1 /* Copyright 2008 - 2016 Freescale Semiconductor Inc.
3 * Redistribution and use in source and binary forms, with or without
4 * modification, are permitted provided that the following conditions are met:
5 * * Redistributions of source code must retain the above copyright
6 * notice, this list of conditions and the following disclaimer.
7 * * Redistributions in binary form must reproduce the above copyright
8 * notice, this list of conditions and the following disclaimer in the
9 * documentation and/or other materials provided with the distribution.
10 * * Neither the name of Freescale Semiconductor nor the
11 * names of its contributors may be used to endorse or promote products
12 * derived from this software without specific prior written permission.
14 * ALTERNATIVELY, this software may be distributed under the terms of the
15 * GNU General Public License ("GPL") as published by the Free Software
16 * Foundation, either version 2 of that License or (at your option) any
19 * THIS SOFTWARE IS PROVIDED BY Freescale Semiconductor ``AS IS'' AND ANY
20 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
21 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
22 * DISCLAIMED. IN NO EVENT SHALL Freescale Semiconductor BE LIABLE FOR ANY
23 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
24 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
25 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
26 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
28 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
31 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
33 #include <linux/init.h>
34 #include <linux/module.h>
35 #include <linux/of_platform.h>
36 #include <linux/of_mdio.h>
37 #include <linux/of_net.h>
39 #include <linux/if_arp.h>
40 #include <linux/if_vlan.h>
41 #include <linux/icmp.h>
43 #include <linux/ipv6.h>
44 #include <linux/udp.h>
45 #include <linux/tcp.h>
46 #include <linux/net.h>
47 #include <linux/skbuff.h>
48 #include <linux/etherdevice.h>
49 #include <linux/if_ether.h>
50 #include <linux/highmem.h>
51 #include <linux/percpu.h>
52 #include <linux/dma-mapping.h>
53 #include <linux/sort.h>
54 #include <soc/fsl/bman.h>
55 #include <soc/fsl/qman.h>
58 #include "fman_port.h"
62 /* CREATE_TRACE_POINTS only needs to be defined once. Other dpaa files
63 * using trace events only need to #include <trace/events/sched.h>
65 #define CREATE_TRACE_POINTS
66 #include "dpaa_eth_trace.h"
68 static int debug = -1;
69 module_param(debug, int, 0444);
70 MODULE_PARM_DESC(debug, "Module/Driver verbosity level (0=none,...,16=all)");
72 static u16 tx_timeout = 1000;
73 module_param(tx_timeout, ushort, 0444);
74 MODULE_PARM_DESC(tx_timeout, "The Tx timeout in ms");
76 #define FM_FD_STAT_RX_ERRORS \
77 (FM_FD_ERR_DMA | FM_FD_ERR_PHYSICAL | \
78 FM_FD_ERR_SIZE | FM_FD_ERR_CLS_DISCARD | \
79 FM_FD_ERR_EXTRACTION | FM_FD_ERR_NO_SCHEME | \
80 FM_FD_ERR_PRS_TIMEOUT | FM_FD_ERR_PRS_ILL_INSTRUCT | \
81 FM_FD_ERR_PRS_HDR_ERR)
83 #define FM_FD_STAT_TX_ERRORS \
84 (FM_FD_ERR_UNSUPPORTED_FORMAT | \
85 FM_FD_ERR_LENGTH | FM_FD_ERR_DMA)
87 #define DPAA_MSG_DEFAULT (NETIF_MSG_DRV | NETIF_MSG_PROBE | \
88 NETIF_MSG_LINK | NETIF_MSG_IFUP | \
91 #define DPAA_INGRESS_CS_THRESHOLD 0x10000000
92 /* Ingress congestion threshold on FMan ports
93 * The size in bytes of the ingress tail-drop threshold on FMan ports.
94 * Traffic piling up above this value will be rejected by QMan and discarded
98 /* Size in bytes of the FQ taildrop threshold */
99 #define DPAA_FQ_TD 0x200000
101 #define DPAA_CS_THRESHOLD_1G 0x06000000
102 /* Egress congestion threshold on 1G ports, range 0x1000 .. 0x10000000
103 * The size in bytes of the egress Congestion State notification threshold on
104 * 1G ports. The 1G dTSECs can quite easily be flooded by cores doing Tx in a
105 * tight loop (e.g. by sending UDP datagrams at "while(1) speed"),
106 * and the larger the frame size, the more acute the problem.
107 * So we have to find a balance between these factors:
108 * - avoiding the device staying congested for a prolonged time (risking
109 * the netdev watchdog to fire - see also the tx_timeout module param);
110 * - affecting performance of protocols such as TCP, which otherwise
111 * behave well under the congestion notification mechanism;
112 * - preventing the Tx cores from tightly-looping (as if the congestion
113 * threshold was too low to be effective);
114 * - running out of memory if the CS threshold is set too high.
117 #define DPAA_CS_THRESHOLD_10G 0x10000000
118 /* The size in bytes of the egress Congestion State notification threshold on
119 * 10G ports, range 0x1000 .. 0x10000000
122 /* Largest value that the FQD's OAL field can hold */
123 #define FSL_QMAN_MAX_OAL 127
125 /* Default alignment for start of data in an Rx FD */
126 #define DPAA_FD_DATA_ALIGNMENT 16
128 /* The DPAA requires 256 bytes reserved and mapped for the SGT */
129 #define DPAA_SGT_SIZE 256
131 /* Values for the L3R field of the FM Parse Results
133 /* L3 Type field: First IP Present IPv4 */
134 #define FM_L3_PARSE_RESULT_IPV4 0x8000
135 /* L3 Type field: First IP Present IPv6 */
136 #define FM_L3_PARSE_RESULT_IPV6 0x4000
137 /* Values for the L4R field of the FM Parse Results */
138 /* L4 Type field: UDP */
139 #define FM_L4_PARSE_RESULT_UDP 0x40
140 /* L4 Type field: TCP */
141 #define FM_L4_PARSE_RESULT_TCP 0x20
143 /* FD status field indicating whether the FM Parser has attempted to validate
144 * the L4 csum of the frame.
145 * Note that having this bit set doesn't necessarily imply that the checksum
146 * is valid. One would have to check the parse results to find that out.
148 #define FM_FD_STAT_L4CV 0x00000004
150 #define DPAA_SGT_MAX_ENTRIES 16 /* maximum number of entries in SG Table */
151 #define DPAA_BUFF_RELEASE_MAX 8 /* maximum number of buffers released at once */
153 #define FSL_DPAA_BPID_INV 0xff
154 #define FSL_DPAA_ETH_MAX_BUF_COUNT 128
155 #define FSL_DPAA_ETH_REFILL_THRESHOLD 80
157 #define DPAA_TX_PRIV_DATA_SIZE 16
158 #define DPAA_PARSE_RESULTS_SIZE sizeof(struct fman_prs_result)
159 #define DPAA_TIME_STAMP_SIZE 8
160 #define DPAA_HASH_RESULTS_SIZE 8
161 #define DPAA_RX_PRIV_DATA_SIZE (u16)(DPAA_TX_PRIV_DATA_SIZE + \
162 dpaa_rx_extra_headroom)
164 #define DPAA_ETH_PCD_RXQ_NUM 128
166 #define DPAA_ENQUEUE_RETRIES 100000
168 enum port_type {RX, TX};
171 struct dpaa_fq *tx_defq;
172 struct dpaa_fq *tx_errq;
173 struct dpaa_fq *rx_defq;
174 struct dpaa_fq *rx_errq;
175 struct dpaa_fq *rx_pcdq;
178 /* All the dpa bps in use at any moment */
179 static struct dpaa_bp *dpaa_bp_array[BM_MAX_NUM_OF_POOLS];
181 /* The raw buffer size must be cacheline aligned */
182 #define DPAA_BP_RAW_SIZE 4096
183 /* When using more than one buffer pool, the raw sizes are as follows:
186 * 3 bp: 1KB, 2KB, 4KB
187 * 4 bp: 1KB, 2KB, 4KB, 8KB
189 static inline size_t bpool_buffer_raw_size(u8 index, u8 cnt)
191 size_t res = DPAA_BP_RAW_SIZE / 4;
194 for (i = (cnt < 3) ? cnt : 3; i < 3 + index; i++)
199 /* FMan-DMA requires 16-byte alignment for Rx buffers, but SKB_DATA_ALIGN is
200 * even stronger (SMP_CACHE_BYTES-aligned), so we just get away with that,
201 * via SKB_WITH_OVERHEAD(). We can't rely on netdev_alloc_frag() giving us
202 * half-page-aligned buffers, so we reserve some more space for start-of-buffer
205 #define dpaa_bp_size(raw_size) SKB_WITH_OVERHEAD((raw_size) - SMP_CACHE_BYTES)
207 static int dpaa_max_frm;
209 static int dpaa_rx_extra_headroom;
211 #define dpaa_get_max_mtu() \
212 (dpaa_max_frm - (VLAN_ETH_HLEN + ETH_FCS_LEN))
214 static int dpaa_netdev_init(struct net_device *net_dev,
215 const struct net_device_ops *dpaa_ops,
218 struct dpaa_priv *priv = netdev_priv(net_dev);
219 struct device *dev = net_dev->dev.parent;
220 struct dpaa_percpu_priv *percpu_priv;
224 /* Although we access another CPU's private data here
225 * we do it at initialization so it is safe
227 for_each_possible_cpu(i) {
228 percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
229 percpu_priv->net_dev = net_dev;
232 net_dev->netdev_ops = dpaa_ops;
233 mac_addr = priv->mac_dev->addr;
235 net_dev->mem_start = priv->mac_dev->res->start;
236 net_dev->mem_end = priv->mac_dev->res->end;
238 net_dev->min_mtu = ETH_MIN_MTU;
239 net_dev->max_mtu = dpaa_get_max_mtu();
241 net_dev->hw_features |= (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM |
242 NETIF_F_LLTX | NETIF_F_RXHASH);
244 net_dev->hw_features |= NETIF_F_SG | NETIF_F_HIGHDMA;
245 /* The kernels enables GSO automatically, if we declare NETIF_F_SG.
246 * For conformity, we'll still declare GSO explicitly.
248 net_dev->features |= NETIF_F_GSO;
249 net_dev->features |= NETIF_F_RXCSUM;
251 net_dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
252 /* we do not want shared skbs on TX */
253 net_dev->priv_flags &= ~IFF_TX_SKB_SHARING;
255 net_dev->features |= net_dev->hw_features;
256 net_dev->vlan_features = net_dev->features;
258 memcpy(net_dev->perm_addr, mac_addr, net_dev->addr_len);
259 memcpy(net_dev->dev_addr, mac_addr, net_dev->addr_len);
261 net_dev->ethtool_ops = &dpaa_ethtool_ops;
263 net_dev->needed_headroom = priv->tx_headroom;
264 net_dev->watchdog_timeo = msecs_to_jiffies(tx_timeout);
266 /* start without the RUNNING flag, phylib controls it later */
267 netif_carrier_off(net_dev);
269 err = register_netdev(net_dev);
271 dev_err(dev, "register_netdev() = %d\n", err);
278 static int dpaa_stop(struct net_device *net_dev)
280 struct mac_device *mac_dev;
281 struct dpaa_priv *priv;
284 priv = netdev_priv(net_dev);
285 mac_dev = priv->mac_dev;
287 netif_tx_stop_all_queues(net_dev);
288 /* Allow the Fman (Tx) port to process in-flight frames before we
289 * try switching it off.
291 usleep_range(5000, 10000);
293 err = mac_dev->stop(mac_dev);
295 netif_err(priv, ifdown, net_dev, "mac_dev->stop() = %d\n",
298 for (i = 0; i < ARRAY_SIZE(mac_dev->port); i++) {
299 error = fman_port_disable(mac_dev->port[i]);
305 phy_disconnect(net_dev->phydev);
306 net_dev->phydev = NULL;
311 static void dpaa_tx_timeout(struct net_device *net_dev)
313 struct dpaa_percpu_priv *percpu_priv;
314 const struct dpaa_priv *priv;
316 priv = netdev_priv(net_dev);
317 percpu_priv = this_cpu_ptr(priv->percpu_priv);
319 netif_crit(priv, timer, net_dev, "Transmit timeout latency: %u ms\n",
320 jiffies_to_msecs(jiffies - dev_trans_start(net_dev)));
322 percpu_priv->stats.tx_errors++;
325 /* Calculates the statistics for the given device by adding the statistics
326 * collected by each CPU.
