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/errno.h>
37#include <linux/ktime.h>
38#include <linux/module.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 "tx.h"
47#include "nic.h" /* indirectly includes ptp.h */
48#include "efx_channels.h"
49
50/* Maximum number of events expected to make up a PTP event */
51#define MAX_EVENT_FRAGS 3
52
53/* Maximum delay, ms, to begin synchronisation */
54#define MAX_SYNCHRONISE_WAIT_MS 2
55
56/* How long, at most, to spend synchronising */
57#define SYNCHRONISE_PERIOD_NS 250000
58
59/* How often to update the shared memory time */
60#define SYNCHRONISATION_GRANULARITY_NS 200
61
62/* Minimum permitted length of a (corrected) synchronisation time */
63#define DEFAULT_MIN_SYNCHRONISATION_NS 120
64
65/* Maximum permitted length of a (corrected) synchronisation time */
66#define MAX_SYNCHRONISATION_NS 1000
67
68/* How many (MC) receive events that can be queued */
69#define MAX_RECEIVE_EVENTS 8
70
71/* Length of (modified) moving average. */
72#define AVERAGE_LENGTH 16
73
74/* How long an unmatched event or packet can be held */
75#define PKT_EVENT_LIFETIME_MS 10
76
77/* How long unused unicast filters can be held */
78#define UCAST_FILTER_EXPIRY_JIFFIES msecs_to_jiffies(30000)
79
80/* Offsets into PTP packet for identification. These offsets are from the
81 * start of the IP header, not the MAC header. Note that neither PTP V1 nor
82 * PTP V2 permit the use of IPV4 options.
83 */
84#define PTP_DPORT_OFFSET 22
85
86#define PTP_V1_VERSION_LENGTH 2
87#define PTP_V1_VERSION_OFFSET 28
88
89#define PTP_V1_SEQUENCE_LENGTH 2
90#define PTP_V1_SEQUENCE_OFFSET 58
91
92/* The minimum length of a PTP V1 packet for offsets, etc. to be valid:
93 * includes IP header.
94 */
95#define PTP_V1_MIN_LENGTH 64
96
97#define PTP_V2_VERSION_LENGTH 1
98#define PTP_V2_VERSION_OFFSET 29
99
100#define PTP_V2_SEQUENCE_LENGTH 2
101#define PTP_V2_SEQUENCE_OFFSET 58
102
103/* The minimum length of a PTP V2 packet for offsets, etc. to be valid:
104 * includes IP header.
105 */
106#define PTP_V2_MIN_LENGTH 63
107
108#define PTP_MIN_LENGTH 63
109
110#define PTP_ADDR_IPV4 0xe0000181 /* 224.0.1.129 */
111
112/* ff0e::181 */
113static const struct in6_addr ptp_addr_ipv6 = { { {
114 0xff, 0x0e, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0x01, 0x81 } } };
115
116/* 01-1B-19-00-00-00 */
117static const u8 ptp_addr_ether[ETH_ALEN] __aligned(2) = {
118 0x01, 0x1b, 0x19, 0x00, 0x00, 0x00 };
119
120#define PTP_EVENT_PORT 319
121#define PTP_GENERAL_PORT 320
122
123/* Annoyingly the format of the version numbers are different between
124 * versions 1 and 2 so it isn't possible to simply look for 1 or 2.
125 */
126#define PTP_VERSION_V1 1
127
128#define PTP_VERSION_V2 2
129#define PTP_VERSION_V2_MASK 0x0f
130
131enum ptp_packet_state {
132 PTP_PACKET_STATE_UNMATCHED = 0,
133 PTP_PACKET_STATE_MATCHED,
134 PTP_PACKET_STATE_TIMED_OUT,
135 PTP_PACKET_STATE_MATCH_UNWANTED
136};
137
138/* NIC synchronised with single word of time only comprising
139 * partial seconds and full nanoseconds: 10^9 ~ 2^30 so 2 bits for seconds.
140 */
141#define MC_NANOSECOND_BITS 30
142#define MC_NANOSECOND_MASK ((1 << MC_NANOSECOND_BITS) - 1)
143#define MC_SECOND_MASK ((1 << (32 - MC_NANOSECOND_BITS)) - 1)
144
145/* Maximum parts-per-billion adjustment that is acceptable */
146#define MAX_PPB 1000000
147
148/* Precalculate scale word to avoid long long division at runtime */
149/* This is equivalent to 2^66 / 10^9. */
150#define PPB_SCALE_WORD ((1LL << (57)) / 1953125LL)
151
152/* How much to shift down after scaling to convert to FP40 */
153#define PPB_SHIFT_FP40 26
154/* ... and FP44. */
155#define PPB_SHIFT_FP44 22
156
157#define PTP_SYNC_ATTEMPTS 4
158
159/**
160 * struct efx_ptp_match - Matching structure, stored in sk_buff's cb area.
161 * @expiry: Time after which the packet should be delivered irrespective of
162 * event arrival.
163 * @state: The state of the packet - whether it is ready for processing or
164 * whether that is of no interest.
165 */
166struct efx_ptp_match {
167 unsigned long expiry;
168 enum ptp_packet_state state;
169};
170
171/**
172 * struct efx_ptp_event_rx - A PTP receive event (from MC)
173 * @link: list of events
174 * @seq0: First part of (PTP) UUID
175 * @seq1: Second part of (PTP) UUID and sequence number
176 * @hwtimestamp: Event timestamp
177 * @expiry: Time which the packet arrived
178 */
179struct efx_ptp_event_rx {
180 struct list_head link;
181 u32 seq0;
182 u32 seq1;
183 ktime_t hwtimestamp;
184 unsigned long expiry;
185};
186
187/**
188 * struct efx_ptp_timeset - Synchronisation between host and MC
189 * @host_start: Host time immediately before hardware timestamp taken
190 * @major: Hardware timestamp, major
191 * @minor: Hardware timestamp, minor
192 * @host_end: Host time immediately after hardware timestamp taken
193 * @wait: Number of NIC clock ticks between hardware timestamp being read and
194 * host end time being seen
195 * @window: Difference of host_end and host_start
196 * @valid: Whether this timeset is valid
197 */
198struct efx_ptp_timeset {
199 u32 host_start;
200 u32 major;
201 u32 minor;
202 u32 host_end;
203 u32 wait;
204 u32 window; /* Derived: end - start, allowing for wrap */
205};
206
207/**
208 * struct efx_ptp_rxfilter - Filter for PTP packets
209 * @list: Node of the list where the filter is added
210 * @ether_type: Network protocol of the filter (ETHER_P_IP / ETHER_P_IPV6)
211 * @loc_port: UDP port of the filter (PTP_EVENT_PORT / PTP_GENERAL_PORT)
212 * @loc_host: IPv4/v6 address of the filter
213 * @expiry: time when the filter expires, in jiffies
214 * @handle: Handle ID for the MCDI filters table
215 */
216struct efx_ptp_rxfilter {
217 struct list_head list;
218 __be16 ether_type;
219 __be16 loc_port;
220 __be32 loc_host[4];
221 unsigned long expiry;
222 int handle;
223};
224
225/**
226 * struct efx_ptp_data - Precision Time Protocol (PTP) state
227 * @efx: The NIC context
228 * @channel: The PTP channel (for Medford and Medford2)
229 * @rxq: Receive SKB queue (awaiting timestamps)
230 * @txq: Transmit SKB queue
231 * @workwq: Work queue for processing pending PTP operations
232 * @work: Work task
233 * @cleanup_work: Work task for periodic cleanup
234 * @reset_required: A serious error has occurred and the PTP task needs to be
235 * reset (disable, enable).
236 * @rxfilters_mcast: Receive filters for multicast PTP packets
237 * @rxfilters_ucast: Receive filters for unicast PTP packets
238 * @config: Current timestamp configuration
239 * @enabled: PTP operation enabled
240 * @mode: Mode in which PTP operating (PTP version)
241 * @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time
242 * @nic_to_kernel_time: Function to convert from NIC to kernel time
243 * @nic_time: contains time details
244 * @nic_time.minor_max: Wrap point for NIC minor times
245 * @nic_time.sync_event_diff_min: Minimum acceptable difference between time
246 * in packet prefix and last MCDI time sync event i.e. how much earlier than
247 * the last sync event time a packet timestamp can be.
248 * @nic_time.sync_event_diff_max: Maximum acceptable difference between time
249 * in packet prefix and last MCDI time sync event i.e. how much later than
250 * the last sync event time a packet timestamp can be.
251 * @nic_time.sync_event_minor_shift: Shift required to make minor time from
252 * field in MCDI time sync event.
253 * @min_synchronisation_ns: Minimum acceptable corrected sync window
254 * @capabilities: Capabilities flags from the NIC
255 * @ts_corrections: contains corrections details
256 * @ts_corrections.ptp_tx: Required driver correction of PTP packet transmit
257 * timestamps
258 * @ts_corrections.ptp_rx: Required driver correction of PTP packet receive
259 * timestamps
260 * @ts_corrections.pps_out: PPS output error (information only)
261 * @ts_corrections.pps_in: Required driver correction of PPS input timestamps
262 * @ts_corrections.general_tx: Required driver correction of general packet
263 * transmit timestamps
264 * @ts_corrections.general_rx: Required driver correction of general packet
265 * receive timestamps
266 * @evt_frags: Partly assembled PTP events
267 * @evt_frag_idx: Current fragment number
268 * @evt_code: Last event code
269 * @start: Address at which MC indicates ready for synchronisation
270 * @host_time_pps: Host time at last PPS
271 * @adjfreq_ppb_shift: Shift required to convert scaled parts-per-billion
272 * frequency adjustment into a fixed point fractional nanosecond format.
