested_jiffies +=
819 (jiffies - priv->cgr_data.congestion_start_jiffies);
820 netif_tx_wake_all_queues(priv->net_dev);
824 static int dpaa_eth_cgr_init(struct dpaa_priv *priv)
826 struct qm_mcc_initcgr initcgr;
830 err = qman_alloc_cgrid(&priv->cgr_data.cgr.cgrid);
832 if (netif_msg_drv(priv))
833 pr_err("%s: Error %d allocating CGR ID\n",
837 priv->cgr_data.cgr.cb = dpaa_eth_cgscn;
839 /* Enable Congestion State Change Notifications and CS taildrop */
840 memset(&initcgr, 0, sizeof(initcgr));
841 initcgr.we_mask = cpu_to_be16(QM_CGR_WE_CSCN_EN | QM_CGR_WE_CS_THRES);
842 initcgr.cgr.cscn_en = QM_CGR_EN;
844 /* Set different thresholds based on the configured MAC speed.
845 * This may turn suboptimal if the MAC is reconfigured at another
846 * speed, so MACs must call dpaa_eth_cgr_set_speed in their link_up
849 if (priv->mac_dev->phylink_config.mac_capabilities & MAC_10000FD)
850 cs_th = DPAA_CS_THRESHOLD_10G;
852 cs_th = DPAA_CS_THRESHOLD_1G;
853 qm_cgr_cs_thres_set64(&initcgr.cgr.cs_thres, cs_th, 1);
855 initcgr.we_mask |= cpu_to_be16(QM_CGR_WE_CSTD_EN);
856 initcgr.cgr.cstd_en = QM_CGR_EN;
858 err = qman_create_cgr(&priv->cgr_data.cgr, QMAN_CGR_FLAG_USE_INIT,
861 if (netif_msg_drv(priv))
862 pr_err("%s: Error %d creating CGR with ID %d\n",
863 __func__, err, priv->cgr_data.cgr.cgrid);
864 qman_release_cgrid(priv->cgr_data.cgr.cgrid);
867 if (netif_msg_drv(priv))
868 pr_debug("Created CGR %d for netdev with hwaddr %pM on QMan channel %d\n",
869 priv->cgr_data.cgr.cgrid, priv->mac_dev->addr,
870 priv->cgr_data.cgr.chan);
876 static void dpaa_eth_cgr_set_speed(struct mac_device *mac_dev, int speed)
878 struct net_device *net_dev = to_net_dev(mac_dev->phylink_config.dev);
879 struct dpaa_priv *priv = netdev_priv(net_dev);
880 struct qm_mcc_initcgr opts = { };
884 opts.we_mask = cpu_to_be16(QM_CGR_WE_CS_THRES);
887 cs_th = DPAA_CS_THRESHOLD_10G;
891 cs_th = DPAA_CS_THRESHOLD_1G;
894 qm_cgr_cs_thres_set64(&opts.cgr.cs_thres, cs_th, 1);
896 err = qman_update_cgr_safe(&priv->cgr_data.cgr, &opts);
898 netdev_err(net_dev, "could not update speed: %d\n", err);
901 static inline void dpaa_setup_ingress(const struct dpaa_priv *priv,
903 const struct qman_fq *template)
905 fq->fq_base = *template;
906 fq->net_dev = priv->net_dev;
908 fq->flags = QMAN_FQ_FLAG_NO_ENQUEUE;
909 fq->channel = priv->channel;
912 static inline void dpaa_setup_egress(const struct dpaa_priv *priv,
914 struct fman_port *port,
915 const struct qman_fq *template)
917 fq->fq_base = *template;
918 fq->net_dev = priv->net_dev;
921 fq->flags = QMAN_FQ_FLAG_TO_DCPORTAL;
922 fq->channel = (u16)fman_port_get_qman_channel_id(port);
924 fq->flags = QMAN_FQ_FLAG_NO_MODIFY;
928 static void dpaa_fq_setup(struct dpaa_priv *priv,
929 const struct dpaa_fq_cbs *fq_cbs,
930 struct fman_port *tx_port)
932 int egress_cnt = 0, conf_cnt = 0, num_portals = 0, portal_cnt = 0, cpu;
933 const cpumask_t *affine_cpus = qman_affine_cpus();
934 u16 channels[NR_CPUS];
937 for_each_cpu_and(cpu, affine_cpus, cpu_online_mask)
938 channels[num_portals++] = qman_affine_channel(cpu);
940 if (num_portals == 0)
941 dev_err(priv->net_dev->dev.parent,
942 "No Qman software (affine) channels found\n");
944 /* Initialize each FQ in the list */
945 list_for_each_entry(fq, &priv->dpaa_fq_list, list) {
946 switch (fq->fq_type) {
947 case FQ_TYPE_RX_DEFAULT:
948 dpaa_setup_ingress(priv, fq, &fq_cbs->rx_defq);
950 case FQ_TYPE_RX_ERROR:
951 dpaa_setup_ingress(priv, fq, &fq_cbs->rx_errq);
956 dpaa_setup_ingress(priv, fq, &fq_cbs->rx_defq);
957 fq->channel = channels[portal_cnt++ % num_portals];
960 dpaa_setup_egress(priv, fq, tx_port,
961 &fq_cbs->egress_ern);
962 /* If we have more Tx queues than the number of cores,
963 * just ignore the extra ones.
965 if (egress_cnt < DPAA_ETH_TXQ_NUM)
966 priv->egress_fqs[egress_cnt++] = &fq->fq_base;
968 case FQ_TYPE_TX_CONF_MQ:
969 priv->conf_fqs[conf_cnt++] = &fq->fq_base;
971 case FQ_TYPE_TX_CONFIRM:
972 dpaa_setup_ingress(priv, fq, &fq_cbs->tx_defq);
974 case FQ_TYPE_TX_ERROR:
975 dpaa_setup_ingress(priv, fq, &fq_cbs->tx_errq);
978 dev_warn(priv->net_dev->dev.parent,
979 "Unknown FQ type detected!\n");
984 /* Make sure all CPUs receive a corresponding Tx queue. */
985 while (egress_cnt < DPAA_ETH_TXQ_NUM) {
986 list_for_each_entry(fq, &priv->dpaa_fq_list, list) {
987 if (fq->fq_type != FQ_TYPE_TX)
989 priv->egress_fqs[egress_cnt++] = &fq->fq_base;
990 if (egress_cnt == DPAA_ETH_TXQ_NUM)
996 static inline int dpaa_tx_fq_to_id(const struct dpaa_priv *priv,
997 struct qman_fq *tx_fq)
1001 for (i = 0; i < DPAA_ETH_TXQ_NUM; i++)
1002 if (priv->egress_fqs[i] == tx_fq)
1008 static int dpaa_fq_init(struct dpaa_fq *dpaa_fq, bool td_enable)
1010 const struct dpaa_priv *priv;
1011 struct qman_fq *confq = NULL;
1012 struct qm_mcc_initfq initfq;
1018 priv = netdev_priv(dpaa_fq->net_dev);
1019 dev = dpaa_fq->net_dev->dev.parent;
1021 if (dpaa_fq->fqid == 0)
1022 dpaa_fq->flags |= QMAN_FQ_FLAG_DYNAMIC_FQID;
1024 dpaa_fq->init = !(dpaa_fq->flags & QMAN_FQ_FLAG_NO_MODIFY);
1026 err = qman_create_fq(dpaa_fq->fqid, dpaa_fq->flags, &dpaa_fq->fq_base);
1028 dev_err(dev, "qman_create_fq() failed\n");
1031 fq = &dpaa_fq->fq_base;
1033 if (dpaa_fq->init) {
1034 memset(&initfq, 0, sizeof(initfq));
1036 initfq.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL);
1037 /* Note: we may get to keep an empty FQ in cache */
1038 initfq.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_PREFERINCACHE);
1040 /* Try to reduce the number of portal interrupts for
1041 * Tx Confirmation FQs.
1043 if (dpaa_fq->fq_type == FQ_TYPE_TX_CONFIRM)
1044 initfq.fqd.fq_ctrl |= cpu_to_be16(QM_FQCTRL_AVOIDBLOCK);
1047 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_DESTWQ);
1049 qm_fqd_set_destwq(&initfq.fqd, dpaa_fq->channel, dpaa_fq->wq);
1051 /* Put all egress queues in a congestion group of their own.
1052 * Sensu stricto, the Tx confirmation queues are Rx FQs,
1053 * rather than Tx - but they nonetheless account for the
1054 * memory footprint on behalf of egress traffic. We therefore
1055 * place them in the netdev's CGR, along with the Tx FQs.
1057 if (dpaa_fq->fq_type == FQ_TYPE_TX ||
1058 dpaa_fq->fq_type == FQ_TYPE_TX_CONFIRM ||
1059 dpaa_fq->fq_type == FQ_TYPE_TX_CONF_MQ) {
1060 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_CGID);
1061 initfq.fqd.fq_ctrl |= cpu_to_be16(QM_FQCTRL_CGE);
1062 initfq.fqd.cgid = (u8)priv->cgr_data.cgr.cgrid;
1063 /* Set a fixed overhead accounting, in an attempt to
1064 * reduce the impact of fixed-size skb shells and the
1065 * driver's needed headroom on system memory. This is
1066 * especially the case when the egress traffic is
1067 * composed of small datagrams.
1068 * Unfortunately, QMan's OAL value is capped to an
1069 * insufficient value, but even that is better than
1070 * no overhead accounting at all.
