Loading...
Note: File does not exist in v3.1.
1/****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2011-2013 Solarflare Communications Inc.
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published
7 * by the Free Software Foundation, incorporated herein by reference.
8 */
9
10/* Theory of operation:
11 *
12 * PTP support is assisted by firmware running on the MC, which provides
13 * the hardware timestamping capabilities. Both transmitted and received
14 * PTP event packets are queued onto internal queues for subsequent processing;
15 * this is because the MC operations are relatively long and would block
16 * block NAPI/interrupt operation.
17 *
18 * Receive event processing:
19 * The event contains the packet's UUID and sequence number, together
20 * with the hardware timestamp. The PTP receive packet queue is searched
21 * for this UUID/sequence number and, if found, put on a pending queue.
22 * Packets not matching are delivered without timestamps (MCDI events will
23 * always arrive after the actual packet).
24 * It is important for the operation of the PTP protocol that the ordering
25 * of packets between the event and general port is maintained.
26 *
27 * Work queue processing:
28 * If work waiting, synchronise host/hardware time
29 *
30 * Transmit: send packet through MC, which returns the transmission time
31 * that is converted to an appropriate timestamp.
32 *
33 * Receive: the packet's reception time is converted to an appropriate
34 * timestamp.
35 */
36#include <linux/ip.h>
37#include <linux/udp.h>
38#include <linux/time.h>
39#include <linux/ktime.h>
40#include <linux/module.h>
41#include <linux/net_tstamp.h>
42#include <linux/pps_kernel.h>
43#include <linux/ptp_clock_kernel.h>
44#include "net_driver.h"
45#include "efx.h"
46#include "mcdi.h"
47#include "mcdi_pcol.h"
48#include "io.h"
49#include "farch_regs.h"
50#include "nic.h"
51
52/* Maximum number of events expected to make up a PTP event */
53#define MAX_EVENT_FRAGS 3
54
55/* Maximum delay, ms, to begin synchronisation */
56#define MAX_SYNCHRONISE_WAIT_MS 2
57
58/* How long, at most, to spend synchronising */
59#define SYNCHRONISE_PERIOD_NS 250000
60
61/* How often to update the shared memory time */
62#define SYNCHRONISATION_GRANULARITY_NS 200
63
64/* Minimum permitted length of a (corrected) synchronisation time */
65#define DEFAULT_MIN_SYNCHRONISATION_NS 120
66
67/* Maximum permitted length of a (corrected) synchronisation time */
68#define MAX_SYNCHRONISATION_NS 1000
69
70/* How many (MC) receive events that can be queued */
71#define MAX_RECEIVE_EVENTS 8
72
73/* Length of (modified) moving average. */
74#define AVERAGE_LENGTH 16
75
76/* How long an unmatched event or packet can be held */
77#define PKT_EVENT_LIFETIME_MS 10
78
79/* Offsets into PTP packet for identification. These offsets are from the
80 * start of the IP header, not the MAC header. Note that neither PTP V1 nor
81 * PTP V2 permit the use of IPV4 options.
82 */
83#define PTP_DPORT_OFFSET 22
84
85#define PTP_V1_VERSION_LENGTH 2
86#define PTP_V1_VERSION_OFFSET 28
87
88#define PTP_V1_UUID_LENGTH 6
89#define PTP_V1_UUID_OFFSET 50
90
91#define PTP_V1_SEQUENCE_LENGTH 2
92#define PTP_V1_SEQUENCE_OFFSET 58
93
94/* The minimum length of a PTP V1 packet for offsets, etc. to be valid:
95 * includes IP header.
96 */
97#define PTP_V1_MIN_LENGTH 64
98
99#define PTP_V2_VERSION_LENGTH 1
100#define PTP_V2_VERSION_OFFSET 29
101
102#define PTP_V2_UUID_LENGTH 8
103#define PTP_V2_UUID_OFFSET 48
104
105/* Although PTP V2 UUIDs are comprised a ClockIdentity (8) and PortNumber (2),
106 * the MC only captures the last six bytes of the clock identity. These values
107 * reflect those, not the ones used in the standard. The standard permits
108 * mapping of V1 UUIDs to V2 UUIDs with these same values.
109 */
110#define PTP_V2_MC_UUID_LENGTH 6
111#define PTP_V2_MC_UUID_OFFSET 50
112
113#define PTP_V2_SEQUENCE_LENGTH 2
114#define PTP_V2_SEQUENCE_OFFSET 58
115
116/* The minimum length of a PTP V2 packet for offsets, etc. to be valid:
117 * includes IP header.
118 */
119#define PTP_V2_MIN_LENGTH 63
120
121#define PTP_MIN_LENGTH 63
122
123#define PTP_ADDRESS 0xe0000181 /* 224.0.1.129 */
124#define PTP_EVENT_PORT 319
125#define PTP_GENERAL_PORT 320
126
127/* Annoyingly the format of the version numbers are different between
128 * versions 1 and 2 so it isn't possible to simply look for 1 or 2.
129 */
130#define PTP_VERSION_V1 1
131
132#define PTP_VERSION_V2 2
133#define PTP_VERSION_V2_MASK 0x0f
134
135enum ptp_packet_state {
136 PTP_PACKET_STATE_UNMATCHED = 0,
137 PTP_PACKET_STATE_MATCHED,
138 PTP_PACKET_STATE_TIMED_OUT,
139 PTP_PACKET_STATE_MATCH_UNWANTED
140};
141
142/* NIC synchronised with single word of time only comprising
143 * partial seconds and full nanoseconds: 10^9 ~ 2^30 so 2 bits for seconds.
144 */
145#define MC_NANOSECOND_BITS 30
146#define MC_NANOSECOND_MASK ((1 << MC_NANOSECOND_BITS) - 1)
147#define MC_SECOND_MASK ((1 << (32 - MC_NANOSECOND_BITS)) - 1)
148
149/* Maximum parts-per-billion adjustment that is acceptable */
150#define MAX_PPB 1000000
151
152/* Precalculate scale word to avoid long long division at runtime */
153/* This is equivalent to 2^66 / 10^9. */
154#define PPB_SCALE_WORD ((1LL << (57)) / 1953125LL)
155
156/* How much to shift down after scaling to convert to FP40 */
157#define PPB_SHIFT_FP40 26
158/* ... and FP44. */
159#define PPB_SHIFT_FP44 22
160
161#define PTP_SYNC_ATTEMPTS 4
162
163/**
164 * struct efx_ptp_match - Matching structure, stored in sk_buff's cb area.
165 * @words: UUID and (partial) sequence number
166 * @expiry: Time after which the packet should be delivered irrespective of
167 * event arrival.
168 * @state: The state of the packet - whether it is ready for processing or
169 * whether that is of no interest.
170 */
171struct efx_ptp_match {
172 u32 words[DIV_ROUND_UP(PTP_V1_UUID_LENGTH, 4)];
173 unsigned long expiry;
174 enum ptp_packet_state state;
175};
176
177/**
178 * struct efx_ptp_event_rx - A PTP receive event (from MC)
179 * @seq0: First part of (PTP) UUID
180 * @seq1: Second part of (PTP) UUID and sequence number
181 * @hwtimestamp: Event timestamp
182 */
183struct efx_ptp_event_rx {
184 struct list_head link;
185 u32 seq0;
186 u32 seq1;
187 ktime_t hwtimestamp;
188 unsigned long expiry;
189};
190
191/**
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
201 */
202struct efx_ptp_timeset {
203 u32 host_start;
204 u32 major;
205 u32 minor;
206 u32 host_end;
207 u32 wait;
208 u32 window; /* Derived: end - start, allowing for wrap */
209};
210
211/**
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
216 * separate events)
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
224 * @work: Work task
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 * @config: Current timestamp configuration
230 * @enabled: PTP operation enabled
231 * @mode: Mode in which PTP operating (PTP version)
232 * @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time
233 * @nic_to_kernel_time: Function to convert from NIC to kernel time
234 * @nic_time.minor_max: Wrap point for NIC minor times
235 * @nic_time.sync_event_diff_min: Minimum acceptable difference between time
236 * in packet prefix and last MCDI time sync event i.e. how much earlier than
237 * the last sync event time a packet timestamp can be.
238 * @nic_time.sync_event_diff_max: Maximum acceptable difference between time
239 * in packet prefix and last MCDI time sync event i.e. how much later than
240 * the last sync event time a packet timestamp can be.
241 * @nic_time.sync_event_minor_shift: Shift required to make minor time from
242 * field in MCDI time sync event.
243 * @min_synchronisation_ns: Minimum acceptable corrected sync window
244 * @capabilities: Capabilities flags from the NIC
245 * @ts_corrections.ptp_tx: Required driver correction of PTP packet transmit
246 * timestamps
247 * @ts_corrections.ptp_rx: Required driver correction of PTP packet receive
248 * timestamps
249 * @ts_corrections.pps_out: PPS output error (information only)
250 * @ts_corrections.pps_in: Required driver correction of PPS input timestamps
251 * @ts_corrections.general_tx: Required driver correction of general packet
252 * transmit timestamps
253 * @ts_corrections.general_rx: Required driver correction of general packet
254 * receive timestamps
255 * @evt_frags: Partly assembled PTP events
256 * @evt_frag_idx: Current fragment number
257 * @evt_code: Last event code
258 * @start: Address at which MC indicates ready for synchronisation
259 * @host_time_pps: Host time at last PPS
260 * @adjfreq_ppb_shift: Shift required to convert scaled parts-per-billion
261 * frequency adjustment into a fixed point fractional nanosecond format.
