xref: /openbmc/linux/drivers/net/ethernet/sfc/ptp.c (revision 4f3db074)
1 /****************************************************************************
2  * Driver for Solarflare network controllers and boards
3  * Copyright 2011-2013 Solarflare Communications Inc.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms of the GNU General Public License version 2 as published
7  * by the Free Software Foundation, incorporated herein by reference.
8  */
9 
10 /* Theory of operation:
11  *
12  * PTP support is assisted by firmware running on the MC, which provides
13  * the hardware timestamping capabilities.  Both transmitted and received
14  * PTP event packets are queued onto internal queues for subsequent processing;
15  * this is because the MC operations are relatively long and would block
16  * block NAPI/interrupt operation.
17  *
18  * Receive event processing:
19  *	The event contains the packet's UUID and sequence number, together
20  *	with the hardware timestamp.  The PTP receive packet queue is searched
21  *	for this UUID/sequence number and, if found, put on a pending queue.
22  *	Packets not matching are delivered without timestamps (MCDI events will
23  *	always arrive after the actual packet).
24  *	It is important for the operation of the PTP protocol that the ordering
25  *	of packets between the event and general port is maintained.
26  *
27  * Work queue processing:
28  *	If work waiting, synchronise host/hardware time
29  *
30  *	Transmit: send packet through MC, which returns the transmission time
31  *	that is converted to an appropriate timestamp.
32  *
33  *	Receive: the packet's reception time is converted to an appropriate
34  *	timestamp.
35  */
36 #include <linux/ip.h>
37 #include <linux/udp.h>
38 #include <linux/time.h>
39 #include <linux/ktime.h>
40 #include <linux/module.h>
41 #include <linux/net_tstamp.h>
42 #include <linux/pps_kernel.h>
43 #include <linux/ptp_clock_kernel.h>
44 #include "net_driver.h"
45 #include "efx.h"
46 #include "mcdi.h"
47 #include "mcdi_pcol.h"
48 #include "io.h"
49 #include "farch_regs.h"
50 #include "nic.h"
51 
52 /* Maximum number of events expected to make up a PTP event */
53 #define	MAX_EVENT_FRAGS			3
54 
55 /* Maximum delay, ms, to begin synchronisation */
56 #define	MAX_SYNCHRONISE_WAIT_MS		2
57 
58 /* How long, at most, to spend synchronising */
59 #define	SYNCHRONISE_PERIOD_NS		250000
60 
61 /* How often to update the shared memory time */
62 #define	SYNCHRONISATION_GRANULARITY_NS	200
63 
64 /* Minimum permitted length of a (corrected) synchronisation time */
65 #define	DEFAULT_MIN_SYNCHRONISATION_NS	120
66 
67 /* Maximum permitted length of a (corrected) synchronisation time */
68 #define	MAX_SYNCHRONISATION_NS		1000
69 
70 /* How many (MC) receive events that can be queued */
71 #define	MAX_RECEIVE_EVENTS		8
72 
73 /* Length of (modified) moving average. */
74 #define	AVERAGE_LENGTH			16
75 
76 /* How long an unmatched event or packet can be held */
77 #define PKT_EVENT_LIFETIME_MS		10
78 
79 /* Offsets into PTP packet for identification.  These offsets are from the
80  * start of the IP header, not the MAC header.  Note that neither PTP V1 nor
81  * PTP V2 permit the use of IPV4 options.
82  */
83 #define PTP_DPORT_OFFSET	22
84 
85 #define PTP_V1_VERSION_LENGTH	2
86 #define PTP_V1_VERSION_OFFSET	28
87 
88 #define PTP_V1_UUID_LENGTH	6
89 #define PTP_V1_UUID_OFFSET	50
90 
91 #define PTP_V1_SEQUENCE_LENGTH	2
92 #define PTP_V1_SEQUENCE_OFFSET	58
93 
94 /* The minimum length of a PTP V1 packet for offsets, etc. to be valid:
95  * includes IP header.
96  */
97 #define	PTP_V1_MIN_LENGTH	64
98 
99 #define PTP_V2_VERSION_LENGTH	1
100 #define PTP_V2_VERSION_OFFSET	29
101 
102 #define PTP_V2_UUID_LENGTH	8
103 #define PTP_V2_UUID_OFFSET	48
104 
105 /* Although PTP V2 UUIDs are comprised a ClockIdentity (8) and PortNumber (2),
106  * the MC only captures the last six bytes of the clock identity. These values
107  * reflect those, not the ones used in the standard.  The standard permits
108  * mapping of V1 UUIDs to V2 UUIDs with these same values.
109  */
110 #define PTP_V2_MC_UUID_LENGTH	6
111 #define PTP_V2_MC_UUID_OFFSET	50
112 
113 #define PTP_V2_SEQUENCE_LENGTH	2
114 #define PTP_V2_SEQUENCE_OFFSET	58
115 
116 /* The minimum length of a PTP V2 packet for offsets, etc. to be valid:
117  * includes IP header.
118  */
119 #define	PTP_V2_MIN_LENGTH	63
120 
121 #define	PTP_MIN_LENGTH		63
122 
123 #define PTP_ADDRESS		0xe0000181	/* 224.0.1.129 */
124 #define PTP_EVENT_PORT		319
125 #define PTP_GENERAL_PORT	320
126 
127 /* Annoyingly the format of the version numbers are different between
128  * versions 1 and 2 so it isn't possible to simply look for 1 or 2.
129  */
130 #define	PTP_VERSION_V1		1
131 
132 #define	PTP_VERSION_V2		2
133 #define	PTP_VERSION_V2_MASK	0x0f
134 
135 enum ptp_packet_state {
136 	PTP_PACKET_STATE_UNMATCHED = 0,
137 	PTP_PACKET_STATE_MATCHED,
138 	PTP_PACKET_STATE_TIMED_OUT,
139 	PTP_PACKET_STATE_MATCH_UNWANTED
140 };
141 
142 /* NIC synchronised with single word of time only comprising
143  * partial seconds and full nanoseconds: 10^9 ~ 2^30 so 2 bits for seconds.
144  */
145 #define	MC_NANOSECOND_BITS	30
146 #define	MC_NANOSECOND_MASK	((1 << MC_NANOSECOND_BITS) - 1)
147 #define	MC_SECOND_MASK		((1 << (32 - MC_NANOSECOND_BITS)) - 1)
148 
149 /* Maximum parts-per-billion adjustment that is acceptable */
150 #define MAX_PPB			1000000
151 
152 /* Number of bits required to hold the above */
153 #define	MAX_PPB_BITS		20
154 
155 /* Number of extra bits allowed when calculating fractional ns.
156  * EXTRA_BITS + MC_CMD_PTP_IN_ADJUST_BITS + MAX_PPB_BITS should
157  * be less than 63.
158  */
159 #define	PPB_EXTRA_BITS		2
160 
161 /* Precalculate scale word to avoid long long division at runtime */
162 #define	PPB_SCALE_WORD	((1LL << (PPB_EXTRA_BITS + MC_CMD_PTP_IN_ADJUST_BITS +\
163 			MAX_PPB_BITS)) / 1000000000LL)
164 
165 #define PTP_SYNC_ATTEMPTS	4
166 
167 /**
168  * struct efx_ptp_match - Matching structure, stored in sk_buff's cb area.
169  * @words: UUID and (partial) sequence number
170  * @expiry: Time after which the packet should be delivered irrespective of
171  *            event arrival.
172  * @state: The state of the packet - whether it is ready for processing or
173  *         whether that is of no interest.
174  */
175 struct efx_ptp_match {
176 	u32 words[DIV_ROUND_UP(PTP_V1_UUID_LENGTH, 4)];
177 	unsigned long expiry;
178 	enum ptp_packet_state state;
179 };
180 
181 /**
182  * struct efx_ptp_event_rx - A PTP receive event (from MC)
183  * @seq0: First part of (PTP) UUID
184  * @seq1: Second part of (PTP) UUID and sequence number
185  * @hwtimestamp: Event timestamp
186  */
187 struct efx_ptp_event_rx {
188 	struct list_head link;
189 	u32 seq0;
190 	u32 seq1;
191 	ktime_t hwtimestamp;
192 	unsigned long expiry;
193 };
194 
195 /**
196  * struct efx_ptp_timeset - Synchronisation between host and MC
197  * @host_start: Host time immediately before hardware timestamp taken
198  * @major: Hardware timestamp, major
199  * @minor: Hardware timestamp, minor
200  * @host_end: Host time immediately after hardware timestamp taken
201  * @wait: Number of NIC clock ticks between hardware timestamp being read and
202  *          host end time being seen
203  * @window: Difference of host_end and host_start
204  * @valid: Whether this timeset is valid
205  */
206 struct efx_ptp_timeset {
207 	u32 host_start;
208 	u32 major;
209 	u32 minor;
210 	u32 host_end;
211 	u32 wait;
212 	u32 window;	/* Derived: end - start, allowing for wrap */
213 };
214 
215 /**
216  * struct efx_ptp_data - Precision Time Protocol (PTP) state
217  * @efx: The NIC context
218  * @channel: The PTP channel (Siena only)
219  * @rx_ts_inline: Flag for whether RX timestamps are inline (else they are
220  *	separate events)
221  * @rxq: Receive queue (awaiting timestamps)
222  * @txq: Transmit queue
223  * @evt_list: List of MC receive events awaiting packets
224  * @evt_free_list: List of free events
225  * @evt_lock: Lock for manipulating evt_list and evt_free_list
226  * @rx_evts: Instantiated events (on evt_list and evt_free_list)
227  * @workwq: Work queue for processing pending PTP operations
228  * @work: Work task
229  * @reset_required: A serious error has occurred and the PTP task needs to be
230  *                  reset (disable, enable).