328 static void dpaa_get_stats64(struct net_device *net_dev,
329 struct rtnl_link_stats64 *s)
331 int numstats = sizeof(struct rtnl_link_stats64) / sizeof(u64);
332 struct dpaa_priv *priv = netdev_priv(net_dev);
333 struct dpaa_percpu_priv *percpu_priv;
334 u64 *netstats = (u64 *)s;
338 for_each_possible_cpu(i) {
339 percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
341 cpustats = (u64 *)&percpu_priv->stats;
343 /* add stats from all CPUs */
344 for (j = 0; j < numstats; j++)
345 netstats[j] += cpustats[j];
349 static int dpaa_setup_tc(struct net_device *net_dev, enum tc_setup_type type,
352 struct dpaa_priv *priv = netdev_priv(net_dev);
353 struct tc_mqprio_qopt *mqprio = type_data;
357 if (type != TC_SETUP_MQPRIO)
360 mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
361 num_tc = mqprio->num_tc;
363 if (num_tc == priv->num_tc)
367 netdev_reset_tc(net_dev);
371 if (num_tc > DPAA_TC_NUM) {
372 netdev_err(net_dev, "Too many traffic classes: max %d supported.\n",
377 netdev_set_num_tc(net_dev, num_tc);
379 for (i = 0; i < num_tc; i++)
380 netdev_set_tc_queue(net_dev, i, DPAA_TC_TXQ_NUM,
381 i * DPAA_TC_TXQ_NUM);
384 priv->num_tc = num_tc ? : 1;
385 netif_set_real_num_tx_queues(net_dev, priv->num_tc * DPAA_TC_TXQ_NUM);
389 static struct mac_device *dpaa_mac_dev_get(struct platform_device *pdev)
391 struct platform_device *of_dev;
392 struct dpaa_eth_data *eth_data;
393 struct device *dpaa_dev, *dev;
394 struct device_node *mac_node;
395 struct mac_device *mac_dev;
397 dpaa_dev = &pdev->dev;
398 eth_data = dpaa_dev->platform_data;
400 return ERR_PTR(-ENODEV);
402 mac_node = eth_data->mac_node;
404 of_dev = of_find_device_by_node(mac_node);
406 dev_err(dpaa_dev, "of_find_device_by_node(%pOF) failed\n",
408 of_node_put(mac_node);
409 return ERR_PTR(-EINVAL);
411 of_node_put(mac_node);
415 mac_dev = dev_get_drvdata(dev);
417 dev_err(dpaa_dev, "dev_get_drvdata(%s) failed\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 (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 err = priv->mac_dev->set_multi(net_dev, priv->mac_dev);
477 netif_err(priv, drv, net_dev, "mac_dev->set_multi() = %d\n",
481 static struct dpaa_bp *dpaa_bpid2pool(int bpid)
483 if (WARN_ON(bpid < 0 || bpid >= BM_MAX_NUM_OF_POOLS))
486 return dpaa_bp_array[bpid];
489 /* checks if this bpool is already allocated */
490 static bool dpaa_bpid2pool_use(int bpid)
492 if (dpaa_bpid2pool(bpid)) {
493 atomic_inc(&dpaa_bp_array[bpid]->refs);
500 /* called only once per bpid by dpaa_bp_alloc_pool() */
501 static void dpaa_bpid2pool_map(int bpid, struct dpaa_bp *dpaa_bp)
503 dpaa_bp_array[bpid] = dpaa_bp;
504 atomic_set(&dpaa_bp->refs, 1);
507 static int dpaa_bp_alloc_pool(struct dpaa_bp *dpaa_bp)
511 if (dpaa_bp->size == 0 || dpaa_bp->config_count == 0) {
512 pr_err("%s: Buffer pool is not properly initialized! Missing size or initial number of buffers\n",
517 /* If the pool is already specified, we only create one per bpid */
518 if (dpaa_bp->bpid != FSL_DPAA_BPID_INV &&
519 dpaa_bpid2pool_use(dpaa_bp->bpid))
522 if (dpaa_bp->bpid == FSL_DPAA_BPID_INV) {
523 dpaa_bp->pool = bman_new_pool();
524 if (!dpaa_bp->pool) {
525 pr_err("%s: bman_new_pool() failed\n",
530 dpaa_bp->bpid = (u8)bman_get_bpid(dpaa_bp->pool);
533 if (dpaa_bp->seed_cb) {
534 err = dpaa_bp->seed_cb(dpaa_bp);
536 goto pool_seed_failed;
539 dpaa_bpid2pool_map(dpaa_bp->bpid, dpaa_bp);
544 pr_err("%s: pool seeding failed\n", __func__);
545 bman_free_pool(dpaa_bp->pool);
550 /* remove and free all the buffers from the given buffer pool */
551 static void dpaa_bp_drain(struct dpaa_bp *bp)
557 struct bm_buffer bmb[8];
560 ret = bman_acquire(bp->pool, bmb, num);
563 /* we have less than 8 buffers left;
564 * drain them one by one
570 /* Pool is fully drained */
576 for (i = 0; i < num; i++)
577 bp->free_buf_cb(bp, &bmb[i]);
581 static void dpaa_bp_free(struct dpaa_bp *dpaa_bp)
583 struct dpaa_bp *bp = dpaa_bpid2pool(dpaa_bp->bpid);
585 /* the mapping between bpid and dpaa_bp is done very late in the
586 * allocation procedure; if something failed before the mapping, the bp
587 * was not configured, therefore we don't need the below instructions
592 if (!atomic_dec_and_test(&bp->refs))
598 dpaa_bp_array[bp->bpid] = NULL;
599 bman_free_pool(bp->pool);
602 static void dpaa_bps_free(struct dpaa_priv *priv)
606 for (i = 0; i < DPAA_BPS_NUM; i++)
607 dpaa_bp_free(priv->dpaa_bps[i]);
610 /* Use multiple WQs for FQ assignment:
611 * - Tx Confirmation queues go to WQ1.
612 * - Rx Error and Tx Error queues go to WQ5 (giving them a better chance
613 * to be scheduled, in case there are many more FQs in WQ6).
614 * - Rx Default goes to WQ6.
615 * - Tx queues go to different WQs depending on their priority. Equal
616 * chunks of NR_CPUS queues go to WQ6 (lowest priority), WQ2, WQ1 and
617 * WQ0 (highest priority).
618 * This ensures that Tx-confirmed buffers are timely released. In particular,
619 * it avoids congestion on the Tx Confirm FQs, which can pile up PFDRs if they
620 * are greatly outnumbered by other FQs in the system, while
621 * dequeue scheduling is round-robin.
623 static inline void dpaa_assign_wq(struct dpaa_fq *fq, int idx)
625 switch (fq->fq_type) {
626 case FQ_TYPE_TX_CONFIRM:
627 case FQ_TYPE_TX_CONF_MQ:
630 case FQ_TYPE_RX_ERROR:
631 case FQ_TYPE_TX_ERROR:
634 case FQ_TYPE_RX_DEFAULT:
639 switch (idx / DPAA_TC_TXQ_NUM) {
641 /* Low priority (best effort) */
645 /* Medium priority */
653 /* Very high priority */
657 WARN(1, "Too many TX FQs: more than %d!\n",
662 WARN(1, "Invalid FQ type %d for FQID %d!\n",
663 fq->fq_type, fq->fqid);
667 static struct dpaa_fq *dpaa_fq_alloc(struct device *dev,
668 u32 start, u32 count,
669 struct list_head *list,
670 enum dpaa_fq_type fq_type)
672 struct dpaa_fq *dpaa_fq;
675 dpaa_fq = devm_kzalloc(dev, sizeof(*dpaa_fq) * count,
680 for (i = 0; i < count; i++) {
681 dpaa_fq[i].fq_type = fq_type;
682 dpaa_fq[i].fqid = start ? start + i : 0;
683 list_add_tail(&dpaa_fq[i].list, list);
686 for (i = 0; i < count; i++)
687 dpaa_assign_wq(dpaa_fq + i, i);
692 static int dpaa_alloc_all_fqs(struct device *dev, struct list_head *list,
693 struct fm_port_fqs *port_fqs)
695 struct dpaa_fq *dpaa_fq;
696 u32 fq_base, fq_base_aligned, i;
698 dpaa_fq = dpaa_fq_alloc(dev, 0, 1, list, FQ_TYPE_RX_ERROR);
700 goto fq_alloc_failed;
702 port_fqs->rx_errq = &dpaa_fq[0];
704 dpaa_fq = dpaa_fq_alloc(dev, 0, 1, list, FQ_TYPE_RX_DEFAULT);
706 goto fq_alloc_failed;
708 port_fqs->rx_defq = &dpaa_fq[0];
710 /* the PCD FQIDs range needs to be aligned for correct operation */
711 if (qman_alloc_fqid_range(&fq_base, 2 * DPAA_ETH_PCD_RXQ_NUM))
712 goto fq_alloc_failed;
714 fq_base_aligned = ALIGN(fq_base, DPAA_ETH_PCD_RXQ_NUM);
716 for (i = fq_base; i < fq_base_aligned; i++)
717 qman_release_fqid(i);
719 for (i = fq_base_aligned + DPAA_ETH_PCD_RXQ_NUM;
720 i < (fq_base + 2 * DPAA_ETH_PCD_RXQ_NUM); i++)
721 qman_release_fqid(i);
723 dpaa_fq = dpaa_fq_alloc(dev, fq_base_aligned, DPAA_ETH_PCD_RXQ_NUM,
724 list, FQ_TYPE_RX_PCD);
726 goto fq_alloc_failed;
728 port_fqs->rx_pcdq = &dpaa_fq[0];
730 if (!dpaa_fq_alloc(dev, 0, DPAA_ETH_TXQ_NUM, list, FQ_TYPE_TX_CONF_MQ))
731 goto fq_alloc_failed;
733 dpaa_fq = dpaa_fq_alloc(dev, 0, 1, list, FQ_TYPE_TX_ERROR);
735 goto fq_alloc_failed;
737 port_fqs->tx_errq = &dpaa_fq[0];
739 dpaa_fq = dpaa_fq_alloc(dev, 0, 1, list, FQ_TYPE_TX_CONFIRM);
741 goto fq_alloc_failed;
743 port_fqs->tx_defq = &dpaa_fq[0];
745 if (!dpaa_fq_alloc(dev, 0, DPAA_ETH_TXQ_NUM, list, FQ_TYPE_TX))
746 goto fq_alloc_failed;
751 dev_err(dev, "dpaa_fq_alloc() failed\n");
755 static u32 rx_pool_channel;
756 static DEFINE_SPINLOCK(rx_pool_channel_init);
758 static int dpaa_get_channel(void)
760 spin_lock(&rx_pool_channel_init);
761 if (!rx_pool_channel) {
765 ret = qman_alloc_pool(&pool);
768 rx_pool_channel = pool;
770 spin_unlock(&rx_pool_channel_init);
771 if (!rx_pool_channel)
773 return rx_pool_channel;
776 static void dpaa_release_channel(void)
778 qman_release_pool(rx_pool_channel);
781 static void dpaa_eth_add_channel(u16 channel)
783 u32 pool = QM_SDQCR_CHANNELS_POOL_CONV(channel);
784 const cpumask_t *cpus = qman_affine_cpus();
785 struct qman_portal *portal;
788 for_each_cpu(cpu, cpus) {
789 portal = qman_get_affine_portal(cpu);
790 qman_p_static_dequeue_add(portal, pool);
794 /* Congestion group state change notification callback.
795 * Stops the device's egress queues while they are congested and
796 * wakes them upon exiting congested state.
797 * Also updates some CGR-related stats.