262 * @current_adjfreq: Current ppb adjustment.
263 * @phc_clock: Pointer to registered phc device (if primary function)
264 * @phc_clock_info: Registration structure for phc device
265 * @pps_work: pps work task for handling pps events
266 * @pps_workwq: pps work queue
267 * @nic_ts_enabled: Flag indicating if NIC generated TS events are handled
268 * @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids
269 * allocations in main data path).
270 * @good_syncs: Number of successful synchronisations.
271 * @fast_syncs: Number of synchronisations requiring short delay
272 * @bad_syncs: Number of failed synchronisations.
273 * @sync_timeouts: Number of synchronisation timeouts
274 * @no_time_syncs: Number of synchronisations with no good times.
275 * @invalid_sync_windows: Number of sync windows with bad durations.
276 * @undersize_sync_windows: Number of corrected sync windows that are too small
277 * @oversize_sync_windows: Number of corrected sync windows that are too large
278 * @rx_no_timestamp: Number of packets received without a timestamp.
279 * @timeset: Last set of synchronisation statistics.
280 * @xmit_skb: Transmit SKB function.
281 */
282struct efx_ptp_data {
283 struct efx_nic *efx;
284 struct efx_channel *channel;
285 bool rx_ts_inline;
286 struct sk_buff_head rxq;
287 struct sk_buff_head txq;
288 struct list_head evt_list;
289 struct list_head evt_free_list;
290 spinlock_t evt_lock;
291 struct efx_ptp_event_rx rx_evts[MAX_RECEIVE_EVENTS];
292 struct workqueue_struct *workwq;
293 struct work_struct work;
294 bool reset_required;
295 u32 rxfilter_event;
296 u32 rxfilter_general;
297 bool rxfilter_installed;
298 struct hwtstamp_config config;
299 bool enabled;
300 unsigned int mode;
301 void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor);
302 ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor,
303 s32 correction);
304 struct {
305 u32 minor_max;
306 u32 sync_event_diff_min;
307 u32 sync_event_diff_max;
308 unsigned int sync_event_minor_shift;
309 } nic_time;
310 unsigned int min_synchronisation_ns;
311 unsigned int capabilities;
312 struct {
313 s32 ptp_tx;
314 s32 ptp_rx;
315 s32 pps_out;
316 s32 pps_in;
317 s32 general_tx;
318 s32 general_rx;
319 } ts_corrections;
320 efx_qword_t evt_frags[MAX_EVENT_FRAGS];
321 int evt_frag_idx;
322 int evt_code;
323 struct efx_buffer start;
324 struct pps_event_time host_time_pps;
325 unsigned int adjfreq_ppb_shift;
326 s64 current_adjfreq;
327 struct ptp_clock *phc_clock;
328 struct ptp_clock_info phc_clock_info;
329 struct work_struct pps_work;
330 struct workqueue_struct *pps_workwq;
331 bool nic_ts_enabled;
332 _MCDI_DECLARE_BUF(txbuf, MC_CMD_PTP_IN_TRANSMIT_LENMAX);
333
334 unsigned int good_syncs;
335 unsigned int fast_syncs;
336 unsigned int bad_syncs;
337 unsigned int sync_timeouts;
338 unsigned int no_time_syncs;
339 unsigned int invalid_sync_windows;
340 unsigned int undersize_sync_windows;
341 unsigned int oversize_sync_windows;
342 unsigned int rx_no_timestamp;
343 struct efx_ptp_timeset
344 timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM];
345 void (*xmit_skb)(struct efx_nic *efx, struct sk_buff *skb);
346};
347
348static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta);
349static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta);
350static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts);
351static int efx_phc_settime(struct ptp_clock_info *ptp,
352 const struct timespec64 *e_ts);
353static int efx_phc_enable(struct ptp_clock_info *ptp,
354 struct ptp_clock_request *request, int on);
355
356bool efx_ptp_use_mac_tx_timestamps(struct efx_nic *efx)
357{
358 struct efx_ef10_nic_data *nic_data = efx->nic_data;
359
360 return ((efx_nic_rev(efx) >= EFX_REV_HUNT_A0) &&
361 (nic_data->datapath_caps2 &
362 (1 << MC_CMD_GET_CAPABILITIES_V2_OUT_TX_MAC_TIMESTAMPING_LBN)
363 ));
364}
365
366/* PTP 'extra' channel is still a traffic channel, but we only create TX queues
367 * if PTP uses MAC TX timestamps, not if PTP uses the MC directly to transmit.
368 */
369static bool efx_ptp_want_txqs(struct efx_channel *channel)
370{
371 return efx_ptp_use_mac_tx_timestamps(channel->efx);
372}
373
374#define PTP_SW_STAT(ext_name, field_name) \
375 { #ext_name, 0, offsetof(struct efx_ptp_data, field_name) }
376#define PTP_MC_STAT(ext_name, mcdi_name) \
377 { #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST }
378static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = {
379 PTP_SW_STAT(ptp_good_syncs, good_syncs),
380 PTP_SW_STAT(ptp_fast_syncs, fast_syncs),
381 PTP_SW_STAT(ptp_bad_syncs, bad_syncs),
382 PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts),
383 PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs),
384 PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows),
385 PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows),
386 PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows),
387 PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp),
388 PTP_MC_STAT(ptp_tx_timestamp_packets, TX),
389 PTP_MC_STAT(ptp_rx_timestamp_packets, RX),
390 PTP_MC_STAT(ptp_timestamp_packets, TS),
391 PTP_MC_STAT(ptp_filter_matches, FM),
392 PTP_MC_STAT(ptp_non_filter_matches, NFM),
393};
394#define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc)
395static const unsigned long efx_ptp_stat_mask[] = {
396 [0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL,
397};
398
399size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
400{
401 if (!efx->ptp_data)
402 return 0;
403
404 return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
405 efx_ptp_stat_mask, strings);
406}
407
408size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats)
409{
410 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
411 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
412 size_t i;
413 int rc;
414
415 if (!efx->ptp_data)
416 return 0;
417
418 /* Copy software statistics */
419 for (i = 0; i < PTP_STAT_COUNT; i++) {
420 if (efx_ptp_stat_desc[i].dma_width)
421 continue;
422 stats[i] = *(unsigned int *)((char *)efx->ptp_data +
423 efx_ptp_stat_desc[i].offset);
424 }
425
426 /* Fetch MC statistics. We *must* fill in all statistics or
427 * risk leaking kernel memory to userland, so if the MCDI
428 * request fails we pretend we got zeroes.
429 */
430 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS);
431 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
432 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
433 outbuf, sizeof(outbuf), NULL);
434 if (rc)
435 memset(outbuf, 0, sizeof(outbuf));
436 efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
437 efx_ptp_stat_mask,
438 stats, _MCDI_PTR(outbuf, 0), false);
439
440 return PTP_STAT_COUNT;
441}
442
443/* For Siena platforms NIC time is s and ns */
444static void efx_ptp_ns_to_s_ns(s64 ns, u32 *nic_major, u32 *nic_minor)
445{
446 struct timespec64 ts = ns_to_timespec64(ns);
447 *nic_major = (u32)ts.tv_sec;
448 *nic_minor = ts.tv_nsec;
449}
450
451static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
452 s32 correction)
453{
454 ktime_t kt = ktime_set(nic_major, nic_minor);
455 if (correction >= 0)
456 kt = ktime_add_ns(kt, (u64)correction);
457 else
458 kt = ktime_sub_ns(kt, (u64)-correction);
459 return kt;
460}
461
462/* To convert from s27 format to ns we multiply then divide by a power of 2.
463 * For the conversion from ns to s27, the operation is also converted to a
464 * multiply and shift.
465 */
466#define S27_TO_NS_SHIFT (27)
467#define NS_TO_S27_MULT (((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC)
468#define NS_TO_S27_SHIFT (63 - S27_TO_NS_SHIFT)
469#define S27_MINOR_MAX (1 << S27_TO_NS_SHIFT)
470
471/* For Huntington platforms NIC time is in seconds and fractions of a second
472 * where the minor register only uses 27 bits in units of 2^-27s.
473 */
474static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
475{
476 struct timespec64 ts = ns_to_timespec64(ns);
477 u32 maj = (u32)ts.tv_sec;
478 u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT +
479 (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT);
480
481 /* The conversion can result in the minor value exceeding the maximum.
482 * In this case, round up to the next second.
