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 */
48 #include "efx_channels.h"
50 /* Maximum number of events expected to make up a PTP event */
51 #define MAX_EVENT_FRAGS 3
53 /* Maximum delay, ms, to begin synchronisation */
54 #define MAX_SYNCHRONISE_WAIT_MS 2
56 /* How long, at most, to spend synchronising */
57 #define SYNCHRONISE_PERIOD_NS 250000
59 /* How often to update the shared memory time */
60 #define SYNCHRONISATION_GRANULARITY_NS 200
62 /* Minimum permitted length of a (corrected) synchronisation time */
63 #define DEFAULT_MIN_SYNCHRONISATION_NS 120
65 /* Maximum permitted length of a (corrected) synchronisation time */
66 #define MAX_SYNCHRONISATION_NS 1000
68 /* How many (MC) receive events that can be queued */
69 #define MAX_RECEIVE_EVENTS 8
71 /* Length of (modified) moving average. */
72 #define AVERAGE_LENGTH 16
74 /* How long an unmatched event or packet can be held */
75 #define PKT_EVENT_LIFETIME_MS 10
77 /* Offsets into PTP packet for identification. These offsets are from the
78 * start of the IP header, not the MAC header. Note that neither PTP V1 nor
79 * PTP V2 permit the use of IPV4 options.
81 #define PTP_DPORT_OFFSET 22
83 #define PTP_V1_VERSION_LENGTH 2
84 #define PTP_V1_VERSION_OFFSET 28
86 #define PTP_V1_UUID_LENGTH 6
87 #define PTP_V1_UUID_OFFSET 50
89 #define PTP_V1_SEQUENCE_LENGTH 2
90 #define PTP_V1_SEQUENCE_OFFSET 58
92 /* The minimum length of a PTP V1 packet for offsets, etc. to be valid:
95 #define PTP_V1_MIN_LENGTH 64
97 #define PTP_V2_VERSION_LENGTH 1
98 #define PTP_V2_VERSION_OFFSET 29
100 #define PTP_V2_UUID_LENGTH 8
101 #define PTP_V2_UUID_OFFSET 48
103 /* Although PTP V2 UUIDs are comprised a ClockIdentity (8) and PortNumber (2),
104 * the MC only captures the last six bytes of the clock identity. These values
105 * reflect those, not the ones used in the standard. The standard permits
106 * mapping of V1 UUIDs to V2 UUIDs with these same values.
108 #define PTP_V2_MC_UUID_LENGTH 6
109 #define PTP_V2_MC_UUID_OFFSET 50
111 #define PTP_V2_SEQUENCE_LENGTH 2
112 #define PTP_V2_SEQUENCE_OFFSET 58
114 /* The minimum length of a PTP V2 packet for offsets, etc. to be valid:
115 * includes IP header.
117 #define PTP_V2_MIN_LENGTH 63
119 #define PTP_MIN_LENGTH 63
121 #define PTP_ADDRESS 0xe0000181 /* 224.0.1.129 */
122 #define PTP_EVENT_PORT 319
123 #define PTP_GENERAL_PORT 320
125 /* Annoyingly the format of the version numbers are different between
126 * versions 1 and 2 so it isn't possible to simply look for 1 or 2.
128 #define PTP_VERSION_V1 1
130 #define PTP_VERSION_V2 2
131 #define PTP_VERSION_V2_MASK 0x0f
133 enum ptp_packet_state {
134 PTP_PACKET_STATE_UNMATCHED = 0,
135 PTP_PACKET_STATE_MATCHED,
136 PTP_PACKET_STATE_TIMED_OUT,
137 PTP_PACKET_STATE_MATCH_UNWANTED
140 /* NIC synchronised with single word of time only comprising
141 * partial seconds and full nanoseconds: 10^9 ~ 2^30 so 2 bits for seconds.
143 #define MC_NANOSECOND_BITS 30
144 #define MC_NANOSECOND_MASK ((1 << MC_NANOSECOND_BITS) - 1)
145 #define MC_SECOND_MASK ((1 << (32 - MC_NANOSECOND_BITS)) - 1)
147 /* Maximum parts-per-billion adjustment that is acceptable */
148 #define MAX_PPB 1000000
150 /* Precalculate scale word to avoid long long division at runtime */
151 /* This is equivalent to 2^66 / 10^9. */
152 #define PPB_SCALE_WORD ((1LL << (57)) / 1953125LL)
154 /* How much to shift down after scaling to convert to FP40 */
155 #define PPB_SHIFT_FP40 26
157 #define PPB_SHIFT_FP44 22
159 #define PTP_SYNC_ATTEMPTS 4
162 * struct efx_ptp_match - Matching structure, stored in sk_buff's cb area.
163 * @words: UUID and (partial) sequence number
164 * @expiry: Time after which the packet should be delivered irrespective of
166 * @state: The state of the packet - whether it is ready for processing or
167 * whether that is of no interest.
169 struct efx_ptp_match {
170 u32 words[DIV_ROUND_UP(PTP_V1_UUID_LENGTH, 4)];
171 unsigned long expiry;
172 enum ptp_packet_state state;
176 * struct efx_ptp_event_rx - A PTP receive event (from MC)
177 * @link: list of events
178 * @seq0: First part of (PTP) UUID
179 * @seq1: Second part of (PTP) UUID and sequence number
180 * @hwtimestamp: Event timestamp
181 * @expiry: Time which the packet arrived
183 struct efx_ptp_event_rx {
184 struct list_head link;
188 unsigned long expiry;
192 * struct efx_ptp_timeset - Synchronisation between host and MC
193 * @host_start: Host time immediately before hardware timestamp taken
194 * @major: Hardware timestamp, major
195 * @minor: Hardware timestamp, minor
196 * @host_end: Host time immediately after hardware timestamp taken
197 * @wait: Number of NIC clock ticks between hardware timestamp being read and
198 * host end time being seen
199 * @window: Difference of host_end and host_start
200 * @valid: Whether this timeset is valid
202 struct efx_ptp_timeset {
208 u32 window; /* Derived: end - start, allowing for wrap */
212 * struct efx_ptp_data - Precision Time Protocol (PTP) state
213 * @efx: The NIC context
214 * @channel: The PTP channel (Siena only)
215 * @rx_ts_inline: Flag for whether RX timestamps are inline (else they are
217 * @rxq: Receive SKB queue (awaiting timestamps)
218 * @txq: Transmit SKB queue
219 * @evt_list: List of MC receive events awaiting packets
220 * @evt_free_list: List of free events
221 * @evt_lock: Lock for manipulating evt_list and evt_free_list
222 * @rx_evts: Instantiated events (on evt_list and evt_free_list)
223 * @workwq: Work queue for processing pending PTP operations
225 * @reset_required: A serious error has occurred and the PTP task needs to be
226 * reset (disable, enable).
227 * @rxfilter_event: Receive filter when operating
228 * @rxfilter_general: Receive filter when operating
229 * @rxfilter_installed: Receive filter installed
230 * @config: Current timestamp configuration
231 * @enabled: PTP operation enabled
232 * @mode: Mode in which PTP operating (PTP version)
233 * @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time
234 * @nic_to_kernel_time: Function to convert from NIC to kernel time
235 * @nic_time: contains time details
236 * @nic_time.minor_max: Wrap point for NIC minor times
237 * @nic_time.sync_event_diff_min: Minimum acceptable difference between time
238 * in packet prefix and last MCDI time sync event i.e. how much earlier than
239 * the last sync event time a packet timestamp can be.
240 * @nic_time.sync_event_diff_max: Maximum acceptable difference between time
241 * in packet prefix and last MCDI time sync event i.e. how much later than
242 * the last sync event time a packet timestamp can be.
243 * @nic_time.sync_event_minor_shift: Shift required to make minor time from
244 * field in MCDI time sync event.