273 * @current_adjfreq: Current ppb adjustment.
274 * @phc_clock: Pointer to registered phc device (if primary function)
275 * @phc_clock_info: Registration structure for phc device
276 * @pps_work: pps work task for handling pps events
277 * @pps_workwq: pps work queue
278 * @nic_ts_enabled: Flag indicating if NIC generated TS events are handled
279 * @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids
280 * allocations in main data path).
281 * @good_syncs: Number of successful synchronisations.
282 * @fast_syncs: Number of synchronisations requiring short delay
283 * @bad_syncs: Number of failed synchronisations.
284 * @sync_timeouts: Number of synchronisation timeouts
285 * @no_time_syncs: Number of synchronisations with no good times.
286 * @invalid_sync_windows: Number of sync windows with bad durations.
287 * @undersize_sync_windows: Number of corrected sync windows that are too small
288 * @oversize_sync_windows: Number of corrected sync windows that are too large
289 * @rx_no_timestamp: Number of packets received without a timestamp.
290 * @timeset: Last set of synchronisation statistics.
291 * @xmit_skb: Transmit SKB function.
292 */
293struct efx_ptp_data {
294 struct efx_nic *efx;
295 struct efx_channel *channel;
296 struct sk_buff_head rxq;
297 struct sk_buff_head txq;
298 struct workqueue_struct *workwq;
299 struct work_struct work;
300 struct delayed_work cleanup_work;
301 bool reset_required;
302 struct list_head rxfilters_mcast;
303 struct list_head rxfilters_ucast;
304 struct hwtstamp_config config;
305 bool enabled;
306 unsigned int mode;
307 void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor);
308 ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor,
309 s32 correction);
310 struct {
311 u32 minor_max;
312 u32 sync_event_diff_min;
313 u32 sync_event_diff_max;
314 unsigned int sync_event_minor_shift;
315 } nic_time;
316 unsigned int min_synchronisation_ns;
317 unsigned int capabilities;
318 struct {
319 s32 ptp_tx;
320 s32 ptp_rx;
321 s32 pps_out;
322 s32 pps_in;
323 s32 general_tx;
324 s32 general_rx;
325 } ts_corrections;
326 efx_qword_t evt_frags[MAX_EVENT_FRAGS];
327 int evt_frag_idx;
328 int evt_code;
329 struct efx_buffer start;
330 struct pps_event_time host_time_pps;
331 unsigned int adjfreq_ppb_shift;
332 s64 current_adjfreq;
333 struct ptp_clock *phc_clock;
334 struct ptp_clock_info phc_clock_info;
335 struct work_struct pps_work;
336 struct workqueue_struct *pps_workwq;
337 bool nic_ts_enabled;
338 efx_dword_t txbuf[MCDI_TX_BUF_LEN(MC_CMD_PTP_IN_TRANSMIT_LENMAX)];
339
340 unsigned int good_syncs;
341 unsigned int fast_syncs;
342 unsigned int bad_syncs;
343 unsigned int sync_timeouts;
344 unsigned int no_time_syncs;
345 unsigned int invalid_sync_windows;
346 unsigned int undersize_sync_windows;
347 unsigned int oversize_sync_windows;
348 unsigned int rx_no_timestamp;
349 struct efx_ptp_timeset
350 timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM];
351 void (*xmit_skb)(struct efx_nic *efx, struct sk_buff *skb);
352};
353
354static int efx_phc_adjfine(struct ptp_clock_info *ptp, long scaled_ppm);
355static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta);
356static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts);
357static int efx_phc_settime(struct ptp_clock_info *ptp,
358 const struct timespec64 *e_ts);
359static int efx_phc_enable(struct ptp_clock_info *ptp,
360 struct ptp_clock_request *request, int on);
361static int efx_ptp_insert_unicast_filter(struct efx_nic *efx,
362 struct sk_buff *skb);
363
364bool efx_ptp_use_mac_tx_timestamps(struct efx_nic *efx)
365{
366 return efx_has_cap(efx, TX_MAC_TIMESTAMPING);
367}
368
369/* PTP 'extra' channel is still a traffic channel, but we only create TX queues
370 * if PTP uses MAC TX timestamps, not if PTP uses the MC directly to transmit.
371 */
372static bool efx_ptp_want_txqs(struct efx_channel *channel)
373{
374 return efx_ptp_use_mac_tx_timestamps(efx: channel->efx);
375}
376
377#define PTP_SW_STAT(ext_name, field_name) \
378 { #ext_name, 0, offsetof(struct efx_ptp_data, field_name) }
379#define PTP_MC_STAT(ext_name, mcdi_name) \
380 { #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST }
381static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = {
382 PTP_SW_STAT(ptp_good_syncs, good_syncs),
383 PTP_SW_STAT(ptp_fast_syncs, fast_syncs),
384 PTP_SW_STAT(ptp_bad_syncs, bad_syncs),
385 PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts),
386 PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs),
387 PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows),
388 PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows),
389 PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows),
390 PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp),
391 PTP_MC_STAT(ptp_tx_timestamp_packets, TX),
392 PTP_MC_STAT(ptp_rx_timestamp_packets, RX),
393 PTP_MC_STAT(ptp_timestamp_packets, TS),
394 PTP_MC_STAT(ptp_filter_matches, FM),
395 PTP_MC_STAT(ptp_non_filter_matches, NFM),
396};
397#define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc)
398static const unsigned long efx_ptp_stat_mask[] = {
399 [0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL,
400};
401
402size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings)
403{
404 if (!efx->ptp_data)
405 return 0;
406
407 return efx_nic_describe_stats(desc: efx_ptp_stat_desc, PTP_STAT_COUNT,
408 mask: efx_ptp_stat_mask, names: strings);
409}
410
411size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats)
412{
413 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN);
414 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN);
415 size_t i;
416 int rc;
417
418 if (!efx->ptp_data)
419 return 0;
420
421 /* Copy software statistics */
422 for (i = 0; i < PTP_STAT_COUNT; i++) {
423 if (efx_ptp_stat_desc[i].dma_width)
424 continue;
425 stats[i] = *(unsigned int *)((char *)efx->ptp_data +
426 efx_ptp_stat_desc[i].offset);
427 }
428
429 /* Fetch MC statistics. We *must* fill in all statistics or
430 * risk leaking kernel memory to userland, so if the MCDI
431 * request fails we pretend we got zeroes.
432 */
433 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS);
434 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
435 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
436 outbuf, outlen: sizeof(outbuf), NULL);
437 if (rc)
438 memset(outbuf, 0, sizeof(outbuf));
439 efx_nic_update_stats(desc: efx_ptp_stat_desc, PTP_STAT_COUNT,
440 mask: efx_ptp_stat_mask,
441 stats, _MCDI_PTR(outbuf, 0), accumulate: false);
442
443 return PTP_STAT_COUNT;
444}
445
446/* To convert from s27 format to ns we multiply then divide by a power of 2.
447 * For the conversion from ns to s27, the operation is also converted to a
448 * multiply and shift.
449 */
450#define S27_TO_NS_SHIFT (27)
451#define NS_TO_S27_MULT (((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC)
452#define NS_TO_S27_SHIFT (63 - S27_TO_NS_SHIFT)
453#define S27_MINOR_MAX (1 << S27_TO_NS_SHIFT)
454
455/* For Huntington platforms NIC time is in seconds and fractions of a second
456 * where the minor register only uses 27 bits in units of 2^-27s.
457 */
458static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor)
459{
460 struct timespec64 ts = ns_to_timespec64(nsec: ns);
461 u32 maj = (u32)ts.tv_sec;
462 u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT +
463 (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT);
464
465 /* The conversion can result in the minor value exceeding the maximum.
466 * In this case, round up to the next second.