483 */
484 if (min >= S27_MINOR_MAX) {
485 min -= S27_MINOR_MAX;
486 maj++;
487 }
488
489 *nic_major = maj;
490 *nic_minor = min;
491}
492
493static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
494{
495 u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC +
496 (1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT);
497 return ktime_set(nic_major, ns);
498}
499
500static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
501 s32 correction)
502{
503 /* Apply the correction and deal with carry */
504 nic_minor += correction;
505 if ((s32)nic_minor < 0) {
506 nic_minor += S27_MINOR_MAX;
507 nic_major--;
508 } else if (nic_minor >= S27_MINOR_MAX) {
509 nic_minor -= S27_MINOR_MAX;
510 nic_major++;
511 }
512
513 return efx_ptp_s27_to_ktime(nic_major, nic_minor);
514}
515
516/* For Medford2 platforms the time is in seconds and quarter nanoseconds. */
517static void efx_ptp_ns_to_s_qns(s64 ns, u32 *nic_major, u32 *nic_minor)
518{
519 struct timespec64 ts = ns_to_timespec64(ns);
520
521 *nic_major = (u32)ts.tv_sec;
522 *nic_minor = ts.tv_nsec * 4;
523}
524
525static ktime_t efx_ptp_s_qns_to_ktime_correction(u32 nic_major, u32 nic_minor,
526 s32 correction)
527{
528 ktime_t kt;
529
530 nic_minor = DIV_ROUND_CLOSEST(nic_minor, 4);
531 correction = DIV_ROUND_CLOSEST(correction, 4);
532
533 kt = ktime_set(nic_major, nic_minor);
534
535 if (correction >= 0)
536 kt = ktime_add_ns(kt, (u64)correction);
537 else
538 kt = ktime_sub_ns(kt, (u64)-correction);
539 return kt;
540}
541
542struct efx_channel *efx_ptp_channel(struct efx_nic *efx)
543{
544 return efx->ptp_data ? efx->ptp_data->channel : NULL;
545}
546
547static u32 last_sync_timestamp_major(struct efx_nic *efx)
548{
549 struct efx_channel *channel = efx_ptp_channel(efx);
550 u32 major = 0;
551
552 if (channel)
553 major = channel->sync_timestamp_major;
554 return major;
555}
556
557/* The 8000 series and later can provide the time from the MAC, which is only
558 * 48 bits long and provides meta-information in the top 2 bits.
559 */
560static ktime_t
561efx_ptp_mac_nic_to_ktime_correction(struct efx_nic *efx,
562 struct efx_ptp_data *ptp,
563 u32 nic_major, u32 nic_minor,
564 s32 correction)
565{
566 ktime_t kt = { 0 };
567
568 if (!(nic_major & 0x80000000)) {
569 WARN_ON_ONCE(nic_major >> 16);
570 /* Use the top bits from the latest sync event. */
571 nic_major &= 0xffff;
572 nic_major |= (last_sync_timestamp_major(efx) & 0xffff0000);
573
574 kt = ptp->nic_to_kernel_time(nic_major, nic_minor,
575 correction);
576 }
577 return kt;
578}
579
580ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue)
581{
582 struct efx_nic *efx = tx_queue->efx;
583 struct efx_ptp_data *ptp = efx->ptp_data;
584 ktime_t kt;
585
586 if (efx_ptp_use_mac_tx_timestamps(efx))
587 kt = efx_ptp_mac_nic_to_ktime_correction(efx, ptp,
588 tx_queue->completed_timestamp_major,
589 tx_queue->completed_timestamp_minor,
590 ptp->ts_corrections.general_tx);
591 else
592 kt = ptp->nic_to_kernel_time(
593 tx_queue->completed_timestamp_major,
594 tx_queue->completed_timestamp_minor,
595 ptp->ts_corrections.general_tx);
596 return kt;
597}
598
599/* Get PTP attributes and set up time conversions */
600static int efx_ptp_get_attributes(struct efx_nic *efx)
601{
602 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN);
603 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN);
604 struct efx_ptp_data *ptp = efx->ptp_data;
605 int rc;
606 u32 fmt;
607 size_t out_len;
608
609 /* Get the PTP attributes. If the NIC doesn't support the operation we
610 * use the default format for compatibility with older NICs i.e.
611 * seconds and nanoseconds.
612 */
613 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES);
614 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
615 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
616 outbuf, sizeof(outbuf), &out_len);
617 if (rc == 0) {
618 fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT);
619 } else if (rc == -EINVAL) {
620 fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS;
621 } else if (rc == -EPERM) {
622 netif_info(efx, probe, efx->net_dev, "no PTP support\n");
623 return rc;
624 } else {
625 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf),
626 outbuf, sizeof(outbuf), rc);
627 return rc;
628 }
629
630 switch (fmt) {
631 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION:
632 ptp->ns_to_nic_time = efx_ptp_ns_to_s27;
633 ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction;
634 ptp->nic_time.minor_max = 1 << 27;
635 ptp->nic_time.sync_event_minor_shift = 19;
636 break;
637 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS:
638 ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns;
639 ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction;
640 ptp->nic_time.minor_max = 1000000000;
641 ptp->nic_time.sync_event_minor_shift = 22;
642 break;
643 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_QTR_NANOSECONDS:
644 ptp->ns_to_nic_time = efx_ptp_ns_to_s_qns;
645 ptp->nic_to_kernel_time = efx_ptp_s_qns_to_ktime_correction;
646 ptp->nic_time.minor_max = 4000000000UL;
647 ptp->nic_time.sync_event_minor_shift = 24;
648 break;
649 default:
650 return -ERANGE;
651 }
652
653 /* Precalculate acceptable difference between the minor time in the
654 * packet prefix and the last MCDI time sync event. We expect the
655 * packet prefix timestamp to be after of sync event by up to one
656 * sync event interval (0.25s) but we allow it to exceed this by a
657 * fuzz factor of (0.1s)
658 */
659 ptp->nic_time.sync_event_diff_min = ptp->nic_time.minor_max
660 - (ptp->nic_time.minor_max / 10);
661 ptp->nic_time.sync_event_diff_max = (ptp->nic_time.minor_max / 4)
662 + (ptp->nic_time.minor_max / 10);
663
664 /* MC_CMD_PTP_OP_GET_ATTRIBUTES has been extended twice from an older
665 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT. The function now may return
666 * a value to use for the minimum acceptable corrected synchronization
667 * window and may return further capabilities.
668 * If we have the extra information store it. For older firmware that
669 * does not implement the extended command use the default value.
670 */
671 if (rc == 0 &&
672 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_CAPABILITIES_OFST)
673 ptp->min_synchronisation_ns =
674 MCDI_DWORD(outbuf,
675 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
676 else
677 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
678
679 if (rc == 0 &&
680 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
681 ptp->capabilities = MCDI_DWORD(outbuf,
682 PTP_OUT_GET_ATTRIBUTES_CAPABILITIES);
683 else
684 ptp->capabilities = 0;
685
686 /* Set up the shift for conversion between frequency
687 * adjustments in parts-per-billion and the fixed-point
688 * fractional ns format that the adapter uses.
689 */
690 if (ptp->capabilities & (1 << MC_CMD_PTP_OUT_GET_ATTRIBUTES_FP44_FREQ_ADJ_LBN))
691 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP44;
692 else
693 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP40;
694
695 return 0;
696}
697
698/* Get PTP timestamp corrections */
699static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
700{
701 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN);
702 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN);
703 int rc;
704 size_t out_len;
705
706 /* Get the timestamp corrections from the NIC. If this operation is
707 * not supported (older NICs) then no correction is required.
708 */
709 MCDI_SET_DWORD(inbuf, PTP_IN_OP,
710 MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS);
711 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
712
713 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
714 outbuf, sizeof(outbuf), &out_len);
715 if (rc == 0) {
716 efx->ptp_data->ts_corrections.ptp_tx = MCDI_DWORD(outbuf,
717 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT);
718 efx->ptp_data->ts_corrections.ptp_rx = MCDI_DWORD(outbuf,
719 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE);
720 efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf,
721 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT);
722 efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf,
723 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN);
724
725 if (out_len >= MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN) {
726 efx->ptp_data->ts_corrections.general_tx = MCDI_DWORD(
727 outbuf,
728 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_TX);
729 efx->ptp_data->ts_corrections.general_rx = MCDI_DWORD(
730 outbuf,
731 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_RX);
732 } else {
733 efx->ptp_data->ts_corrections.general_tx =
734 efx->ptp_data->ts_corrections.ptp_tx;
735 efx->ptp_data->ts_corrections.general_rx =
736 efx->ptp_data->ts_corrections.ptp_rx;
737 }
738 } else if (rc == -EINVAL) {
739 efx->ptp_data->ts_corrections.ptp_tx = 0;
740 efx->ptp_data->ts_corrections.ptp_rx = 0;
741 efx->ptp_data->ts_corrections.pps_out = 0;
742 efx->ptp_data->ts_corrections.pps_in = 0;
743 efx->ptp_data->ts_corrections.general_tx = 0;
744 efx->ptp_data->ts_corrections.general_rx = 0;
745 } else {
746 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), outbuf,
747 sizeof(outbuf), rc);
748 return rc;
749 }
750
751 return 0;
752}
753
754/* Enable MCDI PTP support. */
755static int efx_ptp_enable(struct efx_nic *efx)
756{
757 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
758 MCDI_DECLARE_BUF_ERR(outbuf);
759 int rc;
760
761 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE);
762 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
763 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE,
764 efx->ptp_data->channel ?
765 efx->ptp_data->channel->channel : 0);
766 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode);
767
768 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
769 outbuf, sizeof(outbuf), NULL);
770 rc = (rc == -EALREADY) ? 0 : rc;
771 if (rc)
772 efx_mcdi_display_error(efx, MC_CMD_PTP,
773 MC_CMD_PTP_IN_ENABLE_LEN,
774 outbuf, sizeof(outbuf), rc);
775 return rc;
776}
777
778/* Disable MCDI PTP support.
