1 /* QLogic qede NIC Driver
2  * Copyright (c) 2015-2017  QLogic Corporation
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and /or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 #ifndef _QEDE_H_
33 #define _QEDE_H_
34 #include <linux/compiler.h>
35 #include <linux/version.h>
36 #include <linux/workqueue.h>
37 #include <linux/netdevice.h>
38 #include <linux/interrupt.h>
39 #include <linux/bitmap.h>
40 #include <linux/kernel.h>
41 #include <linux/mutex.h>
42 #include <linux/bpf.h>
43 #include <net/xdp.h>
44 #include <linux/qed/qede_rdma.h>
45 #include <linux/io.h>
46 #ifdef CONFIG_RFS_ACCEL
47 #include <linux/cpu_rmap.h>
48 #endif
49 #include <linux/qed/common_hsi.h>
50 #include <linux/qed/eth_common.h>
51 #include <linux/qed/qed_if.h>
52 #include <linux/qed/qed_chain.h>
53 #include <linux/qed/qed_eth_if.h>
54 
55 #define QEDE_MAJOR_VERSION		8
56 #define QEDE_MINOR_VERSION		33
57 #define QEDE_REVISION_VERSION		0
58 #define QEDE_ENGINEERING_VERSION	20
59 #define DRV_MODULE_VERSION __stringify(QEDE_MAJOR_VERSION) "."	\
60 		__stringify(QEDE_MINOR_VERSION) "."		\
61 		__stringify(QEDE_REVISION_VERSION) "."		\
62 		__stringify(QEDE_ENGINEERING_VERSION)
63 
64 #define DRV_MODULE_SYM		qede
65 
66 struct qede_stats_common {
67 	u64 no_buff_discards;
68 	u64 packet_too_big_discard;
69 	u64 ttl0_discard;
70 	u64 rx_ucast_bytes;
71 	u64 rx_mcast_bytes;
72 	u64 rx_bcast_bytes;
73 	u64 rx_ucast_pkts;
74 	u64 rx_mcast_pkts;
75 	u64 rx_bcast_pkts;
76 	u64 mftag_filter_discards;
77 	u64 mac_filter_discards;
78 	u64 gft_filter_drop;
79 	u64 tx_ucast_bytes;
80 	u64 tx_mcast_bytes;
81 	u64 tx_bcast_bytes;
82 	u64 tx_ucast_pkts;
83 	u64 tx_mcast_pkts;
84 	u64 tx_bcast_pkts;
85 	u64 tx_err_drop_pkts;
86 	u64 coalesced_pkts;
87 	u64 coalesced_events;
88 	u64 coalesced_aborts_num;
89 	u64 non_coalesced_pkts;
90 	u64 coalesced_bytes;
91 
92 	/* port */
93 	u64 rx_64_byte_packets;
94 	u64 rx_65_to_127_byte_packets;
95 	u64 rx_128_to_255_byte_packets;
96 	u64 rx_256_to_511_byte_packets;
97 	u64 rx_512_to_1023_byte_packets;
98 	u64 rx_1024_to_1518_byte_packets;
99 	u64 rx_crc_errors;
100 	u64 rx_mac_crtl_frames;
101 	u64 rx_pause_frames;
102 	u64 rx_pfc_frames;
103 	u64 rx_align_errors;
104 	u64 rx_carrier_errors;
105 	u64 rx_oversize_packets;
106 	u64 rx_jabbers;
107 	u64 rx_undersize_packets;
108 	u64 rx_fragments;
109 	u64 tx_64_byte_packets;
110 	u64 tx_65_to_127_byte_packets;
111 	u64 tx_128_to_255_byte_packets;
112 	u64 tx_256_to_511_byte_packets;
113 	u64 tx_512_to_1023_byte_packets;
114 	u64 tx_1024_to_1518_byte_packets;
115 	u64 tx_pause_frames;
116 	u64 tx_pfc_frames;
117 	u64 brb_truncates;
118 	u64 brb_discards;
119 	u64 tx_mac_ctrl_frames;
120 };
121 
122 struct qede_stats_bb {
123 	u64 rx_1519_to_1522_byte_packets;
124 	u64 rx_1519_to_2047_byte_packets;
125 	u64 rx_2048_to_4095_byte_packets;
126 	u64 rx_4096_to_9216_byte_packets;
