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_RXFILTERS_LEN 5
123 #define PTP_ADDR_IPV4 0xe0000181 /* 224.0.1.129 */
124 #define PTP_ADDR_IPV6 {0xff, 0x0e, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, \
125 0, 0x01, 0x81} /* ff0e::181 */
126 #define PTP_EVENT_PORT 319
127 #define PTP_GENERAL_PORT 320
128 #define PTP_ADDR_ETHER {0x01, 0x1b, 0x19, 0, 0, 0} /* 01-1B-19-00-00-00 */
130 /* Annoyingly the format of the version numbers are different between
131 * versions 1 and 2 so it isn't possible to simply look for 1 or 2.
133 #define PTP_VERSION_V1 1
135 #define PTP_VERSION_V2 2
136 #define PTP_VERSION_V2_MASK 0x0f
138 enum ptp_packet_state {
139 PTP_PACKET_STATE_UNMATCHED = 0,
140 PTP_PACKET_STATE_MATCHED,
141 PTP_PACKET_STATE_TIMED_OUT,
142 PTP_PACKET_STATE_MATCH_UNWANTED
145 /* NIC synchronised with single word of time only comprising
146 * partial seconds and full nanoseconds: 10^9 ~ 2^30 so 2 bits for seconds.
148 #define MC_NANOSECOND_BITS 30
149 #define MC_NANOSECOND_MASK ((1 << MC_NANOSECOND_BITS) - 1)
150 #define MC_SECOND_MASK ((1 << (32 - MC_NANOSECOND_BITS)) - 1)
152 /* Maximum parts-per-billion adjustment that is acceptable */
153 #define MAX_PPB 1000000
155 /* Precalculate scale word to avoid long long division at runtime */
156 /* This is equivalent to 2^66 / 10^9. */
157 #define PPB_SCALE_WORD ((1LL << (57)) / 1953125LL)
159 /* How much to shift down after scaling to convert to FP40 */
160 #define PPB_SHIFT_FP40 26
162 #define PPB_SHIFT_FP44 22
164 #define PTP_SYNC_ATTEMPTS 4
167 * struct efx_ptp_match - Matching structure, stored in sk_buff's cb area.
168 * @words: UUID and (partial) sequence number
169 * @expiry: Time after which the packet should be delivered irrespective of
171 * @state: The state of the packet - whether it is ready for processing or
172 * whether that is of no interest.
174 struct efx_ptp_match {
175 u32 words[DIV_ROUND_UP(PTP_V1_UUID_LENGTH, 4)];
176 unsigned long expiry;
177 enum ptp_packet_state state;
181 * struct efx_ptp_event_rx - A PTP receive event (from MC)
182 * @link: list of events
183 * @seq0: First part of (PTP) UUID
184 * @seq1: Second part of (PTP) UUID and sequence number
185 * @hwtimestamp: Event timestamp
186 * @expiry: Time which the packet arrived
188 struct efx_ptp_event_rx {
189 struct list_head link;
193 unsigned long expiry;
197 * struct efx_ptp_timeset - Synchronisation between host and MC
198 * @host_start: Host time immediately before hardware timestamp taken
199 * @major: Hardware timestamp, major
200 * @minor: Hardware timestamp, minor
201 * @host_end: Host time immediately after hardware timestamp taken
202 * @wait: Number of NIC clock ticks between hardware timestamp being read and
203 * host end time being seen
204 * @window: Difference of host_end and host_start
205 * @valid: Whether this timeset is valid
207 struct efx_ptp_timeset {
213 u32 window; /* Derived: end - start, allowing for wrap */
217 * struct efx_ptp_data - Precision Time Protocol (PTP) state
218 * @efx: The NIC context
219 * @channel: The PTP channel (Siena only)
220 * @rx_ts_inline: Flag for whether RX timestamps are inline (else they are
222 * @rxq: Receive SKB queue (awaiting timestamps)
223 * @txq: Transmit SKB queue
224 * @evt_list: List of MC receive events awaiting packets
225 * @evt_free_list: List of free events
226 * @evt_lock: Lock for manipulating evt_list and evt_free_list
227 * @rx_evts: Instantiated events (on evt_list and evt_free_list)
228 * @workwq: Work queue for processing pending PTP operations
230 * @reset_required: A serious error has occurred and the PTP task needs to be
231 * reset (disable, enable).
232 * @rxfilters: Receive filters when operating
233 * @rxfilters_count: Num of installed rxfilters, should be == PTP_RXFILTERS_LEN
234 * @config: Current timestamp configuration
235 * @enabled: PTP operation enabled
236 * @mode: Mode in which PTP operating (PTP version)
237 * @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time
238 * @nic_to_kernel_time: Function to convert from NIC to kernel time
239 * @nic_time: contains time details
240 * @nic_time.minor_max: Wrap point for NIC minor times
241 * @nic_time.sync_event_diff_min: Minimum acceptable difference between time
242 * in packet prefix and last MCDI time sync event i.e. how much earlier than
243 * the last sync event time a packet timestamp can be.
244 * @nic_time.sync_event_diff_max: Maximum acceptable difference between time
245 * in packet prefix and last MCDI time sync event i.e. how much later than
246 * the last sync event time a packet timestamp can be.
247 * @nic_time.sync_event_minor_shift: Shift required to make minor time from
248 * field in MCDI time sync event.
249 * @min_synchronisation_ns: Minimum acceptable corrected sync window
250 * @capabilities: Capabilities flags from the NIC
251 * @ts_corrections: contains corrections details
252 * @ts_corrections.ptp_tx: Required driver correction of PTP packet transmit
254 * @ts_corrections.ptp_rx: Required driver correction of PTP packet receive
256 * @ts_corrections.pps_out: PPS output error (information only)
257 * @ts_corrections.pps_in: Required driver correction of PPS input timestamps
258 * @ts_corrections.general_tx: Required driver correction of general packet
259 * transmit timestamps
260 * @ts_corrections.general_rx: Required driver correction of general packet
262 * @evt_frags: Partly assembled PTP events
263 * @evt_frag_idx: Current fragment number
264 * @evt_code: Last event code
265 * @start: Address at which MC indicates ready for synchronisation
266 * @host_time_pps: Host time at last PPS
267 * @adjfreq_ppb_shift: Shift required to convert scaled parts-per-billion
268 * frequency adjustment into a fixed point fractional nanosecond format.
269 * @current_adjfreq: Current ppb adjustment.
270 * @phc_clock: Pointer to registered phc device (if primary function)
271 * @phc_clock_info: Registration structure for phc device
272 * @pps_work: pps work task for handling pps events
273 * @pps_workwq: pps work queue
274 * @nic_ts_enabled: Flag indicating if NIC generated TS events are handled
275 * @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids
276 * allocations in main data path).
277 * @good_syncs: Number of successful synchronisations.
278 * @fast_syncs: Number of synchronisations requiring short delay
279 * @bad_syncs: Number of failed synchronisations.
280 * @sync_timeouts: Number of synchronisation timeouts
281 * @no_time_syncs: Number of synchronisations with no good times.
282 * @invalid_sync_windows: Number of sync windows with bad durations.
283 * @undersize_sync_windows: Number of corrected sync windows that are too small
284 * @oversize_sync_windows: Number of corrected sync windows that are too large
285 * @rx_no_timestamp: Number of packets received without a timestamp.
286 * @timeset: Last set of synchronisation statistics.
287 * @xmit_skb: Transmit SKB function.