231  * @rxfilter_event: Receive filter when operating
232  * @rxfilter_general: Receive filter when operating
233  * @config: Current timestamp configuration
234  * @enabled: PTP operation enabled
235  * @mode: Mode in which PTP operating (PTP version)
236  * @time_format: Time format supported by this NIC
237  * @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time
238  * @nic_to_kernel_time: Function to convert from NIC to kernel time
239  * @min_synchronisation_ns: Minimum acceptable corrected sync window
240  * @ts_corrections.tx: Required driver correction of transmit timestamps
241  * @ts_corrections.rx: Required driver correction of receive timestamps
242  * @ts_corrections.pps_out: PPS output error (information only)
243  * @ts_corrections.pps_in: Required driver correction of PPS input timestamps
244  * @evt_frags: Partly assembled PTP events
245  * @evt_frag_idx: Current fragment number
246  * @evt_code: Last event code
247  * @start: Address at which MC indicates ready for synchronisation
248  * @host_time_pps: Host time at last PPS
249  * @current_adjfreq: Current ppb adjustment.
250  * @phc_clock: Pointer to registered phc device (if primary function)
251  * @phc_clock_info: Registration structure for phc device
252  * @pps_work: pps work task for handling pps events
253  * @pps_workwq: pps work queue
254  * @nic_ts_enabled: Flag indicating if NIC generated TS events are handled
255  * @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids
256  *         allocations in main data path).
257  * @good_syncs: Number of successful synchronisations.
258  * @fast_syncs: Number of synchronisations requiring short delay
259  * @bad_syncs: Number of failed synchronisations.
260  * @sync_timeouts: Number of synchronisation timeouts
261  * @no_time_syncs: Number of synchronisations with no good times.
262  * @invalid_sync_windows: Number of sync windows with bad durations.
263  * @undersize_sync_windows: Number of corrected sync windows that are too small
264  * @oversize_sync_windows: Number of corrected sync windows that are too large
265  * @rx_no_timestamp: Number of packets received without a timestamp.
266  * @timeset: Last set of synchronisation statistics.
267  */
268 struct efx_ptp_data {
269 	struct efx_nic *efx;
270 	struct efx_channel *channel;
271 	bool rx_ts_inline;
272 	struct sk_buff_head rxq;
273 	struct sk_buff_head txq;
274 	struct list_head evt_list;
275 	struct list_head evt_free_list;
276 	spinlock_t evt_lock;
277 	struct efx_ptp_event_rx rx_evts[MAX_RECEIVE_EVENTS];
278 	struct workqueue_struct *workwq;
279 	struct work_struct work;
280 	bool reset_required;
281 	u32 rxfilter_event;
282 	u32 rxfilter_general;
283 	bool rxfilter_installed;
284 	struct hwtstamp_config config;
285 	bool enabled;
286 	unsigned int mode;
287 	unsigned int time_format;
288 	void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor);
289 	ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor,
290 				      s32 correction);
291 	unsigned int min_synchronisation_ns;
292 	struct {
293 		s32 tx;
294 		s32 rx;
295 		s32 pps_out;
296 		s32 pps_in;
297 	} ts_corrections;
298 	efx_qword_t evt_frags[MAX_EVENT_FRAGS];
299 	int evt_frag_idx;
300 	int evt_code;
301 	struct efx_buffer start;
302 	struct pps_event_time host_time_pps;
303 	s64 current_adjfreq;
304 	struct ptp_clock *phc_clock;
305 	struct ptp_clock_info phc_clock_info;
306 	struct work_struct pps_work;
307 	struct workqueue_struct *pps_workwq;
308 	bool nic_ts_enabled;
309 	MCDI_DECLARE_BUF(txbuf, MC_CMD_PTP_IN_TRANSMIT_LENMAX);
310 
311 	unsigned int good_syncs;
312 	unsigned int fast_syncs;
313 	unsigned int bad_syncs;
314 	unsigned int sync_timeouts;
315 	unsigned int no_time_syncs;
316 	unsigned int invalid_sync_windows;
317 	unsigned int undersize_sync_windows;
318 	unsigned int oversize_sync_windows;
319 	unsigned int rx_no_timestamp;
320 	struct efx_ptp_timeset
321 	timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM];
322 };
323 
324 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta);
325 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta);
326 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts);
327 static int efx_phc_settime(struct ptp_clock_info *ptp,
328 			   const struct timespec64 *e_ts);
329 static int efx_phc_enable(struct ptp_clock_info *ptp,
330 			  struct ptp_clock_request *request, int on);
331 
332 #define PTP_SW_STAT(ext_name, field_name)				\
333 	{ #ext_name, 0, offsetof(struct efx_ptp_data, field_name) }
334 #define PTP_MC_STAT(ext_name, mcdi_name)				\
335 	{ #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST }
336 static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = {
337 	PTP_SW_STAT(ptp_good_syncs, good_syncs),
338 	PTP_SW_STAT(ptp_fast_syncs, fast_syncs),
339 	PTP_SW_STAT(ptp_bad_syncs, bad_syncs),
340 	PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts),
341 	PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs),
342 	PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows),
343 	PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows),
344 	PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows),
345 	PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp),
346 	PTP_MC_STAT(ptp_tx_timestamp_packets, TX),
347 	PTP_MC_STAT(ptp_rx_timestamp_packets, RX),
348 	PTP_MC_STAT(ptp_timestamp_packets, TS),
349 	PTP_MC_STAT(ptp_filter_matches, FM),
350 	PTP_MC_STAT(ptp_non_filter_matches, NFM),
351 };
352 #define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc)
353 static const unsigned long efx_ptp_stat_mask[] = {
354 	[0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL,
355 };
356 
357 size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
358 {
359 	if (!efx->ptp_data)
360 		return 0;
361 
362 	return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
363 				      efx_ptp_stat_mask, strings);
364 }
365 
366 size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats)
367 {
368 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
369 	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
370 	size_t i;
371 	int rc;
372 
373 	if (!efx->ptp_data)
374 		return 0;
375 
376 	/* Copy software statistics */
377 	for (i = 0; i < PTP_STAT_COUNT; i++) {
378 		if (efx_ptp_stat_desc[i].dma_width)
379 			continue;
380 		stats[i] = *(unsigned int *)((char *)efx->ptp_data +
381 					     efx_ptp_stat_desc[i].offset);
382 	}
383 
384 	/* Fetch MC statistics.  We *must* fill in all statistics or
385 	 * risk leaking kernel memory to userland, so if the MCDI
386 	 * request fails we pretend we got zeroes.
387 	 */
388 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS);
389 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
390 	rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
391 			  outbuf, sizeof(outbuf), NULL);
392 	if (rc) {
393 		netif_err(efx, hw, efx->net_dev,
394 			  "MC_CMD_PTP_OP_STATUS failed (%d)\n", rc);
395 		memset(outbuf, 0, sizeof(outbuf));
396 	}
397 	efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
398 			     efx_ptp_stat_mask,
399 			     stats, _MCDI_PTR(outbuf, 0), false);
400 
401 	return PTP_STAT_COUNT;
402 }
403 
404 /* For Siena platforms NIC time is s and ns */
405 static void efx_ptp_ns_to_s_ns(s64 ns, u32 *nic_major, u32 *nic_minor)
406 {
407 	struct timespec ts = ns_to_timespec(ns);
408 	*nic_major = ts.tv_sec;
409 	*nic_minor = ts.tv_nsec;
410 }
411 
412 static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
413 						s32 correction)
414 {
415 	ktime_t kt = ktime_set(nic_major, nic_minor);
416 	if (correction >= 0)
417 		kt = ktime_add_ns(kt, (u64)correction);
418 	else
419 		kt = ktime_sub_ns(kt, (u64)-correction);
420 	return kt;
421 }
422 
423 /* To convert from s27 format to ns we multiply then divide by a power of 2.
424  * For the conversion from ns to s27, the operation is also converted to a
425  * multiply and shift.