779 *
780 * Note that this function should never rely on the presence of ptp_data -
781 * may be called before that exists.
782 */
783static int efx_ptp_disable(struct efx_nic *efx)
784{
785 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
786 MCDI_DECLARE_BUF_ERR(outbuf);
787 int rc;
788
789 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE);
790 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
791 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
792 outbuf, sizeof(outbuf), NULL);
793 rc = (rc == -EALREADY) ? 0 : rc;
794 /* If we get ENOSYS, the NIC doesn't support PTP, and thus this function
795 * should only have been called during probe.
796 */
797 if (rc == -ENOSYS || rc == -EPERM)
798 netif_info(efx, probe, efx->net_dev, "no PTP support\n");
799 else if (rc)
800 efx_mcdi_display_error(efx, MC_CMD_PTP,
801 MC_CMD_PTP_IN_DISABLE_LEN,
802 outbuf, sizeof(outbuf), rc);
803 return rc;
804}
805
806static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
807{
808 struct sk_buff *skb;
809
810 while ((skb = skb_dequeue(q))) {
811 local_bh_disable();
812 netif_receive_skb(skb);
813 local_bh_enable();
814 }
815}
816
817static void efx_ptp_handle_no_channel(struct efx_nic *efx)
818{
819 netif_err(efx, drv, efx->net_dev,
820 "ERROR: PTP requires MSI-X and 1 additional interrupt"
821 "vector. PTP disabled\n");
822}
823
824/* Repeatedly send the host time to the MC which will capture the hardware
825 * time.
826 */
827static void efx_ptp_send_times(struct efx_nic *efx,
828 struct pps_event_time *last_time)
829{
830 struct pps_event_time now;
831 struct timespec64 limit;
832 struct efx_ptp_data *ptp = efx->ptp_data;
833 int *mc_running = ptp->start.addr;
834
835 pps_get_ts(&now);
836 limit = now.ts_real;
837 timespec64_add_ns(&limit, SYNCHRONISE_PERIOD_NS);
838
839 /* Write host time for specified period or until MC is done */
840 while ((timespec64_compare(&now.ts_real, &limit) < 0) &&
841 READ_ONCE(*mc_running)) {
842 struct timespec64 update_time;
843 unsigned int host_time;
844
845 /* Don't update continuously to avoid saturating the PCIe bus */
846 update_time = now.ts_real;
847 timespec64_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS);
848 do {
849 pps_get_ts(&now);
850 } while ((timespec64_compare(&now.ts_real, &update_time) < 0) &&
851 READ_ONCE(*mc_running));
852
853 /* Synchronise NIC with single word of time only */
854 host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS |
855 now.ts_real.tv_nsec);
856 /* Update host time in NIC memory */
857 efx->type->ptp_write_host_time(efx, host_time);
858 }
859 *last_time = now;
860}
861
862/* Read a timeset from the MC's results and partial process. */
863static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data),
864 struct efx_ptp_timeset *timeset)
865{
866 unsigned start_ns, end_ns;
867
868 timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART);
869 timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR);
870 timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR);
871 timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND),
872 timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
873
874 /* Ignore seconds */
875 start_ns = timeset->host_start & MC_NANOSECOND_MASK;
876 end_ns = timeset->host_end & MC_NANOSECOND_MASK;
877 /* Allow for rollover */
878 if (end_ns < start_ns)
879 end_ns += NSEC_PER_SEC;
880 /* Determine duration of operation */
881 timeset->window = end_ns - start_ns;
882}
883
884/* Process times received from MC.
885 *
886 * Extract times from returned results, and establish the minimum value
887 * seen. The minimum value represents the "best" possible time and events
888 * too much greater than this are rejected - the machine is, perhaps, too
889 * busy. A number of readings are taken so that, hopefully, at least one good
890 * synchronisation will be seen in the results.
891 */
892static int
893efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf),
894 size_t response_length,
895 const struct pps_event_time *last_time)
896{
897 unsigned number_readings =
898 MCDI_VAR_ARRAY_LEN(response_length,
899 PTP_OUT_SYNCHRONIZE_TIMESET);
900 unsigned i;
901 unsigned ngood = 0;
902 unsigned last_good = 0;
903 struct efx_ptp_data *ptp = efx->ptp_data;
904 u32 last_sec;
905 u32 start_sec;
906 struct timespec64 delta;
907 ktime_t mc_time;
908
909 if (number_readings == 0)
910 return -EAGAIN;
911
912 /* Read the set of results and find the last good host-MC
913 * synchronization result. The MC times when it finishes reading the
914 * host time so the corrected window time should be fairly constant
915 * for a given platform. Increment stats for any results that appear
916 * to be erroneous.
917 */
918 for (i = 0; i < number_readings; i++) {
919 s32 window, corrected;
920 struct timespec64 wait;
921
922 efx_ptp_read_timeset(
923 MCDI_ARRAY_STRUCT_PTR(synch_buf,
924 PTP_OUT_SYNCHRONIZE_TIMESET, i),
925 &ptp->timeset[i]);
926
927 wait = ktime_to_timespec64(
928 ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0));
929 window = ptp->timeset[i].window;
930 corrected = window - wait.tv_nsec;
931
932 /* We expect the uncorrected synchronization window to be at
933 * least as large as the interval between host start and end
934 * times. If it is smaller than this then this is mostly likely
935 * to be a consequence of the host's time being adjusted.
936 * Check that the corrected sync window is in a reasonable
937 * range. If it is out of range it is likely to be because an
938 * interrupt or other delay occurred between reading the system
939 * time and writing it to MC memory.
940 */
941 if (window < SYNCHRONISATION_GRANULARITY_NS) {
942 ++ptp->invalid_sync_windows;
943 } else if (corrected >= MAX_SYNCHRONISATION_NS) {
944 ++ptp->oversize_sync_windows;
945 } else if (corrected < ptp->min_synchronisation_ns) {
946 ++ptp->undersize_sync_windows;
947 } else {
948 ngood++;
949 last_good = i;
950 }
951 }
952
953 if (ngood == 0) {
954 netif_warn(efx, drv, efx->net_dev,
955 "PTP no suitable synchronisations\n");
956 return -EAGAIN;
957 }
958
959 /* Calculate delay from last good sync (host time) to last_time.
960 * It is possible that the seconds rolled over between taking
961 * the start reading and the last value written by the host. The
962 * timescales are such that a gap of more than one second is never
963 * expected. delta is *not* normalised.
964 */
965 start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS;
966 last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK;
967 if (start_sec != last_sec &&
968 ((start_sec + 1) & MC_SECOND_MASK) != last_sec) {
969 netif_warn(efx, hw, efx->net_dev,
970 "PTP bad synchronisation seconds\n");
971 return -EAGAIN;
972 }
973 delta.tv_sec = (last_sec - start_sec) & 1;
974 delta.tv_nsec =
975 last_time->ts_real.tv_nsec -
976 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
977
978 /* Convert the NIC time at last good sync into kernel time.
979 * No correction is required - this time is the output of a
980 * firmware process.
981 */
982 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
983 ptp->timeset[last_good].minor, 0);
984
985 /* Calculate delay from NIC top of second to last_time */
986 delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec;
987
988 /* Set PPS timestamp to match NIC top of second */
989 ptp->host_time_pps = *last_time;
990 pps_sub_ts(&ptp->host_time_pps, delta);
991
992 return 0;
993}
994
995/* Synchronize times between the host and the MC */
996static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
997{
998 struct efx_ptp_data *ptp = efx->ptp_data;
999 MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX);
1000 size_t response_length;
1001 int rc;
1002 unsigned long timeout;
1003 struct pps_event_time last_time = {};
1004 unsigned int loops = 0;
1005 int *start = ptp->start.addr;
1006
1007 MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE);
1008 MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0);
1009 MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS,
1010 num_readings);
1011 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
1012 ptp->start.dma_addr);
1013
1014 /* Clear flag that signals MC ready */
1015 WRITE_ONCE(*start, 0);
1016 rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf,
1017 MC_CMD_PTP_IN_SYNCHRONIZE_LEN);
1018 EFX_WARN_ON_ONCE_PARANOID(rc);
1019
1020 /* Wait for start from MCDI (or timeout) */
1021 timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS);
1022 while (!READ_ONCE(*start) && (time_before(jiffies, timeout))) {
1023 udelay(20); /* Usually start MCDI execution quickly */
1024 loops++;
1025 }
1026
1027 if (loops <= 1)
1028 ++ptp->fast_syncs;
1029 if (!time_before(jiffies, timeout))
1030 ++ptp->sync_timeouts;
1031
1032 if (READ_ONCE(*start))
1033 efx_ptp_send_times(efx, &last_time);
1034
1035 /* Collect results */
1036 rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP,
1037 MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
1038 synch_buf, sizeof(synch_buf),
1039 &response_length);
1040 if (rc == 0) {
1041 rc = efx_ptp_process_times(efx, synch_buf, response_length,
1042 &last_time);
1043 if (rc == 0)
1044 ++ptp->good_syncs;
1045 else
1046 ++ptp->no_time_syncs;
1047 }
1048
1049 /* Increment the bad syncs counter if the synchronize fails, whatever
1050 * the reason.
