1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2011-2013 Solarflare Communications Inc.
7 /* Theory of operation:
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.
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.
24 * Work queue processing:
25 * If work waiting, synchronise host/hardware time
27 * Transmit: send packet through MC, which returns the transmission time
28 * that is converted to an appropriate timestamp.
30 * Receive: the packet's reception time is converted to an appropriate
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"
43 #include "mcdi_pcol.h"
45 #include "farch_regs.h"
47 #include "nic.h" /* indirectly includes ptp.h */
49 /* Maximum number of events expected to make up a PTP event */
50 #define MAX_EVENT_FRAGS 3
52 /* Maximum delay, ms, to begin synchronisation */
53 #define MAX_SYNCHRONISE_WAIT_MS 2
55 /* How long, at most, to spend synchronising */
56 #define SYNCHRONISE_PERIOD_NS 250000
58 /* How often to update the shared memory time */
59 #define SYNCHRONISATION_GRANULARITY_NS 200
61 /* Minimum permitted length of a (corrected) synchronisation time */
62 #define DEFAULT_MIN_SYNCHRONISATION_NS 120
64 /* Maximum permitted length of a (corrected) synchronisation time */
65 #define MAX_SYNCHRONISATION_NS 1000
67 /* How many (MC) receive events that can be queued */
68 #define MAX_RECEIVE_EVENTS 8
70 /* Length of (modified) moving average. */
71 #define AVERAGE_LENGTH 16
73 /* How long an unmatched event or packet can be held */
74 #define PKT_EVENT_LIFETIME_MS 10
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.
80 #define PTP_DPORT_OFFSET 22
82 #define PTP_V1_VERSION_LENGTH 2
83 #define PTP_V1_VERSION_OFFSET 28
85 #define PTP_V1_UUID_LENGTH 6
86 #define PTP_V1_UUID_OFFSET 50
88 #define PTP_V1_SEQUENCE_LENGTH 2
89 #define PTP_V1_SEQUENCE_OFFSET 58
91 /* The minimum length of a PTP V1 packet for offsets, etc. to be valid:
94 #define PTP_V1_MIN_LENGTH 64
96 #define PTP_V2_VERSION_LENGTH 1
97 #define PTP_V2_VERSION_OFFSET 29
99 #define PTP_V2_UUID_LENGTH 8
100 #define PTP_V2_UUID_OFFSET 48
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.
107 #define PTP_V2_MC_UUID_LENGTH 6
108 #define PTP_V2_MC_UUID_OFFSET 50
110 #define PTP_V2_SEQUENCE_LENGTH 2
111 #define PTP_V2_SEQUENCE_OFFSET 58
113 /* The minimum length of a PTP V2 packet for offsets, etc. to be valid:
114 * includes IP header.
116 #define PTP_V2_MIN_LENGTH 63
118 #define PTP_MIN_LENGTH 63
120 #define PTP_ADDRESS 0xe0000181 /* 224.0.1.129 */
121 #define PTP_EVENT_PORT 319
122 #define PTP_GENERAL_PORT 320
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.
127 #define PTP_VERSION_V1 1
129 #define PTP_VERSION_V2 2
130 #define PTP_VERSION_V2_MASK 0x0f
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
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.
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)
146 /* Maximum parts-per-billion adjustment that is acceptable */
147 #define MAX_PPB 1000000
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)
153 /* How much to shift down after scaling to convert to FP40 */
154 #define PPB_SHIFT_FP40 26
156 #define PPB_SHIFT_FP44 22
158 #define PTP_SYNC_ATTEMPTS 4
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
165 * @state: The state of the packet - whether it is ready for processing or
166 * whether that is of no interest.
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;
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
182 struct efx_ptp_event_rx {
183 struct list_head link;
187 unsigned long expiry;
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
201 struct efx_ptp_timeset {
207 u32 window; /* Derived: end - start, allowing for wrap */
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
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
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
249 * @ts_corrections.ptp_rx: Required driver correction of PTP packet receive
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
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.
284 struct efx_ptp_data {
286 struct efx_channel *channel;
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;
293 struct efx_ptp_event_rx rx_evts[MAX_RECEIVE_EVENTS];
294 struct workqueue_struct *workwq;
295 struct work_struct work;
298 u32 rxfilter_general;
299 bool rxfilter_installed;
300 struct hwtstamp_config config;
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,
308 u32 sync_event_diff_min;
309 u32 sync_event_diff_max;
310 unsigned int sync_event_minor_shift;
312 unsigned int min_synchronisation_ns;
313 unsigned int capabilities;
322 efx_qword_t evt_frags[MAX_EVENT_FRAGS];
325 struct efx_buffer start;
326 struct pps_event_time host_time_pps;
327 unsigned int adjfreq_ppb_shift;
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;
334 efx_dword_t txbuf[MCDI_TX_BUF_LEN(MC_CMD_PTP_IN_TRANSMIT_LENMAX)];
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);
350 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta);
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);
358 bool efx_ptp_use_mac_tx_timestamps(struct efx_nic *efx)
360 return efx_has_cap(efx, TX_MAC_TIMESTAMPING);
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.