426  */
427 #define S27_TO_NS_SHIFT	(27)
428 #define NS_TO_S27_MULT	(((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC)
429 #define NS_TO_S27_SHIFT	(63 - S27_TO_NS_SHIFT)
430 #define S27_MINOR_MAX	(1 << S27_TO_NS_SHIFT)
431 
432 /* For Huntington platforms NIC time is in seconds and fractions of a second
433  * where the minor register only uses 27 bits in units of 2^-27s.
434  */
435 static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
436 {
437 	struct timespec ts = ns_to_timespec(ns);
438 	u32 maj = ts.tv_sec;
439 	u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT +
440 			 (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT);
441 
442 	/* The conversion can result in the minor value exceeding the maximum.
443 	 * In this case, round up to the next second.
444 	 */
445 	if (min >= S27_MINOR_MAX) {
446 		min -= S27_MINOR_MAX;
447 		maj++;
448 	}
449 
450 	*nic_major = maj;
451 	*nic_minor = min;
452 }
453 
454 static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
455 {
456 	u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC +
457 			(1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT);
458 	return ktime_set(nic_major, ns);
459 }
460 
461 static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
462 					       s32 correction)
463 {
464 	/* Apply the correction and deal with carry */
465 	nic_minor += correction;
466 	if ((s32)nic_minor < 0) {
467 		nic_minor += S27_MINOR_MAX;
468 		nic_major--;
469 	} else if (nic_minor >= S27_MINOR_MAX) {
470 		nic_minor -= S27_MINOR_MAX;
471 		nic_major++;
472 	}
473 
474 	return efx_ptp_s27_to_ktime(nic_major, nic_minor);
475 }
476 
477 /* Get PTP attributes and set up time conversions */
478 static int efx_ptp_get_attributes(struct efx_nic *efx)
479 {
480 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN);
481 	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN);
482 	struct efx_ptp_data *ptp = efx->ptp_data;
483 	int rc;
484 	u32 fmt;
485 	size_t out_len;
486 
487 	/* Get the PTP attributes. If the NIC doesn't support the operation we
488 	 * use the default format for compatibility with older NICs i.e.
489 	 * seconds and nanoseconds.
490 	 */
491 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES);
492 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
493 	rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
494 			  outbuf, sizeof(outbuf), &out_len);
495 	if (rc == 0)
496 		fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT);
497 	else if (rc == -EINVAL)
498 		fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS;
499 	else
500 		return rc;
501 
502 	if (fmt == MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION) {
503 		ptp->ns_to_nic_time = efx_ptp_ns_to_s27;
504 		ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction;
505 	} else if (fmt == MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS) {
506 		ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns;
507 		ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction;
508 	} else {
509 		return -ERANGE;
510 	}
511 
512 	ptp->time_format = fmt;
513 
514 	/* MC_CMD_PTP_OP_GET_ATTRIBUTES is an extended version of an older
515 	 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT that also returns a value
516 	 * to use for the minimum acceptable corrected synchronization window.
517 	 * If we have the extra information store it. For older firmware that
518 	 * does not implement the extended command use the default value.
519 	 */
520 	if (rc == 0 && out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
521 		ptp->min_synchronisation_ns =
522 			MCDI_DWORD(outbuf,
523 				   PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
524 	else
525 		ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
526 
527 	return 0;
528 }
529 
530 /* Get PTP timestamp corrections */
531 static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
532 {
533 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN);
534 	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_LEN);
535 	int rc;
536 
537 	/* Get the timestamp corrections from the NIC. If this operation is
538 	 * not supported (older NICs) then no correction is required.
539 	 */
540 	MCDI_SET_DWORD(inbuf, PTP_IN_OP,
541 		       MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS);
542 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
543 
544 	rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
545 			  outbuf, sizeof(outbuf), NULL);
546 	if (rc == 0) {
547 		efx->ptp_data->ts_corrections.tx = MCDI_DWORD(outbuf,
548 			PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT);
549 		efx->ptp_data->ts_corrections.rx = MCDI_DWORD(outbuf,
550 			PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE);
551 		efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf,
552 			PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT);
553 		efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf,
554 			PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN);
555 	} else if (rc == -EINVAL) {
556 		efx->ptp_data->ts_corrections.tx = 0;
557 		efx->ptp_data->ts_corrections.rx = 0;
558 		efx->ptp_data->ts_corrections.pps_out = 0;
559 		efx->ptp_data->ts_corrections.pps_in = 0;
560 	} else {
561 		return rc;
562 	}
563 
564 	return 0;
565 }
566 
567 /* Enable MCDI PTP support. */
568 static int efx_ptp_enable(struct efx_nic *efx)
569 {
570 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
571 	MCDI_DECLARE_BUF_OUT_OR_ERR(outbuf, 0);
572 	int rc;
573 
574 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE);
575 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
576 	MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE,
577 		       efx->ptp_data->channel ?
578 		       efx->ptp_data->channel->channel : 0);
579 	MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode);
580 
581 	rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
582 				outbuf, sizeof(outbuf), NULL);
583 	rc = (rc == -EALREADY) ? 0 : rc;
584 	if (rc)
585 		efx_mcdi_display_error(efx, MC_CMD_PTP,
586 				       MC_CMD_PTP_IN_ENABLE_LEN,
587 				       outbuf, sizeof(outbuf), rc);
588 	return rc;
589 }
590 
591 /* Disable MCDI PTP support.
592  *
593  * Note that this function should never rely on the presence of ptp_data -
594  * may be called before that exists.
595  */
596 static int efx_ptp_disable(struct efx_nic *efx)
597 {
598 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
599 	MCDI_DECLARE_BUF_OUT_OR_ERR(outbuf, 0);
600 	int rc;
601 
602 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE);
603 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
604 	rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
605 				outbuf, sizeof(outbuf), NULL);
606 	rc = (rc == -EALREADY) ? 0 : rc;
607 	if (rc)
608 		efx_mcdi_display_error(efx, MC_CMD_PTP,
609 				       MC_CMD_PTP_IN_DISABLE_LEN,
610 				       outbuf, sizeof(outbuf), rc);
611 	return rc;
612 }
613 
614 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
615 {
616 	struct sk_buff *skb;
617 
618 	while ((skb = skb_dequeue(q))) {
619 		local_bh_disable();
620 		netif_receive_skb(skb);
621 		local_bh_enable();
622 	}
623 }
624 
625 static void efx_ptp_handle_no_channel(struct efx_nic *efx)
626 {
627 	netif_err(efx, drv, efx->net_dev,
628 		  "ERROR: PTP requires MSI-X and 1 additional interrupt"
629 		  "vector. PTP disabled\n");
630 }
631 
632 /* Repeatedly send the host time to the MC which will capture the hardware
633  * time.
634  */
635 static void efx_ptp_send_times(struct efx_nic *efx,
636 			       struct pps_event_time *last_time)
637 {
638 	struct pps_event_time now;
639 	struct timespec limit;
640 	struct efx_ptp_data *ptp = efx->ptp_data;
641 	struct timespec start;
642 	int *mc_running = ptp->start.addr;
643 
644 	pps_get_ts(&now);
645 	start = now.ts_real;
646 	limit = now.ts_real;
647 	timespec_add_ns(&limit, SYNCHRONISE_PERIOD_NS);
648 
649 	/* Write host time for specified period or until MC is done */
650 	while ((timespec_compare(&now.ts_real, &limit) < 0) &&
651 	       ACCESS_ONCE(*mc_running)) {
652 		struct timespec update_time;
653 		unsigned int host_time;
654 
655 		/* Don't update continuously to avoid saturating the PCIe bus */
656 		update_time = now.ts_real;
657 		timespec_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS);
658 		do {
659 			pps_get_ts(&now);
660 		} while ((timespec_compare(&now.ts_real, &update_time) < 0) &&
661 			 ACCESS_ONCE(*mc_running));
662 
663 		/* Synchronise NIC with single word of time only */
664 		host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS |
665 			     now.ts_real.tv_nsec);
666 		/* Update host time in NIC memory */
667 		efx->type->ptp_write_host_time(efx, host_time);
668 	}
669 	*last_time = now;
670 }
671 
672 /* Read a timeset from the MC's results and partial process. */
673 static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data),
674 				 struct efx_ptp_timeset *timeset)
675 {
676 	unsigned start_ns, end_ns;
677 
678 	timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART);
679 	timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR);
680 	timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR);
681 	timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND),
682 	timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
683 
684 	/* Ignore seconds */
685 	start_ns = timeset->host_start & MC_NANOSECOND_MASK;
686 	end_ns = timeset->host_end & MC_NANOSECOND_MASK;
687 	/* Allow for rollover */
688 	if (end_ns < start_ns)
689 		end_ns += NSEC_PER_SEC;
690 	/* Determine duration of operation */
691 	timeset->window = end_ns - start_ns;
692 }
693 
694 /* Process times received from MC.
