Loading...
Note: File does not exist in v3.1.
1// SPDX-License-Identifier: GPL-2.0-only
2/****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2011-2013 Solarflare Communications Inc.
5 */
6
7/* Theory of operation:
8 *
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.
14 *
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.
23 *
24 * Work queue processing:
25 * If work waiting, synchronise host/hardware time
26 *
27 * Transmit: send packet through MC, which returns the transmission time
28 * that is converted to an appropriate timestamp.
29 *
30 * Receive: the packet's reception time is converted to an appropriate
31 * timestamp.
32 */
33#include <linux/ip.h>
34#include <linux/udp.h>
35#include <linux/time.h>
36#include <linux/ktime.h>
37#include <linux/module.h>
38#include <linux/net_tstamp.h>
39#include <linux/pps_kernel.h>
40#include <linux/ptp_clock_kernel.h>
41#include "net_driver.h"
42#include "efx.h"
43#include "mcdi.h"
44#include "mcdi_pcol.h"
45#include "io.h"
46#include "farch_regs.h"
47#include "nic.h"
48
49/* Maximum number of events expected to make up a PTP event */
50#define MAX_EVENT_FRAGS 3
51
52/* Maximum delay, ms, to begin synchronisation */
53#define MAX_SYNCHRONISE_WAIT_MS 2
54
55/* How long, at most, to spend synchronising */
56#define SYNCHRONISE_PERIOD_NS 250000
57
58/* How often to update the shared memory time */
59#define SYNCHRONISATION_GRANULARITY_NS 200
60
61/* Minimum permitted length of a (corrected) synchronisation time */
62#define DEFAULT_MIN_SYNCHRONISATION_NS 120
63
64/* Maximum permitted length of a (corrected) synchronisation time */
65#define MAX_SYNCHRONISATION_NS 1000
66
67/* How many (MC) receive events that can be queued */
68#define MAX_RECEIVE_EVENTS 8
69
70/* Length of (modified) moving average. */
71#define AVERAGE_LENGTH 16
72
73/* How long an unmatched event or packet can be held */
74#define PKT_EVENT_LIFETIME_MS 10
75
76/* Offsets into PTP packet for identification. These offsets are from the
77 * start of the IP header, not the MAC header. Note that neither PTP V1 nor
78 * PTP V2 permit the use of IPV4 options.
79 */
80#define PTP_DPORT_OFFSET 22
81
82#define PTP_V1_VERSION_LENGTH 2
83#define PTP_V1_VERSION_OFFSET 28
84
85#define PTP_V1_UUID_LENGTH 6
86#define PTP_V1_UUID_OFFSET 50
87
88#define PTP_V1_SEQUENCE_LENGTH 2
89#define PTP_V1_SEQUENCE_OFFSET 58
90
91/* The minimum length of a PTP V1 packet for offsets, etc. to be valid:
92 * includes IP header.
93 */
94#define PTP_V1_MIN_LENGTH 64
95
96#define PTP_V2_VERSION_LENGTH 1
97#define PTP_V2_VERSION_OFFSET 29
98
99#define PTP_V2_UUID_LENGTH 8
100#define PTP_V2_UUID_OFFSET 48
101
102/* Although PTP V2 UUIDs are comprised a ClockIdentity (8) and PortNumber (2),
103 * the MC only captures the last six bytes of the clock identity. These values
104 * reflect those, not the ones used in the standard. The standard permits
105 * mapping of V1 UUIDs to V2 UUIDs with these same values.
106 */
107#define PTP_V2_MC_UUID_LENGTH 6
108#define PTP_V2_MC_UUID_OFFSET 50
109
110#define PTP_V2_SEQUENCE_LENGTH 2
111#define PTP_V2_SEQUENCE_OFFSET 58
112
113/* The minimum length of a PTP V2 packet for offsets, etc. to be valid:
114 * includes IP header.
115 */
116#define PTP_V2_MIN_LENGTH 63
117
118#define PTP_MIN_LENGTH 63
119
120#define PTP_ADDRESS 0xe0000181 /* 224.0.1.129 */
121#define PTP_EVENT_PORT 319
122#define PTP_GENERAL_PORT 320
123
124/* Annoyingly the format of the version numbers are different between
125 * versions 1 and 2 so it isn't possible to simply look for 1 or 2.
126 */
127#define PTP_VERSION_V1 1
128
129#define PTP_VERSION_V2 2
130#define PTP_VERSION_V2_MASK 0x0f
131
132enum ptp_packet_state {
133 PTP_PACKET_STATE_UNMATCHED = 0,
134 PTP_PACKET_STATE_MATCHED,
135 PTP_PACKET_STATE_TIMED_OUT,
136 PTP_PACKET_STATE_MATCH_UNWANTED
137};
138
139/* NIC synchronised with single word of time only comprising
140 * partial seconds and full nanoseconds: 10^9 ~ 2^30 so 2 bits for seconds.
141 */
142#define MC_NANOSECOND_BITS 30
143#define MC_NANOSECOND_MASK ((1 << MC_NANOSECOND_BITS) - 1)
144#define MC_SECOND_MASK ((1 << (32 - MC_NANOSECOND_BITS)) - 1)
145
146/* Maximum parts-per-billion adjustment that is acceptable */
147#define MAX_PPB 1000000
148
149/* Precalculate scale word to avoid long long division at runtime */
150/* This is equivalent to 2^66 / 10^9. */
151#define PPB_SCALE_WORD ((1LL << (57)) / 1953125LL)
152
153/* How much to shift down after scaling to convert to FP40 */
154#define PPB_SHIFT_FP40 26
155/* ... and FP44. */
156#define PPB_SHIFT_FP44 22
157
158#define PTP_SYNC_ATTEMPTS 4
159
160/**
161 * struct efx_ptp_match - Matching structure, stored in sk_buff's cb area.
162 * @words: UUID and (partial) sequence number
163 * @expiry: Time after which the packet should be delivered irrespective of
164 * event arrival.
165 * @state: The state of the packet - whether it is ready for processing or
166 * whether that is of no interest.
167 */
168struct efx_ptp_match {
169 u32 words[DIV_ROUND_UP(PTP_V1_UUID_LENGTH, 4)];
170 unsigned long expiry;
171 enum ptp_packet_state state;
172};
173
174/**
175 * struct efx_ptp_event_rx - A PTP receive event (from MC)
176 * @seq0: First part of (PTP) UUID
177 * @seq1: Second part of (PTP) UUID and sequence number
178 * @hwtimestamp: Event timestamp
179 */
180struct efx_ptp_event_rx {
181 struct list_head link;
182 u32 seq0;
183 u32 seq1;
184 ktime_t hwtimestamp;
185 unsigned long expiry;
186};
187
188/**
189 * struct efx_ptp_timeset - Synchronisation between host and MC
190 * @host_start: Host time immediately before hardware timestamp taken
191 * @major: Hardware timestamp, major
192 * @minor: Hardware timestamp, minor
193 * @host_end: Host time immediately after hardware timestamp taken
194 * @wait: Number of NIC clock ticks between hardware timestamp being read and
195 * host end time being seen
196 * @window: Difference of host_end and host_start
197 * @valid: Whether this timeset is valid
198 */
199struct efx_ptp_timeset {
200 u32 host_start;
201 u32 major;
202 u32 minor;
203 u32 host_end;
204 u32 wait;
205 u32 window; /* Derived: end - start, allowing for wrap */
206};
207
208/**
209 * struct efx_ptp_data - Precision Time Protocol (PTP) state
210 * @efx: The NIC context
211 * @channel: The PTP channel (Siena only)
212 * @rx_ts_inline: Flag for whether RX timestamps are inline (else they are
213 * separate events)
214 * @rxq: Receive SKB queue (awaiting timestamps)
215 * @txq: Transmit SKB queue
216 * @evt_list: List of MC receive events awaiting packets
217 * @evt_free_list: List of free events
218 * @evt_lock: Lock for manipulating evt_list and evt_free_list
219 * @rx_evts: Instantiated events (on evt_list and evt_free_list)
220 * @workwq: Work queue for processing pending PTP operations
221 * @work: Work task
222 * @reset_required: A serious error has occurred and the PTP task needs to be
223 * reset (disable, enable).
224 * @rxfilter_event: Receive filter when operating
225 * @rxfilter_general: Receive filter when operating
226 * @config: Current timestamp configuration
227 * @enabled: PTP operation enabled
228 * @mode: Mode in which PTP operating (PTP version)
229 * @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time
230 * @nic_to_kernel_time: Function to convert from NIC to kernel time
231 * @nic_time.minor_max: Wrap point for NIC minor times
232 * @nic_time.sync_event_diff_min: Minimum acceptable difference between time
233 * in packet prefix and last MCDI time sync event i.e. how much earlier than
234 * the last sync event time a packet timestamp can be.
235 * @nic_time.sync_event_diff_max: Maximum acceptable difference between time
236 * in packet prefix and last MCDI time sync event i.e. how much later than
237 * the last sync event time a packet timestamp can be.
238 * @nic_time.sync_event_minor_shift: Shift required to make minor time from
239 * field in MCDI time sync event.
