1	// SPDX-License-Identifier: GPL-2.0-only
2	/****************************************************************************
3	* Driver for Solarflare network controllers and boards
4	* Copyright 2011-2013 Solarflare Communications Inc.
5	*/
6
7	/* Theory of operation:
8	*
9	* PTP support is assisted by firmware running on the MC, which provides
10	* the hardware timestamping capabilities. Both transmitted and received
11	* PTP event packets are queued onto internal queues for subsequent processing;
12	* this is because the MC operations are relatively long and would block
13	* block NAPI/interrupt operation.
14	*
15	* Receive event processing:
16	* The event contains the packet's UUID and sequence number, together
17	* with the hardware timestamp. The PTP receive packet queue is searched
18	* for this UUID/sequence number and, if found, put on a pending queue.
19	* Packets not matching are delivered without timestamps (MCDI events will
20	* always arrive after the actual packet).
21	* It is important for the operation of the PTP protocol that the ordering
22	* of packets between the event and general port is maintained.
23	*
24	* Work queue processing:
25	* If work waiting, synchronise host/hardware time
26	*
27	* Transmit: send packet through MC, which returns the transmission time
28	* that is converted to an appropriate timestamp.
29	*
30	* Receive: the packet's reception time is converted to an appropriate
31	* timestamp.
32	*/
33	#include <linux/ip.h>
34	#include <linux/udp.h>
35	#include <linux/time.h>
36	#include <linux/ktime.h>
37	#include <linux/module.h>
38	#include <linux/pps_kernel.h>
39	#include <linux/ptp_clock_kernel.h>
40	#include "net_driver.h"
41	#include "efx.h"
42	#include "mcdi.h"
43	#include "mcdi_pcol.h"
44	#include "io.h"
45	#include "farch_regs.h"
46	#include "tx.h"
47	#include "nic.h" /* indirectly includes ptp.h */
48
49	/* Maximum number of events expected to make up a PTP event */
50	#define MAX_EVENT_FRAGS 3
51
52	/* Maximum delay, ms, to begin synchronisation */
53	#define MAX_SYNCHRONISE_WAIT_MS 2
54
55	/* How long, at most, to spend synchronising */
56	#define SYNCHRONISE_PERIOD_NS 250000
57
58	/* How often to update the shared memory time */
59	#define SYNCHRONISATION_GRANULARITY_NS 200
60
61	/* Minimum permitted length of a (corrected) synchronisation time */
62	#define DEFAULT_MIN_SYNCHRONISATION_NS 120
63
64	/* Maximum permitted length of a (corrected) synchronisation time */
65	#define MAX_SYNCHRONISATION_NS 1000
66
67	/* How many (MC) receive events that can be queued */
68	#define MAX_RECEIVE_EVENTS 8
69
70	/* Length of (modified) moving average. */
71	#define AVERAGE_LENGTH 16
72
73	/* How long an unmatched event or packet can be held */
74	#define PKT_EVENT_LIFETIME_MS 10
75
76	/* Offsets into PTP packet for identification. These offsets are from the
77	* start of the IP header, not the MAC header. Note that neither PTP V1 nor
78	* PTP V2 permit the use of IPV4 options.
79	*/
80	#define PTP_DPORT_OFFSET 22
81
82	#define PTP_V1_VERSION_LENGTH 2
83	#define PTP_V1_VERSION_OFFSET 28
84
85	#define PTP_V1_UUID_LENGTH 6
86	#define PTP_V1_UUID_OFFSET 50
87
88	#define PTP_V1_SEQUENCE_LENGTH 2
89	#define PTP_V1_SEQUENCE_OFFSET 58
90
91	/* The minimum length of a PTP V1 packet for offsets, etc. to be valid:
92	* includes IP header.
93	*/
94	#define PTP_V1_MIN_LENGTH 64
95
96	#define PTP_V2_VERSION_LENGTH 1
97	#define PTP_V2_VERSION_OFFSET 29
98
99	#define PTP_V2_UUID_LENGTH 8
100	#define PTP_V2_UUID_OFFSET 48
101
102	/* Although PTP V2 UUIDs are comprised a ClockIdentity (8) and PortNumber (2),
103	* the MC only captures the last six bytes of the clock identity. These values
104	* reflect those, not the ones used in the standard. The standard permits
105	* mapping of V1 UUIDs to V2 UUIDs with these same values.
106	*/
107	#define PTP_V2_MC_UUID_LENGTH 6
108	#define PTP_V2_MC_UUID_OFFSET 50
109
110	#define PTP_V2_SEQUENCE_LENGTH 2
111	#define PTP_V2_SEQUENCE_OFFSET 58
112
113	/* The minimum length of a PTP V2 packet for offsets, etc. to be valid:
114	* includes IP header.
115	*/
116	#define PTP_V2_MIN_LENGTH 63
117
118	#define PTP_MIN_LENGTH 63
119
120	#define PTP_ADDRESS 0xe0000181 /* 224.0.1.129 */
121	#define PTP_EVENT_PORT 319
122	#define PTP_GENERAL_PORT 320
123
124	/* Annoyingly the format of the version numbers are different between
125	* versions 1 and 2 so it isn't possible to simply look for 1 or 2.
126	*/
127	#define PTP_VERSION_V1 1
128
129	#define PTP_VERSION_V2 2
130	#define PTP_VERSION_V2_MASK 0x0f
131
132	enum ptp_packet_state {
133	PTP_PACKET_STATE_UNMATCHED = 0,
134	PTP_PACKET_STATE_MATCHED,
135	PTP_PACKET_STATE_TIMED_OUT,
136	PTP_PACKET_STATE_MATCH_UNWANTED
137	};
138
139	/* NIC synchronised with single word of time only comprising
140	* partial seconds and full nanoseconds: 10^9 ~ 2^30 so 2 bits for seconds.
141	*/
142	#define MC_NANOSECOND_BITS 30
143	#define MC_NANOSECOND_MASK ((1 << MC_NANOSECOND_BITS) - 1)
144	#define MC_SECOND_MASK ((1 << (32 - MC_NANOSECOND_BITS)) - 1)
145
146	/* Maximum parts-per-billion adjustment that is acceptable */
147	#define MAX_PPB 1000000
148
149	/* Precalculate scale word to avoid long long division at runtime */
150	/* This is equivalent to 2^66 / 10^9. */
151	#define PPB_SCALE_WORD ((1LL << (57)) / 1953125LL)
152
153	/* How much to shift down after scaling to convert to FP40 */
154	#define PPB_SHIFT_FP40 26
155	/* ... and FP44. */
156	#define PPB_SHIFT_FP44 22
157
158	#define PTP_SYNC_ATTEMPTS 4
159
160	/**
161	* struct efx_ptp_match - Matching structure, stored in sk_buff's cb area.
162	* @words: UUID and (partial) sequence number
163	* @expiry: Time after which the packet should be delivered irrespective of
164	* event arrival.
165	* @state: The state of the packet - whether it is ready for processing or
166	* whether that is of no interest.
167	*/
168	struct efx_ptp_match {
169	u32 words[DIV_ROUND_UP(PTP_V1_UUID_LENGTH, 4)];
170	unsigned long expiry;
171	enum ptp_packet_state state;
172	};
173
174	/**
175	* struct efx_ptp_event_rx - A PTP receive event (from MC)
176	* @link: list of events
177	* @seq0: First part of (PTP) UUID
178	* @seq1: Second part of (PTP) UUID and sequence number
179	* @hwtimestamp: Event timestamp
180	* @expiry: Time which the packet arrived
181	*/
182	struct efx_ptp_event_rx {
183	struct list_head link;
184	u32 seq0;
185	u32 seq1;
186	ktime_t hwtimestamp;
187	unsigned long expiry;
188	};
189
190	/**
191	* struct efx_ptp_timeset - Synchronisation between host and MC
192	* @host_start: Host time immediately before hardware timestamp taken
193	* @major: Hardware timestamp, major
194	* @minor: Hardware timestamp, minor
195	* @host_end: Host time immediately after hardware timestamp taken
196	* @wait: Number of NIC clock ticks between hardware timestamp being read and
197	* host end time being seen
198	* @window: Difference of host_end and host_start
199	* @valid: Whether this timeset is valid
200	*/
201	struct efx_ptp_timeset {
202	u32 host_start;
203	u32 major;
204	u32 minor;
205	u32 host_end;
206	u32 wait;
207	u32 window; /* Derived: end - start, allowing for wrap */
208	};
209
210	/**
211	* struct efx_ptp_data - Precision Time Protocol (PTP) state
212	* @efx: The NIC context
213	* @channel: The PTP channel (Siena only)
214	* @rx_ts_inline: Flag for whether RX timestamps are inline (else they are
215	* separate events)
216	* @rxq: Receive SKB queue (awaiting timestamps)
217	* @txq: Transmit SKB queue
218	* @evt_list: List of MC receive events awaiting packets
219	* @evt_free_list: List of free events
220	* @evt_lock: Lock for manipulating evt_list and evt_free_list
221	* @rx_evts: Instantiated events (on evt_list and evt_free_list)
222	* @workwq: Work queue for processing pending PTP operations
223	* @work: Work task
224	* @reset_required: A serious error has occurred and the PTP task needs to be
225	* reset (disable, enable).
226	* @rxfilter_event: Receive filter when operating
227	* @rxfilter_general: Receive filter when operating
228	* @rxfilter_installed: Receive filter installed
229	* @config: Current timestamp configuration
230	* @enabled: PTP operation enabled
231	* @mode: Mode in which PTP operating (PTP version)
232	* @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time
233	* @nic_to_kernel_time: Function to convert from NIC to kernel time
234	* @nic_time: contains time details
235	* @nic_time.minor_max: Wrap point for NIC minor times
236	* @nic_time.sync_event_diff_min: Minimum acceptable difference between time
237	* in packet prefix and last MCDI time sync event i.e. how much earlier than
238	* the last sync event time a packet timestamp can be.
239	* @nic_time.sync_event_diff_max: Maximum acceptable difference between time
240	* in packet prefix and last MCDI time sync event i.e. how much later than
241	* the last sync event time a packet timestamp can be.
242	* @nic_time.sync_event_minor_shift: Shift required to make minor time from
243	* field in MCDI time sync event.