366 static bool efx_ptp_want_txqs(struct efx_channel *channel)
368 return efx_ptp_use_mac_tx_timestamps(channel->efx);
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),
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,
396 size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
401 return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
402 efx_ptp_stat_mask, strings);
405 size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats)
407 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
408 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
415 /* Copy software statistics */
416 for (i = 0; i < PTP_STAT_COUNT; i++) {
417 if (efx_ptp_stat_desc[i].dma_width)
419 stats[i] = *(unsigned int *)((char *)efx->ptp_data +
420 efx_ptp_stat_desc[i].offset);
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.
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_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
430 outbuf, sizeof(outbuf), NULL);
432 memset(outbuf, 0, sizeof(outbuf));
433 efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
435 stats, _MCDI_PTR(outbuf, 0), false);
437 return PTP_STAT_COUNT;
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)
443 struct timespec64 ts = ns_to_timespec64(ns);
444 *nic_major = (u32)ts.tv_sec;
445 *nic_minor = ts.tv_nsec;
448 static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
451 ktime_t kt = ktime_set(nic_major, nic_minor);
453 kt = ktime_add_ns(kt, (u64)correction);
455 kt = ktime_sub_ns(kt, (u64)-correction);
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.
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)
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.
471 static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
473 struct timespec64 ts = ns_to_timespec64(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);
478 /* The conversion can result in the minor value exceeding the maximum.
479 * In this case, round up to the next second.
481 if (min >= S27_MINOR_MAX) {
482 min -= S27_MINOR_MAX;
490 static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
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(nic_major, ns);
497 static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
500 /* Apply the correction and deal with carry */
501 nic_minor += correction;
502 if ((s32)nic_minor < 0) {
503 nic_minor += S27_MINOR_MAX;
505 } else if (nic_minor >= S27_MINOR_MAX) {
506 nic_minor -= S27_MINOR_MAX;
510 return efx_ptp_s27_to_ktime(nic_major, nic_minor);
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)
516 struct timespec64 ts = ns_to_timespec64(ns);
518 *nic_major = (u32)ts.tv_sec;
519 *nic_minor = ts.tv_nsec * 4;
522 static ktime_t efx_ptp_s_qns_to_ktime_correction(u32 nic_major, u32 nic_minor,
527 nic_minor = DIV_ROUND_CLOSEST(nic_minor, 4);
528 correction = DIV_ROUND_CLOSEST(correction, 4);
530 kt = ktime_set(nic_major, nic_minor);
533 kt = ktime_add_ns(kt, (u64)correction);
535 kt = ktime_sub_ns(kt, (u64)-correction);
539 struct efx_channel *efx_ptp_channel(struct efx_nic *efx)
541 return efx->ptp_data ? efx->ptp_data->channel : NULL;
544 static u32 last_sync_timestamp_major(struct efx_nic *efx)
546 struct efx_channel *channel = efx_ptp_channel(efx);
550 major = channel->sync_timestamp_major;
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.
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,
567 if (!(nic_major & 0x80000000)) {
568 WARN_ON_ONCE(nic_major >> 16);
570 /* Medford provides 48 bits of timestamp, so we must get the top
571 * 16 bits from the timesync event state.
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
582 * a - b = ((a mod k) - b) mod k
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.
588 sync_timestamp = last_sync_timestamp_major(efx);
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.
596 delta = nic_major - sync_timestamp;
598 /* Recover the fully specified time now, by applying the offset
599 * to the (fully specified) sync time.
601 nic_major = sync_timestamp + delta;
603 kt = ptp->nic_to_kernel_time(nic_major, nic_minor,
609 ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue)
611 struct efx_nic *efx = tx_queue->efx;
612 struct efx_ptp_data *ptp = efx->ptp_data;
615 if (efx_ptp_use_mac_tx_timestamps(efx))
616 kt = efx_ptp_mac_nic_to_ktime_correction(efx, ptp,
617 tx_queue->completed_timestamp_major,
618 tx_queue->completed_timestamp_minor,
619 ptp->ts_corrections.general_tx);
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);
628 /* Get PTP attributes and set up time conversions */
629 static int efx_ptp_get_attributes(struct efx_nic *efx)
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;
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.
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_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
645 outbuf, sizeof(outbuf), &out_len);
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");
654 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf),
655 outbuf, sizeof(outbuf), rc);
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;
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;
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;
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)
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);
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.
701 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_CAPABILITIES_OFST)
702 ptp->min_synchronisation_ns =
704 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
706 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
709 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
710 ptp->capabilities = MCDI_DWORD(outbuf,
711 PTP_OUT_GET_ATTRIBUTES_CAPABILITIES);
713 ptp->capabilities = 0;
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.
719 if (ptp->capabilities & (1 << MC_CMD_PTP_OUT_GET_ATTRIBUTES_FP44_FREQ_ADJ_LBN))
720 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP44;
722 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP40;
727 /* Get PTP timestamp corrections */
728 static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
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);
735 /* Get the timestamp corrections from the NIC. If this operation is
736 * not supported (older NICs) then no correction is required.
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);
742 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
743 outbuf, sizeof(outbuf), &out_len);
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);
754 if (out_len >= MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN) {
755 efx->ptp_data->ts_corrections.general_tx = MCDI_DWORD(
757 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_TX);
758 efx->ptp_data->ts_corrections.general_rx = MCDI_DWORD(
760 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_RX);
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;
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;
775 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), outbuf,
783 /* Enable MCDI PTP support. */
784 static int efx_ptp_enable(struct efx_nic *efx)
786 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
787 MCDI_DECLARE_BUF_ERR(outbuf);
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);
797 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
798 outbuf, sizeof(outbuf), NULL);
799 rc = (rc == -EALREADY) ? 0 : rc;
801 efx_mcdi_display_error(efx, MC_CMD_PTP,
802 MC_CMD_PTP_IN_ENABLE_LEN,
803 outbuf, sizeof(outbuf), rc);
807 /* Disable MCDI PTP support.
