1 /****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2011-2013 Solarflare Communications Inc.
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.
10 /* Theory of operation:
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.
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.
27 * Work queue processing:
28 * If work waiting, synchronise host/hardware time
30 * Transmit: send packet through MC, which returns the transmission time
31 * that is converted to an appropriate timestamp.
33 * Receive: the packet's reception time is converted to an appropriate
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"
47 #include "mcdi_pcol.h"
49 #include "farch_regs.h"
52 /* Maximum number of events expected to make up a PTP event */
53 #define MAX_EVENT_FRAGS 3
55 /* Maximum delay, ms, to begin synchronisation */
56 #define MAX_SYNCHRONISE_WAIT_MS 2
58 /* How long, at most, to spend synchronising */
59 #define SYNCHRONISE_PERIOD_NS 250000
61 /* How often to update the shared memory time */
62 #define SYNCHRONISATION_GRANULARITY_NS 200
64 /* Minimum permitted length of a (corrected) synchronisation time */
65 #define DEFAULT_MIN_SYNCHRONISATION_NS 120
67 /* Maximum permitted length of a (corrected) synchronisation time */
68 #define MAX_SYNCHRONISATION_NS 1000
70 /* How many (MC) receive events that can be queued */
71 #define MAX_RECEIVE_EVENTS 8
73 /* Length of (modified) moving average. */
74 #define AVERAGE_LENGTH 16
76 /* How long an unmatched event or packet can be held */
77 #define PKT_EVENT_LIFETIME_MS 10
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.
83 #define PTP_DPORT_OFFSET 22
85 #define PTP_V1_VERSION_LENGTH 2
86 #define PTP_V1_VERSION_OFFSET 28
88 #define PTP_V1_UUID_LENGTH 6
89 #define PTP_V1_UUID_OFFSET 50
91 #define PTP_V1_SEQUENCE_LENGTH 2
92 #define PTP_V1_SEQUENCE_OFFSET 58
94 /* The minimum length of a PTP V1 packet for offsets, etc. to be valid:
97 #define PTP_V1_MIN_LENGTH 64
99 #define PTP_V2_VERSION_LENGTH 1
100 #define PTP_V2_VERSION_OFFSET 29
102 #define PTP_V2_UUID_LENGTH 8
103 #define PTP_V2_UUID_OFFSET 48
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.
110 #define PTP_V2_MC_UUID_LENGTH 6
111 #define PTP_V2_MC_UUID_OFFSET 50
113 #define PTP_V2_SEQUENCE_LENGTH 2
114 #define PTP_V2_SEQUENCE_OFFSET 58
116 /* The minimum length of a PTP V2 packet for offsets, etc. to be valid:
117 * includes IP header.
119 #define PTP_V2_MIN_LENGTH 63
121 #define PTP_MIN_LENGTH 63
123 #define PTP_ADDRESS 0xe0000181 /* 224.0.1.129 */
124 #define PTP_EVENT_PORT 319
125 #define PTP_GENERAL_PORT 320
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.
130 #define PTP_VERSION_V1 1
132 #define PTP_VERSION_V2 2
133 #define PTP_VERSION_V2_MASK 0x0f
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
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.
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)
149 /* Maximum parts-per-billion adjustment that is acceptable */
150 #define MAX_PPB 1000000
152 /* Number of bits required to hold the above */
153 #define MAX_PPB_BITS 20
155 /* Number of extra bits allowed when calculating fractional ns.
156 * EXTRA_BITS + MC_CMD_PTP_IN_ADJUST_BITS + MAX_PPB_BITS should
159 #define PPB_EXTRA_BITS 2
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)
165 #define PTP_SYNC_ATTEMPTS 4
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
172 * @state: The state of the packet - whether it is ready for processing or
173 * whether that is of no interest.
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;
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
187 struct efx_ptp_event_rx {
188 struct list_head link;
192 unsigned long expiry;
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
206 struct efx_ptp_timeset {
212 u32 window; /* Derived: end - start, allowing for wrap */
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
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
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.