244	* @min_synchronisation_ns: Minimum acceptable corrected sync window
245	* @capabilities: Capabilities flags from the NIC
246	* @ts_corrections: contains corrections details
247	* @ts_corrections.ptp_tx: Required driver correction of PTP packet transmit
248	* timestamps
249	* @ts_corrections.ptp_rx: Required driver correction of PTP packet receive
250	* timestamps
251	* @ts_corrections.pps_out: PPS output error (information only)
252	* @ts_corrections.pps_in: Required driver correction of PPS input timestamps
253	* @ts_corrections.general_tx: Required driver correction of general packet
254	* transmit timestamps
255	* @ts_corrections.general_rx: Required driver correction of general packet
256	* receive timestamps
257	* @evt_frags: Partly assembled PTP events
258	* @evt_frag_idx: Current fragment number
259	* @evt_code: Last event code
260	* @start: Address at which MC indicates ready for synchronisation
261	* @host_time_pps: Host time at last PPS
262	* @adjfreq_ppb_shift: Shift required to convert scaled parts-per-billion
263	* frequency adjustment into a fixed point fractional nanosecond format.
264	* @current_adjfreq: Current ppb adjustment.
265	* @phc_clock: Pointer to registered phc device (if primary function)
266	* @phc_clock_info: Registration structure for phc device
267	* @pps_work: pps work task for handling pps events
268	* @pps_workwq: pps work queue
269	* @nic_ts_enabled: Flag indicating if NIC generated TS events are handled
270	* @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids
271	* allocations in main data path).
272	* @good_syncs: Number of successful synchronisations.
273	* @fast_syncs: Number of synchronisations requiring short delay
274	* @bad_syncs: Number of failed synchronisations.
275	* @sync_timeouts: Number of synchronisation timeouts
276	* @no_time_syncs: Number of synchronisations with no good times.
277	* @invalid_sync_windows: Number of sync windows with bad durations.
278	* @undersize_sync_windows: Number of corrected sync windows that are too small
279	* @oversize_sync_windows: Number of corrected sync windows that are too large
280	* @rx_no_timestamp: Number of packets received without a timestamp.
281	* @timeset: Last set of synchronisation statistics.
282	* @xmit_skb: Transmit SKB function.
283	*/
284	struct efx_ptp_data {
285	struct efx_nic *efx;
286	struct efx_channel *channel;
287	bool rx_ts_inline;
288	struct sk_buff_head rxq;
289	struct sk_buff_head txq;
290	struct list_head evt_list;
291	struct list_head evt_free_list;
292	spinlock_t evt_lock;
293	struct efx_ptp_event_rx rx_evts[MAX_RECEIVE_EVENTS];
294	struct workqueue_struct *workwq;
295	struct work_struct work;
296	bool reset_required;
297	u32 rxfilter_event;
298	u32 rxfilter_general;
299	bool rxfilter_installed;
300	struct kernel_hwtstamp_config config;
301	bool enabled;
302	unsigned int mode;
303	void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor);
304	ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor,
305	s32 correction);
306	struct {
307	u32 minor_max;
308	u32 sync_event_diff_min;
309	u32 sync_event_diff_max;
310	unsigned int sync_event_minor_shift;
311	} nic_time;
312	unsigned int min_synchronisation_ns;
313	unsigned int capabilities;
314	struct {
315	s32 ptp_tx;
316	s32 ptp_rx;
317	s32 pps_out;
318	s32 pps_in;
319	s32 general_tx;
320	s32 general_rx;
321	} ts_corrections;
322	efx_qword_t evt_frags[MAX_EVENT_FRAGS];
323	int evt_frag_idx;
324	int evt_code;
325	struct efx_buffer start;
326	struct pps_event_time host_time_pps;
327	unsigned int adjfreq_ppb_shift;
328	s64 current_adjfreq;
329	struct ptp_clock *phc_clock;
330	struct ptp_clock_info phc_clock_info;
331	struct work_struct pps_work;
332	struct workqueue_struct *pps_workwq;
333	bool nic_ts_enabled;
334	efx_dword_t txbuf[MCDI_TX_BUF_LEN(MC_CMD_PTP_IN_TRANSMIT_LENMAX)];
335
336	unsigned int good_syncs;
337	unsigned int fast_syncs;
338	unsigned int bad_syncs;
339	unsigned int sync_timeouts;
340	unsigned int no_time_syncs;
341	unsigned int invalid_sync_windows;
342	unsigned int undersize_sync_windows;
343	unsigned int oversize_sync_windows;
344	unsigned int rx_no_timestamp;
345	struct efx_ptp_timeset
346	timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM];
347	void (*xmit_skb)(struct efx_nic *efx, struct sk_buff *skb);
348	};
349
350	static int efx_phc_adjfine(struct ptp_clock_info *ptp, long scaled_ppm);
351	static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta);
352	static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts);
353	static int efx_phc_settime(struct ptp_clock_info *ptp,
354	const struct timespec64 *e_ts);
355	static int efx_phc_enable(struct ptp_clock_info *ptp,
356	struct ptp_clock_request *request, int on);
357
358	bool efx_siena_ptp_use_mac_tx_timestamps(struct efx_nic *efx)
359	{
360	return efx_has_cap(efx, TX_MAC_TIMESTAMPING);
361	}
362
363	/* PTP 'extra' channel is still a traffic channel, but we only create TX queues
364	* if PTP uses MAC TX timestamps, not if PTP uses the MC directly to transmit.
365	*/
366	static bool efx_ptp_want_txqs(struct efx_channel *channel)
367	{
368	return efx_siena_ptp_use_mac_tx_timestamps(efx: channel->efx);
369	}
370
371	#define PTP_SW_STAT(ext_name, field_name) \
372	{ #ext_name, 0, offsetof(struct efx_ptp_data, field_name) }
373	#define PTP_MC_STAT(ext_name, mcdi_name) \
374	{ #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST }
375	static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = {
376	PTP_SW_STAT(ptp_good_syncs, good_syncs),
377	PTP_SW_STAT(ptp_fast_syncs, fast_syncs),
378	PTP_SW_STAT(ptp_bad_syncs, bad_syncs),
379	PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts),
380	PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs),
381	PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows),
382	PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows),
383	PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows),
384	PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp),
385	PTP_MC_STAT(ptp_tx_timestamp_packets, TX),
386	PTP_MC_STAT(ptp_rx_timestamp_packets, RX),
387	PTP_MC_STAT(ptp_timestamp_packets, TS),
388	PTP_MC_STAT(ptp_filter_matches, FM),
389	PTP_MC_STAT(ptp_non_filter_matches, NFM),
390	};
391	#define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc)
392	static const unsigned long efx_ptp_stat_mask[] = {
393	[0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL,
394	};
395
396	size_t efx_siena_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
397	{
398	if (!efx->ptp_data)
399	return 0;
400
401	return efx_siena_describe_stats(desc: efx_ptp_stat_desc, PTP_STAT_COUNT,
402	mask: efx_ptp_stat_mask, names: strings);
403	}
404
405	size_t efx_siena_ptp_update_stats(struct efx_nic *efx, u64 *stats)
406	{
407	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
408	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
409	size_t i;
410	int rc;
411
412	if (!efx->ptp_data)
413	return 0;
414
415	/* Copy software statistics */
416	for (i = 0; i < PTP_STAT_COUNT; i++) {
417	if (efx_ptp_stat_desc[i].dma_width)
418	continue;
419	stats[i] = *(unsigned int *)((char *)efx->ptp_data +
420	efx_ptp_stat_desc[i].offset);
421	}
422
423	/* Fetch MC statistics. We *must* fill in all statistics or
424	* risk leaking kernel memory to userland, so if the MCDI
425	* request fails we pretend we got zeroes.
426	*/
427	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS);
428	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
429	rc = efx_siena_mcdi_rpc(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
430	outbuf, outlen: sizeof(outbuf), NULL);
431	if (rc)
432	memset(outbuf, 0, sizeof(outbuf));
433	efx_siena_update_stats(desc: efx_ptp_stat_desc, PTP_STAT_COUNT,
434	mask: efx_ptp_stat_mask,
435	stats, _MCDI_PTR(outbuf, 0), accumulate: false);
436
437	return PTP_STAT_COUNT;
438	}
439
440	/* For Siena platforms NIC time is s and ns */
441	static void efx_ptp_ns_to_s_ns(s64 ns, u32 *nic_major, u32 *nic_minor)
442	{
443	struct timespec64 ts = ns_to_timespec64(nsec: ns);
444	*nic_major = (u32)ts.tv_sec;
445	*nic_minor = ts.tv_nsec;
446	}
447
448	static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
449	s32 correction)
450	{
451	ktime_t kt = ktime_set(secs: nic_major, nsecs: nic_minor);
452	if (correction >= 0)
453	kt = ktime_add_ns(kt, (u64)correction);
454	else
455	kt = ktime_sub_ns(kt, (u64)-correction);
456	return kt;
457	}
458
459	/* To convert from s27 format to ns we multiply then divide by a power of 2.
460	* For the conversion from ns to s27, the operation is also converted to a
461	* multiply and shift.
462	*/
463	#define S27_TO_NS_SHIFT (27)
464	#define NS_TO_S27_MULT (((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC)
465	#define NS_TO_S27_SHIFT (63 - S27_TO_NS_SHIFT)
466	#define S27_MINOR_MAX (1 << S27_TO_NS_SHIFT)
467
468	/* For Huntington platforms NIC time is in seconds and fractions of a second
469	* where the minor register only uses 27 bits in units of 2^-27s.
470	*/
471	static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
472	{
473	struct timespec64 ts = ns_to_timespec64(nsec: ns);
474	u32 maj = (u32)ts.tv_sec;
475	u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT +
476	(1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT);
477
478	/* The conversion can result in the minor value exceeding the maximum.
479	* In this case, round up to the next second.