240 * @min_synchronisation_ns: Minimum acceptable corrected sync window
241 * @capabilities: Capabilities flags from the NIC
242 * @ts_corrections.ptp_tx: Required driver correction of PTP packet transmit
243 * timestamps
244 * @ts_corrections.ptp_rx: Required driver correction of PTP packet receive
245 * timestamps
246 * @ts_corrections.pps_out: PPS output error (information only)
247 * @ts_corrections.pps_in: Required driver correction of PPS input timestamps
248 * @ts_corrections.general_tx: Required driver correction of general packet
249 * transmit timestamps
250 * @ts_corrections.general_rx: Required driver correction of general packet
251 * receive timestamps
252 * @evt_frags: Partly assembled PTP events
253 * @evt_frag_idx: Current fragment number
254 * @evt_code: Last event code
255 * @start: Address at which MC indicates ready for synchronisation
256 * @host_time_pps: Host time at last PPS
257 * @adjfreq_ppb_shift: Shift required to convert scaled parts-per-billion
258 * frequency adjustment into a fixed point fractional nanosecond format.
259 * @current_adjfreq: Current ppb adjustment.
260 * @phc_clock: Pointer to registered phc device (if primary function)
261 * @phc_clock_info: Registration structure for phc device
262 * @pps_work: pps work task for handling pps events
263 * @pps_workwq: pps work queue
264 * @nic_ts_enabled: Flag indicating if NIC generated TS events are handled
265 * @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids
266 * allocations in main data path).
267 * @good_syncs: Number of successful synchronisations.
268 * @fast_syncs: Number of synchronisations requiring short delay
269 * @bad_syncs: Number of failed synchronisations.
270 * @sync_timeouts: Number of synchronisation timeouts
271 * @no_time_syncs: Number of synchronisations with no good times.
272 * @invalid_sync_windows: Number of sync windows with bad durations.
273 * @undersize_sync_windows: Number of corrected sync windows that are too small
274 * @oversize_sync_windows: Number of corrected sync windows that are too large
275 * @rx_no_timestamp: Number of packets received without a timestamp.
276 * @timeset: Last set of synchronisation statistics.
277 * @xmit_skb: Transmit SKB function.
278 */
279struct efx_ptp_data {
280 struct efx_nic *efx;
281 struct efx_channel *channel;
282 bool rx_ts_inline;
283 struct sk_buff_head rxq;
284 struct sk_buff_head txq;
285 struct list_head evt_list;
286 struct list_head evt_free_list;
287 spinlock_t evt_lock;
288 struct efx_ptp_event_rx rx_evts[MAX_RECEIVE_EVENTS];
289 struct workqueue_struct *workwq;
290 struct work_struct work;
291 bool reset_required;
292 u32 rxfilter_event;
293 u32 rxfilter_general;
294 bool rxfilter_installed;
295 struct hwtstamp_config config;
296 bool enabled;
297 unsigned int mode;
298 void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor);
299 ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor,
300 s32 correction);
301 struct {
302 u32 minor_max;
303 u32 sync_event_diff_min;
304 u32 sync_event_diff_max;
305 unsigned int sync_event_minor_shift;
306 } nic_time;
307 unsigned int min_synchronisation_ns;
308 unsigned int capabilities;
309 struct {
310 s32 ptp_tx;
311 s32 ptp_rx;
312 s32 pps_out;
313 s32 pps_in;
314 s32 general_tx;
315 s32 general_rx;
316 } ts_corrections;
317 efx_qword_t evt_frags[MAX_EVENT_FRAGS];
318 int evt_frag_idx;
319 int evt_code;
320 struct efx_buffer start;
321 struct pps_event_time host_time_pps;
322 unsigned int adjfreq_ppb_shift;
323 s64 current_adjfreq;
324 struct ptp_clock *phc_clock;
325 struct ptp_clock_info phc_clock_info;
326 struct work_struct pps_work;
327 struct workqueue_struct *pps_workwq;
328 bool nic_ts_enabled;
329 _MCDI_DECLARE_BUF(txbuf, MC_CMD_PTP_IN_TRANSMIT_LENMAX);
330
331 unsigned int good_syncs;
332 unsigned int fast_syncs;
333 unsigned int bad_syncs;
334 unsigned int sync_timeouts;
335 unsigned int no_time_syncs;
336 unsigned int invalid_sync_windows;
337 unsigned int undersize_sync_windows;
338 unsigned int oversize_sync_windows;
339 unsigned int rx_no_timestamp;
340 struct efx_ptp_timeset
341 timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM];
342 void (*xmit_skb)(struct efx_nic *efx, struct sk_buff *skb);
343};
344
345static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta);
346static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta);
347static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts);
348static int efx_phc_settime(struct ptp_clock_info *ptp,
349 const struct timespec64 *e_ts);
350static int efx_phc_enable(struct ptp_clock_info *ptp,
351 struct ptp_clock_request *request, int on);
352
353bool efx_ptp_use_mac_tx_timestamps(struct efx_nic *efx)
354{
355 struct efx_ef10_nic_data *nic_data = efx->nic_data;
356
357 return ((efx_nic_rev(efx) >= EFX_REV_HUNT_A0) &&
358 (nic_data->datapath_caps2 &
359 (1 << MC_CMD_GET_CAPABILITIES_V2_OUT_TX_MAC_TIMESTAMPING_LBN)
360 ));
361}
362
363/* PTP 'extra' channel is still a traffic channel, but we only create TX queues
364 * if PTP uses MAC TX timestamps, not if PTP uses the MC directly to transmit.
365 */
366static bool efx_ptp_want_txqs(struct efx_channel *channel)
367{
368 return efx_ptp_use_mac_tx_timestamps(channel->efx);
369}
370
371#define PTP_SW_STAT(ext_name, field_name) \
372 { #ext_name, 0, offsetof(struct efx_ptp_data, field_name) }
373#define PTP_MC_STAT(ext_name, mcdi_name) \
374 { #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST }
375static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = {
376 PTP_SW_STAT(ptp_good_syncs, good_syncs),
377 PTP_SW_STAT(ptp_fast_syncs, fast_syncs),
378 PTP_SW_STAT(ptp_bad_syncs, bad_syncs),
379 PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts),
380 PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs),
381 PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows),
382 PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows),
383 PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows),
384 PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp),
385 PTP_MC_STAT(ptp_tx_timestamp_packets, TX),
386 PTP_MC_STAT(ptp_rx_timestamp_packets, RX),
387 PTP_MC_STAT(ptp_timestamp_packets, TS),
388 PTP_MC_STAT(ptp_filter_matches, FM),
389 PTP_MC_STAT(ptp_non_filter_matches, NFM),
390};
391#define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc)
392static const unsigned long efx_ptp_stat_mask[] = {
393 [0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL,
394};
395
396size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
397{
398 if (!efx->ptp_data)
399 return 0;
400
401 return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
402 efx_ptp_stat_mask, strings);
403}
404
405size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats)
406{
407 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
408 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
409 size_t i;
410 int rc;
411
412 if (!efx->ptp_data)
413 return 0;
414
415 /* Copy software statistics */
416 for (i = 0; i < PTP_STAT_COUNT; i++) {
417 if (efx_ptp_stat_desc[i].dma_width)
418 continue;
419 stats[i] = *(unsigned int *)((char *)efx->ptp_data +
420 efx_ptp_stat_desc[i].offset);
421 }
422
423 /* Fetch MC statistics. We *must* fill in all statistics or
424 * risk leaking kernel memory to userland, so if the MCDI
425 * request fails we pretend we got zeroes.
426 */
427 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS);
428 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
429 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
430 outbuf, sizeof(outbuf), NULL);
431 if (rc)
432 memset(outbuf, 0, sizeof(outbuf));
433 efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT,
434 efx_ptp_stat_mask,
435 stats, _MCDI_PTR(outbuf, 0), false);
436
437 return PTP_STAT_COUNT;
438}
439
440/* For Siena platforms NIC time is s and ns */
441static void efx_ptp_ns_to_s_ns(s64 ns, u32 *nic_major, u32 *nic_minor)
442{
443 struct timespec64 ts = ns_to_timespec64(ns);
444 *nic_major = (u32)ts.tv_sec;
445 *nic_minor = ts.tv_nsec;
446}
447
448static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor,
449 s32 correction)
450{
451 ktime_t kt = ktime_set(nic_major, nic_minor);
452 if (correction >= 0)
453 kt = ktime_add_ns(kt, (u64)correction);
454 else
455 kt = ktime_sub_ns(kt, (u64)-correction);
456 return kt;
457}
458
459/* To convert from s27 format to ns we multiply then divide by a power of 2.
460 * For the conversion from ns to s27, the operation is also converted to a
461 * multiply and shift.
462 */
463#define S27_TO_NS_SHIFT (27)
464#define NS_TO_S27_MULT (((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC)
465#define NS_TO_S27_SHIFT (63 - S27_TO_NS_SHIFT)
466#define S27_MINOR_MAX (1 << S27_TO_NS_SHIFT)
467
468/* For Huntington platforms NIC time is in seconds and fractions of a second
469 * where the minor register only uses 27 bits in units of 2^-27s.
470 */
471static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
472{
473 struct timespec64 ts = ns_to_timespec64(ns);
474 u32 maj = (u32)ts.tv_sec;
475 u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT +
476 (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT);
477
478 /* The conversion can result in the minor value exceeding the maximum.
479 * In this case, round up to the next second.