695  *
696  * Extract times from returned results, and establish the minimum value
697  * seen.  The minimum value represents the "best" possible time and events
698  * too much greater than this are rejected - the machine is, perhaps, too
699  * busy. A number of readings are taken so that, hopefully, at least one good
700  * synchronisation will be seen in the results.
701  */
702 static int
703 efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf),
704 		      size_t response_length,
705 		      const struct pps_event_time *last_time)
706 {
707 	unsigned number_readings =
708 		MCDI_VAR_ARRAY_LEN(response_length,
709 				   PTP_OUT_SYNCHRONIZE_TIMESET);
710 	unsigned i;
711 	unsigned ngood = 0;
712 	unsigned last_good = 0;
713 	struct efx_ptp_data *ptp = efx->ptp_data;
714 	u32 last_sec;
715 	u32 start_sec;
716 	struct timespec delta;
717 	ktime_t mc_time;
718 
719 	if (number_readings == 0)
720 		return -EAGAIN;
721 
722 	/* Read the set of results and find the last good host-MC
723 	 * synchronization result. The MC times when it finishes reading the
724 	 * host time so the corrected window time should be fairly constant
725 	 * for a given platform. Increment stats for any results that appear
726 	 * to be erroneous.
727 	 */
728 	for (i = 0; i < number_readings; i++) {
729 		s32 window, corrected;
730 		struct timespec wait;
731 
732 		efx_ptp_read_timeset(
733 			MCDI_ARRAY_STRUCT_PTR(synch_buf,
734 					      PTP_OUT_SYNCHRONIZE_TIMESET, i),
735 			&ptp->timeset[i]);
736 
737 		wait = ktime_to_timespec(
738 			ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0));
739 		window = ptp->timeset[i].window;
740 		corrected = window - wait.tv_nsec;
741 
742 		/* We expect the uncorrected synchronization window to be at
743 		 * least as large as the interval between host start and end
744 		 * times. If it is smaller than this then this is mostly likely
745 		 * to be a consequence of the host's time being adjusted.
746 		 * Check that the corrected sync window is in a reasonable
747 		 * range. If it is out of range it is likely to be because an
748 		 * interrupt or other delay occurred between reading the system
749 		 * time and writing it to MC memory.
750 		 */
751 		if (window < SYNCHRONISATION_GRANULARITY_NS) {
752 			++ptp->invalid_sync_windows;
753 		} else if (corrected >= MAX_SYNCHRONISATION_NS) {
754 			++ptp->oversize_sync_windows;
755 		} else if (corrected < ptp->min_synchronisation_ns) {
756 			++ptp->undersize_sync_windows;
757 		} else {
758 			ngood++;
759 			last_good = i;
760 		}
761 	}
762 
763 	if (ngood == 0) {
764 		netif_warn(efx, drv, efx->net_dev,
765 			   "PTP no suitable synchronisations\n");
766 		return -EAGAIN;
767 	}
768 
769 	/* Calculate delay from last good sync (host time) to last_time.
770 	 * It is possible that the seconds rolled over between taking
771 	 * the start reading and the last value written by the host.  The
772 	 * timescales are such that a gap of more than one second is never
773 	 * expected.  delta is *not* normalised.
774 	 */
775 	start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS;
776 	last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK;
777 	if (start_sec != last_sec &&
778 	    ((start_sec + 1) & MC_SECOND_MASK) != last_sec) {
779 		netif_warn(efx, hw, efx->net_dev,
780 			   "PTP bad synchronisation seconds\n");
781 		return -EAGAIN;
782 	}
783 	delta.tv_sec = (last_sec - start_sec) & 1;
784 	delta.tv_nsec =
785 		last_time->ts_real.tv_nsec -
786 		(ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
787 
788 	/* Convert the NIC time at last good sync into kernel time.
789 	 * No correction is required - this time is the output of a
790 	 * firmware process.
791 	 */
792 	mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
793 					  ptp->timeset[last_good].minor, 0);
794 
795 	/* Calculate delay from NIC top of second to last_time */
796 	delta.tv_nsec += ktime_to_timespec(mc_time).tv_nsec;
797 
798 	/* Set PPS timestamp to match NIC top of second */
799 	ptp->host_time_pps = *last_time;
800 	pps_sub_ts(&ptp->host_time_pps, delta);
801 
802 	return 0;
803 }
804 
805 /* Synchronize times between the host and the MC */
806 static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
807 {
808 	struct efx_ptp_data *ptp = efx->ptp_data;
809 	MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX);
810 	size_t response_length;
811 	int rc;
812 	unsigned long timeout;
813 	struct pps_event_time last_time = {};
814 	unsigned int loops = 0;
815 	int *start = ptp->start.addr;
816 
817 	MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE);
818 	MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0);
819 	MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS,
820 		       num_readings);
821 	MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
822 		       ptp->start.dma_addr);
823 
824 	/* Clear flag that signals MC ready */
825 	ACCESS_ONCE(*start) = 0;
826 	rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf,
827 				MC_CMD_PTP_IN_SYNCHRONIZE_LEN);
828 	EFX_BUG_ON_PARANOID(rc);
829 
830 	/* Wait for start from MCDI (or timeout) */
831 	timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS);
832 	while (!ACCESS_ONCE(*start) && (time_before(jiffies, timeout))) {
833 		udelay(20);	/* Usually start MCDI execution quickly */
834 		loops++;
835 	}
836 
837 	if (loops <= 1)
838 		++ptp->fast_syncs;
839 	if (!time_before(jiffies, timeout))
840 		++ptp->sync_timeouts;
841 
842 	if (ACCESS_ONCE(*start))
843 		efx_ptp_send_times(efx, &last_time);
844 
845 	/* Collect results */
846 	rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP,
847 				 MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
848 				 synch_buf, sizeof(synch_buf),
849 				 &response_length);
850 	if (rc == 0) {
851 		rc = efx_ptp_process_times(efx, synch_buf, response_length,
852 					   &last_time);
853 		if (rc == 0)
854 			++ptp->good_syncs;
855 		else
856 			++ptp->no_time_syncs;
857 	}
858 
859 	/* Increment the bad syncs counter if the synchronize fails, whatever
860 	 * the reason.
861 	 */
862 	if (rc != 0)
863 		++ptp->bad_syncs;
864 
865 	return rc;
866 }
867 
868 /* Transmit a PTP packet, via the MCDI interface, to the wire. */
869 static int efx_ptp_xmit_skb(struct efx_nic *efx, struct sk_buff *skb)
870 {
871 	struct efx_ptp_data *ptp_data = efx->ptp_data;
872 	struct skb_shared_hwtstamps timestamps;
873 	int rc = -EIO;
874 	MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
875 	size_t len;
876 
877 	MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT);
878 	MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0);
879 	MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len);
880 	if (skb_shinfo(skb)->nr_frags != 0) {
881 		rc = skb_linearize(skb);
882 		if (rc != 0)
883 			goto fail;
884 	}
885 
886 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
887 		rc = skb_checksum_help(skb);
888 		if (rc != 0)
889 			goto fail;
890 	}
891 	skb_copy_from_linear_data(skb,
892 				  MCDI_PTR(ptp_data->txbuf,
893 					   PTP_IN_TRANSMIT_PACKET),
894 				  skb->len);
895 	rc = efx_mcdi_rpc(efx, MC_CMD_PTP,
896 			  ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len),
897 			  txtime, sizeof(txtime), &len);
898 	if (rc != 0)
899 		goto fail;
900 
901 	memset(&timestamps, 0, sizeof(timestamps));
902 	timestamps.hwtstamp = ptp_data->nic_to_kernel_time(
903 		MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR),
904 		MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR),
905 		ptp_data->ts_corrections.tx);
906 
907 	skb_tstamp_tx(skb, &timestamps);
908 
909 	rc = 0;
910 
911 fail:
912 	dev_kfree_skb(skb);
913 
914 	return rc;
915 }
916 
917 static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
918 {
919 	struct efx_ptp_data *ptp = efx->ptp_data;
920 	struct list_head *cursor;
921 	struct list_head *next;
922 
923 	if (ptp->rx_ts_inline)
924 		return;
925 
926 	/* Drop time-expired events */
927 	spin_lock_bh(&ptp->evt_lock);
928 	if (!list_empty(&ptp->evt_list)) {
929 		list_for_each_safe(cursor, next, &ptp->evt_list) {
930 			struct efx_ptp_event_rx *evt;
931 
932 			evt = list_entry(cursor, struct efx_ptp_event_rx,
933 					 link);
934 			if (time_after(jiffies, evt->expiry)) {
935 				list_move(&evt->link, &ptp->evt_free_list);
936 				netif_warn(efx, hw, efx->net_dev,
937 					   "PTP rx event dropped\n");
938 			}
939 		}
940 	}
941 	spin_unlock_bh(&ptp->evt_lock);
942 }
943 
944 static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
945 					      struct sk_buff *skb)
946 {
947 	struct efx_ptp_data *ptp = efx->ptp_data;
948 	bool evts_waiting;
949 	struct list_head *cursor;
950 	struct list_head *next;
951 	struct efx_ptp_match *match;
952 	enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED;
953 
954 	WARN_ON_ONCE(ptp->rx_ts_inline);
955 
956 	spin_lock_bh(&ptp->evt_lock);
957 	evts_waiting = !list_empty(&ptp->evt_list);
958 	spin_unlock_bh(&ptp->evt_lock);
959 
960 	if (!evts_waiting)
961 		return PTP_PACKET_STATE_UNMATCHED;
962 
963 	match = (struct efx_ptp_match *)skb->cb;
964 	/* Look for a matching timestamp in the event queue */
965 	spin_lock_bh(&ptp->evt_lock);
966 	list_for_each_safe(cursor, next, &ptp->evt_list) {
967 		struct efx_ptp_event_rx *evt;
968 
969 		evt = list_entry(cursor, struct efx_ptp_event_rx, link);
970 		if ((evt->seq0 == match->words[0]) &&
971 		    (evt->seq1 == match->words[1])) {
972 			struct skb_shared_hwtstamps *timestamps;
973 
974 			/* Match - add in hardware timestamp */
975 			timestamps = skb_hwtstamps(skb);
976 			timestamps->hwtstamp = evt->hwtimestamp;
977 
978 			match->state = PTP_PACKET_STATE_MATCHED;
979 			rc = PTP_PACKET_STATE_MATCHED;
980 			list_move(&evt->link, &ptp->evt_free_list);
981 			break;
982 		}
983 	}
984 	spin_unlock_bh(&ptp->evt_lock);
985 
986 	return rc;
987 }
988 
989 /* Process any queued receive events and corresponding packets
990  *
991  * q is returned with all the packets that are ready for delivery.