809 * Note that this function should never rely on the presence of ptp_data -
810 * may be called before that exists.
812 static int efx_ptp_disable(struct efx_nic *efx)
814 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
815 MCDI_DECLARE_BUF_ERR(outbuf);
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_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
821 outbuf, 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.
826 if (rc == -ENOSYS || rc == -EPERM)
827 pci_info(efx->pci_dev, "no PTP support\n");
829 efx_mcdi_display_error(efx, MC_CMD_PTP,
830 MC_CMD_PTP_IN_DISABLE_LEN,
831 outbuf, sizeof(outbuf), rc);
835 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
839 while ((skb = skb_dequeue(q))) {
841 netif_receive_skb(skb);
846 static void efx_ptp_handle_no_channel(struct efx_nic *efx)
848 netif_err(efx, drv, efx->net_dev,
849 "ERROR: PTP requires MSI-X and 1 additional interrupt"
850 "vector. PTP disabled\n");
853 /* Repeatedly send the host time to the MC which will capture the hardware
856 static void efx_ptp_send_times(struct efx_nic *efx,
857 struct pps_event_time *last_time)
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;
866 timespec64_add_ns(&limit, SYNCHRONISE_PERIOD_NS);
868 /* Write host time for specified period or until MC is done */
869 while ((timespec64_compare(&now.ts_real, &limit) < 0) &&
870 READ_ONCE(*mc_running)) {
871 struct timespec64 update_time;
872 unsigned int host_time;
874 /* Don't update continuously to avoid saturating the PCIe bus */
875 update_time = now.ts_real;
876 timespec64_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS);
879 } while ((timespec64_compare(&now.ts_real, &update_time) < 0) &&
880 READ_ONCE(*mc_running));
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);
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)
895 unsigned start_ns, end_ns;
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);
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;
913 /* Process times received from MC.
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.
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)
926 unsigned number_readings =
927 MCDI_VAR_ARRAY_LEN(response_length,
928 PTP_OUT_SYNCHRONIZE_TIMESET);
931 unsigned last_good = 0;
932 struct efx_ptp_data *ptp = efx->ptp_data;
935 struct timespec64 delta;
938 if (number_readings == 0)
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
947 for (i = 0; i < number_readings; i++) {
948 s32 window, corrected;
949 struct timespec64 wait;
951 efx_ptp_read_timeset(
952 MCDI_ARRAY_STRUCT_PTR(synch_buf,
953 PTP_OUT_SYNCHRONIZE_TIMESET, i),
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;
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.
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;
983 netif_warn(efx, drv, efx->net_dev,
984 "PTP no suitable synchronisations\n");
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.
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");
1002 delta.tv_sec = (last_sec - start_sec) & 1;
1004 last_time->ts_real.tv_nsec -
1005 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
1007 /* Convert the NIC time at last good sync into kernel time.
1008 * No correction is required - this time is the output of a
1011 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
1012 ptp->timeset[last_good].minor, 0);
1014 /* Calculate delay from NIC top of second to last_time */
1015 delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec;
1017 /* Set PPS timestamp to match NIC top of second */
1018 ptp->host_time_pps = *last_time;
1019 pps_sub_ts(&ptp->host_time_pps, delta);
1024 /* Synchronize times between the host and the MC */
1025 static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
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;
1031 unsigned long timeout;
1032 struct pps_event_time last_time = {};
1033 unsigned int loops = 0;
1034 int *start = ptp->start.addr;
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,
1040 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
1041 ptp->start.dma_addr);
1043 /* Clear flag that signals MC ready */
1044 WRITE_ONCE(*start, 0);
1045 rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf,
1046 MC_CMD_PTP_IN_SYNCHRONIZE_LEN);
1047 EFX_WARN_ON_ONCE_PARANOID(rc);
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 */
1058 if (!time_before(jiffies, timeout))
1059 ++ptp->sync_timeouts;
1061 if (READ_ONCE(*start))
1062 efx_ptp_send_times(efx, &last_time);
1064 /* Collect results */