245 * @min_synchronisation_ns: Minimum acceptable corrected sync window
246 * @capabilities: Capabilities flags from the NIC
247 * @ts_corrections: contains corrections details
248 * @ts_corrections.ptp_tx: Required driver correction of PTP packet transmit
250 * @ts_corrections.ptp_rx: Required driver correction of PTP packet receive
252 * @ts_corrections.pps_out: PPS output error (information only)
253 * @ts_corrections.pps_in: Required driver correction of PPS input timestamps
254 * @ts_corrections.general_tx: Required driver correction of general packet
255 * transmit timestamps
256 * @ts_corrections.general_rx: Required driver correction of general packet
258 * @evt_frags: Partly assembled PTP events
259 * @evt_frag_idx: Current fragment number
260 * @evt_code: Last event code
261 * @start: Address at which MC indicates ready for synchronisation
262 * @host_time_pps: Host time at last PPS
263 * @adjfreq_ppb_shift: Shift required to convert scaled parts-per-billion
264 * frequency adjustment into a fixed point fractional nanosecond format.
265 * @current_adjfreq: Current ppb adjustment.
266 * @phc_clock: Pointer to registered phc device (if primary function)
267 * @phc_clock_info: Registration structure for phc device
268 * @pps_work: pps work task for handling pps events
269 * @pps_workwq: pps work queue
270 * @nic_ts_enabled: Flag indicating if NIC generated TS events are handled
271 * @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids
272 * allocations in main data path).
273 * @good_syncs: Number of successful synchronisations.
274 * @fast_syncs: Number of synchronisations requiring short delay
275 * @bad_syncs: Number of failed synchronisations.
276 * @sync_timeouts: Number of synchronisation timeouts
277 * @no_time_syncs: Number of synchronisations with no good times.
278 * @invalid_sync_windows: Number of sync windows with bad durations.
279 * @undersize_sync_windows: Number of corrected sync windows that are too small
280 * @oversize_sync_windows: Number of corrected sync windows that are too large
281 * @rx_no_timestamp: Number of packets received without a timestamp.
282 * @timeset: Last set of synchronisation statistics.
283 * @xmit_skb: Transmit SKB function.
285 struct efx_ptp_data {
287 struct efx_channel *channel;
289 struct sk_buff_head rxq;
290 struct sk_buff_head txq;
291 struct list_head evt_list;
292 struct list_head evt_free_list;
294 struct efx_ptp_event_rx rx_evts[MAX_RECEIVE_EVENTS];
295 struct workqueue_struct *workwq;
296 struct work_struct work;
299 u32 rxfilter_general;
300 bool rxfilter_installed;
301 struct hwtstamp_config config;
304 void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor);
305 ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor,
309 u32 sync_event_diff_min;
310 u32 sync_event_diff_max;
311 unsigned int sync_event_minor_shift;
313 unsigned int min_synchronisation_ns;
314 unsigned int capabilities;
323 efx_qword_t evt_frags[MAX_EVENT_FRAGS];
326 struct efx_buffer start;
327 struct pps_event_time host_time_pps;
328 unsigned int adjfreq_ppb_shift;
330 struct ptp_clock *phc_clock;
331 struct ptp_clock_info phc_clock_info;
332 struct work_struct pps_work;
333 struct workqueue_struct *pps_workwq;
335 efx_dword_t txbuf[MCDI_TX_BUF_LEN(MC_CMD_PTP_IN_TRANSMIT_LENMAX)];
337 unsigned int good_syncs;
338 unsigned int fast_syncs;
339 unsigned int bad_syncs;
340 unsigned int sync_timeouts;
341 unsigned int no_time_syncs;
342 unsigned int invalid_sync_windows;
343 unsigned int undersize_sync_windows;
344 unsigned int oversize_sync_windows;
345 unsigned int rx_no_timestamp;
346 struct efx_ptp_timeset
347 timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM];
348 void (*xmit_skb)(struct efx_nic *efx, struct sk_buff *skb);
351 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta);
352 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta);
353 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts);
354 static int efx_phc_settime(struct ptp_clock_info *ptp,
355 const struct timespec64 *e_ts);
356 static int efx_phc_enable(struct ptp_clock_info *ptp,
357 struct ptp_clock_request *request, int on);
359 bool efx_ptp_use_mac_tx_timestamps(struct efx_nic *efx)
361 return efx_has_cap(efx, TX_MAC_TIMESTAMPING);
364 /* PTP 'extra' channel is still a traffic channel, but we only create TX queues
365 * if PTP uses MAC TX timestamps, not if PTP uses the MC directly to transmit.
367 static bool efx_ptp_want_txqs(struct efx_channel *channel)
369 return efx_ptp_use_mac_tx_timestamps(channel->efx);
372 #define PTP_SW_STAT(ext_name, field_name) \
373 { #ext_name, 0, offsetof(struct efx_ptp_data, field_name) }
374 #define PTP_MC_STAT(ext_name, mcdi_name) \
375 { #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST }
376 static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = {
377 PTP_SW_STAT(ptp_good_syncs, good_syncs),
378 PTP_SW_STAT(ptp_fast_syncs, fast_syncs),
379 PTP_SW_STAT(ptp_bad_syncs, bad_syncs),
380 PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts),
381 PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs),
382 PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows),
383 PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows),
384 PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows),
385 PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp),
386 PTP_MC_STAT(ptp_tx_timestamp_packets, TX),
387 PTP_MC_STAT(ptp_rx_timestamp_packets, RX),
388 PTP_MC_STAT(ptp_timestamp_packets, TS),
389 PTP_MC_STAT(ptp_filter_matches, FM),
390 PTP_MC_STAT(ptp_non_filter_matches, NFM),
392 #define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc)
393 static const unsigned long efx_ptp_stat_mask[] = {
394 [0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL,
397 size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
402 return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
403 efx_ptp_stat_mask, strings);
406 size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats)
408 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
409 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
416 /* Copy software statistics */
417 for (i = 0; i < PTP_STAT_COUNT; i++) {
418 if (efx_ptp_stat_desc[i].dma_width)
420 stats[i] = *(unsigned int *)((char *)efx->ptp_data +
421 efx_ptp_stat_desc[i].offset);
424 /* Fetch MC statistics. We *must* fill in all statistics or
425 * risk leaking kernel memory to userland, so if the MCDI
426 * request fails we pretend we got zeroes.
428 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS);
429 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
430 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
431 outbuf, sizeof(outbuf), NULL);
433 memset(outbuf, 0, sizeof(outbuf));
434 efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
436 stats, _MCDI_PTR(outbuf, 0), false);
438 return PTP_STAT_COUNT;
441 /* For Siena platforms NIC time is s and ns */
442 static void efx_ptp_ns_to_s_ns(s64 ns, u32 *nic_major, u32 *nic_minor)
444 struct timespec64 ts = ns_to_timespec64(ns);
445 *nic_major = (u32)ts.tv_sec;
446 *nic_minor = ts.tv_nsec;
449 static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
452 ktime_t kt = ktime_set(nic_major, nic_minor);
454 kt = ktime_add_ns(kt, (u64)correction);
456 kt = ktime_sub_ns(kt, (u64)-correction);
460 /* To convert from s27 format to ns we multiply then divide by a power of 2.
461 * For the conversion from ns to s27, the operation is also converted to a
462 * multiply and shift.
464 #define S27_TO_NS_SHIFT (27)
465 #define NS_TO_S27_MULT (((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC)
466 #define NS_TO_S27_SHIFT (63 - S27_TO_NS_SHIFT)
467 #define S27_MINOR_MAX (1 << S27_TO_NS_SHIFT)
469 /* For Huntington platforms NIC time is in seconds and fractions of a second
470 * where the minor register only uses 27 bits in units of 2^-27s.
472 static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
474 struct timespec64 ts = ns_to_timespec64(ns);
475 u32 maj = (u32)ts.tv_sec;
476 u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT +
477 (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT);
479 /* The conversion can result in the minor value exceeding the maximum.
480 * In this case, round up to the next second.