1072 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_OAC);
1073 qm_fqd_set_oac(&initfq.fqd, QM_OAC_CG);
1074 qm_fqd_set_oal(&initfq.fqd,
1075 min(sizeof(struct sk_buff) +
1077 (size_t)FSL_QMAN_MAX_OAL));
1081 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_TDTHRESH);
1082 qm_fqd_set_taildrop(&initfq.fqd, DPAA_FQ_TD, 1);
1083 initfq.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_TDE);
1086 if (dpaa_fq->fq_type == FQ_TYPE_TX) {
1087 queue_id = dpaa_tx_fq_to_id(priv, &dpaa_fq->fq_base);
1089 confq = priv->conf_fqs[queue_id];
1092 cpu_to_be16(QM_INITFQ_WE_CONTEXTA);
1093 /* ContextA: OVOM=1(use contextA2 bits instead of ICAD)
1094 * A2V=1 (contextA A2 field is valid)
1095 * A0V=1 (contextA A0 field is valid)
1096 * B0V=1 (contextB field is valid)
1097 * ContextA A2: EBD=1 (deallocate buffers inside FMan)
1098 * ContextB B0(ASPID): 0 (absolute Virtual Storage ID)
1100 qm_fqd_context_a_set64(&initfq.fqd,
1101 0x1e00000080000000ULL);
1105 /* Put all the ingress queues in our "ingress CGR". */
1106 if (priv->use_ingress_cgr &&
1107 (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
1108 dpaa_fq->fq_type == FQ_TYPE_RX_ERROR ||
1109 dpaa_fq->fq_type == FQ_TYPE_RX_PCD)) {
1110 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_CGID);
1111 initfq.fqd.fq_ctrl |= cpu_to_be16(QM_FQCTRL_CGE);
1112 initfq.fqd.cgid = (u8)priv->ingress_cgr.cgrid;
1113 /* Set a fixed overhead accounting, just like for the
1116 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_OAC);
1117 qm_fqd_set_oac(&initfq.fqd, QM_OAC_CG);
1118 qm_fqd_set_oal(&initfq.fqd,
1119 min(sizeof(struct sk_buff) +
1121 (size_t)FSL_QMAN_MAX_OAL));
1124 /* Initialization common to all ingress queues */
1125 if (dpaa_fq->flags & QMAN_FQ_FLAG_NO_ENQUEUE) {
1126 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_CONTEXTA);
1127 initfq.fqd.fq_ctrl |= cpu_to_be16(QM_FQCTRL_HOLDACTIVE |
1128 QM_FQCTRL_CTXASTASHING);
1129 initfq.fqd.context_a.stashing.exclusive =
1130 QM_STASHING_EXCL_DATA | QM_STASHING_EXCL_CTX |
1131 QM_STASHING_EXCL_ANNOTATION;
1132 qm_fqd_set_stashing(&initfq.fqd, 1, 2,
1133 DIV_ROUND_UP(sizeof(struct qman_fq),
1137 err = qman_init_fq(fq, QMAN_INITFQ_FLAG_SCHED, &initfq);
1139 dev_err(dev, "qman_init_fq(%u) = %d\n",
1140 qman_fq_fqid(fq), err);
1141 qman_destroy_fq(fq);
1146 dpaa_fq->fqid = qman_fq_fqid(fq);
1148 if (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
1149 dpaa_fq->fq_type == FQ_TYPE_RX_PCD) {
1150 err = xdp_rxq_info_reg(&dpaa_fq->xdp_rxq, dpaa_fq->net_dev,
1153 dev_err(dev, "xdp_rxq_info_reg() = %d\n", err);
1157 err = xdp_rxq_info_reg_mem_model(&dpaa_fq->xdp_rxq,
1158 MEM_TYPE_PAGE_ORDER0, NULL);
1160 dev_err(dev, "xdp_rxq_info_reg_mem_model() = %d\n",
1162 xdp_rxq_info_unreg(&dpaa_fq->xdp_rxq);
1170 static int dpaa_fq_free_entry(struct device *dev, struct qman_fq *fq)
1172 const struct dpaa_priv *priv;
1173 struct dpaa_fq *dpaa_fq;
1178 dpaa_fq = container_of(fq, struct dpaa_fq, fq_base);
1179 priv = netdev_priv(dpaa_fq->net_dev);
1181 if (dpaa_fq->init) {
1182 err = qman_retire_fq(fq, NULL);
1183 if (err < 0 && netif_msg_drv(priv))
1184 dev_err(dev, "qman_retire_fq(%u) = %d\n",
1185 qman_fq_fqid(fq), err);
1187 error = qman_oos_fq(fq);
1188 if (error < 0 && netif_msg_drv(priv)) {
1189 dev_err(dev, "qman_oos_fq(%u) = %d\n",
1190 qman_fq_fqid(fq), error);
1196 if ((dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
1197 dpaa_fq->fq_type == FQ_TYPE_RX_PCD) &&
1198 xdp_rxq_info_is_reg(&dpaa_fq->xdp_rxq))
1199 xdp_rxq_info_unreg(&dpaa_fq->xdp_rxq);
1201 qman_destroy_fq(fq);
1202 list_del(&dpaa_fq->list);
1207 static int dpaa_fq_free(struct device *dev, struct list_head *list)
1209 struct dpaa_fq *dpaa_fq, *tmp;
1213 list_for_each_entry_safe(dpaa_fq, tmp, list, list) {
1214 error = dpaa_fq_free_entry(dev, (struct qman_fq *)dpaa_fq);
1215 if (error < 0 && err >= 0)
1222 static int dpaa_eth_init_tx_port(struct fman_port *port, struct dpaa_fq *errq,
1223 struct dpaa_fq *defq,
1224 struct dpaa_buffer_layout *buf_layout)
1226 struct fman_buffer_prefix_content buf_prefix_content;
1227 struct fman_port_params params;
1230 memset(¶ms, 0, sizeof(params));
1231 memset(&buf_prefix_content, 0, sizeof(buf_prefix_content));
1233 buf_prefix_content.priv_data_size = buf_layout->priv_data_size;
1234 buf_prefix_content.pass_prs_result = true;
1235 buf_prefix_content.pass_hash_result = true;
1236 buf_prefix_content.pass_time_stamp = true;
1237 buf_prefix_content.data_align = DPAA_FD_DATA_ALIGNMENT;
1239 params.specific_params.non_rx_params.err_fqid = errq->fqid;
1240 params.specific_params.non_rx_params.dflt_fqid = defq->fqid;
1242 err = fman_port_config(port, ¶ms);
1244 pr_err("%s: fman_port_config failed\n", __func__);
1248 err = fman_port_cfg_buf_prefix_content(port, &buf_prefix_content);
1250 pr_err("%s: fman_port_cfg_buf_prefix_content failed\n",
1255 err = fman_port_init(port);
1257 pr_err("%s: fm_port_init failed\n", __func__);
1262 static int dpaa_eth_init_rx_port(struct fman_port *port, struct dpaa_bp *bp,
1263 struct dpaa_fq *errq,
1264 struct dpaa_fq *defq, struct dpaa_fq *pcdq,
1265 struct dpaa_buffer_layout *buf_layout)
1267 struct fman_buffer_prefix_content buf_prefix_content;
1268 struct fman_port_rx_params *rx_p;
1269 struct fman_port_params params;
1272 memset(¶ms, 0, sizeof(params));
1273 memset(&buf_prefix_content, 0, sizeof(buf_prefix_content));
1275 buf_prefix_content.priv_data_size = buf_layout->priv_data_size;
1276 buf_prefix_content.pass_prs_result = true;
1277 buf_prefix_content.pass_hash_result = true;
1278 buf_prefix_content.pass_time_stamp = true;
1279 buf_prefix_content.data_align = DPAA_FD_RX_DATA_ALIGNMENT;
1281 rx_p = ¶ms.specific_params.rx_params;
1282 rx_p->err_fqid = errq->fqid;
1283 rx_p->dflt_fqid = defq->fqid;
1285 rx_p->pcd_base_fqid = pcdq->fqid;
1286 rx_p->pcd_fqs_count = DPAA_ETH_PCD_RXQ_NUM;
1289 rx_p->ext_buf_pools.num_of_pools_used = 1;
1290 rx_p->ext_buf_pools.ext_buf_pool[0].id = bp->bpid;
1291 rx_p->ext_buf_pools.ext_buf_pool[0].size = (u16)bp->size;
1293 err = fman_port_config(port, ¶ms);
1295 pr_err("%s: fman_port_config failed\n", __func__);
1299 err = fman_port_cfg_buf_prefix_content(port, &buf_prefix_content);
1301 pr_err("%s: fman_port_cfg_buf_prefix_content failed\n",
1306 err = fman_port_init(port);
1308 pr_err("%s: fm_port_init failed\n", __func__);
1313 static int dpaa_eth_init_ports(struct mac_device *mac_dev,
1315 struct fm_port_fqs *port_fqs,
1316 struct dpaa_buffer_layout *buf_layout,
1319 struct fman_port *rxport = mac_dev->port[RX];
1320 struct fman_port *txport = mac_dev->port[TX];
1323 err = dpaa_eth_init_tx_port(txport, port_fqs->tx_errq,
1324 port_fqs->tx_defq, &buf_layout[TX]);
1328 err = dpaa_eth_init_rx_port(rxport, bp, port_fqs->rx_errq,
1329 port_fqs->rx_defq, port_fqs->rx_pcdq,
1335 static int dpaa_bman_release(const struct dpaa_bp *dpaa_bp,
1336 struct bm_buffer *bmb, int cnt)
1340 err = bman_release(dpaa_bp->pool, bmb, cnt);
1341 /* Should never occur, address anyway to avoid leaking the buffers */
1342 if (WARN_ON(err) && dpaa_bp->free_buf_cb)
1344 dpaa_bp->free_buf_cb(dpaa_bp, &bmb[cnt]);
1349 static void dpaa_release_sgt_members(struct qm_sg_entry *sgt)
1351 struct bm_buffer bmb[DPAA_BUFF_RELEASE_MAX];
1352 struct dpaa_bp *dpaa_bp;
1355 memset(bmb, 0, sizeof(bmb));
1358 dpaa_bp = dpaa_bpid2pool(sgt[i].bpid);
1364 WARN_ON(qm_sg_entry_is_ext(&sgt[i]));
1366 bm_buffer_set64(&bmb[j], qm_sg_entry_get64(&sgt[i]));
1369 } while (j < ARRAY_SIZE(bmb) &&
1370 !qm_sg_entry_is_final(&sgt[i - 1]) &&
1371 sgt[i - 1].bpid == sgt[i].bpid);
1373 dpaa_bman_release(dpaa_bp, bmb, j);
1374 } while (!qm_sg_entry_is_final(&sgt[i - 1]));
1377 static void dpaa_fd_release(const struct net_device *net_dev,
1378 const struct qm_fd *fd)
1380 struct qm_sg_entry *sgt;
1381 struct dpaa_bp *dpaa_bp;
1382 struct bm_buffer bmb;
1387 bm_buffer_set64(&bmb, qm_fd_addr(fd));
1389 dpaa_bp = dpaa_bpid2pool(fd->bpid);
1393 if (qm_fd_get_format(fd) == qm_fd_sg) {
1394 vaddr = phys_to_virt(qm_fd_addr(fd));
1395 sgt = vaddr + qm_fd_get_offset(fd);
1397 dma_unmap_page(dpaa_bp->priv->rx_dma_dev, qm_fd_addr(fd),
1398 DPAA_BP_RAW_SIZE, DMA_FROM_DEVICE);
1400 dpaa_release_sgt_members(sgt);
1402 addr = dma_map_page(dpaa_bp->priv->rx_dma_dev,
1403 virt_to_page(vaddr), 0, DPAA_BP_RAW_SIZE,
1405 if (dma_mapping_error(dpaa_bp->priv->rx_dma_dev, addr)) {
1406 netdev_err(net_dev, "DMA mapping failed\n");
1409 bm_buffer_set64(&bmb, addr);
1412 dpaa_bman_release(dpaa_bp, &bmb, 1);
1415 static void count_ern(struct dpaa_percpu_priv *percpu_priv,
1416 const union qm_mr_entry *msg)
1418 switch (msg->ern.rc & QM_MR_RC_MASK) {
1419 case QM_MR_RC_CGR_TAILDROP:
1420 percpu_priv->ern_cnt.cg_tdrop++;
1423 percpu_priv->ern_cnt.wred++;
1425 case QM_MR_RC_ERROR:
1426 percpu_priv->ern_cnt.err_cond++;
1428 case QM_MR_RC_ORPWINDOW_EARLY:
1429 percpu_priv->ern_cnt.early_window++;
1431 case QM_MR_RC_ORPWINDOW_LATE:
1432 percpu_priv->ern_cnt.late_window++;
1434 case QM_MR_RC_FQ_TAILDROP:
1435 percpu_priv->ern_cnt.fq_tdrop++;
1437 case QM_MR_RC_ORPWINDOW_RETIRED:
1438 percpu_priv->ern_cnt.fq_retired++;
1440 case QM_MR_RC_ORP_ZERO:
1441 percpu_priv->ern_cnt.orp_zero++;
1446 /* Turn on HW checksum computation for this outgoing frame.
1447 * If the current protocol is not something we support in this regard
1448 * (or if the stack has already computed the SW checksum), we do nothing.
1450 * Returns 0 if all goes well (or HW csum doesn't apply), and a negative value
1453 * Note that this function may modify the fd->cmd field and the skb data buffer
1454 * (the Parse Results area).
1456 static int dpaa_enable_tx_csum(struct dpaa_priv *priv,
1457 struct sk_buff *skb,
1459 void *parse_results)
1461 struct fman_prs_result *parse_result;
1462 u16 ethertype = ntohs(skb->protocol);
1463 struct ipv6hdr *ipv6h = NULL;
1468 if (skb->ip_summed != CHECKSUM_PARTIAL)
1471 /* Note: L3 csum seems to be already computed in sw, but we can't choose
1472 * L4 alone from the FM configuration anyway.