480	*/
481	if (min >= S27_MINOR_MAX) {
482	min -= S27_MINOR_MAX;
483	maj++;
484	}
485
486	*nic_major = maj;
487	*nic_minor = min;
488	}
489
490	static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
491	{
492	u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC +
493	(1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT);
494	return ktime_set(secs: nic_major, nsecs: ns);
495	}
496
497	static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
498	s32 correction)
499	{
500	/* Apply the correction and deal with carry */
501	nic_minor += correction;
502	if ((s32)nic_minor < 0) {
503	nic_minor += S27_MINOR_MAX;
504	nic_major--;
505	} else if (nic_minor >= S27_MINOR_MAX) {
506	nic_minor -= S27_MINOR_MAX;
507	nic_major++;
508	}
509
510	return efx_ptp_s27_to_ktime(nic_major, nic_minor);
511	}
512
513	/* For Medford2 platforms the time is in seconds and quarter nanoseconds. */
514	static void efx_ptp_ns_to_s_qns(s64 ns, u32 *nic_major, u32 *nic_minor)
515	{
516	struct timespec64 ts = ns_to_timespec64(nsec: ns);
517
518	*nic_major = (u32)ts.tv_sec;
519	*nic_minor = ts.tv_nsec * 4;
520	}
521
522	static ktime_t efx_ptp_s_qns_to_ktime_correction(u32 nic_major, u32 nic_minor,
523	s32 correction)
524	{
525	ktime_t kt;
526
527	nic_minor = DIV_ROUND_CLOSEST(nic_minor, 4);
528	correction = DIV_ROUND_CLOSEST(correction, 4);
529
530	kt = ktime_set(secs: nic_major, nsecs: nic_minor);
531
532	if (correction >= 0)
533	kt = ktime_add_ns(kt, (u64)correction);
534	else
535	kt = ktime_sub_ns(kt, (u64)-correction);
536	return kt;
537	}
538
539	struct efx_channel *efx_siena_ptp_channel(struct efx_nic *efx)
540	{
541	return efx->ptp_data ? efx->ptp_data->channel : NULL;
542	}
543
544	static u32 last_sync_timestamp_major(struct efx_nic *efx)
545	{
546	struct efx_channel *channel = efx_siena_ptp_channel(efx);
547	u32 major = 0;
548
549	if (channel)
550	major = channel->sync_timestamp_major;
551	return major;
552	}
553
554	/* The 8000 series and later can provide the time from the MAC, which is only
555	* 48 bits long and provides meta-information in the top 2 bits.
556	*/
557	static ktime_t
558	efx_ptp_mac_nic_to_ktime_correction(struct efx_nic *efx,
559	struct efx_ptp_data *ptp,
560	u32 nic_major, u32 nic_minor,
561	s32 correction)
562	{
563	u32 sync_timestamp;
564	ktime_t kt = { 0 };
565	s16 delta;
566
567	if (!(nic_major & 0x80000000)) {
568	WARN_ON_ONCE(nic_major >> 16);
569
570	/* Medford provides 48 bits of timestamp, so we must get the top
571	* 16 bits from the timesync event state.
572	*
573	* We only have the lower 16 bits of the time now, but we do
574	* have a full resolution timestamp at some point in past. As
575	* long as the difference between the (real) now and the sync
576	* is less than 2^15, then we can reconstruct the difference
577	* between those two numbers using only the lower 16 bits of
578	* each.
579	*
580	* Put another way
581	*
582	* a - b = ((a mod k) - b) mod k
583	*
584	* when -k/2 < (a-b) < k/2. In our case k is 2^16. We know
585	* (a mod k) and b, so can calculate the delta, a - b.
586	*
587	*/
588	sync_timestamp = last_sync_timestamp_major(efx);
589
590	/* Because delta is s16 this does an implicit mask down to
591	* 16 bits which is what we need, assuming
592	* MEDFORD_TX_SECS_EVENT_BITS is 16. delta is signed so that
593	* we can deal with the (unlikely) case of sync timestamps
594	* arriving from the future.
595	*/
596	delta = nic_major - sync_timestamp;
597
598	/* Recover the fully specified time now, by applying the offset
599	* to the (fully specified) sync time.
600	*/
601	nic_major = sync_timestamp + delta;
602
603	kt = ptp->nic_to_kernel_time(nic_major, nic_minor,
604	correction);
605	}
606	return kt;
607	}
608
609	ktime_t efx_siena_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue)
610	{
611	struct efx_nic *efx = tx_queue->efx;
612	struct efx_ptp_data *ptp = efx->ptp_data;
613	ktime_t kt;
614
615	if (efx_siena_ptp_use_mac_tx_timestamps(efx))
616	kt = efx_ptp_mac_nic_to_ktime_correction(efx, ptp,
617	nic_major: tx_queue->completed_timestamp_major,
618	nic_minor: tx_queue->completed_timestamp_minor,
619	correction: ptp->ts_corrections.general_tx);
620	else
621	kt = ptp->nic_to_kernel_time(
622	tx_queue->completed_timestamp_major,
623	tx_queue->completed_timestamp_minor,
624	ptp->ts_corrections.general_tx);
625	return kt;
626	}
627
628	/* Get PTP attributes and set up time conversions */
629	static int efx_ptp_get_attributes(struct efx_nic *efx)
630	{
631	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN);
632	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN);
633	struct efx_ptp_data *ptp = efx->ptp_data;
634	int rc;
635	u32 fmt;
636	size_t out_len;
637
638	/* Get the PTP attributes. If the NIC doesn't support the operation we
639	* use the default format for compatibility with older NICs i.e.
640	* seconds and nanoseconds.
641	*/
642	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES);
643	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
644	rc = efx_siena_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
645	outbuf, outlen: sizeof(outbuf), outlen_actual: &out_len);
646	if (rc == 0) {
647	fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT);
648	} else if (rc == -EINVAL) {
649	fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS;
650	} else if (rc == -EPERM) {
651	pci_info(efx->pci_dev, "no PTP support\n");
652	return rc;
653	} else {
654	efx_siena_mcdi_display_error(efx, MC_CMD_PTP, inlen: sizeof(inbuf),
655	outbuf, outlen: sizeof(outbuf), rc);
656	return rc;
657	}
658
659	switch (fmt) {
660	case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION:
661	ptp->ns_to_nic_time = efx_ptp_ns_to_s27;
662	ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction;
663	ptp->nic_time.minor_max = 1 << 27;
664	ptp->nic_time.sync_event_minor_shift = 19;
665	break;
666	case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS:
667	ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns;
668	ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction;
669	ptp->nic_time.minor_max = 1000000000;
670	ptp->nic_time.sync_event_minor_shift = 22;
671	break;
672	case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_QTR_NANOSECONDS:
673	ptp->ns_to_nic_time = efx_ptp_ns_to_s_qns;
674	ptp->nic_to_kernel_time = efx_ptp_s_qns_to_ktime_correction;
675	ptp->nic_time.minor_max = 4000000000UL;
676	ptp->nic_time.sync_event_minor_shift = 24;
677	break;
678	default:
679	return -ERANGE;
680	}
681
682	/* Precalculate acceptable difference between the minor time in the
683	* packet prefix and the last MCDI time sync event. We expect the
684	* packet prefix timestamp to be after of sync event by up to one
685	* sync event interval (0.25s) but we allow it to exceed this by a
686	* fuzz factor of (0.1s)
687	*/
688	ptp->nic_time.sync_event_diff_min = ptp->nic_time.minor_max
689	- (ptp->nic_time.minor_max / 10);
690	ptp->nic_time.sync_event_diff_max = (ptp->nic_time.minor_max / 4)
691	+ (ptp->nic_time.minor_max / 10);
692
693	/* MC_CMD_PTP_OP_GET_ATTRIBUTES has been extended twice from an older
694	* operation MC_CMD_PTP_OP_GET_TIME_FORMAT. The function now may return
695	* a value to use for the minimum acceptable corrected synchronization
696	* window and may return further capabilities.
697	* If we have the extra information store it. For older firmware that
698	* does not implement the extended command use the default value.
699	*/
700	if (rc == 0 &&
701	out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_CAPABILITIES_OFST)
702	ptp->min_synchronisation_ns =
703	MCDI_DWORD(outbuf,
704	PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
705	else
706	ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
707
708	if (rc == 0 &&
709	out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
710	ptp->capabilities = MCDI_DWORD(outbuf,
711	PTP_OUT_GET_ATTRIBUTES_CAPABILITIES);
712	else
713	ptp->capabilities = 0;
714
715	/* Set up the shift for conversion between frequency
716	* adjustments in parts-per-billion and the fixed-point
717	* fractional ns format that the adapter uses.
718	*/
719	if (ptp->capabilities & (1 << MC_CMD_PTP_OUT_GET_ATTRIBUTES_FP44_FREQ_ADJ_LBN))
720	ptp->adjfreq_ppb_shift = PPB_SHIFT_FP44;
721	else
722	ptp->adjfreq_ppb_shift = PPB_SHIFT_FP40;
723
724	return 0;
725	}
726
727	/* Get PTP timestamp corrections */
728	static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
729	{
730	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN);
731	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN);
732	int rc;
733	size_t out_len;
734
735	/* Get the timestamp corrections from the NIC. If this operation is
736	* not supported (older NICs) then no correction is required.
737	*/
738	MCDI_SET_DWORD(inbuf, PTP_IN_OP,
739	MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS);
740	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
741
742	rc = efx_siena_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
743	outbuf, outlen: sizeof(outbuf), outlen_actual: &out_len);
744	if (rc == 0) {
745	efx->ptp_data->ts_corrections.ptp_tx = MCDI_DWORD(outbuf,
746	PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT);
747	efx->ptp_data->ts_corrections.ptp_rx = MCDI_DWORD(outbuf,
748	PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE);
749	efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf,
750	PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT);
751	efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf,
752	PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN);
753
754	if (out_len >= MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN) {
755	efx->ptp_data->ts_corrections.general_tx = MCDI_DWORD(
756	outbuf,
757	PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_TX);
758	efx->ptp_data->ts_corrections.general_rx = MCDI_DWORD(
759	outbuf,
760	PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_RX);
761	} else {
762	efx->ptp_data->ts_corrections.general_tx =
763	efx->ptp_data->ts_corrections.ptp_tx;
764	efx->ptp_data->ts_corrections.general_rx =
765	efx->ptp_data->ts_corrections.ptp_rx;
766	}
767	} else if (rc == -EINVAL) {
768	efx->ptp_data->ts_corrections.ptp_tx = 0;
769	efx->ptp_data->ts_corrections.ptp_rx = 0;
770	efx->ptp_data->ts_corrections.pps_out = 0;
771	efx->ptp_data->ts_corrections.pps_in = 0;
772	efx->ptp_data->ts_corrections.general_tx = 0;
773	efx->ptp_data->ts_corrections.general_rx = 0;
774	} else {
775	efx_siena_mcdi_display_error(efx, MC_CMD_PTP, inlen: sizeof(inbuf),
776	outbuf, outlen: sizeof(outbuf), rc);
777	return rc;
778	}
779
780	return 0;
781	}
782
783	/* Enable MCDI PTP support. */
784	static int efx_ptp_enable(struct efx_nic *efx)
785	{
786	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
787	MCDI_DECLARE_BUF_ERR(outbuf);
788	int rc;
789
790	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE);
791	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
792	MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE,
793	efx->ptp_data->channel ?
