xref: /openbmc/linux/drivers/net/ethernet/sfc/ptp.c (revision 9cfc5c90)
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 		memset(outbuf, 0, sizeof(outbuf));
394 	efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
395 			     efx_ptp_stat_mask,
396 			     stats, _MCDI_PTR(outbuf, 0), false);
397 
398 	return PTP_STAT_COUNT;
399 }
400 
401 /* For Siena platforms NIC time is s and ns */
402 static void efx_ptp_ns_to_s_ns(s64 ns, u32 *nic_major, u32 *nic_minor)
403 {
404 	struct timespec64 ts = ns_to_timespec64(ns);
405 	*nic_major = (u32)ts.tv_sec;
406 	*nic_minor = ts.tv_nsec;
407 }
408 
409 static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
410 						s32 correction)
411 {
412 	ktime_t kt = ktime_set(nic_major, nic_minor);
413 	if (correction >= 0)
414 		kt = ktime_add_ns(kt, (u64)correction);
415 	else
416 		kt = ktime_sub_ns(kt, (u64)-correction);
417 	return kt;
418 }
419 
420 /* To convert from s27 format to ns we multiply then divide by a power of 2.
421  * For the conversion from ns to s27, the operation is also converted to a
422  * multiply and shift.
423  */
424 #define S27_TO_NS_SHIFT	(27)
425 #define NS_TO_S27_MULT	(((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC)
426 #define NS_TO_S27_SHIFT	(63 - S27_TO_NS_SHIFT)
427 #define S27_MINOR_MAX	(1 << S27_TO_NS_SHIFT)
428 
429 /* For Huntington platforms NIC time is in seconds and fractions of a second
430  * where the minor register only uses 27 bits in units of 2^-27s.
431  */
432 static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
433 {
434 	struct timespec64 ts = ns_to_timespec64(ns);
435 	u32 maj = (u32)ts.tv_sec;
436 	u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT +
437 			 (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT);
438 
439 	/* The conversion can result in the minor value exceeding the maximum.
440 	 * In this case, round up to the next second.
441 	 */
442 	if (min >= S27_MINOR_MAX) {
443 		min -= S27_MINOR_MAX;
444 		maj++;
445 	}
446 
447 	*nic_major = maj;
448 	*nic_minor = min;
449 }
450 
451 static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
452 {
453 	u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC +
454 			(1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT);
455 	return ktime_set(nic_major, ns);
456 }
457 
458 static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
459 					       s32 correction)
460 {
461 	/* Apply the correction and deal with carry */
462 	nic_minor += correction;
463 	if ((s32)nic_minor < 0) {
464 		nic_minor += S27_MINOR_MAX;
465 		nic_major--;
466 	} else if (nic_minor >= S27_MINOR_MAX) {
467 		nic_minor -= S27_MINOR_MAX;
468 		nic_major++;
469 	}
470 
471 	return efx_ptp_s27_to_ktime(nic_major, nic_minor);
472 }
473 
474 /* Get PTP attributes and set up time conversions */
475 static int efx_ptp_get_attributes(struct efx_nic *efx)
476 {
477 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN);
478 	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN);
479 	struct efx_ptp_data *ptp = efx->ptp_data;
480 	int rc;
481 	u32 fmt;
482 	size_t out_len;
483 
484 	/* Get the PTP attributes. If the NIC doesn't support the operation we
485 	 * use the default format for compatibility with older NICs i.e.
486 	 * seconds and nanoseconds.
487 	 */
488 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES);
489 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
490 	rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
491 				outbuf, sizeof(outbuf), &out_len);
492 	if (rc == 0) {
493 		fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT);
494 	} else if (rc == -EINVAL) {
495 		fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS;
496 	} else if (rc == -EPERM) {
497 		netif_info(efx, probe, efx->net_dev, "no PTP support\n");
498 		return rc;
499 	} else {
500 		efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf),
501 				       outbuf, sizeof(outbuf), rc);
502 		return rc;
503 	}
504 
505 	if (fmt == MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION) {
506 		ptp->ns_to_nic_time = efx_ptp_ns_to_s27;
507 		ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction;
508 	} else if (fmt == MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS) {
509 		ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns;
510 		ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction;
511 	} else {
512 		return -ERANGE;
513 	}
514 
515 	ptp->time_format = fmt;
516 
517 	/* MC_CMD_PTP_OP_GET_ATTRIBUTES is an extended version of an older
518 	 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT that also returns a value
519 	 * to use for the minimum acceptable corrected synchronization window.
520 	 * If we have the extra information store it. For older firmware that
521 	 * does not implement the extended command use the default value.
522 	 */
523 	if (rc == 0 && out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
524 		ptp->min_synchronisation_ns =
525 			MCDI_DWORD(outbuf,
526 				   PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
527 	else
528 		ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
529 
530 	return 0;
531 }
532 
533 /* Get PTP timestamp corrections */
534 static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
535 {
536 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN);
537 	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_LEN);
538 	int rc;
539 
540 	/* Get the timestamp corrections from the NIC. If this operation is
541 	 * not supported (older NICs) then no correction is required.
542 	 */
543 	MCDI_SET_DWORD(inbuf, PTP_IN_OP,
544 		       MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS);
545 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
546 
547 	rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
548 				outbuf, sizeof(outbuf), NULL);
549 	if (rc == 0) {
550 		efx->ptp_data->ts_corrections.tx = MCDI_DWORD(outbuf,
551 			PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT);
552 		efx->ptp_data->ts_corrections.rx = MCDI_DWORD(outbuf,
553 			PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE);
554 		efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf,
555 			PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT);
556 		efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf,
557 			PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN);
558 	} else if (rc == -EINVAL) {
559 		efx->ptp_data->ts_corrections.tx = 0;
560 		efx->ptp_data->ts_corrections.rx = 0;
561 		efx->ptp_data->ts_corrections.pps_out = 0;
562 		efx->ptp_data->ts_corrections.pps_in = 0;
563 	} else {
564 		efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), outbuf,
565 				       sizeof(outbuf), rc);
566 		return rc;
567 	}
568 
569 	return 0;
570 }
571 
572 /* Enable MCDI PTP support. */
573 static int efx_ptp_enable(struct efx_nic *efx)
574 {
575 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
576 	MCDI_DECLARE_BUF_ERR(outbuf);
577 	int rc;
578 
579 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE);
580 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
581 	MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE,
582 		       efx->ptp_data->channel ?
583 		       efx->ptp_data->channel->channel : 0);
584 	MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode);
585 
586 	rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
587 				outbuf, sizeof(outbuf), NULL);
588 	rc = (rc == -EALREADY) ? 0 : rc;
589 	if (rc)
590 		efx_mcdi_display_error(efx, MC_CMD_PTP,
591 				       MC_CMD_PTP_IN_ENABLE_LEN,
592 				       outbuf, sizeof(outbuf), rc);
593 	return rc;
594 }
595 
596 /* Disable MCDI PTP support.