992  */
993 static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
994 {
995 	struct efx_ptp_data *ptp = efx->ptp_data;
996 	struct sk_buff *skb;
997 
998 	while ((skb = skb_dequeue(&ptp->rxq))) {
999 		struct efx_ptp_match *match;
1000 
1001 		match = (struct efx_ptp_match *)skb->cb;
1002 		if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) {
1003 			__skb_queue_tail(q, skb);
1004 		} else if (efx_ptp_match_rx(efx, skb) ==
1005 			   PTP_PACKET_STATE_MATCHED) {
1006 			__skb_queue_tail(q, skb);
1007 		} else if (time_after(jiffies, match->expiry)) {
1008 			match->state = PTP_PACKET_STATE_TIMED_OUT;
1009 			++ptp->rx_no_timestamp;
1010 			__skb_queue_tail(q, skb);
1011 		} else {
1012 			/* Replace unprocessed entry and stop */
1013 			skb_queue_head(&ptp->rxq, skb);
1014 			break;
1015 		}
1016 	}
1017 }
1018 
1019 /* Complete processing of a received packet */
1020 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1021 {
1022 	local_bh_disable();
1023 	netif_receive_skb(skb);
1024 	local_bh_enable();
1025 }
1026 
1027 static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
1028 {
1029 	struct efx_ptp_data *ptp = efx->ptp_data;
1030 
1031 	if (ptp->rxfilter_installed) {
1032 		efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1033 					  ptp->rxfilter_general);
1034 		efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1035 					  ptp->rxfilter_event);
1036 		ptp->rxfilter_installed = false;
1037 	}
1038 }
1039 
1040 static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1041 {
1042 	struct efx_ptp_data *ptp = efx->ptp_data;
1043 	struct efx_filter_spec rxfilter;
1044 	int rc;
1045 
1046 	if (!ptp->channel || ptp->rxfilter_installed)
1047 		return 0;
1048 
1049 	/* Must filter on both event and general ports to ensure
1050 	 * that there is no packet re-ordering.
1051 	 */
1052 	efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1053 			   efx_rx_queue_index(
1054 				   efx_channel_get_rx_queue(ptp->channel)));
1055 	rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1056 				       htonl(PTP_ADDRESS),
1057 				       htons(PTP_EVENT_PORT));
1058 	if (rc != 0)
1059 		return rc;
1060 
1061 	rc = efx_filter_insert_filter(efx, &rxfilter, true);
1062 	if (rc < 0)
1063 		return rc;
1064 	ptp->rxfilter_event = rc;
1065 
1066 	efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1067 			   efx_rx_queue_index(
1068 				   efx_channel_get_rx_queue(ptp->channel)));
1069 	rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1070 				       htonl(PTP_ADDRESS),
1071 				       htons(PTP_GENERAL_PORT));
1072 	if (rc != 0)
1073 		goto fail;
1074 
1075 	rc = efx_filter_insert_filter(efx, &rxfilter, true);
1076 	if (rc < 0)
1077 		goto fail;
1078 	ptp->rxfilter_general = rc;
1079 
1080 	ptp->rxfilter_installed = true;
1081 	return 0;
1082 
1083 fail:
1084 	efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1085 				  ptp->rxfilter_event);
1086 	return rc;
1087 }
1088 
1089 static int efx_ptp_start(struct efx_nic *efx)
1090 {
1091 	struct efx_ptp_data *ptp = efx->ptp_data;
1092 	int rc;
1093 
1094 	ptp->reset_required = false;
1095 
1096 	rc = efx_ptp_insert_multicast_filters(efx);
1097 	if (rc)
1098 		return rc;
1099 
1100 	rc = efx_ptp_enable(efx);
1101 	if (rc != 0)
1102 		goto fail;
1103 
1104 	ptp->evt_frag_idx = 0;
1105 	ptp->current_adjfreq = 0;
1106 
1107 	return 0;
1108 
1109 fail:
1110 	efx_ptp_remove_multicast_filters(efx);
1111 	return rc;
1112 }
1113 
1114 static int efx_ptp_stop(struct efx_nic *efx)
1115 {
1116 	struct efx_ptp_data *ptp = efx->ptp_data;
1117 	struct list_head *cursor;
1118 	struct list_head *next;
1119 	int rc;
1120 
1121 	if (ptp == NULL)
1122 		return 0;
1123 
1124 	rc = efx_ptp_disable(efx);
1125 
1126 	efx_ptp_remove_multicast_filters(efx);
1127 
1128 	/* Make sure RX packets are really delivered */
1129 	efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq);
1130 	skb_queue_purge(&efx->ptp_data->txq);
1131 
1132 	/* Drop any pending receive events */
1133 	spin_lock_bh(&efx->ptp_data->evt_lock);
1134 	list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) {
1135 		list_move(cursor, &efx->ptp_data->evt_free_list);
1136 	}
1137 	spin_unlock_bh(&efx->ptp_data->evt_lock);
1138 
1139 	return rc;
1140 }
1141 
1142 static int efx_ptp_restart(struct efx_nic *efx)
1143 {
1144 	if (efx->ptp_data && efx->ptp_data->enabled)
1145 		return efx_ptp_start(efx);
1146 	return 0;
1147 }
1148 
1149 static void efx_ptp_pps_worker(struct work_struct *work)
1150 {
1151 	struct efx_ptp_data *ptp =
1152 		container_of(work, struct efx_ptp_data, pps_work);
1153 	struct efx_nic *efx = ptp->efx;
1154 	struct ptp_clock_event ptp_evt;
1155 
1156 	if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
1157 		return;
1158 
1159 	ptp_evt.type = PTP_CLOCK_PPSUSR;
1160 	ptp_evt.pps_times = ptp->host_time_pps;
1161 	ptp_clock_event(ptp->phc_clock, &ptp_evt);
1162 }
1163 
1164 static void efx_ptp_worker(struct work_struct *work)
1165 {
1166 	struct efx_ptp_data *ptp_data =
1167 		container_of(work, struct efx_ptp_data, work);
1168 	struct efx_nic *efx = ptp_data->efx;
1169 	struct sk_buff *skb;
1170 	struct sk_buff_head tempq;
1171 
1172 	if (ptp_data->reset_required) {
1173 		efx_ptp_stop(efx);
1174 		efx_ptp_start(efx);
1175 		return;
1176 	}
1177 
1178 	efx_ptp_drop_time_expired_events(efx);
1179 
1180 	__skb_queue_head_init(&tempq);
1181 	efx_ptp_process_events(efx, &tempq);
1182 
1183 	while ((skb = skb_dequeue(&ptp_data->txq)))
1184 		efx_ptp_xmit_skb(efx, skb);
1185 
1186 	while ((skb = __skb_dequeue(&tempq)))
1187 		efx_ptp_process_rx(efx, skb);
1188 }
1189 
1190 static const struct ptp_clock_info efx_phc_clock_info = {
1191 	.owner		= THIS_MODULE,
1192 	.name		= "sfc",
1193 	.max_adj	= MAX_PPB,
1194 	.n_alarm	= 0,
1195 	.n_ext_ts	= 0,
1196 	.n_per_out	= 0,
1197 	.n_pins		= 0,
1198 	.pps		= 1,
1199 	.adjfreq	= efx_phc_adjfreq,
1200 	.adjtime	= efx_phc_adjtime,
1201 	.gettime64	= efx_phc_gettime,
1202 	.settime64	= efx_phc_settime,
1203 	.enable		= efx_phc_enable,
1204 };
1205 
1206 /* Initialise PTP state. */
1207 int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1208 {
1209 	struct efx_ptp_data *ptp;
1210 	int rc = 0;
1211 	unsigned int pos;
1212 
1213 	ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL);
1214 	efx->ptp_data = ptp;
1215 	if (!efx->ptp_data)
1216 		return -ENOMEM;
1217 
1218 	ptp->efx = efx;
1219 	ptp->channel = channel;
1220 	ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1221 
1222 	rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL);
1223 	if (rc != 0)
1224 		goto fail1;
1225 
1226 	skb_queue_head_init(&ptp->rxq);
1227 	skb_queue_head_init(&ptp->txq);
1228 	ptp->workwq = create_singlethread_workqueue("sfc_ptp");
1229 	if (!ptp->workwq) {
1230 		rc = -ENOMEM;
1231 		goto fail2;
1232 	}
1233 
1234 	INIT_WORK(&ptp->work, efx_ptp_worker);
1235 	ptp->config.flags = 0;
1236 	ptp->config.tx_type = HWTSTAMP_TX_OFF;
1237 	ptp->config.rx_filter = HWTSTAMP_FILTER_NONE;