480 */
481 if (min >= S27_MINOR_MAX) {
482 min -= S27_MINOR_MAX;
483 maj++;
484 }
485
486 *nic_major = maj;
487 *nic_minor = min;
488}
489
490static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
491{
492 u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC +
493 (1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT);
494 return ktime_set(nic_major, ns);
495}
496
497static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
498 s32 correction)
499{
500 /* Apply the correction and deal with carry */
501 nic_minor += correction;
502 if ((s32)nic_minor < 0) {
503 nic_minor += S27_MINOR_MAX;
504 nic_major--;
505 } else if (nic_minor >= S27_MINOR_MAX) {
506 nic_minor -= S27_MINOR_MAX;
507 nic_major++;
508 }
509
510 return efx_ptp_s27_to_ktime(nic_major, nic_minor);
511}
512
513/* For Medford2 platforms the time is in seconds and quarter nanoseconds. */
514static void efx_ptp_ns_to_s_qns(s64 ns, u32 *nic_major, u32 *nic_minor)
515{
516 struct timespec64 ts = ns_to_timespec64(ns);
517
518 *nic_major = (u32)ts.tv_sec;
519 *nic_minor = ts.tv_nsec * 4;
520}
521
522static ktime_t efx_ptp_s_qns_to_ktime_correction(u32 nic_major, u32 nic_minor,
523 s32 correction)
524{
525 ktime_t kt;
526
527 nic_minor = DIV_ROUND_CLOSEST(nic_minor, 4);
528 correction = DIV_ROUND_CLOSEST(correction, 4);
529
530 kt = ktime_set(nic_major, nic_minor);
531
532 if (correction >= 0)
533 kt = ktime_add_ns(kt, (u64)correction);
534 else
535 kt = ktime_sub_ns(kt, (u64)-correction);
536 return kt;
537}
538
539struct efx_channel *efx_ptp_channel(struct efx_nic *efx)
540{
541 return efx->ptp_data ? efx->ptp_data->channel : NULL;
542}
543
544static u32 last_sync_timestamp_major(struct efx_nic *efx)
545{
546 struct efx_channel *channel = efx_ptp_channel(efx);
547 u32 major = 0;
548
549 if (channel)
550 major = channel->sync_timestamp_major;
551 return major;
552}
553
554/* The 8000 series and later can provide the time from the MAC, which is only
555 * 48 bits long and provides meta-information in the top 2 bits.
556 */
557static ktime_t
558efx_ptp_mac_nic_to_ktime_correction(struct efx_nic *efx,
559 struct efx_ptp_data *ptp,
560 u32 nic_major, u32 nic_minor,
561 s32 correction)
562{
563 ktime_t kt = { 0 };
564
565 if (!(nic_major & 0x80000000)) {
566 WARN_ON_ONCE(nic_major >> 16);
567 /* Use the top bits from the latest sync event. */
568 nic_major &= 0xffff;
569 nic_major |= (last_sync_timestamp_major(efx) & 0xffff0000);
570
571 kt = ptp->nic_to_kernel_time(nic_major, nic_minor,
572 correction);
573 }
574 return kt;
575}
576
577ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue)
578{
579 struct efx_nic *efx = tx_queue->efx;
580 struct efx_ptp_data *ptp = efx->ptp_data;
581 ktime_t kt;
582
583 if (efx_ptp_use_mac_tx_timestamps(efx))
584 kt = efx_ptp_mac_nic_to_ktime_correction(efx, ptp,
585 tx_queue->completed_timestamp_major,
586 tx_queue->completed_timestamp_minor,
587 ptp->ts_corrections.general_tx);
588 else
589 kt = ptp->nic_to_kernel_time(
590 tx_queue->completed_timestamp_major,
591 tx_queue->completed_timestamp_minor,
592 ptp->ts_corrections.general_tx);
593 return kt;
594}
595
596/* Get PTP attributes and set up time conversions */
597static int efx_ptp_get_attributes(struct efx_nic *efx)
598{
599 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN);
600 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN);
601 struct efx_ptp_data *ptp = efx->ptp_data;
602 int rc;
603 u32 fmt;
604 size_t out_len;
605
606 /* Get the PTP attributes. If the NIC doesn't support the operation we
607 * use the default format for compatibility with older NICs i.e.
608 * seconds and nanoseconds.
609 */
610 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES);
611 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
612 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
613 outbuf, sizeof(outbuf), &out_len);
614 if (rc == 0) {
615 fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT);
616 } else if (rc == -EINVAL) {
617 fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS;
618 } else if (rc == -EPERM) {
619 netif_info(efx, probe, efx->net_dev, "no PTP support\n");
620 return rc;
621 } else {
622 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf),
623 outbuf, sizeof(outbuf), rc);
624 return rc;
625 }
626
627 switch (fmt) {
628 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION:
629 ptp->ns_to_nic_time = efx_ptp_ns_to_s27;
630 ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction;
631 ptp->nic_time.minor_max = 1 << 27;
632 ptp->nic_time.sync_event_minor_shift = 19;
633 break;
634 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS:
635 ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns;
636 ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction;
637 ptp->nic_time.minor_max = 1000000000;
638 ptp->nic_time.sync_event_minor_shift = 22;
639 break;
640 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_QTR_NANOSECONDS:
641 ptp->ns_to_nic_time = efx_ptp_ns_to_s_qns;
642 ptp->nic_to_kernel_time = efx_ptp_s_qns_to_ktime_correction;
643 ptp->nic_time.minor_max = 4000000000UL;
644 ptp->nic_time.sync_event_minor_shift = 24;
645 break;
646 default:
647 return -ERANGE;
648 }
649
650 /* Precalculate acceptable difference between the minor time in the
651 * packet prefix and the last MCDI time sync event. We expect the
652 * packet prefix timestamp to be after of sync event by up to one
653 * sync event interval (0.25s) but we allow it to exceed this by a
654 * fuzz factor of (0.1s)
655 */
656 ptp->nic_time.sync_event_diff_min = ptp->nic_time.minor_max
657 - (ptp->nic_time.minor_max / 10);
658 ptp->nic_time.sync_event_diff_max = (ptp->nic_time.minor_max / 4)
659 + (ptp->nic_time.minor_max / 10);
660
661 /* MC_CMD_PTP_OP_GET_ATTRIBUTES has been extended twice from an older
662 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT. The function now may return
663 * a value to use for the minimum acceptable corrected synchronization
664 * window and may return further capabilities.
665 * If we have the extra information store it. For older firmware that
666 * does not implement the extended command use the default value.
667 */
668 if (rc == 0 &&
669 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_CAPABILITIES_OFST)
670 ptp->min_synchronisation_ns =
671 MCDI_DWORD(outbuf,
672 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
673 else
674 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
675
676 if (rc == 0 &&
677 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
678 ptp->capabilities = MCDI_DWORD(outbuf,
679 PTP_OUT_GET_ATTRIBUTES_CAPABILITIES);
680 else
681 ptp->capabilities = 0;
682
683 /* Set up the shift for conversion between frequency
684 * adjustments in parts-per-billion and the fixed-point
685 * fractional ns format that the adapter uses.
686 */
687 if (ptp->capabilities & (1 << MC_CMD_PTP_OUT_GET_ATTRIBUTES_FP44_FREQ_ADJ_LBN))
688 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP44;
689 else
690 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP40;
691
692 return 0;
693}
694
695/* Get PTP timestamp corrections */
696static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
697{
698 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN);
699 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN);
700 int rc;
701 size_t out_len;
702
703 /* Get the timestamp corrections from the NIC. If this operation is
704 * not supported (older NICs) then no correction is required.
705 */
706 MCDI_SET_DWORD(inbuf, PTP_IN_OP,
707 MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS);
708 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
709
710 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
711 outbuf, sizeof(outbuf), &out_len);
712 if (rc == 0) {
713 efx->ptp_data->ts_corrections.ptp_tx = MCDI_DWORD(outbuf,
714 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT);
715 efx->ptp_data->ts_corrections.ptp_rx = MCDI_DWORD(outbuf,
716 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE);
717 efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf,
718 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT);
719 efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf,
720 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN);
721
722 if (out_len >= MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN) {
723 efx->ptp_data->ts_corrections.general_tx = MCDI_DWORD(
724 outbuf,
725 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_TX);
726 efx->ptp_data->ts_corrections.general_rx = MCDI_DWORD(
727 outbuf,
728 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_RX);
729 } else {
730 efx->ptp_data->ts_corrections.general_tx =
731 efx->ptp_data->ts_corrections.ptp_tx;
732 efx->ptp_data->ts_corrections.general_rx =
733 efx->ptp_data->ts_corrections.ptp_rx;
734 }
735 } else if (rc == -EINVAL) {
736 efx->ptp_data->ts_corrections.ptp_tx = 0;
737 efx->ptp_data->ts_corrections.ptp_rx = 0;
738 efx->ptp_data->ts_corrections.pps_out = 0;
739 efx->ptp_data->ts_corrections.pps_in = 0;
740 efx->ptp_data->ts_corrections.general_tx = 0;
741 efx->ptp_data->ts_corrections.general_rx = 0;
742 } else {
743 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), outbuf,
744 sizeof(outbuf), rc);
745 return rc;
746 }
747
748 return 0;
749}
750
751/* Enable MCDI PTP support. */
752static int efx_ptp_enable(struct efx_nic *efx)
753{
754 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
755 MCDI_DECLARE_BUF_ERR(outbuf);
756 int rc;
757
758 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE);
759 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
760 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE,
761 efx->ptp_data->channel ?
762 efx->ptp_data->channel->channel : 0);
763 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode);
764
765 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
766 outbuf, sizeof(outbuf), NULL);
767 rc = (rc == -EALREADY) ? 0 : rc;
768 if (rc)
769 efx_mcdi_display_error(efx, MC_CMD_PTP,
770 MC_CMD_PTP_IN_ENABLE_LEN,
771 outbuf, sizeof(outbuf), rc);
772 return rc;
773}
774
775/* Disable MCDI PTP support.