482 if (min >= S27_MINOR_MAX) {
483 min -= S27_MINOR_MAX;
491 static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
493 u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC +
494 (1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT);
495 return ktime_set(nic_major, ns);
498 static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
501 /* Apply the correction and deal with carry */
502 nic_minor += correction;
503 if ((s32)nic_minor < 0) {
504 nic_minor += S27_MINOR_MAX;
506 } else if (nic_minor >= S27_MINOR_MAX) {
507 nic_minor -= S27_MINOR_MAX;
511 return efx_ptp_s27_to_ktime(nic_major, nic_minor);
514 /* For Medford2 platforms the time is in seconds and quarter nanoseconds. */
515 static void efx_ptp_ns_to_s_qns(s64 ns, u32 *nic_major, u32 *nic_minor)
517 struct timespec64 ts = ns_to_timespec64(ns);
519 *nic_major = (u32)ts.tv_sec;
520 *nic_minor = ts.tv_nsec * 4;
523 static ktime_t efx_ptp_s_qns_to_ktime_correction(u32 nic_major, u32 nic_minor,
528 nic_minor = DIV_ROUND_CLOSEST(nic_minor, 4);
529 correction = DIV_ROUND_CLOSEST(correction, 4);
531 kt = ktime_set(nic_major, nic_minor);
534 kt = ktime_add_ns(kt, (u64)correction);
536 kt = ktime_sub_ns(kt, (u64)-correction);
540 struct efx_channel *efx_ptp_channel(struct efx_nic *efx)
542 return efx->ptp_data ? efx->ptp_data->channel : NULL;
545 void efx_ptp_update_channel(struct efx_nic *efx, struct efx_channel *channel)
548 efx->ptp_data->channel = channel;
551 static u32 last_sync_timestamp_major(struct efx_nic *efx)
553 struct efx_channel *channel = efx_ptp_channel(efx);
557 major = channel->sync_timestamp_major;
561 /* The 8000 series and later can provide the time from the MAC, which is only
562 * 48 bits long and provides meta-information in the top 2 bits.
565 efx_ptp_mac_nic_to_ktime_correction(struct efx_nic *efx,
566 struct efx_ptp_data *ptp,
567 u32 nic_major, u32 nic_minor,
574 if (!(nic_major & 0x80000000)) {
575 WARN_ON_ONCE(nic_major >> 16);
577 /* Medford provides 48 bits of timestamp, so we must get the top
578 * 16 bits from the timesync event state.
580 * We only have the lower 16 bits of the time now, but we do
581 * have a full resolution timestamp at some point in past. As
582 * long as the difference between the (real) now and the sync
583 * is less than 2^15, then we can reconstruct the difference
584 * between those two numbers using only the lower 16 bits of
589 * a - b = ((a mod k) - b) mod k
591 * when -k/2 < (a-b) < k/2. In our case k is 2^16. We know
592 * (a mod k) and b, so can calculate the delta, a - b.
595 sync_timestamp = last_sync_timestamp_major(efx);
597 /* Because delta is s16 this does an implicit mask down to
598 * 16 bits which is what we need, assuming
599 * MEDFORD_TX_SECS_EVENT_BITS is 16. delta is signed so that
600 * we can deal with the (unlikely) case of sync timestamps
601 * arriving from the future.
603 delta = nic_major - sync_timestamp;
605 /* Recover the fully specified time now, by applying the offset
606 * to the (fully specified) sync time.
608 nic_major = sync_timestamp + delta;
610 kt = ptp->nic_to_kernel_time(nic_major, nic_minor,
616 ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue)
618 struct efx_nic *efx = tx_queue->efx;
619 struct efx_ptp_data *ptp = efx->ptp_data;
622 if (efx_ptp_use_mac_tx_timestamps(efx))
623 kt = efx_ptp_mac_nic_to_ktime_correction(efx, ptp,
624 tx_queue->completed_timestamp_major,
625 tx_queue->completed_timestamp_minor,
626 ptp->ts_corrections.general_tx);
628 kt = ptp->nic_to_kernel_time(
629 tx_queue->completed_timestamp_major,
630 tx_queue->completed_timestamp_minor,
631 ptp->ts_corrections.general_tx);
635 /* Get PTP attributes and set up time conversions */
636 static int efx_ptp_get_attributes(struct efx_nic *efx)
638 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN);
639 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN);
640 struct efx_ptp_data *ptp = efx->ptp_data;
645 /* Get the PTP attributes. If the NIC doesn't support the operation we
646 * use the default format for compatibility with older NICs i.e.
647 * seconds and nanoseconds.
649 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES);
650 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
651 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
652 outbuf, sizeof(outbuf), &out_len);
654 fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT);
655 } else if (rc == -EINVAL) {
656 fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS;
657 } else if (rc == -EPERM) {
658 pci_info(efx->pci_dev, "no PTP support\n");
661 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf),
662 outbuf, sizeof(outbuf), rc);
667 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION:
668 ptp->ns_to_nic_time = efx_ptp_ns_to_s27;
669 ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction;
670 ptp->nic_time.minor_max = 1 << 27;
671 ptp->nic_time.sync_event_minor_shift = 19;
673 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS:
674 ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns;
675 ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction;
676 ptp->nic_time.minor_max = 1000000000;
677 ptp->nic_time.sync_event_minor_shift = 22;
679 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_QTR_NANOSECONDS:
680 ptp->ns_to_nic_time = efx_ptp_ns_to_s_qns;
681 ptp->nic_to_kernel_time = efx_ptp_s_qns_to_ktime_correction;
682 ptp->nic_time.minor_max = 4000000000UL;
683 ptp->nic_time.sync_event_minor_shift = 24;
689 /* Precalculate acceptable difference between the minor time in the
690 * packet prefix and the last MCDI time sync event. We expect the
691 * packet prefix timestamp to be after of sync event by up to one
692 * sync event interval (0.25s) but we allow it to exceed this by a
693 * fuzz factor of (0.1s)
695 ptp->nic_time.sync_event_diff_min = ptp->nic_time.minor_max
696 - (ptp->nic_time.minor_max / 10);
697 ptp->nic_time.sync_event_diff_max = (ptp->nic_time.minor_max / 4)
698 + (ptp->nic_time.minor_max / 10);
700 /* MC_CMD_PTP_OP_GET_ATTRIBUTES has been extended twice from an older
701 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT. The function now may return
702 * a value to use for the minimum acceptable corrected synchronization
703 * window and may return further capabilities.
704 * If we have the extra information store it. For older firmware that
705 * does not implement the extended command use the default value.
708 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_CAPABILITIES_OFST)
709 ptp->min_synchronisation_ns =
711 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
713 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
716 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
717 ptp->capabilities = MCDI_DWORD(outbuf,
718 PTP_OUT_GET_ATTRIBUTES_CAPABILITIES);
720 ptp->capabilities = 0;
722 /* Set up the shift for conversion between frequency
723 * adjustments in parts-per-billion and the fixed-point
724 * fractional ns format that the adapter uses.
726 if (ptp->capabilities & (1 << MC_CMD_PTP_OUT_GET_ATTRIBUTES_FP44_FREQ_ADJ_LBN))
727 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP44;
729 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP40;
734 /* Get PTP timestamp corrections */
735 static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
737 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN);
738 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN);
742 /* Get the timestamp corrections from the NIC. If this operation is
743 * not supported (older NICs) then no correction is required.