467 */
468 if (min >= S27_MINOR_MAX) {
469 min -= S27_MINOR_MAX;
470 maj++;
471 }
472
473 *nic_major = maj;
474 *nic_minor = min;
475}
476
477static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor)
478{
479 u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC +
480 (1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT);
481 return ktime_set(secs: nic_major, nsecs: ns);
482}
483
484static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor,
485 s32 correction)
486{
487 /* Apply the correction and deal with carry */
488 nic_minor += correction;
489 if ((s32)nic_minor < 0) {
490 nic_minor += S27_MINOR_MAX;
491 nic_major--;
492 } else if (nic_minor >= S27_MINOR_MAX) {
493 nic_minor -= S27_MINOR_MAX;
494 nic_major++;
495 }
496
497 return efx_ptp_s27_to_ktime(nic_major, nic_minor);
498}
499
500/* For Medford2 platforms the time is in seconds and quarter nanoseconds. */
501static void efx_ptp_ns_to_s_qns(s64 ns, u32 *nic_major, u32 *nic_minor)
502{
503 struct timespec64 ts = ns_to_timespec64(nsec: ns);
504
505 *nic_major = (u32)ts.tv_sec;
506 *nic_minor = ts.tv_nsec * 4;
507}
508
509static ktime_t efx_ptp_s_qns_to_ktime_correction(u32 nic_major, u32 nic_minor,
510 s32 correction)
511{
512 ktime_t kt;
513
514 nic_minor = DIV_ROUND_CLOSEST(nic_minor, 4);
515 correction = DIV_ROUND_CLOSEST(correction, 4);
516
517 kt = ktime_set(secs: nic_major, nsecs: nic_minor);
518
519 if (correction >= 0)
520 kt = ktime_add_ns(kt, (u64)correction);
521 else
522 kt = ktime_sub_ns(kt, (u64)-correction);
523 return kt;
524}
525
526struct efx_channel *efx_ptp_channel(struct efx_nic *efx)
527{
528 return efx->ptp_data ? efx->ptp_data->channel : NULL;
529}
530
531void efx_ptp_update_channel(struct efx_nic *efx, struct efx_channel *channel)
532{
533 if (efx->ptp_data)
534 efx->ptp_data->channel = channel;
535}
536
537static u32 last_sync_timestamp_major(struct efx_nic *efx)
538{
539 struct efx_channel *channel = efx_ptp_channel(efx);
540 u32 major = 0;
541
542 if (channel)
543 major = channel->sync_timestamp_major;
544 return major;
545}
546
547/* The 8000 series and later can provide the time from the MAC, which is only
548 * 48 bits long and provides meta-information in the top 2 bits.
549 */
550static ktime_t
551efx_ptp_mac_nic_to_ktime_correction(struct efx_nic *efx,
552 struct efx_ptp_data *ptp,
553 u32 nic_major, u32 nic_minor,
554 s32 correction)
555{
556 u32 sync_timestamp;
557 ktime_t kt = { 0 };
558 s16 delta;
559
560 if (!(nic_major & 0x80000000)) {
561 WARN_ON_ONCE(nic_major >> 16);
562
563 /* Medford provides 48 bits of timestamp, so we must get the top
564 * 16 bits from the timesync event state.
565 *
566 * We only have the lower 16 bits of the time now, but we do
567 * have a full resolution timestamp at some point in past. As
568 * long as the difference between the (real) now and the sync
569 * is less than 2^15, then we can reconstruct the difference
570 * between those two numbers using only the lower 16 bits of
571 * each.
572 *
573 * Put another way
574 *
575 * a - b = ((a mod k) - b) mod k
576 *
577 * when -k/2 < (a-b) < k/2. In our case k is 2^16. We know
578 * (a mod k) and b, so can calculate the delta, a - b.
579 *
580 */
581 sync_timestamp = last_sync_timestamp_major(efx);
582
583 /* Because delta is s16 this does an implicit mask down to
584 * 16 bits which is what we need, assuming
585 * MEDFORD_TX_SECS_EVENT_BITS is 16. delta is signed so that
586 * we can deal with the (unlikely) case of sync timestamps
587 * arriving from the future.
588 */
589 delta = nic_major - sync_timestamp;
590
591 /* Recover the fully specified time now, by applying the offset
592 * to the (fully specified) sync time.
593 */
594 nic_major = sync_timestamp + delta;
595
596 kt = ptp->nic_to_kernel_time(nic_major, nic_minor,
597 correction);
598 }
599 return kt;
600}
601
602ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue)
603{
604 struct efx_nic *efx = tx_queue->efx;
605 struct efx_ptp_data *ptp = efx->ptp_data;
606 ktime_t kt;
607
608 if (efx_ptp_use_mac_tx_timestamps(efx))
609 kt = efx_ptp_mac_nic_to_ktime_correction(efx, ptp,
610 nic_major: tx_queue->completed_timestamp_major,
611 nic_minor: tx_queue->completed_timestamp_minor,
612 correction: ptp->ts_corrections.general_tx);
613 else
614 kt = ptp->nic_to_kernel_time(
615 tx_queue->completed_timestamp_major,
616 tx_queue->completed_timestamp_minor,
617 ptp->ts_corrections.general_tx);
618 return kt;
619}
620
621/* Get PTP attributes and set up time conversions */
622static int efx_ptp_get_attributes(struct efx_nic *efx)
623{
624 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN);
625 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN);
626 struct efx_ptp_data *ptp = efx->ptp_data;
627 int rc;
628 u32 fmt;
629 size_t out_len;
630
631 /* Get the PTP attributes. If the NIC doesn't support the operation we
632 * use the default format for compatibility with older NICs i.e.
633 * seconds and nanoseconds.
634 */
635 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES);
636 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
637 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
638 outbuf, outlen: sizeof(outbuf), outlen_actual: &out_len);
639 if (rc == 0) {
640 fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT);
641 } else if (rc == -EINVAL) {
642 fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS;
643 } else if (rc == -EPERM) {
644 pci_info(efx->pci_dev, "no PTP support\n");
645 return rc;
646 } else {
647 efx_mcdi_display_error(efx, MC_CMD_PTP, inlen: sizeof(inbuf),
648 outbuf, outlen: sizeof(outbuf), rc);
649 return rc;
650 }
651
652 switch (fmt) {
653 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION:
654 ptp->ns_to_nic_time = efx_ptp_ns_to_s27;
655 ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction;
656 ptp->nic_time.minor_max = 1 << 27;
657 ptp->nic_time.sync_event_minor_shift = 19;
658 break;
659 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_QTR_NANOSECONDS:
660 ptp->ns_to_nic_time = efx_ptp_ns_to_s_qns;
661 ptp->nic_to_kernel_time = efx_ptp_s_qns_to_ktime_correction;
662 ptp->nic_time.minor_max = 4000000000UL;
663 ptp->nic_time.sync_event_minor_shift = 24;
664 break;
665 default:
666 return -ERANGE;
667 }
668
669 /* Precalculate acceptable difference between the minor time in the
670 * packet prefix and the last MCDI time sync event. We expect the
671 * packet prefix timestamp to be after of sync event by up to one
672 * sync event interval (0.25s) but we allow it to exceed this by a
673 * fuzz factor of (0.1s)
674 */
675 ptp->nic_time.sync_event_diff_min = ptp->nic_time.minor_max
676 - (ptp->nic_time.minor_max / 10);
677 ptp->nic_time.sync_event_diff_max = (ptp->nic_time.minor_max / 4)
678 + (ptp->nic_time.minor_max / 10);
679
680 /* MC_CMD_PTP_OP_GET_ATTRIBUTES has been extended twice from an older
681 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT. The function now may return
682 * a value to use for the minimum acceptable corrected synchronization
683 * window and may return further capabilities.
684 * If we have the extra information store it. For older firmware that
685 * does not implement the extended command use the default value.
686 */
687 if (rc == 0 &&
688 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_CAPABILITIES_OFST)
689 ptp->min_synchronisation_ns =
690 MCDI_DWORD(outbuf,
691 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN);
692 else
693 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS;
694
695 if (rc == 0 &&
696 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN)
697 ptp->capabilities = MCDI_DWORD(outbuf,
698 PTP_OUT_GET_ATTRIBUTES_CAPABILITIES);
699 else
700 ptp->capabilities = 0;
701
702 /* Set up the shift for conversion between frequency
703 * adjustments in parts-per-billion and the fixed-point
704 * fractional ns format that the adapter uses.
705 */
706 if (ptp->capabilities & (1 << MC_CMD_PTP_OUT_GET_ATTRIBUTES_FP44_FREQ_ADJ_LBN))
707 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP44;
708 else
709 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP40;
710
711 return 0;
712}
713
714/* Get PTP timestamp corrections */
715static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx)
716{
717 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN);
718 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN);
719 int rc;
720 size_t out_len;
721
722 /* Get the timestamp corrections from the NIC. If this operation is
723 * not supported (older NICs) then no correction is required.
724 */
725 MCDI_SET_DWORD(inbuf, PTP_IN_OP,
726 MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS);
727 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
728
729 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
730 outbuf, outlen: sizeof(outbuf), outlen_actual: &out_len);
731 if (rc == 0) {
732 efx->ptp_data->ts_corrections.ptp_tx = MCDI_DWORD(outbuf,
733 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT);
734 efx->ptp_data->ts_corrections.ptp_rx = MCDI_DWORD(outbuf,
735 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE);
736 efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf,
737 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT);
738 efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf,
739 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN);
740
741 if (out_len >= MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN) {
742 efx->ptp_data->ts_corrections.general_tx = MCDI_DWORD(
743 outbuf,
744 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_TX);
745 efx->ptp_data->ts_corrections.general_rx = MCDI_DWORD(
746 outbuf,
747 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_RX);
748 } else {
749 efx->ptp_data->ts_corrections.general_tx =
750 efx->ptp_data->ts_corrections.ptp_tx;
751 efx->ptp_data->ts_corrections.general_rx =
752 efx->ptp_data->ts_corrections.ptp_rx;
753 }
754 } else if (rc == -EINVAL) {
755 efx->ptp_data->ts_corrections.ptp_tx = 0;
756 efx->ptp_data->ts_corrections.ptp_rx = 0;
757 efx->ptp_data->ts_corrections.pps_out = 0;
758 efx->ptp_data->ts_corrections.pps_in = 0;
759 efx->ptp_data->ts_corrections.general_tx = 0;
760 efx->ptp_data->ts_corrections.general_rx = 0;
761 } else {
762 efx_mcdi_display_error(efx, MC_CMD_PTP, inlen: sizeof(inbuf), outbuf,
763 outlen: sizeof(outbuf), rc);
764 return rc;
765 }
766
767 return 0;
768}
769
770/* Enable MCDI PTP support. */
771static int efx_ptp_enable(struct efx_nic *efx)
772{
773 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN);
774 MCDI_DECLARE_BUF_ERR(outbuf);
775 int rc;
776
777 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE);
778 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
779 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE,
780 efx->ptp_data->channel ?