1475 /* Fill in some fields of the Parse Results array, so the FMan
1476 * can find them as if they came from the FMan Parser.
1478 parse_result = (struct fman_prs_result *)parse_results;
1480 /* If we're dealing with VLAN, get the real Ethernet type */
1481 if (ethertype == ETH_P_8021Q) {
1482 /* We can't always assume the MAC header is set correctly
1483 * by the stack, so reset to beginning of skb->data
1485 skb_reset_mac_header(skb);
1486 ethertype = ntohs(vlan_eth_hdr(skb)->h_vlan_encapsulated_proto);
1489 /* Fill in the relevant L3 parse result fields
1490 * and read the L4 protocol type
1492 switch (ethertype) {
1494 parse_result->l3r = cpu_to_be16(FM_L3_PARSE_RESULT_IPV4);
1497 l4_proto = iph->protocol;
1500 parse_result->l3r = cpu_to_be16(FM_L3_PARSE_RESULT_IPV6);
1501 ipv6h = ipv6_hdr(skb);
1503 l4_proto = ipv6h->nexthdr;
1506 /* We shouldn't even be here */
1507 if (net_ratelimit())
1508 netif_alert(priv, tx_err, priv->net_dev,
1509 "Can't compute HW csum for L3 proto 0x%x\n",
1510 ntohs(skb->protocol));
1515 /* Fill in the relevant L4 parse result fields */
1518 parse_result->l4r = FM_L4_PARSE_RESULT_UDP;
1521 parse_result->l4r = FM_L4_PARSE_RESULT_TCP;
1524 if (net_ratelimit())
1525 netif_alert(priv, tx_err, priv->net_dev,
1526 "Can't compute HW csum for L4 proto 0x%x\n",
1532 /* At index 0 is IPOffset_1 as defined in the Parse Results */
1533 parse_result->ip_off[0] = (u8)skb_network_offset(skb);
1534 parse_result->l4_off = (u8)skb_transport_offset(skb);
1536 /* Enable L3 (and L4, if TCP or UDP) HW checksum. */
1537 fd->cmd |= cpu_to_be32(FM_FD_CMD_RPD | FM_FD_CMD_DTC);
1539 /* On P1023 and similar platforms fd->cmd interpretation could
1540 * be disabled by setting CONTEXT_A bit ICMD; currently this bit
1541 * is not set so we do not need to check; in the future, if/when
1542 * using context_a we need to check this bit
1549 static int dpaa_bp_add_8_bufs(const struct dpaa_bp *dpaa_bp)
1551 struct net_device *net_dev = dpaa_bp->priv->net_dev;
1552 struct bm_buffer bmb[8];
1557 for (i = 0; i < 8; i++) {
1558 p = dev_alloc_pages(0);
1560 netdev_err(net_dev, "dev_alloc_pages() failed\n");
1561 goto release_previous_buffs;
1564 addr = dma_map_page(dpaa_bp->priv->rx_dma_dev, p, 0,
1565 DPAA_BP_RAW_SIZE, DMA_FROM_DEVICE);
1566 if (unlikely(dma_mapping_error(dpaa_bp->priv->rx_dma_dev,
1568 netdev_err(net_dev, "DMA map failed\n");
1569 goto release_previous_buffs;
1573 bm_buffer_set64(&bmb[i], addr);
1577 return dpaa_bman_release(dpaa_bp, bmb, i);
1579 release_previous_buffs:
1580 WARN_ONCE(1, "dpaa_eth: failed to add buffers on Rx\n");
1582 bm_buffer_set64(&bmb[i], 0);
1583 /* Avoid releasing a completely null buffer; bman_release() requires
1584 * at least one buffer.
1592 static int dpaa_bp_seed(struct dpaa_bp *dpaa_bp)
1596 /* Give each CPU an allotment of "config_count" buffers */
1597 for_each_possible_cpu(i) {
1598 int *count_ptr = per_cpu_ptr(dpaa_bp->percpu_count, i);
1601 /* Although we access another CPU's counters here
1602 * we do it at boot time so it is safe
1604 for (j = 0; j < dpaa_bp->config_count; j += 8)
1605 *count_ptr += dpaa_bp_add_8_bufs(dpaa_bp);
1610 /* Add buffers/(pages) for Rx processing whenever bpool count falls below
1613 static int dpaa_eth_refill_bpool(struct dpaa_bp *dpaa_bp, int *countptr)
1615 int count = *countptr;
1618 if (unlikely(count < FSL_DPAA_ETH_REFILL_THRESHOLD)) {
1620 new_bufs = dpaa_bp_add_8_bufs(dpaa_bp);
1621 if (unlikely(!new_bufs)) {
1622 /* Avoid looping forever if we've temporarily
1623 * run out of memory. We'll try again at the
1629 } while (count < FSL_DPAA_ETH_MAX_BUF_COUNT);
1632 if (unlikely(count < FSL_DPAA_ETH_MAX_BUF_COUNT))
1639 static int dpaa_eth_refill_bpools(struct dpaa_priv *priv)
1641 struct dpaa_bp *dpaa_bp;
1644 dpaa_bp = priv->dpaa_bp;
1647 countptr = this_cpu_ptr(dpaa_bp->percpu_count);
1649 return dpaa_eth_refill_bpool(dpaa_bp, countptr);
1652 /* Cleanup function for outgoing frame descriptors that were built on Tx path,
1653 * either contiguous frames or scatter/gather ones.
1654 * Skb freeing is not handled here.
1656 * This function may be called on error paths in the Tx function, so guard
1657 * against cases when not all fd relevant fields were filled in. To avoid
1658 * reading the invalid transmission timestamp for the error paths set ts to
1661 * Return the skb backpointer, since for S/G frames the buffer containing it
1664 * No skb backpointer is set when transmitting XDP frames. Cleanup the buffer
1665 * and return NULL in this case.
1667 static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
1668 const struct qm_fd *fd, bool ts)
1670 const enum dma_data_direction dma_dir = DMA_TO_DEVICE;
1671 struct device *dev = priv->net_dev->dev.parent;
1672 struct skb_shared_hwtstamps shhwtstamps;
1673 dma_addr_t addr = qm_fd_addr(fd);
1674 void *vaddr = phys_to_virt(addr);
1675 const struct qm_sg_entry *sgt;
1676 struct dpaa_eth_swbp *swbp;
1677 struct sk_buff *skb;
1681 if (unlikely(qm_fd_get_format(fd) == qm_fd_sg)) {
1682 dma_unmap_page(priv->tx_dma_dev, addr,
1683 qm_fd_get_offset(fd) + DPAA_SGT_SIZE,
1686 /* The sgt buffer has been allocated with netdev_alloc_frag(),
1689 sgt = vaddr + qm_fd_get_offset(fd);
1691 /* sgt[0] is from lowmem, was dma_map_single()-ed */
1692 dma_unmap_single(priv->tx_dma_dev, qm_sg_addr(&sgt[0]),
1693 qm_sg_entry_get_len(&sgt[0]), dma_dir);
1695 /* remaining pages were mapped with skb_frag_dma_map() */
1696 for (i = 1; (i < DPAA_SGT_MAX_ENTRIES) &&
1697 !qm_sg_entry_is_final(&sgt[i - 1]); i++) {
1698 WARN_ON(qm_sg_entry_is_ext(&sgt[i]));
1700 dma_unmap_page(priv->tx_dma_dev, qm_sg_addr(&sgt[i]),
1701 qm_sg_entry_get_len(&sgt[i]), dma_dir);
1704 dma_unmap_single(priv->tx_dma_dev, addr,
1705 qm_fd_get_offset(fd) + qm_fd_get_length(fd),
1709 swbp = (struct dpaa_eth_swbp *)vaddr;
1712 /* No skb backpointer is set when running XDP. An xdp_frame
1713 * backpointer is saved instead.
1716 xdp_return_frame(swbp->xdpf);
1720 /* DMA unmapping is required before accessing the HW provided info */
1721 if (ts && priv->tx_tstamp &&
1722 skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
1723 memset(&shhwtstamps, 0, sizeof(shhwtstamps));
1725 if (!fman_port_get_tstamp(priv->mac_dev->port[TX], vaddr,
1727 shhwtstamps.hwtstamp = ns_to_ktime(ns);
1728 skb_tstamp_tx(skb, &shhwtstamps);
1730 dev_warn(dev, "fman_port_get_tstamp failed!\n");
1734 if (qm_fd_get_format(fd) == qm_fd_sg)
1735 /* Free the page that we allocated on Tx for the SGT */
1736 free_pages((unsigned long)vaddr, 0);
1741 static u8 rx_csum_offload(const struct dpaa_priv *priv, const struct qm_fd *fd)
1743 /* The parser has run and performed L4 checksum validation.
1744 * We know there were no parser errors (and implicitly no
1745 * L4 csum error), otherwise we wouldn't be here.
1747 if ((priv->net_dev->features & NETIF_F_RXCSUM) &&
1748 (be32_to_cpu(fd->status) & FM_FD_STAT_L4CV))
1749 return CHECKSUM_UNNECESSARY;
1751 /* We're here because either the parser didn't run or the L4 checksum
1752 * was not verified. This may include the case of a UDP frame with
1753 * checksum zero or an L4 proto other than TCP/UDP
1755 return CHECKSUM_NONE;
1758 #define PTR_IS_ALIGNED(x, a) (IS_ALIGNED((unsigned long)(x), (a)))
1760 /* Build a linear skb around the received buffer.
1761 * We are guaranteed there is enough room at the end of the data buffer to
1762 * accommodate the shared info area of the skb.
1764 static struct sk_buff *contig_fd_to_skb(const struct dpaa_priv *priv,
1765 const struct qm_fd *fd)
1767 ssize_t fd_off = qm_fd_get_offset(fd);
1768 dma_addr_t addr = qm_fd_addr(fd);
1769 struct dpaa_bp *dpaa_bp;
1770 struct sk_buff *skb;
1773 vaddr = phys_to_virt(addr);
1774 WARN_ON(!IS_ALIGNED((unsigned long)vaddr, SMP_CACHE_BYTES));
1776 dpaa_bp = dpaa_bpid2pool(fd->bpid);
1780 skb = build_skb(vaddr, dpaa_bp->size +
1781 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
1782 if (WARN_ONCE(!skb, "Build skb failure on Rx\n"))
1784 skb_reserve(skb, fd_off);
1785 skb_put(skb, qm_fd_get_length(fd));
1787 skb->ip_summed = rx_csum_offload(priv, fd);
1792 free_pages((unsigned long)vaddr, 0);
1796 /* Build an skb with the data of the first S/G entry in the linear portion and
1797 * the rest of the frame as skb fragments.
1799 * The page fragment holding the S/G Table is recycled here.