127 	u64 rx_9217_to_16383_byte_packets;
128 	u64 tx_1519_to_2047_byte_packets;
129 	u64 tx_2048_to_4095_byte_packets;
130 	u64 tx_4096_to_9216_byte_packets;
131 	u64 tx_9217_to_16383_byte_packets;
132 	u64 tx_lpi_entry_count;
133 	u64 tx_total_collisions;
134 };
135 
136 struct qede_stats_ah {
137 	u64 rx_1519_to_max_byte_packets;
138 	u64 tx_1519_to_max_byte_packets;
139 };
140 
141 struct qede_stats {
142 	struct qede_stats_common common;
143 
144 	union {
145 		struct qede_stats_bb bb;
146 		struct qede_stats_ah ah;
147 	};
148 };
149 
150 struct qede_vlan {
151 	struct list_head list;
152 	u16 vid;
153 	bool configured;
154 };
155 
156 struct qede_rdma_dev {
157 	struct qedr_dev *qedr_dev;
158 	struct list_head entry;
159 	struct list_head rdma_event_list;
160 	struct workqueue_struct *rdma_wq;
161 };
162 
163 struct qede_ptp;
164 
165 #define QEDE_RFS_MAX_FLTR	256
166 
167 struct qede_dev {
168 	struct qed_dev			*cdev;
169 	struct net_device		*ndev;
170 	struct pci_dev			*pdev;
171 
172 	u32				dp_module;
173 	u8				dp_level;
174 
175 	unsigned long flags;
176 #define QEDE_FLAG_IS_VF			BIT(0)
177 #define IS_VF(edev)	(!!((edev)->flags & QEDE_FLAG_IS_VF))
178 #define QEDE_TX_TIMESTAMPING_EN		BIT(1)
179 #define QEDE_FLAGS_PTP_TX_IN_PRORGESS	BIT(2)
180 
181 	const struct qed_eth_ops	*ops;
182 	struct qede_ptp			*ptp;
183 
184 	struct qed_dev_eth_info dev_info;
185 #define QEDE_MAX_RSS_CNT(edev)	((edev)->dev_info.num_queues)
186 #define QEDE_MAX_TSS_CNT(edev)	((edev)->dev_info.num_queues)
187 #define QEDE_IS_BB(edev) \
188 	((edev)->dev_info.common.dev_type == QED_DEV_TYPE_BB)
189 #define QEDE_IS_AH(edev) \
190 	((edev)->dev_info.common.dev_type == QED_DEV_TYPE_AH)
191 
192 	struct qede_fastpath		*fp_array;
193 	u8				req_num_tx;
194 	u8				fp_num_tx;
195 	u8				req_num_rx;
196 	u8				fp_num_rx;
197 	u16				req_queues;
198 	u16				num_queues;
199 #define QEDE_QUEUE_CNT(edev)	((edev)->num_queues)
200 #define QEDE_RSS_COUNT(edev)	((edev)->num_queues - (edev)->fp_num_tx)
201 #define QEDE_RX_QUEUE_IDX(edev, i)	(i)
202 #define QEDE_TSS_COUNT(edev)	((edev)->num_queues - (edev)->fp_num_rx)
203 
204 	struct qed_int_info		int_info;
205 
206 	/* Smaller private varaiant of the RTNL lock */
207 	struct mutex			qede_lock;
208 	u32				state; /* Protected by qede_lock */
209 	u16				rx_buf_size;
210 	u32				rx_copybreak;
211 
212 	/* L2 header size + 2*VLANs (8 bytes) + LLC SNAP (8 bytes) */
213 #define ETH_OVERHEAD			(ETH_HLEN + 8 + 8)
214 	/* Max supported alignment is 256 (8 shift)
215 	 * minimal alignment shift 6 is optimal for 57xxx HW performance
216 	 */
217 #define QEDE_RX_ALIGN_SHIFT		max(6, min(8, L1_CACHE_SHIFT))
218 	/* We assume skb_build() uses sizeof(struct skb_shared_info) bytes
219 	 * at the end of skb->data, to avoid wasting a full cache line.
220 	 * This reduces memory use (skb->truesize).