745 MCDI_SET_DWORD(inbuf, PTP_IN_OP,
746 MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS);
747 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
749 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
750 outbuf, sizeof(outbuf), &out_len);
752 efx->ptp_data->ts_corrections.ptp_tx = MCDI_DWORD(outbuf,
753 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT);
754 efx->ptp_data->ts_corrections.ptp_rx = MCDI_DWORD(outbuf,
755 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE);
756 efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf,
757 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT);
758 efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf,
759 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN);
761 if (out_len >= MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN) {
762 efx->ptp_data->ts_corrections.general_tx = MCDI_DWORD(
764 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_TX);
765 efx->ptp_data->ts_corrections.general_rx = MCDI_DWORD(
767 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_RX);
769 efx->ptp_data->ts_corrections.general_tx =
770 efx->ptp_data->ts_corrections.ptp_tx;
771 efx->ptp_data->ts_corrections.general_rx =
772 efx->ptp_data->ts_corrections.ptp_rx;
774 } else if (rc == -EINVAL) {
775 efx->ptp_data->ts_corrections.ptp_tx = 0;
776 efx->ptp_data->ts_corrections.ptp_rx = 0;
777 efx->ptp_data->ts_corrections.pps_out = 0;
778 efx->ptp_data->ts_corrections.pps_in = 0;
779 efx->ptp_data->ts_corrections.general_tx = 0;
780 efx->ptp_data->ts_corrections.general_rx = 0;
782 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), outbuf,
790 /* Enable MCDI PTP support. */
791 static int efx_ptp_enable(struct efx_nic *efx)
793 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
794 MCDI_DECLARE_BUF_ERR(outbuf);
797 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE);
798 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
799 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE,
800 efx->ptp_data->channel ?
801 efx->ptp_data->channel->channel : 0);
802 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode);
804 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
805 outbuf, sizeof(outbuf), NULL);
806 rc = (rc == -EALREADY) ? 0 : rc;
808 efx_mcdi_display_error(efx, MC_CMD_PTP,
809 MC_CMD_PTP_IN_ENABLE_LEN,
810 outbuf, sizeof(outbuf), rc);
814 /* Disable MCDI PTP support.
816 * Note that this function should never rely on the presence of ptp_data -
817 * may be called before that exists.
819 static int efx_ptp_disable(struct efx_nic *efx)
821 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
822 MCDI_DECLARE_BUF_ERR(outbuf);
825 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE);
826 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
827 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
828 outbuf, sizeof(outbuf), NULL);
829 rc = (rc == -EALREADY) ? 0 : rc;
830 /* If we get ENOSYS, the NIC doesn't support PTP, and thus this function
831 * should only have been called during probe.
833 if (rc == -ENOSYS || rc == -EPERM)
834 pci_info(efx->pci_dev, "no PTP support\n");
836 efx_mcdi_display_error(efx, MC_CMD_PTP,
837 MC_CMD_PTP_IN_DISABLE_LEN,
838 outbuf, sizeof(outbuf), rc);
842 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
846 while ((skb = skb_dequeue(q))) {
848 netif_receive_skb(skb);
853 static void efx_ptp_handle_no_channel(struct efx_nic *efx)
855 netif_err(efx, drv, efx->net_dev,
856 "ERROR: PTP requires MSI-X and 1 additional interrupt"
857 "vector. PTP disabled\n");
860 /* Repeatedly send the host time to the MC which will capture the hardware
863 static void efx_ptp_send_times(struct efx_nic *efx,
864 struct pps_event_time *last_time)
866 struct pps_event_time now;
867 struct timespec64 limit;
868 struct efx_ptp_data *ptp = efx->ptp_data;
869 int *mc_running = ptp->start.addr;
873 timespec64_add_ns(&limit, SYNCHRONISE_PERIOD_NS);
875 /* Write host time for specified period or until MC is done */
876 while ((timespec64_compare(&now.ts_real, &limit) < 0) &&
877 READ_ONCE(*mc_running)) {
878 struct timespec64 update_time;
879 unsigned int host_time;
881 /* Don't update continuously to avoid saturating the PCIe bus */
882 update_time = now.ts_real;
883 timespec64_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS);
886 } while ((timespec64_compare(&now.ts_real, &update_time) < 0) &&
887 READ_ONCE(*mc_running));
889 /* Synchronise NIC with single word of time only */
890 host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS |
891 now.ts_real.tv_nsec);
892 /* Update host time in NIC memory */
893 efx->type->ptp_write_host_time(efx, host_time);
898 /* Read a timeset from the MC's results and partial process. */
899 static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data),
900 struct efx_ptp_timeset *timeset)
902 unsigned start_ns, end_ns;
904 timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART);
905 timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR);
906 timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR);
907 timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND),
908 timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
911 start_ns = timeset->host_start & MC_NANOSECOND_MASK;
912 end_ns = timeset->host_end & MC_NANOSECOND_MASK;
913 /* Allow for rollover */
914 if (end_ns < start_ns)
915 end_ns += NSEC_PER_SEC;
916 /* Determine duration of operation */
917 timeset->window = end_ns - start_ns;
920 /* Process times received from MC.
922 * Extract times from returned results, and establish the minimum value
923 * seen. The minimum value represents the "best" possible time and events
924 * too much greater than this are rejected - the machine is, perhaps, too
925 * busy. A number of readings are taken so that, hopefully, at least one good
926 * synchronisation will be seen in the results.
929 efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf),
930 size_t response_length,
931 const struct pps_event_time *last_time)
933 unsigned number_readings =
934 MCDI_VAR_ARRAY_LEN(response_length,
935 PTP_OUT_SYNCHRONIZE_TIMESET);
938 unsigned last_good = 0;
939 struct efx_ptp_data *ptp = efx->ptp_data;
942 struct timespec64 delta;
945 if (number_readings == 0)
948 /* Read the set of results and find the last good host-MC
949 * synchronization result. The MC times when it finishes reading the
950 * host time so the corrected window time should be fairly constant
951 * for a given platform. Increment stats for any results that appear
954 for (i = 0; i < number_readings; i++) {
955 s32 window, corrected;
956 struct timespec64 wait;
958 efx_ptp_read_timeset(
959 MCDI_ARRAY_STRUCT_PTR(synch_buf,
960 PTP_OUT_SYNCHRONIZE_TIMESET, i),
963 wait = ktime_to_timespec64(
964 ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0));
965 window = ptp->timeset[i].window;
966 corrected = window - wait.tv_nsec;
968 /* We expect the uncorrected synchronization window to be at
969 * least as large as the interval between host start and end
970 * times. If it is smaller than this then this is mostly likely
971 * to be a consequence of the host's time being adjusted.
972 * Check that the corrected sync window is in a reasonable
973 * range. If it is out of range it is likely to be because an
974 * interrupt or other delay occurred between reading the system
975 * time and writing it to MC memory.
977 if (window < SYNCHRONISATION_GRANULARITY_NS) {
978 ++ptp->invalid_sync_windows;
979 } else if (corrected >= MAX_SYNCHRONISATION_NS) {
980 ++ptp->oversize_sync_windows;
981 } else if (corrected < ptp->min_synchronisation_ns) {
982 ++ptp->undersize_sync_windows;
990 netif_warn(efx, drv, efx->net_dev,
991 "PTP no suitable synchronisations\n");
995 /* Calculate delay from last good sync (host time) to last_time.
996 * It is possible that the seconds rolled over between taking
997 * the start reading and the last value written by the host. The
998 * timescales are such that a gap of more than one second is never
999 * expected. delta is *not* normalised.
1001 start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS;
1002 last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK;
1003 if (start_sec != last_sec &&
1004 ((start_sec + 1) & MC_SECOND_MASK) != last_sec) {
1005 netif_warn(efx, hw, efx->net_dev,
1006 "PTP bad synchronisation seconds\n");
1009 delta.tv_sec = (last_sec - start_sec) & 1;
1011 last_time->ts_real.tv_nsec -
1012 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
1014 /* Convert the NIC time at last good sync into kernel time.