597  *
598  * Note that this function should never rely on the presence of ptp_data -
599  * may be called before that exists.
600  */
601 static int efx_ptp_disable(struct efx_nic *efx)
602 {
603 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
604 	MCDI_DECLARE_BUF_ERR(outbuf);
605 	int rc;
606 
607 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE);
608 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
609 	rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
610 				outbuf, sizeof(outbuf), NULL);
611 	rc = (rc == -EALREADY) ? 0 : rc;
612 	/* If we get ENOSYS, the NIC doesn't support PTP, and thus this function
613 	 * should only have been called during probe.
614 	 */
615 	if (rc == -ENOSYS || rc == -EPERM)
616 		netif_info(efx, probe, efx->net_dev, "no PTP support\n");
617 	else if (rc)
618 		efx_mcdi_display_error(efx, MC_CMD_PTP,
619 				       MC_CMD_PTP_IN_DISABLE_LEN,
620 				       outbuf, sizeof(outbuf), rc);
621 	return rc;
622 }
623 
624 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
625 {
626 	struct sk_buff *skb;
627 
628 	while ((skb = skb_dequeue(q))) {
629 		local_bh_disable();
630 		netif_receive_skb(skb);
631 		local_bh_enable();
632 	}
633 }
634 
635 static void efx_ptp_handle_no_channel(struct efx_nic *efx)
636 {
637 	netif_err(efx, drv, efx->net_dev,
638 		  "ERROR: PTP requires MSI-X and 1 additional interrupt"
639 		  "vector. PTP disabled\n");
640 }
641 
642 /* Repeatedly send the host time to the MC which will capture the hardware
643  * time.
644  */
645 static void efx_ptp_send_times(struct efx_nic *efx,
646 			       struct pps_event_time *last_time)
647 {
648 	struct pps_event_time now;
649 	struct timespec64 limit;
650 	struct efx_ptp_data *ptp = efx->ptp_data;
651 	struct timespec64 start;
652 	int *mc_running = ptp->start.addr;
653 
654 	pps_get_ts(&now);
655 	start = now.ts_real;
656 	limit = now.ts_real;
657 	timespec64_add_ns(&limit, SYNCHRONISE_PERIOD_NS);
658 
659 	/* Write host time for specified period or until MC is done */
660 	while ((timespec64_compare(&now.ts_real, &limit) < 0) &&
661 	       ACCESS_ONCE(*mc_running)) {
662 		struct timespec64 update_time;
663 		unsigned int host_time;
664 
665 		/* Don't update continuously to avoid saturating the PCIe bus */
666 		update_time = now.ts_real;
667 		timespec64_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS);
668 		do {
669 			pps_get_ts(&now);
670 		} while ((timespec64_compare(&now.ts_real, &update_time) < 0) &&
671 			 ACCESS_ONCE(*mc_running));
672 
673 		/* Synchronise NIC with single word of time only */
674 		host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS |
675 			     now.ts_real.tv_nsec);
676 		/* Update host time in NIC memory */
677 		efx->type->ptp_write_host_time(efx, host_time);
678 	}
679 	*last_time = now;
680 }
681 
682 /* Read a timeset from the MC's results and partial process. */
683 static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data),
684 				 struct efx_ptp_timeset *timeset)
685 {
686 	unsigned start_ns, end_ns;
687 
688 	timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART);
689 	timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR);
690 	timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR);
691 	timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND),
692 	timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
693 
694 	/* Ignore seconds */
695 	start_ns = timeset->host_start & MC_NANOSECOND_MASK;
696 	end_ns = timeset->host_end & MC_NANOSECOND_MASK;
697 	/* Allow for rollover */
698 	if (end_ns < start_ns)
699 		end_ns += NSEC_PER_SEC;
700 	/* Determine duration of operation */
701 	timeset->window = end_ns - start_ns;
702 }
703 
704 /* Process times received from MC.
705  *
706  * Extract times from returned results, and establish the minimum value
707  * seen.  The minimum value represents the "best" possible time and events
708  * too much greater than this are rejected - the machine is, perhaps, too
709  * busy. A number of readings are taken so that, hopefully, at least one good
710  * synchronisation will be seen in the results.
711  */
712 static int
713 efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf),
714 		      size_t response_length,
715 		      const struct pps_event_time *last_time)
716 {
717 	unsigned number_readings =
718 		MCDI_VAR_ARRAY_LEN(response_length,
719 				   PTP_OUT_SYNCHRONIZE_TIMESET);
720 	unsigned i;
721 	unsigned ngood = 0;
722 	unsigned last_good = 0;
723 	struct efx_ptp_data *ptp = efx->ptp_data;
724 	u32 last_sec;
725 	u32 start_sec;
726 	struct timespec64 delta;
727 	ktime_t mc_time;
728 
729 	if (number_readings == 0)
730 		return -EAGAIN;
731 
732 	/* Read the set of results and find the last good host-MC
733 	 * synchronization result. The MC times when it finishes reading the
734 	 * host time so the corrected window time should be fairly constant
735 	 * for a given platform. Increment stats for any results that appear
736 	 * to be erroneous.
737 	 */
738 	for (i = 0; i < number_readings; i++) {
739 		s32 window, corrected;
740 		struct timespec64 wait;
741 
742 		efx_ptp_read_timeset(
743 			MCDI_ARRAY_STRUCT_PTR(synch_buf,
744 					      PTP_OUT_SYNCHRONIZE_TIMESET, i),
745 			&ptp->timeset[i]);
746 
747 		wait = ktime_to_timespec64(
748 			ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0));
749 		window = ptp->timeset[i].window;
750 		corrected = window - wait.tv_nsec;
751 
752 		/* We expect the uncorrected synchronization window to be at
753 		 * least as large as the interval between host start and end
754 		 * times. If it is smaller than this then this is mostly likely
755 		 * to be a consequence of the host's time being adjusted.
756 		 * Check that the corrected sync window is in a reasonable
757 		 * range. If it is out of range it is likely to be because an
758 		 * interrupt or other delay occurred between reading the system
759 		 * time and writing it to MC memory.
760 		 */
761 		if (window < SYNCHRONISATION_GRANULARITY_NS) {
762 			++ptp->invalid_sync_windows;
763 		} else if (corrected >= MAX_SYNCHRONISATION_NS) {
764 			++ptp->oversize_sync_windows;
765 		} else if (corrected < ptp->min_synchronisation_ns) {
766 			++ptp->undersize_sync_windows;
767 		} else {
768 			ngood++;
769 			last_good = i;
770 		}
771 	}
772 
773 	if (ngood == 0) {
774 		netif_warn(efx, drv, efx->net_dev,
775 			   "PTP no suitable synchronisations\n");
776 		return -EAGAIN;
777 	}
778 
779 	/* Calculate delay from last good sync (host time) to last_time.