794	efx->ptp_data->channel->channel : 0);
795	MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode);
796
797	rc = efx_siena_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
798	outbuf, outlen: sizeof(outbuf), NULL);
799	rc = (rc == -EALREADY) ? 0 : rc;
800	if (rc)
801	efx_siena_mcdi_display_error(efx, MC_CMD_PTP,
802	MC_CMD_PTP_IN_ENABLE_LEN,
803	outbuf, outlen: sizeof(outbuf), rc);
804	return rc;
805	}
806
807	/* Disable MCDI PTP support.
808	*
809	* Note that this function should never rely on the presence of ptp_data -
810	* may be called before that exists.
811	*/
812	static int efx_ptp_disable(struct efx_nic *efx)
813	{
814	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
815	MCDI_DECLARE_BUF_ERR(outbuf);
816	int rc;
817
818	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE);
819	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
820	rc = efx_siena_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
821	outbuf, outlen: sizeof(outbuf), NULL);
822	rc = (rc == -EALREADY) ? 0 : rc;
823	/* If we get ENOSYS, the NIC doesn't support PTP, and thus this function
824	* should only have been called during probe.
825	*/
826	if (rc == -ENOSYS || rc == -EPERM)
827	pci_info(efx->pci_dev, "no PTP support\n");
828	else if (rc)
829	efx_siena_mcdi_display_error(efx, MC_CMD_PTP,
830	MC_CMD_PTP_IN_DISABLE_LEN,
831	outbuf, outlen: sizeof(outbuf), rc);
832	return rc;
833	}
834
835	static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
836	{
837	struct sk_buff *skb;
838
839	while ((skb = skb_dequeue(list: q))) {
840	local_bh_disable();
841	netif_receive_skb(skb);
842	local_bh_enable();
843	}
844	}
845
846	static void efx_ptp_handle_no_channel(struct efx_nic *efx)
847	{
848	netif_err(efx, drv, efx->net_dev,
849	"ERROR: PTP requires MSI-X and 1 additional interrupt"
850	"vector. PTP disabled\n");
851	}
852
853	/* Repeatedly send the host time to the MC which will capture the hardware
854	* time.
855	*/
856	static void efx_ptp_send_times(struct efx_nic *efx,
857	struct pps_event_time *last_time)
858	{
859	struct pps_event_time now;
860	struct timespec64 limit;
861	struct efx_ptp_data *ptp = efx->ptp_data;
862	int *mc_running = ptp->start.addr;
863
864	pps_get_ts(ts: &now);
865	limit = now.ts_real;
866	timespec64_add_ns(a: &limit, SYNCHRONISE_PERIOD_NS);
867
868	/* Write host time for specified period or until MC is done */
869	while ((timespec64_compare(lhs: &now.ts_real, rhs: &limit) < 0) &&
870	READ_ONCE(*mc_running)) {
871	struct timespec64 update_time;
872	unsigned int host_time;
873
874	/* Don't update continuously to avoid saturating the PCIe bus */
875	update_time = now.ts_real;
876	timespec64_add_ns(a: &update_time, SYNCHRONISATION_GRANULARITY_NS);
877	do {
878	pps_get_ts(ts: &now);
879	} while ((timespec64_compare(lhs: &now.ts_real, rhs: &update_time) < 0) &&
880	READ_ONCE(*mc_running));
881
882	/* Synchronise NIC with single word of time only */
883	host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS |
884	now.ts_real.tv_nsec);
885	/* Update host time in NIC memory */
886	efx->type->ptp_write_host_time(efx, host_time);
887	}
888	*last_time = now;
889	}
890
891	/* Read a timeset from the MC's results and partial process. */
892	static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data),
893	struct efx_ptp_timeset *timeset)
894	{
895	unsigned start_ns, end_ns;
896
897	timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART);
898	timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR);
899	timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR);
900	timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND),
901	timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
902
903	/* Ignore seconds */
904	start_ns = timeset->host_start & MC_NANOSECOND_MASK;
905	end_ns = timeset->host_end & MC_NANOSECOND_MASK;
906	/* Allow for rollover */
907	if (end_ns < start_ns)
908	end_ns += NSEC_PER_SEC;
909	/* Determine duration of operation */
910	timeset->window = end_ns - start_ns;
911	}
912
913	/* Process times received from MC.
914	*
915	* Extract times from returned results, and establish the minimum value
916	* seen. The minimum value represents the "best" possible time and events
917	* too much greater than this are rejected - the machine is, perhaps, too
918	* busy. A number of readings are taken so that, hopefully, at least one good
919	* synchronisation will be seen in the results.
920	*/
921	static int
922	efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf),
923	size_t response_length,
924	const struct pps_event_time *last_time)
925	{
926	unsigned number_readings =
927	MCDI_VAR_ARRAY_LEN(response_length,
928	PTP_OUT_SYNCHRONIZE_TIMESET);
929	unsigned i;
930	unsigned ngood = 0;
931	unsigned last_good = 0;
932	struct efx_ptp_data *ptp = efx->ptp_data;
933	u32 last_sec;
934	u32 start_sec;
935	struct timespec64 delta;
936	ktime_t mc_time;
937
938	if (number_readings == 0)
939	return -EAGAIN;
940
941	/* Read the set of results and find the last good host-MC
942	* synchronization result. The MC times when it finishes reading the
943	* host time so the corrected window time should be fairly constant
944	* for a given platform. Increment stats for any results that appear
945	* to be erroneous.
946	*/
947	for (i = 0; i < number_readings; i++) {
948	s32 window, corrected;
949	struct timespec64 wait;
950
951	efx_ptp_read_timeset(
952	MCDI_ARRAY_STRUCT_PTR(synch_buf,
953	PTP_OUT_SYNCHRONIZE_TIMESET, i),
954	timeset: &ptp->timeset[i]);
955
956	wait = ktime_to_timespec64(
957	ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0));
958	window = ptp->timeset[i].window;
959	corrected = window - wait.tv_nsec;
960
961	/* We expect the uncorrected synchronization window to be at
962	* least as large as the interval between host start and end
963	* times. If it is smaller than this then this is mostly likely
964	* to be a consequence of the host's time being adjusted.
965	* Check that the corrected sync window is in a reasonable
966	* range. If it is out of range it is likely to be because an
967	* interrupt or other delay occurred between reading the system
968	* time and writing it to MC memory.
969	*/
970	if (window < SYNCHRONISATION_GRANULARITY_NS) {
971	++ptp->invalid_sync_windows;
972	} else if (corrected >= MAX_SYNCHRONISATION_NS) {
973	++ptp->oversize_sync_windows;
974	} else if (corrected < ptp->min_synchronisation_ns) {
975	++ptp->undersize_sync_windows;
976	} else {
977	ngood++;
978	last_good = i;
979	}
980	}
981
982	if (ngood == 0) {
983	netif_warn(efx, drv, efx->net_dev,
984	"PTP no suitable synchronisations\n");
985	return -EAGAIN;
986	}
987
988	/* Calculate delay from last good sync (host time) to last_time.
989	* It is possible that the seconds rolled over between taking
990	* the start reading and the last value written by the host. The
991	* timescales are such that a gap of more than one second is never
992	* expected. delta is *not* normalised.
993	*/
994	start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS;
995	last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK;
996	if (start_sec != last_sec &&
997	((start_sec + 1) & MC_SECOND_MASK) != last_sec) {
998	netif_warn(efx, hw, efx->net_dev,
999	"PTP bad synchronisation seconds\n");
1000	return -EAGAIN;
1001	}
1002	delta.tv_sec = (last_sec - start_sec) & 1;
1003	delta.tv_nsec =
1004	last_time->ts_real.tv_nsec -
1005	(ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
1006
1007	/* Convert the NIC time at last good sync into kernel time.
1008	* No correction is required - this time is the output of a
1009	* firmware process.
1010	*/
1011	mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
1012	ptp->timeset[last_good].minor, 0);
1013
1014	/* Calculate delay from NIC top of second to last_time */
1015	delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec;
1016
1017	/* Set PPS timestamp to match NIC top of second */
1018	ptp->host_time_pps = *last_time;
1019	pps_sub_ts(ts: &ptp->host_time_pps, delta);
1020
1021	return 0;
1022	}
1023
1024	/* Synchronize times between the host and the MC */
1025	static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
1026	{
1027	struct efx_ptp_data *ptp = efx->ptp_data;
1028	MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX);
1029	size_t response_length;
1030	int rc;
1031	unsigned long timeout;
1032	struct pps_event_time last_time = {};
1033	unsigned int loops = 0;
1034	int *start = ptp->start.addr;
1035
1036	MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE);
1037	MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0);
1038	MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS,
1039	num_readings);
1040	MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
1041	ptp->start.dma_addr);
1042
1043	/* Clear flag that signals MC ready */
1044	WRITE_ONCE(*start, 0);
1045	rc = efx_siena_mcdi_rpc_start(efx, MC_CMD_PTP, inbuf: synch_buf,
1046	MC_CMD_PTP_IN_SYNCHRONIZE_LEN);
1047	EFX_WARN_ON_ONCE_PARANOID(rc);
1048
1049	/* Wait for start from MCDI (or timeout) */
1050	timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS);
1051	while (!READ_ONCE(*start) && (time_before(jiffies, timeout))) {
1052	udelay(20); /* Usually start MCDI execution quickly */
1053	loops++;
1054	}
1055
1056	if (loops <= 1)
1057	++ptp->fast_syncs;
1058	if (!time_before(jiffies, timeout))
1059	++ptp->sync_timeouts;
1060
1061	if (READ_ONCE(*start))
1062	efx_ptp_send_times(efx, last_time: &last_time);
1063
1064	/* Collect results */
1065	rc = efx_siena_mcdi_rpc_finish(efx, MC_CMD_PTP,
1066	MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
1067	outbuf: synch_buf, outlen: sizeof(synch_buf),
1068	outlen_actual: &response_length);
1069	if (rc == 0) {
1070	rc = efx_ptp_process_times(efx, synch_buf, response_length,
1071	last_time: &last_time);
1072	if (rc == 0)
1073	++ptp->good_syncs;
1074	else
1075	++ptp->no_time_syncs;
1076	}
1077
1078	/* Increment the bad syncs counter if the synchronize fails, whatever
1079	* the reason.