1051 */
1052 if (rc != 0)
1053 ++ptp->bad_syncs;
1054
1055 return rc;
1056}
1057
1058/* Transmit a PTP packet via the dedicated hardware timestamped queue. */
1059static void efx_ptp_xmit_skb_queue(struct efx_nic *efx, struct sk_buff *skb)
1060{
1061 struct efx_ptp_data *ptp_data = efx->ptp_data;
1062 struct efx_tx_queue *tx_queue;
1063 u8 type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0;
1064
1065 tx_queue = &ptp_data->channel->tx_queue[type];
1066 if (tx_queue && tx_queue->timestamping) {
1067 efx_enqueue_skb(tx_queue, skb);
1068 } else {
1069 WARN_ONCE(1, "PTP channel has no timestamped tx queue\n");
1070 dev_kfree_skb_any(skb);
1071 }
1072}
1073
1074/* Transmit a PTP packet, via the MCDI interface, to the wire. */
1075static void efx_ptp_xmit_skb_mc(struct efx_nic *efx, struct sk_buff *skb)
1076{
1077 struct efx_ptp_data *ptp_data = efx->ptp_data;
1078 struct skb_shared_hwtstamps timestamps;
1079 int rc = -EIO;
1080 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
1081 size_t len;
1082
1083 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT);
1084 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0);
1085 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len);
1086 if (skb_shinfo(skb)->nr_frags != 0) {
1087 rc = skb_linearize(skb);
1088 if (rc != 0)
1089 goto fail;
1090 }
1091
1092 if (skb->ip_summed == CHECKSUM_PARTIAL) {
1093 rc = skb_checksum_help(skb);
1094 if (rc != 0)
1095 goto fail;
1096 }
1097 skb_copy_from_linear_data(skb,
1098 MCDI_PTR(ptp_data->txbuf,
1099 PTP_IN_TRANSMIT_PACKET),
1100 skb->len);
1101 rc = efx_mcdi_rpc(efx, MC_CMD_PTP,
1102 ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len),
1103 txtime, sizeof(txtime), &len);
1104 if (rc != 0)
1105 goto fail;
1106
1107 memset(×tamps, 0, sizeof(timestamps));
1108 timestamps.hwtstamp = ptp_data->nic_to_kernel_time(
1109 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR),
1110 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR),
1111 ptp_data->ts_corrections.ptp_tx);
1112
1113 skb_tstamp_tx(skb, ×tamps);
1114
1115 rc = 0;
1116
1117fail:
1118 dev_kfree_skb_any(skb);
1119
1120 return;
1121}
1122
1123static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
1124{
1125 struct efx_ptp_data *ptp = efx->ptp_data;
1126 struct list_head *cursor;
1127 struct list_head *next;
1128
1129 if (ptp->rx_ts_inline)
1130 return;
1131
1132 /* Drop time-expired events */
1133 spin_lock_bh(&ptp->evt_lock);
1134 if (!list_empty(&ptp->evt_list)) {
1135 list_for_each_safe(cursor, next, &ptp->evt_list) {
1136 struct efx_ptp_event_rx *evt;
1137
1138 evt = list_entry(cursor, struct efx_ptp_event_rx,
1139 link);
1140 if (time_after(jiffies, evt->expiry)) {
1141 list_move(&evt->link, &ptp->evt_free_list);
1142 netif_warn(efx, hw, efx->net_dev,
1143 "PTP rx event dropped\n");
1144 }
1145 }
1146 }
1147 spin_unlock_bh(&ptp->evt_lock);
1148}
1149
1150static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
1151 struct sk_buff *skb)
1152{
1153 struct efx_ptp_data *ptp = efx->ptp_data;
1154 bool evts_waiting;
1155 struct list_head *cursor;
1156 struct list_head *next;
1157 struct efx_ptp_match *match;
1158 enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED;
1159
1160 WARN_ON_ONCE(ptp->rx_ts_inline);
1161
1162 spin_lock_bh(&ptp->evt_lock);
1163 evts_waiting = !list_empty(&ptp->evt_list);
1164 spin_unlock_bh(&ptp->evt_lock);
1165
1166 if (!evts_waiting)
1167 return PTP_PACKET_STATE_UNMATCHED;
1168
1169 match = (struct efx_ptp_match *)skb->cb;
1170 /* Look for a matching timestamp in the event queue */
1171 spin_lock_bh(&ptp->evt_lock);
1172 list_for_each_safe(cursor, next, &ptp->evt_list) {
1173 struct efx_ptp_event_rx *evt;
1174
1175 evt = list_entry(cursor, struct efx_ptp_event_rx, link);
1176 if ((evt->seq0 == match->words[0]) &&
1177 (evt->seq1 == match->words[1])) {
1178 struct skb_shared_hwtstamps *timestamps;
1179
1180 /* Match - add in hardware timestamp */
1181 timestamps = skb_hwtstamps(skb);
1182 timestamps->hwtstamp = evt->hwtimestamp;
1183
1184 match->state = PTP_PACKET_STATE_MATCHED;
1185 rc = PTP_PACKET_STATE_MATCHED;
1186 list_move(&evt->link, &ptp->evt_free_list);
1187 break;
1188 }
1189 }
1190 spin_unlock_bh(&ptp->evt_lock);
1191
1192 return rc;
1193}
1194
1195/* Process any queued receive events and corresponding packets
1196 *
1197 * q is returned with all the packets that are ready for delivery.
1198 */
1199static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
1200{
1201 struct efx_ptp_data *ptp = efx->ptp_data;
1202 struct sk_buff *skb;
1203
1204 while ((skb = skb_dequeue(&ptp->rxq))) {
1205 struct efx_ptp_match *match;
1206
1207 match = (struct efx_ptp_match *)skb->cb;
1208 if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) {
1209 __skb_queue_tail(q, skb);
1210 } else if (efx_ptp_match_rx(efx, skb) ==
1211 PTP_PACKET_STATE_MATCHED) {
1212 __skb_queue_tail(q, skb);
1213 } else if (time_after(jiffies, match->expiry)) {
1214 match->state = PTP_PACKET_STATE_TIMED_OUT;
1215 ++ptp->rx_no_timestamp;
1216 __skb_queue_tail(q, skb);
1217 } else {
1218 /* Replace unprocessed entry and stop */
1219 skb_queue_head(&ptp->rxq, skb);
1220 break;
1221 }
1222 }
1223}
1224
1225/* Complete processing of a received packet */
1226static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1227{
1228 local_bh_disable();
1229 netif_receive_skb(skb);
1230 local_bh_enable();
1231}
1232
1233static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
1234{
1235 struct efx_ptp_data *ptp = efx->ptp_data;
1236
1237 if (ptp->rxfilter_installed) {
1238 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1239 ptp->rxfilter_general);
1240 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1241 ptp->rxfilter_event);
1242 ptp->rxfilter_installed = false;
1243 }
1244}
1245
1246static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1247{
1248 struct efx_ptp_data *ptp = efx->ptp_data;
1249 struct efx_filter_spec rxfilter;
1250 int rc;
1251
1252 if (!ptp->channel || ptp->rxfilter_installed)
1253 return 0;
1254
1255 /* Must filter on both event and general ports to ensure
1256 * that there is no packet re-ordering.