1238 	INIT_LIST_HEAD(&ptp->evt_list);
1239 	INIT_LIST_HEAD(&ptp->evt_free_list);
1240 	spin_lock_init(&ptp->evt_lock);
1241 	for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++)
1242 		list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list);
1243 
1244 	/* Get the NIC PTP attributes and set up time conversions */
1245 	rc = efx_ptp_get_attributes(efx);
1246 	if (rc < 0)
1247 		goto fail3;
1248 
1249 	/* Get the timestamp corrections */
1250 	rc = efx_ptp_get_timestamp_corrections(efx);
1251 	if (rc < 0)
1252 		goto fail3;
1253 
1254 	if (efx->mcdi->fn_flags &
1255 	    (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) {
1256 		ptp->phc_clock_info = efx_phc_clock_info;
1257 		ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info,
1258 						    &efx->pci_dev->dev);
1259 		if (IS_ERR(ptp->phc_clock)) {
1260 			rc = PTR_ERR(ptp->phc_clock);
1261 			goto fail3;
1262 		}
1263 
1264 		INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1265 		ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
1266 		if (!ptp->pps_workwq) {
1267 			rc = -ENOMEM;
1268 			goto fail4;
1269 		}
1270 	}
1271 	ptp->nic_ts_enabled = false;
1272 
1273 	return 0;
1274 fail4:
1275 	ptp_clock_unregister(efx->ptp_data->phc_clock);
1276 
1277 fail3:
1278 	destroy_workqueue(efx->ptp_data->workwq);
1279 
1280 fail2:
1281 	efx_nic_free_buffer(efx, &ptp->start);
1282 
1283 fail1:
1284 	kfree(efx->ptp_data);
1285 	efx->ptp_data = NULL;
1286 
1287 	return rc;
1288 }
1289 
1290 /* Initialise PTP channel.
1291  *
1292  * Setting core_index to zero causes the queue to be initialised and doesn't
1293  * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
1294  */
1295 static int efx_ptp_probe_channel(struct efx_channel *channel)
1296 {
1297 	struct efx_nic *efx = channel->efx;
1298 
1299 	channel->irq_moderation = 0;
1300 	channel->rx_queue.core_index = 0;
1301 
1302 	return efx_ptp_probe(efx, channel);
1303 }
1304 
1305 void efx_ptp_remove(struct efx_nic *efx)
1306 {
1307 	if (!efx->ptp_data)
1308 		return;
1309 
1310 	(void)efx_ptp_disable(efx);
1311 
1312 	cancel_work_sync(&efx->ptp_data->work);
1313 	cancel_work_sync(&efx->ptp_data->pps_work);
1314 
1315 	skb_queue_purge(&efx->ptp_data->rxq);
1316 	skb_queue_purge(&efx->ptp_data->txq);
1317 
1318 	if (efx->ptp_data->phc_clock) {
1319 		destroy_workqueue(efx->ptp_data->pps_workwq);
1320 		ptp_clock_unregister(efx->ptp_data->phc_clock);
1321 	}
1322 
1323 	destroy_workqueue(efx->ptp_data->workwq);
1324 
1325 	efx_nic_free_buffer(efx, &efx->ptp_data->start);
1326 	kfree(efx->ptp_data);
1327 }
1328 
1329 static void efx_ptp_remove_channel(struct efx_channel *channel)
1330 {
1331 	efx_ptp_remove(channel->efx);
1332 }
1333 
1334 static void efx_ptp_get_channel_name(struct efx_channel *channel,
1335 				     char *buf, size_t len)
1336 {
1337 	snprintf(buf, len, "%s-ptp", channel->efx->name);
1338 }
1339 
1340 /* Determine whether this packet should be processed by the PTP module
1341  * or transmitted conventionally.
1342  */
1343 bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1344 {
1345 	return efx->ptp_data &&
1346 		efx->ptp_data->enabled &&
1347 		skb->len >= PTP_MIN_LENGTH &&
1348 		skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1349 		likely(skb->protocol == htons(ETH_P_IP)) &&
1350 		skb_transport_header_was_set(skb) &&
1351 		skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1352 		ip_hdr(skb)->protocol == IPPROTO_UDP &&
1353 		skb_headlen(skb) >=
1354 		skb_transport_offset(skb) + sizeof(struct udphdr) &&
1355 		udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1356 }
1357 
1358 /* Receive a PTP packet.  Packets are queued until the arrival of
1359  * the receive timestamp from the MC - this will probably occur after the
1360  * packet arrival because of the processing in the MC.
1361  */
1362 static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1363 {
1364 	struct efx_nic *efx = channel->efx;
1365 	struct efx_ptp_data *ptp = efx->ptp_data;
1366 	struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1367 	u8 *match_data_012, *match_data_345;
1368 	unsigned int version;
1369 	u8 *data;
1370 
1371 	match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1372 
1373 	/* Correct version? */
1374 	if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1375 		if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1376 			return false;
1377 		}
1378 		data = skb->data;
1379 		version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1380 		if (version != PTP_VERSION_V1) {
1381 			return false;
1382 		}
1383 
1384 		/* PTP V1 uses all six bytes of the UUID to match the packet
1385 		 * to the timestamp
1386 		 */
1387 		match_data_012 = data + PTP_V1_UUID_OFFSET;
1388 		match_data_345 = data + PTP_V1_UUID_OFFSET + 3;
1389 	} else {
1390 		if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1391 			return false;
1392 		}
1393 		data = skb->data;
1394 		version = data[PTP_V2_VERSION_OFFSET];
1395 		if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1396 			return false;
1397 		}
1398 
1399 		/* The original V2 implementation uses bytes 2-7 of
1400 		 * the UUID to match the packet to the timestamp. This
1401 		 * discards two of the bytes of the MAC address used
1402 		 * to create the UUID (SF bug 33070).  The PTP V2
1403 		 * enhanced mode fixes this issue and uses bytes 0-2
1404 		 * and byte 5-7 of the UUID.
1405 		 */
1406 		match_data_345 = data + PTP_V2_UUID_OFFSET + 5;
1407 		if (ptp->mode == MC_CMD_PTP_MODE_V2) {
1408 			match_data_012 = data + PTP_V2_UUID_OFFSET + 2;
1409 		} else {
1410 			match_data_012 = data + PTP_V2_UUID_OFFSET + 0;
1411 			BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED);
1412 		}
1413 	}
1414 
1415 	/* Does this packet require timestamping? */
1416 	if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1417 		match->state = PTP_PACKET_STATE_UNMATCHED;
1418 
1419 		/* We expect the sequence number to be in the same position in
1420 		 * the packet for PTP V1 and V2
1421 		 */
1422 		BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1423 		BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1424 
1425 		/* Extract UUID/Sequence information */
1426 		match->words[0] = (match_data_012[0]         |
1427 				   (match_data_012[1] << 8)  |
1428 				   (match_data_012[2] << 16) |
1429 				   (match_data_345[0] << 24));
1430 		match->words[1] = (match_data_345[1]         |
1431 				   (match_data_345[2] << 8)  |
1432 				   (data[PTP_V1_SEQUENCE_OFFSET +
1433 					 PTP_V1_SEQUENCE_LENGTH - 1] <<
1434 				    16));
1435 	} else {
1436 		match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1437 	}
1438 
1439 	skb_queue_tail(&ptp->rxq, skb);
1440 	queue_work(ptp->workwq, &ptp->work);
1441 
1442 	return true;
1443 }
1444 
1445 /* Transmit a PTP packet.  This has to be transmitted by the MC
1446  * itself, through an MCDI call.  MCDI calls aren't permitted
1447  * in the transmit path so defer the actual transmission to a suitable worker.