1080	*/
1081	if (rc != 0)
1082	++ptp->bad_syncs;
1083
1084	return rc;
1085	}
1086
1087	/* Transmit a PTP packet via the dedicated hardware timestamped queue. */
1088	static void efx_ptp_xmit_skb_queue(struct efx_nic *efx, struct sk_buff *skb)
1089	{
1090	struct efx_ptp_data *ptp_data = efx->ptp_data;
1091	u8 type = efx_tx_csum_type_skb(skb);
1092	struct efx_tx_queue *tx_queue;
1093
1094	tx_queue = efx_channel_get_tx_queue(channel: ptp_data->channel, type);
1095	if (tx_queue && tx_queue->timestamping) {
1096	efx_enqueue_skb(tx_queue, skb);
1097	} else {
1098	WARN_ONCE(1, "PTP channel has no timestamped tx queue\n");
1099	dev_kfree_skb_any(skb);
1100	}
1101	}
1102
1103	/* Transmit a PTP packet, via the MCDI interface, to the wire. */
1104	static void efx_ptp_xmit_skb_mc(struct efx_nic *efx, struct sk_buff *skb)
1105	{
1106	struct efx_ptp_data *ptp_data = efx->ptp_data;
1107	struct skb_shared_hwtstamps timestamps;
1108	int rc = -EIO;
1109	MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
1110	size_t len;
1111
1112	MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT);
1113	MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0);
1114	MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len);
1115	if (skb_shinfo(skb)->nr_frags != 0) {
1116	rc = skb_linearize(skb);
1117	if (rc != 0)
1118	goto fail;
1119	}
1120
1121	if (skb->ip_summed == CHECKSUM_PARTIAL) {
1122	rc = skb_checksum_help(skb);
1123	if (rc != 0)
1124	goto fail;
1125	}
1126	skb_copy_from_linear_data(skb,
1127	MCDI_PTR(ptp_data->txbuf,
1128	PTP_IN_TRANSMIT_PACKET),
1129	len: skb->len);
1130	rc = efx_siena_mcdi_rpc(efx, MC_CMD_PTP, inbuf: ptp_data->txbuf,
1131	MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len), outbuf: txtime,
1132	outlen: sizeof(txtime), outlen_actual: &len);
1133	if (rc != 0)
1134	goto fail;
1135
1136	memset(&timestamps, 0, sizeof(timestamps));
1137	timestamps.hwtstamp = ptp_data->nic_to_kernel_time(
1138	MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR),
1139	MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR),
1140	ptp_data->ts_corrections.ptp_tx);
1141
1142	skb_tstamp_tx(orig_skb: skb, hwtstamps: &timestamps);
1143
1144	rc = 0;
1145
1146	fail:
1147	dev_kfree_skb_any(skb);
1148
1149	return;
1150	}
1151
1152	static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
1153	{
1154	struct efx_ptp_data *ptp = efx->ptp_data;
1155	struct list_head *cursor;
1156	struct list_head *next;
1157
1158	if (ptp->rx_ts_inline)
1159	return;
1160
1161	/* Drop time-expired events */
1162	spin_lock_bh(lock: &ptp->evt_lock);
1163	list_for_each_safe(cursor, next, &ptp->evt_list) {
1164	struct efx_ptp_event_rx *evt;
1165
1166	evt = list_entry(cursor, struct efx_ptp_event_rx,
1167	link);
1168	if (time_after(jiffies, evt->expiry)) {
1169	list_move(list: &evt->link, head: &ptp->evt_free_list);
1170	netif_warn(efx, hw, efx->net_dev,
1171	"PTP rx event dropped\n");
1172	}
1173	}
1174	spin_unlock_bh(lock: &ptp->evt_lock);
1175	}
1176
1177	static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
1178	struct sk_buff *skb)
1179	{
1180	struct efx_ptp_data *ptp = efx->ptp_data;
1181	bool evts_waiting;
1182	struct list_head *cursor;
1183	struct list_head *next;
1184	struct efx_ptp_match *match;
1185	enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED;
1186
1187	WARN_ON_ONCE(ptp->rx_ts_inline);
1188
1189	spin_lock_bh(lock: &ptp->evt_lock);
1190	evts_waiting = !list_empty(head: &ptp->evt_list);
1191	spin_unlock_bh(lock: &ptp->evt_lock);
1192
1193	if (!evts_waiting)
1194	return PTP_PACKET_STATE_UNMATCHED;
1195
1196	match = (struct efx_ptp_match *)skb->cb;
1197	/* Look for a matching timestamp in the event queue */
1198	spin_lock_bh(lock: &ptp->evt_lock);
1199	list_for_each_safe(cursor, next, &ptp->evt_list) {
1200	struct efx_ptp_event_rx *evt;
1201
1202	evt = list_entry(cursor, struct efx_ptp_event_rx, link);
1203	if ((evt->seq0 == match->words[0]) &&
1204	(evt->seq1 == match->words[1])) {
1205	struct skb_shared_hwtstamps *timestamps;
1206
1207	/* Match - add in hardware timestamp */
1208	timestamps = skb_hwtstamps(skb);
1209	timestamps->hwtstamp = evt->hwtimestamp;
1210
1211	match->state = PTP_PACKET_STATE_MATCHED;
1212	rc = PTP_PACKET_STATE_MATCHED;
1213	list_move(list: &evt->link, head: &ptp->evt_free_list);
1214	break;
1215	}
1216	}
1217	spin_unlock_bh(lock: &ptp->evt_lock);
1218
1219	return rc;
1220	}
1221
1222	/* Process any queued receive events and corresponding packets
1223	*
1224	* q is returned with all the packets that are ready for delivery.
1225	*/
1226	static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
1227	{
1228	struct efx_ptp_data *ptp = efx->ptp_data;
1229	struct sk_buff *skb;
1230
1231	while ((skb = skb_dequeue(list: &ptp->rxq))) {
1232	struct efx_ptp_match *match;
1233
1234	match = (struct efx_ptp_match *)skb->cb;
1235	if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) {
1236	__skb_queue_tail(list: q, newsk: skb);
1237	} else if (efx_ptp_match_rx(efx, skb) ==
1238	PTP_PACKET_STATE_MATCHED) {
1239	__skb_queue_tail(list: q, newsk: skb);
1240	} else if (time_after(jiffies, match->expiry)) {
1241	match->state = PTP_PACKET_STATE_TIMED_OUT;
1242	++ptp->rx_no_timestamp;
1243	__skb_queue_tail(list: q, newsk: skb);
1244	} else {
1245	/* Replace unprocessed entry and stop */
1246	skb_queue_head(list: &ptp->rxq, newsk: skb);
1247	break;
1248	}
1249	}
1250	}
1251
1252	/* Complete processing of a received packet */
1253	static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1254	{
1255	local_bh_disable();
1256	netif_receive_skb(skb);
1257	local_bh_enable();
1258	}
1259
1260	static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
1261	{
1262	struct efx_ptp_data *ptp = efx->ptp_data;
1263
1264	if (ptp->rxfilter_installed) {
1265	efx_filter_remove_id_safe(efx, priority: EFX_FILTER_PRI_REQUIRED,
1266	filter_id: ptp->rxfilter_general);
1267	efx_filter_remove_id_safe(efx, priority: EFX_FILTER_PRI_REQUIRED,
1268	filter_id: ptp->rxfilter_event);
1269	ptp->rxfilter_installed = false;
1270	}
1271	}
1272
1273	static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1274	{
1275	struct efx_ptp_data *ptp = efx->ptp_data;
1276	struct efx_filter_spec rxfilter;
1277	int rc;
1278
1279	if (!ptp->channel || ptp->rxfilter_installed)
1280	return 0;
1281
1282	/* Must filter on both event and general ports to ensure
1283	* that there is no packet re-ordering.