780 	 * It is possible that the seconds rolled over between taking
781 	 * the start reading and the last value written by the host.  The
782 	 * timescales are such that a gap of more than one second is never
783 	 * expected.  delta is *not* normalised.
784 	 */
785 	start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS;
786 	last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK;
787 	if (start_sec != last_sec &&
788 	    ((start_sec + 1) & MC_SECOND_MASK) != last_sec) {
789 		netif_warn(efx, hw, efx->net_dev,
790 			   "PTP bad synchronisation seconds\n");
791 		return -EAGAIN;
792 	}
793 	delta.tv_sec = (last_sec - start_sec) & 1;
794 	delta.tv_nsec =
795 		last_time->ts_real.tv_nsec -
796 		(ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
797 
798 	/* Convert the NIC time at last good sync into kernel time.
799 	 * No correction is required - this time is the output of a
800 	 * firmware process.
801 	 */
802 	mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
803 					  ptp->timeset[last_good].minor, 0);
804 
805 	/* Calculate delay from NIC top of second to last_time */
806 	delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec;
807 
808 	/* Set PPS timestamp to match NIC top of second */
809 	ptp->host_time_pps = *last_time;
810 	pps_sub_ts(&ptp->host_time_pps, delta);
811 
812 	return 0;
813 }
814 
815 /* Synchronize times between the host and the MC */
816 static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
817 {
818 	struct efx_ptp_data *ptp = efx->ptp_data;
819 	MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX);
820 	size_t response_length;
821 	int rc;
822 	unsigned long timeout;
823 	struct pps_event_time last_time = {};
824 	unsigned int loops = 0;
825 	int *start = ptp->start.addr;
826 
827 	MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE);
828 	MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0);
829 	MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS,
830 		       num_readings);
831 	MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
832 		       ptp->start.dma_addr);
833 
834 	/* Clear flag that signals MC ready */
835 	ACCESS_ONCE(*start) = 0;
836 	rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf,
837 				MC_CMD_PTP_IN_SYNCHRONIZE_LEN);
838 	EFX_BUG_ON_PARANOID(rc);
839 
840 	/* Wait for start from MCDI (or timeout) */
841 	timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS);
842 	while (!ACCESS_ONCE(*start) && (time_before(jiffies, timeout))) {
843 		udelay(20);	/* Usually start MCDI execution quickly */
844 		loops++;
845 	}
846 
847 	if (loops <= 1)
848 		++ptp->fast_syncs;
849 	if (!time_before(jiffies, timeout))
850 		++ptp->sync_timeouts;
851 
852 	if (ACCESS_ONCE(*start))
853 		efx_ptp_send_times(efx, &last_time);
854 
855 	/* Collect results */
856 	rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP,
857 				 MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
858 				 synch_buf, sizeof(synch_buf),
859 				 &response_length);
860 	if (rc == 0) {
861 		rc = efx_ptp_process_times(efx, synch_buf, response_length,
862 					   &last_time);
863 		if (rc == 0)
864 			++ptp->good_syncs;
865 		else
866 			++ptp->no_time_syncs;
867 	}
868 
869 	/* Increment the bad syncs counter if the synchronize fails, whatever
870 	 * the reason.
871 	 */
872 	if (rc != 0)
873 		++ptp->bad_syncs;
874 
875 	return rc;
876 }
877 
878 /* Transmit a PTP packet, via the MCDI interface, to the wire. */
879 static int efx_ptp_xmit_skb(struct efx_nic *efx, struct sk_buff *skb)
880 {
881 	struct efx_ptp_data *ptp_data = efx->ptp_data;
882 	struct skb_shared_hwtstamps timestamps;
883 	int rc = -EIO;
884 	MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
885 	size_t len;
886 
887 	MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT);
888 	MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0);
889 	MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len);
890 	if (skb_shinfo(skb)->nr_frags != 0) {
891 		rc = skb_linearize(skb);
892 		if (rc != 0)
893 			goto fail;
894 	}
895 
896 	if (skb->ip_summed == CHECKSUM_PARTIAL) {
897 		rc = skb_checksum_help(skb);
898 		if (rc != 0)
899 			goto fail;
900 	}
901 	skb_copy_from_linear_data(skb,
902 				  MCDI_PTR(ptp_data->txbuf,
903 					   PTP_IN_TRANSMIT_PACKET),
904 				  skb->len);
905 	rc = efx_mcdi_rpc(efx, MC_CMD_PTP,
906 			  ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len),
907 			  txtime, sizeof(txtime), &len);
908 	if (rc != 0)
909 		goto fail;
910 
911 	memset(&timestamps, 0, sizeof(timestamps));
912 	timestamps.hwtstamp = ptp_data->nic_to_kernel_time(
913 		MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR),
914 		MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR),
915 		ptp_data->ts_corrections.tx);
916 
917 	skb_tstamp_tx(skb, &timestamps);
918 
919 	rc = 0;
920 
921 fail:
922 	dev_kfree_skb(skb);
923 
924 	return rc;
925 }
926 
927 static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
928 {
929 	struct efx_ptp_data *ptp = efx->ptp_data;
930 	struct list_head *cursor;
931 	struct list_head *next;
932 
933 	if (ptp->rx_ts_inline)
934 		return;
935 
936 	/* Drop time-expired events */
937 	spin_lock_bh(&ptp->evt_lock);
938 	if (!list_empty(&ptp->evt_list)) {
939 		list_for_each_safe(cursor, next, &ptp->evt_list) {
940 			struct efx_ptp_event_rx *evt;
941 
942 			evt = list_entry(cursor, struct efx_ptp_event_rx,
943 					 link);
944 			if (time_after(jiffies, evt->expiry)) {
945 				list_move(&evt->link, &ptp->evt_free_list);
946 				netif_warn(efx, hw, efx->net_dev,
947 					   "PTP rx event dropped\n");
948 			}
949 		}
950 	}
951 	spin_unlock_bh(&ptp->evt_lock);
952 }
953 
954 static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
955 					      struct sk_buff *skb)
956 {
957 	struct efx_ptp_data *ptp = efx->ptp_data;
958 	bool evts_waiting;
959 	struct list_head *cursor;
960 	struct list_head *next;
961 	struct efx_ptp_match *match;
962 	enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED;
963 
964 	WARN_ON_ONCE(ptp->rx_ts_inline);
965 
966 	spin_lock_bh(&ptp->evt_lock);
967 	evts_waiting = !list_empty(&ptp->evt_list);
968 	spin_unlock_bh(&ptp->evt_lock);
969 
970 	if (!evts_waiting)
971 		return PTP_PACKET_STATE_UNMATCHED;
972 
973 	match = (struct efx_ptp_match *)skb->cb;
974 	/* Look for a matching timestamp in the event queue */
975 	spin_lock_bh(&ptp->evt_lock);
976 	list_for_each_safe(cursor, next, &ptp->evt_list) {
977 		struct efx_ptp_event_rx *evt;
978 
979 		evt = list_entry(cursor, struct efx_ptp_event_rx, link);
980 		if ((evt->seq0 == match->words[0]) &&
981 		    (evt->seq1 == match->words[1])) {
982 			struct skb_shared_hwtstamps *timestamps;
983 
984 			/* Match - add in hardware timestamp */
985 			timestamps = skb_hwtstamps(skb);
986 			timestamps->hwtstamp = evt->hwtimestamp;
987 
988 			match->state = PTP_PACKET_STATE_MATCHED;
989 			rc = PTP_PACKET_STATE_MATCHED;
990 			list_move(&evt->link, &ptp->evt_free_list);
991 			break;
992 		}
993 	}
994 	spin_unlock_bh(&ptp->evt_lock);
995 
996 	return rc;
997 }
998 
999 /* Process any queued receive events and corresponding packets
1000  *
1001  * q is returned with all the packets that are ready for delivery.