221 	 */
222 #define QEDE_FW_RX_ALIGN_END					\
223 	max_t(u64, 1UL << QEDE_RX_ALIGN_SHIFT,			\
224 	      SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
225 
226 	struct qede_stats		stats;
227 #define QEDE_RSS_INDIR_INITED	BIT(0)
228 #define QEDE_RSS_KEY_INITED	BIT(1)
229 #define QEDE_RSS_CAPS_INITED	BIT(2)
230 	u32 rss_params_inited; /* bit-field to track initialized rss params */
231 	u16 rss_ind_table[128];
232 	u32 rss_key[10];
233 	u8 rss_caps;
234 
235 	u16			q_num_rx_buffers; /* Must be a power of two */
236 	u16			q_num_tx_buffers; /* Must be a power of two */
237 
238 	bool gro_disable;
239 	struct list_head vlan_list;
240 	u16 configured_vlans;
241 	u16 non_configured_vlans;
242 	bool accept_any_vlan;
243 	struct delayed_work		sp_task;
244 	unsigned long			sp_flags;
245 	u16				vxlan_dst_port;
246 	u16				geneve_dst_port;
247 
248 	struct qede_arfs		*arfs;
249 	bool				wol_enabled;
250 
251 	struct qede_rdma_dev		rdma_info;
252 
253 	struct bpf_prog *xdp_prog;
254 };
255 
256 enum QEDE_STATE {
257 	QEDE_STATE_CLOSED,
258 	QEDE_STATE_OPEN,
259 };
260 
261 #define HILO_U64(hi, lo)		((((u64)(hi)) << 32) + (lo))
262 
263 #define	MAX_NUM_TC	8
264 #define	MAX_NUM_PRI	8
265 
266 /* The driver supports the new build_skb() API:
267  * RX ring buffer contains pointer to kmalloc() data only,
268  * skb are built only after the frame was DMA-ed.
269  */
270 struct sw_rx_data {
271 	struct page *data;
272 	dma_addr_t mapping;
273 	unsigned int page_offset;
274 };
275 
276 enum qede_agg_state {
277 	QEDE_AGG_STATE_NONE  = 0,
278 	QEDE_AGG_STATE_START = 1,
279 	QEDE_AGG_STATE_ERROR = 2
280 };
281 
282 struct qede_agg_info {
283 	/* rx_buf is a data buffer that can be placed / consumed from rx bd
284 	 * chain. It has two purposes: We will preallocate the data buffer
285 	 * for each aggregation when we open the interface and will place this
286 	 * buffer on the rx-bd-ring when we receive TPA_START. We don't want
287 	 * to be in a state where allocation fails, as we can't reuse the
288 	 * consumer buffer in the rx-chain since FW may still be writing to it
289 	 * (since header needs to be modified for TPA).
290 	 * The second purpose is to keep a pointer to the bd buffer during
291 	 * aggregation.
292 	 */
293 	struct sw_rx_data buffer;
294 	struct sk_buff *skb;
295 
296 	/* We need some structs from the start cookie until termination */
297 	u16 vlan_tag;
298 
299 	bool tpa_start_fail;
300 	u8 state;
301 	u8 frag_id;
302 
303 	u8 tunnel_type;
304 };
305 
306 struct qede_rx_queue {
307 	__le16 *hw_cons_ptr;
308 	void __iomem *hw_rxq_prod_addr;
309 
310 	/* Required for the allocation of replacement buffers */
311 	struct device *dev;
312 
313 	struct bpf_prog *xdp_prog;
314 
315 	u16 sw_rx_cons;
316 	u16 sw_rx_prod;
317 
318 	u16 filled_buffers;
319 	u8 data_direction;
320 	u8 rxq_id;
321 
322 	/* Used once per each NAPI run */
323 	u16 num_rx_buffers;
324 
325 	u16 rx_headroom;
326 
327 	u32 rx_buf_size;
328 	u32 rx_buf_seg_size;
329 
330 	struct sw_rx_data *sw_rx_ring;
331 	struct qed_chain rx_bd_ring;
332 	struct qed_chain rx_comp_ring ____cacheline_aligned;
333 
334 	/* GRO */
335 	struct qede_agg_info tpa_info[ETH_TPA_MAX_AGGS_NUM];
336 
337 	/* Used once per each NAPI run */
338 	u64 rcv_pkts;
339 
340 	u64 rx_hw_errors;
341 	u64 rx_alloc_errors;
342 	u64 rx_ip_frags;
343 
344 	u64 xdp_no_pass;
345 
346 	void *handle;
347 	struct xdp_rxq_info xdp_rxq;
348 };
349 
350 union db_prod {
351 	struct eth_db_data data;
352 	u32		raw;
353 };
354 
355 struct sw_tx_bd {
356 	struct sk_buff *skb;
357 	u8 flags;
358 /* Set on the first BD descriptor when there is a split BD */
359 #define QEDE_TSO_SPLIT_BD		BIT(0)
360 };
361 
362 struct sw_tx_xdp {
363 	struct page *page;
364 	dma_addr_t mapping;
365 };
366 
367 struct qede_tx_queue {
368 	u8 is_xdp;
369 	bool is_legacy;
370 	u16 sw_tx_cons;
371 	u16 sw_tx_prod;
372 	u16 num_tx_buffers; /* Slowpath only */
373 
374 	u64 xmit_pkts;
375 	u64 stopped_cnt;
376 
377 	__le16 *hw_cons_ptr;
378 
379 	/* Needed for the mapping of packets */
380 	struct device *dev;
381 
382 	void __iomem *doorbell_addr;
383 	union db_prod tx_db;
384 	int index; /* Slowpath only */
385 #define QEDE_TXQ_XDP_TO_IDX(edev, txq)	((txq)->index - \
386 					 QEDE_MAX_TSS_CNT(edev))
387 #define QEDE_TXQ_IDX_TO_XDP(edev, idx)	((idx) + QEDE_MAX_TSS_CNT(edev))
388 
389 	/* Regular Tx requires skb + metadata for release purpose,
390 	 * while XDP requires the pages and the mapped address.