268 struct efx_ptp_data {
270 struct efx_channel *channel;
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;
277 struct efx_ptp_event_rx rx_evts[MAX_RECEIVE_EVENTS];
278 struct workqueue_struct *workwq;
279 struct work_struct work;
282 u32 rxfilter_general;
283 bool rxfilter_installed;
284 struct hwtstamp_config config;
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,
291 unsigned int min_synchronisation_ns;
298 efx_qword_t evt_frags[MAX_EVENT_FRAGS];
301 struct efx_buffer start;
302 struct pps_event_time host_time_pps;
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;
309 _MCDI_DECLARE_BUF(txbuf, MC_CMD_PTP_IN_TRANSMIT_LENMAX);
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];
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);
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),
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,
357 size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
362 return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
363 efx_ptp_stat_mask, strings);
366 size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats)
368 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
369 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
376 /* Copy software statistics */
377 for (i = 0; i < PTP_STAT_COUNT; i++) {
378 if (efx_ptp_stat_desc[i].dma_width)
380 stats[i] = *(unsigned int *)((char *)efx->ptp_data +
381 efx_ptp_stat_desc[i].offset);
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.
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);
393 memset(outbuf, 0, sizeof(outbuf));
394 efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
396 stats, _MCDI_PTR(outbuf, 0), false);
398 return PTP_STAT_COUNT;
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)
404 struct timespec64 ts = ns_to_timespec64(ns);
405 *nic_major = (u32)ts.tv_sec;
406 *nic_minor = ts.tv_nsec;
409 static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
412 ktime_t kt = ktime_set(nic_major, nic_minor);
414 kt = ktime_add_ns(kt, (u64)correction);
416 kt = ktime_sub_ns(kt, (u64)-correction);
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.
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)
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.
432 static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
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);
439 /* The conversion can result in the minor value exceeding the maximum.
440 * In this case, round up to the next second.
442 if (min >= S27_MINOR_MAX) {
443 min -= S27_MINOR_MAX;
451 static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
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);
458 static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
461 /* Apply the correction and deal with carry */
462 nic_minor += correction;
463 if ((s32)nic_minor < 0) {
464 nic_minor += S27_MINOR_MAX;
466 } else if (nic_minor >= S27_MINOR_MAX) {
467 nic_minor -= S27_MINOR_MAX;
471 return efx_ptp_s27_to_ktime(nic_major, nic_minor);
474 /* Get PTP attributes and set up time conversions */
475 static int efx_ptp_get_attributes(struct efx_nic *efx)
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;
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.
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);
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 pci_info(efx->pci_dev, "no PTP support\n");
500 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf),
501 outbuf, sizeof(outbuf), rc);
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;
515 ptp->time_format = fmt;
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.
523 if (rc == 0 && out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
524 ptp->min_synchronisation_ns =
526 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
528 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
533 /* Get PTP timestamp corrections */
534 static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
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);
540 /* Get the timestamp corrections from the NIC. If this operation is
541 * not supported (older NICs) then no correction is required.
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);
547 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
548 outbuf, sizeof(outbuf), NULL);
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;
564 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), outbuf,
572 /* Enable MCDI PTP support. */
573 static int efx_ptp_enable(struct efx_nic *efx)
575 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
576 MCDI_DECLARE_BUF_ERR(outbuf);
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);
586 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
587 outbuf, sizeof(outbuf), NULL);
588 rc = (rc == -EALREADY) ? 0 : rc;
590 efx_mcdi_display_error(efx, MC_CMD_PTP,
591 MC_CMD_PTP_IN_ENABLE_LEN,
592 outbuf, sizeof(outbuf), rc);
596 /* Disable MCDI PTP support.
598 * Note that this function should never rely on the presence of ptp_data -
599 * may be called before that exists.
601 static int efx_ptp_disable(struct efx_nic *efx)
603 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
604 MCDI_DECLARE_BUF_ERR(outbuf);
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.
615 if (rc == -ENOSYS || rc == -EPERM)
616 pci_info(efx->pci_dev, "no PTP support\n");
618 efx_mcdi_display_error(efx, MC_CMD_PTP,
619 MC_CMD_PTP_IN_DISABLE_LEN,
620 outbuf, sizeof(outbuf), rc);
624 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
628 while ((skb = skb_dequeue(q))) {
630 netif_receive_skb(skb);
635 static void efx_ptp_handle_no_channel(struct efx_nic *efx)
637 netif_err(efx, drv, efx->net_dev,
638 "ERROR: PTP requires MSI-X and 1 additional interrupt"
639 "vector. PTP disabled\n");
642 /* Repeatedly send the host time to the MC which will capture the hardware
645 static void efx_ptp_send_times(struct efx_nic *efx,
646 struct pps_event_time *last_time)
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;
657 timespec64_add_ns(&limit, SYNCHRONISE_PERIOD_NS);
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;
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);
670 } while ((timespec64_compare(&now.ts_real, &update_time) < 0) &&
671 ACCESS_ONCE(*mc_running));
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);
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)
686 unsigned start_ns, end_ns;
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);
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;
704 /* Process times received from MC.