1002  */
1003 static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
1004 {
1005 	struct efx_ptp_data *ptp = efx->ptp_data;
1006 	struct sk_buff *skb;
1007 
1008 	while ((skb = skb_dequeue(&ptp->rxq))) {
1009 		struct efx_ptp_match *match;
1010 
1011 		match = (struct efx_ptp_match *)skb->cb;
1012 		if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) {
1013 			__skb_queue_tail(q, skb);
1014 		} else if (efx_ptp_match_rx(efx, skb) ==
1015 			   PTP_PACKET_STATE_MATCHED) {
1016 			__skb_queue_tail(q, skb);
1017 		} else if (time_after(jiffies, match->expiry)) {
1018 			match->state = PTP_PACKET_STATE_TIMED_OUT;
1019 			++ptp->rx_no_timestamp;
1020 			__skb_queue_tail(q, skb);
1021 		} else {
1022 			/* Replace unprocessed entry and stop */
1023 			skb_queue_head(&ptp->rxq, skb);
1024 			break;
1025 		}
1026 	}
1027 }
1028 
1029 /* Complete processing of a received packet */
1030 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1031 {
1032 	local_bh_disable();
1033 	netif_receive_skb(skb);
1034 	local_bh_enable();
1035 }
1036 
1037 static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
1038 {
1039 	struct efx_ptp_data *ptp = efx->ptp_data;
1040 
1041 	if (ptp->rxfilter_installed) {
1042 		efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1043 					  ptp->rxfilter_general);
1044 		efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1045 					  ptp->rxfilter_event);
1046 		ptp->rxfilter_installed = false;
1047 	}
1048 }
1049 
1050 static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1051 {
1052 	struct efx_ptp_data *ptp = efx->ptp_data;
1053 	struct efx_filter_spec rxfilter;
1054 	int rc;
1055 
1056 	if (!ptp->channel || ptp->rxfilter_installed)
1057 		return 0;
1058 
1059 	/* Must filter on both event and general ports to ensure
1060 	 * that there is no packet re-ordering.
1061 	 */
1062 	efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1063 			   efx_rx_queue_index(
1064 				   efx_channel_get_rx_queue(ptp->channel)));
1065 	rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1066 				       htonl(PTP_ADDRESS),
1067 				       htons(PTP_EVENT_PORT));
1068 	if (rc != 0)
1069 		return rc;
1070 
1071 	rc = efx_filter_insert_filter(efx, &rxfilter, true);
1072 	if (rc < 0)
1073 		return rc;
1074 	ptp->rxfilter_event = rc;
1075 
1076 	efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1077 			   efx_rx_queue_index(
1078 				   efx_channel_get_rx_queue(ptp->channel)));
1079 	rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1080 				       htonl(PTP_ADDRESS),
1081 				       htons(PTP_GENERAL_PORT));
1082 	if (rc != 0)
1083 		goto fail;
1084 
1085 	rc = efx_filter_insert_filter(efx, &rxfilter, true);
1086 	if (rc < 0)
1087 		goto fail;
1088 	ptp->rxfilter_general = rc;
1089 
1090 	ptp->rxfilter_installed = true;
1091 	return 0;
1092 
1093 fail:
1094 	efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1095 				  ptp->rxfilter_event);
1096 	return rc;
1097 }
1098 
1099 static int efx_ptp_start(struct efx_nic *efx)
1100 {
1101 	struct efx_ptp_data *ptp = efx->ptp_data;
1102 	int rc;
1103 
1104 	ptp->reset_required = false;
1105 
1106 	rc = efx_ptp_insert_multicast_filters(efx);
1107 	if (rc)
1108 		return rc;
1109 
1110 	rc = efx_ptp_enable(efx);
1111 	if (rc != 0)
1112 		goto fail;
1113 
1114 	ptp->evt_frag_idx = 0;
1115 	ptp->current_adjfreq = 0;
1116 
1117 	return 0;
1118 
1119 fail:
1120 	efx_ptp_remove_multicast_filters(efx);
1121 	return rc;
1122 }
1123 
1124 static int efx_ptp_stop(struct efx_nic *efx)
1125 {
1126 	struct efx_ptp_data *ptp = efx->ptp_data;
1127 	struct list_head *cursor;
1128 	struct list_head *next;
1129 	int rc;
1130 
1131 	if (ptp == NULL)
1132 		return 0;
1133 
1134 	rc = efx_ptp_disable(efx);
1135 
1136 	efx_ptp_remove_multicast_filters(efx);
1137 
1138 	/* Make sure RX packets are really delivered */
1139 	efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq);
1140 	skb_queue_purge(&efx->ptp_data->txq);
1141 
1142 	/* Drop any pending receive events */
1143 	spin_lock_bh(&efx->ptp_data->evt_lock);
1144 	list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) {
1145 		list_move(cursor, &efx->ptp_data->evt_free_list);
1146 	}
1147 	spin_unlock_bh(&efx->ptp_data->evt_lock);
1148 
1149 	return rc;
1150 }
1151 
1152 static int efx_ptp_restart(struct efx_nic *efx)
1153 {
1154 	if (efx->ptp_data && efx->ptp_data->enabled)
1155 		return efx_ptp_start(efx);
1156 	return 0;
1157 }
1158 
1159 static void efx_ptp_pps_worker(struct work_struct *work)
1160 {
1161 	struct efx_ptp_data *ptp =
1162 		container_of(work, struct efx_ptp_data, pps_work);
1163 	struct efx_nic *efx = ptp->efx;
1164 	struct ptp_clock_event ptp_evt;
1165 
1166 	if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
1167 		return;
1168 
1169 	ptp_evt.type = PTP_CLOCK_PPSUSR;
1170 	ptp_evt.pps_times = ptp->host_time_pps;
1171 	ptp_clock_event(ptp->phc_clock, &ptp_evt);
1172 }
1173 
1174 static void efx_ptp_worker(struct work_struct *work)
1175 {
1176 	struct efx_ptp_data *ptp_data =
1177 		container_of(work, struct efx_ptp_data, work);
1178 	struct efx_nic *efx = ptp_data->efx;
1179 	struct sk_buff *skb;
1180 	struct sk_buff_head tempq;
1181 
1182 	if (ptp_data->reset_required) {
1183 		efx_ptp_stop(efx);
1184 		efx_ptp_start(efx);
1185 		return;
1186 	}
1187 
1188 	efx_ptp_drop_time_expired_events(efx);
1189 
1190 	__skb_queue_head_init(&tempq);