1065 rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP,
1066 MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
1067 synch_buf, sizeof(synch_buf),
1070 rc = efx_ptp_process_times(efx, synch_buf, response_length,
1075 ++ptp->no_time_syncs;
1078 /* Increment the bad syncs counter if the synchronize fails, whatever
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)
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;
1094 tx_queue = efx_channel_get_tx_queue(ptp_data->channel, type);
1095 if (tx_queue && tx_queue->timestamping) {
1096 efx_enqueue_skb(tx_queue, skb);
1098 WARN_ONCE(1, "PTP channel has no timestamped tx queue\n");
1099 dev_kfree_skb_any(skb);
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)
1106 struct efx_ptp_data *ptp_data = efx->ptp_data;
1107 struct skb_shared_hwtstamps timestamps;
1109 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
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);
1121 if (skb->ip_summed == CHECKSUM_PARTIAL) {
1122 rc = skb_checksum_help(skb);
1126 skb_copy_from_linear_data(skb,
1127 MCDI_PTR(ptp_data->txbuf,
1128 PTP_IN_TRANSMIT_PACKET),
1130 rc = efx_mcdi_rpc(efx, MC_CMD_PTP,
1131 ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len),
1132 txtime, sizeof(txtime), &len);
1136 memset(×tamps, 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);
1142 skb_tstamp_tx(skb, ×tamps);
1147 dev_kfree_skb_any(skb);
1152 static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
1154 struct efx_ptp_data *ptp = efx->ptp_data;
1155 struct list_head *cursor;
1156 struct list_head *next;
1158 if (ptp->rx_ts_inline)
1161 /* Drop time-expired events */
1162 spin_lock_bh(&ptp->evt_lock);
1163 list_for_each_safe(cursor, next, &ptp->evt_list) {
1164 struct efx_ptp_event_rx *evt;
1166 evt = list_entry(cursor, struct efx_ptp_event_rx,
1168 if (time_after(jiffies, evt->expiry)) {
1169 list_move(&evt->link, &ptp->evt_free_list);
1170 netif_warn(efx, hw, efx->net_dev,
1171 "PTP rx event dropped\n");
1174 spin_unlock_bh(&ptp->evt_lock);
1177 static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
1178 struct sk_buff *skb)
1180 struct efx_ptp_data *ptp = efx->ptp_data;
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;
1187 WARN_ON_ONCE(ptp->rx_ts_inline);
1189 spin_lock_bh(&ptp->evt_lock);
1190 evts_waiting = !list_empty(&ptp->evt_list);
1191 spin_unlock_bh(&ptp->evt_lock);
1194 return PTP_PACKET_STATE_UNMATCHED;
1196 match = (struct efx_ptp_match *)skb->cb;
1197 /* Look for a matching timestamp in the event queue */
1198 spin_lock_bh(&ptp->evt_lock);
1199 list_for_each_safe(cursor, next, &ptp->evt_list) {
1200 struct efx_ptp_event_rx *evt;
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;
1207 /* Match - add in hardware timestamp */
1208 timestamps = skb_hwtstamps(skb);
1209 timestamps->hwtstamp = evt->hwtimestamp;
1211 match->state = PTP_PACKET_STATE_MATCHED;
1212 rc = PTP_PACKET_STATE_MATCHED;
1213 list_move(&evt->link, &ptp->evt_free_list);
1217 spin_unlock_bh(&ptp->evt_lock);
1222 /* Process any queued receive events and corresponding packets
1224 * q is returned with all the packets that are ready for delivery.
1226 static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
1228 struct efx_ptp_data *ptp = efx->ptp_data;
1229 struct sk_buff *skb;
1231 while ((skb = skb_dequeue(&ptp->rxq))) {
1232 struct efx_ptp_match *match;
1234 match = (struct efx_ptp_match *)skb->cb;
1235 if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) {
1236 __skb_queue_tail(q, skb);
1237 } else if (efx_ptp_match_rx(efx, skb) ==
1238 PTP_PACKET_STATE_MATCHED) {
1239 __skb_queue_tail(q, 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(q, skb);
1245 /* Replace unprocessed entry and stop */
1246 skb_queue_head(&ptp->rxq, skb);
1252 /* Complete processing of a received packet */
1253 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1256 netif_receive_skb(skb);
1260 static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
1262 struct efx_ptp_data *ptp = efx->ptp_data;
1264 if (ptp->rxfilter_installed) {
1265 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1266 ptp->rxfilter_general);
1267 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1268 ptp->rxfilter_event);
1269 ptp->rxfilter_installed = false;
1273 static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1275 struct efx_ptp_data *ptp = efx->ptp_data;
1276 struct efx_filter_spec rxfilter;
1279 if (!ptp->channel || ptp->rxfilter_installed)
1282 /* Must filter on both event and general ports to ensure
1283 * that there is no packet re-ordering.