776 *
777 * Note that this function should never rely on the presence of ptp_data -
778 * may be called before that exists.
779 */
780static int efx_ptp_disable(struct efx_nic *efx)
781{
782 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
783 MCDI_DECLARE_BUF_ERR(outbuf);
784 int rc;
785
786 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE);
787 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
788 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
789 outbuf, sizeof(outbuf), NULL);
790 rc = (rc == -EALREADY) ? 0 : rc;
791 /* If we get ENOSYS, the NIC doesn't support PTP, and thus this function
792 * should only have been called during probe.
793 */
794 if (rc == -ENOSYS || rc == -EPERM)
795 netif_info(efx, probe, efx->net_dev, "no PTP support\n");
796 else if (rc)
797 efx_mcdi_display_error(efx, MC_CMD_PTP,
798 MC_CMD_PTP_IN_DISABLE_LEN,
799 outbuf, sizeof(outbuf), rc);
800 return rc;
801}
802
803static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
804{
805 struct sk_buff *skb;
806
807 while ((skb = skb_dequeue(q))) {
808 local_bh_disable();
809 netif_receive_skb(skb);
810 local_bh_enable();
811 }
812}
813
814static void efx_ptp_handle_no_channel(struct efx_nic *efx)
815{
816 netif_err(efx, drv, efx->net_dev,
817 "ERROR: PTP requires MSI-X and 1 additional interrupt"
818 "vector. PTP disabled\n");
819}
820
821/* Repeatedly send the host time to the MC which will capture the hardware
822 * time.
823 */
824static void efx_ptp_send_times(struct efx_nic *efx,
825 struct pps_event_time *last_time)
826{
827 struct pps_event_time now;
828 struct timespec64 limit;
829 struct efx_ptp_data *ptp = efx->ptp_data;
830 int *mc_running = ptp->start.addr;
831
832 pps_get_ts(&now);
833 limit = now.ts_real;
834 timespec64_add_ns(&limit, SYNCHRONISE_PERIOD_NS);
835
836 /* Write host time for specified period or until MC is done */
837 while ((timespec64_compare(&now.ts_real, &limit) < 0) &&
838 READ_ONCE(*mc_running)) {
839 struct timespec64 update_time;
840 unsigned int host_time;
841
842 /* Don't update continuously to avoid saturating the PCIe bus */
843 update_time = now.ts_real;
844 timespec64_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS);
845 do {
846 pps_get_ts(&now);
847 } while ((timespec64_compare(&now.ts_real, &update_time) < 0) &&
848 READ_ONCE(*mc_running));
849
850 /* Synchronise NIC with single word of time only */
851 host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS |
852 now.ts_real.tv_nsec);
853 /* Update host time in NIC memory */
854 efx->type->ptp_write_host_time(efx, host_time);
855 }
856 *last_time = now;
857}
858
859/* Read a timeset from the MC's results and partial process. */
860static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data),
861 struct efx_ptp_timeset *timeset)
862{
863 unsigned start_ns, end_ns;
864
865 timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART);
866 timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR);
867 timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR);
868 timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND),
869 timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
870
871 /* Ignore seconds */
872 start_ns = timeset->host_start & MC_NANOSECOND_MASK;
873 end_ns = timeset->host_end & MC_NANOSECOND_MASK;
874 /* Allow for rollover */
875 if (end_ns < start_ns)
876 end_ns += NSEC_PER_SEC;
877 /* Determine duration of operation */
878 timeset->window = end_ns - start_ns;
879}
880
881/* Process times received from MC.
882 *
883 * Extract times from returned results, and establish the minimum value
884 * seen. The minimum value represents the "best" possible time and events
885 * too much greater than this are rejected - the machine is, perhaps, too
886 * busy. A number of readings are taken so that, hopefully, at least one good
887 * synchronisation will be seen in the results.
888 */
889static int
890efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf),
891 size_t response_length,
892 const struct pps_event_time *last_time)
893{
894 unsigned number_readings =
895 MCDI_VAR_ARRAY_LEN(response_length,
896 PTP_OUT_SYNCHRONIZE_TIMESET);
897 unsigned i;
898 unsigned ngood = 0;
899 unsigned last_good = 0;
900 struct efx_ptp_data *ptp = efx->ptp_data;
901 u32 last_sec;
902 u32 start_sec;
903 struct timespec64 delta;
904 ktime_t mc_time;
905
906 if (number_readings == 0)
907 return -EAGAIN;
908
909 /* Read the set of results and find the last good host-MC
910 * synchronization result. The MC times when it finishes reading the
911 * host time so the corrected window time should be fairly constant
912 * for a given platform. Increment stats for any results that appear
913 * to be erroneous.
914 */
915 for (i = 0; i < number_readings; i++) {
916 s32 window, corrected;
917 struct timespec64 wait;
918
919 efx_ptp_read_timeset(
920 MCDI_ARRAY_STRUCT_PTR(synch_buf,
921 PTP_OUT_SYNCHRONIZE_TIMESET, i),
922 &ptp->timeset[i]);
923
924 wait = ktime_to_timespec64(
925 ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0));
926 window = ptp->timeset[i].window;
927 corrected = window - wait.tv_nsec;
928
929 /* We expect the uncorrected synchronization window to be at
930 * least as large as the interval between host start and end
931 * times. If it is smaller than this then this is mostly likely
932 * to be a consequence of the host's time being adjusted.
933 * Check that the corrected sync window is in a reasonable
934 * range. If it is out of range it is likely to be because an
935 * interrupt or other delay occurred between reading the system
936 * time and writing it to MC memory.
937 */
938 if (window < SYNCHRONISATION_GRANULARITY_NS) {
939 ++ptp->invalid_sync_windows;
940 } else if (corrected >= MAX_SYNCHRONISATION_NS) {
941 ++ptp->oversize_sync_windows;
942 } else if (corrected < ptp->min_synchronisation_ns) {
943 ++ptp->undersize_sync_windows;
944 } else {
945 ngood++;
946 last_good = i;
947 }
948 }
949
950 if (ngood == 0) {
951 netif_warn(efx, drv, efx->net_dev,
952 "PTP no suitable synchronisations\n");
953 return -EAGAIN;
954 }
955
956 /* Calculate delay from last good sync (host time) to last_time.
957 * It is possible that the seconds rolled over between taking
958 * the start reading and the last value written by the host. The
959 * timescales are such that a gap of more than one second is never
960 * expected. delta is *not* normalised.
961 */
962 start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS;
963 last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK;
964 if (start_sec != last_sec &&
965 ((start_sec + 1) & MC_SECOND_MASK) != last_sec) {
966 netif_warn(efx, hw, efx->net_dev,
967 "PTP bad synchronisation seconds\n");
968 return -EAGAIN;
969 }
970 delta.tv_sec = (last_sec - start_sec) & 1;
971 delta.tv_nsec =
972 last_time->ts_real.tv_nsec -
973 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
974
975 /* Convert the NIC time at last good sync into kernel time.
976 * No correction is required - this time is the output of a
977 * firmware process.
978 */
979 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
980 ptp->timeset[last_good].minor, 0);
981
982 /* Calculate delay from NIC top of second to last_time */
983 delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec;
984
985 /* Set PPS timestamp to match NIC top of second */
986 ptp->host_time_pps = *last_time;
987 pps_sub_ts(&ptp->host_time_pps, delta);
988
989 return 0;
990}
991
992/* Synchronize times between the host and the MC */
993static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
994{
995 struct efx_ptp_data *ptp = efx->ptp_data;
996 MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX);
997 size_t response_length;
998 int rc;
999 unsigned long timeout;
1000 struct pps_event_time last_time = {};
1001 unsigned int loops = 0;
1002 int *start = ptp->start.addr;
1003
1004 MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE);
1005 MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0);
1006 MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS,
1007 num_readings);
1008 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
1009 ptp->start.dma_addr);
1010
1011 /* Clear flag that signals MC ready */
1012 WRITE_ONCE(*start, 0);
1013 rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf,
1014 MC_CMD_PTP_IN_SYNCHRONIZE_LEN);
1015 EFX_WARN_ON_ONCE_PARANOID(rc);
1016
1017 /* Wait for start from MCDI (or timeout) */
1018 timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS);
1019 while (!READ_ONCE(*start) && (time_before(jiffies, timeout))) {
1020 udelay(20); /* Usually start MCDI execution quickly */
1021 loops++;
1022 }
1023
1024 if (loops <= 1)
1025 ++ptp->fast_syncs;
1026 if (!time_before(jiffies, timeout))
1027 ++ptp->sync_timeouts;
1028
1029 if (READ_ONCE(*start))
1030 efx_ptp_send_times(efx, &last_time);
1031
1032 /* Collect results */
1033 rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP,
1034 MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
1035 synch_buf, sizeof(synch_buf),
1036 &response_length);
1037 if (rc == 0) {
1038 rc = efx_ptp_process_times(efx, synch_buf, response_length,
1039 &last_time);
1040 if (rc == 0)
1041 ++ptp->good_syncs;
1042 else
1043 ++ptp->no_time_syncs;
1044 }
1045
1046 /* Increment the bad syncs counter if the synchronize fails, whatever
1047 * the reason.