799 static void dpaa_eth_cgscn(struct qman_portal *qm, struct qman_cgr *cgr,
802 struct dpaa_priv *priv = (struct dpaa_priv *)container_of(cgr,
803 struct dpaa_priv, cgr_data.cgr);
806 priv->cgr_data.congestion_start_jiffies = jiffies;
807 netif_tx_stop_all_queues(priv->net_dev);
808 priv->cgr_data.cgr_congested_count++;
810 priv->cgr_data.congested_jiffies +=
811 (jiffies - priv->cgr_data.congestion_start_jiffies);
812 netif_tx_wake_all_queues(priv->net_dev);
816 static int dpaa_eth_cgr_init(struct dpaa_priv *priv)
818 struct qm_mcc_initcgr initcgr;
822 err = qman_alloc_cgrid(&priv->cgr_data.cgr.cgrid);
824 if (netif_msg_drv(priv))
825 pr_err("%s: Error %d allocating CGR ID\n",
829 priv->cgr_data.cgr.cb = dpaa_eth_cgscn;
831 /* Enable Congestion State Change Notifications and CS taildrop */
832 memset(&initcgr, 0, sizeof(initcgr));
833 initcgr.we_mask = cpu_to_be16(QM_CGR_WE_CSCN_EN | QM_CGR_WE_CS_THRES);
834 initcgr.cgr.cscn_en = QM_CGR_EN;
836 /* Set different thresholds based on the MAC speed.
837 * This may turn suboptimal if the MAC is reconfigured at a speed
838 * lower than its max, e.g. if a dTSEC later negotiates a 100Mbps link.
839 * In such cases, we ought to reconfigure the threshold, too.
841 if (priv->mac_dev->if_support & SUPPORTED_10000baseT_Full)
842 cs_th = DPAA_CS_THRESHOLD_10G;
844 cs_th = DPAA_CS_THRESHOLD_1G;
845 qm_cgr_cs_thres_set64(&initcgr.cgr.cs_thres, cs_th, 1);
847 initcgr.we_mask |= cpu_to_be16(QM_CGR_WE_CSTD_EN);
848 initcgr.cgr.cstd_en = QM_CGR_EN;
850 err = qman_create_cgr(&priv->cgr_data.cgr, QMAN_CGR_FLAG_USE_INIT,
853 if (netif_msg_drv(priv))
854 pr_err("%s: Error %d creating CGR with ID %d\n",
855 __func__, err, priv->cgr_data.cgr.cgrid);
856 qman_release_cgrid(priv->cgr_data.cgr.cgrid);
859 if (netif_msg_drv(priv))
860 pr_debug("Created CGR %d for netdev with hwaddr %pM on QMan channel %d\n",
861 priv->cgr_data.cgr.cgrid, priv->mac_dev->addr,
862 priv->cgr_data.cgr.chan);
868 static inline void dpaa_setup_ingress(const struct dpaa_priv *priv,
870 const struct qman_fq *template)
872 fq->fq_base = *template;
873 fq->net_dev = priv->net_dev;
875 fq->flags = QMAN_FQ_FLAG_NO_ENQUEUE;
876 fq->channel = priv->channel;
879 static inline void dpaa_setup_egress(const struct dpaa_priv *priv,
881 struct fman_port *port,
882 const struct qman_fq *template)
884 fq->fq_base = *template;
885 fq->net_dev = priv->net_dev;
888 fq->flags = QMAN_FQ_FLAG_TO_DCPORTAL;
889 fq->channel = (u16)fman_port_get_qman_channel_id(port);
891 fq->flags = QMAN_FQ_FLAG_NO_MODIFY;
895 static void dpaa_fq_setup(struct dpaa_priv *priv,
896 const struct dpaa_fq_cbs *fq_cbs,
897 struct fman_port *tx_port)
899 int egress_cnt = 0, conf_cnt = 0, num_portals = 0, portal_cnt = 0, cpu;
900 const cpumask_t *affine_cpus = qman_affine_cpus();
901 u16 channels[NR_CPUS];
904 for_each_cpu(cpu, affine_cpus)
905 channels[num_portals++] = qman_affine_channel(cpu);
907 if (num_portals == 0)
908 dev_err(priv->net_dev->dev.parent,
909 "No Qman software (affine) channels found");
911 /* Initialize each FQ in the list */
912 list_for_each_entry(fq, &priv->dpaa_fq_list, list) {
913 switch (fq->fq_type) {
914 case FQ_TYPE_RX_DEFAULT:
915 dpaa_setup_ingress(priv, fq, &fq_cbs->rx_defq);
917 case FQ_TYPE_RX_ERROR:
918 dpaa_setup_ingress(priv, fq, &fq_cbs->rx_errq);
923 dpaa_setup_ingress(priv, fq, &fq_cbs->rx_defq);
924 fq->channel = channels[portal_cnt++ % num_portals];
927 dpaa_setup_egress(priv, fq, tx_port,
928 &fq_cbs->egress_ern);
929 /* If we have more Tx queues than the number of cores,
930 * just ignore the extra ones.
932 if (egress_cnt < DPAA_ETH_TXQ_NUM)
933 priv->egress_fqs[egress_cnt++] = &fq->fq_base;
935 case FQ_TYPE_TX_CONF_MQ:
936 priv->conf_fqs[conf_cnt++] = &fq->fq_base;
938 case FQ_TYPE_TX_CONFIRM:
939 dpaa_setup_ingress(priv, fq, &fq_cbs->tx_defq);
941 case FQ_TYPE_TX_ERROR:
942 dpaa_setup_ingress(priv, fq, &fq_cbs->tx_errq);
945 dev_warn(priv->net_dev->dev.parent,
946 "Unknown FQ type detected!\n");
951 /* Make sure all CPUs receive a corresponding Tx queue. */
952 while (egress_cnt < DPAA_ETH_TXQ_NUM) {
953 list_for_each_entry(fq, &priv->dpaa_fq_list, list) {
954 if (fq->fq_type != FQ_TYPE_TX)
956 priv->egress_fqs[egress_cnt++] = &fq->fq_base;
957 if (egress_cnt == DPAA_ETH_TXQ_NUM)
963 static inline int dpaa_tx_fq_to_id(const struct dpaa_priv *priv,
964 struct qman_fq *tx_fq)
968 for (i = 0; i < DPAA_ETH_TXQ_NUM; i++)
969 if (priv->egress_fqs[i] == tx_fq)
975 static int dpaa_fq_init(struct dpaa_fq *dpaa_fq, bool td_enable)
977 const struct dpaa_priv *priv;
978 struct qman_fq *confq = NULL;
979 struct qm_mcc_initfq initfq;
985 priv = netdev_priv(dpaa_fq->net_dev);
986 dev = dpaa_fq->net_dev->dev.parent;
988 if (dpaa_fq->fqid == 0)
989 dpaa_fq->flags |= QMAN_FQ_FLAG_DYNAMIC_FQID;
991 dpaa_fq->init = !(dpaa_fq->flags & QMAN_FQ_FLAG_NO_MODIFY);
993 err = qman_create_fq(dpaa_fq->fqid, dpaa_fq->flags, &dpaa_fq->fq_base);
995 dev_err(dev, "qman_create_fq() failed\n");
998 fq = &dpaa_fq->fq_base;
1000 if (dpaa_fq->init) {
1001 memset(&initfq, 0, sizeof(initfq));
1003 initfq.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL);
1004 /* Note: we may get to keep an empty FQ in cache */
1005 initfq.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_PREFERINCACHE);
1007 /* Try to reduce the number of portal interrupts for
1008 * Tx Confirmation FQs.
1010 if (dpaa_fq->fq_type == FQ_TYPE_TX_CONFIRM)
1011 initfq.fqd.fq_ctrl |= cpu_to_be16(QM_FQCTRL_AVOIDBLOCK);
1014 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_DESTWQ);
1016 qm_fqd_set_destwq(&initfq.fqd, dpaa_fq->channel, dpaa_fq->wq);
1018 /* Put all egress queues in a congestion group of their own.
1019 * Sensu stricto, the Tx confirmation queues are Rx FQs,
1020 * rather than Tx - but they nonetheless account for the
1021 * memory footprint on behalf of egress traffic. We therefore
1022 * place them in the netdev's CGR, along with the Tx FQs.
1024 if (dpaa_fq->fq_type == FQ_TYPE_TX ||
1025 dpaa_fq->fq_type == FQ_TYPE_TX_CONFIRM ||
1026 dpaa_fq->fq_type == FQ_TYPE_TX_CONF_MQ) {
1027 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_CGID);
1028 initfq.fqd.fq_ctrl |= cpu_to_be16(QM_FQCTRL_CGE);
1029 initfq.fqd.cgid = (u8)priv->cgr_data.cgr.cgrid;
1030 /* Set a fixed overhead accounting, in an attempt to
1031 * reduce the impact of fixed-size skb shells and the
1032 * driver's needed headroom on system memory. This is
1033 * especially the case when the egress traffic is
1034 * composed of small datagrams.
1035 * Unfortunately, QMan's OAL value is capped to an
1036 * insufficient value, but even that is better than
1037 * no overhead accounting at all.