1285 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1287 efx_channel_get_rx_queue(ptp->channel)));
1288 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1290 htons(PTP_EVENT_PORT));
1294 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1297 ptp->rxfilter_event = rc;
1299 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1301 efx_channel_get_rx_queue(ptp->channel)));
1302 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1304 htons(PTP_GENERAL_PORT));
1308 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1311 ptp->rxfilter_general = rc;
1313 ptp->rxfilter_installed = true;
1317 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1318 ptp->rxfilter_event);
1322 static int efx_ptp_start(struct efx_nic *efx)
1324 struct efx_ptp_data *ptp = efx->ptp_data;
1327 ptp->reset_required = false;
1329 rc = efx_ptp_insert_multicast_filters(efx);
1333 rc = efx_ptp_enable(efx);
1337 ptp->evt_frag_idx = 0;
1338 ptp->current_adjfreq = 0;
1343 efx_ptp_remove_multicast_filters(efx);
1347 static int efx_ptp_stop(struct efx_nic *efx)
1349 struct efx_ptp_data *ptp = efx->ptp_data;
1350 struct list_head *cursor;
1351 struct list_head *next;
1357 rc = efx_ptp_disable(efx);
1359 efx_ptp_remove_multicast_filters(efx);
1361 /* Make sure RX packets are really delivered */
1362 efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq);
1363 skb_queue_purge(&efx->ptp_data->txq);
1365 /* Drop any pending receive events */
1366 spin_lock_bh(&efx->ptp_data->evt_lock);
1367 list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) {
1368 list_move(cursor, &efx->ptp_data->evt_free_list);
1370 spin_unlock_bh(&efx->ptp_data->evt_lock);
1375 static int efx_ptp_restart(struct efx_nic *efx)
1377 if (efx->ptp_data && efx->ptp_data->enabled)
1378 return efx_ptp_start(efx);
1382 static void efx_ptp_pps_worker(struct work_struct *work)
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;
1389 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
1392 ptp_evt.type = PTP_CLOCK_PPSUSR;
1393 ptp_evt.pps_times = ptp->host_time_pps;
1394 ptp_clock_event(ptp->phc_clock, &ptp_evt);
1397 static void efx_ptp_worker(struct work_struct *work)
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;
1405 if (ptp_data->reset_required) {
1411 efx_ptp_drop_time_expired_events(efx);
1413 __skb_queue_head_init(&tempq);
1414 efx_ptp_process_events(efx, &tempq);
1416 while ((skb = skb_dequeue(&ptp_data->txq)))
1417 ptp_data->xmit_skb(efx, skb);
1419 while ((skb = __skb_dequeue(&tempq)))
1420 efx_ptp_process_rx(efx, skb);
1423 static const struct ptp_clock_info efx_phc_clock_info = {
1424 .owner = THIS_MODULE,
1432 .adjfreq = efx_phc_adjfreq,
1433 .adjtime = efx_phc_adjtime,
1434 .gettime64 = efx_phc_gettime,
1435 .settime64 = efx_phc_settime,
1436 .enable = efx_phc_enable,
1439 /* Initialise PTP state. */
1440 int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1442 struct efx_ptp_data *ptp;
1446 ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL);
1447 efx->ptp_data = ptp;
1452 ptp->channel = channel;
1453 ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1455 rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL);
1459 skb_queue_head_init(&ptp->rxq);
1460 skb_queue_head_init(&ptp->txq);
1461 ptp->workwq = create_singlethread_workqueue("sfc_ptp");
1467 if (efx_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;
1472 ptp->xmit_skb = efx_ptp_xmit_skb_mc;
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(&ptp->evt_list);
1480 INIT_LIST_HEAD(&ptp->evt_free_list);
1481 spin_lock_init(&ptp->evt_lock);
1482 for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++)
1483 list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list);
1485 /* Get the NIC PTP attributes and set up time conversions */
1486 rc = efx_ptp_get_attributes(efx);
1490 /* Get the timestamp corrections */
1491 rc = efx_ptp_get_timestamp_corrections(efx);
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(&ptp->phc_clock_info,
1499 &efx->pci_dev->dev);
1500 if (IS_ERR(ptp->phc_clock)) {
1501 rc = PTR_ERR(ptp->phc_clock);
1503 } else if (ptp->phc_clock) {
1504 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1505 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
1506 if (!ptp->pps_workwq) {
1512 ptp->nic_ts_enabled = false;
1516 ptp_clock_unregister(efx->ptp_data->phc_clock);
1519 destroy_workqueue(efx->ptp_data->workwq);
1522 efx_nic_free_buffer(efx, &ptp->start);
1525 kfree(efx->ptp_data);
1526 efx->ptp_data = NULL;
1531 /* Initialise PTP channel.
1533 * Setting core_index to zero causes the queue to be initialised and doesn't
1534 * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
1536 static int efx_ptp_probe_channel(struct efx_channel *channel)
1538 struct efx_nic *efx = channel->efx;
1541 channel->irq_moderation_us = 0;
1542 channel->rx_queue.core_index = 0;
1544 rc = efx_ptp_probe(efx, channel);
1545 /* Failure to probe PTP is not fatal; this channel will just not be
1546 * used for anything.
1547 * In the case of EPERM, efx_ptp_probe will print its own message (in
1548 * efx_ptp_get_attributes()), so we don't need to.
1550 if (rc && rc != -EPERM)
1551 netif_warn(efx, drv, efx->net_dev,
1552 "Failed to probe PTP, rc=%d\n", rc);
1556 void efx_ptp_remove(struct efx_nic *efx)
1561 (void)efx_ptp_disable(efx);
1563 cancel_work_sync(&efx->ptp_data->work);
1564 if (efx->ptp_data->pps_workwq)
1565 cancel_work_sync(&efx->ptp_data->pps_work);
1567 skb_queue_purge(&efx->ptp_data->rxq);
1568 skb_queue_purge(&efx->ptp_data->txq);
1570 if (efx->ptp_data->phc_clock) {
1571 destroy_workqueue(efx->ptp_data->pps_workwq);
1572 ptp_clock_unregister(efx->ptp_data->phc_clock);
1575 destroy_workqueue(efx->ptp_data->workwq);
1577 efx_nic_free_buffer(efx, &efx->ptp_data->start);
1578 kfree(efx->ptp_data);
1579 efx->ptp_data = NULL;
1582 static void efx_ptp_remove_channel(struct efx_channel *channel)
1584 efx_ptp_remove(channel->efx);
1587 static void efx_ptp_get_channel_name(struct efx_channel *channel,
1588 char *buf, size_t len)
1590 snprintf(buf, len, "%s-ptp", channel->efx->name);
1593 /* Determine whether this packet should be processed by the PTP module
1594 * or transmitted conventionally.