1048 */
1049 if (rc != 0)
1050 ++ptp->bad_syncs;
1051
1052 return rc;
1053}
1054
1055/* Transmit a PTP packet via the dedicated hardware timestamped queue. */
1056static void efx_ptp_xmit_skb_queue(struct efx_nic *efx, struct sk_buff *skb)
1057{
1058 struct efx_ptp_data *ptp_data = efx->ptp_data;
1059 struct efx_tx_queue *tx_queue;
1060 u8 type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0;
1061
1062 tx_queue = &ptp_data->channel->tx_queue[type];
1063 if (tx_queue && tx_queue->timestamping) {
1064 efx_enqueue_skb(tx_queue, skb);
1065 } else {
1066 WARN_ONCE(1, "PTP channel has no timestamped tx queue\n");
1067 dev_kfree_skb_any(skb);
1068 }
1069}
1070
1071/* Transmit a PTP packet, via the MCDI interface, to the wire. */
1072static void efx_ptp_xmit_skb_mc(struct efx_nic *efx, struct sk_buff *skb)
1073{
1074 struct efx_ptp_data *ptp_data = efx->ptp_data;
1075 struct skb_shared_hwtstamps timestamps;
1076 int rc = -EIO;
1077 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
1078 size_t len;
1079
1080 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT);
1081 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0);
1082 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len);
1083 if (skb_shinfo(skb)->nr_frags != 0) {
1084 rc = skb_linearize(skb);
1085 if (rc != 0)
1086 goto fail;
1087 }
1088
1089 if (skb->ip_summed == CHECKSUM_PARTIAL) {
1090 rc = skb_checksum_help(skb);
1091 if (rc != 0)
1092 goto fail;
1093 }
1094 skb_copy_from_linear_data(skb,
1095 MCDI_PTR(ptp_data->txbuf,
1096 PTP_IN_TRANSMIT_PACKET),
1097 skb->len);
1098 rc = efx_mcdi_rpc(efx, MC_CMD_PTP,
1099 ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len),
1100 txtime, sizeof(txtime), &len);
1101 if (rc != 0)
1102 goto fail;
1103
1104 memset(×tamps, 0, sizeof(timestamps));
1105 timestamps.hwtstamp = ptp_data->nic_to_kernel_time(
1106 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR),
1107 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR),
1108 ptp_data->ts_corrections.ptp_tx);
1109
1110 skb_tstamp_tx(skb, ×tamps);
1111
1112 rc = 0;
1113
1114fail:
1115 dev_kfree_skb_any(skb);
1116
1117 return;
1118}
1119
1120static void efx_ptp_drop_time_expired_events(struct efx_nic *efx)
1121{
1122 struct efx_ptp_data *ptp = efx->ptp_data;
1123 struct list_head *cursor;
1124 struct list_head *next;
1125
1126 if (ptp->rx_ts_inline)
1127 return;
1128
1129 /* Drop time-expired events */
1130 spin_lock_bh(&ptp->evt_lock);
1131 if (!list_empty(&ptp->evt_list)) {
1132 list_for_each_safe(cursor, next, &ptp->evt_list) {
1133 struct efx_ptp_event_rx *evt;
1134
1135 evt = list_entry(cursor, struct efx_ptp_event_rx,
1136 link);
1137 if (time_after(jiffies, evt->expiry)) {
1138 list_move(&evt->link, &ptp->evt_free_list);
1139 netif_warn(efx, hw, efx->net_dev,
1140 "PTP rx event dropped\n");
1141 }
1142 }
1143 }
1144 spin_unlock_bh(&ptp->evt_lock);
1145}
1146
1147static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx,
1148 struct sk_buff *skb)
1149{
1150 struct efx_ptp_data *ptp = efx->ptp_data;
1151 bool evts_waiting;
1152 struct list_head *cursor;
1153 struct list_head *next;
1154 struct efx_ptp_match *match;
1155 enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED;
1156
1157 WARN_ON_ONCE(ptp->rx_ts_inline);
1158
1159 spin_lock_bh(&ptp->evt_lock);
1160 evts_waiting = !list_empty(&ptp->evt_list);
1161 spin_unlock_bh(&ptp->evt_lock);
1162
1163 if (!evts_waiting)
1164 return PTP_PACKET_STATE_UNMATCHED;
1165
1166 match = (struct efx_ptp_match *)skb->cb;
1167 /* Look for a matching timestamp in the event queue */
1168 spin_lock_bh(&ptp->evt_lock);
1169 list_for_each_safe(cursor, next, &ptp->evt_list) {
1170 struct efx_ptp_event_rx *evt;
1171
1172 evt = list_entry(cursor, struct efx_ptp_event_rx, link);
1173 if ((evt->seq0 == match->words[0]) &&
1174 (evt->seq1 == match->words[1])) {
1175 struct skb_shared_hwtstamps *timestamps;
1176
1177 /* Match - add in hardware timestamp */
1178 timestamps = skb_hwtstamps(skb);
1179 timestamps->hwtstamp = evt->hwtimestamp;
1180
1181 match->state = PTP_PACKET_STATE_MATCHED;
1182 rc = PTP_PACKET_STATE_MATCHED;
1183 list_move(&evt->link, &ptp->evt_free_list);
1184 break;
1185 }
1186 }
1187 spin_unlock_bh(&ptp->evt_lock);
1188
1189 return rc;
1190}
1191
1192/* Process any queued receive events and corresponding packets
1193 *
1194 * q is returned with all the packets that are ready for delivery.
1195 */
1196static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
1197{
1198 struct efx_ptp_data *ptp = efx->ptp_data;
1199 struct sk_buff *skb;
1200
1201 while ((skb = skb_dequeue(&ptp->rxq))) {
1202 struct efx_ptp_match *match;
1203
1204 match = (struct efx_ptp_match *)skb->cb;
1205 if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) {
1206 __skb_queue_tail(q, skb);
1207 } else if (efx_ptp_match_rx(efx, skb) ==
1208 PTP_PACKET_STATE_MATCHED) {
1209 __skb_queue_tail(q, skb);
1210 } else if (time_after(jiffies, match->expiry)) {
1211 match->state = PTP_PACKET_STATE_TIMED_OUT;
1212 ++ptp->rx_no_timestamp;
1213 __skb_queue_tail(q, skb);
1214 } else {
1215 /* Replace unprocessed entry and stop */
1216 skb_queue_head(&ptp->rxq, skb);
1217 break;
1218 }
1219 }
1220}
1221
1222/* Complete processing of a received packet */
1223static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1224{
1225 local_bh_disable();
1226 netif_receive_skb(skb);
1227 local_bh_enable();
1228}
1229
1230static void efx_ptp_remove_multicast_filters(struct efx_nic *efx)
1231{
1232 struct efx_ptp_data *ptp = efx->ptp_data;
1233
1234 if (ptp->rxfilter_installed) {
1235 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1236 ptp->rxfilter_general);
1237 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1238 ptp->rxfilter_event);
1239 ptp->rxfilter_installed = false;
1240 }
1241}
1242
1243static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1244{
1245 struct efx_ptp_data *ptp = efx->ptp_data;
1246 struct efx_filter_spec rxfilter;
1247 int rc;
1248
1249 if (!ptp->channel || ptp->rxfilter_installed)
1250 return 0;
1251
1252 /* Must filter on both event and general ports to ensure
1253 * that there is no packet re-ordering.