781 efx->ptp_data->channel->channel : 0);
782 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode);
783
784 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
785 outbuf, outlen: sizeof(outbuf), NULL);
786 rc = (rc == -EALREADY) ? 0 : rc;
787 if (rc)
788 efx_mcdi_display_error(efx, MC_CMD_PTP,
789 MC_CMD_PTP_IN_ENABLE_LEN,
790 outbuf, outlen: sizeof(outbuf), rc);
791 return rc;
792}
793
794/* Disable MCDI PTP support.
795 *
796 * Note that this function should never rely on the presence of ptp_data -
797 * may be called before that exists.
798 */
799static int efx_ptp_disable(struct efx_nic *efx)
800{
801 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN);
802 MCDI_DECLARE_BUF_ERR(outbuf);
803 int rc;
804
805 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE);
806 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
807 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
808 outbuf, outlen: sizeof(outbuf), NULL);
809 rc = (rc == -EALREADY) ? 0 : rc;
810 /* If we get ENOSYS, the NIC doesn't support PTP, and thus this function
811 * should only have been called during probe.
812 */
813 if (rc == -ENOSYS || rc == -EPERM)
814 pci_info(efx->pci_dev, "no PTP support\n");
815 else if (rc)
816 efx_mcdi_display_error(efx, MC_CMD_PTP,
817 MC_CMD_PTP_IN_DISABLE_LEN,
818 outbuf, outlen: sizeof(outbuf), rc);
819 return rc;
820}
821
822static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q)
823{
824 struct sk_buff *skb;
825
826 while ((skb = skb_dequeue(list: q))) {
827 local_bh_disable();
828 netif_receive_skb(skb);
829 local_bh_enable();
830 }
831}
832
833static void efx_ptp_handle_no_channel(struct efx_nic *efx)
834{
835 netif_err(efx, drv, efx->net_dev,
836 "ERROR: PTP requires MSI-X and 1 additional interrupt"
837 "vector. PTP disabled\n");
838}
839
840/* Repeatedly send the host time to the MC which will capture the hardware
841 * time.
842 */
843static void efx_ptp_send_times(struct efx_nic *efx,
844 struct pps_event_time *last_time)
845{
846 struct pps_event_time now;
847 struct timespec64 limit;
848 struct efx_ptp_data *ptp = efx->ptp_data;
849 int *mc_running = ptp->start.addr;
850
851 pps_get_ts(ts: &now);
852 limit = now.ts_real;
853 timespec64_add_ns(a: &limit, SYNCHRONISE_PERIOD_NS);
854
855 /* Write host time for specified period or until MC is done */
856 while ((timespec64_compare(lhs: &now.ts_real, rhs: &limit) < 0) &&
857 READ_ONCE(*mc_running)) {
858 struct timespec64 update_time;
859 unsigned int host_time;
860
861 /* Don't update continuously to avoid saturating the PCIe bus */
862 update_time = now.ts_real;
863 timespec64_add_ns(a: &update_time, SYNCHRONISATION_GRANULARITY_NS);
864 do {
865 pps_get_ts(ts: &now);
866 } while ((timespec64_compare(lhs: &now.ts_real, rhs: &update_time) < 0) &&
867 READ_ONCE(*mc_running));
868
869 /* Synchronise NIC with single word of time only */
870 host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS |
871 now.ts_real.tv_nsec);
872 /* Update host time in NIC memory */
873 efx->type->ptp_write_host_time(efx, host_time);
874 }
875 *last_time = now;
876}
877
878/* Read a timeset from the MC's results and partial process. */
879static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data),
880 struct efx_ptp_timeset *timeset)
881{
882 unsigned start_ns, end_ns;
883
884 timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART);
885 timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR);
886 timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR);
887 timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND),
888 timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS);
889
890 /* Ignore seconds */
891 start_ns = timeset->host_start & MC_NANOSECOND_MASK;
892 end_ns = timeset->host_end & MC_NANOSECOND_MASK;
893 /* Allow for rollover */
894 if (end_ns < start_ns)
895 end_ns += NSEC_PER_SEC;
896 /* Determine duration of operation */
897 timeset->window = end_ns - start_ns;
898}
899
900/* Process times received from MC.
901 *
902 * Extract times from returned results, and establish the minimum value
903 * seen. The minimum value represents the "best" possible time and events
904 * too much greater than this are rejected - the machine is, perhaps, too
905 * busy. A number of readings are taken so that, hopefully, at least one good
906 * synchronisation will be seen in the results.
907 */
908static int
909efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf),
910 size_t response_length,
911 const struct pps_event_time *last_time)
912{
913 unsigned number_readings =
914 MCDI_VAR_ARRAY_LEN(response_length,
915 PTP_OUT_SYNCHRONIZE_TIMESET);
916 unsigned i;
917 unsigned ngood = 0;
918 unsigned last_good = 0;
919 struct efx_ptp_data *ptp = efx->ptp_data;
920 u32 last_sec;
921 u32 start_sec;
922 struct timespec64 delta;
923 ktime_t mc_time;
924
925 if (number_readings == 0)
926 return -EAGAIN;
927
928 /* Read the set of results and find the last good host-MC
929 * synchronization result. The MC times when it finishes reading the
930 * host time so the corrected window time should be fairly constant
931 * for a given platform. Increment stats for any results that appear
932 * to be erroneous.
933 */
934 for (i = 0; i < number_readings; i++) {
935 s32 window, corrected;
936 struct timespec64 wait;
937
938 efx_ptp_read_timeset(
939 MCDI_ARRAY_STRUCT_PTR(synch_buf,
940 PTP_OUT_SYNCHRONIZE_TIMESET, i),
941 timeset: &ptp->timeset[i]);
942
943 wait = ktime_to_timespec64(
944 ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0));
945 window = ptp->timeset[i].window;
946 corrected = window - wait.tv_nsec;
947
948 /* We expect the uncorrected synchronization window to be at
949 * least as large as the interval between host start and end
950 * times. If it is smaller than this then this is mostly likely
951 * to be a consequence of the host's time being adjusted.
952 * Check that the corrected sync window is in a reasonable
953 * range. If it is out of range it is likely to be because an
954 * interrupt or other delay occurred between reading the system
955 * time and writing it to MC memory.
956 */
957 if (window < SYNCHRONISATION_GRANULARITY_NS) {
958 ++ptp->invalid_sync_windows;
959 } else if (corrected >= MAX_SYNCHRONISATION_NS) {
960 ++ptp->oversize_sync_windows;
961 } else if (corrected < ptp->min_synchronisation_ns) {
962 ++ptp->undersize_sync_windows;
963 } else {
964 ngood++;
965 last_good = i;
966 }
967 }
968
969 if (ngood == 0) {
970 netif_warn(efx, drv, efx->net_dev,
971 "PTP no suitable synchronisations\n");
972 return -EAGAIN;
973 }
974
975 /* Calculate delay from last good sync (host time) to last_time.
976 * It is possible that the seconds rolled over between taking
977 * the start reading and the last value written by the host. The
978 * timescales are such that a gap of more than one second is never
979 * expected. delta is *not* normalised.
980 */
981 start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS;
982 last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK;
983 if (start_sec != last_sec &&
984 ((start_sec + 1) & MC_SECOND_MASK) != last_sec) {
985 netif_warn(efx, hw, efx->net_dev,
986 "PTP bad synchronisation seconds\n");
987 return -EAGAIN;
988 }
989 delta.tv_sec = (last_sec - start_sec) & 1;
990 delta.tv_nsec =
991 last_time->ts_real.tv_nsec -
992 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK);
993
994 /* Convert the NIC time at last good sync into kernel time.
995 * No correction is required - this time is the output of a
996 * firmware process.