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.
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)
717 unsigned number_readings =
718 MCDI_VAR_ARRAY_LEN(response_length,
719 PTP_OUT_SYNCHRONIZE_TIMESET);
722 unsigned last_good = 0;
723 struct efx_ptp_data *ptp = efx->ptp_data;
726 struct timespec64 delta;
729 if (number_readings == 0)
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
738 for (i = 0; i < number_readings; i++) {
739 s32 window, corrected;
740 struct timespec64 wait;
742 efx_ptp_read_timeset(
743 MCDI_ARRAY_STRUCT_PTR(synch_buf,
744 PTP_OUT_SYNCHRONIZE_TIMESET, i),
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;
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.
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;
774 netif_warn(efx, drv, efx->net_dev,
775 "PTP no suitable synchronisations\n");
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.
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");
793 delta.tv_sec = (last_sec - start_sec) & 1;
795 last_time->ts_real.tv_nsec -
796 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
798 /* Convert the NIC time at last good sync into kernel time.
799 * No correction is required - this time is the output of a
802 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
803 ptp->timeset[last_good].minor, 0);
805 /* Calculate delay from NIC top of second to last_time */
806 delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec;
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);
815 /* Synchronize times between the host and the MC */
816 static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
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;
822 unsigned long timeout;
823 struct pps_event_time last_time = {};
824 unsigned int loops = 0;
825 int *start = ptp->start.addr;
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,
831 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
832 ptp->start.dma_addr);
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);
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 */
849 if (!time_before(jiffies, timeout))
850 ++ptp->sync_timeouts;
852 if (ACCESS_ONCE(*start))
853 efx_ptp_send_times(efx, &last_time);
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),
861 rc = efx_ptp_process_times(efx, synch_buf, response_length,
866 ++ptp->no_time_syncs;
869 /* Increment the bad syncs counter if the synchronize fails, whatever
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)
881 struct efx_ptp_data *ptp_data = efx->ptp_data;
882 struct skb_shared_hwtstamps timestamps;
884 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
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);
896 if (skb->ip_summed == CHECKSUM_PARTIAL) {
897 rc = skb_checksum_help(skb);
901 skb_copy_from_linear_data(skb,
902 MCDI_PTR(ptp_data->txbuf,
903 PTP_IN_TRANSMIT_PACKET),
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);
911 memset(×tamps, 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);
917 skb_tstamp_tx(skb, ×tamps);
927 static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
929 struct efx_ptp_data *ptp = efx->ptp_data;
930 struct list_head *cursor;
931 struct list_head *next;
933 if (ptp->rx_ts_inline)
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;
942 evt = list_entry(cursor, struct efx_ptp_event_rx,
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");
951 spin_unlock_bh(&ptp->evt_lock);
954 static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
957 struct efx_ptp_data *ptp = efx->ptp_data;
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;
964 WARN_ON_ONCE(ptp->rx_ts_inline);
966 spin_lock_bh(&ptp->evt_lock);
967 evts_waiting = !list_empty(&ptp->evt_list);
968 spin_unlock_bh(&ptp->evt_lock);
971 return PTP_PACKET_STATE_UNMATCHED;
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;
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;
984 /* Match - add in hardware timestamp */
985 timestamps = skb_hwtstamps(skb);
986 timestamps->hwtstamp = evt->hwtimestamp;
988 match->state = PTP_PACKET_STATE_MATCHED;
989 rc = PTP_PACKET_STATE_MATCHED;
990 list_move(&evt->link, &ptp->evt_free_list);
994 spin_unlock_bh(&ptp->evt_lock);
999 /* Process any queued receive events and corresponding packets
1001 * q is returned with all the packets that are ready for delivery.