1596 bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1598 return efx->ptp_data &&
1599 efx->ptp_data->enabled &&
1600 skb->len >= PTP_MIN_LENGTH &&
1601 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1602 likely(skb->protocol == htons(ETH_P_IP)) &&
1603 skb_transport_header_was_set(skb) &&
1604 skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1605 ip_hdr(skb)->protocol == IPPROTO_UDP &&
1607 skb_transport_offset(skb) + sizeof(struct udphdr) &&
1608 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1611 /* Receive a PTP packet. Packets are queued until the arrival of
1612 * the receive timestamp from the MC - this will probably occur after the
1613 * packet arrival because of the processing in the MC.
1615 static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1617 struct efx_nic *efx = channel->efx;
1618 struct efx_ptp_data *ptp = efx->ptp_data;
1619 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1620 u8 *match_data_012, *match_data_345;
1621 unsigned int version;
1624 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1626 /* Correct version? */
1627 if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1628 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1632 version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1633 if (version != PTP_VERSION_V1) {
1637 /* PTP V1 uses all six bytes of the UUID to match the packet
1640 match_data_012 = data + PTP_V1_UUID_OFFSET;
1641 match_data_345 = data + PTP_V1_UUID_OFFSET + 3;
1643 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1647 version = data[PTP_V2_VERSION_OFFSET];
1648 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1652 /* The original V2 implementation uses bytes 2-7 of
1653 * the UUID to match the packet to the timestamp. This
1654 * discards two of the bytes of the MAC address used
1655 * to create the UUID (SF bug 33070). The PTP V2
1656 * enhanced mode fixes this issue and uses bytes 0-2
1657 * and byte 5-7 of the UUID.
1659 match_data_345 = data + PTP_V2_UUID_OFFSET + 5;
1660 if (ptp->mode == MC_CMD_PTP_MODE_V2) {
1661 match_data_012 = data + PTP_V2_UUID_OFFSET + 2;
1663 match_data_012 = data + PTP_V2_UUID_OFFSET + 0;
1664 BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED);
1668 /* Does this packet require timestamping? */
1669 if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1670 match->state = PTP_PACKET_STATE_UNMATCHED;
1672 /* We expect the sequence number to be in the same position in
1673 * the packet for PTP V1 and V2
1675 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1676 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1678 /* Extract UUID/Sequence information */
1679 match->words[0] = (match_data_012[0] |
1680 (match_data_012[1] << 8) |
1681 (match_data_012[2] << 16) |
1682 (match_data_345[0] << 24));
1683 match->words[1] = (match_data_345[1] |
1684 (match_data_345[2] << 8) |
1685 (data[PTP_V1_SEQUENCE_OFFSET +
1686 PTP_V1_SEQUENCE_LENGTH - 1] <<
1689 match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1692 skb_queue_tail(&ptp->rxq, skb);
1693 queue_work(ptp->workwq, &ptp->work);
1698 /* Transmit a PTP packet. This has to be transmitted by the MC
1699 * itself, through an MCDI call. MCDI calls aren't permitted
1700 * in the transmit path so defer the actual transmission to a suitable worker.
1702 int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1704 struct efx_ptp_data *ptp = efx->ptp_data;
1706 skb_queue_tail(&ptp->txq, skb);
1708 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1709 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1710 efx_xmit_hwtstamp_pending(skb);
1711 queue_work(ptp->workwq, &ptp->work);
1713 return NETDEV_TX_OK;
1716 int efx_ptp_get_mode(struct efx_nic *efx)
1718 return efx->ptp_data->mode;
1721 int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1722 unsigned int new_mode)
1724 if ((enable_wanted != efx->ptp_data->enabled) ||
1725 (enable_wanted && (efx->ptp_data->mode != new_mode))) {
1728 if (enable_wanted) {
1729 /* Change of mode requires disable */
1730 if (efx->ptp_data->enabled &&
1731 (efx->ptp_data->mode != new_mode)) {
1732 efx->ptp_data->enabled = false;
1733 rc = efx_ptp_stop(efx);
1738 /* Set new operating mode and establish
1739 * baseline synchronisation, which must
1742 efx->ptp_data->mode = new_mode;
1743 if (netif_running(efx->net_dev))
1744 rc = efx_ptp_start(efx);
1746 rc = efx_ptp_synchronize(efx,
1747 PTP_SYNC_ATTEMPTS * 2);
1752 rc = efx_ptp_stop(efx);
1758 efx->ptp_data->enabled = enable_wanted;
1764 static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
1771 if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1772 (init->tx_type != HWTSTAMP_TX_ON))
1775 rc = efx->type->ptp_set_ts_config(efx, init);
1779 efx->ptp_data->config = *init;
1783 void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
1785 struct efx_ptp_data *ptp = efx->ptp_data;
1786 struct efx_nic *primary = efx->primary;
1793 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1794 SOF_TIMESTAMPING_RX_HARDWARE |
1795 SOF_TIMESTAMPING_RAW_HARDWARE);
1796 /* Check licensed features. If we don't have the license for TX
1797 * timestamps, the NIC will not support them.