289 struct efx_ptp_data {
291 struct efx_channel *channel;
293 struct sk_buff_head rxq;
294 struct sk_buff_head txq;
295 struct list_head evt_list;
296 struct list_head evt_free_list;
298 struct efx_ptp_event_rx rx_evts[MAX_RECEIVE_EVENTS];
299 struct workqueue_struct *workwq;
300 struct work_struct work;
302 u32 rxfilters[PTP_RXFILTERS_LEN];
303 size_t rxfilters_count;
304 struct hwtstamp_config config;
307 void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor);
308 ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor,
312 u32 sync_event_diff_min;
313 u32 sync_event_diff_max;
314 unsigned int sync_event_minor_shift;
316 unsigned int min_synchronisation_ns;
317 unsigned int capabilities;
326 efx_qword_t evt_frags[MAX_EVENT_FRAGS];
329 struct efx_buffer start;
330 struct pps_event_time host_time_pps;
331 unsigned int adjfreq_ppb_shift;
333 struct ptp_clock *phc_clock;
334 struct ptp_clock_info phc_clock_info;
335 struct work_struct pps_work;
336 struct workqueue_struct *pps_workwq;
338 efx_dword_t txbuf[MCDI_TX_BUF_LEN(MC_CMD_PTP_IN_TRANSMIT_LENMAX)];
340 unsigned int good_syncs;
341 unsigned int fast_syncs;
342 unsigned int bad_syncs;
343 unsigned int sync_timeouts;
344 unsigned int no_time_syncs;
345 unsigned int invalid_sync_windows;
346 unsigned int undersize_sync_windows;
347 unsigned int oversize_sync_windows;
348 unsigned int rx_no_timestamp;
349 struct efx_ptp_timeset
350 timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM];
351 void (*xmit_skb)(struct efx_nic *efx, struct sk_buff *skb);
354 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta);
355 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta);
356 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts);
357 static int efx_phc_settime(struct ptp_clock_info *ptp,
358 const struct timespec64 *e_ts);
359 static int efx_phc_enable(struct ptp_clock_info *ptp,
360 struct ptp_clock_request *request, int on);
362 bool efx_ptp_use_mac_tx_timestamps(struct efx_nic *efx)
364 return efx_has_cap(efx, TX_MAC_TIMESTAMPING);
367 /* PTP 'extra' channel is still a traffic channel, but we only create TX queues
368 * if PTP uses MAC TX timestamps, not if PTP uses the MC directly to transmit.
370 static bool efx_ptp_want_txqs(struct efx_channel *channel)
372 return efx_ptp_use_mac_tx_timestamps(channel->efx);
375 #define PTP_SW_STAT(ext_name, field_name) \
376 { #ext_name, 0, offsetof(struct efx_ptp_data, field_name) }
377 #define PTP_MC_STAT(ext_name, mcdi_name) \
378 { #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST }
379 static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = {
380 PTP_SW_STAT(ptp_good_syncs, good_syncs),
381 PTP_SW_STAT(ptp_fast_syncs, fast_syncs),
382 PTP_SW_STAT(ptp_bad_syncs, bad_syncs),
383 PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts),
384 PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs),
385 PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows),
386 PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows),
387 PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows),
388 PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp),
389 PTP_MC_STAT(ptp_tx_timestamp_packets, TX),
390 PTP_MC_STAT(ptp_rx_timestamp_packets, RX),
391 PTP_MC_STAT(ptp_timestamp_packets, TS),
392 PTP_MC_STAT(ptp_filter_matches, FM),
393 PTP_MC_STAT(ptp_non_filter_matches, NFM),
395 #define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc)
396 static const unsigned long efx_ptp_stat_mask[] = {
397 [0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL,
400 size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
405 return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
406 efx_ptp_stat_mask, strings);
409 size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats)
411 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
412 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
419 /* Copy software statistics */
420 for (i = 0; i < PTP_STAT_COUNT; i++) {
421 if (efx_ptp_stat_desc[i].dma_width)
423 stats[i] = *(unsigned int *)((char *)efx->ptp_data +
424 efx_ptp_stat_desc[i].offset);
427 /* Fetch MC statistics. We *must* fill in all statistics or
428 * risk leaking kernel memory to userland, so if the MCDI
429 * request fails we pretend we got zeroes.
431 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS);
432 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
433 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
434 outbuf, sizeof(outbuf), NULL);
436 memset(outbuf, 0, sizeof(outbuf));
437 efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
439 stats, _MCDI_PTR(outbuf, 0), false);
441 return PTP_STAT_COUNT;
444 /* For Siena platforms NIC time is s and ns */
445 static void efx_ptp_ns_to_s_ns(s64 ns, u32 *nic_major, u32 *nic_minor)
447 struct timespec64 ts = ns_to_timespec64(ns);
448 *nic_major = (u32)ts.tv_sec;
449 *nic_minor = ts.tv_nsec;
452 static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
455 ktime_t kt = ktime_set(nic_major, nic_minor);
457 kt = ktime_add_ns(kt, (u64)correction);
459 kt = ktime_sub_ns(kt, (u64)-correction);
463 /* To convert from s27 format to ns we multiply then divide by a power of 2.
464 * For the conversion from ns to s27, the operation is also converted to a
465 * multiply and shift.
467 #define S27_TO_NS_SHIFT (27)
468 #define NS_TO_S27_MULT (((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC)
469 #define NS_TO_S27_SHIFT (63 - S27_TO_NS_SHIFT)
470 #define S27_MINOR_MAX (1 << S27_TO_NS_SHIFT)
472 /* For Huntington platforms NIC time is in seconds and fractions of a second
473 * where the minor register only uses 27 bits in units of 2^-27s.
475 static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
477 struct timespec64 ts = ns_to_timespec64(ns);
478 u32 maj = (u32)ts.tv_sec;
479 u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT +
480 (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT);
482 /* The conversion can result in the minor value exceeding the maximum.
483 * In this case, round up to the next second.
485 if (min >= S27_MINOR_MAX) {
486 min -= S27_MINOR_MAX;
494 static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
496 u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC +
497 (1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT);
498 return ktime_set(nic_major, ns);
501 static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
504 /* Apply the correction and deal with carry */
505 nic_minor += correction;
506 if ((s32)nic_minor < 0) {
507 nic_minor += S27_MINOR_MAX;
509 } else if (nic_minor >= S27_MINOR_MAX) {
510 nic_minor -= S27_MINOR_MAX;
514 return efx_ptp_s27_to_ktime(nic_major, nic_minor);
517 /* For Medford2 platforms the time is in seconds and quarter nanoseconds. */
518 static void efx_ptp_ns_to_s_qns(s64 ns, u32 *nic_major, u32 *nic_minor)
520 struct timespec64 ts = ns_to_timespec64(ns);
522 *nic_major = (u32)ts.tv_sec;
523 *nic_minor = ts.tv_nsec * 4;
526 static ktime_t efx_ptp_s_qns_to_ktime_correction(u32 nic_major, u32 nic_minor,
531 nic_minor = DIV_ROUND_CLOSEST(nic_minor, 4);
532 correction = DIV_ROUND_CLOSEST(correction, 4);
534 kt = ktime_set(nic_major, nic_minor);
537 kt = ktime_add_ns(kt, (u64)correction);
539 kt = ktime_sub_ns(kt, (u64)-correction);
543 struct efx_channel *efx_ptp_channel(struct efx_nic *efx)
545 return efx->ptp_data ? efx->ptp_data->channel : NULL;
548 void efx_ptp_update_channel(struct efx_nic *efx, struct efx_channel *channel)
551 efx->ptp_data->channel = channel;
554 static u32 last_sync_timestamp_major(struct efx_nic *efx)
556 struct efx_channel *channel = efx_ptp_channel(efx);
560 major = channel->sync_timestamp_major;
564 /* The 8000 series and later can provide the time from the MAC, which is only
565 * 48 bits long and provides meta-information in the top 2 bits.
568 efx_ptp_mac_nic_to_ktime_correction(struct efx_nic *efx,
569 struct efx_ptp_data *ptp,
570 u32 nic_major, u32 nic_minor,
577 if (!(nic_major & 0x80000000)) {
578 WARN_ON_ONCE(nic_major >> 16);
580 /* Medford provides 48 bits of timestamp, so we must get the top
581 * 16 bits from the timesync event state.