1003 static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
1005 struct efx_ptp_data *ptp = efx->ptp_data;
1006 struct sk_buff *skb;
1008 while ((skb = skb_dequeue(&ptp->rxq))) {
1009 struct efx_ptp_match *match;
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);
1022 /* Replace unprocessed entry and stop */
1023 skb_queue_head(&ptp->rxq, skb);
1029 /* Complete processing of a received packet */
1030 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1033 netif_receive_skb(skb);
1037 static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
1039 struct efx_ptp_data *ptp = efx->ptp_data;
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;
1050 static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1052 struct efx_ptp_data *ptp = efx->ptp_data;
1053 struct efx_filter_spec rxfilter;
1056 if (!ptp->channel || ptp->rxfilter_installed)
1059 /* Must filter on both event and general ports to ensure
1060 * that there is no packet re-ordering.
1062 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1064 efx_channel_get_rx_queue(ptp->channel)));
1065 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1067 htons(PTP_EVENT_PORT));
1071 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1074 ptp->rxfilter_event = rc;
1076 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1078 efx_channel_get_rx_queue(ptp->channel)));
1079 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1081 htons(PTP_GENERAL_PORT));
1085 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1088 ptp->rxfilter_general = rc;
1090 ptp->rxfilter_installed = true;
1094 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1095 ptp->rxfilter_event);
1099 static int efx_ptp_start(struct efx_nic *efx)
1101 struct efx_ptp_data *ptp = efx->ptp_data;
1104 ptp->reset_required = false;
1106 rc = efx_ptp_insert_multicast_filters(efx);
1110 rc = efx_ptp_enable(efx);
1114 ptp->evt_frag_idx = 0;
1115 ptp->current_adjfreq = 0;
1120 efx_ptp_remove_multicast_filters(efx);
1124 static int efx_ptp_stop(struct efx_nic *efx)
1126 struct efx_ptp_data *ptp = efx->ptp_data;
1127 struct list_head *cursor;
1128 struct list_head *next;
1134 rc = efx_ptp_disable(efx);
1136 efx_ptp_remove_multicast_filters(efx);
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);
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);
1147 spin_unlock_bh(&efx->ptp_data->evt_lock);
1152 static int efx_ptp_restart(struct efx_nic *efx)
1154 if (efx->ptp_data && efx->ptp_data->enabled)
1155 return efx_ptp_start(efx);
1159 static void efx_ptp_pps_worker(struct work_struct *work)
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;
1166 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
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);
1174 static void efx_ptp_worker(struct work_struct *work)
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;
1182 if (ptp_data->reset_required) {
1188 efx_ptp_drop_time_expired_events(efx);
1190 __skb_queue_head_init(&tempq);
1191 efx_ptp_process_events(efx, &tempq);
1193 while ((skb = skb_dequeue(&ptp_data->txq)))
1194 efx_ptp_xmit_skb(efx, skb);
1196 while ((skb = __skb_dequeue(&tempq)))
1197 efx_ptp_process_rx(efx, skb);
1200 static const struct ptp_clock_info efx_phc_clock_info = {
1201 .owner = THIS_MODULE,
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,
1216 /* Initialise PTP state. */
1217 int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1219 struct efx_ptp_data *ptp;
1223 ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL);
1224 efx->ptp_data = ptp;
1229 ptp->channel = channel;
1230 ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1232 rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL);
1236 skb_queue_head_init(&ptp->rxq);
1237 skb_queue_head_init(&ptp->txq);
1238 ptp->workwq = create_singlethread_workqueue("sfc_ptp");
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);
1254 /* Get the NIC PTP attributes and set up time conversions */
1255 rc = efx_ptp_get_attributes(efx);
1259 /* Get the timestamp corrections */
1260 rc = efx_ptp_get_timestamp_corrections(efx);
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);
1272 } else if (ptp->phc_clock) {
1273 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1274 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
1275 if (!ptp->pps_workwq) {
1281 ptp->nic_ts_enabled = false;
1285 ptp_clock_unregister(efx->ptp_data->phc_clock);
1288 destroy_workqueue(efx->ptp_data->workwq);
1291 efx_nic_free_buffer(efx, &ptp->start);
1294 kfree(efx->ptp_data);
1295 efx->ptp_data = NULL;
1300 /* Initialise PTP channel.