1801 static struct sk_buff *sg_fd_to_skb(const struct dpaa_priv *priv,
1802 const struct qm_fd *fd)
1804 ssize_t fd_off = qm_fd_get_offset(fd);
1805 dma_addr_t addr = qm_fd_addr(fd);
1806 const struct qm_sg_entry *sgt;
1807 struct page *page, *head_page;
1808 struct dpaa_bp *dpaa_bp;
1809 void *vaddr, *sg_vaddr;
1810 int frag_off, frag_len;
1811 struct sk_buff *skb;
1818 vaddr = phys_to_virt(addr);
1819 WARN_ON(!IS_ALIGNED((unsigned long)vaddr, SMP_CACHE_BYTES));
1821 /* Iterate through the SGT entries and add data buffers to the skb */
1822 sgt = vaddr + fd_off;
1824 for (i = 0; i < DPAA_SGT_MAX_ENTRIES; i++) {
1825 /* Extension bit is not supported */
1826 WARN_ON(qm_sg_entry_is_ext(&sgt[i]));
1828 sg_addr = qm_sg_addr(&sgt[i]);
1829 sg_vaddr = phys_to_virt(sg_addr);
1830 WARN_ON(!PTR_IS_ALIGNED(sg_vaddr, SMP_CACHE_BYTES));
1832 dma_unmap_page(priv->rx_dma_dev, sg_addr,
1833 DPAA_BP_RAW_SIZE, DMA_FROM_DEVICE);
1835 /* We may use multiple Rx pools */
1836 dpaa_bp = dpaa_bpid2pool(sgt[i].bpid);
1841 sz = dpaa_bp->size +
1842 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1843 skb = build_skb(sg_vaddr, sz);
1847 skb->ip_summed = rx_csum_offload(priv, fd);
1849 /* Make sure forwarded skbs will have enough space
1850 * on Tx, if extra headers are added.
1852 WARN_ON(fd_off != priv->rx_headroom);
1853 skb_reserve(skb, fd_off);
1854 skb_put(skb, qm_sg_entry_get_len(&sgt[i]));
1856 /* Not the first S/G entry; all data from buffer will
1857 * be added in an skb fragment; fragment index is offset
1858 * by one since first S/G entry was incorporated in the
1859 * linear part of the skb.
1861 * Caution: 'page' may be a tail page.
1863 page = virt_to_page(sg_vaddr);
1864 head_page = virt_to_head_page(sg_vaddr);
1866 /* Compute offset in (possibly tail) page */
1867 page_offset = ((unsigned long)sg_vaddr &
1869 (page_address(page) - page_address(head_page));
1870 /* page_offset only refers to the beginning of sgt[i];
1871 * but the buffer itself may have an internal offset.
1873 frag_off = qm_sg_entry_get_off(&sgt[i]) + page_offset;
1874 frag_len = qm_sg_entry_get_len(&sgt[i]);
1875 /* skb_add_rx_frag() does no checking on the page; if
1876 * we pass it a tail page, we'll end up with
1877 * bad page accounting and eventually with segafults.
1879 skb_add_rx_frag(skb, i - 1, head_page, frag_off,
1880 frag_len, dpaa_bp->size);
1883 /* Update the pool count for the current {cpu x bpool} */
1884 count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
1887 if (qm_sg_entry_is_final(&sgt[i]))
1890 WARN_ONCE(i == DPAA_SGT_MAX_ENTRIES, "No final bit on SGT\n");
1892 /* free the SG table buffer */
1893 free_pages((unsigned long)vaddr, 0);
1898 /* free all the SG entries */
1899 for (j = 0; j < DPAA_SGT_MAX_ENTRIES ; j++) {
1900 sg_addr = qm_sg_addr(&sgt[j]);
1901 sg_vaddr = phys_to_virt(sg_addr);
1902 /* all pages 0..i were unmaped */
1904 dma_unmap_page(priv->rx_dma_dev, qm_sg_addr(&sgt[j]),
1905 DPAA_BP_RAW_SIZE, DMA_FROM_DEVICE);
1906 free_pages((unsigned long)sg_vaddr, 0);
1907 /* counters 0..i-1 were decremented */
1909 dpaa_bp = dpaa_bpid2pool(sgt[j].bpid);
1911 count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
1916 if (qm_sg_entry_is_final(&sgt[j]))
1919 /* free the SGT fragment */
1920 free_pages((unsigned long)vaddr, 0);
1925 static int skb_to_contig_fd(struct dpaa_priv *priv,
1926 struct sk_buff *skb, struct qm_fd *fd,
1929 struct net_device *net_dev = priv->net_dev;
1930 enum dma_data_direction dma_dir;
1931 struct dpaa_eth_swbp *swbp;
1932 unsigned char *buff_start;
1936 /* We are guaranteed to have at least tx_headroom bytes
1937 * available, so just use that for offset.
1939 fd->bpid = FSL_DPAA_BPID_INV;
1940 buff_start = skb->data - priv->tx_headroom;
1941 dma_dir = DMA_TO_DEVICE;
1943 swbp = (struct dpaa_eth_swbp *)buff_start;
1946 /* Enable L3/L4 hardware checksum computation.
1948 * We must do this before dma_map_single(DMA_TO_DEVICE), because we may
1949 * need to write into the skb.
1951 err = dpaa_enable_tx_csum(priv, skb, fd,
1952 buff_start + DPAA_TX_PRIV_DATA_SIZE);
1953 if (unlikely(err < 0)) {
1954 if (net_ratelimit())
1955 netif_err(priv, tx_err, net_dev, "HW csum error: %d\n",
1960 /* Fill in the rest of the FD fields */
1961 qm_fd_set_contig(fd, priv->tx_headroom, skb->len);
1962 fd->cmd |= cpu_to_be32(FM_FD_CMD_FCO);
1964 /* Map the entire buffer size that may be seen by FMan, but no more */
1965 addr = dma_map_single(priv->tx_dma_dev, buff_start,
1966 priv->tx_headroom + skb->len, dma_dir);
1967 if (unlikely(dma_mapping_error(priv->tx_dma_dev, addr))) {
1968 if (net_ratelimit())
1969 netif_err(priv, tx_err, net_dev, "dma_map_single() failed\n");
1972 qm_fd_addr_set64(fd, addr);
1977 static int skb_to_sg_fd(struct dpaa_priv *priv,
1978 struct sk_buff *skb, struct qm_fd *fd)
1980 const enum dma_data_direction dma_dir = DMA_TO_DEVICE;
1981 const int nr_frags = skb_shinfo(skb)->nr_frags;
1982 struct net_device *net_dev = priv->net_dev;
1983 struct dpaa_eth_swbp *swbp;
1984 struct qm_sg_entry *sgt;
1992 /* get a page to store the SGTable */
1993 p = dev_alloc_pages(0);
1995 netdev_err(net_dev, "dev_alloc_pages() failed\n");
1998 buff_start = page_address(p);
2000 /* Enable L3/L4 hardware checksum computation.
2002 * We must do this before dma_map_single(DMA_TO_DEVICE), because we may
2003 * need to write into the skb.
2005 err = dpaa_enable_tx_csum(priv, skb, fd,
2006 buff_start + DPAA_TX_PRIV_DATA_SIZE);
2007 if (unlikely(err < 0)) {
2008 if (net_ratelimit())
2009 netif_err(priv, tx_err, net_dev, "HW csum error: %d\n",
2014 /* SGT[0] is used by the linear part */
2015 sgt = (struct qm_sg_entry *)(buff_start + priv->tx_headroom);
2016 frag_len = skb_headlen(skb);
2017 qm_sg_entry_set_len(&sgt[0], frag_len);
2018 sgt[0].bpid = FSL_DPAA_BPID_INV;
2020 addr = dma_map_single(priv->tx_dma_dev, skb->data,
2021 skb_headlen(skb), dma_dir);
2022 if (unlikely(dma_mapping_error(priv->tx_dma_dev, addr))) {
2023 netdev_err(priv->net_dev, "DMA mapping failed\n");
2025 goto sg0_map_failed;
2027 qm_sg_entry_set64(&sgt[0], addr);
2029 /* populate the rest of SGT entries */
2030 for (i = 0; i < nr_frags; i++) {
2031 frag = &skb_shinfo(skb)->frags[i];
2032 frag_len = skb_frag_size(frag);
2033 WARN_ON(!skb_frag_page(frag));
2034 addr = skb_frag_dma_map(priv->tx_dma_dev, frag, 0,
2036 if (unlikely(dma_mapping_error(priv->tx_dma_dev, addr))) {
2037 netdev_err(priv->net_dev, "DMA mapping failed\n");
2042 qm_sg_entry_set_len(&sgt[i + 1], frag_len);
2043 sgt[i + 1].bpid = FSL_DPAA_BPID_INV;
2044 sgt[i + 1].offset = 0;
2046 /* keep the offset in the address */
2047 qm_sg_entry_set64(&sgt[i + 1], addr);
2050 /* Set the final bit in the last used entry of the SGT */
2051 qm_sg_entry_set_f(&sgt[nr_frags], frag_len);
2053 /* set fd offset to priv->tx_headroom */
2054 qm_fd_set_sg(fd, priv->tx_headroom, skb->len);
2056 /* DMA map the SGT page */
2057 swbp = (struct dpaa_eth_swbp *)buff_start;
2060 addr = dma_map_page(priv->tx_dma_dev, p, 0,
2061 priv->tx_headroom + DPAA_SGT_SIZE, dma_dir);
2062 if (unlikely(dma_mapping_error(priv->tx_dma_dev, addr))) {
2063 netdev_err(priv->net_dev, "DMA mapping failed\n");
2065 goto sgt_map_failed;
2068 fd->bpid = FSL_DPAA_BPID_INV;
2069 fd->cmd |= cpu_to_be32(FM_FD_CMD_FCO);
2070 qm_fd_addr_set64(fd, addr);
2076 for (j = 0; j < i; j++)
2077 dma_unmap_page(priv->tx_dma_dev, qm_sg_addr(&sgt[j]),
2078 qm_sg_entry_get_len(&sgt[j]), dma_dir);
2081 free_pages((unsigned long)buff_start, 0);
2086 static inline int dpaa_xmit(struct dpaa_priv *priv,
2087 struct rtnl_link_stats64 *percpu_stats,
2091 struct qman_fq *egress_fq;
2094 egress_fq = priv->egress_fqs[queue];
2095 if (fd->bpid == FSL_DPAA_BPID_INV)
2096 fd->cmd |= cpu_to_be32(qman_fq_fqid(priv->conf_fqs[queue]));
2098 /* Trace this Tx fd */
2099 trace_dpaa_tx_fd(priv->net_dev, egress_fq, fd);
2101 for (i = 0; i < DPAA_ENQUEUE_RETRIES; i++) {
2102 err = qman_enqueue(egress_fq, fd);
2107 if (unlikely(err < 0)) {
2108 percpu_stats->tx_fifo_errors++;
2112 percpu_stats->tx_packets++;
2113 percpu_stats->tx_bytes += qm_fd_get_length(fd);
2118 #ifdef CONFIG_DPAA_ERRATUM_A050385
2119 static int dpaa_a050385_wa_skb(struct net_device *net_dev, struct sk_buff **s)
2121 struct dpaa_priv *priv = netdev_priv(net_dev);
2122 struct sk_buff *new_skb, *skb = *s;
2123 unsigned char *start, i;
2125 /* check linear buffer alignment */
2126 if (!PTR_IS_ALIGNED(skb->data, DPAA_A050385_ALIGN))
2129 /* linear buffers just need to have an aligned start */
2130 if (!skb_is_nonlinear(skb))
2133 /* linear data size for nonlinear skbs needs to be aligned */
2134 if (!IS_ALIGNED(skb_headlen(skb), DPAA_A050385_ALIGN))
2137 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2138 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2140 /* all fragments need to have aligned start addresses */
2141 if (!IS_ALIGNED(skb_frag_off(frag), DPAA_A050385_ALIGN))
2144 /* all but last fragment need to have aligned sizes */
2145 if (!IS_ALIGNED(skb_frag_size(frag), DPAA_A050385_ALIGN) &&
2146 (i < skb_shinfo(skb)->nr_frags - 1))
2153 /* copy all the skb content into a new linear buffer */
2154 new_skb = netdev_alloc_skb(net_dev, skb->len + DPAA_A050385_ALIGN - 1 +
2159 /* NET_SKB_PAD bytes already reserved, adding up to tx_headroom */
2160 skb_reserve(new_skb, priv->tx_headroom - NET_SKB_PAD);
2162 /* Workaround for DPAA_A050385 requires data start to be aligned */
2163 start = PTR_ALIGN(new_skb->data, DPAA_A050385_ALIGN);
2164 if (start - new_skb->data)
2165 skb_reserve(new_skb, start - new_skb->data);
2167 skb_put(new_skb, skb->len);
2168 skb_copy_bits(skb, 0, new_skb->data, skb->len);
2169 skb_copy_header(new_skb, skb);
2170 new_skb->dev = skb->dev;
2172 /* Copy relevant timestamp info from the old skb to the new */
2173 if (priv->tx_tstamp) {
2174 skb_shinfo(new_skb)->tx_flags = skb_shinfo(skb)->tx_flags;
2175 skb_shinfo(new_skb)->hwtstamps = skb_shinfo(skb)->hwtstamps;
2176 skb_shinfo(new_skb)->tskey = skb_shinfo(skb)->tskey;
2178 skb_set_owner_w(new_skb, skb->sk);
2181 /* We move the headroom when we align it so we have to reset the
2182 * network and transport header offsets relative to the new data
2183 * pointer. The checksum offload relies on these offsets.