1257 */
1258 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1259 efx_rx_queue_index(
1260 efx_channel_get_rx_queue(ptp->channel)));
1261 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1262 htonl(PTP_ADDRESS),
1263 htons(PTP_EVENT_PORT));
1264 if (rc != 0)
1265 return rc;
1266
1267 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1268 if (rc < 0)
1269 return rc;
1270 ptp->rxfilter_event = rc;
1271
1272 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1273 efx_rx_queue_index(
1274 efx_channel_get_rx_queue(ptp->channel)));
1275 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1276 htonl(PTP_ADDRESS),
1277 htons(PTP_GENERAL_PORT));
1278 if (rc != 0)
1279 goto fail;
1280
1281 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1282 if (rc < 0)
1283 goto fail;
1284 ptp->rxfilter_general = rc;
1285
1286 ptp->rxfilter_installed = true;
1287 return 0;
1288
1289fail:
1290 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1291 ptp->rxfilter_event);
1292 return rc;
1293}
1294
1295static int efx_ptp_start(struct efx_nic *efx)
1296{
1297 struct efx_ptp_data *ptp = efx->ptp_data;
1298 int rc;
1299
1300 ptp->reset_required = false;
1301
1302 rc = efx_ptp_insert_multicast_filters(efx);
1303 if (rc)
1304 return rc;
1305
1306 rc = efx_ptp_enable(efx);
1307 if (rc != 0)
1308 goto fail;
1309
1310 ptp->evt_frag_idx = 0;
1311 ptp->current_adjfreq = 0;
1312
1313 return 0;
1314
1315fail:
1316 efx_ptp_remove_multicast_filters(efx);
1317 return rc;
1318}
1319
1320static int efx_ptp_stop(struct efx_nic *efx)
1321{
1322 struct efx_ptp_data *ptp = efx->ptp_data;
1323 struct list_head *cursor;
1324 struct list_head *next;
1325 int rc;
1326
1327 if (ptp == NULL)
1328 return 0;
1329
1330 rc = efx_ptp_disable(efx);
1331
1332 efx_ptp_remove_multicast_filters(efx);
1333
1334 /* Make sure RX packets are really delivered */
1335 efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq);
1336 skb_queue_purge(&efx->ptp_data->txq);
1337
1338 /* Drop any pending receive events */
1339 spin_lock_bh(&efx->ptp_data->evt_lock);
1340 list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) {
1341 list_move(cursor, &efx->ptp_data->evt_free_list);
1342 }
1343 spin_unlock_bh(&efx->ptp_data->evt_lock);
1344
1345 return rc;
1346}
1347
1348static int efx_ptp_restart(struct efx_nic *efx)
1349{
1350 if (efx->ptp_data && efx->ptp_data->enabled)
1351 return efx_ptp_start(efx);
1352 return 0;
1353}
1354
1355static void efx_ptp_pps_worker(struct work_struct *work)
1356{
1357 struct efx_ptp_data *ptp =
1358 container_of(work, struct efx_ptp_data, pps_work);
1359 struct efx_nic *efx = ptp->efx;
1360 struct ptp_clock_event ptp_evt;
1361
1362 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
1363 return;
1364
1365 ptp_evt.type = PTP_CLOCK_PPSUSR;
1366 ptp_evt.pps_times = ptp->host_time_pps;
1367 ptp_clock_event(ptp->phc_clock, &ptp_evt);
1368}
1369
1370static void efx_ptp_worker(struct work_struct *work)
1371{
1372 struct efx_ptp_data *ptp_data =
1373 container_of(work, struct efx_ptp_data, work);
1374 struct efx_nic *efx = ptp_data->efx;
1375 struct sk_buff *skb;
1376 struct sk_buff_head tempq;
1377
1378 if (ptp_data->reset_required) {
1379 efx_ptp_stop(efx);
1380 efx_ptp_start(efx);
1381 return;
1382 }
1383
1384 efx_ptp_drop_time_expired_events(efx);
1385
1386 __skb_queue_head_init(&tempq);
1387 efx_ptp_process_events(efx, &tempq);
1388
1389 while ((skb = skb_dequeue(&ptp_data->txq)))
1390 ptp_data->xmit_skb(efx, skb);
1391
1392 while ((skb = __skb_dequeue(&tempq)))
1393 efx_ptp_process_rx(efx, skb);
1394}
1395
1396static const struct ptp_clock_info efx_phc_clock_info = {
1397 .owner = THIS_MODULE,
1398 .name = "sfc",
1399 .max_adj = MAX_PPB,
1400 .n_alarm = 0,
1401 .n_ext_ts = 0,
1402 .n_per_out = 0,
1403 .n_pins = 0,
1404 .pps = 1,
1405 .adjfreq = efx_phc_adjfreq,
1406 .adjtime = efx_phc_adjtime,
1407 .gettime64 = efx_phc_gettime,
1408 .settime64 = efx_phc_settime,
1409 .enable = efx_phc_enable,
1410};
1411
1412/* Initialise PTP state. */
1413int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1414{
1415 struct efx_ptp_data *ptp;
1416 int rc = 0;
1417 unsigned int pos;
1418
1419 ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL);
1420 efx->ptp_data = ptp;
1421 if (!efx->ptp_data)
1422 return -ENOMEM;
1423
1424 ptp->efx = efx;
1425 ptp->channel = channel;
1426 ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1427
1428 rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL);
1429 if (rc != 0)
1430 goto fail1;
1431
1432 skb_queue_head_init(&ptp->rxq);
1433 skb_queue_head_init(&ptp->txq);
1434 ptp->workwq = create_singlethread_workqueue("sfc_ptp");
1435 if (!ptp->workwq) {
1436 rc = -ENOMEM;
1437 goto fail2;
1438 }
1439
1440 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1441 ptp->xmit_skb = efx_ptp_xmit_skb_queue;
1442 /* Request sync events on this channel. */
1443 channel->sync_events_state = SYNC_EVENTS_QUIESCENT;
1444 } else {
1445 ptp->xmit_skb = efx_ptp_xmit_skb_mc;
1446 }
1447
1448 INIT_WORK(&ptp->work, efx_ptp_worker);
1449 ptp->config.flags = 0;
1450 ptp->config.tx_type = HWTSTAMP_TX_OFF;
1451 ptp->config.rx_filter = HWTSTAMP_FILTER_NONE;
1452 INIT_LIST_HEAD(&ptp->evt_list);
1453 INIT_LIST_HEAD(&ptp->evt_free_list);
1454 spin_lock_init(&ptp->evt_lock);
1455 for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++)
1456 list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list);
1457
1458 /* Get the NIC PTP attributes and set up time conversions */
1459 rc = efx_ptp_get_attributes(efx);
1460 if (rc < 0)
1461 goto fail3;
1462
1463 /* Get the timestamp corrections */
1464 rc = efx_ptp_get_timestamp_corrections(efx);
1465 if (rc < 0)
1466 goto fail3;
1467
1468 if (efx->mcdi->fn_flags &
1469 (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) {
1470 ptp->phc_clock_info = efx_phc_clock_info;
1471 ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info,
1472 &efx->pci_dev->dev);
1473 if (IS_ERR(ptp->phc_clock)) {
1474 rc = PTR_ERR(ptp->phc_clock);
1475 goto fail3;
1476 } else if (ptp->phc_clock) {
1477 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1478 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
1479 if (!ptp->pps_workwq) {
1480 rc = -ENOMEM;
1481 goto fail4;
1482 }
1483 }
1484 }
1485 ptp->nic_ts_enabled = false;
1486
1487 return 0;
1488fail4:
1489 ptp_clock_unregister(efx->ptp_data->phc_clock);
1490
1491fail3:
1492 destroy_workqueue(efx->ptp_data->workwq);
1493
1494fail2:
1495 efx_nic_free_buffer(efx, &ptp->start);
1496
1497fail1:
1498 kfree(efx->ptp_data);
1499 efx->ptp_data = NULL;
1500
1501 return rc;
1502}
1503
1504/* Initialise PTP channel.
1505 *
1506 * Setting core_index to zero causes the queue to be initialised and doesn't
1507 * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
1508 */
1509static int efx_ptp_probe_channel(struct efx_channel *channel)
1510{
1511 struct efx_nic *efx = channel->efx;
1512 int rc;
1513
1514 channel->irq_moderation_us = 0;
1515 channel->rx_queue.core_index = 0;
1516
1517 rc = efx_ptp_probe(efx, channel);
1518 /* Failure to probe PTP is not fatal; this channel will just not be
1519 * used for anything.
1520 * In the case of EPERM, efx_ptp_probe will print its own message (in
1521 * efx_ptp_get_attributes()), so we don't need to.
1522 */
1523 if (rc && rc != -EPERM)
1524 netif_warn(efx, drv, efx->net_dev,
1525 "Failed to probe PTP, rc=%d\n", rc);
1526 return 0;
1527}
1528
1529void efx_ptp_remove(struct efx_nic *efx)
1530{
1531 if (!efx->ptp_data)
1532 return;
1533
1534 (void)efx_ptp_disable(efx);
1535
1536 cancel_work_sync(&efx->ptp_data->work);
1537 cancel_work_sync(&efx->ptp_data->pps_work);
1538
1539 skb_queue_purge(&efx->ptp_data->rxq);
1540 skb_queue_purge(&efx->ptp_data->txq);
1541
1542 if (efx->ptp_data->phc_clock) {
1543 destroy_workqueue(efx->ptp_data->pps_workwq);
1544 ptp_clock_unregister(efx->ptp_data->phc_clock);
1545 }
1546
1547 destroy_workqueue(efx->ptp_data->workwq);
1548
1549 efx_nic_free_buffer(efx, &efx->ptp_data->start);
1550 kfree(efx->ptp_data);
1551 efx->ptp_data = NULL;
1552}
1553
1554static void efx_ptp_remove_channel(struct efx_channel *channel)
1555{
1556 efx_ptp_remove(channel->efx);
1557}
1558
1559static void efx_ptp_get_channel_name(struct efx_channel *channel,
1560 char *buf, size_t len)
1561{
1562 snprintf(buf, len, "%s-ptp", channel->efx->name);
1563}
1564
1565/* Determine whether this packet should be processed by the PTP module
1566 * or transmitted conventionally.
1567 */
1568bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1569{
1570 return efx->ptp_data &&
1571 efx->ptp_data->enabled &&
1572 skb->len >= PTP_MIN_LENGTH &&
1573 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1574 likely(skb->protocol == htons(ETH_P_IP)) &&
1575 skb_transport_header_was_set(skb) &&
1576 skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1577 ip_hdr(skb)->protocol == IPPROTO_UDP &&
1578 skb_headlen(skb) >=
1579 skb_transport_offset(skb) + sizeof(struct udphdr) &&
1580 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1581}
1582
1583/* Receive a PTP packet. Packets are queued until the arrival of
1584 * the receive timestamp from the MC - this will probably occur after the
1585 * packet arrival because of the processing in the MC.