1015 * No correction is required - this time is the output of a
1018 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
1019 ptp->timeset[last_good].minor, 0);
1021 /* Calculate delay from NIC top of second to last_time */
1022 delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec;
1024 /* Set PPS timestamp to match NIC top of second */
1025 ptp->host_time_pps = *last_time;
1026 pps_sub_ts(&ptp->host_time_pps, delta);
1031 /* Synchronize times between the host and the MC */
1032 static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
1034 struct efx_ptp_data *ptp = efx->ptp_data;
1035 MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX);
1036 size_t response_length;
1038 unsigned long timeout;
1039 struct pps_event_time last_time = {};
1040 unsigned int loops = 0;
1041 int *start = ptp->start.addr;
1043 MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE);
1044 MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0);
1045 MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS,
1047 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
1048 ptp->start.dma_addr);
1050 /* Clear flag that signals MC ready */
1051 WRITE_ONCE(*start, 0);
1052 rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf,
1053 MC_CMD_PTP_IN_SYNCHRONIZE_LEN);
1054 EFX_WARN_ON_ONCE_PARANOID(rc);
1056 /* Wait for start from MCDI (or timeout) */
1057 timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS);
1058 while (!READ_ONCE(*start) && (time_before(jiffies, timeout))) {
1059 udelay(20); /* Usually start MCDI execution quickly */
1065 if (!time_before(jiffies, timeout))
1066 ++ptp->sync_timeouts;
1068 if (READ_ONCE(*start))
1069 efx_ptp_send_times(efx, &last_time);
1071 /* Collect results */
1072 rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP,
1073 MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
1074 synch_buf, sizeof(synch_buf),
1077 rc = efx_ptp_process_times(efx, synch_buf, response_length,
1082 ++ptp->no_time_syncs;
1085 /* Increment the bad syncs counter if the synchronize fails, whatever
1094 /* Transmit a PTP packet via the dedicated hardware timestamped queue. */
1095 static void efx_ptp_xmit_skb_queue(struct efx_nic *efx, struct sk_buff *skb)
1097 struct efx_ptp_data *ptp_data = efx->ptp_data;
1098 u8 type = efx_tx_csum_type_skb(skb);
1099 struct efx_tx_queue *tx_queue;
1101 tx_queue = efx_channel_get_tx_queue(ptp_data->channel, type);
1102 if (tx_queue && tx_queue->timestamping) {
1103 /* This code invokes normal driver TX code which is always
1104 * protected from softirqs when called from generic TX code,
1105 * which in turn disables preemption. Look at __dev_queue_xmit
1106 * which uses rcu_read_lock_bh disabling preemption for RCU
1107 * plus disabling softirqs. We do not need RCU reader
1110 * Although it is theoretically safe for current PTP TX/RX code
1111 * running without disabling softirqs, there are three good
1112 * reasond for doing so:
1114 * 1) The code invoked is mainly implemented for non-PTP
1115 * packets and it is always executed with softirqs
1117 * 2) This being a single PTP packet, better to not
1118 * interrupt its processing by softirqs which can lead
1119 * to high latencies.
1120 * 3) netdev_xmit_more checks preemption is disabled and
1121 * triggers a BUG_ON if not.
1124 efx_enqueue_skb(tx_queue, skb);
1127 WARN_ONCE(1, "PTP channel has no timestamped tx queue\n");
1128 dev_kfree_skb_any(skb);
1132 /* Transmit a PTP packet, via the MCDI interface, to the wire. */
1133 static void efx_ptp_xmit_skb_mc(struct efx_nic *efx, struct sk_buff *skb)
1135 struct efx_ptp_data *ptp_data = efx->ptp_data;
1136 struct skb_shared_hwtstamps timestamps;
1138 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
1141 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT);
1142 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0);
1143 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len);
1144 if (skb_shinfo(skb)->nr_frags != 0) {
1145 rc = skb_linearize(skb);
1150 if (skb->ip_summed == CHECKSUM_PARTIAL) {
1151 rc = skb_checksum_help(skb);
1155 skb_copy_from_linear_data(skb,
1156 MCDI_PTR(ptp_data->txbuf,
1157 PTP_IN_TRANSMIT_PACKET),
1159 rc = efx_mcdi_rpc(efx, MC_CMD_PTP,
1160 ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len),
1161 txtime, sizeof(txtime), &len);
1165 memset(×tamps, 0, sizeof(timestamps));
1166 timestamps.hwtstamp = ptp_data->nic_to_kernel_time(
1167 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR),
1168 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR),
1169 ptp_data->ts_corrections.ptp_tx);
1171 skb_tstamp_tx(skb, ×tamps);
1176 dev_kfree_skb_any(skb);
1181 static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
1183 struct efx_ptp_data *ptp = efx->ptp_data;
1184 struct list_head *cursor;
1185 struct list_head *next;
1187 if (ptp->rx_ts_inline)
1190 /* Drop time-expired events */
1191 spin_lock_bh(&ptp->evt_lock);
1192 list_for_each_safe(cursor, next, &ptp->evt_list) {
1193 struct efx_ptp_event_rx *evt;
1195 evt = list_entry(cursor, struct efx_ptp_event_rx,
1197 if (time_after(jiffies, evt->expiry)) {
1198 list_move(&evt->link, &ptp->evt_free_list);
1199 netif_warn(efx, hw, efx->net_dev,
1200 "PTP rx event dropped\n");
1203 spin_unlock_bh(&ptp->evt_lock);
1206 static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
1207 struct sk_buff *skb)
1209 struct efx_ptp_data *ptp = efx->ptp_data;
1211 struct list_head *cursor;
1212 struct list_head *next;
1213 struct efx_ptp_match *match;
1214 enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED;
1216 WARN_ON_ONCE(ptp->rx_ts_inline);
1218 spin_lock_bh(&ptp->evt_lock);
1219 evts_waiting = !list_empty(&ptp->evt_list);
1220 spin_unlock_bh(&ptp->evt_lock);
1223 return PTP_PACKET_STATE_UNMATCHED;
1225 match = (struct efx_ptp_match *)skb->cb;
1226 /* Look for a matching timestamp in the event queue */
1227 spin_lock_bh(&ptp->evt_lock);
1228 list_for_each_safe(cursor, next, &ptp->evt_list) {
1229 struct efx_ptp_event_rx *evt;
1231 evt = list_entry(cursor, struct efx_ptp_event_rx, link);
1232 if ((evt->seq0 == match->words[0]) &&
1233 (evt->seq1 == match->words[1])) {
1234 struct skb_shared_hwtstamps *timestamps;
1236 /* Match - add in hardware timestamp */
1237 timestamps = skb_hwtstamps(skb);
1238 timestamps->hwtstamp = evt->hwtimestamp;
1240 match->state = PTP_PACKET_STATE_MATCHED;
1241 rc = PTP_PACKET_STATE_MATCHED;
1242 list_move(&evt->link, &ptp->evt_free_list);
1246 spin_unlock_bh(&ptp->evt_lock);
1251 /* Process any queued receive events and corresponding packets
1253 * q is returned with all the packets that are ready for delivery.
1255 static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
1257 struct efx_ptp_data *ptp = efx->ptp_data;
1258 struct sk_buff *skb;
1260 while ((skb = skb_dequeue(&ptp->rxq))) {
1261 struct efx_ptp_match *match;
1263 match = (struct efx_ptp_match *)skb->cb;
1264 if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) {
1265 __skb_queue_tail(q, skb);
1266 } else if (efx_ptp_match_rx(efx, skb) ==
1267 PTP_PACKET_STATE_MATCHED) {
1268 __skb_queue_tail(q, skb);
1269 } else if (time_after(jiffies, match->expiry)) {
1270 match->state = PTP_PACKET_STATE_TIMED_OUT;
1271 ++ptp->rx_no_timestamp;
1272 __skb_queue_tail(q, skb);
1274 /* Replace unprocessed entry and stop */
1275 skb_queue_head(&ptp->rxq, skb);
1281 /* Complete processing of a received packet */
1282 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1285 netif_receive_skb(skb);
1289 static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
1291 struct efx_ptp_data *ptp = efx->ptp_data;
1293 if (ptp->rxfilter_installed) {
1294 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1295 ptp->rxfilter_general);
1296 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1297 ptp->rxfilter_event);
1298 ptp->rxfilter_installed = false;
1302 static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1304 struct efx_ptp_data *ptp = efx->ptp_data;
1305 struct efx_filter_spec rxfilter;
1308 if (!ptp->channel || ptp->rxfilter_installed)
1311 /* Must filter on both event and general ports to ensure
1312 * that there is no packet re-ordering.