2185 skb_set_network_header(new_skb, skb_network_offset(skb));
2186 skb_set_transport_header(new_skb, skb_transport_offset(skb));
2194 static int dpaa_a050385_wa_xdpf(struct dpaa_priv *priv,
2195 struct xdp_frame **init_xdpf)
2197 struct xdp_frame *new_xdpf, *xdpf = *init_xdpf;
2198 void *new_buff, *aligned_data;
2203 /* Check the data alignment and make sure the headroom is large
2204 * enough to store the xdpf backpointer. Use an aligned headroom
2207 * Due to alignment constraints, we give XDP access to the full 256
2208 * byte frame headroom. If the XDP program uses all of it, copy the
2209 * data to a new buffer and make room for storing the backpointer.
2211 if (PTR_IS_ALIGNED(xdpf->data, DPAA_FD_DATA_ALIGNMENT) &&
2212 xdpf->headroom >= priv->tx_headroom) {
2213 xdpf->headroom = priv->tx_headroom;
2217 /* Try to move the data inside the buffer just enough to align it and
2218 * store the xdpf backpointer. If the available headroom isn't large
2219 * enough, resort to allocating a new buffer and copying the data.
2221 aligned_data = PTR_ALIGN_DOWN(xdpf->data, DPAA_FD_DATA_ALIGNMENT);
2222 data_shift = xdpf->data - aligned_data;
2224 /* The XDP frame's headroom needs to be large enough to accommodate
2225 * shifting the data as well as storing the xdpf backpointer.
2227 if (xdpf->headroom >= data_shift + priv->tx_headroom) {
2228 memmove(aligned_data, xdpf->data, xdpf->len);
2229 xdpf->data = aligned_data;
2230 xdpf->headroom = priv->tx_headroom;
2234 /* The new xdp_frame is stored in the new buffer. Reserve enough space
2235 * in the headroom for storing it along with the driver's private
2236 * info. The headroom needs to be aligned to DPAA_FD_DATA_ALIGNMENT to
2237 * guarantee the data's alignment in the buffer.
2239 headroom = ALIGN(sizeof(*new_xdpf) + priv->tx_headroom,
2240 DPAA_FD_DATA_ALIGNMENT);
2242 /* Assure the extended headroom and data don't overflow the buffer,
2243 * while maintaining the mandatory tailroom.
2245 if (headroom + xdpf->len > DPAA_BP_RAW_SIZE -
2246 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
2249 p = dev_alloc_pages(0);
2253 /* Copy the data to the new buffer at a properly aligned offset */
2254 new_buff = page_address(p);
2255 memcpy(new_buff + headroom, xdpf->data, xdpf->len);
2257 /* Create an XDP frame around the new buffer in a similar fashion
2258 * to xdp_convert_buff_to_frame.
2260 new_xdpf = new_buff;
2261 new_xdpf->data = new_buff + headroom;
2262 new_xdpf->len = xdpf->len;
2263 new_xdpf->headroom = priv->tx_headroom;
2264 new_xdpf->frame_sz = DPAA_BP_RAW_SIZE;
2265 new_xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
2267 /* Release the initial buffer */
2268 xdp_return_frame_rx_napi(xdpf);
2270 *init_xdpf = new_xdpf;
2276 dpaa_start_xmit(struct sk_buff *skb, struct net_device *net_dev)
2278 const int queue_mapping = skb_get_queue_mapping(skb);
2279 bool nonlinear = skb_is_nonlinear(skb);
2280 struct rtnl_link_stats64 *percpu_stats;
2281 struct dpaa_percpu_priv *percpu_priv;
2282 struct netdev_queue *txq;
2283 struct dpaa_priv *priv;
2288 priv = netdev_priv(net_dev);
2289 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2290 percpu_stats = &percpu_priv->stats;
2292 qm_fd_clear_fd(&fd);
2295 /* We're going to store the skb backpointer at the beginning
2296 * of the data buffer, so we need a privately owned skb
2298 * We've made sure skb is not shared in dev->priv_flags,
2299 * we need to verify the skb head is not cloned
2301 if (skb_cow_head(skb, priv->tx_headroom))
2304 WARN_ON(skb_is_nonlinear(skb));
2307 /* MAX_SKB_FRAGS is equal or larger than our dpaa_SGT_MAX_ENTRIES;
2308 * make sure we don't feed FMan with more fragments than it supports.
2310 if (unlikely(nonlinear &&
2311 (skb_shinfo(skb)->nr_frags >= DPAA_SGT_MAX_ENTRIES))) {
2312 /* If the egress skb contains more fragments than we support
2313 * we have no choice but to linearize it ourselves.
2315 if (__skb_linearize(skb))
2318 nonlinear = skb_is_nonlinear(skb);
2321 #ifdef CONFIG_DPAA_ERRATUM_A050385
2322 if (unlikely(fman_has_errata_a050385())) {
2323 if (dpaa_a050385_wa_skb(net_dev, &skb))
2325 nonlinear = skb_is_nonlinear(skb);
2330 /* Just create a S/G fd based on the skb */
2331 err = skb_to_sg_fd(priv, skb, &fd);
2332 percpu_priv->tx_frag_skbuffs++;
2334 /* Create a contig FD from this skb */
2335 err = skb_to_contig_fd(priv, skb, &fd, &offset);
2337 if (unlikely(err < 0))
2338 goto skb_to_fd_failed;
2340 txq = netdev_get_tx_queue(net_dev, queue_mapping);
2342 /* LLTX requires to do our own update of trans_start */
2343 txq_trans_cond_update(txq);
2345 if (priv->tx_tstamp && skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
2346 fd.cmd |= cpu_to_be32(FM_FD_CMD_UPD);
2347 skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
2350 if (likely(dpaa_xmit(priv, percpu_stats, queue_mapping, &fd) == 0))
2351 return NETDEV_TX_OK;
2353 dpaa_cleanup_tx_fd(priv, &fd, false);
2356 percpu_stats->tx_errors++;
2358 return NETDEV_TX_OK;
2361 static void dpaa_rx_error(struct net_device *net_dev,
2362 const struct dpaa_priv *priv,
2363 struct dpaa_percpu_priv *percpu_priv,
2364 const struct qm_fd *fd,
2367 if (net_ratelimit())
2368 netif_err(priv, hw, net_dev, "Err FD status = 0x%08x\n",
2369 be32_to_cpu(fd->status) & FM_FD_STAT_RX_ERRORS);
2371 percpu_priv->stats.rx_errors++;
2373 if (be32_to_cpu(fd->status) & FM_FD_ERR_DMA)
2374 percpu_priv->rx_errors.dme++;
2375 if (be32_to_cpu(fd->status) & FM_FD_ERR_PHYSICAL)
2376 percpu_priv->rx_errors.fpe++;
2377 if (be32_to_cpu(fd->status) & FM_FD_ERR_SIZE)
2378 percpu_priv->rx_errors.fse++;
2379 if (be32_to_cpu(fd->status) & FM_FD_ERR_PRS_HDR_ERR)
2380 percpu_priv->rx_errors.phe++;
2382 dpaa_fd_release(net_dev, fd);
2385 static void dpaa_tx_error(struct net_device *net_dev,
2386 const struct dpaa_priv *priv,
2387 struct dpaa_percpu_priv *percpu_priv,
2388 const struct qm_fd *fd,
2391 struct sk_buff *skb;
2393 if (net_ratelimit())
2394 netif_warn(priv, hw, net_dev, "FD status = 0x%08x\n",
2395 be32_to_cpu(fd->status) & FM_FD_STAT_TX_ERRORS);
2397 percpu_priv->stats.tx_errors++;
2399 skb = dpaa_cleanup_tx_fd(priv, fd, false);
2403 static int dpaa_eth_poll(struct napi_struct *napi, int budget)
2405 struct dpaa_napi_portal *np =
2406 container_of(napi, struct dpaa_napi_portal, napi);
2411 cleaned = qman_p_poll_dqrr(np->p, budget);
2413 if (np->xdp_act & XDP_REDIRECT)
2416 if (cleaned < budget) {
2417 napi_complete_done(napi, cleaned);
2418 qman_p_irqsource_add(np->p, QM_PIRQ_DQRI);
2419 } else if (np->down) {
2420 qman_p_irqsource_add(np->p, QM_PIRQ_DQRI);
2426 static void dpaa_tx_conf(struct net_device *net_dev,
2427 const struct dpaa_priv *priv,
2428 struct dpaa_percpu_priv *percpu_priv,
2429 const struct qm_fd *fd,
2432 struct sk_buff *skb;
2434 if (unlikely(be32_to_cpu(fd->status) & FM_FD_STAT_TX_ERRORS)) {
2435 if (net_ratelimit())
2436 netif_warn(priv, hw, net_dev, "FD status = 0x%08x\n",
2437 be32_to_cpu(fd->status) &
2438 FM_FD_STAT_TX_ERRORS);
2440 percpu_priv->stats.tx_errors++;
2443 percpu_priv->tx_confirm++;
2445 skb = dpaa_cleanup_tx_fd(priv, fd, true);
2450 static inline int dpaa_eth_napi_schedule(struct dpaa_percpu_priv *percpu_priv,
2451 struct qman_portal *portal, bool sched_napi)
2454 /* Disable QMan IRQ and invoke NAPI */
2455 qman_p_irqsource_remove(portal, QM_PIRQ_DQRI);
2457 percpu_priv->np.p = portal;
2458 napi_schedule(&percpu_priv->np.napi);
2459 percpu_priv->in_interrupt++;
2465 static enum qman_cb_dqrr_result rx_error_dqrr(struct qman_portal *portal,
2467 const struct qm_dqrr_entry *dq,
2470 struct dpaa_fq *dpaa_fq = container_of(fq, struct dpaa_fq, fq_base);
2471 struct dpaa_percpu_priv *percpu_priv;
2472 struct net_device *net_dev;
2473 struct dpaa_bp *dpaa_bp;
2474 struct dpaa_priv *priv;
2476 net_dev = dpaa_fq->net_dev;
2477 priv = netdev_priv(net_dev);
2478 dpaa_bp = dpaa_bpid2pool(dq->fd.bpid);
2480 return qman_cb_dqrr_consume;
2482 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2484 if (dpaa_eth_napi_schedule(percpu_priv, portal, sched_napi))
2485 return qman_cb_dqrr_stop;
2487 dpaa_eth_refill_bpools(priv);
2488 dpaa_rx_error(net_dev, priv, percpu_priv, &dq->fd, fq->fqid);
2490 return qman_cb_dqrr_consume;
2493 static int dpaa_xdp_xmit_frame(struct net_device *net_dev,
2494 struct xdp_frame *xdpf)
2496 struct dpaa_priv *priv = netdev_priv(net_dev);
2497 struct rtnl_link_stats64 *percpu_stats;
2498 struct dpaa_percpu_priv *percpu_priv;
2499 struct dpaa_eth_swbp *swbp;
2500 struct netdev_queue *txq;
2506 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2507 percpu_stats = &percpu_priv->stats;
2509 #ifdef CONFIG_DPAA_ERRATUM_A050385
2510 if (unlikely(fman_has_errata_a050385())) {
2511 if (dpaa_a050385_wa_xdpf(priv, &xdpf)) {
2518 if (xdpf->headroom < DPAA_TX_PRIV_DATA_SIZE) {
2523 buff_start = xdpf->data - xdpf->headroom;
2525 /* Leave empty the skb backpointer at the start of the buffer.