1039 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_OAC);
1040 qm_fqd_set_oac(&initfq.fqd, QM_OAC_CG);
1041 qm_fqd_set_oal(&initfq.fqd,
1042 min(sizeof(struct sk_buff) +
1044 (size_t)FSL_QMAN_MAX_OAL));
1048 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_TDTHRESH);
1049 qm_fqd_set_taildrop(&initfq.fqd, DPAA_FQ_TD, 1);
1050 initfq.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_TDE);
1053 if (dpaa_fq->fq_type == FQ_TYPE_TX) {
1054 queue_id = dpaa_tx_fq_to_id(priv, &dpaa_fq->fq_base);
1056 confq = priv->conf_fqs[queue_id];
1059 cpu_to_be16(QM_INITFQ_WE_CONTEXTA);
1060 /* ContextA: OVOM=1(use contextA2 bits instead of ICAD)
1061 * A2V=1 (contextA A2 field is valid)
1062 * A0V=1 (contextA A0 field is valid)
1063 * B0V=1 (contextB field is valid)
1064 * ContextA A2: EBD=1 (deallocate buffers inside FMan)
1065 * ContextB B0(ASPID): 0 (absolute Virtual Storage ID)
1067 qm_fqd_context_a_set64(&initfq.fqd,
1068 0x1e00000080000000ULL);
1072 /* Put all the ingress queues in our "ingress CGR". */
1073 if (priv->use_ingress_cgr &&
1074 (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
1075 dpaa_fq->fq_type == FQ_TYPE_RX_ERROR ||
1076 dpaa_fq->fq_type == FQ_TYPE_RX_PCD)) {
1077 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_CGID);
1078 initfq.fqd.fq_ctrl |= cpu_to_be16(QM_FQCTRL_CGE);
1079 initfq.fqd.cgid = (u8)priv->ingress_cgr.cgrid;
1080 /* Set a fixed overhead accounting, just like for the
1083 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_OAC);
1084 qm_fqd_set_oac(&initfq.fqd, QM_OAC_CG);
1085 qm_fqd_set_oal(&initfq.fqd,
1086 min(sizeof(struct sk_buff) +
1088 (size_t)FSL_QMAN_MAX_OAL));
1091 /* Initialization common to all ingress queues */
1092 if (dpaa_fq->flags & QMAN_FQ_FLAG_NO_ENQUEUE) {
1093 initfq.we_mask |= cpu_to_be16(QM_INITFQ_WE_CONTEXTA);
1094 initfq.fqd.fq_ctrl |= cpu_to_be16(QM_FQCTRL_HOLDACTIVE |
1095 QM_FQCTRL_CTXASTASHING);
1096 initfq.fqd.context_a.stashing.exclusive =
1097 QM_STASHING_EXCL_DATA | QM_STASHING_EXCL_CTX |
1098 QM_STASHING_EXCL_ANNOTATION;
1099 qm_fqd_set_stashing(&initfq.fqd, 1, 2,
1100 DIV_ROUND_UP(sizeof(struct qman_fq),
1104 err = qman_init_fq(fq, QMAN_INITFQ_FLAG_SCHED, &initfq);
1106 dev_err(dev, "qman_init_fq(%u) = %d\n",
1107 qman_fq_fqid(fq), err);
1108 qman_destroy_fq(fq);
1113 dpaa_fq->fqid = qman_fq_fqid(fq);
1118 static int dpaa_fq_free_entry(struct device *dev, struct qman_fq *fq)
1120 const struct dpaa_priv *priv;
1121 struct dpaa_fq *dpaa_fq;
1126 dpaa_fq = container_of(fq, struct dpaa_fq, fq_base);
1127 priv = netdev_priv(dpaa_fq->net_dev);
1129 if (dpaa_fq->init) {
1130 err = qman_retire_fq(fq, NULL);
1131 if (err < 0 && netif_msg_drv(priv))
1132 dev_err(dev, "qman_retire_fq(%u) = %d\n",
1133 qman_fq_fqid(fq), err);
1135 error = qman_oos_fq(fq);
1136 if (error < 0 && netif_msg_drv(priv)) {
1137 dev_err(dev, "qman_oos_fq(%u) = %d\n",
1138 qman_fq_fqid(fq), error);
1144 qman_destroy_fq(fq);
1145 list_del(&dpaa_fq->list);
1150 static int dpaa_fq_free(struct device *dev, struct list_head *list)
1152 struct dpaa_fq *dpaa_fq, *tmp;
1156 list_for_each_entry_safe(dpaa_fq, tmp, list, list) {
1157 error = dpaa_fq_free_entry(dev, (struct qman_fq *)dpaa_fq);
1158 if (error < 0 && err >= 0)
1165 static int dpaa_eth_init_tx_port(struct fman_port *port, struct dpaa_fq *errq,
1166 struct dpaa_fq *defq,
1167 struct dpaa_buffer_layout *buf_layout)
1169 struct fman_buffer_prefix_content buf_prefix_content;
1170 struct fman_port_params params;
1173 memset(¶ms, 0, sizeof(params));
1174 memset(&buf_prefix_content, 0, sizeof(buf_prefix_content));
1176 buf_prefix_content.priv_data_size = buf_layout->priv_data_size;
1177 buf_prefix_content.pass_prs_result = true;
1178 buf_prefix_content.pass_hash_result = true;
1179 buf_prefix_content.pass_time_stamp = false;
1180 buf_prefix_content.data_align = DPAA_FD_DATA_ALIGNMENT;
1182 params.specific_params.non_rx_params.err_fqid = errq->fqid;
1183 params.specific_params.non_rx_params.dflt_fqid = defq->fqid;
1185 err = fman_port_config(port, ¶ms);
1187 pr_err("%s: fman_port_config failed\n", __func__);
1191 err = fman_port_cfg_buf_prefix_content(port, &buf_prefix_content);
1193 pr_err("%s: fman_port_cfg_buf_prefix_content failed\n",
1198 err = fman_port_init(port);
1200 pr_err("%s: fm_port_init failed\n", __func__);
1205 static int dpaa_eth_init_rx_port(struct fman_port *port, struct dpaa_bp **bps,
1206 size_t count, struct dpaa_fq *errq,
1207 struct dpaa_fq *defq, struct dpaa_fq *pcdq,
1208 struct dpaa_buffer_layout *buf_layout)
1210 struct fman_buffer_prefix_content buf_prefix_content;
1211 struct fman_port_rx_params *rx_p;
1212 struct fman_port_params params;
1215 memset(¶ms, 0, sizeof(params));
1216 memset(&buf_prefix_content, 0, sizeof(buf_prefix_content));
1218 buf_prefix_content.priv_data_size = buf_layout->priv_data_size;
1219 buf_prefix_content.pass_prs_result = true;
1220 buf_prefix_content.pass_hash_result = true;
1221 buf_prefix_content.pass_time_stamp = false;
1222 buf_prefix_content.data_align = DPAA_FD_DATA_ALIGNMENT;
1224 rx_p = ¶ms.specific_params.rx_params;
1225 rx_p->err_fqid = errq->fqid;
1226 rx_p->dflt_fqid = defq->fqid;
1228 rx_p->pcd_base_fqid = pcdq->fqid;
1229 rx_p->pcd_fqs_count = DPAA_ETH_PCD_RXQ_NUM;
1232 count = min(ARRAY_SIZE(rx_p->ext_buf_pools.ext_buf_pool), count);
1233 rx_p->ext_buf_pools.num_of_pools_used = (u8)count;
1234 for (i = 0; i < count; i++) {
1235 rx_p->ext_buf_pools.ext_buf_pool[i].id = bps[i]->bpid;
1236 rx_p->ext_buf_pools.ext_buf_pool[i].size = (u16)bps[i]->size;
1239 err = fman_port_config(port, ¶ms);
1241 pr_err("%s: fman_port_config failed\n", __func__);
1245 err = fman_port_cfg_buf_prefix_content(port, &buf_prefix_content);
1247 pr_err("%s: fman_port_cfg_buf_prefix_content failed\n",
1252 err = fman_port_init(port);
1254 pr_err("%s: fm_port_init failed\n", __func__);
1259 static int dpaa_eth_init_ports(struct mac_device *mac_dev,
1260 struct dpaa_bp **bps, size_t count,
1261 struct fm_port_fqs *port_fqs,
1262 struct dpaa_buffer_layout *buf_layout,
1265 struct fman_port *rxport = mac_dev->port[RX];
1266 struct fman_port *txport = mac_dev->port[TX];
1269 err = dpaa_eth_init_tx_port(txport, port_fqs->tx_errq,
1270 port_fqs->tx_defq, &buf_layout[TX]);
1274 err = dpaa_eth_init_rx_port(rxport, bps, count, port_fqs->rx_errq,
1275 port_fqs->rx_defq, port_fqs->rx_pcdq,
1281 static int dpaa_bman_release(const struct dpaa_bp *dpaa_bp,
1282 struct bm_buffer *bmb, int cnt)
1286 err = bman_release(dpaa_bp->pool, bmb, cnt);
1287 /* Should never occur, address anyway to avoid leaking the buffers */
1288 if (unlikely(WARN_ON(err)) && dpaa_bp->free_buf_cb)
1290 dpaa_bp->free_buf_cb(dpaa_bp, &bmb[cnt]);
1295 static void dpaa_release_sgt_members(struct qm_sg_entry *sgt)
1297 struct bm_buffer bmb[DPAA_BUFF_RELEASE_MAX];
1298 struct dpaa_bp *dpaa_bp;
1301 memset(bmb, 0, sizeof(bmb));
1304 dpaa_bp = dpaa_bpid2pool(sgt[i].bpid);
1310 WARN_ON(qm_sg_entry_is_ext(&sgt[i]));
1312 bm_buffer_set64(&bmb[j], qm_sg_entry_get64(&sgt[i]));
1315 } while (j < ARRAY_SIZE(bmb) &&
1316 !qm_sg_entry_is_final(&sgt[i - 1]) &&
1317 sgt[i - 1].bpid == sgt[i].bpid);
1319 dpaa_bman_release(dpaa_bp, bmb, j);
1320 } while (!qm_sg_entry_is_final(&sgt[i - 1]));
1323 static void dpaa_fd_release(const struct net_device *net_dev,
1324 const struct qm_fd *fd)
1326 struct qm_sg_entry *sgt;
1327 struct dpaa_bp *dpaa_bp;
1328 struct bm_buffer bmb;
1333 bm_buffer_set64(&bmb, qm_fd_addr(fd));
1335 dpaa_bp = dpaa_bpid2pool(fd->bpid);
1339 if (qm_fd_get_format(fd) == qm_fd_sg) {
1340 vaddr = phys_to_virt(qm_fd_addr(fd));
1341 sgt = vaddr + qm_fd_get_offset(fd);
1343 dma_unmap_single(dpaa_bp->dev, qm_fd_addr(fd), dpaa_bp->size,
1346 dpaa_release_sgt_members(sgt);
1348 addr = dma_map_single(dpaa_bp->dev, vaddr, dpaa_bp->size,
1350 if (dma_mapping_error(dpaa_bp->dev, addr)) {
1351 dev_err(dpaa_bp->dev, "DMA mapping failed");
1354 bm_buffer_set64(&bmb, addr);
1357 dpaa_bman_release(dpaa_bp, &bmb, 1);
1360 static void count_ern(struct dpaa_percpu_priv *percpu_priv,
1361 const union qm_mr_entry *msg)
1363 switch (msg->ern.rc & QM_MR_RC_MASK) {
1364 case QM_MR_RC_CGR_TAILDROP:
1365 percpu_priv->ern_cnt.cg_tdrop++;
1368 percpu_priv->ern_cnt.wred++;
1370 case QM_MR_RC_ERROR:
1371 percpu_priv->ern_cnt.err_cond++;
1373 case QM_MR_RC_ORPWINDOW_EARLY:
1374 percpu_priv->ern_cnt.early_window++;
1376 case QM_MR_RC_ORPWINDOW_LATE:
1377 percpu_priv->ern_cnt.late_window++;
1379 case QM_MR_RC_FQ_TAILDROP:
1380 percpu_priv->ern_cnt.fq_tdrop++;
1382 case QM_MR_RC_ORPWINDOW_RETIRED:
1383 percpu_priv->ern_cnt.fq_retired++;
1385 case QM_MR_RC_ORP_ZERO:
1386 percpu_priv->ern_cnt.orp_zero++;
1391 /* Turn on HW checksum computation for this outgoing frame.
1392 * If the current protocol is not something we support in this regard
1393 * (or if the stack has already computed the SW checksum), we do nothing.
1395 * Returns 0 if all goes well (or HW csum doesn't apply), and a negative value
1398 * Note that this function may modify the fd->cmd field and the skb data buffer
1399 * (the Parse Results area).
1401 static int dpaa_enable_tx_csum(struct dpaa_priv *priv,
1402 struct sk_buff *skb,
1404 char *parse_results)
1406 struct fman_prs_result *parse_result;
1407 u16 ethertype = ntohs(skb->protocol);
1408 struct ipv6hdr *ipv6h = NULL;
1413 if (skb->ip_summed != CHECKSUM_PARTIAL)
1416 /* Note: L3 csum seems to be already computed in sw, but we can't choose
1417 * L4 alone from the FM configuration anyway.
1420 /* Fill in some fields of the Parse Results array, so the FMan
1421 * can find them as if they came from the FMan Parser.
1423 parse_result = (struct fman_prs_result *)parse_results;
1425 /* If we're dealing with VLAN, get the real Ethernet type */
1426 if (ethertype == ETH_P_8021Q) {
1427 /* We can't always assume the MAC header is set correctly
1428 * by the stack, so reset to beginning of skb->data
1430 skb_reset_mac_header(skb);
1431 ethertype = ntohs(vlan_eth_hdr(skb)->h_vlan_encapsulated_proto);
1434 /* Fill in the relevant L3 parse result fields
1435 * and read the L4 protocol type
1437 switch (ethertype) {
1439 parse_result->l3r = cpu_to_be16(FM_L3_PARSE_RESULT_IPV4);
1442 l4_proto = iph->protocol;
1445 parse_result->l3r = cpu_to_be16(FM_L3_PARSE_RESULT_IPV6);
1446 ipv6h = ipv6_hdr(skb);
1448 l4_proto = ipv6h->nexthdr;
1451 /* We shouldn't even be here */
1452 if (net_ratelimit())
1453 netif_alert(priv, tx_err, priv->net_dev,
1454 "Can't compute HW csum for L3 proto 0x%x\n",
1455 ntohs(skb->protocol));
1460 /* Fill in the relevant L4 parse result fields */
1463 parse_result->l4r = FM_L4_PARSE_RESULT_UDP;
1466 parse_result->l4r = FM_L4_PARSE_RESULT_TCP;
1469 if (net_ratelimit())
1470 netif_alert(priv, tx_err, priv->net_dev,
1471 "Can't compute HW csum for L4 proto 0x%x\n",
1477 /* At index 0 is IPOffset_1 as defined in the Parse Results */
1478 parse_result->ip_off[0] = (u8)skb_network_offset(skb);
1479 parse_result->l4_off = (u8)skb_transport_offset(skb);
1481 /* Enable L3 (and L4, if TCP or UDP) HW checksum. */
1482 fd->cmd |= cpu_to_be32(FM_FD_CMD_RPD | FM_FD_CMD_DTC);
1484 /* On P1023 and similar platforms fd->cmd interpretation could
1485 * be disabled by setting CONTEXT_A bit ICMD; currently this bit
1486 * is not set so we do not need to check; in the future, if/when
1487 * using context_a we need to check this bit
1494 static int dpaa_bp_add_8_bufs(const struct dpaa_bp *dpaa_bp)
1496 struct device *dev = dpaa_bp->dev;
1497 struct bm_buffer bmb[8];
1502 for (i = 0; i < 8; i++) {
1503 new_buf = netdev_alloc_frag(dpaa_bp->raw_size);
1504 if (unlikely(!new_buf)) {
1505 dev_err(dev, "netdev_alloc_frag() failed, size %zu\n",
1507 goto release_previous_buffs;
1509 new_buf = PTR_ALIGN(new_buf, SMP_CACHE_BYTES);
1511 addr = dma_map_single(dev, new_buf,
1512 dpaa_bp->size, DMA_FROM_DEVICE);
1513 if (unlikely(dma_mapping_error(dev, addr))) {
1514 dev_err(dpaa_bp->dev, "DMA map failed");
1515 goto release_previous_buffs;
1519 bm_buffer_set64(&bmb[i], addr);
1523 return dpaa_bman_release(dpaa_bp, bmb, i);
1525 release_previous_buffs:
1526 WARN_ONCE(1, "dpaa_eth: failed to add buffers on Rx\n");
1528 bm_buffer_set64(&bmb[i], 0);
1529 /* Avoid releasing a completely null buffer; bman_release() requires
1530 * at least one buffer.