391 	 */
392 	union {
393 		struct sw_tx_bd *skbs;
394 		struct sw_tx_xdp *xdp;
395 	} sw_tx_ring;
396 
397 	struct qed_chain tx_pbl;
398 
399 	/* Slowpath; Should be kept in end [unless missing padding] */
400 	void *handle;
401 };
402 
403 #define BD_UNMAP_ADDR(bd)		HILO_U64(le32_to_cpu((bd)->addr.hi), \
404 						 le32_to_cpu((bd)->addr.lo))
405 #define BD_SET_UNMAP_ADDR_LEN(bd, maddr, len)				\
406 	do {								\
407 		(bd)->addr.hi = cpu_to_le32(upper_32_bits(maddr));	\
408 		(bd)->addr.lo = cpu_to_le32(lower_32_bits(maddr));	\
409 		(bd)->nbytes = cpu_to_le16(len);			\
410 	} while (0)
411 #define BD_UNMAP_LEN(bd)		(le16_to_cpu((bd)->nbytes))
412 
413 struct qede_fastpath {
414 	struct qede_dev	*edev;
415 #define QEDE_FASTPATH_TX	BIT(0)
416 #define QEDE_FASTPATH_RX	BIT(1)
417 #define QEDE_FASTPATH_XDP	BIT(2)
418 #define QEDE_FASTPATH_COMBINED	(QEDE_FASTPATH_TX | QEDE_FASTPATH_RX)
419 	u8			type;
420 	u8			id;
421 	u8			xdp_xmit;
422 	struct napi_struct	napi;
423 	struct qed_sb_info	*sb_info;
424 	struct qede_rx_queue	*rxq;
425 	struct qede_tx_queue	*txq;
426 	struct qede_tx_queue	*xdp_tx;
427 
428 #define VEC_NAME_SIZE	(sizeof(((struct net_device *)0)->name) + 8)
429 	char	name[VEC_NAME_SIZE];
430 };
431 
432 /* Debug print definitions */
433 #define DP_NAME(edev) ((edev)->ndev->name)
434 
435 #define XMIT_PLAIN		0
436 #define XMIT_L4_CSUM		BIT(0)
437 #define XMIT_LSO		BIT(1)
438 #define XMIT_ENC		BIT(2)
439 #define XMIT_ENC_GSO_L4_CSUM	BIT(3)
440 
441 #define QEDE_CSUM_ERROR			BIT(0)
442 #define QEDE_CSUM_UNNECESSARY		BIT(1)
443 #define QEDE_TUNN_CSUM_UNNECESSARY	BIT(2)
444 
445 #define QEDE_SP_RX_MODE			1
446 
447 #ifdef CONFIG_RFS_ACCEL
448 int qede_rx_flow_steer(struct net_device *dev, const struct sk_buff *skb,
449 		       u16 rxq_index, u32 flow_id);
450 #define QEDE_SP_ARFS_CONFIG	4
451 #define QEDE_SP_TASK_POLL_DELAY	(5 * HZ)
452 #endif
453 
454 void qede_process_arfs_filters(struct qede_dev *edev, bool free_fltr);
455 void qede_poll_for_freeing_arfs_filters(struct qede_dev *edev);
456 void qede_arfs_filter_op(void *dev, void *filter, u8 fw_rc);
457 void qede_free_arfs(struct qede_dev *edev);
458 int qede_alloc_arfs(struct qede_dev *edev);
459 int qede_add_cls_rule(struct qede_dev *edev, struct ethtool_rxnfc *info);
460 int qede_del_cls_rule(struct qede_dev *edev, struct ethtool_rxnfc *info);
461 int qede_get_cls_rule_entry(struct qede_dev *edev, struct ethtool_rxnfc *cmd);
462 int qede_get_cls_rule_all(struct qede_dev *edev, struct ethtool_rxnfc *info,
463 			  u32 *rule_locs);
464 int qede_get_arfs_filter_count(struct qede_dev *edev);
465 
466 struct qede_reload_args {
467 	void (*func)(struct qede_dev *edev, struct qede_reload_args *args);
468 	union {
469 		netdev_features_t features;
470 		struct bpf_prog *new_prog;
471 		u16 mtu;
472 	} u;
473 };
474 
475 /* Datapath functions definition */
476 netdev_tx_t qede_start_xmit(struct sk_buff *skb, struct net_device *ndev);
477 netdev_features_t qede_features_check(struct sk_buff *skb,
478 				      struct net_device *dev,