997 */
998 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major,
999 ptp->timeset[last_good].minor, 0);
1000
1001 /* Calculate delay from NIC top of second to last_time */
1002 delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec;
1003
1004 /* Set PPS timestamp to match NIC top of second */
1005 ptp->host_time_pps = *last_time;
1006 pps_sub_ts(ts: &ptp->host_time_pps, delta);
1007
1008 return 0;
1009}
1010
1011/* Synchronize times between the host and the MC */
1012static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings)
1013{
1014 struct efx_ptp_data *ptp = efx->ptp_data;
1015 MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX);
1016 size_t response_length;
1017 int rc;
1018 unsigned long timeout;
1019 struct pps_event_time last_time = {};
1020 unsigned int loops = 0;
1021 int *start = ptp->start.addr;
1022
1023 MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE);
1024 MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0);
1025 MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS,
1026 num_readings);
1027 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR,
1028 ptp->start.dma_addr);
1029
1030 /* Clear flag that signals MC ready */
1031 WRITE_ONCE(*start, 0);
1032 rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, inbuf: synch_buf,
1033 MC_CMD_PTP_IN_SYNCHRONIZE_LEN);
1034 EFX_WARN_ON_ONCE_PARANOID(rc);
1035
1036 /* Wait for start from MCDI (or timeout) */
1037 timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS);
1038 while (!READ_ONCE(*start) && (time_before(jiffies, timeout))) {
1039 udelay(20); /* Usually start MCDI execution quickly */
1040 loops++;
1041 }
1042
1043 if (loops <= 1)
1044 ++ptp->fast_syncs;
1045 if (!time_before(jiffies, timeout))
1046 ++ptp->sync_timeouts;
1047
1048 if (READ_ONCE(*start))
1049 efx_ptp_send_times(efx, last_time: &last_time);
1050
1051 /* Collect results */
1052 rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP,
1053 MC_CMD_PTP_IN_SYNCHRONIZE_LEN,
1054 outbuf: synch_buf, outlen: sizeof(synch_buf),
1055 outlen_actual: &response_length);
1056 if (rc == 0) {
1057 rc = efx_ptp_process_times(efx, synch_buf, response_length,
1058 last_time: &last_time);
1059 if (rc == 0)
1060 ++ptp->good_syncs;
1061 else
1062 ++ptp->no_time_syncs;
1063 }
1064
1065 /* Increment the bad syncs counter if the synchronize fails, whatever
1066 * the reason.
1067 */
1068 if (rc != 0)
1069 ++ptp->bad_syncs;
1070
1071 return rc;
1072}
1073
1074/* Transmit a PTP packet via the dedicated hardware timestamped queue. */
1075static void efx_ptp_xmit_skb_queue(struct efx_nic *efx, struct sk_buff *skb)
1076{
1077 struct efx_ptp_data *ptp_data = efx->ptp_data;
1078 u8 type = efx_tx_csum_type_skb(skb);
1079 struct efx_tx_queue *tx_queue;
1080
1081 tx_queue = efx_channel_get_tx_queue(channel: ptp_data->channel, type);
1082 if (tx_queue && tx_queue->timestamping) {
1083 skb_get(skb);
1084
1085 /* This code invokes normal driver TX code which is always
1086 * protected from softirqs when called from generic TX code,
1087 * which in turn disables preemption. Look at __dev_queue_xmit
1088 * which uses rcu_read_lock_bh disabling preemption for RCU
1089 * plus disabling softirqs. We do not need RCU reader
1090 * protection here.
1091 *
1092 * Although it is theoretically safe for current PTP TX/RX code
1093 * running without disabling softirqs, there are three good
1094 * reasond for doing so:
1095 *
1096 * 1) The code invoked is mainly implemented for non-PTP
1097 * packets and it is always executed with softirqs
1098 * disabled.
1099 * 2) This being a single PTP packet, better to not
1100 * interrupt its processing by softirqs which can lead
1101 * to high latencies.
1102 * 3) netdev_xmit_more checks preemption is disabled and
1103 * triggers a BUG_ON if not.
1104 */
1105 local_bh_disable();
1106 efx_enqueue_skb(tx_queue, skb);
1107 local_bh_enable();
1108
1109 /* We need to add the filters after enqueuing the packet.
1110 * Otherwise, there's high latency in sending back the
1111 * timestamp, causing ptp4l timeouts
1112 */
1113 efx_ptp_insert_unicast_filter(efx, skb);
1114 dev_consume_skb_any(skb);
1115 } else {
1116 WARN_ONCE(1, "PTP channel has no timestamped tx queue\n");
1117 dev_kfree_skb_any(skb);
1118 }
1119}
1120
1121/* Transmit a PTP packet, via the MCDI interface, to the wire. */
1122static void efx_ptp_xmit_skb_mc(struct efx_nic *efx, struct sk_buff *skb)
1123{
1124 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN);
1125 struct efx_ptp_data *ptp_data = efx->ptp_data;
1126 struct skb_shared_hwtstamps timestamps;
1127 size_t len;
1128 int rc;
1129
1130 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT);
1131 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0);
1132 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len);
1133 if (skb_shinfo(skb)->nr_frags != 0) {
1134 rc = skb_linearize(skb);
1135 if (rc != 0)
1136 goto fail;
1137 }
1138
1139 if (skb->ip_summed == CHECKSUM_PARTIAL) {
1140 rc = skb_checksum_help(skb);
1141 if (rc != 0)
1142 goto fail;
1143 }
1144 skb_copy_from_linear_data(skb,
1145 MCDI_PTR(ptp_data->txbuf,
1146 PTP_IN_TRANSMIT_PACKET),
1147 len: skb->len);
1148 rc = efx_mcdi_rpc(efx, MC_CMD_PTP,
1149 inbuf: ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len),
1150 outbuf: txtime, outlen: sizeof(txtime), outlen_actual: &len);
1151 if (rc != 0)
1152 goto fail;
1153
1154 memset(&timestamps, 0, sizeof(timestamps));
1155 timestamps.hwtstamp = ptp_data->nic_to_kernel_time(
1156 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR),
1157 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR),
1158 ptp_data->ts_corrections.ptp_tx);
1159
1160 skb_tstamp_tx(orig_skb: skb, hwtstamps: &timestamps);
1161
1162 /* Add the filters after sending back the timestamp to avoid delaying it
1163 * or ptp4l may timeout.
1164 */
1165 efx_ptp_insert_unicast_filter(efx, skb);
1166
1167fail:
1168 dev_kfree_skb_any(skb);
1169
1170 return;
1171}
1172
1173/* Process any queued receive events and corresponding packets
1174 *
1175 * q is returned with all the packets that are ready for delivery.
1176 */
1177static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q)
1178{
1179 struct efx_ptp_data *ptp = efx->ptp_data;
1180 struct sk_buff *skb;
1181
1182 while ((skb = skb_dequeue(list: &ptp->rxq))) {
1183 struct efx_ptp_match *match;
1184
1185 match = (struct efx_ptp_match *)skb->cb;
1186 if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) {
1187 __skb_queue_tail(list: q, newsk: skb);
1188 } else if (time_after(jiffies, match->expiry)) {
1189 match->state = PTP_PACKET_STATE_TIMED_OUT;
1190 ++ptp->rx_no_timestamp;
1191 __skb_queue_tail(list: q, newsk: skb);
1192 } else {
1193 /* Replace unprocessed entry and stop */
1194 skb_queue_head(list: &ptp->rxq, newsk: skb);
1195 break;
1196 }
1197 }
1198}
1199
1200/* Complete processing of a received packet */
1201static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb)
1202{
1203 local_bh_disable();
1204 netif_receive_skb(skb);
1205 local_bh_enable();
1206}
1207
1208static struct efx_ptp_rxfilter *
1209efx_ptp_find_filter(struct list_head *filter_list, struct efx_filter_spec *spec)
1210{
1211 struct efx_ptp_rxfilter *rxfilter;
1212
1213 list_for_each_entry(rxfilter, filter_list, list) {
1214 if (rxfilter->ether_type == spec->ether_type &&
1215 rxfilter->loc_port == spec->loc_port &&
1216 !memcmp(p: rxfilter->loc_host, q: spec->loc_host, size: sizeof(spec->loc_host)))
1217 return rxfilter;
1218 }
1219
1220 return NULL;
1221}
1222
1223static void efx_ptp_remove_one_filter(struct efx_nic *efx,
1224 struct efx_ptp_rxfilter *rxfilter)
1225{
1226 efx_filter_remove_id_safe(efx, priority: EFX_FILTER_PRI_REQUIRED,