583 * We only have the lower 16 bits of the time now, but we do
584 * have a full resolution timestamp at some point in past. As
585 * long as the difference between the (real) now and the sync
586 * is less than 2^15, then we can reconstruct the difference
587 * between those two numbers using only the lower 16 bits of
592 * a - b = ((a mod k) - b) mod k
594 * when -k/2 < (a-b) < k/2. In our case k is 2^16. We know
595 * (a mod k) and b, so can calculate the delta, a - b.
598 sync_timestamp = last_sync_timestamp_major(efx);
600 /* Because delta is s16 this does an implicit mask down to
601 * 16 bits which is what we need, assuming
602 * MEDFORD_TX_SECS_EVENT_BITS is 16. delta is signed so that
603 * we can deal with the (unlikely) case of sync timestamps
604 * arriving from the future.
606 delta = nic_major - sync_timestamp;
608 /* Recover the fully specified time now, by applying the offset
609 * to the (fully specified) sync time.
611 nic_major = sync_timestamp + delta;
613 kt = ptp->nic_to_kernel_time(nic_major, nic_minor,
619 ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue)
621 struct efx_nic *efx = tx_queue->efx;
622 struct efx_ptp_data *ptp = efx->ptp_data;
625 if (efx_ptp_use_mac_tx_timestamps(efx))
626 kt = efx_ptp_mac_nic_to_ktime_correction(efx, ptp,
627 tx_queue->completed_timestamp_major,
628 tx_queue->completed_timestamp_minor,
629 ptp->ts_corrections.general_tx);
631 kt = ptp->nic_to_kernel_time(
632 tx_queue->completed_timestamp_major,
633 tx_queue->completed_timestamp_minor,
634 ptp->ts_corrections.general_tx);
638 /* Get PTP attributes and set up time conversions */
639 static int efx_ptp_get_attributes(struct efx_nic *efx)
641 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN);
642 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN);
643 struct efx_ptp_data *ptp = efx->ptp_data;
648 /* Get the PTP attributes. If the NIC doesn't support the operation we
649 * use the default format for compatibility with older NICs i.e.
650 * seconds and nanoseconds.
652 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES);
653 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
654 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
655 outbuf, sizeof(outbuf), &out_len);
657 fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT);
658 } else if (rc == -EINVAL) {
659 fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS;
660 } else if (rc == -EPERM) {
661 pci_info(efx->pci_dev, "no PTP support\n");
664 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf),
665 outbuf, sizeof(outbuf), rc);
670 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION:
671 ptp->ns_to_nic_time = efx_ptp_ns_to_s27;
672 ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction;
673 ptp->nic_time.minor_max = 1 << 27;
674 ptp->nic_time.sync_event_minor_shift = 19;
676 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS:
677 ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns;
678 ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction;
679 ptp->nic_time.minor_max = 1000000000;
680 ptp->nic_time.sync_event_minor_shift = 22;
682 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_QTR_NANOSECONDS:
683 ptp->ns_to_nic_time = efx_ptp_ns_to_s_qns;
684 ptp->nic_to_kernel_time = efx_ptp_s_qns_to_ktime_correction;
685 ptp->nic_time.minor_max = 4000000000UL;
686 ptp->nic_time.sync_event_minor_shift = 24;
692 /* Precalculate acceptable difference between the minor time in the
693 * packet prefix and the last MCDI time sync event. We expect the
694 * packet prefix timestamp to be after of sync event by up to one
695 * sync event interval (0.25s) but we allow it to exceed this by a
696 * fuzz factor of (0.1s)
698 ptp->nic_time.sync_event_diff_min = ptp->nic_time.minor_max
699 - (ptp->nic_time.minor_max / 10);
700 ptp->nic_time.sync_event_diff_max = (ptp->nic_time.minor_max / 4)
701 + (ptp->nic_time.minor_max / 10);
703 /* MC_CMD_PTP_OP_GET_ATTRIBUTES has been extended twice from an older
704 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT. The function now may return
705 * a value to use for the minimum acceptable corrected synchronization
706 * window and may return further capabilities.
707 * If we have the extra information store it. For older firmware that
708 * does not implement the extended command use the default value.
711 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_CAPABILITIES_OFST)
712 ptp->min_synchronisation_ns =
714 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
716 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
719 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
720 ptp->capabilities = MCDI_DWORD(outbuf,
721 PTP_OUT_GET_ATTRIBUTES_CAPABILITIES);
723 ptp->capabilities = 0;
725 /* Set up the shift for conversion between frequency
726 * adjustments in parts-per-billion and the fixed-point
727 * fractional ns format that the adapter uses.
729 if (ptp->capabilities & (1 << MC_CMD_PTP_OUT_GET_ATTRIBUTES_FP44_FREQ_ADJ_LBN))
730 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP44;
732 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP40;
737 /* Get PTP timestamp corrections */
738 static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
740 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN);
741 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN);
745 /* Get the timestamp corrections from the NIC. If this operation is
746 * not supported (older NICs) then no correction is required.
748 MCDI_SET_DWORD(inbuf, PTP_IN_OP,
749 MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS);
750 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
752 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
753 outbuf, sizeof(outbuf), &out_len);
755 efx->ptp_data->ts_corrections.ptp_tx = MCDI_DWORD(outbuf,
756 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT);
757 efx->ptp_data->ts_corrections.ptp_rx = MCDI_DWORD(outbuf,
758 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE);
759 efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf,
760 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT);
761 efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf,
762 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN);
764 if (out_len >= MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN) {
765 efx->ptp_data->ts_corrections.general_tx = MCDI_DWORD(
767 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_TX);
768 efx->ptp_data->ts_corrections.general_rx = MCDI_DWORD(
770 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_RX);
772 efx->ptp_data->ts_corrections.general_tx =
773 efx->ptp_data->ts_corrections.ptp_tx;
774 efx->ptp_data->ts_corrections.general_rx =
775 efx->ptp_data->ts_corrections.ptp_rx;
777 } else if (rc == -EINVAL) {
778 efx->ptp_data->ts_corrections.ptp_tx = 0;
779 efx->ptp_data->ts_corrections.ptp_rx = 0;
780 efx->ptp_data->ts_corrections.pps_out = 0;
781 efx->ptp_data->ts_corrections.pps_in = 0;
782 efx->ptp_data->ts_corrections.general_tx = 0;
783 efx->ptp_data->ts_corrections.general_rx = 0;
785 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), outbuf,
793 /* Enable MCDI PTP support. */
794 static int efx_ptp_enable(struct efx_nic *efx)
796 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
797 MCDI_DECLARE_BUF_ERR(outbuf);
800 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE);
801 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
802 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE,
803 efx->ptp_data->channel ?
804 efx->ptp_data->channel->channel : 0);
805 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode);
807 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
808 outbuf, sizeof(outbuf), NULL);
809 rc = (rc == -EALREADY) ? 0 : rc;
811 efx_mcdi_display_error(efx, MC_CMD_PTP,
812 MC_CMD_PTP_IN_ENABLE_LEN,
813 outbuf, sizeof(outbuf), rc);
817 /* Disable MCDI PTP support.
819 * Note that this function should never rely on the presence of ptp_data -
820 * may be called before that exists.
822 static int efx_ptp_disable(struct efx_nic *efx)
824 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
825 MCDI_DECLARE_BUF_ERR(outbuf);
828 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE);
829 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
830 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
831 outbuf, sizeof(outbuf), NULL);
832 rc = (rc == -EALREADY) ? 0 : rc;
833 /* If we get ENOSYS, the NIC doesn't support PTP, and thus this function
834 * should only have been called during probe.