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.
1305 static int efx_ptp_probe_channel(struct efx_channel *channel)
1307 struct efx_nic *efx = channel->efx;
1309 channel->irq_moderation_us = 0;
1310 channel->rx_queue.core_index = 0;
1312 return efx_ptp_probe(efx, channel);
1315 void efx_ptp_remove(struct efx_nic *efx)
1320 (void)efx_ptp_disable(efx);
1322 cancel_work_sync(&efx->ptp_data->work);
1323 if (efx->ptp_data->pps_workwq)
1324 cancel_work_sync(&efx->ptp_data->pps_work);
1326 skb_queue_purge(&efx->ptp_data->rxq);
1327 skb_queue_purge(&efx->ptp_data->txq);
1329 if (efx->ptp_data->phc_clock) {
1330 destroy_workqueue(efx->ptp_data->pps_workwq);
1331 ptp_clock_unregister(efx->ptp_data->phc_clock);
1334 destroy_workqueue(efx->ptp_data->workwq);
1336 efx_nic_free_buffer(efx, &efx->ptp_data->start);
1337 kfree(efx->ptp_data);
1340 static void efx_ptp_remove_channel(struct efx_channel *channel)
1342 efx_ptp_remove(channel->efx);
1345 static void efx_ptp_get_channel_name(struct efx_channel *channel,
1346 char *buf, size_t len)
1348 snprintf(buf, len, "%s-ptp", channel->efx->name);
1351 /* Determine whether this packet should be processed by the PTP module
1352 * or transmitted conventionally.
1354 bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1356 return efx->ptp_data &&
1357 efx->ptp_data->enabled &&
1358 skb->len >= PTP_MIN_LENGTH &&
1359 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1360 likely(skb->protocol == htons(ETH_P_IP)) &&
1361 skb_transport_header_was_set(skb) &&
1362 skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1363 ip_hdr(skb)->protocol == IPPROTO_UDP &&
1365 skb_transport_offset(skb) + sizeof(struct udphdr) &&
1366 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1369 /* Receive a PTP packet. Packets are queued until the arrival of
1370 * the receive timestamp from the MC - this will probably occur after the
1371 * packet arrival because of the processing in the MC.
1373 static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1375 struct efx_nic *efx = channel->efx;
1376 struct efx_ptp_data *ptp = efx->ptp_data;
1377 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1378 u8 *match_data_012, *match_data_345;
1379 unsigned int version;
1382 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1384 /* Correct version? */
1385 if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1386 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1390 version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1391 if (version != PTP_VERSION_V1) {
1395 /* PTP V1 uses all six bytes of the UUID to match the packet
1398 match_data_012 = data + PTP_V1_UUID_OFFSET;
1399 match_data_345 = data + PTP_V1_UUID_OFFSET + 3;
1401 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1405 version = data[PTP_V2_VERSION_OFFSET];
1406 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1410 /* The original V2 implementation uses bytes 2-7 of
1411 * the UUID to match the packet to the timestamp. This
1412 * discards two of the bytes of the MAC address used
1413 * to create the UUID (SF bug 33070). The PTP V2
1414 * enhanced mode fixes this issue and uses bytes 0-2
1415 * and byte 5-7 of the UUID.
1417 match_data_345 = data + PTP_V2_UUID_OFFSET + 5;
1418 if (ptp->mode == MC_CMD_PTP_MODE_V2) {
1419 match_data_012 = data + PTP_V2_UUID_OFFSET + 2;
1421 match_data_012 = data + PTP_V2_UUID_OFFSET + 0;
1422 BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED);
1426 /* Does this packet require timestamping? */
1427 if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1428 match->state = PTP_PACKET_STATE_UNMATCHED;
1430 /* We expect the sequence number to be in the same position in
1431 * the packet for PTP V1 and V2
1433 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1434 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1436 /* Extract UUID/Sequence information */
1437 match->words[0] = (match_data_012[0] |
1438 (match_data_012[1] << 8) |
1439 (match_data_012[2] << 16) |
1440 (match_data_345[0] << 24));
1441 match->words[1] = (match_data_345[1] |
1442 (match_data_345[2] << 8) |
1443 (data[PTP_V1_SEQUENCE_OFFSET +
1444 PTP_V1_SEQUENCE_LENGTH - 1] <<
1447 match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1450 skb_queue_tail(&ptp->rxq, skb);
1451 queue_work(ptp->workwq, &ptp->work);
1456 /* Transmit a PTP packet. This has to be transmitted by the MC
1457 * itself, through an MCDI call. MCDI calls aren't permitted
1458 * in the transmit path so defer the actual transmission to a suitable worker.