1191 	efx_ptp_process_events(efx, &tempq);
1192 
1193 	while ((skb = skb_dequeue(&ptp_data->txq)))
1194 		efx_ptp_xmit_skb(efx, skb);
1195 
1196 	while ((skb = __skb_dequeue(&tempq)))
1197 		efx_ptp_process_rx(efx, skb);
1198 }
1199 
1200 static const struct ptp_clock_info efx_phc_clock_info = {
1201 	.owner		= THIS_MODULE,
1202 	.name		= "sfc",
1203 	.max_adj	= MAX_PPB,
1204 	.n_alarm	= 0,
1205 	.n_ext_ts	= 0,
1206 	.n_per_out	= 0,
1207 	.n_pins		= 0,
1208 	.pps		= 1,
1209 	.adjfreq	= efx_phc_adjfreq,
1210 	.adjtime	= efx_phc_adjtime,
1211 	.gettime64	= efx_phc_gettime,
1212 	.settime64	= efx_phc_settime,
1213 	.enable		= efx_phc_enable,
1214 };
1215 
1216 /* Initialise PTP state. */
1217 int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1218 {
1219 	struct efx_ptp_data *ptp;
1220 	int rc = 0;
1221 	unsigned int pos;
1222 
1223 	ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL);
1224 	efx->ptp_data = ptp;
1225 	if (!efx->ptp_data)
1226 		return -ENOMEM;
1227 
1228 	ptp->efx = efx;
1229 	ptp->channel = channel;
1230 	ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1231 
1232 	rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL);
1233 	if (rc != 0)
1234 		goto fail1;
1235 
1236 	skb_queue_head_init(&ptp->rxq);
1237 	skb_queue_head_init(&ptp->txq);
1238 	ptp->workwq = create_singlethread_workqueue("sfc_ptp");
1239 	if (!ptp->workwq) {
1240 		rc = -ENOMEM;
1241 		goto fail2;
1242 	}
1243 
1244 	INIT_WORK(&ptp->work, efx_ptp_worker);
1245 	ptp->config.flags = 0;
1246 	ptp->config.tx_type = HWTSTAMP_TX_OFF;
1247 	ptp->config.rx_filter = HWTSTAMP_FILTER_NONE;
1248 	INIT_LIST_HEAD(&ptp->evt_list);
1249 	INIT_LIST_HEAD(&ptp->evt_free_list);
1250 	spin_lock_init(&ptp->evt_lock);
1251 	for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++)
1252 		list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list);
1253 
1254 	/* Get the NIC PTP attributes and set up time conversions */
1255 	rc = efx_ptp_get_attributes(efx);
1256 	if (rc < 0)
1257 		goto fail3;
1258 
1259 	/* Get the timestamp corrections */
1260 	rc = efx_ptp_get_timestamp_corrections(efx);
1261 	if (rc < 0)
1262 		goto fail3;
1263 
1264 	if (efx->mcdi->fn_flags &
1265 	    (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) {
1266 		ptp->phc_clock_info = efx_phc_clock_info;
1267 		ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info,
1268 						    &efx->pci_dev->dev);
1269 		if (IS_ERR(ptp->phc_clock)) {
1270 			rc = PTR_ERR(ptp->phc_clock);
1271 			goto fail3;
1272 		}
1273 
1274 		INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1275 		ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
1276 		if (!ptp->pps_workwq) {
1277 			rc = -ENOMEM;
1278 			goto fail4;
1279 		}
1280 	}
1281 	ptp->nic_ts_enabled = false;
1282 
1283 	return 0;
1284 fail4:
1285 	ptp_clock_unregister(efx->ptp_data->phc_clock);
1286 
1287 fail3:
1288 	destroy_workqueue(efx->ptp_data->workwq);
1289 
1290 fail2:
1291 	efx_nic_free_buffer(efx, &ptp->start);
1292 
1293 fail1:
1294 	kfree(efx->ptp_data);
1295 	efx->ptp_data = NULL;
1296 
1297 	return rc;
1298 }
1299 
1300 /* Initialise PTP channel.
1301  *
1302  * Setting core_index to zero causes the queue to be initialised and doesn't
1303  * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
1304  */
1305 static int efx_ptp_probe_channel(struct efx_channel *channel)
1306 {
1307 	struct efx_nic *efx = channel->efx;
1308 
1309 	channel->irq_moderation = 0;
1310 	channel->rx_queue.core_index = 0;
1311 
1312 	return efx_ptp_probe(efx, channel);
1313 }
1314 
1315 void efx_ptp_remove(struct efx_nic *efx)
1316 {
1317 	if (!efx->ptp_data)
1318 		return;
1319 
1320 	(void)efx_ptp_disable(efx);
1321 
1322 	cancel_work_sync(&efx->ptp_data->work);
1323 	cancel_work_sync(&efx->ptp_data->pps_work);
1324 
1325 	skb_queue_purge(&efx->ptp_data->rxq);
1326 	skb_queue_purge(&efx->ptp_data->txq);
1327 
1328 	if (efx->ptp_data->phc_clock) {
1329 		destroy_workqueue(efx->ptp_data->pps_workwq);
1330 		ptp_clock_unregister(efx->ptp_data->phc_clock);
1331 	}
1332 
1333 	destroy_workqueue(efx->ptp_data->workwq);
1334 
1335 	efx_nic_free_buffer(efx, &efx->ptp_data->start);
1336 	kfree(efx->ptp_data);
1337 }
1338 
1339 static void efx_ptp_remove_channel(struct efx_channel *channel)
1340 {
1341 	efx_ptp_remove(channel->efx);
1342 }
1343 
1344 static void efx_ptp_get_channel_name(struct efx_channel *channel,
1345 				     char *buf, size_t len)
1346 {
1347 	snprintf(buf, len, "%s-ptp", channel->efx->name);
1348 }
1349 
1350 /* Determine whether this packet should be processed by the PTP module
1351  * or transmitted conventionally.
1352  */
1353 bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1354 {
1355 	return efx->ptp_data &&
1356 		efx->ptp_data->enabled &&
1357 		skb->len >= PTP_MIN_LENGTH &&
1358 		skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1359 		likely(skb->protocol == htons(ETH_P_IP)) &&
1360 		skb_transport_header_was_set(skb) &&
1361 		skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1362 		ip_hdr(skb)->protocol == IPPROTO_UDP &&
1363 		skb_headlen(skb) >=
1364 		skb_transport_offset(skb) + sizeof(struct udphdr) &&
1365 		udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1366 }
1367 
1368 /* Receive a PTP packet.  Packets are queued until the arrival of
1369  * the receive timestamp from the MC - this will probably occur after the
1370  * packet arrival because of the processing in the MC.