836 if (rc == -ENOSYS || rc == -EPERM)
837 pci_info(efx->pci_dev, "no PTP support\n");
839 efx_mcdi_display_error(efx, MC_CMD_PTP,
840 MC_CMD_PTP_IN_DISABLE_LEN,
841 outbuf, sizeof(outbuf), rc);
845 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
849 while ((skb = skb_dequeue(q))) {
851 netif_receive_skb(skb);
856 static void efx_ptp_handle_no_channel(struct efx_nic *efx)
858 netif_err(efx, drv, efx->net_dev,
859 "ERROR: PTP requires MSI-X and 1 additional interrupt"
860 "vector. PTP disabled\n");
863 /* Repeatedly send the host time to the MC which will capture the hardware
866 static void efx_ptp_send_times(struct efx_nic *efx,
867 struct pps_event_time *last_time)
869 struct pps_event_time now;
870 struct timespec64 limit;
871 struct efx_ptp_data *ptp = efx->ptp_data;
872 int *mc_running = ptp->start.addr;
876 timespec64_add_ns(&limit, SYNCHRONISE_PERIOD_NS);
878 /* Write host time for specified period or until MC is done */
879 while ((timespec64_compare(&now.ts_real, &limit) < 0) &&
880 READ_ONCE(*mc_running)) {
881 struct timespec64 update_time;
882 unsigned int host_time;
884 /* Don't update continuously to avoid saturating the PCIe bus */
885 update_time = now.ts_real;
886 timespec64_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS);
889 } while ((timespec64_compare(&now.ts_real, &update_time) < 0) &&
890 READ_ONCE(*mc_running));
892 /* Synchronise NIC with single word of time only */
893 host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS |
894 now.ts_real.tv_nsec);
895 /* Update host time in NIC memory */
896 efx->type->ptp_write_host_time(efx, host_time);
901 /* Read a timeset from the MC's results and partial process. */
902 static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data),
903 struct efx_ptp_timeset *timeset)
905 unsigned start_ns, end_ns;
907 timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART);
908 timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR);
909 timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR);
910 timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND),
911 timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
914 start_ns = timeset->host_start & MC_NANOSECOND_MASK;
915 end_ns = timeset->host_end & MC_NANOSECOND_MASK;
916 /* Allow for rollover */
917 if (end_ns < start_ns)
918 end_ns += NSEC_PER_SEC;
919 /* Determine duration of operation */
920 timeset->window = end_ns - start_ns;
923 /* Process times received from MC.
925 * Extract times from returned results, and establish the minimum value
926 * seen. The minimum value represents the "best" possible time and events
927 * too much greater than this are rejected - the machine is, perhaps, too
928 * busy. A number of readings are taken so that, hopefully, at least one good
929 * synchronisation will be seen in the results.
932 efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf),
933 size_t response_length,
934 const struct pps_event_time *last_time)
936 unsigned number_readings =
937 MCDI_VAR_ARRAY_LEN(response_length,
938 PTP_OUT_SYNCHRONIZE_TIMESET);
941 unsigned last_good = 0;
942 struct efx_ptp_data *ptp = efx->ptp_data;
945 struct timespec64 delta;
948 if (number_readings == 0)
951 /* Read the set of results and find the last good host-MC
952 * synchronization result. The MC times when it finishes reading the
953 * host time so the corrected window time should be fairly constant
954 * for a given platform. Increment stats for any results that appear
957 for (i = 0; i < number_readings; i++) {
958 s32 window, corrected;
959 struct timespec64 wait;
961 efx_ptp_read_timeset(
962 MCDI_ARRAY_STRUCT_PTR(synch_buf,
963 PTP_OUT_SYNCHRONIZE_TIMESET, i),
966 wait = ktime_to_timespec64(
967 ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0));
968 window = ptp->timeset[i].window;
969 corrected = window - wait.tv_nsec;
971 /* We expect the uncorrected synchronization window to be at
972 * least as large as the interval between host start and end
973 * times. If it is smaller than this then this is mostly likely
974 * to be a consequence of the host's time being adjusted.
975 * Check that the corrected sync window is in a reasonable
976 * range. If it is out of range it is likely to be because an
977 * interrupt or other delay occurred between reading the system
978 * time and writing it to MC memory.
980 if (window < SYNCHRONISATION_GRANULARITY_NS) {
981 ++ptp->invalid_sync_windows;
982 } else if (corrected >= MAX_SYNCHRONISATION_NS) {
983 ++ptp->oversize_sync_windows;
984 } else if (corrected < ptp->min_synchronisation_ns) {
985 ++ptp->undersize_sync_windows;
993 netif_warn(efx, drv, efx->net_dev,
994 "PTP no suitable synchronisations\n");
998 /* Calculate delay from last good sync (host time) to last_time.
999 * It is possible that the seconds rolled over between taking
1000 * the start reading and the last value written by the host. The
1001 * timescales are such that a gap of more than one second is never
1002 * expected. delta is *not* normalised.
1004 start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS;
1005 last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK;
1006 if (start_sec != last_sec &&
1007 ((start_sec + 1) & MC_SECOND_MASK) != last_sec) {
1008 netif_warn(efx, hw, efx->net_dev,
1009 "PTP bad synchronisation seconds\n");
1012 delta.tv_sec = (last_sec - start_sec) & 1;
1014 last_time->ts_real.tv_nsec -
1015 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
1017 /* Convert the NIC time at last good sync into kernel time.
1018 * No correction is required - this time is the output of a
1021 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
1022 ptp->timeset[last_good].minor, 0);
1024 /* Calculate delay from NIC top of second to last_time */
1025 delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec;
1027 /* Set PPS timestamp to match NIC top of second */
1028 ptp->host_time_pps = *last_time;
1029 pps_sub_ts(&ptp->host_time_pps, delta);
1034 /* Synchronize times between the host and the MC */
1035 static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
1037 struct efx_ptp_data *ptp = efx->ptp_data;
1038 MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX);
1039 size_t response_length;
1041 unsigned long timeout;
1042 struct pps_event_time last_time = {};
1043 unsigned int loops = 0;
1044 int *start = ptp->start.addr;
1046 MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE);
1047 MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0);
1048 MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS,
1050 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
1051 ptp->start.dma_addr);
1053 /* Clear flag that signals MC ready */
1054 WRITE_ONCE(*start, 0);
1055 rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf,
1056 MC_CMD_PTP_IN_SYNCHRONIZE_LEN);
1057 EFX_WARN_ON_ONCE_PARANOID(rc);
1059 /* Wait for start from MCDI (or timeout) */
1060 timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS);
1061 while (!READ_ONCE(*start) && (time_before(jiffies, timeout))) {
1062 udelay(20); /* Usually start MCDI execution quickly */
1068 if (!time_before(jiffies, timeout))
1069 ++ptp->sync_timeouts;
1071 if (READ_ONCE(*start))
1072 efx_ptp_send_times(efx, &last_time);
1074 /* Collect results */
1075 rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP,
1076 MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
1077 synch_buf, sizeof(synch_buf),
1080 rc = efx_ptp_process_times(efx, synch_buf, response_length,
1085 ++ptp->no_time_syncs;
1088 /* Increment the bad syncs counter if the synchronize fails, whatever
1097 /* Transmit a PTP packet via the dedicated hardware timestamped queue. */
1098 static void efx_ptp_xmit_skb_queue(struct efx_nic *efx, struct sk_buff *skb)
1100 struct efx_ptp_data *ptp_data = efx->ptp_data;
1101 u8 type = efx_tx_csum_type_skb(skb);
1102 struct efx_tx_queue *tx_queue;
1104 tx_queue = efx_channel_get_tx_queue(ptp_data->channel, type);
1105 if (tx_queue && tx_queue->timestamping) {
1106 /* This code invokes normal driver TX code which is always
1107 * protected from softirqs when called from generic TX code,
1108 * which in turn disables preemption. Look at __dev_queue_xmit
1109 * which uses rcu_read_lock_bh disabling preemption for RCU
1110 * plus disabling softirqs. We do not need RCU reader
1113 * Although it is theoretically safe for current PTP TX/RX code
1114 * running without disabling softirqs, there are three good
1115 * reasond for doing so:
1117 * 1) The code invoked is mainly implemented for non-PTP
1118 * packets and it is always executed with softirqs
1120 * 2) This being a single PTP packet, better to not
1121 * interrupt its processing by softirqs which can lead
1122 * to high latencies.