1254 */
1255 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1256 efx_rx_queue_index(
1257 efx_channel_get_rx_queue(ptp->channel)));
1258 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1259 htonl(PTP_ADDRESS),
1260 htons(PTP_EVENT_PORT));
1261 if (rc != 0)
1262 return rc;
1263
1264 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1265 if (rc < 0)
1266 return rc;
1267 ptp->rxfilter_event = rc;
1268
1269 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0,
1270 efx_rx_queue_index(
1271 efx_channel_get_rx_queue(ptp->channel)));
1272 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP,
1273 htonl(PTP_ADDRESS),
1274 htons(PTP_GENERAL_PORT));
1275 if (rc != 0)
1276 goto fail;
1277
1278 rc = efx_filter_insert_filter(efx, &rxfilter, true);
1279 if (rc < 0)
1280 goto fail;
1281 ptp->rxfilter_general = rc;
1282
1283 ptp->rxfilter_installed = true;
1284 return 0;
1285
1286fail:
1287 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
1288 ptp->rxfilter_event);
1289 return rc;
1290}
1291
1292static int efx_ptp_start(struct efx_nic *efx)
1293{
1294 struct efx_ptp_data *ptp = efx->ptp_data;
1295 int rc;
1296
1297 ptp->reset_required = false;
1298
1299 rc = efx_ptp_insert_multicast_filters(efx);
1300 if (rc)
1301 return rc;
1302
1303 rc = efx_ptp_enable(efx);
1304 if (rc != 0)
1305 goto fail;
1306
1307 ptp->evt_frag_idx = 0;
1308 ptp->current_adjfreq = 0;
1309
1310 return 0;
1311
1312fail:
1313 efx_ptp_remove_multicast_filters(efx);
1314 return rc;
1315}
1316
1317static int efx_ptp_stop(struct efx_nic *efx)
1318{
1319 struct efx_ptp_data *ptp = efx->ptp_data;
1320 struct list_head *cursor;
1321 struct list_head *next;
1322 int rc;
1323
1324 if (ptp == NULL)
1325 return 0;
1326
1327 rc = efx_ptp_disable(efx);
1328
1329 efx_ptp_remove_multicast_filters(efx);
1330
1331 /* Make sure RX packets are really delivered */
1332 efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq);
1333 skb_queue_purge(&efx->ptp_data->txq);
1334
1335 /* Drop any pending receive events */
1336 spin_lock_bh(&efx->ptp_data->evt_lock);
1337 list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) {
1338 list_move(cursor, &efx->ptp_data->evt_free_list);
1339 }
1340 spin_unlock_bh(&efx->ptp_data->evt_lock);
1341
1342 return rc;
1343}
1344
1345static int efx_ptp_restart(struct efx_nic *efx)
1346{
1347 if (efx->ptp_data && efx->ptp_data->enabled)
1348 return efx_ptp_start(efx);
1349 return 0;
1350}
1351
1352static void efx_ptp_pps_worker(struct work_struct *work)
1353{
1354 struct efx_ptp_data *ptp =
1355 container_of(work, struct efx_ptp_data, pps_work);
1356 struct efx_nic *efx = ptp->efx;
1357 struct ptp_clock_event ptp_evt;
1358
1359 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
1360 return;
1361
1362 ptp_evt.type = PTP_CLOCK_PPSUSR;
1363 ptp_evt.pps_times = ptp->host_time_pps;
1364 ptp_clock_event(ptp->phc_clock, &ptp_evt);
1365}
1366
1367static void efx_ptp_worker(struct work_struct *work)
1368{
1369 struct efx_ptp_data *ptp_data =
1370 container_of(work, struct efx_ptp_data, work);
1371 struct efx_nic *efx = ptp_data->efx;
1372 struct sk_buff *skb;
1373 struct sk_buff_head tempq;
1374
1375 if (ptp_data->reset_required) {
1376 efx_ptp_stop(efx);
1377 efx_ptp_start(efx);
1378 return;
1379 }
1380
1381 efx_ptp_drop_time_expired_events(efx);
1382
1383 __skb_queue_head_init(&tempq);
1384 efx_ptp_process_events(efx, &tempq);
1385
1386 while ((skb = skb_dequeue(&ptp_data->txq)))
1387 ptp_data->xmit_skb(efx, skb);
1388
1389 while ((skb = __skb_dequeue(&tempq)))
1390 efx_ptp_process_rx(efx, skb);
1391}
1392
1393static const struct ptp_clock_info efx_phc_clock_info = {
1394 .owner = THIS_MODULE,
1395 .name = "sfc",
1396 .max_adj = MAX_PPB,
1397 .n_alarm = 0,
1398 .n_ext_ts = 0,
1399 .n_per_out = 0,
1400 .n_pins = 0,
1401 .pps = 1,
1402 .adjfreq = efx_phc_adjfreq,
1403 .adjtime = efx_phc_adjtime,
1404 .gettime64 = efx_phc_gettime,
1405 .settime64 = efx_phc_settime,
1406 .enable = efx_phc_enable,
1407};
1408
1409/* Initialise PTP state. */
1410int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1411{
1412 struct efx_ptp_data *ptp;
1413 int rc = 0;
1414 unsigned int pos;
1415
1416 ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL);
1417 efx->ptp_data = ptp;
1418 if (!efx->ptp_data)
1419 return -ENOMEM;
1420
1421 ptp->efx = efx;
1422 ptp->channel = channel;
1423 ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1424
1425 rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL);
1426 if (rc != 0)
1427 goto fail1;
1428
1429 skb_queue_head_init(&ptp->rxq);
1430 skb_queue_head_init(&ptp->txq);
1431 ptp->workwq = create_singlethread_workqueue("sfc_ptp");
1432 if (!ptp->workwq) {
1433 rc = -ENOMEM;
1434 goto fail2;
1435 }
1436
1437 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1438 ptp->xmit_skb = efx_ptp_xmit_skb_queue;
1439 /* Request sync events on this channel. */
1440 channel->sync_events_state = SYNC_EVENTS_QUIESCENT;
1441 } else {
1442 ptp->xmit_skb = efx_ptp_xmit_skb_mc;
1443 }
1444
1445 INIT_WORK(&ptp->work, efx_ptp_worker);
1446 ptp->config.flags = 0;
1447 ptp->config.tx_type = HWTSTAMP_TX_OFF;
1448 ptp->config.rx_filter = HWTSTAMP_FILTER_NONE;
1449 INIT_LIST_HEAD(&ptp->evt_list);
1450 INIT_LIST_HEAD(&ptp->evt_free_list);
1451 spin_lock_init(&ptp->evt_lock);
1452 for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++)
1453 list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list);
1454
1455 /* Get the NIC PTP attributes and set up time conversions */
1456 rc = efx_ptp_get_attributes(efx);
1457 if (rc < 0)
1458 goto fail3;
1459
1460 /* Get the timestamp corrections */
1461 rc = efx_ptp_get_timestamp_corrections(efx);
1462 if (rc < 0)
1463 goto fail3;
1464
1465 if (efx->mcdi->fn_flags &
1466 (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) {
1467 ptp->phc_clock_info = efx_phc_clock_info;
1468 ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info,
1469 &efx->pci_dev->dev);
1470 if (IS_ERR(ptp->phc_clock)) {
1471 rc = PTR_ERR(ptp->phc_clock);
1472 goto fail3;
1473 } else if (ptp->phc_clock) {
1474 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1475 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
1476 if (!ptp->pps_workwq) {
1477 rc = -ENOMEM;
1478 goto fail4;
1479 }
1480 }
1481 }
1482 ptp->nic_ts_enabled = false;
1483
1484 return 0;
1485fail4:
1486 ptp_clock_unregister(efx->ptp_data->phc_clock);
1487
1488fail3:
1489 destroy_workqueue(efx->ptp_data->workwq);
1490
1491fail2:
1492 efx_nic_free_buffer(efx, &ptp->start);
1493
1494fail1:
1495 kfree(efx->ptp_data);
1496 efx->ptp_data = NULL;
1497
1498 return rc;
1499}
1500
1501/* Initialise PTP channel.
1502 *
1503 * Setting core_index to zero causes the queue to be initialised and doesn't
1504 * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
1505 */
1506static int efx_ptp_probe_channel(struct efx_channel *channel)
1507{
1508 struct efx_nic *efx = channel->efx;
1509 int rc;
1510
1511 channel->irq_moderation_us = 0;
1512 channel->rx_queue.core_index = 0;
1513
1514 rc = efx_ptp_probe(efx, channel);
1515 /* Failure to probe PTP is not fatal; this channel will just not be
1516 * used for anything.
1517 * In the case of EPERM, efx_ptp_probe will print its own message (in
1518 * efx_ptp_get_attributes()), so we don't need to.
1519 */
1520 if (rc && rc != -EPERM)
1521 netif_warn(efx, drv, efx->net_dev,
1522 "Failed to probe PTP, rc=%d\n", rc);
1523 return 0;
1524}
1525
1526void efx_ptp_remove(struct efx_nic *efx)
1527{
1528 if (!efx->ptp_data)
1529 return;
1530
1531 (void)efx_ptp_disable(efx);
1532
1533 cancel_work_sync(&efx->ptp_data->work);
1534 if (efx->ptp_data->pps_workwq)
1535 cancel_work_sync(&efx->ptp_data->pps_work);
1536
1537 skb_queue_purge(&efx->ptp_data->rxq);
1538 skb_queue_purge(&efx->ptp_data->txq);
1539
1540 if (efx->ptp_data->phc_clock) {
1541 destroy_workqueue(efx->ptp_data->pps_workwq);
1542 ptp_clock_unregister(efx->ptp_data->phc_clock);
1543 }
1544
1545 destroy_workqueue(efx->ptp_data->workwq);
1546
1547 efx_nic_free_buffer(efx, &efx->ptp_data->start);
1548 kfree(efx->ptp_data);
1549 efx->ptp_data = NULL;
1550}
1551
1552static void efx_ptp_remove_channel(struct efx_channel *channel)
1553{
1554 efx_ptp_remove(channel->efx);
1555}
1556
1557static void efx_ptp_get_channel_name(struct efx_channel *channel,
1558 char *buf, size_t len)
1559{
1560 snprintf(buf, len, "%s-ptp", channel->efx->name);
1561}
1562
1563/* Determine whether this packet should be processed by the PTP module
1564 * or transmitted conventionally.
1565 */
1566bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1567{
1568 return efx->ptp_data &&
1569 efx->ptp_data->enabled &&
1570 skb->len >= PTP_MIN_LENGTH &&
1571 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1572 likely(skb->protocol == htons(ETH_P_IP)) &&
1573 skb_transport_header_was_set(skb) &&
1574 skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1575 ip_hdr(skb)->protocol == IPPROTO_UDP &&
1576 skb_headlen(skb) >=
1577 skb_transport_offset(skb) + sizeof(struct udphdr) &&
1578 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1579}
1580
1581/* Receive a PTP packet. Packets are queued until the arrival of
1582 * the receive timestamp from the MC - this will probably occur after the
1583 * packet arrival because of the processing in the MC.