1227 filter_id: rxfilter->handle);
1228 list_del(entry: &rxfilter->list);
1229 kfree(objp: rxfilter);
1230}
1231
1232static void efx_ptp_remove_filters(struct efx_nic *efx,
1233 struct list_head *filter_list)
1234{
1235 struct efx_ptp_rxfilter *rxfilter, *tmp;
1236
1237 list_for_each_entry_safe(rxfilter, tmp, filter_list, list)
1238 efx_ptp_remove_one_filter(efx, rxfilter);
1239}
1240
1241static void efx_ptp_init_filter(struct efx_nic *efx,
1242 struct efx_filter_spec *rxfilter)
1243{
1244 struct efx_channel *channel = efx->ptp_data->channel;
1245 struct efx_rx_queue *queue = efx_channel_get_rx_queue(channel);
1246
1247 efx_filter_init_rx(spec: rxfilter, priority: EFX_FILTER_PRI_REQUIRED, flags: 0,
1248 rxq_id: efx_rx_queue_index(rx_queue: queue));
1249}
1250
1251static int efx_ptp_insert_filter(struct efx_nic *efx,
1252 struct list_head *filter_list,
1253 struct efx_filter_spec *spec,
1254 unsigned long expiry)
1255{
1256 struct efx_ptp_data *ptp = efx->ptp_data;
1257 struct efx_ptp_rxfilter *rxfilter;
1258 int rc;
1259
1260 rxfilter = efx_ptp_find_filter(filter_list, spec);
1261 if (rxfilter) {
1262 rxfilter->expiry = expiry;
1263 return 0;
1264 }
1265
1266 rxfilter = kzalloc(size: sizeof(*rxfilter), GFP_KERNEL);
1267 if (!rxfilter)
1268 return -ENOMEM;
1269
1270 rc = efx_filter_insert_filter(efx, spec, replace_equal: true);
1271 if (rc < 0)
1272 goto fail;
1273
1274 rxfilter->handle = rc;
1275 rxfilter->ether_type = spec->ether_type;
1276 rxfilter->loc_port = spec->loc_port;
1277 memcpy(rxfilter->loc_host, spec->loc_host, sizeof(spec->loc_host));
1278 rxfilter->expiry = expiry;
1279 list_add(new: &rxfilter->list, head: filter_list);
1280
1281 queue_delayed_work(wq: ptp->workwq, dwork: &ptp->cleanup_work,
1282 UCAST_FILTER_EXPIRY_JIFFIES + 1);
1283
1284 return 0;
1285
1286fail:
1287 kfree(objp: rxfilter);
1288 return rc;
1289}
1290
1291static int efx_ptp_insert_ipv4_filter(struct efx_nic *efx,
1292 struct list_head *filter_list,
1293 __be32 addr, u16 port,
1294 unsigned long expiry)
1295{
1296 struct efx_filter_spec spec;
1297
1298 efx_ptp_init_filter(efx, rxfilter: &spec);
1299 efx_filter_set_ipv4_local(spec: &spec, IPPROTO_UDP, host: addr, htons(port));
1300 return efx_ptp_insert_filter(efx, filter_list, spec: &spec, expiry);
1301}
1302
1303static int efx_ptp_insert_ipv6_filter(struct efx_nic *efx,
1304 struct list_head *filter_list,
1305 const struct in6_addr *addr, u16 port,
1306 unsigned long expiry)
1307{
1308 struct efx_filter_spec spec;
1309
1310 efx_ptp_init_filter(efx, rxfilter: &spec);
1311 efx_filter_set_ipv6_local(spec: &spec, IPPROTO_UDP, host: addr, htons(port));
1312 return efx_ptp_insert_filter(efx, filter_list, spec: &spec, expiry);
1313}
1314
1315static int efx_ptp_insert_eth_multicast_filter(struct efx_nic *efx)
1316{
1317 struct efx_ptp_data *ptp = efx->ptp_data;
1318 struct efx_filter_spec spec;
1319
1320 efx_ptp_init_filter(efx, rxfilter: &spec);
1321 efx_filter_set_eth_local(spec: &spec, vid: EFX_FILTER_VID_UNSPEC, addr: ptp_addr_ether);
1322 spec.match_flags |= EFX_FILTER_MATCH_ETHER_TYPE;
1323 spec.ether_type = htons(ETH_P_1588);
1324 return efx_ptp_insert_filter(efx, filter_list: &ptp->rxfilters_mcast, spec: &spec, expiry: 0);
1325}
1326
1327static int efx_ptp_insert_multicast_filters(struct efx_nic *efx)
1328{
1329 struct efx_ptp_data *ptp = efx->ptp_data;
1330 int rc;
1331
1332 if (!ptp->channel || !list_empty(head: &ptp->rxfilters_mcast))
1333 return 0;
1334
1335 /* Must filter on both event and general ports to ensure
1336 * that there is no packet re-ordering.
1337 */
1338 rc = efx_ptp_insert_ipv4_filter(efx, filter_list: &ptp->rxfilters_mcast,
1339 htonl(PTP_ADDR_IPV4), PTP_EVENT_PORT,
1340 expiry: 0);
1341 if (rc < 0)
1342 goto fail;
1343
1344 rc = efx_ptp_insert_ipv4_filter(efx, filter_list: &ptp->rxfilters_mcast,
1345 htonl(PTP_ADDR_IPV4), PTP_GENERAL_PORT,
1346 expiry: 0);
1347 if (rc < 0)
1348 goto fail;
1349
1350 /* if the NIC supports hw timestamps by the MAC, we can support
1351 * PTP over IPv6 and Ethernet
1352 */
1353 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1354 rc = efx_ptp_insert_ipv6_filter(efx, filter_list: &ptp->rxfilters_mcast,
1355 addr: &ptp_addr_ipv6, PTP_EVENT_PORT, expiry: 0);
1356 if (rc < 0)
1357 goto fail;
1358
1359 rc = efx_ptp_insert_ipv6_filter(efx, filter_list: &ptp->rxfilters_mcast,
1360 addr: &ptp_addr_ipv6, PTP_GENERAL_PORT, expiry: 0);
1361 if (rc < 0)
1362 goto fail;
1363
1364 rc = efx_ptp_insert_eth_multicast_filter(efx);
1365
1366 /* Not all firmware variants support this filter */
1367 if (rc < 0 && rc != -EPROTONOSUPPORT)
1368 goto fail;
1369 }
1370
1371 return 0;
1372
1373fail:
1374 efx_ptp_remove_filters(efx, filter_list: &ptp->rxfilters_mcast);
1375 return rc;
1376}
1377
1378static bool efx_ptp_valid_unicast_event_pkt(struct sk_buff *skb)
1379{
1380 if (skb->protocol == htons(ETH_P_IP)) {
1381 return ip_hdr(skb)->daddr != htonl(PTP_ADDR_IPV4) &&
1382 ip_hdr(skb)->protocol == IPPROTO_UDP &&
1383 udp_hdr(skb)->source == htons(PTP_EVENT_PORT);
1384 } else if (skb->protocol == htons(ETH_P_IPV6)) {
1385 return !ipv6_addr_equal(a1: &ipv6_hdr(skb)->daddr, a2: &ptp_addr_ipv6) &&
1386 ipv6_hdr(skb)->nexthdr == IPPROTO_UDP &&
1387 udp_hdr(skb)->source == htons(PTP_EVENT_PORT);
1388 }
1389 return false;
1390}
1391
1392static int efx_ptp_insert_unicast_filter(struct efx_nic *efx,
1393 struct sk_buff *skb)
1394{
1395 struct efx_ptp_data *ptp = efx->ptp_data;
1396 unsigned long expiry;
1397 int rc;
1398
1399 if (!efx_ptp_valid_unicast_event_pkt(skb))
1400 return -EINVAL;
1401
1402 expiry = jiffies + UCAST_FILTER_EXPIRY_JIFFIES;
1403
1404 if (skb->protocol == htons(ETH_P_IP)) {
1405 __be32 addr = ip_hdr(skb)->saddr;
1406
1407 rc = efx_ptp_insert_ipv4_filter(efx, filter_list: &ptp->rxfilters_ucast,
1408 addr, PTP_EVENT_PORT, expiry);
1409 if (rc < 0)
1410 goto out;
1411
1412 rc = efx_ptp_insert_ipv4_filter(efx, filter_list: &ptp->rxfilters_ucast,
1413 addr, PTP_GENERAL_PORT, expiry);
1414 } else if (efx_ptp_use_mac_tx_timestamps(efx)) {
1415 /* IPv6 PTP only supported by devices with MAC hw timestamp */
1416 struct in6_addr *addr = &ipv6_hdr(skb)->saddr;
1417
1418 rc = efx_ptp_insert_ipv6_filter(efx, filter_list: &ptp->rxfilters_ucast,
1419 addr, PTP_EVENT_PORT, expiry);
1420 if (rc < 0)
1421 goto out;
1422
1423 rc = efx_ptp_insert_ipv6_filter(efx, filter_list: &ptp->rxfilters_ucast,
1424 addr, PTP_GENERAL_PORT, expiry);
1425 } else {
1426 return -EOPNOTSUPP;
1427 }
1428
1429out:
1430 return rc;
1431}
1432
1433static int efx_ptp_start(struct efx_nic *efx)
1434{
1435 struct efx_ptp_data *ptp = efx->ptp_data;
1436 int rc;
1437
1438 ptp->reset_required = false;
1439
1440 rc = efx_ptp_insert_multicast_filters(efx);
1441 if (rc)
1442 return rc;
1443
1444 rc = efx_ptp_enable(efx);
1445 if (rc != 0)
1446 goto fail;
1447
1448 ptp->evt_frag_idx = 0;
1449 ptp->current_adjfreq = 0;
1450
1451 return 0;
1452
1453fail:
1454 efx_ptp_remove_filters(efx, filter_list: &ptp->rxfilters_mcast);
1455 return rc;
1456}
1457
1458static int efx_ptp_stop(struct efx_nic *efx)
1459{
1460 struct efx_ptp_data *ptp = efx->ptp_data;
1461 int rc;
1462
1463 if (ptp == NULL)
1464 return 0;
1465
1466 rc = efx_ptp_disable(efx);
1467
1468 efx_ptp_remove_filters(efx, filter_list: &ptp->rxfilters_mcast);
1469 efx_ptp_remove_filters(efx, filter_list: &ptp->rxfilters_ucast);
1470
1471 /* Make sure RX packets are really delivered */
1472 efx_ptp_deliver_rx_queue(q: &efx->ptp_data->rxq);
1473 skb_queue_purge(list: &efx->ptp_data->txq);
1474
1475 return rc;
1476}
1477
1478static int efx_ptp_restart(struct efx_nic *efx)
1479{
1480 if (efx->ptp_data && efx->ptp_data->enabled)