1123 * 3) netdev_xmit_more checks preemption is disabled and
1124 * triggers a BUG_ON if not.
1127 efx_enqueue_skb(tx_queue, skb);
1130 WARN_ONCE(1, "PTP channel has no timestamped tx queue\n");
1131 dev_kfree_skb_any(skb);
1135 /* Transmit a PTP packet, via the MCDI interface, to the wire. */
1136 static void efx_ptp_xmit_skb_mc(struct efx_nic *efx, struct sk_buff *skb)
1138 struct efx_ptp_data *ptp_data = efx->ptp_data;
1139 struct skb_shared_hwtstamps timestamps;
1141 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
1144 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT);
1145 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0);
1146 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len);
1147 if (skb_shinfo(skb)->nr_frags != 0) {
1148 rc = skb_linearize(skb);
1153 if (skb->ip_summed == CHECKSUM_PARTIAL) {
1154 rc = skb_checksum_help(skb);
1158 skb_copy_from_linear_data(skb,
1159 MCDI_PTR(ptp_data->txbuf,
1160 PTP_IN_TRANSMIT_PACKET),
1162 rc = efx_mcdi_rpc(efx, MC_CMD_PTP,
1163 ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len),
1164 txtime, sizeof(txtime), &len);
1168 memset(×tamps, 0, sizeof(timestamps));
1169 timestamps.hwtstamp = ptp_data->nic_to_kernel_time(
1170 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR),
1171 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR),
1172 ptp_data->ts_corrections.ptp_tx);
1174 skb_tstamp_tx(skb, ×tamps);
1179 dev_kfree_skb_any(skb);
1184 static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
1186 struct efx_ptp_data *ptp = efx->ptp_data;
1187 struct list_head *cursor;
1188 struct list_head *next;
1190 if (ptp->rx_ts_inline)
1193 /* Drop time-expired events */
1194 spin_lock_bh(&ptp->evt_lock);
1195 list_for_each_safe(cursor, next, &ptp->evt_list) {
1196 struct efx_ptp_event_rx *evt;
1198 evt = list_entry(cursor, struct efx_ptp_event_rx,
1200 if (time_after(jiffies, evt->expiry)) {
1201 list_move(&evt->link, &ptp->evt_free_list);
1202 netif_warn(efx, hw, efx->net_dev,
1203 "PTP rx event dropped\n");
1206 spin_unlock_bh(&ptp->evt_lock);
1209 static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
1210 struct sk_buff *skb)
1212 struct efx_ptp_data *ptp = efx->ptp_data;
1214 struct list_head *cursor;
1215 struct list_head *next;
1216 struct efx_ptp_match *match;
1217 enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED;
1219 WARN_ON_ONCE(ptp->rx_ts_inline);
1221 spin_lock_bh(&ptp->evt_lock);
1222 evts_waiting = !list_empty(&ptp->evt_list);
1223 spin_unlock_bh(&ptp->evt_lock);
1226 return PTP_PACKET_STATE_UNMATCHED;
1228 match = (struct efx_ptp_match *)skb->cb;
1229 /* Look for a matching timestamp in the event queue */
1230 spin_lock_bh(&ptp->evt_lock);
1231 list_for_each_safe(cursor, next, &ptp->evt_list) {
1232 struct efx_ptp_event_rx *evt;
1234 evt = list_entry(cursor, struct efx_ptp_event_rx, link);
1235 if ((evt->seq0 == match->words[0]) &&
1236 (evt->seq1 == match->words[1])) {
1237 struct skb_shared_hwtstamps *timestamps;
1239 /* Match - add in hardware timestamp */
1240 timestamps = skb_hwtstamps(skb);
1241 timestamps->hwtstamp = evt->hwtimestamp;
1243 match->state = PTP_PACKET_STATE_MATCHED;
1244 rc = PTP_PACKET_STATE_MATCHED;
1245 list_move(&evt->link, &ptp->evt_free_list);
1249 spin_unlock_bh(&ptp->evt_lock);
1254 /* Process any queued receive events and corresponding packets
1256 * q is returned with all the packets that are ready for delivery.
1258 static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
1260 struct efx_ptp_data *ptp = efx->ptp_data;
1261 struct sk_buff *skb;
1263 while ((skb = skb_dequeue(&ptp->rxq))) {
1264 struct efx_ptp_match *match;
1266 match = (struct efx_ptp_match *)skb->cb;
1267 if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) {
1268 __skb_queue_tail(q, skb);
1269 } else if (efx_ptp_match_rx(efx, skb) ==
1270 PTP_PACKET_STATE_MATCHED) {
1271 __skb_queue_tail(q, skb);
1272 } else if (time_after(jiffies, match->expiry)) {
1273 match->state = PTP_PACKET_STATE_TIMED_OUT;
1274 ++ptp->rx_no_timestamp;
1275 __skb_queue_tail(q, skb);
1277 /* Replace unprocessed entry and stop */
1278 skb_queue_head(&ptp->rxq, skb);
1284 /* Complete processing of a received packet */
1285 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1288 netif_receive_skb(skb);
1292 static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
1294 struct efx_ptp_data *ptp = efx->ptp_data;
1296 while (ptp->rxfilters_count) {
1297 ptp->rxfilters_count--;
1298 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1299 ptp->rxfilters[ptp->rxfilters_count]);
1303 static void efx_ptp_init_filter(struct efx_nic *efx,
1304 struct efx_filter_spec *rxfilter)
1306 struct efx_channel *channel = efx->ptp_data->channel;
1307 struct efx_rx_queue *queue = efx_channel_get_rx_queue(channel);
1309 efx_filter_init_rx(rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1310 efx_rx_queue_index(queue));
1313 static int efx_ptp_insert_filter(struct efx_nic *efx,
1314 struct efx_filter_spec *rxfilter)
1316 struct efx_ptp_data *ptp = efx->ptp_data;
1318 int rc = efx_filter_insert_filter(efx, rxfilter, true);
1321 ptp->rxfilters[ptp->rxfilters_count] = rc;
1322 ptp->rxfilters_count++;
1326 static int efx_ptp_insert_ipv4_filter(struct efx_nic *efx, u16 port)
1328 struct efx_filter_spec rxfilter;
1330 efx_ptp_init_filter(efx, &rxfilter);
1331 efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP, htonl(PTP_ADDR_IPV4),
1333 return efx_ptp_insert_filter(efx, &rxfilter);
1336 static int efx_ptp_insert_ipv6_filter(struct efx_nic *efx, u16 port)
1338 const struct in6_addr addr = {{PTP_ADDR_IPV6}};
1339 struct efx_filter_spec rxfilter;
1341 efx_ptp_init_filter(efx, &rxfilter);
1342 efx_filter_set_ipv6_local(&rxfilter, IPPROTO_UDP, &addr, htons(port));
1343 return efx_ptp_insert_filter(efx, &rxfilter);
1346 static int efx_ptp_insert_eth_filter(struct efx_nic *efx)
1348 const u8 addr[ETH_ALEN] = PTP_ADDR_ETHER;
1349 struct efx_filter_spec rxfilter;
1351 efx_ptp_init_filter(efx, &rxfilter);
1352 efx_filter_set_eth_local(&rxfilter, EFX_FILTER_VID_UNSPEC, addr);
1353 rxfilter.match_flags |= EFX_FILTER_MATCH_ETHER_TYPE;
1354 rxfilter.ether_type = htons(ETH_P_1588);
1355 return efx_ptp_insert_filter(efx, &rxfilter);
1358 static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1360 struct efx_ptp_data *ptp = efx->ptp_data;
1363 if (!ptp->channel || ptp->rxfilters_count)
1366 /* Must filter on both event and general ports to ensure
1367 * that there is no packet re-ordering.