1584 */
1585static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1586{
1587 struct efx_nic *efx = channel->efx;
1588 struct efx_ptp_data *ptp = efx->ptp_data;
1589 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1590 u8 *match_data_012, *match_data_345;
1591 unsigned int version;
1592 u8 *data;
1593
1594 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1595
1596 /* Correct version? */
1597 if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1598 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1599 return false;
1600 }
1601 data = skb->data;
1602 version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1603 if (version != PTP_VERSION_V1) {
1604 return false;
1605 }
1606
1607 /* PTP V1 uses all six bytes of the UUID to match the packet
1608 * to the timestamp
1609 */
1610 match_data_012 = data + PTP_V1_UUID_OFFSET;
1611 match_data_345 = data + PTP_V1_UUID_OFFSET + 3;
1612 } else {
1613 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1614 return false;
1615 }
1616 data = skb->data;
1617 version = data[PTP_V2_VERSION_OFFSET];
1618 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1619 return false;
1620 }
1621
1622 /* The original V2 implementation uses bytes 2-7 of
1623 * the UUID to match the packet to the timestamp. This
1624 * discards two of the bytes of the MAC address used
1625 * to create the UUID (SF bug 33070). The PTP V2
1626 * enhanced mode fixes this issue and uses bytes 0-2
1627 * and byte 5-7 of the UUID.
1628 */
1629 match_data_345 = data + PTP_V2_UUID_OFFSET + 5;
1630 if (ptp->mode == MC_CMD_PTP_MODE_V2) {
1631 match_data_012 = data + PTP_V2_UUID_OFFSET + 2;
1632 } else {
1633 match_data_012 = data + PTP_V2_UUID_OFFSET + 0;
1634 BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED);
1635 }
1636 }
1637
1638 /* Does this packet require timestamping? */
1639 if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1640 match->state = PTP_PACKET_STATE_UNMATCHED;
1641
1642 /* We expect the sequence number to be in the same position in
1643 * the packet for PTP V1 and V2
1644 */
1645 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1646 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1647
1648 /* Extract UUID/Sequence information */
1649 match->words[0] = (match_data_012[0] |
1650 (match_data_012[1] << 8) |
1651 (match_data_012[2] << 16) |
1652 (match_data_345[0] << 24));
1653 match->words[1] = (match_data_345[1] |
1654 (match_data_345[2] << 8) |
1655 (data[PTP_V1_SEQUENCE_OFFSET +
1656 PTP_V1_SEQUENCE_LENGTH - 1] <<
1657 16));
1658 } else {
1659 match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1660 }
1661
1662 skb_queue_tail(&ptp->rxq, skb);
1663 queue_work(ptp->workwq, &ptp->work);
1664
1665 return true;
1666}
1667
1668/* Transmit a PTP packet. This has to be transmitted by the MC
1669 * itself, through an MCDI call. MCDI calls aren't permitted
1670 * in the transmit path so defer the actual transmission to a suitable worker.
1671 */
1672int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1673{
1674 struct efx_ptp_data *ptp = efx->ptp_data;
1675
1676 skb_queue_tail(&ptp->txq, skb);
1677
1678 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1679 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1680 efx_xmit_hwtstamp_pending(skb);
1681 queue_work(ptp->workwq, &ptp->work);
1682
1683 return NETDEV_TX_OK;
1684}
1685
1686int efx_ptp_get_mode(struct efx_nic *efx)
1687{
1688 return efx->ptp_data->mode;
1689}
1690
1691int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1692 unsigned int new_mode)
1693{
1694 if ((enable_wanted != efx->ptp_data->enabled) ||
1695 (enable_wanted && (efx->ptp_data->mode != new_mode))) {
1696 int rc = 0;
1697
1698 if (enable_wanted) {
1699 /* Change of mode requires disable */
1700 if (efx->ptp_data->enabled &&
1701 (efx->ptp_data->mode != new_mode)) {
1702 efx->ptp_data->enabled = false;
1703 rc = efx_ptp_stop(efx);
1704 if (rc != 0)
1705 return rc;
1706 }
1707
1708 /* Set new operating mode and establish
1709 * baseline synchronisation, which must
1710 * succeed.
1711 */
1712 efx->ptp_data->mode = new_mode;
1713 if (netif_running(efx->net_dev))
1714 rc = efx_ptp_start(efx);
1715 if (rc == 0) {
1716 rc = efx_ptp_synchronize(efx,
1717 PTP_SYNC_ATTEMPTS * 2);
1718 if (rc != 0)
1719 efx_ptp_stop(efx);
1720 }
1721 } else {
1722 rc = efx_ptp_stop(efx);
1723 }
1724
1725 if (rc != 0)
1726 return rc;
1727
1728 efx->ptp_data->enabled = enable_wanted;
1729 }
1730
1731 return 0;
1732}
1733
1734static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
1735{
1736 int rc;
1737
1738 if (init->flags)
1739 return -EINVAL;
1740
1741 if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1742 (init->tx_type != HWTSTAMP_TX_ON))
1743 return -ERANGE;
1744
1745 rc = efx->type->ptp_set_ts_config(efx, init);
1746 if (rc)
1747 return rc;
1748
1749 efx->ptp_data->config = *init;
1750 return 0;
1751}
1752
1753void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
1754{
1755 struct efx_ptp_data *ptp = efx->ptp_data;
1756 struct efx_nic *primary = efx->primary;
1757
1758 ASSERT_RTNL();
1759
1760 if (!ptp)
1761 return;
1762
1763 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1764 SOF_TIMESTAMPING_RX_HARDWARE |
1765 SOF_TIMESTAMPING_RAW_HARDWARE);
1766 /* Check licensed features. If we don't have the license for TX
1767 * timestamps, the NIC will not support them.
1768 */
1769 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1770 struct efx_ef10_nic_data *nic_data = efx->nic_data;
1771
1772 if (!(nic_data->licensed_features &
1773 (1 << LICENSED_V3_FEATURES_TX_TIMESTAMPS_LBN)))
1774 ts_info->so_timestamping &=
1775 ~SOF_TIMESTAMPING_TX_HARDWARE;
1776 }
1777 if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1778 ts_info->phc_index =
1779 ptp_clock_index(primary->ptp_data->phc_clock);
1780 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1781 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1782}
1783
1784int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1785{
1786 struct hwtstamp_config config;
1787 int rc;
1788
1789 /* Not a PTP enabled port */
1790 if (!efx->ptp_data)
1791 return -EOPNOTSUPP;
1792
1793 if (copy_from_user(&config, ifr->ifr_data, sizeof(config)))
1794 return -EFAULT;
1795
1796 rc = efx_ptp_ts_init(efx, &config);
1797 if (rc != 0)
1798 return rc;
1799
1800 return copy_to_user(ifr->ifr_data, &config, sizeof(config))
1801 ? -EFAULT : 0;
1802}
1803
1804int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1805{
1806 if (!efx->ptp_data)
1807 return -EOPNOTSUPP;
1808
1809 return copy_to_user(ifr->ifr_data, &efx->ptp_data->config,
1810 sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
1811}
1812
1813static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1814{
1815 struct efx_ptp_data *ptp = efx->ptp_data;
1816
1817 netif_err(efx, hw, efx->net_dev,
1818 "PTP unexpected event length: got %d expected %d\n",
1819 ptp->evt_frag_idx, expected_frag_len);
1820 ptp->reset_required = true;
1821 queue_work(ptp->workwq, &ptp->work);
1822}
1823
1824/* Process a completed receive event. Put it on the event queue and
1825 * start worker thread. This is required because event and their
1826 * correspoding packets may come in either order.
1827 */
1828static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp)
1829{
1830 struct efx_ptp_event_rx *evt = NULL;
1831
1832 if (WARN_ON_ONCE(ptp->rx_ts_inline))
1833 return;
1834
1835 if (ptp->evt_frag_idx != 3) {
1836 ptp_event_failure(efx, 3);
1837 return;
1838 }
1839
1840 spin_lock_bh(&ptp->evt_lock);
1841 if (!list_empty(&ptp->evt_free_list)) {
1842 evt = list_first_entry(&ptp->evt_free_list,
1843 struct efx_ptp_event_rx, link);
1844 list_del(&evt->link);
1845
1846 evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA);
1847 evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2],
1848 MCDI_EVENT_SRC) |
1849 (EFX_QWORD_FIELD(ptp->evt_frags[1],
1850 MCDI_EVENT_SRC) << 8) |
1851 (EFX_QWORD_FIELD(ptp->evt_frags[0],
1852 MCDI_EVENT_SRC) << 16));
1853 evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time(
1854 EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA),
1855 EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA),
1856 ptp->ts_corrections.ptp_rx);
1857 evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1858 list_add_tail(&evt->link, &ptp->evt_list);
1859
1860 queue_work(ptp->workwq, &ptp->work);
1861 } else if (net_ratelimit()) {
1862 /* Log a rate-limited warning message. */
1863 netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n");
1864 }
1865 spin_unlock_bh(&ptp->evt_lock);
1866}
1867
1868static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1869{
1870 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1871 if (ptp->evt_frag_idx != 1) {
1872 ptp_event_failure(efx, 1);
1873 return;
1874 }
1875
1876 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1877}
1878
1879static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1880{
1881 if (ptp->nic_ts_enabled)
1882 queue_work(ptp->pps_workwq, &ptp->pps_work);
1883}
1884
1885void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1886{
1887 struct efx_ptp_data *ptp = efx->ptp_data;
1888 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
1889
1890 if (!ptp) {
1891 if (!efx->ptp_warned) {
1892 netif_warn(efx, drv, efx->net_dev,
1893 "Received PTP event but PTP not set up\n");
1894 efx->ptp_warned = true;
1895 }
1896 return;
1897 }
1898
1899 if (!ptp->enabled)
1900 return;
1901
1902 if (ptp->evt_frag_idx == 0) {
1903 ptp->evt_code = code;
1904 } else if (ptp->evt_code != code) {
1905 netif_err(efx, hw, efx->net_dev,
1906 "PTP out of sequence event %d\n", code);
1907 ptp->evt_frag_idx = 0;
1908 }
1909
1910 ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
1911 if (!MCDI_EVENT_FIELD(*ev, CONT)) {
1912 /* Process resulting event */
1913 switch (code) {
1914 case MCDI_EVENT_CODE_PTP_RX:
1915 ptp_event_rx(efx, ptp);
1916 break;
1917 case MCDI_EVENT_CODE_PTP_FAULT:
1918 ptp_event_fault(efx, ptp);
1919 break;
1920 case MCDI_EVENT_CODE_PTP_PPS:
1921 ptp_event_pps(efx, ptp);
1922 break;
1923 default:
1924 netif_err(efx, hw, efx->net_dev,
1925 "PTP unknown event %d\n", code);
1926 break;
1927 }
1928 ptp->evt_frag_idx = 0;
1929 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
1930 netif_err(efx, hw, efx->net_dev,
1931 "PTP too many event fragments\n");
1932 ptp->evt_frag_idx = 0;
1933 }
1934}
1935
1936void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
1937{
1938 struct efx_nic *efx = channel->efx;
1939 struct efx_ptp_data *ptp = efx->ptp_data;
1940
1941 /* When extracting the sync timestamp minor value, we should discard
1942 * the least significant two bits. These are not required in order
1943 * to reconstruct full-range timestamps and they are optionally used
1944 * to report status depending on the options supplied when subscribing
1945 * for sync events.