1460 int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1462 struct efx_ptp_data *ptp = efx->ptp_data;
1464 skb_queue_tail(&ptp->txq, skb);
1466 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1467 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1468 efx_xmit_hwtstamp_pending(skb);
1469 queue_work(ptp->workwq, &ptp->work);
1471 return NETDEV_TX_OK;
1474 int efx_ptp_get_mode(struct efx_nic *efx)
1476 return efx->ptp_data->mode;
1479 int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1480 unsigned int new_mode)
1482 if ((enable_wanted != efx->ptp_data->enabled) ||
1483 (enable_wanted && (efx->ptp_data->mode != new_mode))) {
1486 if (enable_wanted) {
1487 /* Change of mode requires disable */
1488 if (efx->ptp_data->enabled &&
1489 (efx->ptp_data->mode != new_mode)) {
1490 efx->ptp_data->enabled = false;
1491 rc = efx_ptp_stop(efx);
1496 /* Set new operating mode and establish
1497 * baseline synchronisation, which must
1500 efx->ptp_data->mode = new_mode;
1501 if (netif_running(efx->net_dev))
1502 rc = efx_ptp_start(efx);
1504 rc = efx_ptp_synchronize(efx,
1505 PTP_SYNC_ATTEMPTS * 2);
1510 rc = efx_ptp_stop(efx);
1516 efx->ptp_data->enabled = enable_wanted;
1522 static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
1529 if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1530 (init->tx_type != HWTSTAMP_TX_ON))
1533 rc = efx->type->ptp_set_ts_config(efx, init);
1537 efx->ptp_data->config = *init;
1541 void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
1543 struct efx_ptp_data *ptp = efx->ptp_data;
1544 struct efx_nic *primary = efx->primary;
1551 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1552 SOF_TIMESTAMPING_RX_HARDWARE |
1553 SOF_TIMESTAMPING_RAW_HARDWARE);
1554 if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1555 ts_info->phc_index =
1556 ptp_clock_index(primary->ptp_data->phc_clock);
1557 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1558 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1561 int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1563 struct hwtstamp_config config;
1566 /* Not a PTP enabled port */
1570 if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
1573 rc = efx_ptp_ts_init(efx, &config);
1577 return copy_to_user(ifr->ifr_data, &config, sizeof(config))
1581 int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1586 return copy_to_user(ifr->ifr_data, &efx->ptp_data->config,
1587 sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
1590 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1592 struct efx_ptp_data *ptp = efx->ptp_data;
1594 netif_err(efx, hw, efx->net_dev,
1595 "PTP unexpected event length: got %d expected %d\n",
1596 ptp->evt_frag_idx, expected_frag_len);
1597 ptp->reset_required = true;
1598 queue_work(ptp->workwq, &ptp->work);
1601 /* Process a completed receive event. Put it on the event queue and
1602 * start worker thread. This is required because event and their
1603 * correspoding packets may come in either order.