1586 */
1587static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1588{
1589 struct efx_nic *efx = channel->efx;
1590 struct efx_ptp_data *ptp = efx->ptp_data;
1591 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1592 u8 *match_data_012, *match_data_345;
1593 unsigned int version;
1594 u8 *data;
1595
1596 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1597
1598 /* Correct version? */
1599 if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1600 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1601 return false;
1602 }
1603 data = skb->data;
1604 version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1605 if (version != PTP_VERSION_V1) {
1606 return false;
1607 }
1608
1609 /* PTP V1 uses all six bytes of the UUID to match the packet
1610 * to the timestamp
1611 */
1612 match_data_012 = data + PTP_V1_UUID_OFFSET;
1613 match_data_345 = data + PTP_V1_UUID_OFFSET + 3;
1614 } else {
1615 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1616 return false;
1617 }
1618 data = skb->data;
1619 version = data[PTP_V2_VERSION_OFFSET];
1620 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1621 return false;
1622 }
1623
1624 /* The original V2 implementation uses bytes 2-7 of
1625 * the UUID to match the packet to the timestamp. This
1626 * discards two of the bytes of the MAC address used
1627 * to create the UUID (SF bug 33070). The PTP V2
1628 * enhanced mode fixes this issue and uses bytes 0-2
1629 * and byte 5-7 of the UUID.
1630 */
1631 match_data_345 = data + PTP_V2_UUID_OFFSET + 5;
1632 if (ptp->mode == MC_CMD_PTP_MODE_V2) {
1633 match_data_012 = data + PTP_V2_UUID_OFFSET + 2;
1634 } else {
1635 match_data_012 = data + PTP_V2_UUID_OFFSET + 0;
1636 BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED);
1637 }
1638 }
1639
1640 /* Does this packet require timestamping? */
1641 if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1642 match->state = PTP_PACKET_STATE_UNMATCHED;
1643
1644 /* We expect the sequence number to be in the same position in
1645 * the packet for PTP V1 and V2
1646 */
1647 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1648 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1649
1650 /* Extract UUID/Sequence information */
1651 match->words[0] = (match_data_012[0] |
1652 (match_data_012[1] << 8) |
1653 (match_data_012[2] << 16) |
1654 (match_data_345[0] << 24));
1655 match->words[1] = (match_data_345[1] |
1656 (match_data_345[2] << 8) |
1657 (data[PTP_V1_SEQUENCE_OFFSET +
1658 PTP_V1_SEQUENCE_LENGTH - 1] <<
1659 16));
1660 } else {
1661 match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1662 }
1663
1664 skb_queue_tail(&ptp->rxq, skb);
1665 queue_work(ptp->workwq, &ptp->work);
1666
1667 return true;
1668}
1669
1670/* Transmit a PTP packet. This has to be transmitted by the MC
1671 * itself, through an MCDI call. MCDI calls aren't permitted
1672 * in the transmit path so defer the actual transmission to a suitable worker.
1673 */
1674int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1675{
1676 struct efx_ptp_data *ptp = efx->ptp_data;
1677
1678 skb_queue_tail(&ptp->txq, skb);
1679
1680 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1681 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1682 efx_xmit_hwtstamp_pending(skb);
1683 queue_work(ptp->workwq, &ptp->work);
1684
1685 return NETDEV_TX_OK;
1686}
1687
1688int efx_ptp_get_mode(struct efx_nic *efx)
1689{
1690 return efx->ptp_data->mode;
1691}
1692
1693int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1694 unsigned int new_mode)
1695{
1696 if ((enable_wanted != efx->ptp_data->enabled) ||
1697 (enable_wanted && (efx->ptp_data->mode != new_mode))) {
1698 int rc = 0;
1699
1700 if (enable_wanted) {
1701 /* Change of mode requires disable */
1702 if (efx->ptp_data->enabled &&
1703 (efx->ptp_data->mode != new_mode)) {
1704 efx->ptp_data->enabled = false;
1705 rc = efx_ptp_stop(efx);
1706 if (rc != 0)
1707 return rc;
1708 }
1709
1710 /* Set new operating mode and establish
1711 * baseline synchronisation, which must
1712 * succeed.
1713 */
1714 efx->ptp_data->mode = new_mode;
1715 if (netif_running(efx->net_dev))
1716 rc = efx_ptp_start(efx);
1717 if (rc == 0) {
1718 rc = efx_ptp_synchronize(efx,
1719 PTP_SYNC_ATTEMPTS * 2);
1720 if (rc != 0)
1721 efx_ptp_stop(efx);
1722 }
1723 } else {
1724 rc = efx_ptp_stop(efx);
1725 }
1726
1727 if (rc != 0)
1728 return rc;
1729
1730 efx->ptp_data->enabled = enable_wanted;
1731 }
1732
1733 return 0;
1734}
1735
1736static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
1737{
1738 int rc;
1739
1740 if (init->flags)
1741 return -EINVAL;
1742
1743 if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1744 (init->tx_type != HWTSTAMP_TX_ON))
1745 return -ERANGE;
1746
1747 rc = efx->type->ptp_set_ts_config(efx, init);
1748 if (rc)
1749 return rc;
1750
1751 efx->ptp_data->config = *init;
1752 return 0;
1753}
1754
1755void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
1756{
1757 struct efx_ptp_data *ptp = efx->ptp_data;
1758 struct efx_nic *primary = efx->primary;
1759
1760 ASSERT_RTNL();
1761
1762 if (!ptp)
1763 return;
1764
1765 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1766 SOF_TIMESTAMPING_RX_HARDWARE |
1767 SOF_TIMESTAMPING_RAW_HARDWARE);
1768 /* Check licensed features. If we don't have the license for TX
1769 * timestamps, the NIC will not support them.
1770 */
1771 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1772 struct efx_ef10_nic_data *nic_data = efx->nic_data;
1773
1774 if (!(nic_data->licensed_features &
1775 (1 << LICENSED_V3_FEATURES_TX_TIMESTAMPS_LBN)))
1776 ts_info->so_timestamping &=
1777 ~SOF_TIMESTAMPING_TX_HARDWARE;
1778 }
1779 if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1780 ts_info->phc_index =
1781 ptp_clock_index(primary->ptp_data->phc_clock);
1782 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1783 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1784}
1785
1786int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1787{
1788 struct hwtstamp_config config;
1789 int rc;
1790
1791 /* Not a PTP enabled port */
1792 if (!efx->ptp_data)
1793 return -EOPNOTSUPP;
1794
1795 if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
1796 return -EFAULT;
1797
1798 rc = efx_ptp_ts_init(efx, &config);
1799 if (rc != 0)
1800 return rc;
1801
1802 return copy_to_user(ifr->ifr_data, &config, sizeof(config))
1803 ? -EFAULT : 0;
1804}
1805
1806int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1807{
1808 if (!efx->ptp_data)
1809 return -EOPNOTSUPP;
1810
1811 return copy_to_user(ifr->ifr_data, &efx->ptp_data->config,
1812 sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
1813}
1814
1815static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1816{
1817 struct efx_ptp_data *ptp = efx->ptp_data;
1818
1819 netif_err(efx, hw, efx->net_dev,
1820 "PTP unexpected event length: got %d expected %d\n",
1821 ptp->evt_frag_idx, expected_frag_len);
1822 ptp->reset_required = true;
1823 queue_work(ptp->workwq, &ptp->work);
1824}
1825
1826/* Process a completed receive event. Put it on the event queue and
1827 * start worker thread. This is required because event and their
1828 * correspoding packets may come in either order.
1829 */
1830static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
1831{
1832 struct efx_ptp_event_rx *evt = NULL;
1833
1834 if (WARN_ON_ONCE(ptp->rx_ts_inline))
1835 return;
1836
1837 if (ptp->evt_frag_idx != 3) {
1838 ptp_event_failure(efx, 3);
1839 return;
1840 }
1841
1842 spin_lock_bh(&ptp->evt_lock);
1843 if (!list_empty(&ptp->evt_free_list)) {
1844 evt = list_first_entry(&ptp->evt_free_list,
1845 struct efx_ptp_event_rx, link);
1846 list_del(&evt->link);
1847
1848 evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA);
1849 evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2],
1850 MCDI_EVENT_SRC) |
1851 (EFX_QWORD_FIELD(ptp->evt_frags[1],
1852 MCDI_EVENT_SRC) << 8) |
1853 (EFX_QWORD_FIELD(ptp->evt_frags[0],
1854 MCDI_EVENT_SRC) << 16));
1855 evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
1856 EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
1857 EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
1858 ptp->ts_corrections.ptp_rx);
1859 evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1860 list_add_tail(&evt->link, &ptp->evt_list);
1861
1862 queue_work(ptp->workwq, &ptp->work);
1863 } else if (net_ratelimit()) {
1864 /* Log a rate-limited warning message. */
1865 netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n");
1866 }
1867 spin_unlock_bh(&ptp->evt_lock);
1868}
1869
1870static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1871{
1872 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1873 if (ptp->evt_frag_idx != 1) {
1874 ptp_event_failure(efx, 1);
1875 return;
1876 }
1877
1878 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1879}
1880
1881static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1882{
1883 if (ptp->nic_ts_enabled)
1884 queue_work(ptp->pps_workwq, &ptp->pps_work);
1885}
1886
1887void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1888{
1889 struct efx_ptp_data *ptp = efx->ptp_data;
1890 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
1891
1892 if (!ptp) {
1893 if (!efx->ptp_warned) {
1894 netif_warn(efx, drv, efx->net_dev,
1895 "Received PTP event but PTP not set up\n");
1896 efx->ptp_warned = true;
1897 }
1898 return;
1899 }
1900
1901 if (!ptp->enabled)
1902 return;
1903
1904 if (ptp->evt_frag_idx == 0) {
1905 ptp->evt_code = code;
1906 } else if (ptp->evt_code != code) {
1907 netif_err(efx, hw, efx->net_dev,
1908 "PTP out of sequence event %d\n", code);
1909 ptp->evt_frag_idx = 0;
1910 }
1911
1912 ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
1913 if (!MCDI_EVENT_FIELD(*ev, CONT)) {
1914 /* Process resulting event */
1915 switch (code) {
1916 case MCDI_EVENT_CODE_PTP_RX:
1917 ptp_event_rx(efx, ptp);
1918 break;
1919 case MCDI_EVENT_CODE_PTP_FAULT:
1920 ptp_event_fault(efx, ptp);
1921 break;
1922 case MCDI_EVENT_CODE_PTP_PPS:
1923 ptp_event_pps(efx, ptp);
1924 break;
1925 default:
1926 netif_err(efx, hw, efx->net_dev,
1927 "PTP unknown event %d\n", code);
1928 break;
1929 }
1930 ptp->evt_frag_idx = 0;
1931 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
1932 netif_err(efx, hw, efx->net_dev,
1933 "PTP too many event fragments\n");
1934 ptp->evt_frag_idx = 0;
1935 }
1936}
1937
1938void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
1939{
1940 struct efx_nic *efx = channel->efx;
1941 struct efx_ptp_data *ptp = efx->ptp_data;
1942
1943 /* When extracting the sync timestamp minor value, we should discard
1944 * the least significant two bits. These are not required in order
1945 * to reconstruct full-range timestamps and they are optionally used
1946 * to report status depending on the options supplied when subscribing
1947 * for sync events.