1314 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1316 efx_channel_get_rx_queue(ptp->channel)));
1317 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1319 htons(PTP_EVENT_PORT));
1323 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1326 ptp->rxfilter_event = rc;
1328 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1330 efx_channel_get_rx_queue(ptp->channel)));
1331 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1333 htons(PTP_GENERAL_PORT));
1337 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1340 ptp->rxfilter_general = rc;
1342 ptp->rxfilter_installed = true;
1346 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1347 ptp->rxfilter_event);
1351 static int efx_ptp_start(struct efx_nic *efx)
1353 struct efx_ptp_data *ptp = efx->ptp_data;
1356 ptp->reset_required = false;
1358 rc = efx_ptp_insert_multicast_filters(efx);
1362 rc = efx_ptp_enable(efx);
1366 ptp->evt_frag_idx = 0;
1367 ptp->current_adjfreq = 0;
1372 efx_ptp_remove_multicast_filters(efx);
1376 static int efx_ptp_stop(struct efx_nic *efx)
1378 struct efx_ptp_data *ptp = efx->ptp_data;
1379 struct list_head *cursor;
1380 struct list_head *next;
1386 rc = efx_ptp_disable(efx);
1388 efx_ptp_remove_multicast_filters(efx);
1390 /* Make sure RX packets are really delivered */
1391 efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq);
1392 skb_queue_purge(&efx->ptp_data->txq);
1394 /* Drop any pending receive events */
1395 spin_lock_bh(&efx->ptp_data->evt_lock);
1396 list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) {
1397 list_move(cursor, &efx->ptp_data->evt_free_list);
1399 spin_unlock_bh(&efx->ptp_data->evt_lock);
1404 static int efx_ptp_restart(struct efx_nic *efx)
1406 if (efx->ptp_data && efx->ptp_data->enabled)
1407 return efx_ptp_start(efx);
1411 static void efx_ptp_pps_worker(struct work_struct *work)
1413 struct efx_ptp_data *ptp =
1414 container_of(work, struct efx_ptp_data, pps_work);
1415 struct efx_nic *efx = ptp->efx;
1416 struct ptp_clock_event ptp_evt;
1418 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
1421 ptp_evt.type = PTP_CLOCK_PPSUSR;
1422 ptp_evt.pps_times = ptp->host_time_pps;
1423 ptp_clock_event(ptp->phc_clock, &ptp_evt);
1426 static void efx_ptp_worker(struct work_struct *work)
1428 struct efx_ptp_data *ptp_data =
1429 container_of(work, struct efx_ptp_data, work);
1430 struct efx_nic *efx = ptp_data->efx;
1431 struct sk_buff *skb;
1432 struct sk_buff_head tempq;
1434 if (ptp_data->reset_required) {
1440 efx_ptp_drop_time_expired_events(efx);
1442 __skb_queue_head_init(&tempq);
1443 efx_ptp_process_events(efx, &tempq);
1445 while ((skb = skb_dequeue(&ptp_data->txq)))
1446 ptp_data->xmit_skb(efx, skb);
1448 while ((skb = __skb_dequeue(&tempq)))
1449 efx_ptp_process_rx(efx, skb);
1452 static const struct ptp_clock_info efx_phc_clock_info = {
1453 .owner = THIS_MODULE,
1461 .adjfreq = efx_phc_adjfreq,
1462 .adjtime = efx_phc_adjtime,
1463 .gettime64 = efx_phc_gettime,
1464 .settime64 = efx_phc_settime,
1465 .enable = efx_phc_enable,
1468 /* Initialise PTP state. */
1469 int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1471 struct efx_ptp_data *ptp;
1475 if (efx->ptp_data) {
1476 efx->ptp_data->channel = channel;
1480 ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL);
1481 efx->ptp_data = ptp;
1486 ptp->channel = channel;
1487 ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1489 rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL);
1493 skb_queue_head_init(&ptp->rxq);
1494 skb_queue_head_init(&ptp->txq);
1495 ptp->workwq = create_singlethread_workqueue("sfc_ptp");
1501 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1502 ptp->xmit_skb = efx_ptp_xmit_skb_queue;
1503 /* Request sync events on this channel. */
1504 channel->sync_events_state = SYNC_EVENTS_QUIESCENT;
1506 ptp->xmit_skb = efx_ptp_xmit_skb_mc;
1509 INIT_WORK(&ptp->work, efx_ptp_worker);
1510 ptp->config.flags = 0;
1511 ptp->config.tx_type = HWTSTAMP_TX_OFF;
1512 ptp->config.rx_filter = HWTSTAMP_FILTER_NONE;
1513 INIT_LIST_HEAD(&ptp->evt_list);
1514 INIT_LIST_HEAD(&ptp->evt_free_list);
1515 spin_lock_init(&ptp->evt_lock);
1516 for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++)
1517 list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list);
1519 /* Get the NIC PTP attributes and set up time conversions */
1520 rc = efx_ptp_get_attributes(efx);
1524 /* Get the timestamp corrections */
1525 rc = efx_ptp_get_timestamp_corrections(efx);
1529 if (efx->mcdi->fn_flags &
1530 (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) {
1531 ptp->phc_clock_info = efx_phc_clock_info;
1532 ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info,
1533 &efx->pci_dev->dev);
1534 if (IS_ERR(ptp->phc_clock)) {
1535 rc = PTR_ERR(ptp->phc_clock);
1537 } else if (ptp->phc_clock) {
1538 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1539 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
1540 if (!ptp->pps_workwq) {
1546 ptp->nic_ts_enabled = false;
1550 ptp_clock_unregister(efx->ptp_data->phc_clock);
1553 destroy_workqueue(efx->ptp_data->workwq);
1556 efx_nic_free_buffer(efx, &ptp->start);
1559 kfree(efx->ptp_data);
1560 efx->ptp_data = NULL;
1565 /* Initialise PTP channel.
1567 * Setting core_index to zero causes the queue to be initialised and doesn't
1568 * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
1570 static int efx_ptp_probe_channel(struct efx_channel *channel)
1572 struct efx_nic *efx = channel->efx;
1575 channel->irq_moderation_us = 0;
1576 channel->rx_queue.core_index = 0;
1578 rc = efx_ptp_probe(efx, channel);
1579 /* Failure to probe PTP is not fatal; this channel will just not be
1580 * used for anything.
1581 * In the case of EPERM, efx_ptp_probe will print its own message (in
1582 * efx_ptp_get_attributes()), so we don't need to.
1584 if (rc && rc != -EPERM)
1585 netif_warn(efx, drv, efx->net_dev,
1586 "Failed to probe PTP, rc=%d\n", rc);
1590 void efx_ptp_remove(struct efx_nic *efx)
1595 (void)efx_ptp_disable(efx);
1597 cancel_work_sync(&efx->ptp_data->work);
1598 if (efx->ptp_data->pps_workwq)
1599 cancel_work_sync(&efx->ptp_data->pps_work);
1601 skb_queue_purge(&efx->ptp_data->rxq);
1602 skb_queue_purge(&efx->ptp_data->txq);
1604 if (efx->ptp_data->phc_clock) {
1605 destroy_workqueue(efx->ptp_data->pps_workwq);
1606 ptp_clock_unregister(efx->ptp_data->phc_clock);
1609 destroy_workqueue(efx->ptp_data->workwq);
1611 efx_nic_free_buffer(efx, &efx->ptp_data->start);
1612 kfree(efx->ptp_data);
1613 efx->ptp_data = NULL;
1616 static void efx_ptp_remove_channel(struct efx_channel *channel)
1618 efx_ptp_remove(channel->efx);
1621 static void efx_ptp_get_channel_name(struct efx_channel *channel,
1622 char *buf, size_t len)
1624 snprintf(buf, len, "%s-ptp", channel->efx->name);
1627 /* Determine whether this packet should be processed by the PTP module
1628 * or transmitted conventionally.