1605 static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
1607 struct efx_ptp_event_rx *evt = NULL;
1609 if (WARN_ON_ONCE(ptp->rx_ts_inline))
1612 if (ptp->evt_frag_idx != 3) {
1613 ptp_event_failure(efx, 3);
1617 spin_lock_bh(&ptp->evt_lock);
1618 if (!list_empty(&ptp->evt_free_list)) {
1619 evt = list_first_entry(&ptp->evt_free_list,
1620 struct efx_ptp_event_rx, link);
1621 list_del(&evt->link);
1623 evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA);
1624 evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2],
1626 (EFX_QWORD_FIELD(ptp->evt_frags[1],
1627 MCDI_EVENT_SRC) << 8) |
1628 (EFX_QWORD_FIELD(ptp->evt_frags[0],
1629 MCDI_EVENT_SRC) << 16));
1630 evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
1631 EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
1632 EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
1633 ptp->ts_corrections.rx);
1634 evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1635 list_add_tail(&evt->link, &ptp->evt_list);
1637 queue_work(ptp->workwq, &ptp->work);
1638 } else if (net_ratelimit()) {
1639 /* Log a rate-limited warning message. */
1640 netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n");
1642 spin_unlock_bh(&ptp->evt_lock);
1645 static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1647 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1648 if (ptp->evt_frag_idx != 1) {
1649 ptp_event_failure(efx, 1);
1653 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1656 static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1658 if (ptp->nic_ts_enabled)
1659 queue_work(ptp->pps_workwq, &ptp->pps_work);
1662 void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1664 struct efx_ptp_data *ptp = efx->ptp_data;
1665 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
1668 if (net_ratelimit())
1669 netif_warn(efx, drv, efx->net_dev,
1670 "Received PTP event but PTP not set up\n");
1677 if (ptp->evt_frag_idx == 0) {
1678 ptp->evt_code = code;
1679 } else if (ptp->evt_code != code) {
1680 netif_err(efx, hw, efx->net_dev,
1681 "PTP out of sequence event %d\n", code);
1682 ptp->evt_frag_idx = 0;
1685 ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
1686 if (!MCDI_EVENT_FIELD(*ev, CONT)) {
1687 /* Process resulting event */
1689 case MCDI_EVENT_CODE_PTP_RX:
1690 ptp_event_rx(efx, ptp);
1692 case MCDI_EVENT_CODE_PTP_FAULT:
1693 ptp_event_fault(efx, ptp);
1695 case MCDI_EVENT_CODE_PTP_PPS:
1696 ptp_event_pps(efx, ptp);
1699 netif_err(efx, hw, efx->net_dev,
1700 "PTP unknown event %d\n", code);
1703 ptp->evt_frag_idx = 0;
1704 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
1705 netif_err(efx, hw, efx->net_dev,
1706 "PTP too many event fragments\n");
1707 ptp->evt_frag_idx = 0;
1711 void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
1713 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
1714 channel->sync_timestamp_minor =
1715 MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_26_19) << 19;
1716 /* if sync events have been disabled then we want to silently ignore
1717 * this event, so throw away result.
1719 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
1723 /* make some assumptions about the time representation rather than abstract it,
1724 * since we currently only support one type of inline timestamping and only on
1727 #define MINOR_TICKS_PER_SECOND 0x8000000
1728 /* Fuzz factor for sync events to be out of order with RX events */
1729 #define FUZZ (MINOR_TICKS_PER_SECOND / 10)
1730 #define EXPECTED_SYNC_EVENTS_PER_SECOND 4
1732 static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
1734 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
1735 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset));
1737 const u8 *data = eh + efx->rx_packet_ts_offset;
1738 return (u32)data[0] |
1740 (u32)data[2] << 16 |
1745 void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
1746 struct sk_buff *skb)
1748 struct efx_nic *efx = channel->efx;
1749 u32 pkt_timestamp_major, pkt_timestamp_minor;
1751 struct skb_shared_hwtstamps *timestamps;
1753 pkt_timestamp_minor = (efx_rx_buf_timestamp_minor(efx,
1754 skb_mac_header(skb)) +
1755 (u32) efx->ptp_data->ts_corrections.rx) &
1756 (MINOR_TICKS_PER_SECOND - 1);
1758 /* get the difference between the packet and sync timestamps,
1761 diff = (pkt_timestamp_minor - channel->sync_timestamp_minor) &
1762 (MINOR_TICKS_PER_SECOND - 1);
1763 /* do we roll over a second boundary and need to carry the one? */
1764 carry = channel->sync_timestamp_minor + diff > MINOR_TICKS_PER_SECOND ?
1767 if (diff <= MINOR_TICKS_PER_SECOND / EXPECTED_SYNC_EVENTS_PER_SECOND +
1769 /* packet is ahead of the sync event by a quarter of a second or
1770 * less (allowing for fuzz)
1772 pkt_timestamp_major = channel->sync_timestamp_major + carry;
1773 } else if (diff >= MINOR_TICKS_PER_SECOND - FUZZ) {
1774 /* packet is behind the sync event but within the fuzz factor.
1775 * This means the RX packet and sync event crossed as they were
1776 * placed on the event queue, which can sometimes happen.