1369 rc = efx_ptp_insert_ipv4_filter(efx, PTP_EVENT_PORT);
1373 rc = efx_ptp_insert_ipv4_filter(efx, PTP_GENERAL_PORT);
1377 /* if the NIC supports hw timestamps by the MAC, we can support
1378 * PTP over IPv6 and Ethernet
1380 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1381 rc = efx_ptp_insert_ipv6_filter(efx, PTP_EVENT_PORT);
1385 rc = efx_ptp_insert_ipv6_filter(efx, PTP_GENERAL_PORT);
1389 rc = efx_ptp_insert_eth_filter(efx);
1397 efx_ptp_remove_multicast_filters(efx);
1401 static int efx_ptp_start(struct efx_nic *efx)
1403 struct efx_ptp_data *ptp = efx->ptp_data;
1406 ptp->reset_required = false;
1408 rc = efx_ptp_insert_multicast_filters(efx);
1412 rc = efx_ptp_enable(efx);
1416 ptp->evt_frag_idx = 0;
1417 ptp->current_adjfreq = 0;
1422 efx_ptp_remove_multicast_filters(efx);
1426 static int efx_ptp_stop(struct efx_nic *efx)
1428 struct efx_ptp_data *ptp = efx->ptp_data;
1429 struct list_head *cursor;
1430 struct list_head *next;
1436 rc = efx_ptp_disable(efx);
1438 efx_ptp_remove_multicast_filters(efx);
1440 /* Make sure RX packets are really delivered */
1441 efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq);
1442 skb_queue_purge(&efx->ptp_data->txq);
1444 /* Drop any pending receive events */
1445 spin_lock_bh(&efx->ptp_data->evt_lock);
1446 list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) {
1447 list_move(cursor, &efx->ptp_data->evt_free_list);
1449 spin_unlock_bh(&efx->ptp_data->evt_lock);
1454 static int efx_ptp_restart(struct efx_nic *efx)
1456 if (efx->ptp_data && efx->ptp_data->enabled)
1457 return efx_ptp_start(efx);
1461 static void efx_ptp_pps_worker(struct work_struct *work)
1463 struct efx_ptp_data *ptp =
1464 container_of(work, struct efx_ptp_data, pps_work);
1465 struct efx_nic *efx = ptp->efx;
1466 struct ptp_clock_event ptp_evt;
1468 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
1471 ptp_evt.type = PTP_CLOCK_PPSUSR;
1472 ptp_evt.pps_times = ptp->host_time_pps;
1473 ptp_clock_event(ptp->phc_clock, &ptp_evt);
1476 static void efx_ptp_worker(struct work_struct *work)
1478 struct efx_ptp_data *ptp_data =
1479 container_of(work, struct efx_ptp_data, work);
1480 struct efx_nic *efx = ptp_data->efx;
1481 struct sk_buff *skb;
1482 struct sk_buff_head tempq;
1484 if (ptp_data->reset_required) {
1490 efx_ptp_drop_time_expired_events(efx);
1492 __skb_queue_head_init(&tempq);
1493 efx_ptp_process_events(efx, &tempq);
1495 while ((skb = skb_dequeue(&ptp_data->txq)))
1496 ptp_data->xmit_skb(efx, skb);
1498 while ((skb = __skb_dequeue(&tempq)))
1499 efx_ptp_process_rx(efx, skb);
1502 static const struct ptp_clock_info efx_phc_clock_info = {
1503 .owner = THIS_MODULE,
1511 .adjfreq = efx_phc_adjfreq,
1512 .adjtime = efx_phc_adjtime,
1513 .gettime64 = efx_phc_gettime,
1514 .settime64 = efx_phc_settime,
1515 .enable = efx_phc_enable,
1518 /* Initialise PTP state. */
1519 int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1521 struct efx_ptp_data *ptp;
1525 if (efx->ptp_data) {
1526 efx->ptp_data->channel = channel;
1530 ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL);
1531 efx->ptp_data = ptp;
1536 ptp->channel = channel;
1537 ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1539 rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL);
1543 skb_queue_head_init(&ptp->rxq);
1544 skb_queue_head_init(&ptp->txq);
1545 ptp->workwq = create_singlethread_workqueue("sfc_ptp");
1551 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1552 ptp->xmit_skb = efx_ptp_xmit_skb_queue;
1553 /* Request sync events on this channel. */
1554 channel->sync_events_state = SYNC_EVENTS_QUIESCENT;
1556 ptp->xmit_skb = efx_ptp_xmit_skb_mc;
1559 INIT_WORK(&ptp->work, efx_ptp_worker);
1560 ptp->config.flags = 0;
1561 ptp->config.tx_type = HWTSTAMP_TX_OFF;
1562 ptp->config.rx_filter = HWTSTAMP_FILTER_NONE;
1563 INIT_LIST_HEAD(&ptp->evt_list);
1564 INIT_LIST_HEAD(&ptp->evt_free_list);
1565 spin_lock_init(&ptp->evt_lock);
1566 for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++)
1567 list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list);
1569 /* Get the NIC PTP attributes and set up time conversions */
1570 rc = efx_ptp_get_attributes(efx);
1574 /* Get the timestamp corrections */
1575 rc = efx_ptp_get_timestamp_corrections(efx);
1579 if (efx->mcdi->fn_flags &
1580 (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) {
1581 ptp->phc_clock_info = efx_phc_clock_info;
1582 ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info,
1583 &efx->pci_dev->dev);
1584 if (IS_ERR(ptp->phc_clock)) {
1585 rc = PTR_ERR(ptp->phc_clock);
1587 } else if (ptp->phc_clock) {
1588 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1589 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
1590 if (!ptp->pps_workwq) {
1596 ptp->nic_ts_enabled = false;
1600 ptp_clock_unregister(efx->ptp_data->phc_clock);
1603 destroy_workqueue(efx->ptp_data->workwq);
1606 efx_nic_free_buffer(efx, &ptp->start);
1609 kfree(efx->ptp_data);
1610 efx->ptp_data = NULL;
1615 /* Initialise PTP channel.
1617 * Setting core_index to zero causes the queue to be initialised and doesn't
1618 * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
1620 static int efx_ptp_probe_channel(struct efx_channel *channel)
1622 struct efx_nic *efx = channel->efx;
1625 channel->irq_moderation_us = 0;
1626 channel->rx_queue.core_index = 0;
1628 rc = efx_ptp_probe(efx, channel);
1629 /* Failure to probe PTP is not fatal; this channel will just not be
1630 * used for anything.
1631 * In the case of EPERM, efx_ptp_probe will print its own message (in
1632 * efx_ptp_get_attributes()), so we don't need to.
1634 if (rc && rc != -EPERM)
1635 netif_warn(efx, drv, efx->net_dev,
1636 "Failed to probe PTP, rc=%d\n", rc);
1640 void efx_ptp_remove(struct efx_nic *efx)
1645 (void)efx_ptp_disable(efx);
1647 cancel_work_sync(&efx->ptp_data->work);
1648 if (efx->ptp_data->pps_workwq)
1649 cancel_work_sync(&efx->ptp_data->pps_work);
1651 skb_queue_purge(&efx->ptp_data->rxq);
1652 skb_queue_purge(&efx->ptp_data->txq);
1654 if (efx->ptp_data->phc_clock) {
1655 destroy_workqueue(efx->ptp_data->pps_workwq);
1656 ptp_clock_unregister(efx->ptp_data->phc_clock);
1659 destroy_workqueue(efx->ptp_data->workwq);
1661 efx_nic_free_buffer(efx, &efx->ptp_data->start);
1662 kfree(efx->ptp_data);
1663 efx->ptp_data = NULL;
1666 static void efx_ptp_remove_channel(struct efx_channel *channel)
1668 efx_ptp_remove(channel->efx);
1671 static void efx_ptp_get_channel_name(struct efx_channel *channel,
1672 char *buf, size_t len)
1674 snprintf(buf, len, "%s-ptp", channel->efx->name);
1677 /* Determine whether this packet should be processed by the PTP module
1678 * or transmitted conventionally.