1481 return efx_ptp_start(efx);
1482 return 0;
1483}
1484
1485static void efx_ptp_pps_worker(struct work_struct *work)
1486{
1487 struct efx_ptp_data *ptp =
1488 container_of(work, struct efx_ptp_data, pps_work);
1489 struct efx_nic *efx = ptp->efx;
1490 struct ptp_clock_event ptp_evt;
1491
1492 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS))
1493 return;
1494
1495 ptp_evt.type = PTP_CLOCK_PPSUSR;
1496 ptp_evt.pps_times = ptp->host_time_pps;
1497 ptp_clock_event(ptp: ptp->phc_clock, event: &ptp_evt);
1498}
1499
1500static void efx_ptp_worker(struct work_struct *work)
1501{
1502 struct efx_ptp_data *ptp_data =
1503 container_of(work, struct efx_ptp_data, work);
1504 struct efx_nic *efx = ptp_data->efx;
1505 struct sk_buff *skb;
1506 struct sk_buff_head tempq;
1507
1508 if (ptp_data->reset_required) {
1509 efx_ptp_stop(efx);
1510 efx_ptp_start(efx);
1511 return;
1512 }
1513
1514 __skb_queue_head_init(list: &tempq);
1515 efx_ptp_process_events(efx, q: &tempq);
1516
1517 while ((skb = skb_dequeue(list: &ptp_data->txq)))
1518 ptp_data->xmit_skb(efx, skb);
1519
1520 while ((skb = __skb_dequeue(list: &tempq)))
1521 efx_ptp_process_rx(efx, skb);
1522}
1523
1524static void efx_ptp_cleanup_worker(struct work_struct *work)
1525{
1526 struct efx_ptp_data *ptp =
1527 container_of(work, struct efx_ptp_data, cleanup_work.work);
1528 struct efx_ptp_rxfilter *rxfilter, *tmp;
1529
1530 list_for_each_entry_safe(rxfilter, tmp, &ptp->rxfilters_ucast, list) {
1531 if (time_is_before_jiffies(rxfilter->expiry))
1532 efx_ptp_remove_one_filter(efx: ptp->efx, rxfilter);
1533 }
1534
1535 if (!list_empty(head: &ptp->rxfilters_ucast)) {
1536 queue_delayed_work(wq: ptp->workwq, dwork: &ptp->cleanup_work,
1537 UCAST_FILTER_EXPIRY_JIFFIES + 1);
1538 }
1539}
1540
1541static const struct ptp_clock_info efx_phc_clock_info = {
1542 .owner = THIS_MODULE,
1543 .name = "sfc",
1544 .max_adj = MAX_PPB,
1545 .n_alarm = 0,
1546 .n_ext_ts = 0,
1547 .n_per_out = 0,
1548 .n_pins = 0,
1549 .pps = 1,
1550 .adjfine = efx_phc_adjfine,
1551 .adjtime = efx_phc_adjtime,
1552 .gettime64 = efx_phc_gettime,
1553 .settime64 = efx_phc_settime,
1554 .enable = efx_phc_enable,
1555};
1556
1557/* Initialise PTP state. */
1558int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel)
1559{
1560 struct efx_ptp_data *ptp;
1561 int rc = 0;
1562
1563 if (efx->ptp_data) {
1564 efx->ptp_data->channel = channel;
1565 return 0;
1566 }
1567
1568 ptp = kzalloc(size: sizeof(struct efx_ptp_data), GFP_KERNEL);
1569 efx->ptp_data = ptp;
1570 if (!efx->ptp_data)
1571 return -ENOMEM;
1572
1573 ptp->efx = efx;
1574 ptp->channel = channel;
1575
1576 rc = efx_nic_alloc_buffer(efx, buffer: &ptp->start, len: sizeof(int), GFP_KERNEL);
1577 if (rc != 0)
1578 goto fail1;
1579
1580 skb_queue_head_init(list: &ptp->rxq);
1581 skb_queue_head_init(list: &ptp->txq);
1582 ptp->workwq = create_singlethread_workqueue("sfc_ptp");
1583 if (!ptp->workwq) {
1584 rc = -ENOMEM;
1585 goto fail2;
1586 }
1587
1588 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1589 ptp->xmit_skb = efx_ptp_xmit_skb_queue;
1590 /* Request sync events on this channel. */
1591 channel->sync_events_state = SYNC_EVENTS_QUIESCENT;
1592 } else {
1593 ptp->xmit_skb = efx_ptp_xmit_skb_mc;
1594 }
1595
1596 INIT_WORK(&ptp->work, efx_ptp_worker);
1597 INIT_DELAYED_WORK(&ptp->cleanup_work, efx_ptp_cleanup_worker);
1598 ptp->config.flags = 0;
1599 ptp->config.tx_type = HWTSTAMP_TX_OFF;
1600 ptp->config.rx_filter = HWTSTAMP_FILTER_NONE;
1601 INIT_LIST_HEAD(list: &ptp->rxfilters_mcast);
1602 INIT_LIST_HEAD(list: &ptp->rxfilters_ucast);
1603
1604 /* Get the NIC PTP attributes and set up time conversions */
1605 rc = efx_ptp_get_attributes(efx);
1606 if (rc < 0)
1607 goto fail3;
1608
1609 /* Get the timestamp corrections */
1610 rc = efx_ptp_get_timestamp_corrections(efx);
1611 if (rc < 0)
1612 goto fail3;
1613
1614 if (efx->mcdi->fn_flags &
1615 (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) {
1616 ptp->phc_clock_info = efx_phc_clock_info;
1617 ptp->phc_clock = ptp_clock_register(info: &ptp->phc_clock_info,
1618 parent: &efx->pci_dev->dev);
1619 if (IS_ERR(ptr: ptp->phc_clock)) {
1620 rc = PTR_ERR(ptr: ptp->phc_clock);
1621 goto fail3;
1622 } else if (ptp->phc_clock) {
1623 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker);
1624 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps");
1625 if (!ptp->pps_workwq) {
1626 rc = -ENOMEM;
1627 goto fail4;
1628 }
1629 }
1630 }
1631 ptp->nic_ts_enabled = false;
1632
1633 return 0;
1634fail4:
1635 ptp_clock_unregister(ptp: efx->ptp_data->phc_clock);
1636
1637fail3:
1638 destroy_workqueue(wq: efx->ptp_data->workwq);
1639
1640fail2:
1641 efx_nic_free_buffer(efx, buffer: &ptp->start);
1642
1643fail1:
1644 kfree(objp: efx->ptp_data);
1645 efx->ptp_data = NULL;
1646
1647 return rc;
1648}
1649
1650/* Initialise PTP channel.
1651 *
1652 * Setting core_index to zero causes the queue to be initialised and doesn't
1653 * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue.
1654 */
1655static int efx_ptp_probe_channel(struct efx_channel *channel)
1656{
1657 struct efx_nic *efx = channel->efx;
1658 int rc;
1659
1660 channel->irq_moderation_us = 0;
1661 channel->rx_queue.core_index = 0;
1662
1663 rc = efx_ptp_probe(efx, channel);
1664 /* Failure to probe PTP is not fatal; this channel will just not be
1665 * used for anything.
1666 * In the case of EPERM, efx_ptp_probe will print its own message (in
1667 * efx_ptp_get_attributes()), so we don't need to.
1668 */
1669 if (rc && rc != -EPERM)
1670 netif_warn(efx, drv, efx->net_dev,
1671 "Failed to probe PTP, rc=%d\n", rc);
1672 return 0;
1673}
1674
1675void efx_ptp_remove(struct efx_nic *efx)
1676{
1677 if (!efx->ptp_data)
1678 return;
1679
1680 (void)efx_ptp_disable(efx);
1681
1682 cancel_work_sync(work: &efx->ptp_data->work);
1683 cancel_delayed_work_sync(dwork: &efx->ptp_data->cleanup_work);
1684 if (efx->ptp_data->pps_workwq)
1685 cancel_work_sync(work: &efx->ptp_data->pps_work);
1686
1687 skb_queue_purge(list: &efx->ptp_data->rxq);
1688 skb_queue_purge(list: &efx->ptp_data->txq);
1689
1690 if (efx->ptp_data->phc_clock) {
1691 destroy_workqueue(wq: efx->ptp_data->pps_workwq);
1692 ptp_clock_unregister(ptp: efx->ptp_data->phc_clock);
1693 }
1694
1695 destroy_workqueue(wq: efx->ptp_data->workwq);
1696
1697 efx_nic_free_buffer(efx, buffer: &efx->ptp_data->start);
1698 kfree(objp: efx->ptp_data);
1699 efx->ptp_data = NULL;
1700}
1701
1702static void efx_ptp_remove_channel(struct efx_channel *channel)
1703{
1704 efx_ptp_remove(efx: channel->efx);
1705}
1706
1707static void efx_ptp_get_channel_name(struct efx_channel *channel,
1708 char *buf, size_t len)
1709{
1710 snprintf(buf, size: len, fmt: "%s-ptp", channel->efx->name);
1711}
1712
1713/* Determine whether this packet should be processed by the PTP module
1714 * or transmitted conventionally.
1715 */
1716bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1717{
1718 return efx->ptp_data &&
1719 efx->ptp_data->enabled &&
1720 skb->len >= PTP_MIN_LENGTH &&
1721 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM &&
1722 likely(skb->protocol == htons(ETH_P_IP)) &&
1723 skb_transport_header_was_set(skb) &&
1724 skb_network_header_len(skb) >= sizeof(struct iphdr) &&
1725 ip_hdr(skb)->protocol == IPPROTO_UDP &&
1726 skb_headlen(skb) >=
1727 skb_transport_offset(skb) + sizeof(struct udphdr) &&
1728 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT);
1729}
1730
1731/* Receive a PTP packet. Packets are queued until the arrival of
1732 * the receive timestamp from the MC - this will probably occur after the
1733 * packet arrival because of the processing in the MC.