1630 bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1632 return efx->ptp_data &&
1633 efx->ptp_data->enabled &&
1634 skb->len >= PTP_MIN_LENGTH &&
1635 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1636 likely(skb->protocol == htons(ETH_P_IP)) &&
1637 skb_transport_header_was_set(skb) &&
1638 skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1639 ip_hdr(skb)->protocol == IPPROTO_UDP &&
1641 skb_transport_offset(skb) + sizeof(struct udphdr) &&
1642 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1645 /* Receive a PTP packet. Packets are queued until the arrival of
1646 * the receive timestamp from the MC - this will probably occur after the
1647 * packet arrival because of the processing in the MC.
1649 static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1651 struct efx_nic *efx = channel->efx;
1652 struct efx_ptp_data *ptp = efx->ptp_data;
1653 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1654 u8 *match_data_012, *match_data_345;
1655 unsigned int version;
1658 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1660 /* Correct version? */
1661 if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1662 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1666 version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1667 if (version != PTP_VERSION_V1) {
1671 /* PTP V1 uses all six bytes of the UUID to match the packet
1674 match_data_012 = data + PTP_V1_UUID_OFFSET;
1675 match_data_345 = data + PTP_V1_UUID_OFFSET + 3;
1677 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1681 version = data[PTP_V2_VERSION_OFFSET];
1682 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1686 /* The original V2 implementation uses bytes 2-7 of
1687 * the UUID to match the packet to the timestamp. This
1688 * discards two of the bytes of the MAC address used
1689 * to create the UUID (SF bug 33070). The PTP V2
1690 * enhanced mode fixes this issue and uses bytes 0-2
1691 * and byte 5-7 of the UUID.
1693 match_data_345 = data + PTP_V2_UUID_OFFSET + 5;
1694 if (ptp->mode == MC_CMD_PTP_MODE_V2) {
1695 match_data_012 = data + PTP_V2_UUID_OFFSET + 2;
1697 match_data_012 = data + PTP_V2_UUID_OFFSET + 0;
1698 BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED);
1702 /* Does this packet require timestamping? */
1703 if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1704 match->state = PTP_PACKET_STATE_UNMATCHED;
1706 /* We expect the sequence number to be in the same position in
1707 * the packet for PTP V1 and V2
1709 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1710 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1712 /* Extract UUID/Sequence information */
1713 match->words[0] = (match_data_012[0] |
1714 (match_data_012[1] << 8) |
1715 (match_data_012[2] << 16) |
1716 (match_data_345[0] << 24));
1717 match->words[1] = (match_data_345[1] |
1718 (match_data_345[2] << 8) |
1719 (data[PTP_V1_SEQUENCE_OFFSET +
1720 PTP_V1_SEQUENCE_LENGTH - 1] <<
1723 match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1726 skb_queue_tail(&ptp->rxq, skb);
1727 queue_work(ptp->workwq, &ptp->work);
1732 /* Transmit a PTP packet. This has to be transmitted by the MC
1733 * itself, through an MCDI call. MCDI calls aren't permitted
1734 * in the transmit path so defer the actual transmission to a suitable worker.
1736 int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1738 struct efx_ptp_data *ptp = efx->ptp_data;
1740 skb_queue_tail(&ptp->txq, skb);
1742 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1743 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1744 efx_xmit_hwtstamp_pending(skb);
1745 queue_work(ptp->workwq, &ptp->work);
1747 return NETDEV_TX_OK;
1750 int efx_ptp_get_mode(struct efx_nic *efx)
1752 return efx->ptp_data->mode;
1755 int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1756 unsigned int new_mode)
1758 if ((enable_wanted != efx->ptp_data->enabled) ||
1759 (enable_wanted && (efx->ptp_data->mode != new_mode))) {
1762 if (enable_wanted) {
1763 /* Change of mode requires disable */
1764 if (efx->ptp_data->enabled &&
1765 (efx->ptp_data->mode != new_mode)) {
1766 efx->ptp_data->enabled = false;
1767 rc = efx_ptp_stop(efx);
1772 /* Set new operating mode and establish
1773 * baseline synchronisation, which must
1776 efx->ptp_data->mode = new_mode;
1777 if (netif_running(efx->net_dev))
1778 rc = efx_ptp_start(efx);
1780 rc = efx_ptp_synchronize(efx,
1781 PTP_SYNC_ATTEMPTS * 2);
1786 rc = efx_ptp_stop(efx);
1792 efx->ptp_data->enabled = enable_wanted;
1798 static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
1802 if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1803 (init->tx_type != HWTSTAMP_TX_ON))
1806 rc = efx->type->ptp_set_ts_config(efx, init);
1810 efx->ptp_data->config = *init;
1814 void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
1816 struct efx_ptp_data *ptp = efx->ptp_data;
1817 struct efx_nic *primary = efx->primary;
1824 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1825 SOF_TIMESTAMPING_RX_HARDWARE |
1826 SOF_TIMESTAMPING_RAW_HARDWARE);
1827 /* Check licensed features. If we don't have the license for TX
1828 * timestamps, the NIC will not support them.
1830 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1831 struct efx_ef10_nic_data *nic_data = efx->nic_data;
1833 if (!(nic_data->licensed_features &
1834 (1 << LICENSED_V3_FEATURES_TX_TIMESTAMPS_LBN)))
1835 ts_info->so_timestamping &=
1836 ~SOF_TIMESTAMPING_TX_HARDWARE;
1838 if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1839 ts_info->phc_index =
1840 ptp_clock_index(primary->ptp_data->phc_clock);
1841 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1842 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1845 int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1847 struct hwtstamp_config config;
1850 /* Not a PTP enabled port */
1854 if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
1857 rc = efx_ptp_ts_init(efx, &config);
1861 return copy_to_user(ifr->ifr_data, &config, sizeof(config))
1865 int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1870 return copy_to_user(ifr->ifr_data, &efx->ptp_data->config,
1871 sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
1874 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1876 struct efx_ptp_data *ptp = efx->ptp_data;
1878 netif_err(efx, hw, efx->net_dev,
1879 "PTP unexpected event length: got %d expected %d\n",
1880 ptp->evt_frag_idx, expected_frag_len);
1881 ptp->reset_required = true;
1882 queue_work(ptp->workwq, &ptp->work);
1885 /* Process a completed receive event. Put it on the event queue and
1886 * start worker thread. This is required because event and their
1887 * correspoding packets may come in either order.