1371  */
1372 static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1373 {
1374 	struct efx_nic *efx = channel->efx;
1375 	struct efx_ptp_data *ptp = efx->ptp_data;
1376 	struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1377 	u8 *match_data_012, *match_data_345;
1378 	unsigned int version;
1379 	u8 *data;
1380 
1381 	match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1382 
1383 	/* Correct version? */
1384 	if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1385 		if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1386 			return false;
1387 		}
1388 		data = skb->data;
1389 		version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1390 		if (version != PTP_VERSION_V1) {
1391 			return false;
1392 		}
1393 
1394 		/* PTP V1 uses all six bytes of the UUID to match the packet
1395 		 * to the timestamp
1396 		 */
1397 		match_data_012 = data + PTP_V1_UUID_OFFSET;
1398 		match_data_345 = data + PTP_V1_UUID_OFFSET + 3;
1399 	} else {
1400 		if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1401 			return false;
1402 		}
1403 		data = skb->data;
1404 		version = data[PTP_V2_VERSION_OFFSET];
1405 		if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1406 			return false;
1407 		}
1408 
1409 		/* The original V2 implementation uses bytes 2-7 of
1410 		 * the UUID to match the packet to the timestamp. This
1411 		 * discards two of the bytes of the MAC address used
1412 		 * to create the UUID (SF bug 33070).  The PTP V2
1413 		 * enhanced mode fixes this issue and uses bytes 0-2
1414 		 * and byte 5-7 of the UUID.
1415 		 */
1416 		match_data_345 = data + PTP_V2_UUID_OFFSET + 5;
1417 		if (ptp->mode == MC_CMD_PTP_MODE_V2) {
1418 			match_data_012 = data + PTP_V2_UUID_OFFSET + 2;
1419 		} else {
1420 			match_data_012 = data + PTP_V2_UUID_OFFSET + 0;
1421 			BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED);
1422 		}
1423 	}
1424 
1425 	/* Does this packet require timestamping? */
1426 	if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1427 		match->state = PTP_PACKET_STATE_UNMATCHED;
1428 
1429 		/* We expect the sequence number to be in the same position in
1430 		 * the packet for PTP V1 and V2
1431 		 */
1432 		BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1433 		BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1434 
1435 		/* Extract UUID/Sequence information */
1436 		match->words[0] = (match_data_012[0]         |
1437 				   (match_data_012[1] << 8)  |
1438 				   (match_data_012[2] << 16) |
1439 				   (match_data_345[0] << 24));
1440 		match->words[1] = (match_data_345[1]         |
1441 				   (match_data_345[2] << 8)  |
1442 				   (data[PTP_V1_SEQUENCE_OFFSET +
1443 					 PTP_V1_SEQUENCE_LENGTH - 1] <<
1444 				    16));
1445 	} else {
1446 		match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1447 	}
1448 
1449 	skb_queue_tail(&ptp->rxq, skb);
1450 	queue_work(ptp->workwq, &ptp->work);
1451 
1452 	return true;
1453 }
1454 
1455 /* Transmit a PTP packet.  This has to be transmitted by the MC
1456  * itself, through an MCDI call.  MCDI calls aren't permitted
1457  * in the transmit path so defer the actual transmission to a suitable worker.
1458  */
1459 int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1460 {
1461 	struct efx_ptp_data *ptp = efx->ptp_data;
1462 
1463 	skb_queue_tail(&ptp->txq, skb);
1464 
1465 	if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1466 	    (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1467 		efx_xmit_hwtstamp_pending(skb);
1468 	queue_work(ptp->workwq, &ptp->work);
1469 
1470 	return NETDEV_TX_OK;
1471 }
1472 
1473 int efx_ptp_get_mode(struct efx_nic *efx)
1474 {
1475 	return efx->ptp_data->mode;
1476 }
1477 
1478 int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1479 			unsigned int new_mode)
1480 {
1481 	if ((enable_wanted != efx->ptp_data->enabled) ||
1482 	    (enable_wanted && (efx->ptp_data->mode != new_mode))) {
1483 		int rc = 0;
1484 
1485 		if (enable_wanted) {
1486 			/* Change of mode requires disable */
1487 			if (efx->ptp_data->enabled &&
1488 			    (efx->ptp_data->mode != new_mode)) {
1489 				efx->ptp_data->enabled = false;
1490 				rc = efx_ptp_stop(efx);
1491 				if (rc != 0)
1492 					return rc;
1493 			}
1494 
1495 			/* Set new operating mode and establish
1496 			 * baseline synchronisation, which must
1497 			 * succeed.
1498 			 */
1499 			efx->ptp_data->mode = new_mode;
1500 			if (netif_running(efx->net_dev))
1501 				rc = efx_ptp_start(efx);
1502 			if (rc == 0) {
1503 				rc = efx_ptp_synchronize(efx,
1504 							 PTP_SYNC_ATTEMPTS * 2);
1505 				if (rc != 0)
1506 					efx_ptp_stop(efx);
1507 			}
1508 		} else {
1509 			rc = efx_ptp_stop(efx);
1510 		}
1511 
1512 		if (rc != 0)
1513 			return rc;
1514 
1515 		efx->ptp_data->enabled = enable_wanted;
1516 	}
1517 
1518 	return 0;
1519 }
1520 
1521 static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
1522 {
1523 	int rc;
1524 
1525 	if (init->flags)
1526 		return -EINVAL;
1527 
1528 	if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1529 	    (init->tx_type != HWTSTAMP_TX_ON))
1530 		return -ERANGE;
1531 
1532 	rc = efx->type->ptp_set_ts_config(efx, init);
1533 	if (rc)
1534 		return rc;
1535 
1536 	efx->ptp_data->config = *init;
1537 	return 0;
1538 }
1539 
1540 void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
1541 {
1542 	struct efx_ptp_data *ptp = efx->ptp_data;
1543 	struct efx_nic *primary = efx->primary;
1544 
1545 	ASSERT_RTNL();
1546 
1547 	if (!ptp)
1548 		return;
1549 
1550 	ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1551 				     SOF_TIMESTAMPING_RX_HARDWARE |
1552 				     SOF_TIMESTAMPING_RAW_HARDWARE);
1553 	if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1554 		ts_info->phc_index =
1555 			ptp_clock_index(primary->ptp_data->phc_clock);
1556 	ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1557 	ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1558 }
1559 
1560 int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1561 {
1562 	struct hwtstamp_config config;
1563 	int rc;
1564 
1565 	/* Not a PTP enabled port */
1566 	if (!efx->ptp_data)
1567 		return -EOPNOTSUPP;
1568 
1569 	if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
1570 		return -EFAULT;
1571 
1572 	rc = efx_ptp_ts_init(efx, &config);
1573 	if (rc != 0)
1574 		return rc;
1575 
1576 	return copy_to_user(ifr->ifr_data, &config, sizeof(config))
1577 		? -EFAULT : 0;
1578 }
1579 
1580 int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1581 {
1582 	if (!efx->ptp_data)
1583 		return -EOPNOTSUPP;
1584 
1585 	return copy_to_user(ifr->ifr_data, &efx->ptp_data->config,
1586 			    sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
1587 }
1588 
1589 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1590 {
1591 	struct efx_ptp_data *ptp = efx->ptp_data;
1592 
1593 	netif_err(efx, hw, efx->net_dev,
1594 		"PTP unexpected event length: got %d expected %d\n",
1595 		ptp->evt_frag_idx, expected_frag_len);
1596 	ptp->reset_required = true;
1597 	queue_work(ptp->workwq, &ptp->work);
1598 }
1599 
1600 /* Process a completed receive event.  Put it on the event queue and
1601  * start worker thread.  This is required because event and their
1602  * correspoding packets may come in either order.