1778 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
1780 /* it's outside tolerance in both directions. this might be
1781 * indicative of us missing sync events for some reason, so
1782 * we'll call it an error rather than risk giving a bogus
1785 netif_vdbg(efx, drv, efx->net_dev,
1786 "packet timestamp %x too far from sync event %x:%x\n",
1787 pkt_timestamp_minor, channel->sync_timestamp_major,
1788 channel->sync_timestamp_minor);
1792 /* attach the timestamps to the skb */
1793 timestamps = skb_hwtstamps(skb);
1794 timestamps->hwtstamp =
1795 efx_ptp_s27_to_ktime(pkt_timestamp_major, pkt_timestamp_minor);
1798 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
1800 struct efx_ptp_data *ptp_data = container_of(ptp,
1801 struct efx_ptp_data,
1803 struct efx_nic *efx = ptp_data->efx;
1804 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
1808 if (delta > MAX_PPB)
1810 else if (delta < -MAX_PPB)
1813 /* Convert ppb to fixed point ns. */
1814 adjustment_ns = (((s64)delta * PPB_SCALE_WORD) >>
1815 (PPB_EXTRA_BITS + MAX_PPB_BITS));
1817 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
1818 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
1819 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
1820 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
1821 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
1822 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj),
1827 ptp_data->current_adjfreq = adjustment_ns;
1831 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
1833 u32 nic_major, nic_minor;
1834 struct efx_ptp_data *ptp_data = container_of(ptp,
1835 struct efx_ptp_data,
1837 struct efx_nic *efx = ptp_data->efx;
1838 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
1840 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
1842 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
1843 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
1844 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
1845 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
1846 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
1847 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
1851 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
1853 struct efx_ptp_data *ptp_data = container_of(ptp,
1854 struct efx_ptp_data,
1856 struct efx_nic *efx = ptp_data->efx;
1857 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
1858 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
1862 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
1863 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
1865 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
1866 outbuf, sizeof(outbuf), NULL);
1870 kt = ptp_data->nic_to_kernel_time(
1871 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
1872 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
1873 *ts = ktime_to_timespec64(kt);
1877 static int efx_phc_settime(struct ptp_clock_info *ptp,
1878 const struct timespec64 *e_ts)
1880 /* Get the current NIC time, efx_phc_gettime.
1881 * Subtract from the desired time to get the offset
1882 * call efx_phc_adjtime with the offset
1885 struct timespec64 time_now;
1886 struct timespec64 delta;
1888 rc = efx_phc_gettime(ptp, &time_now);
1892 delta = timespec64_sub(*e_ts, time_now);
1894 rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta));
1901 static int efx_phc_enable(struct ptp_clock_info *ptp,
1902 struct ptp_clock_request *request,
1905 struct efx_ptp_data *ptp_data = container_of(ptp,
1906 struct efx_ptp_data,
1908 if (request->type != PTP_CLK_REQ_PPS)
1911 ptp_data->nic_ts_enabled = !!enable;
1915 static const struct efx_channel_type efx_ptp_channel_type = {
1916 .handle_no_channel = efx_ptp_handle_no_channel,
1917 .pre_probe = efx_ptp_probe_channel,
1918 .post_remove = efx_ptp_remove_channel,
1919 .get_name = efx_ptp_get_channel_name,
1920 /* no copy operation; there is no need to reallocate this channel */
1921 .receive_skb = efx_ptp_rx,
1922 .keep_eventq = false,
1925 void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
1927 /* Check whether PTP is implemented on this NIC. The DISABLE
1928 * operation will succeed if and only if it is implemented.
1930 if (efx_ptp_disable(efx) == 0)
1931 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
1932 &efx_ptp_channel_type;
1935 void efx_ptp_start_datapath(struct efx_nic *efx)
1937 if (efx_ptp_restart(efx))
1938 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
1939 /* re-enable timestamping if it was previously enabled */
1940 if (efx->type->ptp_set_ts_sync_events)
1941 efx->type->ptp_set_ts_sync_events(efx, true, true);
1944 void efx_ptp_stop_datapath(struct efx_nic *efx)
1946 /* temporarily disable timestamping */
1947 if (efx->type->ptp_set_ts_sync_events)
1948 efx->type->ptp_set_ts_sync_events(efx, false, true);