1799 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1800 struct efx_ef10_nic_data *nic_data = efx->nic_data;
1802 if (!(nic_data->licensed_features &
1803 (1 << LICENSED_V3_FEATURES_TX_TIMESTAMPS_LBN)))
1804 ts_info->so_timestamping &=
1805 ~SOF_TIMESTAMPING_TX_HARDWARE;
1807 if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1808 ts_info->phc_index =
1809 ptp_clock_index(primary->ptp_data->phc_clock);
1810 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1811 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1814 int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1816 struct hwtstamp_config config;
1819 /* Not a PTP enabled port */
1823 if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
1826 rc = efx_ptp_ts_init(efx, &config);
1830 return copy_to_user(ifr->ifr_data, &config, sizeof(config))
1834 int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1839 return copy_to_user(ifr->ifr_data, &efx->ptp_data->config,
1840 sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
1843 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1845 struct efx_ptp_data *ptp = efx->ptp_data;
1847 netif_err(efx, hw, efx->net_dev,
1848 "PTP unexpected event length: got %d expected %d\n",
1849 ptp->evt_frag_idx, expected_frag_len);
1850 ptp->reset_required = true;
1851 queue_work(ptp->workwq, &ptp->work);
1854 /* Process a completed receive event. Put it on the event queue and
1855 * start worker thread. This is required because event and their
1856 * correspoding packets may come in either order.
1858 static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
1860 struct efx_ptp_event_rx *evt = NULL;
1862 if (WARN_ON_ONCE(ptp->rx_ts_inline))
1865 if (ptp->evt_frag_idx != 3) {
1866 ptp_event_failure(efx, 3);
1870 spin_lock_bh(&ptp->evt_lock);
1871 if (!list_empty(&ptp->evt_free_list)) {
1872 evt = list_first_entry(&ptp->evt_free_list,
1873 struct efx_ptp_event_rx, link);
1874 list_del(&evt->link);
1876 evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA);
1877 evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2],
1879 (EFX_QWORD_FIELD(ptp->evt_frags[1],
1880 MCDI_EVENT_SRC) << 8) |
1881 (EFX_QWORD_FIELD(ptp->evt_frags[0],
1882 MCDI_EVENT_SRC) << 16));
1883 evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
1884 EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
1885 EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
1886 ptp->ts_corrections.ptp_rx);
1887 evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1888 list_add_tail(&evt->link, &ptp->evt_list);
1890 queue_work(ptp->workwq, &ptp->work);
1891 } else if (net_ratelimit()) {
1892 /* Log a rate-limited warning message. */
1893 netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n");
1895 spin_unlock_bh(&ptp->evt_lock);
1898 static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1900 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1901 if (ptp->evt_frag_idx != 1) {
1902 ptp_event_failure(efx, 1);
1906 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1909 static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1911 if (ptp->nic_ts_enabled)
1912 queue_work(ptp->pps_workwq, &ptp->pps_work);
1915 void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1917 struct efx_ptp_data *ptp = efx->ptp_data;
1918 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
1921 if (!efx->ptp_warned) {
1922 netif_warn(efx, drv, efx->net_dev,
1923 "Received PTP event but PTP not set up\n");
1924 efx->ptp_warned = true;
1932 if (ptp->evt_frag_idx == 0) {
1933 ptp->evt_code = code;
1934 } else if (ptp->evt_code != code) {
1935 netif_err(efx, hw, efx->net_dev,
1936 "PTP out of sequence event %d\n", code);
1937 ptp->evt_frag_idx = 0;
1940 ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
1941 if (!MCDI_EVENT_FIELD(*ev, CONT)) {
1942 /* Process resulting event */
1944 case MCDI_EVENT_CODE_PTP_RX:
1945 ptp_event_rx(efx, ptp);
1947 case MCDI_EVENT_CODE_PTP_FAULT:
1948 ptp_event_fault(efx, ptp);
1950 case MCDI_EVENT_CODE_PTP_PPS:
1951 ptp_event_pps(efx, ptp);
1954 netif_err(efx, hw, efx->net_dev,
1955 "PTP unknown event %d\n", code);
1958 ptp->evt_frag_idx = 0;
1959 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
1960 netif_err(efx, hw, efx->net_dev,
1961 "PTP too many event fragments\n");
1962 ptp->evt_frag_idx = 0;
1966 void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
1968 struct efx_nic *efx = channel->efx;
1969 struct efx_ptp_data *ptp = efx->ptp_data;
1971 /* When extracting the sync timestamp minor value, we should discard
1972 * the least significant two bits. These are not required in order
1973 * to reconstruct full-range timestamps and they are optionally used
1974 * to report status depending on the options supplied when subscribing
1977 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
1978 channel->sync_timestamp_minor =
1979 (MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_MS_8BITS) & 0xFC)
1980 << ptp->nic_time.sync_event_minor_shift;
1982 /* if sync events have been disabled then we want to silently ignore
1983 * this event, so throw away result.
1985 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
1989 static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
1991 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
1992 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset));
1994 const u8 *data = eh + efx->rx_packet_ts_offset;
1995 return (u32)data[0] |
1997 (u32)data[2] << 16 |
2002 void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
2003 struct sk_buff *skb)
2005 struct efx_nic *efx = channel->efx;
2006 struct efx_ptp_data *ptp = efx->ptp_data;
2007 u32 pkt_timestamp_major, pkt_timestamp_minor;
2009 struct skb_shared_hwtstamps *timestamps;
2011 if (channel->sync_events_state != SYNC_EVENTS_VALID)
2014 pkt_timestamp_minor = efx_rx_buf_timestamp_minor(efx, skb_mac_header(skb));
2016 /* get the difference between the packet and sync timestamps,
2019 diff = pkt_timestamp_minor - channel->sync_timestamp_minor;
2020 if (pkt_timestamp_minor < channel->sync_timestamp_minor)
2021 diff += ptp->nic_time.minor_max;
2023 /* do we roll over a second boundary and need to carry the one? */
2024 carry = (channel->sync_timestamp_minor >= ptp->nic_time.minor_max - diff) ?