1603  */
1604 static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
1605 {
1606 	struct efx_ptp_event_rx *evt = NULL;
1607 
1608 	if (WARN_ON_ONCE(ptp->rx_ts_inline))
1609 		return;
1610 
1611 	if (ptp->evt_frag_idx != 3) {
1612 		ptp_event_failure(efx, 3);
1613 		return;
1614 	}
1615 
1616 	spin_lock_bh(&ptp->evt_lock);
1617 	if (!list_empty(&ptp->evt_free_list)) {
1618 		evt = list_first_entry(&ptp->evt_free_list,
1619 				       struct efx_ptp_event_rx, link);
1620 		list_del(&evt->link);
1621 
1622 		evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA);
1623 		evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2],
1624 					     MCDI_EVENT_SRC)        |
1625 			     (EFX_QWORD_FIELD(ptp->evt_frags[1],
1626 					      MCDI_EVENT_SRC) << 8) |
1627 			     (EFX_QWORD_FIELD(ptp->evt_frags[0],
1628 					      MCDI_EVENT_SRC) << 16));
1629 		evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
1630 			EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
1631 			EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
1632 			ptp->ts_corrections.rx);
1633 		evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1634 		list_add_tail(&evt->link, &ptp->evt_list);
1635 
1636 		queue_work(ptp->workwq, &ptp->work);
1637 	} else if (net_ratelimit()) {
1638 		/* Log a rate-limited warning message. */
1639 		netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n");
1640 	}
1641 	spin_unlock_bh(&ptp->evt_lock);
1642 }
1643 
1644 static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1645 {
1646 	int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1647 	if (ptp->evt_frag_idx != 1) {
1648 		ptp_event_failure(efx, 1);
1649 		return;
1650 	}
1651 
1652 	netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1653 }
1654 
1655 static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1656 {
1657 	if (ptp->nic_ts_enabled)
1658 		queue_work(ptp->pps_workwq, &ptp->pps_work);
1659 }
1660 
1661 void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1662 {
1663 	struct efx_ptp_data *ptp = efx->ptp_data;
1664 	int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
1665 
1666 	if (!ptp) {
1667 		if (net_ratelimit())
1668 			netif_warn(efx, drv, efx->net_dev,
1669 				   "Received PTP event but PTP not set up\n");
1670 		return;
1671 	}
1672 
1673 	if (!ptp->enabled)
1674 		return;
1675 
1676 	if (ptp->evt_frag_idx == 0) {
1677 		ptp->evt_code = code;
1678 	} else if (ptp->evt_code != code) {
1679 		netif_err(efx, hw, efx->net_dev,
1680 			  "PTP out of sequence event %d\n", code);
1681 		ptp->evt_frag_idx = 0;
1682 	}
1683 
1684 	ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
1685 	if (!MCDI_EVENT_FIELD(*ev, CONT)) {
1686 		/* Process resulting event */
1687 		switch (code) {
1688 		case MCDI_EVENT_CODE_PTP_RX:
1689 			ptp_event_rx(efx, ptp);
1690 			break;
1691 		case MCDI_EVENT_CODE_PTP_FAULT:
1692 			ptp_event_fault(efx, ptp);
1693 			break;
1694 		case MCDI_EVENT_CODE_PTP_PPS:
1695 			ptp_event_pps(efx, ptp);
1696 			break;
1697 		default:
1698 			netif_err(efx, hw, efx->net_dev,
1699 				  "PTP unknown event %d\n", code);
1700 			break;
1701 		}
1702 		ptp->evt_frag_idx = 0;
1703 	} else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
1704 		netif_err(efx, hw, efx->net_dev,
1705 			  "PTP too many event fragments\n");
1706 		ptp->evt_frag_idx = 0;
1707 	}
1708 }
1709 
1710 void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
1711 {
1712 	channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
1713 	channel->sync_timestamp_minor =
1714 		MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_26_19) << 19;
1715 	/* if sync events have been disabled then we want to silently ignore
1716 	 * this event, so throw away result.
1717 	 */
1718 	(void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
1719 		       SYNC_EVENTS_VALID);
1720 }
1721 
1722 /* make some assumptions about the time representation rather than abstract it,
1723  * since we currently only support one type of inline timestamping and only on
1724  * EF10.
1725  */
1726 #define MINOR_TICKS_PER_SECOND 0x8000000
1727 /* Fuzz factor for sync events to be out of order with RX events */
1728 #define FUZZ (MINOR_TICKS_PER_SECOND / 10)
1729 #define EXPECTED_SYNC_EVENTS_PER_SECOND 4
1730 
1731 static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
1732 {
1733 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
1734 	return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset));
1735 #else
1736 	const u8 *data = eh + efx->rx_packet_ts_offset;
1737 	return (u32)data[0]       |
1738 	       (u32)data[1] << 8  |
1739 	       (u32)data[2] << 16 |
1740 	       (u32)data[3] << 24;
1741 #endif
1742 }
1743 
1744 void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
1745 				   struct sk_buff *skb)
1746 {
1747 	struct efx_nic *efx = channel->efx;
1748 	u32 pkt_timestamp_major, pkt_timestamp_minor;
1749 	u32 diff, carry;
1750 	struct skb_shared_hwtstamps *timestamps;
1751 
1752 	pkt_timestamp_minor = (efx_rx_buf_timestamp_minor(efx,
1753 							  skb_mac_header(skb)) +
1754 			       (u32) efx->ptp_data->ts_corrections.rx) &
1755 			      (MINOR_TICKS_PER_SECOND - 1);
1756 
1757 	/* get the difference between the packet and sync timestamps,
1758 	 * modulo one second
1759 	 */
1760 	diff = (pkt_timestamp_minor - channel->sync_timestamp_minor) &
1761 		(MINOR_TICKS_PER_SECOND - 1);
1762 	/* do we roll over a second boundary and need to carry the one? */
1763 	carry = channel->sync_timestamp_minor + diff > MINOR_TICKS_PER_SECOND ?