1538 static int dpaa_bp_seed(struct dpaa_bp *dpaa_bp)
1542 /* Give each CPU an allotment of "config_count" buffers */
1543 for_each_possible_cpu(i) {
1544 int *count_ptr = per_cpu_ptr(dpaa_bp->percpu_count, i);
1547 /* Although we access another CPU's counters here
1548 * we do it at boot time so it is safe
1550 for (j = 0; j < dpaa_bp->config_count; j += 8)
1551 *count_ptr += dpaa_bp_add_8_bufs(dpaa_bp);
1556 /* Add buffers/(pages) for Rx processing whenever bpool count falls below
1559 static int dpaa_eth_refill_bpool(struct dpaa_bp *dpaa_bp, int *countptr)
1561 int count = *countptr;
1564 if (unlikely(count < FSL_DPAA_ETH_REFILL_THRESHOLD)) {
1566 new_bufs = dpaa_bp_add_8_bufs(dpaa_bp);
1567 if (unlikely(!new_bufs)) {
1568 /* Avoid looping forever if we've temporarily
1569 * run out of memory. We'll try again at the
1575 } while (count < FSL_DPAA_ETH_MAX_BUF_COUNT);
1578 if (unlikely(count < FSL_DPAA_ETH_MAX_BUF_COUNT))
1585 static int dpaa_eth_refill_bpools(struct dpaa_priv *priv)
1587 struct dpaa_bp *dpaa_bp;
1591 for (i = 0; i < DPAA_BPS_NUM; i++) {
1592 dpaa_bp = priv->dpaa_bps[i];
1595 countptr = this_cpu_ptr(dpaa_bp->percpu_count);
1596 res = dpaa_eth_refill_bpool(dpaa_bp, countptr);
1603 /* Cleanup function for outgoing frame descriptors that were built on Tx path,
1604 * either contiguous frames or scatter/gather ones.
1605 * Skb freeing is not handled here.
1607 * This function may be called on error paths in the Tx function, so guard
1608 * against cases when not all fd relevant fields were filled in.
1610 * Return the skb backpointer, since for S/G frames the buffer containing it
1613 static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
1614 const struct qm_fd *fd)
1616 const enum dma_data_direction dma_dir = DMA_TO_DEVICE;
1617 struct device *dev = priv->net_dev->dev.parent;
1618 dma_addr_t addr = qm_fd_addr(fd);
1619 const struct qm_sg_entry *sgt;
1620 struct sk_buff **skbh, *skb;
1623 skbh = (struct sk_buff **)phys_to_virt(addr);
1626 if (unlikely(qm_fd_get_format(fd) == qm_fd_sg)) {
1627 nr_frags = skb_shinfo(skb)->nr_frags;
1628 dma_unmap_single(dev, addr,
1629 qm_fd_get_offset(fd) + DPAA_SGT_SIZE,
1632 /* The sgt buffer has been allocated with netdev_alloc_frag(),
1635 sgt = phys_to_virt(addr + qm_fd_get_offset(fd));
1637 /* sgt[0] is from lowmem, was dma_map_single()-ed */
1638 dma_unmap_single(dev, qm_sg_addr(&sgt[0]),
1639 qm_sg_entry_get_len(&sgt[0]), dma_dir);
1641 /* remaining pages were mapped with skb_frag_dma_map() */
1642 for (i = 1; i <= nr_frags; i++) {
1643 WARN_ON(qm_sg_entry_is_ext(&sgt[i]));
1645 dma_unmap_page(dev, qm_sg_addr(&sgt[i]),
1646 qm_sg_entry_get_len(&sgt[i]), dma_dir);
1649 /* Free the page frag that we allocated on Tx */
1650 skb_free_frag(phys_to_virt(addr));
1652 dma_unmap_single(dev, addr,
1653 skb_tail_pointer(skb) - (u8 *)skbh, dma_dir);
1659 static u8 rx_csum_offload(const struct dpaa_priv *priv, const struct qm_fd *fd)
1661 /* The parser has run and performed L4 checksum validation.
1662 * We know there were no parser errors (and implicitly no
1663 * L4 csum error), otherwise we wouldn't be here.
1665 if ((priv->net_dev->features & NETIF_F_RXCSUM) &&
1666 (be32_to_cpu(fd->status) & FM_FD_STAT_L4CV))
1667 return CHECKSUM_UNNECESSARY;
1669 /* We're here because either the parser didn't run or the L4 checksum
1670 * was not verified. This may include the case of a UDP frame with
1671 * checksum zero or an L4 proto other than TCP/UDP
1673 return CHECKSUM_NONE;
1676 /* Build a linear skb around the received buffer.
1677 * We are guaranteed there is enough room at the end of the data buffer to
1678 * accommodate the shared info area of the skb.
1680 static struct sk_buff *contig_fd_to_skb(const struct dpaa_priv *priv,
1681 const struct qm_fd *fd)
1683 ssize_t fd_off = qm_fd_get_offset(fd);
1684 dma_addr_t addr = qm_fd_addr(fd);
1685 struct dpaa_bp *dpaa_bp;
1686 struct sk_buff *skb;
1689 vaddr = phys_to_virt(addr);
1690 WARN_ON(!IS_ALIGNED((unsigned long)vaddr, SMP_CACHE_BYTES));
1692 dpaa_bp = dpaa_bpid2pool(fd->bpid);
1696 skb = build_skb(vaddr, dpaa_bp->size +
1697 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
1698 if (unlikely(!skb)) {
1699 WARN_ONCE(1, "Build skb failure on Rx\n");
1702 WARN_ON(fd_off != priv->rx_headroom);
1703 skb_reserve(skb, fd_off);
1704 skb_put(skb, qm_fd_get_length(fd));
1706 skb->ip_summed = rx_csum_offload(priv, fd);
1711 skb_free_frag(vaddr);
1715 /* Build an skb with the data of the first S/G entry in the linear portion and
1716 * the rest of the frame as skb fragments.
1718 * The page fragment holding the S/G Table is recycled here.
1720 static struct sk_buff *sg_fd_to_skb(const struct dpaa_priv *priv,
1721 const struct qm_fd *fd)
1723 ssize_t fd_off = qm_fd_get_offset(fd);
1724 dma_addr_t addr = qm_fd_addr(fd);
1725 const struct qm_sg_entry *sgt;
1726 struct page *page, *head_page;
1727 struct dpaa_bp *dpaa_bp;
1728 void *vaddr, *sg_vaddr;
1729 int frag_off, frag_len;
1730 struct sk_buff *skb;
1737 vaddr = phys_to_virt(addr);
1738 WARN_ON(!IS_ALIGNED((unsigned long)vaddr, SMP_CACHE_BYTES));
1740 /* Iterate through the SGT entries and add data buffers to the skb */
1741 sgt = vaddr + fd_off;
1742 for (i = 0; i < DPAA_SGT_MAX_ENTRIES; i++) {
1743 /* Extension bit is not supported */
1744 WARN_ON(qm_sg_entry_is_ext(&sgt[i]));
1746 sg_addr = qm_sg_addr(&sgt[i]);
1747 sg_vaddr = phys_to_virt(sg_addr);
1748 WARN_ON(!IS_ALIGNED((unsigned long)sg_vaddr,
1751 /* We may use multiple Rx pools */
1752 dpaa_bp = dpaa_bpid2pool(sgt[i].bpid);
1756 count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
1757 dma_unmap_single(dpaa_bp->dev, sg_addr, dpaa_bp->size,
1760 sz = dpaa_bp->size +
1761 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1762 skb = build_skb(sg_vaddr, sz);
1763 if (WARN_ON(unlikely(!skb)))
1766 skb->ip_summed = rx_csum_offload(priv, fd);
1768 /* Make sure forwarded skbs will have enough space
1769 * on Tx, if extra headers are added.
1771 WARN_ON(fd_off != priv->rx_headroom);
1772 skb_reserve(skb, fd_off);
1773 skb_put(skb, qm_sg_entry_get_len(&sgt[i]));
1775 /* Not the first S/G entry; all data from buffer will
1776 * be added in an skb fragment; fragment index is offset
1777 * by one since first S/G entry was incorporated in the
1778 * linear part of the skb.
1780 * Caution: 'page' may be a tail page.
1782 page = virt_to_page(sg_vaddr);
1783 head_page = virt_to_head_page(sg_vaddr);
1785 /* Compute offset in (possibly tail) page */
1786 page_offset = ((unsigned long)sg_vaddr &
1788 (page_address(page) - page_address(head_page));
1789 /* page_offset only refers to the beginning of sgt[i];
1790 * but the buffer itself may have an internal offset.
1792 frag_off = qm_sg_entry_get_off(&sgt[i]) + page_offset;
1793 frag_len = qm_sg_entry_get_len(&sgt[i]);
1794 /* skb_add_rx_frag() does no checking on the page; if
1795 * we pass it a tail page, we'll end up with
1796 * bad page accounting and eventually with segafults.
1798 skb_add_rx_frag(skb, i - 1, head_page, frag_off,
1799 frag_len, dpaa_bp->size);
1801 /* Update the pool count for the current {cpu x bpool} */
1804 if (qm_sg_entry_is_final(&sgt[i]))
1807 WARN_ONCE(i == DPAA_SGT_MAX_ENTRIES, "No final bit on SGT\n");
1809 /* free the SG table buffer */
1810 skb_free_frag(vaddr);
1815 /* compensate sw bpool counter changes */
1816 for (i--; i >= 0; i--) {
1817 dpaa_bp = dpaa_bpid2pool(sgt[i].bpid);
1819 count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
1823 /* free all the SG entries */
1824 for (i = 0; i < DPAA_SGT_MAX_ENTRIES ; i++) {
1825 sg_addr = qm_sg_addr(&sgt[i]);
1826 sg_vaddr = phys_to_virt(sg_addr);
1827 skb_free_frag(sg_vaddr);
1828 dpaa_bp = dpaa_bpid2pool(sgt[i].bpid);
1830 count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
1834 if (qm_sg_entry_is_final(&sgt[i]))
1837 /* free the SGT fragment */
1838 skb_free_frag(vaddr);
1843 static int skb_to_contig_fd(struct dpaa_priv *priv,
1844 struct sk_buff *skb, struct qm_fd *fd,
1847 struct net_device *net_dev = priv->net_dev;
1848 struct device *dev = net_dev->dev.parent;
1849 enum dma_data_direction dma_dir;
1850 unsigned char *buffer_start;
1851 struct sk_buff **skbh;
1855 /* We are guaranteed to have at least tx_headroom bytes
1856 * available, so just use that for offset.
1858 fd->bpid = FSL_DPAA_BPID_INV;
1859 buffer_start = skb->data - priv->tx_headroom;
1860 dma_dir = DMA_TO_DEVICE;
1862 skbh = (struct sk_buff **)buffer_start;
1865 /* Enable L3/L4 hardware checksum computation.
1867 * We must do this before dma_map_single(DMA_TO_DEVICE), because we may
1868 * need to write into the skb.