1284	*/
1285	efx_filter_init_rx(spec: &rxfilter, priority: EFX_FILTER_PRI_REQUIRED, flags: 0,
1286	rxq_id: efx_rx_queue_index(
1287	rx_queue: efx_channel_get_rx_queue(channel: ptp->channel)));
1288	rc = efx_filter_set_ipv4_local(spec: &rxfilter, IPPROTO_UDP,
1289	htonl(PTP_ADDRESS),
1290	htons(PTP_EVENT_PORT));
1291	if (rc != 0)
1292	return rc;
1293
1294	rc = efx_filter_insert_filter(efx, spec: &rxfilter, replace_equal: true);
1295	if (rc < 0)
1296	return rc;
1297	ptp->rxfilter_event = rc;
1298
1299	efx_filter_init_rx(spec: &rxfilter, priority: EFX_FILTER_PRI_REQUIRED, flags: 0,
1300	rxq_id: efx_rx_queue_index(
1301	rx_queue: efx_channel_get_rx_queue(channel: ptp->channel)));
1302	rc = efx_filter_set_ipv4_local(spec: &rxfilter, IPPROTO_UDP,
1303	htonl(PTP_ADDRESS),
1304	htons(PTP_GENERAL_PORT));
1305	if (rc != 0)
1306	goto fail;
1307
1308	rc = efx_filter_insert_filter(efx, spec: &rxfilter, replace_equal: true);
1309	if (rc < 0)
1310	goto fail;
1311	ptp->rxfilter_general = rc;
1312
1313	ptp->rxfilter_installed = true;
1314	return 0;
1315
1316	fail:
1317	efx_filter_remove_id_safe(efx, priority: EFX_FILTER_PRI_REQUIRED,
1318	filter_id: ptp->rxfilter_event);
1319	return rc;
1320	}
1321
1322	static int efx_ptp_start(struct efx_nic *efx)
1323	{
1324	struct efx_ptp_data *ptp = efx->ptp_data;
1325	int rc;
1326
1327	ptp->reset_required = false;
1328
1329	rc = efx_ptp_insert_multicast_filters(efx);
1330	if (rc)
1331	return rc;
1332
1333	rc = efx_ptp_enable(efx);
1334	if (rc != 0)
1335	goto fail;
1336
1337	ptp->evt_frag_idx = 0;
1338	ptp->current_adjfreq = 0;
1339
1340	return 0;
1341
1342	fail:
1343	efx_ptp_remove_multicast_filters(efx);
1344	return rc;
1345	}
1346
1347	static int efx_ptp_stop(struct efx_nic *efx)
1348	{
1349	struct efx_ptp_data *ptp = efx->ptp_data;
1350	struct list_head *cursor;
1351	struct list_head *next;
1352	int rc;
1353
1354	if (ptp == NULL)
1355	return 0;
1356
1357	rc = efx_ptp_disable(efx);
1358
1359	efx_ptp_remove_multicast_filters(efx);
1360
1361	/* Make sure RX packets are really delivered */
1362	efx_ptp_deliver_rx_queue(q: &efx->ptp_data->rxq);
1363	skb_queue_purge(list: &efx->ptp_data->txq);
1364
1365	/* Drop any pending receive events */
1366	spin_lock_bh(lock: &efx->ptp_data->evt_lock);
1367	list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) {
1368	list_move(list: cursor, head: &efx->ptp_data->evt_free_list);
1369	}
1370	spin_unlock_bh(lock: &efx->ptp_data->evt_lock);
1371
1372	return rc;
1373	}
1374
1375	static int efx_ptp_restart(struct efx_nic *efx)
1376	{
1377	if (efx->ptp_data && efx->ptp_data->enabled)
1378	return efx_ptp_start(efx);
1379	return 0;
1380	}
1381
1382	static void efx_ptp_pps_worker(struct work_struct *work)
1383	{
1384	struct efx_ptp_data *ptp =
1385	container_of(work, struct efx_ptp_data, pps_work);
1386	struct efx_nic *efx = ptp->efx;
1387	struct ptp_clock_event ptp_evt;
1388
1389	if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
1390	return;
1391
1392	ptp_evt.type = PTP_CLOCK_PPSUSR;
1393	ptp_evt.pps_times = ptp->host_time_pps;
1394	ptp_clock_event(ptp: ptp->phc_clock, event: &ptp_evt);
1395	}
1396
1397	static void efx_ptp_worker(struct work_struct *work)
1398	{
1399	struct efx_ptp_data *ptp_data =
1400	container_of(work, struct efx_ptp_data, work);
1401	struct efx_nic *efx = ptp_data->efx;
1402	struct sk_buff *skb;
1403	struct sk_buff_head tempq;
1404
1405	if (ptp_data->reset_required) {
1406	efx_ptp_stop(efx);
1407	efx_ptp_start(efx);
1408	return;
1409	}
1410
1411	efx_ptp_drop_time_expired_events(efx);
1412
1413	__skb_queue_head_init(list: &tempq);
1414	efx_ptp_process_events(efx, q: &tempq);
1415
1416	while ((skb = skb_dequeue(list: &ptp_data->txq)))
1417	ptp_data->xmit_skb(efx, skb);
1418
1419	while ((skb = __skb_dequeue(list: &tempq)))
1420	efx_ptp_process_rx(efx, skb);
1421	}
1422
1423	static const struct ptp_clock_info efx_phc_clock_info = {
1424	.owner = THIS_MODULE,
1425	.name = "sfc_siena",
1426	.max_adj = MAX_PPB,
1427	.n_alarm = 0,
1428	.n_ext_ts = 0,
1429	.n_per_out = 0,
1430	.n_pins = 0,
1431	.pps = 1,
1432	.adjfine = efx_phc_adjfine,
1433	.adjtime = efx_phc_adjtime,
1434	.gettime64 = efx_phc_gettime,
1435	.settime64 = efx_phc_settime,
1436	.enable = efx_phc_enable,
1437	};
1438
1439	/* Initialise PTP state. */
1440	static int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1441	{
1442	struct efx_ptp_data *ptp;
1443	int rc = 0;
1444	unsigned int pos;
1445
1446	ptp = kzalloc(size: sizeof(struct efx_ptp_data), GFP_KERNEL);
1447	efx->ptp_data = ptp;
1448	if (!efx->ptp_data)
1449	return -ENOMEM;
1450
1451	ptp->efx = efx;
1452	ptp->channel = channel;
1453	ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1454
1455	rc = efx_siena_alloc_buffer(efx, buffer: &ptp->start, len: sizeof(int), GFP_KERNEL);
1456	if (rc != 0)
1457	goto fail1;
1458
1459	skb_queue_head_init(list: &ptp->rxq);
1460	skb_queue_head_init(list: &ptp->txq);
1461	ptp->workwq = create_singlethread_workqueue("sfc_siena_ptp");
1462	if (!ptp->workwq) {
1463	rc = -ENOMEM;
1464	goto fail2;
1465	}
1466
1467	if (efx_siena_ptp_use_mac_tx_timestamps(efx)) {
1468	ptp->xmit_skb = efx_ptp_xmit_skb_queue;
1469	/* Request sync events on this channel. */
1470	channel->sync_events_state = SYNC_EVENTS_QUIESCENT;
1471	} else {
1472	ptp->xmit_skb = efx_ptp_xmit_skb_mc;
1473	}
1474
1475	INIT_WORK(&ptp->work, efx_ptp_worker);
1476	ptp->config.flags = 0;
1477	ptp->config.tx_type = HWTSTAMP_TX_OFF;
1478	ptp->config.rx_filter = HWTSTAMP_FILTER_NONE;
1479	INIT_LIST_HEAD(list: &ptp->evt_list);
1480	INIT_LIST_HEAD(list: &ptp->evt_free_list);
1481	spin_lock_init(&ptp->evt_lock);
1482	for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++)
1483	list_add(new: &ptp->rx_evts[pos].link, head: &ptp->evt_free_list);
1484
1485	/* Get the NIC PTP attributes and set up time conversions */
1486	rc = efx_ptp_get_attributes(efx);
1487	if (rc < 0)
1488	goto fail3;
1489
1490	/* Get the timestamp corrections */
1491	rc = efx_ptp_get_timestamp_corrections(efx);
1492	if (rc < 0)
1493	goto fail3;
1494
1495	if (efx->mcdi->fn_flags &
1496	(1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) {
1497	ptp->phc_clock_info = efx_phc_clock_info;
1498	ptp->phc_clock = ptp_clock_register(info: &ptp->phc_clock_info,
1499	parent: &efx->pci_dev->dev);
1500	if (IS_ERR(ptr: ptp->phc_clock)) {
1501	rc = PTR_ERR(ptr: ptp->phc_clock);
1502	goto fail3;
1503	} else if (ptp->phc_clock) {
1504	INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1505	ptp->pps_workwq =
1506	create_singlethread_workqueue("sfc_siena_pps");
1507	if (!ptp->pps_workwq) {
1508	rc = -ENOMEM;
1509	goto fail4;
1510	}
1511	}
1512	}
1513	ptp->nic_ts_enabled = false;
1514
1515	return 0;
1516	fail4:
1517	ptp_clock_unregister(ptp: efx->ptp_data->phc_clock);
1518
1519	fail3:
1520	destroy_workqueue(wq: efx->ptp_data->workwq);
1521
1522	fail2:
1523	efx_siena_free_buffer(efx, buffer: &ptp->start);
1524
1525	fail1:
1526	kfree(objp: efx->ptp_data);
1527	efx->ptp_data = NULL;
1528
1529	return rc;
1530	}
1531
1532	/* Initialise PTP channel.
1533	*
1534	* Setting core_index to zero causes the queue to be initialised and doesn't
1535	* overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
1536	*/
1537	static int efx_ptp_probe_channel(struct efx_channel *channel)
1538	{
1539	struct efx_nic *efx = channel->efx;
1540	int rc;
1541
1542	channel->irq_moderation_us = 0;
1543	channel->rx_queue.core_index = 0;
1544
1545	rc = efx_ptp_probe(efx, channel);
1546	/* Failure to probe PTP is not fatal; this channel will just not be
1547	* used for anything.
1548	* In the case of EPERM, efx_ptp_probe will print its own message (in
1549	* efx_ptp_get_attributes()), so we don't need to.
1550	*/
1551	if (rc && rc != -EPERM)
1552	netif_warn(efx, drv, efx->net_dev,
1553	"Failed to probe PTP, rc=%d\n", rc);
1554	return 0;
1555	}
1556
1557	static void efx_ptp_remove(struct efx_nic *efx)
1558	{
1559	if (!efx->ptp_data)
1560	return;
1561
1562	(void)efx_ptp_disable(efx);
1563
1564	cancel_work_sync(work: &efx->ptp_data->work);
1565	if (efx->ptp_data->pps_workwq)
1566	cancel_work_sync(work: &efx->ptp_data->pps_work);
1567
1568	skb_queue_purge(list: &efx->ptp_data->rxq);
1569	skb_queue_purge(list: &efx->ptp_data->txq);
1570
1571	if (efx->ptp_data->phc_clock) {
1572	destroy_workqueue(wq: efx->ptp_data->pps_workwq);
1573	ptp_clock_unregister(ptp: efx->ptp_data->phc_clock);
1574	}
1575
1576	destroy_workqueue(wq: efx->ptp_data->workwq);
1577
1578	efx_siena_free_buffer(efx, buffer: &efx->ptp_data->start);
1579	kfree(objp: efx->ptp_data);
1580	efx->ptp_data = NULL;
1581	}
1582
1583	static void efx_ptp_remove_channel(struct efx_channel *channel)
1584	{
1585	efx_ptp_remove(efx: channel->efx);
1586	}
1587
1588	static void efx_ptp_get_channel_name(struct efx_channel *channel,
1589	char *buf, size_t len)
1590	{
1591	snprintf(buf, size: len, fmt: "%s-ptp", channel->efx->name);
1592	}
1593
1594	/* Determine whether this packet should be processed by the PTP module
1595	* or transmitted conventionally.
1596	*/
1597	bool efx_siena_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1598	{
1599	return efx->ptp_data &&
1600	efx->ptp_data->enabled &&
1601	skb->len >= PTP_MIN_LENGTH &&
1602	skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1603	likely(skb->protocol == htons(ETH_P_IP)) &&
1604	skb_transport_header_was_set(skb) &&
1605	skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1606	ip_hdr(skb)->protocol == IPPROTO_UDP &&
1607	skb_headlen(skb) >=
1608	skb_transport_offset(skb) + sizeof(struct udphdr) &&
1609	udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1610	}
1611
1612	/* Receive a PTP packet. Packets are queued until the arrival of
1613	* the receive timestamp from the MC - this will probably occur after the
1614	* packet arrival because of the processing in the MC.