479 				      netdev_features_t features);
480 void qede_tx_log_print(struct qede_dev *edev, struct qede_fastpath *fp);
481 int qede_alloc_rx_buffer(struct qede_rx_queue *rxq, bool allow_lazy);
482 int qede_free_tx_pkt(struct qede_dev *edev,
483 		     struct qede_tx_queue *txq, int *len);
484 int qede_poll(struct napi_struct *napi, int budget);
485 irqreturn_t qede_msix_fp_int(int irq, void *fp_cookie);
486 
487 /* Filtering function definitions */
488 void qede_force_mac(void *dev, u8 *mac, bool forced);
489 void qede_udp_ports_update(void *dev, u16 vxlan_port, u16 geneve_port);
490 int qede_set_mac_addr(struct net_device *ndev, void *p);
491 
492 int qede_vlan_rx_add_vid(struct net_device *dev, __be16 proto, u16 vid);
493 int qede_vlan_rx_kill_vid(struct net_device *dev, __be16 proto, u16 vid);
494 void qede_vlan_mark_nonconfigured(struct qede_dev *edev);
495 int qede_configure_vlan_filters(struct qede_dev *edev);
496 
497 netdev_features_t qede_fix_features(struct net_device *dev,
498 				    netdev_features_t features);
499 int qede_set_features(struct net_device *dev, netdev_features_t features);
500 void qede_set_rx_mode(struct net_device *ndev);
501 void qede_config_rx_mode(struct net_device *ndev);
502 void qede_fill_rss_params(struct qede_dev *edev,
503 			  struct qed_update_vport_rss_params *rss, u8 *update);
504 
505 void qede_udp_tunnel_add(struct net_device *dev, struct udp_tunnel_info *ti);
506 void qede_udp_tunnel_del(struct net_device *dev, struct udp_tunnel_info *ti);
507 
508 int qede_xdp(struct net_device *dev, struct netdev_bpf *xdp);
509 
510 #ifdef CONFIG_DCB
511 void qede_set_dcbnl_ops(struct net_device *ndev);
512 #endif
513 
514 void qede_config_debug(uint debug, u32 *p_dp_module, u8 *p_dp_level);
515 void qede_set_ethtool_ops(struct net_device *netdev);
516 void qede_reload(struct qede_dev *edev,
517 		 struct qede_reload_args *args, bool is_locked);
518 int qede_change_mtu(struct net_device *dev, int new_mtu);
519 void qede_fill_by_demand_stats(struct qede_dev *edev);
520 void __qede_lock(struct qede_dev *edev);
521 void __qede_unlock(struct qede_dev *edev);
522 bool qede_has_rx_work(struct qede_rx_queue *rxq);
523 int qede_txq_has_work(struct qede_tx_queue *txq);
524 void qede_recycle_rx_bd_ring(struct qede_rx_queue *rxq, u8 count);
525 void qede_update_rx_prod(struct qede_dev *edev, struct qede_rx_queue *rxq);
526 
527 #define RX_RING_SIZE_POW	13
528 #define RX_RING_SIZE		((u16)BIT(RX_RING_SIZE_POW))
529 #define NUM_RX_BDS_MAX		(RX_RING_SIZE - 1)
530 #define NUM_RX_BDS_MIN		128
531 #define NUM_RX_BDS_DEF		((u16)BIT(10) - 1)
532 
533 #define TX_RING_SIZE_POW	13
534 #define TX_RING_SIZE		((u16)BIT(TX_RING_SIZE_POW))
535 #define NUM_TX_BDS_MAX		(TX_RING_SIZE - 1)
536 #define NUM_TX_BDS_MIN		128
537 #define NUM_TX_BDS_DEF		NUM_TX_BDS_MAX
538 
539 #define QEDE_MIN_PKT_LEN		64
540 #define QEDE_RX_HDR_SIZE		256
541 #define QEDE_MAX_JUMBO_PACKET_SIZE	9600
542 #define	for_each_queue(i) for (i = 0; i < edev->num_queues; i++)
543 
544 #endif /* _QEDE_H_ */
545