1448  */
1449 int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1450 {
1451 	struct efx_ptp_data *ptp = efx->ptp_data;
1452 
1453 	skb_queue_tail(&ptp->txq, skb);
1454 
1455 	if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1456 	    (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1457 		efx_xmit_hwtstamp_pending(skb);
1458 	queue_work(ptp->workwq, &ptp->work);
1459 
1460 	return NETDEV_TX_OK;
1461 }
1462 
1463 int efx_ptp_get_mode(struct efx_nic *efx)
1464 {
1465 	return efx->ptp_data->mode;
1466 }
1467 
1468 int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1469 			unsigned int new_mode)
1470 {
1471 	if ((enable_wanted != efx->ptp_data->enabled) ||
1472 	    (enable_wanted && (efx->ptp_data->mode != new_mode))) {
1473 		int rc = 0;
1474 
1475 		if (enable_wanted) {
1476 			/* Change of mode requires disable */
1477 			if (efx->ptp_data->enabled &&
1478 			    (efx->ptp_data->mode != new_mode)) {
1479 				efx->ptp_data->enabled = false;
1480 				rc = efx_ptp_stop(efx);
1481 				if (rc != 0)
1482 					return rc;
1483 			}
1484 
1485 			/* Set new operating mode and establish
1486 			 * baseline synchronisation, which must
1487 			 * succeed.
1488 			 */
1489 			efx->ptp_data->mode = new_mode;
1490 			if (netif_running(efx->net_dev))
1491 				rc = efx_ptp_start(efx);
1492 			if (rc == 0) {
1493 				rc = efx_ptp_synchronize(efx,
1494 							 PTP_SYNC_ATTEMPTS * 2);
1495 				if (rc != 0)
1496 					efx_ptp_stop(efx);
1497 			}
1498 		} else {
1499 			rc = efx_ptp_stop(efx);
1500 		}
1501 
1502 		if (rc != 0)
1503 			return rc;
1504 
1505 		efx->ptp_data->enabled = enable_wanted;
1506 	}
1507 
1508 	return 0;
1509 }
1510 
1511 static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
1512 {
1513 	int rc;
1514 
1515 	if (init->flags)
1516 		return -EINVAL;
1517 
1518 	if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1519 	    (init->tx_type != HWTSTAMP_TX_ON))
1520 		return -ERANGE;
1521 
1522 	rc = efx->type->ptp_set_ts_config(efx, init);
1523 	if (rc)
1524 		return rc;
1525 
1526 	efx->ptp_data->config = *init;
1527 	return 0;
1528 }
1529 
1530 void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
1531 {
1532 	struct efx_ptp_data *ptp = efx->ptp_data;
1533 	struct efx_nic *primary = efx->primary;
1534 
1535 	ASSERT_RTNL();
1536 
1537 	if (!ptp)
1538 		return;
1539 
1540 	ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1541 				     SOF_TIMESTAMPING_RX_HARDWARE |
1542 				     SOF_TIMESTAMPING_RAW_HARDWARE);
1543 	if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1544 		ts_info->phc_index =
1545 			ptp_clock_index(primary->ptp_data->phc_clock);
1546 	ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1547 	ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1548 }
1549 
1550 int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1551 {
1552 	struct hwtstamp_config config;
1553 	int rc;
1554 
1555 	/* Not a PTP enabled port */
1556 	if (!efx->ptp_data)
1557 		return -EOPNOTSUPP;
1558 
1559 	if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
1560 		return -EFAULT;
1561 
1562 	rc = efx_ptp_ts_init(efx, &config);
1563 	if (rc != 0)
1564 		return rc;
1565 
1566 	return copy_to_user(ifr->ifr_data, &config, sizeof(config))
1567 		? -EFAULT : 0;
1568 }
1569 
1570 int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1571 {
1572 	if (!efx->ptp_data)
1573 		return -EOPNOTSUPP;
1574 
1575 	return copy_to_user(ifr->ifr_data, &efx->ptp_data->config,
1576 			    sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
1577 }
1578 
1579 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1580 {
1581 	struct efx_ptp_data *ptp = efx->ptp_data;
1582 
1583 	netif_err(efx, hw, efx->net_dev,
1584 		"PTP unexpected event length: got %d expected %d\n",
1585 		ptp->evt_frag_idx, expected_frag_len);
1586 	ptp->reset_required = true;
1587 	queue_work(ptp->workwq, &ptp->work);
1588 }
1589 
1590 /* Process a completed receive event.  Put it on the event queue and
1591  * start worker thread.  This is required because event and their
1592  * correspoding packets may come in either order.
1593  */
1594 static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
1595 {
1596 	struct efx_ptp_event_rx *evt = NULL;
1597 
1598 	if (WARN_ON_ONCE(ptp->rx_ts_inline))
1599 		return;
1600 
1601 	if (ptp->evt_frag_idx != 3) {
1602 		ptp_event_failure(efx, 3);
1603 		return;
1604 	}
1605 
1606 	spin_lock_bh(&ptp->evt_lock);
1607 	if (!list_empty(&ptp->evt_free_list)) {
1608 		evt = list_first_entry(&ptp->evt_free_list,
1609 				       struct efx_ptp_event_rx, link);
1610 		list_del(&evt->link);
1611 
1612 		evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA);
1613 		evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2],
1614 					     MCDI_EVENT_SRC)        |
1615 			     (EFX_QWORD_FIELD(ptp->evt_frags[1],
1616 					      MCDI_EVENT_SRC) << 8) |
1617 			     (EFX_QWORD_FIELD(ptp->evt_frags[0],
1618 					      MCDI_EVENT_SRC) << 16));
1619 		evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
1620 			EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
1621 			EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
1622 			ptp->ts_corrections.rx);
1623 		evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1624 		list_add_tail(&evt->link, &ptp->evt_list);
1625 
1626 		queue_work(ptp->workwq, &ptp->work);
1627 	} else if (net_ratelimit()) {
1628 		/* Log a rate-limited warning message. */
1629 		netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n");
1630 	}
1631 	spin_unlock_bh(&ptp->evt_lock);
1632 }
1633 
1634 static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1635 {
1636 	int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1637 	if (ptp->evt_frag_idx != 1) {
1638 		ptp_event_failure(efx, 1);
1639 		return;
1640 	}
1641 
1642 	netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1643 }
1644 
1645 static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1646 {
1647 	if (ptp->nic_ts_enabled)
1648 		queue_work(ptp->pps_workwq, &ptp->pps_work);
1649 }
1650 
1651 void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1652 {
1653 	struct efx_ptp_data *ptp = efx->ptp_data;
1654 	int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
1655 
1656 	if (!ptp) {
1657 		if (net_ratelimit())
1658 			netif_warn(efx, drv, efx->net_dev,
1659 				   "Received PTP event but PTP not set up\n");
1660 		return;
1661 	}
1662 
1663 	if (!ptp->enabled)
1664 		return;
1665 
1666 	if (ptp->evt_frag_idx == 0) {
1667 		ptp->evt_code = code;
1668 	} else if (ptp->evt_code != code) {
1669 		netif_err(efx, hw, efx->net_dev,
1670 			  "PTP out of sequence event %d\n", code);
1671 		ptp->evt_frag_idx = 0;
1672 	}
1673 
1674 	ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
1675 	if (!MCDI_EVENT_FIELD(*ev, CONT)) {
1676 		/* Process resulting event */
1677 		switch (code) {
1678 		case MCDI_EVENT_CODE_PTP_RX:
1679 			ptp_event_rx(efx, ptp);
1680 			break;
1681 		case MCDI_EVENT_CODE_PTP_FAULT:
1682 			ptp_event_fault(efx, ptp);
1683 			break;
1684 		case MCDI_EVENT_CODE_PTP_PPS:
1685 			ptp_event_pps(efx, ptp);
1686 			break;
1687 		default:
1688 			netif_err(efx, hw, efx->net_dev,
1689 				  "PTP unknown event %d\n", code);
1690 			break;
1691 		}
1692 		ptp->evt_frag_idx = 0;
1693 	} else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
1694 		netif_err(efx, hw, efx->net_dev,
1695 			  "PTP too many event fragments\n");
1696 		ptp->evt_frag_idx = 0;
1697 	}
1698 }
1699 
1700 void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
1701 {
1702 	channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
1703 	channel->sync_timestamp_minor =
1704 		MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_26_19) << 19;
1705 	/* if sync events have been disabled then we want to silently ignore
1706 	 * this event, so throw away result.