1615	*/
1616	static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1617	{
1618	struct efx_nic *efx = channel->efx;
1619	struct efx_ptp_data *ptp = efx->ptp_data;
1620	struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1621	u8 *match_data_012, *match_data_345;
1622	unsigned int version;
1623	u8 *data;
1624
1625	match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1626
1627	/* Correct version? */
1628	if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1629	if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1630	return false;
1631	}
1632	data = skb->data;
1633	version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1634	if (version != PTP_VERSION_V1) {
1635	return false;
1636	}
1637
1638	/* PTP V1 uses all six bytes of the UUID to match the packet
1639	* to the timestamp
1640	*/
1641	match_data_012 = data + PTP_V1_UUID_OFFSET;
1642	match_data_345 = data + PTP_V1_UUID_OFFSET + 3;
1643	} else {
1644	if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1645	return false;
1646	}
1647	data = skb->data;
1648	version = data[PTP_V2_VERSION_OFFSET];
1649	if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1650	return false;
1651	}
1652
1653	/* The original V2 implementation uses bytes 2-7 of
1654	* the UUID to match the packet to the timestamp. This
1655	* discards two of the bytes of the MAC address used
1656	* to create the UUID (SF bug 33070). The PTP V2
1657	* enhanced mode fixes this issue and uses bytes 0-2
1658	* and byte 5-7 of the UUID.
1659	*/
1660	match_data_345 = data + PTP_V2_UUID_OFFSET + 5;
1661	if (ptp->mode == MC_CMD_PTP_MODE_V2) {
1662	match_data_012 = data + PTP_V2_UUID_OFFSET + 2;
1663	} else {
1664	match_data_012 = data + PTP_V2_UUID_OFFSET + 0;
1665	BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED);
1666	}
1667	}
1668
1669	/* Does this packet require timestamping? */
1670	if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1671	match->state = PTP_PACKET_STATE_UNMATCHED;
1672
1673	/* We expect the sequence number to be in the same position in
1674	* the packet for PTP V1 and V2
1675	*/
1676	BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1677	BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1678
1679	/* Extract UUID/Sequence information */
1680	match->words[0] = (match_data_012[0] |
1681	(match_data_012[1] << 8) |
1682	(match_data_012[2] << 16) |
1683	(match_data_345[0] << 24));
1684	match->words[1] = (match_data_345[1] |
1685	(match_data_345[2] << 8) |
1686	(data[PTP_V1_SEQUENCE_OFFSET +
1687	PTP_V1_SEQUENCE_LENGTH - 1] <<
1688	16));
1689	} else {
1690	match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1691	}
1692
1693	skb_queue_tail(list: &ptp->rxq, newsk: skb);
1694	queue_work(wq: ptp->workwq, work: &ptp->work);
1695
1696	return true;
1697	}
1698
1699	/* Transmit a PTP packet. This has to be transmitted by the MC
1700	* itself, through an MCDI call. MCDI calls aren't permitted
1701	* in the transmit path so defer the actual transmission to a suitable worker.
1702	*/
1703	int efx_siena_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1704	{
1705	struct efx_ptp_data *ptp = efx->ptp_data;
1706
1707	skb_queue_tail(list: &ptp->txq, newsk: skb);
1708
1709	if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1710	(skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1711	efx_xmit_hwtstamp_pending(skb);
1712	queue_work(wq: ptp->workwq, work: &ptp->work);
1713
1714	return NETDEV_TX_OK;
1715	}
1716
1717	int efx_siena_ptp_get_mode(struct efx_nic *efx)
1718	{
1719	return efx->ptp_data->mode;
1720	}
1721
1722	int efx_siena_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1723	unsigned int new_mode)
1724	{
1725	if ((enable_wanted != efx->ptp_data->enabled) ||
1726	(enable_wanted && (efx->ptp_data->mode != new_mode))) {
1727	int rc = 0;
1728
1729	if (enable_wanted) {
1730	/* Change of mode requires disable */
1731	if (efx->ptp_data->enabled &&
1732	(efx->ptp_data->mode != new_mode)) {
1733	efx->ptp_data->enabled = false;
1734	rc = efx_ptp_stop(efx);
1735	if (rc != 0)
1736	return rc;
1737	}
1738
1739	/* Set new operating mode and establish
1740	* baseline synchronisation, which must
1741	* succeed.
1742	*/
1743	efx->ptp_data->mode = new_mode;
1744	if (netif_running(dev: efx->net_dev))
1745	rc = efx_ptp_start(efx);
1746	if (rc == 0) {
1747	rc = efx_ptp_synchronize(efx,
1748	PTP_SYNC_ATTEMPTS * 2);
1749	if (rc != 0)
1750	efx_ptp_stop(efx);
1751	}
1752	} else {
1753	rc = efx_ptp_stop(efx);
1754	}
1755
1756	if (rc != 0)
1757	return rc;
1758
1759	efx->ptp_data->enabled = enable_wanted;
1760	}
1761
1762	return 0;
1763	}
1764
1765	static int efx_ptp_ts_init(struct efx_nic *efx,
1766	struct kernel_hwtstamp_config *init)
1767	{
1768	int rc;
1769
1770	if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1771	(init->tx_type != HWTSTAMP_TX_ON))
1772	return -ERANGE;
1773
1774	rc = efx->type->ptp_set_ts_config(efx, init);
1775	if (rc)
1776	return rc;
1777
1778	efx->ptp_data->config = *init;
1779	return 0;
1780	}
1781
1782	void efx_siena_ptp_get_ts_info(struct efx_nic *efx,
1783	struct ethtool_ts_info *ts_info)
1784	{
1785	struct efx_ptp_data *ptp = efx->ptp_data;
1786	struct efx_nic *primary = efx->primary;
1787
1788	ASSERT_RTNL();
1789
1790	if (!ptp)
1791	return;
1792
1793	ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1794	SOF_TIMESTAMPING_RX_HARDWARE |
1795	SOF_TIMESTAMPING_RAW_HARDWARE);
1796	if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1797	ts_info->phc_index =
1798	ptp_clock_index(ptp: primary->ptp_data->phc_clock);
1799	ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1800	ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1801	}
1802
1803	int efx_siena_ptp_set_ts_config(struct efx_nic *efx,
1804	struct kernel_hwtstamp_config *config,
1805	struct netlink_ext_ack __always_unused *extack)
1806	{
1807	/* Not a PTP enabled port */
1808	if (!efx->ptp_data)
1809	return -EOPNOTSUPP;
1810
1811	return efx_ptp_ts_init(efx, init: config);
1812	}
1813
1814	int efx_siena_ptp_get_ts_config(struct efx_nic *efx,
1815	struct kernel_hwtstamp_config *config)
1816	{
1817	/* Not a PTP enabled port */
1818	if (!efx->ptp_data)
1819	return -EOPNOTSUPP;
1820
1821	*config = efx->ptp_data->config;
1822	return 0;
1823	}
1824
1825	static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1826	{
1827	struct efx_ptp_data *ptp = efx->ptp_data;
1828
1829	netif_err(efx, hw, efx->net_dev,
1830	"PTP unexpected event length: got %d expected %d\n",
1831	ptp->evt_frag_idx, expected_frag_len);
1832	ptp->reset_required = true;
1833	queue_work(wq: ptp->workwq, work: &ptp->work);
1834	}
1835
1836	/* Process a completed receive event. Put it on the event queue and
1837	* start worker thread. This is required because event and their
1838	* correspoding packets may come in either order.
1839	*/
1840	static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
1841	{
1842	struct efx_ptp_event_rx *evt = NULL;
1843
1844	if (WARN_ON_ONCE(ptp->rx_ts_inline))
1845	return;
1846
1847	if (ptp->evt_frag_idx != 3) {
1848	ptp_event_failure(efx, expected_frag_len: 3);
1849	return;
1850	}
1851
1852	spin_lock_bh(lock: &ptp->evt_lock);
1853	if (!list_empty(head: &ptp->evt_free_list)) {
1854	evt = list_first_entry(&ptp->evt_free_list,
1855	struct efx_ptp_event_rx, link);
1856	list_del(entry: &evt->link);
1857
1858	evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA);
1859	evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2],
1860	MCDI_EVENT_SRC) |
1861	(EFX_QWORD_FIELD(ptp->evt_frags[1],
1862	MCDI_EVENT_SRC) << 8) |
1863	(EFX_QWORD_FIELD(ptp->evt_frags[0],
1864	MCDI_EVENT_SRC) << 16));
1865	evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
1866	EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
1867	EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
1868	ptp->ts_corrections.ptp_rx);
1869	evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1870	list_add_tail(new: &evt->link, head: &ptp->evt_list);
1871
1872	queue_work(wq: ptp->workwq, work: &ptp->work);
1873	} else if (net_ratelimit()) {
1874	/* Log a rate-limited warning message. */
1875	netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n");
1876	}
1877	spin_unlock_bh(lock: &ptp->evt_lock);
1878	}
1879
1880	static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1881	{
1882	int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1883	if (ptp->evt_frag_idx != 1) {
1884	ptp_event_failure(efx, expected_frag_len: 1);
1885	return;
1886	}
1887
1888	netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1889	}
1890
1891	static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1892	{
1893	if (ptp->nic_ts_enabled)
1894	queue_work(wq: ptp->pps_workwq, work: &ptp->pps_work);
1895	}
1896
1897	void efx_siena_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1898	{
1899	struct efx_ptp_data *ptp = efx->ptp_data;
1900	int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
1901
1902	if (!ptp) {
1903	if (!efx->ptp_warned) {
1904	netif_warn(efx, drv, efx->net_dev,
1905	"Received PTP event but PTP not set up\n");
1906	efx->ptp_warned = true;
1907	}
1908	return;
1909	}
1910
1911	if (!ptp->enabled)
1912	return;
1913
1914	if (ptp->evt_frag_idx == 0) {
1915	ptp->evt_code = code;
1916	} else if (ptp->evt_code != code) {
1917	netif_err(efx, hw, efx->net_dev,
1918	"PTP out of sequence event %d\n", code);
1919	ptp->evt_frag_idx = 0;
1920	}
1921
1922	ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
1923	if (!MCDI_EVENT_FIELD(*ev, CONT)) {
1924	/* Process resulting event */
1925	switch (code) {
1926	case MCDI_EVENT_CODE_PTP_RX:
1927	ptp_event_rx(efx, ptp);
1928	break;
1929	case MCDI_EVENT_CODE_PTP_FAULT:
1930	ptp_event_fault(efx, ptp);
1931	break;
1932	case MCDI_EVENT_CODE_PTP_PPS:
1933	ptp_event_pps(efx, ptp);
1934	break;
1935	default:
1936	netif_err(efx, hw, efx->net_dev,
1937	"PTP unknown event %d\n", code);
1938	break;
1939	}
1940	ptp->evt_frag_idx = 0;
1941	} else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
1942	netif_err(efx, hw, efx->net_dev,
1943	"PTP too many event fragments\n");
1944	ptp->evt_frag_idx = 0;
1945	}
1946	}
1947
1948	void efx_siena_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
1949	{
1950	struct efx_nic *efx = channel->efx;
1951	struct efx_ptp_data *ptp = efx->ptp_data;
1952
1953	/* When extracting the sync timestamp minor value, we should discard
1954	* the least significant two bits. These are not required in order
1955	* to reconstruct full-range timestamps and they are optionally used
1956	* to report status depending on the options supplied when subscribing
1957	* for sync events.