2027 if (diff <= ptp->nic_time.sync_event_diff_max) {
2028 /* packet is ahead of the sync event by a quarter of a second or
2029 * less (allowing for fuzz)
2031 pkt_timestamp_major = channel->sync_timestamp_major + carry;
2032 } else if (diff >= ptp->nic_time.sync_event_diff_min) {
2033 /* packet is behind the sync event but within the fuzz factor.
2034 * This means the RX packet and sync event crossed as they were
2035 * placed on the event queue, which can sometimes happen.
2037 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
2039 /* it's outside tolerance in both directions. this might be
2040 * indicative of us missing sync events for some reason, so
2041 * we'll call it an error rather than risk giving a bogus
2044 netif_vdbg(efx, drv, efx->net_dev,
2045 "packet timestamp %x too far from sync event %x:%x\n",
2046 pkt_timestamp_minor, channel->sync_timestamp_major,
2047 channel->sync_timestamp_minor);
2051 /* attach the timestamps to the skb */
2052 timestamps = skb_hwtstamps(skb);
2053 timestamps->hwtstamp =
2054 ptp->nic_to_kernel_time(pkt_timestamp_major,
2055 pkt_timestamp_minor,
2056 ptp->ts_corrections.general_rx);
2059 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
2061 struct efx_ptp_data *ptp_data = container_of(ptp,
2062 struct efx_ptp_data,
2064 struct efx_nic *efx = ptp_data->efx;
2065 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
2069 if (delta > MAX_PPB)
2071 else if (delta < -MAX_PPB)
2074 /* Convert ppb to fixed point ns taking care to round correctly. */
2075 adjustment_ns = ((s64)delta * PPB_SCALE_WORD +
2076 (1 << (ptp_data->adjfreq_ppb_shift - 1))) >>
2077 ptp_data->adjfreq_ppb_shift;
2079 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2080 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
2081 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
2082 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
2083 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
2084 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj),
2089 ptp_data->current_adjfreq = adjustment_ns;
2093 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
2095 u32 nic_major, nic_minor;
2096 struct efx_ptp_data *ptp_data = container_of(ptp,
2097 struct efx_ptp_data,
2099 struct efx_nic *efx = ptp_data->efx;
2100 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
2102 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
2104 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2105 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2106 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
2107 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
2108 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
2109 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2113 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
2115 struct efx_ptp_data *ptp_data = container_of(ptp,
2116 struct efx_ptp_data,
2118 struct efx_nic *efx = ptp_data->efx;
2119 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
2120 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
2124 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
2125 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2127 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2128 outbuf, sizeof(outbuf), NULL);
2132 kt = ptp_data->nic_to_kernel_time(
2133 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
2134 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
2135 *ts = ktime_to_timespec64(kt);
2139 static int efx_phc_settime(struct ptp_clock_info *ptp,
2140 const struct timespec64 *e_ts)
2142 /* Get the current NIC time, efx_phc_gettime.
2143 * Subtract from the desired time to get the offset
2144 * call efx_phc_adjtime with the offset
2147 struct timespec64 time_now;
2148 struct timespec64 delta;
2150 rc = efx_phc_gettime(ptp, &time_now);
2154 delta = timespec64_sub(*e_ts, time_now);
2156 rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta));
2163 static int efx_phc_enable(struct ptp_clock_info *ptp,
2164 struct ptp_clock_request *request,
2167 struct efx_ptp_data *ptp_data = container_of(ptp,
2168 struct efx_ptp_data,
2170 if (request->type != PTP_CLK_REQ_PPS)
2173 ptp_data->nic_ts_enabled = !!enable;
2177 static const struct efx_channel_type efx_ptp_channel_type = {
2178 .handle_no_channel = efx_ptp_handle_no_channel,
2179 .pre_probe = efx_ptp_probe_channel,
2180 .post_remove = efx_ptp_remove_channel,
2181 .get_name = efx_ptp_get_channel_name,
2182 /* no copy operation; there is no need to reallocate this channel */
2183 .receive_skb = efx_ptp_rx,
2184 .want_txqs = efx_ptp_want_txqs,
2185 .keep_eventq = false,
2188 void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
2190 /* Check whether PTP is implemented on this NIC. The DISABLE
2191 * operation will succeed if and only if it is implemented.
2193 if (efx_ptp_disable(efx) == 0)
2194 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
2195 &efx_ptp_channel_type;
2198 void efx_ptp_start_datapath(struct efx_nic *efx)
2200 if (efx_ptp_restart(efx))
2201 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
2202 /* re-enable timestamping if it was previously enabled */
2203 if (efx->type->ptp_set_ts_sync_events)
2204 efx->type->ptp_set_ts_sync_events(efx, true, true);
2207 void efx_ptp_stop_datapath(struct efx_nic *efx)
2209 /* temporarily disable timestamping */
2210 if (efx->type->ptp_set_ts_sync_events)
2211 efx->type->ptp_set_ts_sync_events(efx, false, true);