1764 		1 : 0;
1765 
1766 	if (diff <= MINOR_TICKS_PER_SECOND / EXPECTED_SYNC_EVENTS_PER_SECOND +
1767 		    FUZZ) {
1768 		/* packet is ahead of the sync event by a quarter of a second or
1769 		 * less (allowing for fuzz)
1770 		 */
1771 		pkt_timestamp_major = channel->sync_timestamp_major + carry;
1772 	} else if (diff >= MINOR_TICKS_PER_SECOND - FUZZ) {
1773 		/* packet is behind the sync event but within the fuzz factor.
1774 		 * This means the RX packet and sync event crossed as they were
1775 		 * placed on the event queue, which can sometimes happen.
1776 		 */
1777 		pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
1778 	} else {
1779 		/* it's outside tolerance in both directions. this might be
1780 		 * indicative of us missing sync events for some reason, so
1781 		 * we'll call it an error rather than risk giving a bogus
1782 		 * timestamp.
1783 		 */
1784 		netif_vdbg(efx, drv, efx->net_dev,
1785 			  "packet timestamp %x too far from sync event %x:%x\n",
1786 			  pkt_timestamp_minor, channel->sync_timestamp_major,
1787 			  channel->sync_timestamp_minor);
1788 		return;
1789 	}
1790 
1791 	/* attach the timestamps to the skb */
1792 	timestamps = skb_hwtstamps(skb);
1793 	timestamps->hwtstamp =
1794 		efx_ptp_s27_to_ktime(pkt_timestamp_major, pkt_timestamp_minor);
1795 }
1796 
1797 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
1798 {
1799 	struct efx_ptp_data *ptp_data = container_of(ptp,
1800 						     struct efx_ptp_data,
1801 						     phc_clock_info);
1802 	struct efx_nic *efx = ptp_data->efx;
1803 	MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
1804 	s64 adjustment_ns;
1805 	int rc;
1806 
1807 	if (delta > MAX_PPB)
1808 		delta = MAX_PPB;
1809 	else if (delta < -MAX_PPB)
1810 		delta = -MAX_PPB;
1811 
1812 	/* Convert ppb to fixed point ns. */
1813 	adjustment_ns = (((s64)delta * PPB_SCALE_WORD) >>
1814 			 (PPB_EXTRA_BITS + MAX_PPB_BITS));
1815 
1816 	MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
1817 	MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
1818 	MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
1819 	MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
1820 	MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
1821 	rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj),
1822 			  NULL, 0, NULL);
1823 	if (rc != 0)
1824 		return rc;
1825 
1826 	ptp_data->current_adjfreq = adjustment_ns;
1827 	return 0;
1828 }
1829 
1830 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
1831 {
1832 	u32 nic_major, nic_minor;
1833 	struct efx_ptp_data *ptp_data = container_of(ptp,
1834 						     struct efx_ptp_data,
1835 						     phc_clock_info);
1836 	struct efx_nic *efx = ptp_data->efx;
1837 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
1838 
1839 	efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
1840 
1841 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
1842 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
1843 	MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
1844 	MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
1845 	MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
1846 	return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
1847 			    NULL, 0, NULL);
1848 }
1849 
1850 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
1851 {
1852 	struct efx_ptp_data *ptp_data = container_of(ptp,
1853 						     struct efx_ptp_data,
1854 						     phc_clock_info);
1855 	struct efx_nic *efx = ptp_data->efx;
1856 	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
1857 	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
1858 	int rc;
1859 	ktime_t kt;
1860 
1861 	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
1862 	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
1863 
1864 	rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
1865 			  outbuf, sizeof(outbuf), NULL);
1866 	if (rc != 0)
1867 		return rc;
1868 
1869 	kt = ptp_data->nic_to_kernel_time(
1870 		MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
1871 		MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
1872 	*ts = ktime_to_timespec64(kt);
1873 	return 0;
1874 }
1875 
1876 static int efx_phc_settime(struct ptp_clock_info *ptp,
1877 			   const struct timespec64 *e_ts)
1878 {
1879 	/* Get the current NIC time, efx_phc_gettime.
1880 	 * Subtract from the desired time to get the offset
1881 	 * call efx_phc_adjtime with the offset
1882 	 */
1883 	int rc;
1884 	struct timespec64 time_now;
1885 	struct timespec64 delta;
1886 
1887 	rc = efx_phc_gettime(ptp, &time_now);
1888 	if (rc != 0)
1889 		return rc;
1890 
1891 	delta = timespec64_sub(*e_ts, time_now);
1892 
1893 	rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta));
1894 	if (rc != 0)
1895 		return rc;
1896 
1897 	return 0;
1898 }
1899 
1900 static int efx_phc_enable(struct ptp_clock_info *ptp,
1901 			  struct ptp_clock_request *request,
1902 			  int enable)
1903 {
1904 	struct efx_ptp_data *ptp_data = container_of(ptp,
1905 						     struct efx_ptp_data,
1906 						     phc_clock_info);
1907 	if (request->type != PTP_CLK_REQ_PPS)
1908 		return -EOPNOTSUPP;
1909 
1910 	ptp_data->nic_ts_enabled = !!enable;
1911 	return 0;
1912 }
1913 
1914 static const struct efx_channel_type efx_ptp_channel_type = {
1915 	.handle_no_channel	= efx_ptp_handle_no_channel,
1916 	.pre_probe		= efx_ptp_probe_channel,
1917 	.post_remove		= efx_ptp_remove_channel,
1918 	.get_name		= efx_ptp_get_channel_name,
1919 	/* no copy operation; there is no need to reallocate this channel */
1920 	.receive_skb		= efx_ptp_rx,
1921 	.keep_eventq		= false,
1922 };
1923 
1924 void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
1925 {
1926 	/* Check whether PTP is implemented on this NIC.  The DISABLE
1927 	 * operation will succeed if and only if it is implemented.
1928 	 */
1929 	if (efx_ptp_disable(efx) == 0)
1930 		efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
1931 			&efx_ptp_channel_type;
1932 }
1933 
1934 void efx_ptp_start_datapath(struct efx_nic *efx)
1935 {
1936 	if (efx_ptp_restart(efx))
1937 		netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
1938 	/* re-enable timestamping if it was previously enabled */
1939 	if (efx->type->ptp_set_ts_sync_events)
1940 		efx->type->ptp_set_ts_sync_events(efx, true, true);
1941 }
1942 
1943 void efx_ptp_stop_datapath(struct efx_nic *efx)
1944 {
1945 	/* temporarily disable timestamping */
1946 	if (efx->type->ptp_set_ts_sync_events)
1947 		efx->type->ptp_set_ts_sync_events(efx, false, true);
1948 	efx_ptp_stop(efx);
1949 }
1950