1870 err = dpaa_enable_tx_csum(priv, skb, fd,
1871 ((char *)skbh) + DPAA_TX_PRIV_DATA_SIZE);
1872 if (unlikely(err < 0)) {
1873 if (net_ratelimit())
1874 netif_err(priv, tx_err, net_dev, "HW csum error: %d\n",
1879 /* Fill in the rest of the FD fields */
1880 qm_fd_set_contig(fd, priv->tx_headroom, skb->len);
1881 fd->cmd |= cpu_to_be32(FM_FD_CMD_FCO);
1883 /* Map the entire buffer size that may be seen by FMan, but no more */
1884 addr = dma_map_single(dev, skbh,
1885 skb_tail_pointer(skb) - buffer_start, dma_dir);
1886 if (unlikely(dma_mapping_error(dev, addr))) {
1887 if (net_ratelimit())
1888 netif_err(priv, tx_err, net_dev, "dma_map_single() failed\n");
1891 qm_fd_addr_set64(fd, addr);
1896 static int skb_to_sg_fd(struct dpaa_priv *priv,
1897 struct sk_buff *skb, struct qm_fd *fd)
1899 const enum dma_data_direction dma_dir = DMA_TO_DEVICE;
1900 const int nr_frags = skb_shinfo(skb)->nr_frags;
1901 struct net_device *net_dev = priv->net_dev;
1902 struct device *dev = net_dev->dev.parent;
1903 struct qm_sg_entry *sgt;
1904 struct sk_buff **skbh;
1912 /* get a page frag to store the SGTable */
1913 sz = SKB_DATA_ALIGN(priv->tx_headroom + DPAA_SGT_SIZE);
1914 sgt_buf = netdev_alloc_frag(sz);
1915 if (unlikely(!sgt_buf)) {
1916 netdev_err(net_dev, "netdev_alloc_frag() failed for size %d\n",
1921 /* Enable L3/L4 hardware checksum computation.
1923 * We must do this before dma_map_single(DMA_TO_DEVICE), because we may
1924 * need to write into the skb.
1926 err = dpaa_enable_tx_csum(priv, skb, fd,
1927 sgt_buf + DPAA_TX_PRIV_DATA_SIZE);
1928 if (unlikely(err < 0)) {
1929 if (net_ratelimit())
1930 netif_err(priv, tx_err, net_dev, "HW csum error: %d\n",
1935 /* SGT[0] is used by the linear part */
1936 sgt = (struct qm_sg_entry *)(sgt_buf + priv->tx_headroom);
1937 frag_len = skb_headlen(skb);
1938 qm_sg_entry_set_len(&sgt[0], frag_len);
1939 sgt[0].bpid = FSL_DPAA_BPID_INV;
1941 addr = dma_map_single(dev, skb->data,
1942 skb_headlen(skb), dma_dir);
1943 if (unlikely(dma_mapping_error(dev, addr))) {
1944 dev_err(dev, "DMA mapping failed");
1946 goto sg0_map_failed;
1948 qm_sg_entry_set64(&sgt[0], addr);
1950 /* populate the rest of SGT entries */
1951 for (i = 0; i < nr_frags; i++) {
1952 frag = &skb_shinfo(skb)->frags[i];
1953 frag_len = frag->size;
1954 WARN_ON(!skb_frag_page(frag));
1955 addr = skb_frag_dma_map(dev, frag, 0,
1957 if (unlikely(dma_mapping_error(dev, addr))) {
1958 dev_err(dev, "DMA mapping failed");
1963 qm_sg_entry_set_len(&sgt[i + 1], frag_len);
1964 sgt[i + 1].bpid = FSL_DPAA_BPID_INV;
1965 sgt[i + 1].offset = 0;
1967 /* keep the offset in the address */
1968 qm_sg_entry_set64(&sgt[i + 1], addr);
1971 /* Set the final bit in the last used entry of the SGT */
1972 qm_sg_entry_set_f(&sgt[nr_frags], frag_len);
1974 qm_fd_set_sg(fd, priv->tx_headroom, skb->len);
1976 /* DMA map the SGT page */
1977 buffer_start = (void *)sgt - priv->tx_headroom;
1978 skbh = (struct sk_buff **)buffer_start;
1981 addr = dma_map_single(dev, buffer_start,
1982 priv->tx_headroom + DPAA_SGT_SIZE, dma_dir);
1983 if (unlikely(dma_mapping_error(dev, addr))) {
1984 dev_err(dev, "DMA mapping failed");
1986 goto sgt_map_failed;
1989 fd->bpid = FSL_DPAA_BPID_INV;
1990 fd->cmd |= cpu_to_be32(FM_FD_CMD_FCO);
1991 qm_fd_addr_set64(fd, addr);
1997 for (j = 0; j < i; j++)
1998 dma_unmap_page(dev, qm_sg_addr(&sgt[j]),
1999 qm_sg_entry_get_len(&sgt[j]), dma_dir);
2002 skb_free_frag(sgt_buf);
2007 static inline int dpaa_xmit(struct dpaa_priv *priv,
2008 struct rtnl_link_stats64 *percpu_stats,
2012 struct qman_fq *egress_fq;
2015 egress_fq = priv->egress_fqs[queue];
2016 if (fd->bpid == FSL_DPAA_BPID_INV)
2017 fd->cmd |= cpu_to_be32(qman_fq_fqid(priv->conf_fqs[queue]));
2019 /* Trace this Tx fd */
2020 trace_dpaa_tx_fd(priv->net_dev, egress_fq, fd);
2022 for (i = 0; i < DPAA_ENQUEUE_RETRIES; i++) {
2023 err = qman_enqueue(egress_fq, fd);
2028 if (unlikely(err < 0)) {
2029 percpu_stats->tx_fifo_errors++;
2033 percpu_stats->tx_packets++;
2034 percpu_stats->tx_bytes += qm_fd_get_length(fd);
2040 dpaa_start_xmit(struct sk_buff *skb, struct net_device *net_dev)
2042 const int queue_mapping = skb_get_queue_mapping(skb);
2043 bool nonlinear = skb_is_nonlinear(skb);
2044 struct rtnl_link_stats64 *percpu_stats;
2045 struct dpaa_percpu_priv *percpu_priv;
2046 struct dpaa_priv *priv;
2051 priv = netdev_priv(net_dev);
2052 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2053 percpu_stats = &percpu_priv->stats;
2055 qm_fd_clear_fd(&fd);
2058 /* We're going to store the skb backpointer at the beginning
2059 * of the data buffer, so we need a privately owned skb
2061 * We've made sure skb is not shared in dev->priv_flags,
2062 * we need to verify the skb head is not cloned
2064 if (skb_cow_head(skb, priv->tx_headroom))
2067 WARN_ON(skb_is_nonlinear(skb));
2070 /* MAX_SKB_FRAGS is equal or larger than our dpaa_SGT_MAX_ENTRIES;
2071 * make sure we don't feed FMan with more fragments than it supports.
2074 likely(skb_shinfo(skb)->nr_frags < DPAA_SGT_MAX_ENTRIES)) {
2075 /* Just create a S/G fd based on the skb */
2076 err = skb_to_sg_fd(priv, skb, &fd);
2077 percpu_priv->tx_frag_skbuffs++;
2079 /* If the egress skb contains more fragments than we support
2080 * we have no choice but to linearize it ourselves.
2082 if (unlikely(nonlinear) && __skb_linearize(skb))
2085 /* Finally, create a contig FD from this skb */
2086 err = skb_to_contig_fd(priv, skb, &fd, &offset);
2088 if (unlikely(err < 0))
2089 goto skb_to_fd_failed;
2091 if (likely(dpaa_xmit(priv, percpu_stats, queue_mapping, &fd) == 0))
2092 return NETDEV_TX_OK;
2094 dpaa_cleanup_tx_fd(priv, &fd);
2097 percpu_stats->tx_errors++;
2099 return NETDEV_TX_OK;
2102 static void dpaa_rx_error(struct net_device *net_dev,
2103 const struct dpaa_priv *priv,
2104 struct dpaa_percpu_priv *percpu_priv,
2105 const struct qm_fd *fd,
2108 if (net_ratelimit())
2109 netif_err(priv, hw, net_dev, "Err FD status = 0x%08x\n",
2110 be32_to_cpu(fd->status) & FM_FD_STAT_RX_ERRORS);
2112 percpu_priv->stats.rx_errors++;
2114 if (be32_to_cpu(fd->status) & FM_FD_ERR_DMA)
2115 percpu_priv->rx_errors.dme++;
2116 if (be32_to_cpu(fd->status) & FM_FD_ERR_PHYSICAL)
2117 percpu_priv->rx_errors.fpe++;
2118 if (be32_to_cpu(fd->status) & FM_FD_ERR_SIZE)
2119 percpu_priv->rx_errors.fse++;
2120 if (be32_to_cpu(fd->status) & FM_FD_ERR_PRS_HDR_ERR)
2121 percpu_priv->rx_errors.phe++;
2123 dpaa_fd_release(net_dev, fd);
2126 static void dpaa_tx_error(struct net_device *net_dev,
2127 const struct dpaa_priv *priv,
2128 struct dpaa_percpu_priv *percpu_priv,
2129 const struct qm_fd *fd,
2132 struct sk_buff *skb;
2134 if (net_ratelimit())
2135 netif_warn(priv, hw, net_dev, "FD status = 0x%08x\n",
2136 be32_to_cpu(fd->status) & FM_FD_STAT_TX_ERRORS);
2138 percpu_priv->stats.tx_errors++;
2140 skb = dpaa_cleanup_tx_fd(priv, fd);
2144 static int dpaa_eth_poll(struct napi_struct *napi, int budget)
2146 struct dpaa_napi_portal *np =
2147 container_of(napi, struct dpaa_napi_portal, napi);
2149 int cleaned = qman_p_poll_dqrr(np->p, budget);
2151 if (cleaned < budget) {
2152 napi_complete_done(napi, cleaned);
2153 qman_p_irqsource_add(np->p, QM_PIRQ_DQRI);
2155 } else if (np->down) {
2156 qman_p_irqsource_add(np->p, QM_PIRQ_DQRI);
2162 static void dpaa_tx_conf(struct net_device *net_dev,
2163 const struct dpaa_priv *priv,
2164 struct dpaa_percpu_priv *percpu_priv,
2165 const struct qm_fd *fd,
2168 struct sk_buff *skb;
2170 if (unlikely(be32_to_cpu(fd->status) & FM_FD_STAT_TX_ERRORS)) {
2171 if (net_ratelimit())
2172 netif_warn(priv, hw, net_dev, "FD status = 0x%08x\n",
2173 be32_to_cpu(fd->status) &
2174 FM_FD_STAT_TX_ERRORS);
2176 percpu_priv->stats.tx_errors++;
2179 percpu_priv->tx_confirm++;
2181 skb = dpaa_cleanup_tx_fd(priv, fd);
2186 static inline int dpaa_eth_napi_schedule(struct dpaa_percpu_priv *percpu_priv,
2187 struct qman_portal *portal)
2189 if (unlikely(in_irq() || !in_serving_softirq())) {
2190 /* Disable QMan IRQ and invoke NAPI */
2191 qman_p_irqsource_remove(portal, QM_PIRQ_DQRI);
2193 percpu_priv->np.p = portal;
2194 napi_schedule(&percpu_priv->np.napi);
2195 percpu_priv->in_interrupt++;
2201 static enum qman_cb_dqrr_result rx_error_dqrr(struct qman_portal *portal,
2203 const struct qm_dqrr_entry *dq)
2205 struct dpaa_fq *dpaa_fq = container_of(fq, struct dpaa_fq, fq_base);
2206 struct dpaa_percpu_priv *percpu_priv;
2207 struct net_device *net_dev;
2208 struct dpaa_bp *dpaa_bp;
2209 struct dpaa_priv *priv;
2211 net_dev = dpaa_fq->net_dev;
2212 priv = netdev_priv(net_dev);
2213 dpaa_bp = dpaa_bpid2pool(dq->fd.bpid);
2215 return qman_cb_dqrr_consume;
2217 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2219 if (dpaa_eth_napi_schedule(percpu_priv, portal))
2220 return qman_cb_dqrr_stop;
2222 if (dpaa_eth_refill_bpools(priv))
2223 /* Unable to refill the buffer pool due to insufficient
2224 * system memory. Just release the frame back into the pool,
2225 * otherwise we'll soon end up with an empty buffer pool.