1889 static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
1891 struct efx_ptp_event_rx *evt = NULL;
1893 if (WARN_ON_ONCE(ptp->rx_ts_inline))
1896 if (ptp->evt_frag_idx != 3) {
1897 ptp_event_failure(efx, 3);
1901 spin_lock_bh(&ptp->evt_lock);
1902 if (!list_empty(&ptp->evt_free_list)) {
1903 evt = list_first_entry(&ptp->evt_free_list,
1904 struct efx_ptp_event_rx, link);
1905 list_del(&evt->link);
1907 evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA);
1908 evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2],
1910 (EFX_QWORD_FIELD(ptp->evt_frags[1],
1911 MCDI_EVENT_SRC) << 8) |
1912 (EFX_QWORD_FIELD(ptp->evt_frags[0],
1913 MCDI_EVENT_SRC) << 16));
1914 evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
1915 EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
1916 EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
1917 ptp->ts_corrections.ptp_rx);
1918 evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1919 list_add_tail(&evt->link, &ptp->evt_list);
1921 queue_work(ptp->workwq, &ptp->work);
1922 } else if (net_ratelimit()) {
1923 /* Log a rate-limited warning message. */
1924 netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n");
1926 spin_unlock_bh(&ptp->evt_lock);
1929 static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1931 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1932 if (ptp->evt_frag_idx != 1) {
1933 ptp_event_failure(efx, 1);
1937 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1940 static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1942 if (ptp->nic_ts_enabled)
1943 queue_work(ptp->pps_workwq, &ptp->pps_work);
1946 void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1948 struct efx_ptp_data *ptp = efx->ptp_data;
1949 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
1952 if (!efx->ptp_warned) {
1953 netif_warn(efx, drv, efx->net_dev,
1954 "Received PTP event but PTP not set up\n");
1955 efx->ptp_warned = true;
1963 if (ptp->evt_frag_idx == 0) {
1964 ptp->evt_code = code;
1965 } else if (ptp->evt_code != code) {
1966 netif_err(efx, hw, efx->net_dev,
1967 "PTP out of sequence event %d\n", code);
1968 ptp->evt_frag_idx = 0;
1971 ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
1972 if (!MCDI_EVENT_FIELD(*ev, CONT)) {
1973 /* Process resulting event */
1975 case MCDI_EVENT_CODE_PTP_RX:
1976 ptp_event_rx(efx, ptp);
1978 case MCDI_EVENT_CODE_PTP_FAULT:
1979 ptp_event_fault(efx, ptp);
1981 case MCDI_EVENT_CODE_PTP_PPS:
1982 ptp_event_pps(efx, ptp);
1985 netif_err(efx, hw, efx->net_dev,
1986 "PTP unknown event %d\n", code);
1989 ptp->evt_frag_idx = 0;
1990 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
1991 netif_err(efx, hw, efx->net_dev,
1992 "PTP too many event fragments\n");
1993 ptp->evt_frag_idx = 0;
1997 void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
1999 struct efx_nic *efx = channel->efx;
2000 struct efx_ptp_data *ptp = efx->ptp_data;
2002 /* When extracting the sync timestamp minor value, we should discard
2003 * the least significant two bits. These are not required in order
2004 * to reconstruct full-range timestamps and they are optionally used
2005 * to report status depending on the options supplied when subscribing
2008 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
2009 channel->sync_timestamp_minor =
2010 (MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_MS_8BITS) & 0xFC)
2011 << ptp->nic_time.sync_event_minor_shift;
2013 /* if sync events have been disabled then we want to silently ignore
2014 * this event, so throw away result.
2016 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
2020 static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
2022 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
2023 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset));
2025 const u8 *data = eh + efx->rx_packet_ts_offset;
2026 return (u32)data[0] |
2028 (u32)data[2] << 16 |
2033 void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
2034 struct sk_buff *skb)
2036 struct efx_nic *efx = channel->efx;
2037 struct efx_ptp_data *ptp = efx->ptp_data;
2038 u32 pkt_timestamp_major, pkt_timestamp_minor;
2040 struct skb_shared_hwtstamps *timestamps;
2042 if (channel->sync_events_state != SYNC_EVENTS_VALID)
2045 pkt_timestamp_minor = efx_rx_buf_timestamp_minor(efx, skb_mac_header(skb));
2047 /* get the difference between the packet and sync timestamps,
2050 diff = pkt_timestamp_minor - channel->sync_timestamp_minor;
2051 if (pkt_timestamp_minor < channel->sync_timestamp_minor)
2052 diff += ptp->nic_time.minor_max;
2054 /* do we roll over a second boundary and need to carry the one? */
2055 carry = (channel->sync_timestamp_minor >= ptp->nic_time.minor_max - diff) ?
2058 if (diff <= ptp->nic_time.sync_event_diff_max) {
2059 /* packet is ahead of the sync event by a quarter of a second or
2060 * less (allowing for fuzz)
2062 pkt_timestamp_major = channel->sync_timestamp_major + carry;
2063 } else if (diff >= ptp->nic_time.sync_event_diff_min) {
2064 /* packet is behind the sync event but within the fuzz factor.
2065 * This means the RX packet and sync event crossed as they were
2066 * placed on the event queue, which can sometimes happen.
2068 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
2070 /* it's outside tolerance in both directions. this might be
2071 * indicative of us missing sync events for some reason, so
2072 * we'll call it an error rather than risk giving a bogus
2075 netif_vdbg(efx, drv, efx->net_dev,
2076 "packet timestamp %x too far from sync event %x:%x\n",
2077 pkt_timestamp_minor, channel->sync_timestamp_major,
2078 channel->sync_timestamp_minor);
2082 /* attach the timestamps to the skb */
2083 timestamps = skb_hwtstamps(skb);
2084 timestamps->hwtstamp =
2085 ptp->nic_to_kernel_time(pkt_timestamp_major,
2086 pkt_timestamp_minor,
2087 ptp->ts_corrections.general_rx);
2090 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
2092 struct efx_ptp_data *ptp_data = container_of(ptp,
2093 struct efx_ptp_data,
2095 struct efx_nic *efx = ptp_data->efx;
2096 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
2100 if (delta > MAX_PPB)
2102 else if (delta < -MAX_PPB)
2105 /* Convert ppb to fixed point ns taking care to round correctly. */
2106 adjustment_ns = ((s64)delta * PPB_SCALE_WORD +
2107 (1 << (ptp_data->adjfreq_ppb_shift - 1))) >>
2108 ptp_data->adjfreq_ppb_shift;
2110 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2111 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
2112 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
2113 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
2114 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
2115 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj),
2120 ptp_data->current_adjfreq = adjustment_ns;
2124 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
2126 u32 nic_major, nic_minor;
2127 struct efx_ptp_data *ptp_data = container_of(ptp,
2128 struct efx_ptp_data,
2130 struct efx_nic *efx = ptp_data->efx;
2131 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
2133 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
2135 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2136 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2137 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
2138 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
2139 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
2140 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2144 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
2146 struct efx_ptp_data *ptp_data = container_of(ptp,
2147 struct efx_ptp_data,
2149 struct efx_nic *efx = ptp_data->efx;
2150 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
2151 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
2155 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
2156 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2158 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2159 outbuf, sizeof(outbuf), NULL);
2163 kt = ptp_data->nic_to_kernel_time(
2164 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
2165 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
2166 *ts = ktime_to_timespec64(kt);
2170 static int efx_phc_settime(struct ptp_clock_info *ptp,
2171 const struct timespec64 *e_ts)
2173 /* Get the current NIC time, efx_phc_gettime.
2174 * Subtract from the desired time to get the offset
2175 * call efx_phc_adjtime with the offset
2178 struct timespec64 time_now;
2179 struct timespec64 delta;
2181 rc = efx_phc_gettime(ptp, &time_now);
2185 delta = timespec64_sub(*e_ts, time_now);
2187 rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta));
2194 static int efx_phc_enable(struct ptp_clock_info *ptp,
2195 struct ptp_clock_request *request,
2198 struct efx_ptp_data *ptp_data = container_of(ptp,
2199 struct efx_ptp_data,
2201 if (request->type != PTP_CLK_REQ_PPS)
2204 ptp_data->nic_ts_enabled = !!enable;
2208 static const struct efx_channel_type efx_ptp_channel_type = {
2209 .handle_no_channel = efx_ptp_handle_no_channel,
2210 .pre_probe = efx_ptp_probe_channel,
2211 .post_remove = efx_ptp_remove_channel,
2212 .get_name = efx_ptp_get_channel_name,
2213 .copy = efx_copy_channel,
2214 .receive_skb = efx_ptp_rx,
2215 .want_txqs = efx_ptp_want_txqs,
2216 .keep_eventq = false,
2219 void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
2221 /* Check whether PTP is implemented on this NIC. The DISABLE
2222 * operation will succeed if and only if it is implemented.
2224 if (efx_ptp_disable(efx) == 0)
2225 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
2226 &efx_ptp_channel_type;
2229 void efx_ptp_start_datapath(struct efx_nic *efx)
2231 if (efx_ptp_restart(efx))
2232 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
2233 /* re-enable timestamping if it was previously enabled */
2234 if (efx->type->ptp_set_ts_sync_events)
2235 efx->type->ptp_set_ts_sync_events(efx, true, true);
2238 void efx_ptp_stop_datapath(struct efx_nic *efx)
2240 /* temporarily disable timestamping */
2241 if (efx->type->ptp_set_ts_sync_events)
2242 efx->type->ptp_set_ts_sync_events(efx, false, true);