1946 */
1947 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
1948 channel->sync_timestamp_minor =
1949 (MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_MS_8BITS) & 0xFC)
1950 << ptp->nic_time.sync_event_minor_shift;
1951
1952 /* if sync events have been disabled then we want to silently ignore
1953 * this event, so throw away result.
1954 */
1955 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
1956 SYNC_EVENTS_VALID);
1957}
1958
1959static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
1960{
1961#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
1962 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset));
1963#else
1964 const u8 *data = eh + efx->rx_packet_ts_offset;
1965 return (u32)data[0] |
1966 (u32)data[1] << 8 |
1967 (u32)data[2] << 16 |
1968 (u32)data[3] << 24;
1969#endif
1970}
1971
1972void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
1973 struct sk_buff *skb)
1974{
1975 struct efx_nic *efx = channel->efx;
1976 struct efx_ptp_data *ptp = efx->ptp_data;
1977 u32 pkt_timestamp_major, pkt_timestamp_minor;
1978 u32 diff, carry;
1979 struct skb_shared_hwtstamps *timestamps;
1980
1981 if (channel->sync_events_state != SYNC_EVENTS_VALID)
1982 return;
1983
1984 pkt_timestamp_minor = efx_rx_buf_timestamp_minor(efx, skb_mac_header(skb));
1985
1986 /* get the difference between the packet and sync timestamps,
1987 * modulo one second
1988 */
1989 diff = pkt_timestamp_minor - channel->sync_timestamp_minor;
1990 if (pkt_timestamp_minor < channel->sync_timestamp_minor)
1991 diff += ptp->nic_time.minor_max;
1992
1993 /* do we roll over a second boundary and need to carry the one? */
1994 carry = (channel->sync_timestamp_minor >= ptp->nic_time.minor_max - diff) ?
1995 1 : 0;
1996
1997 if (diff <= ptp->nic_time.sync_event_diff_max) {
1998 /* packet is ahead of the sync event by a quarter of a second or
1999 * less (allowing for fuzz)
2000 */
2001 pkt_timestamp_major = channel->sync_timestamp_major + carry;
2002 } else if (diff >= ptp->nic_time.sync_event_diff_min) {
2003 /* packet is behind the sync event but within the fuzz factor.
2004 * This means the RX packet and sync event crossed as they were
2005 * placed on the event queue, which can sometimes happen.
2006 */
2007 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
2008 } else {
2009 /* it's outside tolerance in both directions. this might be
2010 * indicative of us missing sync events for some reason, so
2011 * we'll call it an error rather than risk giving a bogus
2012 * timestamp.
2013 */
2014 netif_vdbg(efx, drv, efx->net_dev,
2015 "packet timestamp %x too far from sync event %x:%x\n",
2016 pkt_timestamp_minor, channel->sync_timestamp_major,
2017 channel->sync_timestamp_minor);
2018 return;
2019 }
2020
2021 /* attach the timestamps to the skb */
2022 timestamps = skb_hwtstamps(skb);
2023 timestamps->hwtstamp =
2024 ptp->nic_to_kernel_time(pkt_timestamp_major,
2025 pkt_timestamp_minor,
2026 ptp->ts_corrections.general_rx);
2027}
2028
2029static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta)
2030{
2031 struct efx_ptp_data *ptp_data = container_of(ptp,
2032 struct efx_ptp_data,
2033 phc_clock_info);
2034 struct efx_nic *efx = ptp_data->efx;
2035 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
2036 s64 adjustment_ns;
2037 int rc;
2038
2039 if (delta > MAX_PPB)
2040 delta = MAX_PPB;
2041 else if (delta < -MAX_PPB)
2042 delta = -MAX_PPB;
2043
2044 /* Convert ppb to fixed point ns taking care to round correctly. */
2045 adjustment_ns = ((s64)delta * PPB_SCALE_WORD +
2046 (1 << (ptp_data->adjfreq_ppb_shift - 1))) >>
2047 ptp_data->adjfreq_ppb_shift;
2048
2049 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2050 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
2051 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
2052 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
2053 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
2054 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj),
2055 NULL, 0, NULL);
2056 if (rc != 0)
2057 return rc;
2058
2059 ptp_data->current_adjfreq = adjustment_ns;
2060 return 0;
2061}
2062
2063static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
2064{
2065 u32 nic_major, nic_minor;
2066 struct efx_ptp_data *ptp_data = container_of(ptp,
2067 struct efx_ptp_data,
2068 phc_clock_info);
2069 struct efx_nic *efx = ptp_data->efx;
2070 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
2071
2072 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
2073
2074 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2075 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2076 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
2077 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
2078 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
2079 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2080 NULL, 0, NULL);
2081}
2082
2083static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
2084{
2085 struct efx_ptp_data *ptp_data = container_of(ptp,
2086 struct efx_ptp_data,
2087 phc_clock_info);
2088 struct efx_nic *efx = ptp_data->efx;
2089 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
2090 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
2091 int rc;
2092 ktime_t kt;
2093
2094 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
2095 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2096
2097 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf),
2098 outbuf, sizeof(outbuf), NULL);
2099 if (rc != 0)
2100 return rc;
2101
2102 kt = ptp_data->nic_to_kernel_time(
2103 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
2104 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
2105 *ts = ktime_to_timespec64(kt);
2106 return 0;
2107}
2108
2109static int efx_phc_settime(struct ptp_clock_info *ptp,
2110 const struct timespec64 *e_ts)
2111{
2112 /* Get the current NIC time, efx_phc_gettime.
2113 * Subtract from the desired time to get the offset
2114 * call efx_phc_adjtime with the offset
2115 */
2116 int rc;
2117 struct timespec64 time_now;
2118 struct timespec64 delta;
2119
2120 rc = efx_phc_gettime(ptp, &time_now);
2121 if (rc != 0)
2122 return rc;
2123
2124 delta = timespec64_sub(*e_ts, time_now);
2125
2126 rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta));
2127 if (rc != 0)
2128 return rc;
2129
2130 return 0;
2131}
2132
2133static int efx_phc_enable(struct ptp_clock_info *ptp,
2134 struct ptp_clock_request *request,
2135 int enable)
2136{
2137 struct efx_ptp_data *ptp_data = container_of(ptp,
2138 struct efx_ptp_data,
2139 phc_clock_info);
2140 if (request->type != PTP_CLK_REQ_PPS)
2141 return -EOPNOTSUPP;
2142
2143 ptp_data->nic_ts_enabled = !!enable;
2144 return 0;
2145}
2146
2147static const struct efx_channel_type efx_ptp_channel_type = {
2148 .handle_no_channel = efx_ptp_handle_no_channel,
2149 .pre_probe = efx_ptp_probe_channel,
2150 .post_remove = efx_ptp_remove_channel,
2151 .get_name = efx_ptp_get_channel_name,
2152 /* no copy operation; there is no need to reallocate this channel */
2153 .receive_skb = efx_ptp_rx,
2154 .want_txqs = efx_ptp_want_txqs,
2155 .keep_eventq = false,
2156};
2157
2158void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
2159{
2160 /* Check whether PTP is implemented on this NIC. The DISABLE
2161 * operation will succeed if and only if it is implemented.
2162 */
2163 if (efx_ptp_disable(efx) == 0)
2164 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
2165 &efx_ptp_channel_type;
2166}
2167
2168void efx_ptp_start_datapath(struct efx_nic *efx)
2169{
2170 if (efx_ptp_restart(efx))
2171 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
2172 /* re-enable timestamping if it was previously enabled */
2173 if (efx->type->ptp_set_ts_sync_events)
2174 efx->type->ptp_set_ts_sync_events(efx, true, true);
2175}
2176
2177void efx_ptp_stop_datapath(struct efx_nic *efx)
2178{
2179 /* temporarily disable timestamping */
2180 if (efx->type->ptp_set_ts_sync_events)
2181 efx->type->ptp_set_ts_sync_events(efx, false, true);
2182 efx_ptp_stop(efx);
2183}