1958	*/
1959	channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
1960	channel->sync_timestamp_minor =
1961	(MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_MS_8BITS) & 0xFC)
1962	<< ptp->nic_time.sync_event_minor_shift;
1963
1964	/* if sync events have been disabled then we want to silently ignore
1965	* this event, so throw away result.
1966	*/
1967	(void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
1968	SYNC_EVENTS_VALID);
1969	}
1970
1971	static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
1972	{
1973	#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
1974	return __le32_to_cpup(p: (const __le32 *)(eh + efx->rx_packet_ts_offset));
1975	#else
1976	const u8 *data = eh + efx->rx_packet_ts_offset;
1977	return (u32)data[0] |
1978	(u32)data[1] << 8 |
1979	(u32)data[2] << 16 |
1980	(u32)data[3] << 24;
1981	#endif
1982	}
1983
1984	void __efx_siena_rx_skb_attach_timestamp(struct efx_channel *channel,
1985	struct sk_buff *skb)
1986	{
1987	struct efx_nic *efx = channel->efx;
1988	struct efx_ptp_data *ptp = efx->ptp_data;
1989	u32 pkt_timestamp_major, pkt_timestamp_minor;
1990	u32 diff, carry;
1991	struct skb_shared_hwtstamps *timestamps;
1992
1993	if (channel->sync_events_state != SYNC_EVENTS_VALID)
1994	return;
1995
1996	pkt_timestamp_minor = efx_rx_buf_timestamp_minor(efx, eh: skb_mac_header(skb));
1997
1998	/* get the difference between the packet and sync timestamps,
1999	* modulo one second
2000	*/
2001	diff = pkt_timestamp_minor - channel->sync_timestamp_minor;
2002	if (pkt_timestamp_minor < channel->sync_timestamp_minor)
2003	diff += ptp->nic_time.minor_max;
2004
2005	/* do we roll over a second boundary and need to carry the one? */
2006	carry = (channel->sync_timestamp_minor >= ptp->nic_time.minor_max - diff) ?
2007	1 : 0;
2008
2009	if (diff <= ptp->nic_time.sync_event_diff_max) {
2010	/* packet is ahead of the sync event by a quarter of a second or
2011	* less (allowing for fuzz)
2012	*/
2013	pkt_timestamp_major = channel->sync_timestamp_major + carry;
2014	} else if (diff >= ptp->nic_time.sync_event_diff_min) {
2015	/* packet is behind the sync event but within the fuzz factor.
2016	* This means the RX packet and sync event crossed as they were
2017	* placed on the event queue, which can sometimes happen.
2018	*/
2019	pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
2020	} else {
2021	/* it's outside tolerance in both directions. this might be
2022	* indicative of us missing sync events for some reason, so
2023	* we'll call it an error rather than risk giving a bogus
2024	* timestamp.
2025	*/
2026	netif_vdbg(efx, drv, efx->net_dev,
2027	"packet timestamp %x too far from sync event %x:%x\n",
2028	pkt_timestamp_minor, channel->sync_timestamp_major,
2029	channel->sync_timestamp_minor);
2030	return;
2031	}
2032
2033	/* attach the timestamps to the skb */
2034	timestamps = skb_hwtstamps(skb);
2035	timestamps->hwtstamp =
2036	ptp->nic_to_kernel_time(pkt_timestamp_major,
2037	pkt_timestamp_minor,
2038	ptp->ts_corrections.general_rx);
2039	}
2040
2041	static int efx_phc_adjfine(struct ptp_clock_info *ptp, long scaled_ppm)
2042	{
2043	struct efx_ptp_data *ptp_data = container_of(ptp,
2044	struct efx_ptp_data,
2045	phc_clock_info);
2046	s32 delta = scaled_ppm_to_ppb(ppm: scaled_ppm);
2047	struct efx_nic *efx = ptp_data->efx;
2048	MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
2049	s64 adjustment_ns;
2050	int rc;
2051
2052	if (delta > MAX_PPB)
2053	delta = MAX_PPB;
2054	else if (delta < -MAX_PPB)
2055	delta = -MAX_PPB;
2056
2057	/* Convert ppb to fixed point ns taking care to round correctly. */
2058	adjustment_ns = ((s64)delta * PPB_SCALE_WORD +
2059	(1 << (ptp_data->adjfreq_ppb_shift - 1))) >>
2060	ptp_data->adjfreq_ppb_shift;
2061
2062	MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2063	MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
2064	MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
2065	MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
2066	MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
2067	rc = efx_siena_mcdi_rpc(efx, MC_CMD_PTP, inbuf: inadj, inlen: sizeof(inadj),
2068	NULL, outlen: 0, NULL);
2069	if (rc != 0)
2070	return rc;
2071
2072	ptp_data->current_adjfreq = adjustment_ns;
2073	return 0;
2074	}
2075
2076	static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
2077	{
2078	u32 nic_major, nic_minor;
2079	struct efx_ptp_data *ptp_data = container_of(ptp,
2080	struct efx_ptp_data,
2081	phc_clock_info);
2082	struct efx_nic *efx = ptp_data->efx;
2083	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
2084
2085	efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
2086
2087	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2088	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2089	MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
2090	MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
2091	MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
2092	return efx_siena_mcdi_rpc(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
2093	NULL, outlen: 0, NULL);
2094	}
2095
2096	static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
2097	{
2098	struct efx_ptp_data *ptp_data = container_of(ptp,
2099	struct efx_ptp_data,
2100	phc_clock_info);
2101	struct efx_nic *efx = ptp_data->efx;
2102	MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
2103	MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
2104	int rc;
2105	ktime_t kt;
2106
2107	MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
2108	MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2109
2110	rc = efx_siena_mcdi_rpc(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
2111	outbuf, outlen: sizeof(outbuf), NULL);
2112	if (rc != 0)
2113	return rc;
2114
2115	kt = ptp_data->nic_to_kernel_time(
2116	MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
2117	MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
2118	*ts = ktime_to_timespec64(kt);
2119	return 0;
2120	}
2121
2122	static int efx_phc_settime(struct ptp_clock_info *ptp,
2123	const struct timespec64 *e_ts)
2124	{
2125	/* Get the current NIC time, efx_phc_gettime.
2126	* Subtract from the desired time to get the offset
2127	* call efx_phc_adjtime with the offset
2128	*/
2129	int rc;
2130	struct timespec64 time_now;
2131	struct timespec64 delta;
2132
2133	rc = efx_phc_gettime(ptp, ts: &time_now);
2134	if (rc != 0)
2135	return rc;
2136
2137	delta = timespec64_sub(lhs: *e_ts, rhs: time_now);
2138
2139	rc = efx_phc_adjtime(ptp, delta: timespec64_to_ns(ts: &delta));
2140	if (rc != 0)
2141	return rc;
2142
2143	return 0;
2144	}
2145
2146	static int efx_phc_enable(struct ptp_clock_info *ptp,
2147	struct ptp_clock_request *request,
2148	int enable)
2149	{
2150	struct efx_ptp_data *ptp_data = container_of(ptp,
2151	struct efx_ptp_data,
2152	phc_clock_info);
2153	if (request->type != PTP_CLK_REQ_PPS)
2154	return -EOPNOTSUPP;
2155
2156	ptp_data->nic_ts_enabled = !!enable;
2157	return 0;
2158	}
2159
2160	static const struct efx_channel_type efx_ptp_channel_type = {
2161	.handle_no_channel = efx_ptp_handle_no_channel,
2162	.pre_probe = efx_ptp_probe_channel,
2163	.post_remove = efx_ptp_remove_channel,
2164	.get_name = efx_ptp_get_channel_name,
2165	/* no copy operation; there is no need to reallocate this channel */
2166	.receive_skb = efx_ptp_rx,
2167	.want_txqs = efx_ptp_want_txqs,
2168	.keep_eventq = false,
2169	};
2170
2171	void efx_siena_ptp_defer_probe_with_channel(struct efx_nic *efx)
2172	{
2173	/* Check whether PTP is implemented on this NIC. The DISABLE
2174	* operation will succeed if and only if it is implemented.
2175	*/
2176	if (efx_ptp_disable(efx) == 0)
2177	efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
2178	&efx_ptp_channel_type;
2179	}
2180
2181	void efx_siena_ptp_start_datapath(struct efx_nic *efx)
2182	{
2183	if (efx_ptp_restart(efx))
2184	netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
2185	/* re-enable timestamping if it was previously enabled */
2186	if (efx->type->ptp_set_ts_sync_events)
2187	efx->type->ptp_set_ts_sync_events(efx, true, true);
2188	}
2189
2190	void efx_siena_ptp_stop_datapath(struct efx_nic *efx)
2191	{
2192	/* temporarily disable timestamping */
2193	if (efx->type->ptp_set_ts_sync_events)
2194	efx->type->ptp_set_ts_sync_events(efx, false, true);
2195	efx_ptp_stop(efx);
2196	}
2197