2526 * Save the XDP frame for easy cleanup on confirmation.
2528 swbp = (struct dpaa_eth_swbp *)buff_start;
2532 qm_fd_clear_fd(&fd);
2533 fd.bpid = FSL_DPAA_BPID_INV;
2534 fd.cmd |= cpu_to_be32(FM_FD_CMD_FCO);
2535 qm_fd_set_contig(&fd, xdpf->headroom, xdpf->len);
2537 addr = dma_map_single(priv->tx_dma_dev, buff_start,
2538 xdpf->headroom + xdpf->len,
2540 if (unlikely(dma_mapping_error(priv->tx_dma_dev, addr))) {
2545 qm_fd_addr_set64(&fd, addr);
2547 /* Bump the trans_start */
2548 txq = netdev_get_tx_queue(net_dev, smp_processor_id());
2549 txq_trans_cond_update(txq);
2551 err = dpaa_xmit(priv, percpu_stats, smp_processor_id(), &fd);
2553 dma_unmap_single(priv->tx_dma_dev, addr,
2554 qm_fd_get_offset(&fd) + qm_fd_get_length(&fd),
2562 percpu_stats->tx_errors++;
2566 static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
2567 struct dpaa_fq *dpaa_fq, unsigned int *xdp_meta_len)
2569 ssize_t fd_off = qm_fd_get_offset(fd);
2570 struct bpf_prog *xdp_prog;
2571 struct xdp_frame *xdpf;
2572 struct xdp_buff xdp;
2576 xdp_prog = READ_ONCE(priv->xdp_prog);
2580 xdp_init_buff(&xdp, DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE,
2582 xdp_prepare_buff(&xdp, vaddr + fd_off - XDP_PACKET_HEADROOM,
2583 XDP_PACKET_HEADROOM, qm_fd_get_length(fd), true);
2585 /* We reserve a fixed headroom of 256 bytes under the erratum and we
2586 * offer it all to XDP programs to use. If no room is left for the
2587 * xdpf backpointer on TX, we will need to copy the data.
2588 * Disable metadata support since data realignments might be required
2589 * and the information can be lost.
2591 #ifdef CONFIG_DPAA_ERRATUM_A050385
2592 if (unlikely(fman_has_errata_a050385())) {
2593 xdp_set_data_meta_invalid(&xdp);
2594 xdp.data_hard_start = vaddr;
2595 xdp.frame_sz = DPAA_BP_RAW_SIZE;
2599 xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
2601 /* Update the length and the offset of the FD */
2602 qm_fd_set_contig(fd, xdp.data - vaddr, xdp.data_end - xdp.data);
2606 #ifdef CONFIG_DPAA_ERRATUM_A050385
2607 *xdp_meta_len = xdp_data_meta_unsupported(&xdp) ? 0 :
2608 xdp.data - xdp.data_meta;
2610 *xdp_meta_len = xdp.data - xdp.data_meta;
2614 /* We can access the full headroom when sending the frame
2617 xdp.data_hard_start = vaddr;
2618 xdp.frame_sz = DPAA_BP_RAW_SIZE;
2619 xdpf = xdp_convert_buff_to_frame(&xdp);
2620 if (unlikely(!xdpf)) {
2621 free_pages((unsigned long)vaddr, 0);
2625 if (dpaa_xdp_xmit_frame(priv->net_dev, xdpf))
2626 xdp_return_frame_rx_napi(xdpf);
2630 /* Allow redirect to use the full headroom */
2631 xdp.data_hard_start = vaddr;
2632 xdp.frame_sz = DPAA_BP_RAW_SIZE;
2634 err = xdp_do_redirect(priv->net_dev, &xdp, xdp_prog);
2636 trace_xdp_exception(priv->net_dev, xdp_prog, xdp_act);
2637 free_pages((unsigned long)vaddr, 0);
2641 bpf_warn_invalid_xdp_action(priv->net_dev, xdp_prog, xdp_act);
2644 trace_xdp_exception(priv->net_dev, xdp_prog, xdp_act);
2647 /* Free the buffer */
2648 free_pages((unsigned long)vaddr, 0);
2655 static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
2657 const struct qm_dqrr_entry *dq,
2660 bool ts_valid = false, hash_valid = false;
2661 struct skb_shared_hwtstamps *shhwtstamps;
2662 unsigned int skb_len, xdp_meta_len = 0;
2663 struct rtnl_link_stats64 *percpu_stats;
2664 struct dpaa_percpu_priv *percpu_priv;
2665 const struct qm_fd *fd = &dq->fd;
2666 dma_addr_t addr = qm_fd_addr(fd);
2667 struct dpaa_napi_portal *np;
2668 enum qm_fd_format fd_format;
2669 struct net_device *net_dev;
2670 u32 fd_status, hash_offset;
2671 struct qm_sg_entry *sgt;
2672 struct dpaa_bp *dpaa_bp;
2673 struct dpaa_fq *dpaa_fq;
2674 struct dpaa_priv *priv;
2675 struct sk_buff *skb;
2682 dpaa_fq = container_of(fq, struct dpaa_fq, fq_base);
2683 fd_status = be32_to_cpu(fd->status);
2684 fd_format = qm_fd_get_format(fd);
2685 net_dev = dpaa_fq->net_dev;
2686 priv = netdev_priv(net_dev);
2687 dpaa_bp = dpaa_bpid2pool(dq->fd.bpid);
2689 return qman_cb_dqrr_consume;
2691 /* Trace the Rx fd */
2692 trace_dpaa_rx_fd(net_dev, fq, &dq->fd);
2694 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2695 percpu_stats = &percpu_priv->stats;
2696 np = &percpu_priv->np;
2698 if (unlikely(dpaa_eth_napi_schedule(percpu_priv, portal, sched_napi)))
2699 return qman_cb_dqrr_stop;
2701 /* Make sure we didn't run out of buffers */
2702 if (unlikely(dpaa_eth_refill_bpools(priv))) {
2703 /* Unable to refill the buffer pool due to insufficient
2704 * system memory. Just release the frame back into the pool,
2705 * otherwise we'll soon end up with an empty buffer pool.
2707 dpaa_fd_release(net_dev, &dq->fd);
2708 return qman_cb_dqrr_consume;
2711 if (unlikely(fd_status & FM_FD_STAT_RX_ERRORS) != 0) {
2712 if (net_ratelimit())
2713 netif_warn(priv, hw, net_dev, "FD status = 0x%08x\n",
2714 fd_status & FM_FD_STAT_RX_ERRORS);
2716 percpu_stats->rx_errors++;
2717 dpaa_fd_release(net_dev, fd);
2718 return qman_cb_dqrr_consume;
2721 dma_unmap_page(dpaa_bp->priv->rx_dma_dev, addr, DPAA_BP_RAW_SIZE,
2724 /* prefetch the first 64 bytes of the frame or the SGT start */
2725 vaddr = phys_to_virt(addr);
2726 prefetch(vaddr + qm_fd_get_offset(fd));
2728 /* The only FD types that we may receive are contig and S/G */
2729 WARN_ON((fd_format != qm_fd_contig) && (fd_format != qm_fd_sg));
2731 /* Account for either the contig buffer or the SGT buffer (depending on
2732 * which case we were in) having been removed from the pool.
2734 count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
2737 /* Extract the timestamp stored in the headroom before running XDP */
2738 if (priv->rx_tstamp) {
2739 if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr, &ns))
2742 WARN_ONCE(1, "fman_port_get_tstamp failed!\n");
2745 /* Extract the hash stored in the headroom before running XDP */
2746 if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use &&
2747 !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
2749 hash = be32_to_cpu(*(u32 *)(vaddr + hash_offset));
2753 if (likely(fd_format == qm_fd_contig)) {
2754 xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
2755 dpaa_fq, &xdp_meta_len);
2756 np->xdp_act |= xdp_act;
2757 if (xdp_act != XDP_PASS) {
2758 percpu_stats->rx_packets++;
2759 percpu_stats->rx_bytes += qm_fd_get_length(fd);
2760 return qman_cb_dqrr_consume;
2762 skb = contig_fd_to_skb(priv, fd);
2764 /* XDP doesn't support S/G frames. Return the fragments to the
2765 * buffer pool and release the SGT.
2767 if (READ_ONCE(priv->xdp_prog)) {
2768 WARN_ONCE(1, "S/G frames not supported under XDP\n");
2769 sgt = vaddr + qm_fd_get_offset(fd);
2770 dpaa_release_sgt_members(sgt);
2771 free_pages((unsigned long)vaddr, 0);
2772 return qman_cb_dqrr_consume;
2774 skb = sg_fd_to_skb(priv, fd);
2777 return qman_cb_dqrr_consume;
2780 skb_metadata_set(skb, xdp_meta_len);
2782 /* Set the previously extracted timestamp */
2784 shhwtstamps = skb_hwtstamps(skb);
2785 memset(shhwtstamps, 0, sizeof(*shhwtstamps));
2786 shhwtstamps->hwtstamp = ns_to_ktime(ns);
2789 skb->protocol = eth_type_trans(skb, net_dev);
2791 /* Set the previously extracted hash */
2793 enum pkt_hash_types type;
2795 /* if L4 exists, it was used in the hash generation */
2796 type = be32_to_cpu(fd->status) & FM_FD_STAT_L4CV ?