1680 bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1682 return efx->ptp_data &&
1683 efx->ptp_data->enabled &&
1684 skb->len >= PTP_MIN_LENGTH &&
1685 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1686 likely(skb->protocol == htons(ETH_P_IP)) &&
1687 skb_transport_header_was_set(skb) &&
1688 skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1689 ip_hdr(skb)->protocol == IPPROTO_UDP &&
1691 skb_transport_offset(skb) + sizeof(struct udphdr) &&
1692 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1695 /* Receive a PTP packet. Packets are queued until the arrival of
1696 * the receive timestamp from the MC - this will probably occur after the
1697 * packet arrival because of the processing in the MC.
1699 static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1701 struct efx_nic *efx = channel->efx;
1702 struct efx_ptp_data *ptp = efx->ptp_data;
1703 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1704 u8 *match_data_012, *match_data_345;
1705 unsigned int version;
1708 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1710 /* Correct version? */
1711 if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1712 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1716 version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1717 if (version != PTP_VERSION_V1) {
1721 /* PTP V1 uses all six bytes of the UUID to match the packet
1724 match_data_012 = data + PTP_V1_UUID_OFFSET;
1725 match_data_345 = data + PTP_V1_UUID_OFFSET + 3;
1727 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1731 version = data[PTP_V2_VERSION_OFFSET];
1732 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1736 /* The original V2 implementation uses bytes 2-7 of
1737 * the UUID to match the packet to the timestamp. This
1738 * discards two of the bytes of the MAC address used
1739 * to create the UUID (SF bug 33070). The PTP V2
1740 * enhanced mode fixes this issue and uses bytes 0-2
1741 * and byte 5-7 of the UUID.
1743 match_data_345 = data + PTP_V2_UUID_OFFSET + 5;
1744 if (ptp->mode == MC_CMD_PTP_MODE_V2) {
1745 match_data_012 = data + PTP_V2_UUID_OFFSET + 2;
1747 match_data_012 = data + PTP_V2_UUID_OFFSET + 0;
1748 BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED);
1752 /* Does this packet require timestamping? */
1753 if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1754 match->state = PTP_PACKET_STATE_UNMATCHED;
1756 /* We expect the sequence number to be in the same position in
1757 * the packet for PTP V1 and V2
1759 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1760 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1762 /* Extract UUID/Sequence information */
1763 match->words[0] = (match_data_012[0] |
1764 (match_data_012[1] << 8) |
1765 (match_data_012[2] << 16) |
1766 (match_data_345[0] << 24));
1767 match->words[1] = (match_data_345[1] |
1768 (match_data_345[2] << 8) |
1769 (data[PTP_V1_SEQUENCE_OFFSET +
1770 PTP_V1_SEQUENCE_LENGTH - 1] <<
1773 match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1776 skb_queue_tail(&ptp->rxq, skb);
1777 queue_work(ptp->workwq, &ptp->work);
1782 /* Transmit a PTP packet. This has to be transmitted by the MC
1783 * itself, through an MCDI call. MCDI calls aren't permitted
1784 * in the transmit path so defer the actual transmission to a suitable worker.
1786 int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1788 struct efx_ptp_data *ptp = efx->ptp_data;
1790 skb_queue_tail(&ptp->txq, skb);
1792 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1793 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1794 efx_xmit_hwtstamp_pending(skb);
1795 queue_work(ptp->workwq, &ptp->work);
1797 return NETDEV_TX_OK;
1800 int efx_ptp_get_mode(struct efx_nic *efx)
1802 return efx->ptp_data->mode;
1805 int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1806 unsigned int new_mode)
1808 if ((enable_wanted != efx->ptp_data->enabled) ||
1809 (enable_wanted && (efx->ptp_data->mode != new_mode))) {
1812 if (enable_wanted) {
1813 /* Change of mode requires disable */
1814 if (efx->ptp_data->enabled &&
1815 (efx->ptp_data->mode != new_mode)) {
1816 efx->ptp_data->enabled = false;
1817 rc = efx_ptp_stop(efx);
1822 /* Set new operating mode and establish
1823 * baseline synchronisation, which must
1826 efx->ptp_data->mode = new_mode;
1827 if (netif_running(efx->net_dev))
1828 rc = efx_ptp_start(efx);
1830 rc = efx_ptp_synchronize(efx,
1831 PTP_SYNC_ATTEMPTS * 2);
1836 rc = efx_ptp_stop(efx);
1842 efx->ptp_data->enabled = enable_wanted;
1848 static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
1852 if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1853 (init->tx_type != HWTSTAMP_TX_ON))
1856 rc = efx->type->ptp_set_ts_config(efx, init);
1860 efx->ptp_data->config = *init;
1864 void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
1866 struct efx_ptp_data *ptp = efx->ptp_data;
1867 struct efx_nic *primary = efx->primary;
1874 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1875 SOF_TIMESTAMPING_RX_HARDWARE |
1876 SOF_TIMESTAMPING_RAW_HARDWARE);
1877 /* Check licensed features. If we don't have the license for TX
1878 * timestamps, the NIC will not support them.
1880 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1881 struct efx_ef10_nic_data *nic_data = efx->nic_data;
1883 if (!(nic_data->licensed_features &
1884 (1 << LICENSED_V3_FEATURES_TX_TIMESTAMPS_LBN)))
1885 ts_info->so_timestamping &=
1886 ~SOF_TIMESTAMPING_TX_HARDWARE;
1888 if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1889 ts_info->phc_index =
1890 ptp_clock_index(primary->ptp_data->phc_clock);
1891 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1892 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1895 int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1897 struct hwtstamp_config config;
1900 /* Not a PTP enabled port */
1904 if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
1907 rc = efx_ptp_ts_init(efx, &config);
1911 return copy_to_user(ifr->ifr_data, &config, sizeof(config))
1915 int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1920 return copy_to_user(ifr->ifr_data, &efx->ptp_data->config,
1921 sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
1924 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1926 struct efx_ptp_data *ptp = efx->ptp_data;
1928 netif_err(efx, hw, efx->net_dev,
1929 "PTP unexpected event length: got %d expected %d\n",
1930 ptp->evt_frag_idx, expected_frag_len);
1931 ptp->reset_required = true;
1932 queue_work(ptp->workwq, &ptp->work);
1935 /* Process a completed receive event. Put it on the event queue and
1936 * start worker thread. This is required because event and their
1937 * correspoding packets may come in either order.