2227 dpaa_fd_release(net_dev, &dq->fd);
2229 dpaa_rx_error(net_dev, priv, percpu_priv, &dq->fd, fq->fqid);
2231 return qman_cb_dqrr_consume;
2234 static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
2236 const struct qm_dqrr_entry *dq)
2238 struct rtnl_link_stats64 *percpu_stats;
2239 struct dpaa_percpu_priv *percpu_priv;
2240 const struct qm_fd *fd = &dq->fd;
2241 dma_addr_t addr = qm_fd_addr(fd);
2242 enum qm_fd_format fd_format;
2243 struct net_device *net_dev;
2244 u32 fd_status, hash_offset;
2245 struct dpaa_bp *dpaa_bp;
2246 struct dpaa_priv *priv;
2247 unsigned int skb_len;
2248 struct sk_buff *skb;
2252 fd_status = be32_to_cpu(fd->status);
2253 fd_format = qm_fd_get_format(fd);
2254 net_dev = ((struct dpaa_fq *)fq)->net_dev;
2255 priv = netdev_priv(net_dev);
2256 dpaa_bp = dpaa_bpid2pool(dq->fd.bpid);
2258 return qman_cb_dqrr_consume;
2260 /* Trace the Rx fd */
2261 trace_dpaa_rx_fd(net_dev, fq, &dq->fd);
2263 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2264 percpu_stats = &percpu_priv->stats;
2266 if (unlikely(dpaa_eth_napi_schedule(percpu_priv, portal)))
2267 return qman_cb_dqrr_stop;
2269 /* Make sure we didn't run out of buffers */
2270 if (unlikely(dpaa_eth_refill_bpools(priv))) {
2271 /* Unable to refill the buffer pool due to insufficient
2272 * system memory. Just release the frame back into the pool,
2273 * otherwise we'll soon end up with an empty buffer pool.
2275 dpaa_fd_release(net_dev, &dq->fd);
2276 return qman_cb_dqrr_consume;
2279 if (unlikely(fd_status & FM_FD_STAT_RX_ERRORS) != 0) {
2280 if (net_ratelimit())
2281 netif_warn(priv, hw, net_dev, "FD status = 0x%08x\n",
2282 fd_status & FM_FD_STAT_RX_ERRORS);
2284 percpu_stats->rx_errors++;
2285 dpaa_fd_release(net_dev, fd);
2286 return qman_cb_dqrr_consume;
2289 dpaa_bp = dpaa_bpid2pool(fd->bpid);
2291 return qman_cb_dqrr_consume;
2293 dma_unmap_single(dpaa_bp->dev, addr, dpaa_bp->size, DMA_FROM_DEVICE);
2295 /* prefetch the first 64 bytes of the frame or the SGT start */
2296 vaddr = phys_to_virt(addr);
2297 prefetch(vaddr + qm_fd_get_offset(fd));
2299 /* The only FD types that we may receive are contig and S/G */
2300 WARN_ON((fd_format != qm_fd_contig) && (fd_format != qm_fd_sg));
2302 /* Account for either the contig buffer or the SGT buffer (depending on
2303 * which case we were in) having been removed from the pool.
2305 count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
2308 if (likely(fd_format == qm_fd_contig))
2309 skb = contig_fd_to_skb(priv, fd);
2311 skb = sg_fd_to_skb(priv, fd);
2313 return qman_cb_dqrr_consume;
2315 skb->protocol = eth_type_trans(skb, net_dev);
2317 if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use &&
2318 !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
2320 enum pkt_hash_types type;
2322 /* if L4 exists, it was used in the hash generation */
2323 type = be32_to_cpu(fd->status) & FM_FD_STAT_L4CV ?
2324 PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3;
2325 skb_set_hash(skb, be32_to_cpu(*(u32 *)(vaddr + hash_offset)),
2331 if (unlikely(netif_receive_skb(skb) == NET_RX_DROP)) {
2332 percpu_stats->rx_dropped++;
2333 return qman_cb_dqrr_consume;
2336 percpu_stats->rx_packets++;
2337 percpu_stats->rx_bytes += skb_len;
2339 return qman_cb_dqrr_consume;
2342 static enum qman_cb_dqrr_result conf_error_dqrr(struct qman_portal *portal,
2344 const struct qm_dqrr_entry *dq)
2346 struct dpaa_percpu_priv *percpu_priv;
2347 struct net_device *net_dev;
2348 struct dpaa_priv *priv;
2350 net_dev = ((struct dpaa_fq *)fq)->net_dev;
2351 priv = netdev_priv(net_dev);
2353 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2355 if (dpaa_eth_napi_schedule(percpu_priv, portal))
2356 return qman_cb_dqrr_stop;
2358 dpaa_tx_error(net_dev, priv, percpu_priv, &dq->fd, fq->fqid);
2360 return qman_cb_dqrr_consume;
2363 static enum qman_cb_dqrr_result conf_dflt_dqrr(struct qman_portal *portal,
2365 const struct qm_dqrr_entry *dq)
2367 struct dpaa_percpu_priv *percpu_priv;
2368 struct net_device *net_dev;
2369 struct dpaa_priv *priv;
2371 net_dev = ((struct dpaa_fq *)fq)->net_dev;
2372 priv = netdev_priv(net_dev);
2375 trace_dpaa_tx_conf_fd(net_dev, fq, &dq->fd);
2377 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2379 if (dpaa_eth_napi_schedule(percpu_priv, portal))
2380 return qman_cb_dqrr_stop;
2382 dpaa_tx_conf(net_dev, priv, percpu_priv, &dq->fd, fq->fqid);
2384 return qman_cb_dqrr_consume;
2387 static void egress_ern(struct qman_portal *portal,
2389 const union qm_mr_entry *msg)
2391 const struct qm_fd *fd = &msg->ern.fd;
2392 struct dpaa_percpu_priv *percpu_priv;
2393 const struct dpaa_priv *priv;
2394 struct net_device *net_dev;
2395 struct sk_buff *skb;
2397 net_dev = ((struct dpaa_fq *)fq)->net_dev;
2398 priv = netdev_priv(net_dev);
2399 percpu_priv = this_cpu_ptr(priv->percpu_priv);
2401 percpu_priv->stats.tx_dropped++;
2402 percpu_priv->stats.tx_fifo_errors++;
2403 count_ern(percpu_priv, msg);
2405 skb = dpaa_cleanup_tx_fd(priv, fd);
2406 dev_kfree_skb_any(skb);
2409 static const struct dpaa_fq_cbs dpaa_fq_cbs = {
2410 .rx_defq = { .cb = { .dqrr = rx_default_dqrr } },
2411 .tx_defq = { .cb = { .dqrr = conf_dflt_dqrr } },
2412 .rx_errq = { .cb = { .dqrr = rx_error_dqrr } },
2413 .tx_errq = { .cb = { .dqrr = conf_error_dqrr } },
2414 .egress_ern = { .cb = { .ern = egress_ern } }
2417 static void dpaa_eth_napi_enable(struct dpaa_priv *priv)
2419 struct dpaa_percpu_priv *percpu_priv;
2422 for_each_possible_cpu(i) {
2423 percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
2425 percpu_priv->np.down = 0;
2426 napi_enable(&percpu_priv->np.napi);
2430 static void dpaa_eth_napi_disable(struct dpaa_priv *priv)
2432 struct dpaa_percpu_priv *percpu_priv;
2435 for_each_possible_cpu(i) {
2436 percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
2438 percpu_priv->np.down = 1;
2439 napi_disable(&percpu_priv->np.napi);
2443 static int dpaa_open(struct net_device *net_dev)
2445 struct mac_device *mac_dev;
2446 struct dpaa_priv *priv;
2449 priv = netdev_priv(net_dev);
2450 mac_dev = priv->mac_dev;
2451 dpaa_eth_napi_enable(priv);
2453 net_dev->phydev = mac_dev->init_phy(net_dev, priv->mac_dev);
2454 if (!net_dev->phydev) {
2455 netif_err(priv, ifup, net_dev, "init_phy() failed\n");
2457 goto phy_init_failed;
2460 for (i = 0; i < ARRAY_SIZE(mac_dev->port); i++) {
2461 err = fman_port_enable(mac_dev->port[i]);
2463 goto mac_start_failed;
2466 err = priv->mac_dev->start(mac_dev);
2468 netif_err(priv, ifup, net_dev, "mac_dev->start() = %d\n", err);
2469 goto mac_start_failed;
2472 netif_tx_start_all_queues(net_dev);
2477 for (i = 0; i < ARRAY_SIZE(mac_dev->port); i++)
2478 fman_port_disable(mac_dev->port[i]);
2481 dpaa_eth_napi_disable(priv);
2486 static int dpaa_eth_stop(struct net_device *net_dev)
2488 struct dpaa_priv *priv;
2491 err = dpaa_stop(net_dev);
2493 priv = netdev_priv(net_dev);
2494 dpaa_eth_napi_disable(priv);
2499 static int dpaa_ioctl(struct net_device *net_dev, struct ifreq *rq, int cmd)
2501 if (!net_dev->phydev)
2503 return phy_mii_ioctl(net_dev->phydev, rq, cmd);
2506 static const struct net_device_ops dpaa_ops = {
2507 .ndo_open = dpaa_open,
2508 .ndo_start_xmit = dpaa_start_xmit,
2509 .ndo_stop = dpaa_eth_stop,
2510 .ndo_tx_timeout = dpaa_tx_timeout,
2511 .ndo_get_stats64 = dpaa_get_stats64,
2512 .ndo_set_mac_address = dpaa_set_mac_address,
2513 .ndo_validate_addr = eth_validate_addr,
2514 .ndo_set_rx_mode = dpaa_set_rx_mode,
2515 .ndo_do_ioctl = dpaa_ioctl,
2516 .ndo_setup_tc = dpaa_setup_tc,
2519 static int dpaa_napi_add(struct net_device *net_dev)
2521 struct dpaa_priv *priv = netdev_priv(net_dev);
2522 struct dpaa_percpu_priv *percpu_priv;
2525 for_each_possible_cpu(cpu) {
2526 percpu_priv = per_cpu_ptr(priv->percpu_priv, cpu);
2528 netif_napi_add(net_dev, &percpu_priv->np.napi,
2529 dpaa_eth_poll, NAPI_POLL_WEIGHT);
2535 static void dpaa_napi_del(struct net_device *net_dev)
2537 struct dpaa_priv *priv = netdev_priv(net_dev);
2538 struct dpaa_percpu_priv *percpu_priv;
2541 for_each_possible_cpu(cpu) {
2542 percpu_priv = per_cpu_ptr(priv->percpu_priv, cpu);
2544 netif_napi_del(&percpu_priv->np.napi);
2548 static inline void dpaa_bp_free_pf(const struct dpaa_bp *bp,
2549 struct bm_buffer *bmb)
2551 dma_addr_t addr = bm_buf_addr(bmb);
2553 dma_unmap_single(bp->dev, addr, bp->size, DMA_FROM_DEVICE);
2555 skb_free_frag(phys_to_virt(addr));
2558 /* Alloc the dpaa_bp struct and configure default values */
2559 static struct dpaa_bp *dpaa_bp_alloc(struct device *dev)
2561 struct dpaa_bp *dpaa_bp;
2563 dpaa_bp = devm_kzalloc(dev, sizeof(*dpaa_bp), GFP_KERNEL);
2565 return ERR_PTR(-ENOMEM);
2567 dpaa_bp->bpid = FSL_DPAA_BPID_INV;
2568 dpaa_bp->percpu_count = devm_alloc_percpu(dev, *dpaa_bp->percpu_count);
2569 if (!dpaa_bp->percpu_count)
2570 return ERR_PTR(-ENOMEM);
2572 dpaa_bp->config_count = FSL_DPAA_ETH_MAX_BUF_COUNT;
2574 dpaa_bp->seed_cb = dpaa_bp_seed;
2575 dpaa_bp->free_buf_cb = dpaa_bp_free_pf;
2580 /* Place all ingress FQs (Rx Default, Rx Error) in a dedicated CGR.
2581 * We won't be sending congestion notifications to FMan; for now, we just use
2582 * this CGR to generate enqueue rejections to FMan in order to drop the frames
2583 * before they reach our ingress queues and eat up memory.