1948 */
1949 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
1950 channel->sync_timestamp_minor =
1951 (MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_MS_8BITS) & 0xFC)
1952 << ptp->nic_time.sync_event_minor_shift;
1953
1954 /* if sync events have been disabled then we want to silently ignore
1955 * this event, so throw away result.
1956 */
1957 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
1958 SYNC_EVENTS_VALID);
1959}
1960
1961static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
1962{
1963#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
1964 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset));
1965#else
1966 const u8 *data = eh + efx->rx_packet_ts_offset;
1967 return (u32)data[0] |
1968 (u32)data[1] << 8 |
1969 (u32)data[2] << 16 |
1970 (u32)data[3] << 24;
1971#endif
1972}
1973
1974void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
1975 struct sk_buff *skb)
1976{
1977 struct efx_nic *efx = channel->efx;
1978 struct efx_ptp_data *ptp = efx->ptp_data;
1979 u32 pkt_timestamp_major, pkt_timestamp_minor;
1980 u32 diff, carry;
1981 struct skb_shared_hwtstamps *timestamps;
1982
1983 if (channel->sync_events_state != SYNC_EVENTS_VALID)
1984 return;
1985
1986 pkt_timestamp_minor = efx_rx_buf_timestamp_minor(efx, skb_mac_header(skb));
1987
1988 /* get the difference between the packet and sync timestamps,
1989 * modulo one second
1990 */
1991 diff = pkt_timestamp_minor - channel->sync_timestamp_minor;
1992 if (pkt_timestamp_minor < channel->sync_timestamp_minor)
1993 diff += ptp->nic_time.minor_max;
1994
1995 /* do we roll over a second boundary and need to carry the one? */
1996 carry = (channel->sync_timestamp_minor >= ptp->nic_time.minor_max - diff) ?
1997 1 : 0;
1998
1999 if (diff <= ptp->nic_time.sync_event_diff_max) {
2000 /* packet is ahead of the sync event by a quarter of a second or
2001 * less (allowing for fuzz)
2002 */
2003 pkt_timestamp_major = channel->sync_timestamp_major + carry;
2004 } else if (diff >= ptp->nic_time.sync_event_diff_min) {
2005 /* packet is behind the sync event but within the fuzz factor.
2006 * This means the RX packet and sync event crossed as they were
2007 * placed on the event queue, which can sometimes happen.
2008 */
2009 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
2010 } else {
2011 /* it's outside tolerance in both directions. this might be
2012 * indicative of us missing sync events for some reason, so
2013 * we'll call it an error rather than risk giving a bogus
2014 * timestamp.
2015 */
2016 netif_vdbg(efx, drv, efx->net_dev,
2017 "packet timestamp %x too far from sync event %x:%x\n",
2018 pkt_timestamp_minor, channel->sync_timestamp_major,
2019 channel->sync_timestamp_minor);
2020 return;
2021 }
2022
2023 /* attach the timestamps to the skb */
2024 timestamps = skb_hwtstamps(skb);
2025 timestamps->hwtstamp =
2026 ptp->nic_to_kernel_time(pkt_timestamp_major,
2027 pkt_timestamp_minor,
2028 ptp->ts_corrections.general_rx);
2029}
2030
2031static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
2032{
2033 struct efx_ptp_data *ptp_data = container_of(ptp,
2034 struct efx_ptp_data,
2035 phc_clock_info);
2036 struct efx_nic *efx = ptp_data->efx;
2037 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
2038 s64 adjustment_ns;
2039 int rc;
2040
2041 if (delta > MAX_PPB)
2042 delta = MAX_PPB;
2043 else if (delta < -MAX_PPB)
2044 delta = -MAX_PPB;
2045
2046 /* Convert ppb to fixed point ns taking care to round correctly. */
2047 adjustment_ns = ((s64)delta * PPB_SCALE_WORD +
2048 (1 << (ptp_data->adjfreq_ppb_shift - 1))) >>
2049 ptp_data->adjfreq_ppb_shift;
2050
2051 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2052 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
2053 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
2054 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
2055 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
2056 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj),
2057 NULL, 0, NULL);
2058 if (rc != 0)
2059 return rc;
2060
2061 ptp_data->current_adjfreq = adjustment_ns;
2062 return 0;
2063}
2064
2065static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
2066{
2067 u32 nic_major, nic_minor;
2068 struct efx_ptp_data *ptp_data = container_of(ptp,
2069 struct efx_ptp_data,
2070 phc_clock_info);
2071 struct efx_nic *efx = ptp_data->efx;
2072 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
2073
2074 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
2075
2076 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2077 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2078 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
2079 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
2080 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
2081 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2082 NULL, 0, NULL);
2083}
2084
2085static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
2086{
2087 struct efx_ptp_data *ptp_data = container_of(ptp,
2088 struct efx_ptp_data,
2089 phc_clock_info);
2090 struct efx_nic *efx = ptp_data->efx;
2091 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
2092 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
2093 int rc;
2094 ktime_t kt;
2095
2096 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
2097 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2098
2099 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2100 outbuf, sizeof(outbuf), NULL);
2101 if (rc != 0)
2102 return rc;
2103
2104 kt = ptp_data->nic_to_kernel_time(
2105 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
2106 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
2107 *ts = ktime_to_timespec64(kt);
2108 return 0;
2109}
2110
2111static int efx_phc_settime(struct ptp_clock_info *ptp,
2112 const struct timespec64 *e_ts)
2113{
2114 /* Get the current NIC time, efx_phc_gettime.
2115 * Subtract from the desired time to get the offset
2116 * call efx_phc_adjtime with the offset
2117 */
2118 int rc;
2119 struct timespec64 time_now;
2120 struct timespec64 delta;
2121
2122 rc = efx_phc_gettime(ptp, &time_now);
2123 if (rc != 0)
2124 return rc;
2125
2126 delta = timespec64_sub(*e_ts, time_now);
2127
2128 rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta));
2129 if (rc != 0)
2130 return rc;
2131
2132 return 0;
2133}
2134
2135static int efx_phc_enable(struct ptp_clock_info *ptp,
2136 struct ptp_clock_request *request,
2137 int enable)
2138{
2139 struct efx_ptp_data *ptp_data = container_of(ptp,
2140 struct efx_ptp_data,
2141 phc_clock_info);
2142 if (request->type != PTP_CLK_REQ_PPS)
2143 return -EOPNOTSUPP;
2144
2145 ptp_data->nic_ts_enabled = !!enable;
2146 return 0;
2147}
2148
2149static const struct efx_channel_type efx_ptp_channel_type = {
2150 .handle_no_channel = efx_ptp_handle_no_channel,
2151 .pre_probe = efx_ptp_probe_channel,
2152 .post_remove = efx_ptp_remove_channel,
2153 .get_name = efx_ptp_get_channel_name,
2154 /* no copy operation; there is no need to reallocate this channel */
2155 .receive_skb = efx_ptp_rx,
2156 .want_txqs = efx_ptp_want_txqs,
2157 .keep_eventq = false,
2158};
2159
2160void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
2161{
2162 /* Check whether PTP is implemented on this NIC. The DISABLE
2163 * operation will succeed if and only if it is implemented.
2164 */
2165 if (efx_ptp_disable(efx) == 0)
2166 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
2167 &efx_ptp_channel_type;
2168}
2169
2170void efx_ptp_start_datapath(struct efx_nic *efx)
2171{
2172 if (efx_ptp_restart(efx))
2173 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
2174 /* re-enable timestamping if it was previously enabled */
2175 if (efx->type->ptp_set_ts_sync_events)
2176 efx->type->ptp_set_ts_sync_events(efx, true, true);
2177}
2178
2179void efx_ptp_stop_datapath(struct efx_nic *efx)
2180{
2181 /* temporarily disable timestamping */
2182 if (efx->type->ptp_set_ts_sync_events)
2183 efx->type->ptp_set_ts_sync_events(efx, false, true);
2184 efx_ptp_stop(efx);
2185}