1939 static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
1941 struct efx_ptp_event_rx *evt = NULL;
1943 if (WARN_ON_ONCE(ptp->rx_ts_inline))
1946 if (ptp->evt_frag_idx != 3) {
1947 ptp_event_failure(efx, 3);
1951 spin_lock_bh(&ptp->evt_lock);
1952 if (!list_empty(&ptp->evt_free_list)) {
1953 evt = list_first_entry(&ptp->evt_free_list,
1954 struct efx_ptp_event_rx, link);
1955 list_del(&evt->link);
1957 evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA);
1958 evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2],
1960 (EFX_QWORD_FIELD(ptp->evt_frags[1],
1961 MCDI_EVENT_SRC) << 8) |
1962 (EFX_QWORD_FIELD(ptp->evt_frags[0],
1963 MCDI_EVENT_SRC) << 16));
1964 evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
1965 EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
1966 EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
1967 ptp->ts_corrections.ptp_rx);
1968 evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1969 list_add_tail(&evt->link, &ptp->evt_list);
1971 queue_work(ptp->workwq, &ptp->work);
1972 } else if (net_ratelimit()) {
1973 /* Log a rate-limited warning message. */
1974 netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n");
1976 spin_unlock_bh(&ptp->evt_lock);
1979 static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1981 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1982 if (ptp->evt_frag_idx != 1) {
1983 ptp_event_failure(efx, 1);
1987 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1990 static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1992 if (ptp->nic_ts_enabled)
1993 queue_work(ptp->pps_workwq, &ptp->pps_work);
1996 void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1998 struct efx_ptp_data *ptp = efx->ptp_data;
1999 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
2002 if (!efx->ptp_warned) {
2003 netif_warn(efx, drv, efx->net_dev,
2004 "Received PTP event but PTP not set up\n");
2005 efx->ptp_warned = true;
2013 if (ptp->evt_frag_idx == 0) {
2014 ptp->evt_code = code;
2015 } else if (ptp->evt_code != code) {
2016 netif_err(efx, hw, efx->net_dev,
2017 "PTP out of sequence event %d\n", code);
2018 ptp->evt_frag_idx = 0;
2021 ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
2022 if (!MCDI_EVENT_FIELD(*ev, CONT)) {
2023 /* Process resulting event */
2025 case MCDI_EVENT_CODE_PTP_RX:
2026 ptp_event_rx(efx, ptp);
2028 case MCDI_EVENT_CODE_PTP_FAULT:
2029 ptp_event_fault(efx, ptp);
2031 case MCDI_EVENT_CODE_PTP_PPS:
2032 ptp_event_pps(efx, ptp);
2035 netif_err(efx, hw, efx->net_dev,
2036 "PTP unknown event %d\n", code);
2039 ptp->evt_frag_idx = 0;
2040 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
2041 netif_err(efx, hw, efx->net_dev,
2042 "PTP too many event fragments\n");
2043 ptp->evt_frag_idx = 0;
2047 void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
2049 struct efx_nic *efx = channel->efx;
2050 struct efx_ptp_data *ptp = efx->ptp_data;
2052 /* When extracting the sync timestamp minor value, we should discard
2053 * the least significant two bits. These are not required in order
2054 * to reconstruct full-range timestamps and they are optionally used
2055 * to report status depending on the options supplied when subscribing
2058 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
2059 channel->sync_timestamp_minor =
2060 (MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_MS_8BITS) & 0xFC)
2061 << ptp->nic_time.sync_event_minor_shift;
2063 /* if sync events have been disabled then we want to silently ignore
2064 * this event, so throw away result.
2066 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
2070 static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
2072 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
2073 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset));
2075 const u8 *data = eh + efx->rx_packet_ts_offset;
2076 return (u32)data[0] |
2078 (u32)data[2] << 16 |
2083 void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
2084 struct sk_buff *skb)
2086 struct efx_nic *efx = channel->efx;
2087 struct efx_ptp_data *ptp = efx->ptp_data;
2088 u32 pkt_timestamp_major, pkt_timestamp_minor;
2090 struct skb_shared_hwtstamps *timestamps;
2092 if (channel->sync_events_state != SYNC_EVENTS_VALID)
2095 pkt_timestamp_minor = efx_rx_buf_timestamp_minor(efx, skb_mac_header(skb));
2097 /* get the difference between the packet and sync timestamps,
2100 diff = pkt_timestamp_minor - channel->sync_timestamp_minor;
2101 if (pkt_timestamp_minor < channel->sync_timestamp_minor)
2102 diff += ptp->nic_time.minor_max;
2104 /* do we roll over a second boundary and need to carry the one? */
2105 carry = (channel->sync_timestamp_minor >= ptp->nic_time.minor_max - diff) ?
2108 if (diff <= ptp->nic_time.sync_event_diff_max) {
2109 /* packet is ahead of the sync event by a quarter of a second or
2110 * less (allowing for fuzz)
2112 pkt_timestamp_major = channel->sync_timestamp_major + carry;
2113 } else if (diff >= ptp->nic_time.sync_event_diff_min) {
2114 /* packet is behind the sync event but within the fuzz factor.
2115 * This means the RX packet and sync event crossed as they were
2116 * placed on the event queue, which can sometimes happen.
2118 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
2120 /* it's outside tolerance in both directions. this might be
2121 * indicative of us missing sync events for some reason, so
2122 * we'll call it an error rather than risk giving a bogus
2125 netif_vdbg(efx, drv, efx->net_dev,
2126 "packet timestamp %x too far from sync event %x:%x\n",
2127 pkt_timestamp_minor, channel->sync_timestamp_major,
2128 channel->sync_timestamp_minor);
2132 /* attach the timestamps to the skb */
2133 timestamps = skb_hwtstamps(skb);
2134 timestamps->hwtstamp =
2135 ptp->nic_to_kernel_time(pkt_timestamp_major,
2136 pkt_timestamp_minor,
2137 ptp->ts_corrections.general_rx);
2140 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
2142 struct efx_ptp_data *ptp_data = container_of(ptp,
2143 struct efx_ptp_data,
2145 struct efx_nic *efx = ptp_data->efx;
2146 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
2150 if (delta > MAX_PPB)
2152 else if (delta < -MAX_PPB)
2155 /* Convert ppb to fixed point ns taking care to round correctly. */
2156 adjustment_ns = ((s64)delta * PPB_SCALE_WORD +
2157 (1 << (ptp_data->adjfreq_ppb_shift - 1))) >>
2158 ptp_data->adjfreq_ppb_shift;
2160 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2161 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
2162 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
2163 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
2164 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
2165 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj),
2170 ptp_data->current_adjfreq = adjustment_ns;
2174 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
2176 u32 nic_major, nic_minor;
2177 struct efx_ptp_data *ptp_data = container_of(ptp,
2178 struct efx_ptp_data,
2180 struct efx_nic *efx = ptp_data->efx;
2181 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
2183 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
2185 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2186 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2187 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
2188 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
2189 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
2190 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2194 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
2196 struct efx_ptp_data *ptp_data = container_of(ptp,
2197 struct efx_ptp_data,
2199 struct efx_nic *efx = ptp_data->efx;
2200 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
2201 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
2205 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
2206 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2208 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2209 outbuf, sizeof(outbuf), NULL);
2213 kt = ptp_data->nic_to_kernel_time(
2214 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
2215 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
2216 *ts = ktime_to_timespec64(kt);
2220 static int efx_phc_settime(struct ptp_clock_info *ptp,
2221 const struct timespec64 *e_ts)
2223 /* Get the current NIC time, efx_phc_gettime.
2224 * Subtract from the desired time to get the offset
2225 * call efx_phc_adjtime with the offset
2228 struct timespec64 time_now;
2229 struct timespec64 delta;
2231 rc = efx_phc_gettime(ptp, &time_now);
2235 delta = timespec64_sub(*e_ts, time_now);
2237 rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta));
2244 static int efx_phc_enable(struct ptp_clock_info *ptp,
2245 struct ptp_clock_request *request,
2248 struct efx_ptp_data *ptp_data = container_of(ptp,
2249 struct efx_ptp_data,
2251 if (request->type != PTP_CLK_REQ_PPS)
2254 ptp_data->nic_ts_enabled = !!enable;
2258 static const struct efx_channel_type efx_ptp_channel_type = {
2259 .handle_no_channel = efx_ptp_handle_no_channel,
2260 .pre_probe = efx_ptp_probe_channel,
2261 .post_remove = efx_ptp_remove_channel,
2262 .get_name = efx_ptp_get_channel_name,
2263 .copy = efx_copy_channel,
2264 .receive_skb = efx_ptp_rx,
2265 .want_txqs = efx_ptp_want_txqs,
2266 .keep_eventq = false,
2269 void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
2271 /* Check whether PTP is implemented on this NIC. The DISABLE
2272 * operation will succeed if and only if it is implemented.
2274 if (efx_ptp_disable(efx) == 0)
2275 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
2276 &efx_ptp_channel_type;
2279 void efx_ptp_start_datapath(struct efx_nic *efx)
2281 if (efx_ptp_restart(efx))
2282 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
2283 /* re-enable timestamping if it was previously enabled */
2284 if (efx->type->ptp_set_ts_sync_events)
2285 efx->type->ptp_set_ts_sync_events(efx, true, true);
2288 void efx_ptp_stop_datapath(struct efx_nic *efx)
2290 /* temporarily disable timestamping */
2291 if (efx->type->ptp_set_ts_sync_events)
2292 efx->type->ptp_set_ts_sync_events(efx, false, true);