2797 PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3;
2798 skb_set_hash(skb, hash, type);
2803 if (unlikely(netif_receive_skb(skb) == NET_RX_DROP)) {
2804 percpu_stats->rx_dropped++;
2805 return qman_cb_dqrr_consume;
2808 percpu_stats->rx_packets++;
2809 percpu_stats->rx_bytes += skb_len;
2811 return qman_cb_dqrr_consume;
2814 static enum qman_cb_dqrr_result conf_error_dqrr(struct qman_portal *portal,
2816 const struct qm_dqrr_entry *dq,
2819 struct dpaa_percpu_priv *percpu_priv;
2820 struct net_device *net_dev;
2821 struct dpaa_priv *priv;
2823 net_dev = ((struct dpaa_fq *)fq)->net_dev;
2824 priv = netdev_priv(net_dev);
2826 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2828 if (dpaa_eth_napi_schedule(percpu_priv, portal, sched_napi))
2829 return qman_cb_dqrr_stop;
2831 dpaa_tx_error(net_dev, priv, percpu_priv, &dq->fd, fq->fqid);
2833 return qman_cb_dqrr_consume;
2836 static enum qman_cb_dqrr_result conf_dflt_dqrr(struct qman_portal *portal,
2838 const struct qm_dqrr_entry *dq,
2841 struct dpaa_percpu_priv *percpu_priv;
2842 struct net_device *net_dev;
2843 struct dpaa_priv *priv;
2845 net_dev = ((struct dpaa_fq *)fq)->net_dev;
2846 priv = netdev_priv(net_dev);
2849 trace_dpaa_tx_conf_fd(net_dev, fq, &dq->fd);
2851 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2853 if (dpaa_eth_napi_schedule(percpu_priv, portal, sched_napi))
2854 return qman_cb_dqrr_stop;
2856 dpaa_tx_conf(net_dev, priv, percpu_priv, &dq->fd, fq->fqid);
2858 return qman_cb_dqrr_consume;
2861 static void egress_ern(struct qman_portal *portal,
2863 const union qm_mr_entry *msg)
2865 const struct qm_fd *fd = &msg->ern.fd;
2866 struct dpaa_percpu_priv *percpu_priv;
2867 const struct dpaa_priv *priv;
2868 struct net_device *net_dev;
2869 struct sk_buff *skb;
2871 net_dev = ((struct dpaa_fq *)fq)->net_dev;
2872 priv = netdev_priv(net_dev);
2873 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2875 percpu_priv->stats.tx_dropped++;
2876 percpu_priv->stats.tx_fifo_errors++;
2877 count_ern(percpu_priv, msg);
2879 skb = dpaa_cleanup_tx_fd(priv, fd, false);
2880 dev_kfree_skb_any(skb);
2883 static const struct dpaa_fq_cbs dpaa_fq_cbs = {
2884 .rx_defq = { .cb = { .dqrr = rx_default_dqrr } },
2885 .tx_defq = { .cb = { .dqrr = conf_dflt_dqrr } },
2886 .rx_errq = { .cb = { .dqrr = rx_error_dqrr } },
2887 .tx_errq = { .cb = { .dqrr = conf_error_dqrr } },
2888 .egress_ern = { .cb = { .ern = egress_ern } }
2891 static void dpaa_eth_napi_enable(struct dpaa_priv *priv)
2893 struct dpaa_percpu_priv *percpu_priv;
2896 for_each_online_cpu(i) {
2897 percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
2899 percpu_priv->np.down = false;
2900 napi_enable(&percpu_priv->np.napi);
2904 static void dpaa_eth_napi_disable(struct dpaa_priv *priv)
2906 struct dpaa_percpu_priv *percpu_priv;
2909 for_each_online_cpu(i) {
2910 percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
2912 percpu_priv->np.down = true;
2913 napi_disable(&percpu_priv->np.napi);
2917 static int dpaa_open(struct net_device *net_dev)
2919 struct mac_device *mac_dev;
2920 struct dpaa_priv *priv;
2923 priv = netdev_priv(net_dev);
2924 mac_dev = priv->mac_dev;
2925 dpaa_eth_napi_enable(priv);
2927 err = phylink_of_phy_connect(mac_dev->phylink,
2928 mac_dev->dev->of_node, 0);
2930 goto phy_init_failed;
2932 for (i = 0; i < ARRAY_SIZE(mac_dev->port); i++) {
2933 err = fman_port_enable(mac_dev->port[i]);
2935 goto mac_start_failed;
2938 err = priv->mac_dev->enable(mac_dev->fman_mac);
2940 netif_err(priv, ifup, net_dev, "mac_dev->enable() = %d\n", err);
2941 goto mac_start_failed;
2943 phylink_start(mac_dev->phylink);
2945 netif_tx_start_all_queues(net_dev);
2950 for (i = 0; i < ARRAY_SIZE(mac_dev->port); i++)
2951 fman_port_disable(mac_dev->port[i]);
2952 phylink_disconnect_phy(mac_dev->phylink);
2955 dpaa_eth_napi_disable(priv);
2960 static int dpaa_eth_stop(struct net_device *net_dev)
2962 struct dpaa_priv *priv;
2965 err = dpaa_stop(net_dev);
2967 priv = netdev_priv(net_dev);
2968 dpaa_eth_napi_disable(priv);
2973 static bool xdp_validate_mtu(struct dpaa_priv *priv, int mtu)
2975 int max_contig_data = priv->dpaa_bp->size - priv->rx_headroom;
2977 /* We do not support S/G fragments when XDP is enabled.
2978 * Limit the MTU in relation to the buffer size.
2980 if (mtu + VLAN_ETH_HLEN + ETH_FCS_LEN > max_contig_data) {
2981 dev_warn(priv->net_dev->dev.parent,
2982 "The maximum MTU for XDP is %d\n",
2983 max_contig_data - VLAN_ETH_HLEN - ETH_FCS_LEN);
2990 static int dpaa_change_mtu(struct net_device *net_dev, int new_mtu)
2992 struct dpaa_priv *priv = netdev_priv(net_dev);
2994 if (priv->xdp_prog && !xdp_validate_mtu(priv, new_mtu))
2997 net_dev->mtu = new_mtu;
3001 static int dpaa_setup_xdp(struct net_device *net_dev, struct netdev_bpf *bpf)
3003 struct dpaa_priv *priv = netdev_priv(net_dev);
3004 struct bpf_prog *old_prog;
3008 /* S/G fragments are not supported in XDP-mode */
3009 if (bpf->prog && !xdp_validate_mtu(priv, net_dev->mtu)) {
3010 NL_SET_ERR_MSG_MOD(bpf->extack, "MTU too large for XDP");
3014 up = netif_running(net_dev);
3017 dpaa_eth_stop(net_dev);
3019 old_prog = xchg(&priv->xdp_prog, bpf->prog);
3021 bpf_prog_put(old_prog);
3024 err = dpaa_open(net_dev);
3026 NL_SET_ERR_MSG_MOD(bpf->extack, "dpaa_open() failed");
3034 static int dpaa_xdp(struct net_device *net_dev, struct netdev_bpf *xdp)
3036 switch (xdp->command) {
3037 case XDP_SETUP_PROG:
3038 return dpaa_setup_xdp(net_dev, xdp);
3044 static int dpaa_xdp_xmit(struct net_device *net_dev, int n,
3045 struct xdp_frame **frames, u32 flags)
3047 struct xdp_frame *xdpf;
3050 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
3053 if (!netif_running(net_dev))
3056 for (i = 0; i < n; i++) {
3058 if (dpaa_xdp_xmit_frame(net_dev, xdpf))
3066 static int dpaa_ts_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
3068 struct dpaa_priv *priv = netdev_priv(dev);
3069 struct hwtstamp_config config;
3071 if (copy_from_user(&config, rq->ifr_data, sizeof(config)))
3074 switch (config.tx_type) {
3075 case HWTSTAMP_TX_OFF:
3076 /* Couldn't disable rx/tx timestamping separately.
3079 priv->tx_tstamp = false;
3081 case HWTSTAMP_TX_ON:
3082 priv->mac_dev->set_tstamp(priv->mac_dev->fman_mac, true);
3083 priv->tx_tstamp = true;
3089 if (config.rx_filter == HWTSTAMP_FILTER_NONE) {
3090 /* Couldn't disable rx/tx timestamping separately.
3093 priv->rx_tstamp = false;
3095 priv->mac_dev->set_tstamp(priv->mac_dev->fman_mac, true);
3096 priv->rx_tstamp = true;
3097 /* TS is set for all frame types, not only those requested */
3098 config.rx_filter = HWTSTAMP_FILTER_ALL;
3101 return copy_to_user(rq->ifr_data, &config, sizeof(config)) ?
3105 static int dpaa_ioctl(struct net_device *net_dev, struct ifreq *rq, int cmd)
3108 struct dpaa_priv *priv = netdev_priv(net_dev);
3110 if (cmd == SIOCGMIIREG) {
3111 if (net_dev->phydev)
3112 return phylink_mii_ioctl(priv->mac_dev->phylink, rq,
3116 if (cmd == SIOCSHWTSTAMP)
3117 return dpaa_ts_ioctl(net_dev, rq, cmd);
3122 static const struct net_device_ops dpaa_ops = {
3123 .ndo_open = dpaa_open,
3124 .ndo_start_xmit = dpaa_start_xmit,
3125 .ndo_stop = dpaa_eth_stop,
3126 .ndo_tx_timeout = dpaa_tx_timeout,
3127 .ndo_get_stats64 = dpaa_get_stats64,
3128 .ndo_change_carrier = fixed_phy_change_carrier,
3129 .ndo_set_mac_address = dpaa_set_mac_address,
3130 .ndo_validate_addr = eth_validate_addr,
3131 .ndo_set_rx_mode = dpaa_set_rx_mode,
3132 .ndo_eth_ioctl = dpaa_ioctl,
3133 .ndo_setup_tc = dpaa_setup_tc,
3134 .ndo_change_mtu = dpaa_change_mtu,
3135 .ndo_bpf = dpaa_xdp,
3136 .ndo_xdp_xmit = dpaa_xdp_xmit,
3139 static int dpaa_napi_add(struct net_device *net_dev)
3141 struct dpaa_priv *priv = netdev_priv(net_dev);
3142 struct dpaa_percpu_priv *percpu_priv;
3145 for_each_possible_cpu(cpu) {
3146 percpu_priv = per_cpu_ptr(priv->percpu_priv, cpu);
3148 netif_napi_add(net_dev, &percpu_priv->np.napi, dpaa_eth_poll);
3154 static void dpaa_napi_del(struct net_device *net_dev)
3156 struct dpaa_priv *priv = netdev_priv(net_dev);
3157 struct dpaa_percpu_priv *percpu_priv;
3160 for_each_possible_cpu(cpu) {
3161 percpu_priv = per_cpu_ptr(priv->percpu_priv, cpu);
3163 netif_napi_del(&percpu_priv->np.napi);
3167 static inline void dpaa_bp_free_pf(const struct dpaa_bp *bp,
3168 struct bm_buffer *bmb)
3170 dma_addr_t addr = bm_buf_addr(bmb);
3172 dma_unmap_page(bp->priv->rx_dma_dev, addr, DPAA_BP_RAW_SIZE,
3175 skb_free_frag(phys_to_virt(addr));
3178 /* Alloc the dpaa_bp struct and configure default values */
3179 static struct dpaa_bp *dpaa_bp_alloc(struct device *dev)
3181 struct dpaa_bp *dpaa_bp;
3183 dpaa_bp = devm_kzalloc(dev, sizeof(*dpaa_bp), GFP_KERNEL);
3185 return ERR_PTR(-ENOMEM);
3187 dpaa_bp->bpid = FSL_DPAA_BPID_INV;
3188 dpaa_bp->percpu_count = devm_alloc_percpu(dev, *dpaa_bp->percpu_count);
3189 if (!dpaa_bp->percpu_count)
3190 return ERR_PTR(-ENOMEM);
3192 dpaa_bp->config_count = FSL_DPAA_ETH_MAX_BUF_COUNT;
3194 dpaa_bp->seed_cb = dpaa_bp_seed;
3195 dpaa_bp->free_buf_cb = dpaa_bp_free_pf;
3200 /* Place all ingress FQs (Rx Default, Rx Error) in a dedicated CGR.
3201 * We won't be sending congestion notifications to FMan; for now, we just use
3202 * this CGR to generate enqueue rejections to FMan in order to drop the frames
3203 * before they reach our ingress queues and eat up memory.