2585 static int dpaa_ingress_cgr_init(struct dpaa_priv *priv)
2587 struct qm_mcc_initcgr initcgr;
2591 err = qman_alloc_cgrid(&priv->ingress_cgr.cgrid);
2593 if (netif_msg_drv(priv))
2594 pr_err("Error %d allocating CGR ID\n", err);
2598 /* Enable CS TD, but disable Congestion State Change Notifications. */
2599 memset(&initcgr, 0, sizeof(initcgr));
2600 initcgr.we_mask = cpu_to_be16(QM_CGR_WE_CS_THRES);
2601 initcgr.cgr.cscn_en = QM_CGR_EN;
2602 cs_th = DPAA_INGRESS_CS_THRESHOLD;
2603 qm_cgr_cs_thres_set64(&initcgr.cgr.cs_thres, cs_th, 1);
2605 initcgr.we_mask |= cpu_to_be16(QM_CGR_WE_CSTD_EN);
2606 initcgr.cgr.cstd_en = QM_CGR_EN;
2608 /* This CGR will be associated with the SWP affined to the current CPU.
2609 * However, we'll place all our ingress FQs in it.
2611 err = qman_create_cgr(&priv->ingress_cgr, QMAN_CGR_FLAG_USE_INIT,
2614 if (netif_msg_drv(priv))
2615 pr_err("Error %d creating ingress CGR with ID %d\n",
2616 err, priv->ingress_cgr.cgrid);
2617 qman_release_cgrid(priv->ingress_cgr.cgrid);
2620 if (netif_msg_drv(priv))
2621 pr_debug("Created ingress CGR %d for netdev with hwaddr %pM\n",
2622 priv->ingress_cgr.cgrid, priv->mac_dev->addr);
2624 priv->use_ingress_cgr = true;
2630 static const struct of_device_id dpaa_match[];
2632 static inline u16 dpaa_get_headroom(struct dpaa_buffer_layout *bl)
2636 /* The frame headroom must accommodate:
2637 * - the driver private data area
2638 * - parse results, hash results, timestamp if selected
2639 * If either hash results or time stamp are selected, both will
2640 * be copied to/from the frame headroom, as TS is located between PR and
2641 * HR in the IC and IC copy size has a granularity of 16bytes
2642 * (see description of FMBM_RICP and FMBM_TICP registers in DPAARM)
2644 * Also make sure the headroom is a multiple of data_align bytes
2646 headroom = (u16)(bl->priv_data_size + DPAA_PARSE_RESULTS_SIZE +
2647 DPAA_TIME_STAMP_SIZE + DPAA_HASH_RESULTS_SIZE);
2649 return ALIGN(headroom, DPAA_FD_DATA_ALIGNMENT);
2652 static int dpaa_eth_probe(struct platform_device *pdev)
2654 struct dpaa_bp *dpaa_bps[DPAA_BPS_NUM] = {NULL};
2655 struct dpaa_percpu_priv *percpu_priv;
2656 struct net_device *net_dev = NULL;
2657 struct dpaa_fq *dpaa_fq, *tmp;
2658 struct dpaa_priv *priv = NULL;
2659 struct fm_port_fqs port_fqs;
2660 struct mac_device *mac_dev;
2661 int err = 0, i, channel;
2666 /* Allocate this early, so we can store relevant information in
2669 net_dev = alloc_etherdev_mq(sizeof(*priv), DPAA_ETH_TXQ_NUM);
2671 dev_err(dev, "alloc_etherdev_mq() failed\n");
2672 goto alloc_etherdev_mq_failed;
2675 /* Do this here, so we can be verbose early */
2676 SET_NETDEV_DEV(net_dev, dev);
2677 dev_set_drvdata(dev, net_dev);
2679 priv = netdev_priv(net_dev);
2680 priv->net_dev = net_dev;
2682 priv->msg_enable = netif_msg_init(debug, DPAA_MSG_DEFAULT);
2684 mac_dev = dpaa_mac_dev_get(pdev);
2685 if (IS_ERR(mac_dev)) {
2686 dev_err(dev, "dpaa_mac_dev_get() failed\n");
2687 err = PTR_ERR(mac_dev);
2688 goto mac_probe_failed;
2691 /* If fsl_fm_max_frm is set to a higher value than the all-common 1500,
2692 * we choose conservatively and let the user explicitly set a higher
2693 * MTU via ifconfig. Otherwise, the user may end up with different MTUs
2695 * If on the other hand fsl_fm_max_frm has been chosen below 1500,
2696 * start with the maximum allowed.
2698 net_dev->mtu = min(dpaa_get_max_mtu(), ETH_DATA_LEN);
2700 netdev_dbg(net_dev, "Setting initial MTU on net device: %d\n",
2703 priv->buf_layout[RX].priv_data_size = DPAA_RX_PRIV_DATA_SIZE; /* Rx */
2704 priv->buf_layout[TX].priv_data_size = DPAA_TX_PRIV_DATA_SIZE; /* Tx */
2706 /* device used for DMA mapping */
2707 set_dma_ops(dev, get_dma_ops(&pdev->dev));
2708 err = dma_coerce_mask_and_coherent(dev, DMA_BIT_MASK(40));
2710 dev_err(dev, "dma_coerce_mask_and_coherent() failed\n");
2711 goto dev_mask_failed;
2715 for (i = 0; i < DPAA_BPS_NUM; i++) {
2718 dpaa_bps[i] = dpaa_bp_alloc(dev);
2719 if (IS_ERR(dpaa_bps[i]))
2720 return PTR_ERR(dpaa_bps[i]);
2721 /* the raw size of the buffers used for reception */
2722 dpaa_bps[i]->raw_size = bpool_buffer_raw_size(i, DPAA_BPS_NUM);
2723 /* avoid runtime computations by keeping the usable size here */
2724 dpaa_bps[i]->size = dpaa_bp_size(dpaa_bps[i]->raw_size);
2725 dpaa_bps[i]->dev = dev;
2727 err = dpaa_bp_alloc_pool(dpaa_bps[i]);
2729 dpaa_bps_free(priv);
2730 priv->dpaa_bps[i] = NULL;
2731 goto bp_create_failed;
2733 priv->dpaa_bps[i] = dpaa_bps[i];
2736 INIT_LIST_HEAD(&priv->dpaa_fq_list);
2738 memset(&port_fqs, 0, sizeof(port_fqs));
2740 err = dpaa_alloc_all_fqs(dev, &priv->dpaa_fq_list, &port_fqs);
2742 dev_err(dev, "dpaa_alloc_all_fqs() failed\n");
2743 goto fq_probe_failed;
2746 priv->mac_dev = mac_dev;
2748 channel = dpaa_get_channel();
2750 dev_err(dev, "dpaa_get_channel() failed\n");
2752 goto get_channel_failed;
2755 priv->channel = (u16)channel;
2757 /* Start a thread that will walk the CPUs with affine portals
2758 * and add this pool channel to each's dequeue mask.
2760 dpaa_eth_add_channel(priv->channel);
2762 dpaa_fq_setup(priv, &dpaa_fq_cbs, priv->mac_dev->port[TX]);
2764 /* Create a congestion group for this netdev, with
2765 * dynamically-allocated CGR ID.
2766 * Must be executed after probing the MAC, but before
2767 * assigning the egress FQs to the CGRs.
2769 err = dpaa_eth_cgr_init(priv);
2771 dev_err(dev, "Error initializing CGR\n");
2772 goto tx_cgr_init_failed;
2775 err = dpaa_ingress_cgr_init(priv);
2777 dev_err(dev, "Error initializing ingress CGR\n");
2778 goto rx_cgr_init_failed;
2781 /* Add the FQs to the interface, and make them active */
2782 list_for_each_entry_safe(dpaa_fq, tmp, &priv->dpaa_fq_list, list) {
2783 err = dpaa_fq_init(dpaa_fq, false);
2785 goto fq_alloc_failed;
2788 priv->tx_headroom = dpaa_get_headroom(&priv->buf_layout[TX]);
2789 priv->rx_headroom = dpaa_get_headroom(&priv->buf_layout[RX]);
2791 /* All real interfaces need their ports initialized */
2792 err = dpaa_eth_init_ports(mac_dev, dpaa_bps, DPAA_BPS_NUM, &port_fqs,
2793 &priv->buf_layout[0], dev);
2795 goto init_ports_failed;
2797 /* Rx traffic distribution based on keygen hashing defaults to on */
2798 priv->keygen_in_use = true;
2800 priv->percpu_priv = devm_alloc_percpu(dev, *priv->percpu_priv);
2801 if (!priv->percpu_priv) {
2802 dev_err(dev, "devm_alloc_percpu() failed\n");
2804 goto alloc_percpu_failed;
2806 for_each_possible_cpu(i) {
2807 percpu_priv = per_cpu_ptr(priv->percpu_priv, i);
2808 memset(percpu_priv, 0, sizeof(*percpu_priv));
2812 netif_set_real_num_tx_queues(net_dev, priv->num_tc * DPAA_TC_TXQ_NUM);
2814 /* Initialize NAPI */
2815 err = dpaa_napi_add(net_dev);
2817 goto napi_add_failed;
2819 err = dpaa_netdev_init(net_dev, &dpaa_ops, tx_timeout);
2821 goto netdev_init_failed;
2823 dpaa_eth_sysfs_init(&net_dev->dev);
2825 netif_info(priv, probe, net_dev, "Probed interface %s\n",
2832 dpaa_napi_del(net_dev);
2833 alloc_percpu_failed:
2835 dpaa_fq_free(dev, &priv->dpaa_fq_list);
2837 qman_delete_cgr_safe(&priv->ingress_cgr);
2838 qman_release_cgrid(priv->ingress_cgr.cgrid);
2840 qman_delete_cgr_safe(&priv->cgr_data.cgr);
2841 qman_release_cgrid(priv->cgr_data.cgr.cgrid);
2844 dpaa_bps_free(priv);
2849 dev_set_drvdata(dev, NULL);
2850 free_netdev(net_dev);
2851 alloc_etherdev_mq_failed:
2852 for (i = 0; i < DPAA_BPS_NUM && dpaa_bps[i]; i++) {
2853 if (atomic_read(&dpaa_bps[i]->refs) == 0)
2854 devm_kfree(dev, dpaa_bps[i]);
2859 static int dpaa_remove(struct platform_device *pdev)
2861 struct net_device *net_dev;
2862 struct dpaa_priv *priv;
2867 net_dev = dev_get_drvdata(dev);
2869 priv = netdev_priv(net_dev);
2871 dpaa_eth_sysfs_remove(dev);
2873 dev_set_drvdata(dev, NULL);
2874 unregister_netdev(net_dev);
2876 err = dpaa_fq_free(dev, &priv->dpaa_fq_list);
2878 qman_delete_cgr_safe(&priv->ingress_cgr);
2879 qman_release_cgrid(priv->ingress_cgr.cgrid);
2880 qman_delete_cgr_safe(&priv->cgr_data.cgr);
2881 qman_release_cgrid(priv->cgr_data.cgr.cgrid);
2883 dpaa_napi_del(net_dev);
2885 dpaa_bps_free(priv);
2887 free_netdev(net_dev);
2892 static const struct platform_device_id dpaa_devtype[] = {
2894 .name = "dpaa-ethernet",
2899 MODULE_DEVICE_TABLE(platform, dpaa_devtype);
2901 static struct platform_driver dpaa_driver = {
2903 .name = KBUILD_MODNAME,
2905 .id_table = dpaa_devtype,
2906 .probe = dpaa_eth_probe,
2907 .remove = dpaa_remove
2910 static int __init dpaa_load(void)
2914 pr_debug("FSL DPAA Ethernet driver\n");
2916 /* initialize dpaa_eth mirror values */
2917 dpaa_rx_extra_headroom = fman_get_rx_extra_headroom();
2918 dpaa_max_frm = fman_get_max_frm();
2920 err = platform_driver_register(&dpaa_driver);
2922 pr_err("Error, platform_driver_register() = %d\n", err);
2926 module_init(dpaa_load);
2928 static void __exit dpaa_unload(void)
2930 platform_driver_unregister(&dpaa_driver);
2932 /* Only one channel is used and needs to be released after all
2933 * interfaces are removed
2935 dpaa_release_channel();
2937 module_exit(dpaa_unload);
2939 MODULE_LICENSE("Dual BSD/GPL");
2940 MODULE_DESCRIPTION("FSL DPAA Ethernet driver");