1707 	 */
1708 	(void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
1709 		       SYNC_EVENTS_VALID);
1710 }
1711 
1712 /* make some assumptions about the time representation rather than abstract it,
1713  * since we currently only support one type of inline timestamping and only on
1714  * EF10.
1715  */
1716 #define MINOR_TICKS_PER_SECOND 0x8000000
1717 /* Fuzz factor for sync events to be out of order with RX events */
1718 #define FUZZ (MINOR_TICKS_PER_SECOND / 10)
1719 #define EXPECTED_SYNC_EVENTS_PER_SECOND 4
1720 
1721 static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
1722 {
1723 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
1724 	return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset));
1725 #else
1726 	const u8 *data = eh + efx->rx_packet_ts_offset;
1727 	return (u32)data[0]       |
1728 	       (u32)data[1] << 8  |
1729 	       (u32)data[2] << 16 |
1730 	       (u32)data[3] << 24;
1731 #endif
1732 }
1733 
1734 void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
1735 				   struct sk_buff *skb)
1736 {
1737 	struct efx_nic *efx = channel->efx;
1738 	u32 pkt_timestamp_major, pkt_timestamp_minor;
1739 	u32 diff, carry;
1740 	struct skb_shared_hwtstamps *timestamps;
1741 
1742 	pkt_timestamp_minor = (efx_rx_buf_timestamp_minor(efx,
1743 							  skb_mac_header(skb)) +
1744 			       (u32) efx->ptp_data->ts_corrections.rx) &
1745 			      (MINOR_TICKS_PER_SECOND - 1);
1746 
1747 	/* get the difference between the packet and sync timestamps,
1748 	 * modulo one second
1749 	 */
1750 	diff = (pkt_timestamp_minor - channel->sync_timestamp_minor) &
1751 		(MINOR_TICKS_PER_SECOND - 1);
1752 	/* do we roll over a second boundary and need to carry the one? */
1753 	carry = channel->sync_timestamp_minor + diff > MINOR_TICKS_PER_SECOND ?
1754 		1 : 0;
1755 
1756 	if (diff <= MINOR_TICKS_PER_SECOND / EXPECTED_SYNC_EVENTS_PER_SECOND +
1757 		    FUZZ) {
1758 		/* packet is ahead of the sync event by a quarter of a second or
1759 		 * less (allowing for fuzz)
1760 		 */
1761 		pkt_timestamp_major = channel->sync_timestamp_major + carry;
1762 	} else if (diff >= MINOR_TICKS_PER_SECOND - FUZZ) {
1763 		/* packet is behind the sync event but within the fuzz factor.
1764 		 * This means the RX packet and sync event crossed as they were
1765 		 * placed on the event queue, which can sometimes happen.
1766 		 */
1767 		pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
1768 	} else {
1769 		/* it's outside tolerance in both directions. this might be
1770 		 * indicative of us missing sync events for some reason, so
1771 		 * we'll call it an error rather than risk giving a bogus
1772 		 * timestamp.
1773 		 */
1774 		netif_vdbg(efx, drv, efx->net_dev,
1775 			  "packet timestamp %x too far from sync event %x:%x\n",
1776 			  pkt_timestamp_minor, channel->sync_timestamp_major,
1777 			  channel->sync_timestamp_minor);
1778 		return;
1779 	}
1780 
1781 	/* attach the timestamps to the skb */
1782 	timestamps = skb_hwtstamps(skb);
1783 	timestamps->hwtstamp =
1784 		efx_ptp_s27_to_ktime(pkt_timestamp_major, pkt_timestamp_minor);
1785 }
1786 
1787 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
1788 {
1789 	struct efx_ptp_data *ptp_data = container_of(ptp,
1790 						     struct efx_ptp_data,
1791 						     phc_clock_info);
1792 	struct efx_nic *efx = ptp_data->efx;
1793 	MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
1794 	s64 adjustment_ns;
1795 	int rc;
1796 
1797 	if (delta > MAX_PPB)
1798 		delta = MAX_PPB;
1799 	else if (delta < -MAX_PPB)
1800 		delta = -MAX_PPB;
1801 
1802 	/* Convert ppb to fixed point ns. */
1803 	adjustment_ns = (((s64)delta * PPB_SCALE_WORD) >>
1804 			 (PPB_EXTRA_BITS + MAX_PPB_BITS));
1805 
1806 	MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
1807 	MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
1808 	MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
1809 	MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
1810 	MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
1811 	rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj),
1812 			  NULL, 0, NULL);
1813 	if (rc != 0)
1814 		return rc;
1815 
1816 	ptp_data->current_adjfreq = adjustment_ns;
1817 	return 0;
1818 }
1819 
1820 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
1821 {
1822 	u32 nic_major, nic_minor;
1823 	struct efx_ptp_data *ptp_data = container_of(ptp,
1824 						     struct efx_ptp_data,
1825 						     phc_clock_info);
1826 	struct efx_nic *efx = ptp_data->efx;
1827 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
1828 
1829 	efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
1830 
1831 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
1832 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
1833 	MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
1834 	MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
1835 	MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
1836 	return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
1837 			    NULL, 0, NULL);
1838 }
1839 
1840 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
1841 {
1842 	struct efx_ptp_data *ptp_data = container_of(ptp,
1843 						     struct efx_ptp_data,
1844 						     phc_clock_info);
1845 	struct efx_nic *efx = ptp_data->efx;
1846 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
1847 	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
1848 	int rc;
1849 	ktime_t kt;
1850 
1851 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
1852 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
1853 
1854 	rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
1855 			  outbuf, sizeof(outbuf), NULL);
1856 	if (rc != 0)
1857 		return rc;
1858 
1859 	kt = ptp_data->nic_to_kernel_time(
1860 		MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
1861 		MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
1862 	*ts = ktime_to_timespec64(kt);
1863 	return 0;
1864 }
1865 
1866 static int efx_phc_settime(struct ptp_clock_info *ptp,
1867 			   const struct timespec64 *e_ts)
1868 {
1869 	/* Get the current NIC time, efx_phc_gettime.
1870 	 * Subtract from the desired time to get the offset
1871 	 * call efx_phc_adjtime with the offset
1872 	 */
1873 	int rc;
1874 	struct timespec64 time_now;
1875 	struct timespec64 delta;
1876 
1877 	rc = efx_phc_gettime(ptp, &time_now);
1878 	if (rc != 0)
1879 		return rc;
1880 
1881 	delta = timespec64_sub(*e_ts, time_now);
1882 
1883 	rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta));
1884 	if (rc != 0)
1885 		return rc;
1886 
1887 	return 0;
1888 }
1889 
1890 static int efx_phc_enable(struct ptp_clock_info *ptp,
1891 			  struct ptp_clock_request *request,
1892 			  int enable)
1893 {
1894 	struct efx_ptp_data *ptp_data = container_of(ptp,
1895 						     struct efx_ptp_data,
1896 						     phc_clock_info);
1897 	if (request->type != PTP_CLK_REQ_PPS)
1898 		return -EOPNOTSUPP;
1899 
1900 	ptp_data->nic_ts_enabled = !!enable;
1901 	return 0;
1902 }
1903 
1904 static const struct efx_channel_type efx_ptp_channel_type = {
1905 	.handle_no_channel	= efx_ptp_handle_no_channel,
1906 	.pre_probe		= efx_ptp_probe_channel,
1907 	.post_remove		= efx_ptp_remove_channel,
1908 	.get_name		= efx_ptp_get_channel_name,
1909 	/* no copy operation; there is no need to reallocate this channel */
1910 	.receive_skb		= efx_ptp_rx,
1911 	.keep_eventq		= false,
1912 };
1913 
1914 void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
1915 {
1916 	/* Check whether PTP is implemented on this NIC.  The DISABLE
1917 	 * operation will succeed if and only if it is implemented.
1918 	 */
1919 	if (efx_ptp_disable(efx) == 0)
1920 		efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
1921 			&efx_ptp_channel_type;
1922 }
1923 
1924 void efx_ptp_start_datapath(struct efx_nic *efx)
1925 {
1926 	if (efx_ptp_restart(efx))
1927 		netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
1928 	/* re-enable timestamping if it was previously enabled */
1929 	if (efx->type->ptp_set_ts_sync_events)
1930 		efx->type->ptp_set_ts_sync_events(efx, true, true);
1931 }
1932 
1933 void efx_ptp_stop_datapath(struct efx_nic *efx)
1934 {
1935 	/* temporarily disable timestamping */
1936 	if (efx->type->ptp_set_ts_sync_events)
1937 		efx->type->ptp_set_ts_sync_events(efx, false, true);
1938 	efx_ptp_stop(efx);
1939 }
1940