3205 static int dpaa_ingress_cgr_init(struct dpaa_priv *priv)
3207 struct qm_mcc_initcgr initcgr;
3211 err = qman_alloc_cgrid(&priv->ingress_cgr.cgrid);
3213 if (netif_msg_drv(priv))
3214 pr_err("Error %d allocating CGR ID\n", err);
3218 /* Enable CS TD, but disable Congestion State Change Notifications. */
3219 memset(&initcgr, 0, sizeof(initcgr));
3220 initcgr.we_mask = cpu_to_be16(QM_CGR_WE_CS_THRES);
3221 initcgr.cgr.cscn_en = QM_CGR_EN;
3222 cs_th = DPAA_INGRESS_CS_THRESHOLD;
3223 qm_cgr_cs_thres_set64(&initcgr.cgr.cs_thres, cs_th, 1);
3225 initcgr.we_mask |= cpu_to_be16(QM_CGR_WE_CSTD_EN);
3226 initcgr.cgr.cstd_en = QM_CGR_EN;
3228 /* This CGR will be associated with the SWP affined to the current CPU.
3229 * However, we'll place all our ingress FQs in it.
3231 err = qman_create_cgr(&priv->ingress_cgr, QMAN_CGR_FLAG_USE_INIT,
3234 if (netif_msg_drv(priv))
3235 pr_err("Error %d creating ingress CGR with ID %d\n",
3236 err, priv->ingress_cgr.cgrid);
3237 qman_release_cgrid(priv->ingress_cgr.cgrid);
3240 if (netif_msg_drv(priv))
3241 pr_debug("Created ingress CGR %d for netdev with hwaddr %pM\n",
3242 priv->ingress_cgr.cgrid, priv->mac_dev->addr);
3244 priv->use_ingress_cgr = true;
3250 static u16 dpaa_get_headroom(struct dpaa_buffer_layout *bl,
3251 enum port_type port)
3255 /* The frame headroom must accommodate:
3256 * - the driver private data area
3257 * - parse results, hash results, timestamp if selected
3258 * If either hash results or time stamp are selected, both will
3259 * be copied to/from the frame headroom, as TS is located between PR and
3260 * HR in the IC and IC copy size has a granularity of 16bytes
3261 * (see description of FMBM_RICP and FMBM_TICP registers in DPAARM)
3263 * Also make sure the headroom is a multiple of data_align bytes
3265 headroom = (u16)(bl[port].priv_data_size + DPAA_HWA_SIZE);
3268 #ifdef CONFIG_DPAA_ERRATUM_A050385
3269 if (unlikely(fman_has_errata_a050385()))
3270 headroom = XDP_PACKET_HEADROOM;
3273 return ALIGN(headroom, DPAA_FD_RX_DATA_ALIGNMENT);
3275 return ALIGN(headroom, DPAA_FD_DATA_ALIGNMENT);
3279 static int dpaa_eth_probe(struct platform_device *pdev)
3281 struct net_device *net_dev = NULL;
3282 struct dpaa_bp *dpaa_bp = NULL;
3283 struct dpaa_fq *dpaa_fq, *tmp;
3284 struct dpaa_priv *priv = NULL;
3285 struct fm_port_fqs port_fqs;
3286 struct mac_device *mac_dev;
3287 int err = 0, channel;
3292 err = bman_is_probed();
3294 return -EPROBE_DEFER;
3296 dev_err(dev, "failing probe due to bman probe error\n");
3299 err = qman_is_probed();
3301 return -EPROBE_DEFER;
3303 dev_err(dev, "failing probe due to qman probe error\n");
3306 err = bman_portals_probed();
3308 return -EPROBE_DEFER;
3311 "failing probe due to bman portals probe error\n");
3314 err = qman_portals_probed();
3316 return -EPROBE_DEFER;
3319 "failing probe due to qman portals probe error\n");
3323 /* Allocate this early, so we can store relevant information in
3326 net_dev = alloc_etherdev_mq(sizeof(*priv), DPAA_ETH_TXQ_NUM);
3328 dev_err(dev, "alloc_etherdev_mq() failed\n");
3332 /* Do this here, so we can be verbose early */
3333 SET_NETDEV_DEV(net_dev, dev->parent);
3334 dev_set_drvdata(dev, net_dev);
3336 priv = netdev_priv(net_dev);
3337 priv->net_dev = net_dev;
3339 priv->msg_enable = netif_msg_init(debug, DPAA_MSG_DEFAULT);
3341 mac_dev = dpaa_mac_dev_get(pdev);
3342 if (IS_ERR(mac_dev)) {
3343 netdev_err(net_dev, "dpaa_mac_dev_get() failed\n");
3344 err = PTR_ERR(mac_dev);
3348 /* Devices used for DMA mapping */
3349 priv->rx_dma_dev = fman_port_get_device(mac_dev->port[RX]);
3350 priv->tx_dma_dev = fman_port_get_device(mac_dev->port[TX]);
3351 err = dma_coerce_mask_and_coherent(priv->rx_dma_dev, DMA_BIT_MASK(40));
3353 err = dma_coerce_mask_and_coherent(priv->tx_dma_dev,
3356 netdev_err(net_dev, "dma_coerce_mask_and_coherent() failed\n");
3360 /* If fsl_fm_max_frm is set to a higher value than the all-common 1500,
3361 * we choose conservatively and let the user explicitly set a higher
3362 * MTU via ifconfig. Otherwise, the user may end up with different MTUs
3364 * If on the other hand fsl_fm_max_frm has been chosen below 1500,
3365 * start with the maximum allowed.
3367 net_dev->mtu = min(dpaa_get_max_mtu(), ETH_DATA_LEN);
3369 netdev_dbg(net_dev, "Setting initial MTU on net device: %d\n",
3372 priv->buf_layout[RX].priv_data_size = DPAA_RX_PRIV_DATA_SIZE; /* Rx */
3373 priv->buf_layout[TX].priv_data_size = DPAA_TX_PRIV_DATA_SIZE; /* Tx */
3376 dpaa_bp = dpaa_bp_alloc(dev);
3377 if (IS_ERR(dpaa_bp)) {
3378 err = PTR_ERR(dpaa_bp);
3381 /* the raw size of the buffers used for reception */
3382 dpaa_bp->raw_size = DPAA_BP_RAW_SIZE;
3383 /* avoid runtime computations by keeping the usable size here */
3384 dpaa_bp->size = dpaa_bp_size(dpaa_bp->raw_size);
3385 dpaa_bp->priv = priv;
3387 err = dpaa_bp_alloc_pool(dpaa_bp);
3390 priv->dpaa_bp = dpaa_bp;
3392 INIT_LIST_HEAD(&priv->dpaa_fq_list);
3394 memset(&port_fqs, 0, sizeof(port_fqs));
3396 err = dpaa_alloc_all_fqs(dev, &priv->dpaa_fq_list, &port_fqs);
3398 dev_err(dev, "dpaa_alloc_all_fqs() failed\n");
3402 priv->mac_dev = mac_dev;
3404 channel = dpaa_get_channel();
3406 dev_err(dev, "dpaa_get_channel() failed\n");
3411 priv->channel = (u16)channel;
3413 /* Walk the CPUs with affine portals
3414 * and add this pool channel to each's dequeue mask.
3416 dpaa_eth_add_channel(priv->channel, &pdev->dev);
3418 dpaa_fq_setup(priv, &dpaa_fq_cbs, priv->mac_dev->port[TX]);
3420 /* Create a congestion group for this netdev, with
3421 * dynamically-allocated CGR ID.
3422 * Must be executed after probing the MAC, but before
3423 * assigning the egress FQs to the CGRs.
3425 err = dpaa_eth_cgr_init(priv);
3427 dev_err(dev, "Error initializing CGR\n");
3431 err = dpaa_ingress_cgr_init(priv);
3433 dev_err(dev, "Error initializing ingress CGR\n");
3434 goto delete_egress_cgr;
3437 /* Add the FQs to the interface, and make them active */
3438 list_for_each_entry_safe(dpaa_fq, tmp, &priv->dpaa_fq_list, list) {
3439 err = dpaa_fq_init(dpaa_fq, false);
3444 priv->tx_headroom = dpaa_get_headroom(priv->buf_layout, TX);
3445 priv->rx_headroom = dpaa_get_headroom(priv->buf_layout, RX);
3447 /* All real interfaces need their ports initialized */
3448 err = dpaa_eth_init_ports(mac_dev, dpaa_bp, &port_fqs,
3449 &priv->buf_layout[0], dev);
3453 /* Rx traffic distribution based on keygen hashing defaults to on */
3454 priv->keygen_in_use = true;
3456 priv->percpu_priv = devm_alloc_percpu(dev, *priv->percpu_priv);
3457 if (!priv->percpu_priv) {
3458 dev_err(dev, "devm_alloc_percpu() failed\n");
3464 netif_set_real_num_tx_queues(net_dev, priv->num_tc * DPAA_TC_TXQ_NUM);
3466 /* Initialize NAPI */
3467 err = dpaa_napi_add(net_dev);
3469 goto delete_dpaa_napi;
3471 err = dpaa_netdev_init(net_dev, &dpaa_ops, tx_timeout);
3473 goto delete_dpaa_napi;
3475 dpaa_eth_sysfs_init(&net_dev->dev);
3477 netif_info(priv, probe, net_dev, "Probed interface %s\n",
3483 dpaa_napi_del(net_dev);
3485 dpaa_fq_free(dev, &priv->dpaa_fq_list);
3486 qman_delete_cgr_safe(&priv->ingress_cgr);
3487 qman_release_cgrid(priv->ingress_cgr.cgrid);
3489 qman_delete_cgr_safe(&priv->cgr_data.cgr);
3490 qman_release_cgrid(priv->cgr_data.cgr.cgrid);
3492 dpaa_bps_free(priv);
3494 dev_set_drvdata(dev, NULL);
3495 free_netdev(net_dev);
3500 static int dpaa_remove(struct platform_device *pdev)
3502 struct net_device *net_dev;
3503 struct dpaa_priv *priv;
3508 net_dev = dev_get_drvdata(dev);
3510 priv = netdev_priv(net_dev);
3512 dpaa_eth_sysfs_remove(dev);
3514 dev_set_drvdata(dev, NULL);
3515 unregister_netdev(net_dev);
3516 phylink_destroy(priv->mac_dev->phylink);
3518 err = dpaa_fq_free(dev, &priv->dpaa_fq_list);
3520 qman_delete_cgr_safe(&priv->ingress_cgr);
3521 qman_release_cgrid(priv->ingress_cgr.cgrid);
3522 qman_delete_cgr_safe(&priv->cgr_data.cgr);
3523 qman_release_cgrid(priv->cgr_data.cgr.cgrid);
3525 dpaa_napi_del(net_dev);
3527 dpaa_bps_free(priv);
3529 free_netdev(net_dev);
3534 static const struct platform_device_id dpaa_devtype[] = {
3536 .name = "dpaa-ethernet",
3541 MODULE_DEVICE_TABLE(platform, dpaa_devtype);
3543 static struct platform_driver dpaa_driver = {
3545 .name = KBUILD_MODNAME,
3547 .id_table = dpaa_devtype,
3548 .probe = dpaa_eth_probe,
3549 .remove = dpaa_remove
3552 static int __init dpaa_load(void)
3556 pr_debug("FSL DPAA Ethernet driver\n");
3558 /* initialize dpaa_eth mirror values */
3559 dpaa_rx_extra_headroom = fman_get_rx_extra_headroom();
3560 dpaa_max_frm = fman_get_max_frm();
3562 err = platform_driver_register(&dpaa_driver);
3564 pr_err("Error, platform_driver_register() = %d\n", err);
3568 module_init(dpaa_load);
3570 static void __exit dpaa_unload(void)
3572 platform_driver_unregister(&dpaa_driver);
3574 /* Only one channel is used and needs to be released after all
3575 * interfaces are removed
3577 dpaa_release_channel();
3579 module_exit(dpaa_unload);
3581 MODULE_LICENSE("Dual BSD/GPL");
3582 MODULE_DESCRIPTION("FSL DPAA Ethernet driver");