1734 */
1735static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb)
1736{
1737 struct efx_nic *efx = channel->efx;
1738 struct efx_ptp_data *ptp = efx->ptp_data;
1739 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb;
1740 unsigned int version;
1741 u8 *data;
1742
1743 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS);
1744
1745 /* Correct version? */
1746 if (ptp->mode == MC_CMD_PTP_MODE_V1) {
1747 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) {
1748 return false;
1749 }
1750 data = skb->data;
1751 version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]);
1752 if (version != PTP_VERSION_V1) {
1753 return false;
1754 }
1755 } else {
1756 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) {
1757 return false;
1758 }
1759 data = skb->data;
1760 version = data[PTP_V2_VERSION_OFFSET];
1761 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) {
1762 return false;
1763 }
1764 }
1765
1766 /* Does this packet require timestamping? */
1767 if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) {
1768 match->state = PTP_PACKET_STATE_UNMATCHED;
1769
1770 /* We expect the sequence number to be in the same position in
1771 * the packet for PTP V1 and V2
1772 */
1773 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET);
1774 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH);
1775 } else {
1776 match->state = PTP_PACKET_STATE_MATCH_UNWANTED;
1777 }
1778
1779 skb_queue_tail(list: &ptp->rxq, newsk: skb);
1780 queue_work(wq: ptp->workwq, work: &ptp->work);
1781
1782 return true;
1783}
1784
1785/* Transmit a PTP packet. This has to be transmitted by the MC
1786 * itself, through an MCDI call. MCDI calls aren't permitted
1787 * in the transmit path so defer the actual transmission to a suitable worker.
1788 */
1789int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb)
1790{
1791 struct efx_ptp_data *ptp = efx->ptp_data;
1792
1793 skb_queue_tail(list: &ptp->txq, newsk: skb);
1794
1795 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) &&
1796 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM))
1797 efx_xmit_hwtstamp_pending(skb);
1798 queue_work(wq: ptp->workwq, work: &ptp->work);
1799
1800 return NETDEV_TX_OK;
1801}
1802
1803int efx_ptp_get_mode(struct efx_nic *efx)
1804{
1805 return efx->ptp_data->mode;
1806}
1807
1808int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted,
1809 unsigned int new_mode)
1810{
1811 if ((enable_wanted != efx->ptp_data->enabled) ||
1812 (enable_wanted && (efx->ptp_data->mode != new_mode))) {
1813 int rc = 0;
1814
1815 if (enable_wanted) {
1816 /* Change of mode requires disable */
1817 if (efx->ptp_data->enabled &&
1818 (efx->ptp_data->mode != new_mode)) {
1819 efx->ptp_data->enabled = false;
1820 rc = efx_ptp_stop(efx);
1821 if (rc != 0)
1822 return rc;
1823 }
1824
1825 /* Set new operating mode and establish
1826 * baseline synchronisation, which must
1827 * succeed.
1828 */
1829 efx->ptp_data->mode = new_mode;
1830 if (netif_running(dev: efx->net_dev))
1831 rc = efx_ptp_start(efx);
1832 if (rc == 0) {
1833 rc = efx_ptp_synchronize(efx,
1834 PTP_SYNC_ATTEMPTS * 2);
1835 if (rc != 0)
1836 efx_ptp_stop(efx);
1837 }
1838 } else {
1839 rc = efx_ptp_stop(efx);
1840 }
1841
1842 if (rc != 0)
1843 return rc;
1844
1845 efx->ptp_data->enabled = enable_wanted;
1846 }
1847
1848 return 0;
1849}
1850
1851static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init)
1852{
1853 int rc;
1854
1855 if ((init->tx_type != HWTSTAMP_TX_OFF) &&
1856 (init->tx_type != HWTSTAMP_TX_ON))
1857 return -ERANGE;
1858
1859 rc = efx->type->ptp_set_ts_config(efx, init);
1860 if (rc)
1861 return rc;
1862
1863 efx->ptp_data->config = *init;
1864 return 0;
1865}
1866
1867void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info)
1868{
1869 struct efx_ptp_data *ptp = efx->ptp_data;
1870 struct efx_nic *primary = efx->primary;
1871
1872 ASSERT_RTNL();
1873
1874 if (!ptp)
1875 return;
1876
1877 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE |
1878 SOF_TIMESTAMPING_RX_HARDWARE |
1879 SOF_TIMESTAMPING_RAW_HARDWARE);
1880 /* Check licensed features. If we don't have the license for TX
1881 * timestamps, the NIC will not support them.
1882 */
1883 if (efx_ptp_use_mac_tx_timestamps(efx)) {
1884 struct efx_ef10_nic_data *nic_data = efx->nic_data;
1885
1886 if (!(nic_data->licensed_features &
1887 (1 << LICENSED_V3_FEATURES_TX_TIMESTAMPS_LBN)))
1888 ts_info->so_timestamping &=
1889 ~SOF_TIMESTAMPING_TX_HARDWARE;
1890 }
1891 if (primary && primary->ptp_data && primary->ptp_data->phc_clock)
1892 ts_info->phc_index =
1893 ptp_clock_index(ptp: primary->ptp_data->phc_clock);
1894 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON;
1895 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters;
1896}
1897
1898int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1899{
1900 struct hwtstamp_config config;
1901 int rc;
1902
1903 /* Not a PTP enabled port */
1904 if (!efx->ptp_data)
1905 return -EOPNOTSUPP;
1906
1907 if (copy_from_user(to: &config, from: ifr->ifr_data, n: sizeof(config)))
1908 return -EFAULT;
1909
1910 rc = efx_ptp_ts_init(efx, init: &config);
1911 if (rc != 0)
1912 return rc;
1913
1914 return copy_to_user(to: ifr->ifr_data, from: &config, n: sizeof(config))
1915 ? -EFAULT : 0;
1916}
1917
1918int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr)
1919{
1920 if (!efx->ptp_data)
1921 return -EOPNOTSUPP;
1922
1923 return copy_to_user(to: ifr->ifr_data, from: &efx->ptp_data->config,
1924 n: sizeof(efx->ptp_data->config)) ? -EFAULT : 0;
1925}
1926
1927static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len)
1928{
1929 struct efx_ptp_data *ptp = efx->ptp_data;
1930
1931 netif_err(efx, hw, efx->net_dev,
1932 "PTP unexpected event length: got %d expected %d\n",
1933 ptp->evt_frag_idx, expected_frag_len);
1934 ptp->reset_required = true;
1935 queue_work(wq: ptp->workwq, work: &ptp->work);
1936}
1937
1938static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp)
1939{
1940 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA);
1941 if (ptp->evt_frag_idx != 1) {
1942 ptp_event_failure(efx, expected_frag_len: 1);
1943 return;
1944 }
1945
1946 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code);
1947}
1948
1949static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp)
1950{
1951 if (ptp->nic_ts_enabled)
1952 queue_work(wq: ptp->pps_workwq, work: &ptp->pps_work);
1953}
1954
1955void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev)
1956{
1957 struct efx_ptp_data *ptp = efx->ptp_data;
1958 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE);
1959
1960 if (!ptp) {
1961 if (!efx->ptp_warned) {
1962 netif_warn(efx, drv, efx->net_dev,
1963 "Received PTP event but PTP not set up\n");
1964 efx->ptp_warned = true;
1965 }
1966 return;
1967 }
1968
1969 if (!ptp->enabled)
1970 return;
1971
1972 if (ptp->evt_frag_idx == 0) {
1973 ptp->evt_code = code;
1974 } else if (ptp->evt_code != code) {
1975 netif_err(efx, hw, efx->net_dev,
1976 "PTP out of sequence event %d\n", code);
1977 ptp->evt_frag_idx = 0;
1978 }
1979
1980 ptp->evt_frags[ptp->evt_frag_idx++] = *ev;
1981 if (!MCDI_EVENT_FIELD(*ev, CONT)) {
1982 /* Process resulting event */
1983 switch (code) {
1984 case MCDI_EVENT_CODE_PTP_FAULT:
1985 ptp_event_fault(efx, ptp);
1986 break;
1987 case MCDI_EVENT_CODE_PTP_PPS:
1988 ptp_event_pps(efx, ptp);
1989 break;
1990 default:
1991 netif_err(efx, hw, efx->net_dev,
1992 "PTP unknown event %d\n", code);
1993 break;
1994 }
1995 ptp->evt_frag_idx = 0;
1996 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) {
1997 netif_err(efx, hw, efx->net_dev,
1998 "PTP too many event fragments\n");
1999 ptp->evt_frag_idx = 0;
2000 }
2001}
2002
2003void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev)
2004{
2005 struct efx_nic *efx = channel->efx;
2006 struct efx_ptp_data *ptp = efx->ptp_data;
2007
2008 /* When extracting the sync timestamp minor value, we should discard
2009 * the least significant two bits. These are not required in order
2010 * to reconstruct full-range timestamps and they are optionally used
2011 * to report status depending on the options supplied when subscribing
2012 * for sync events.
2013 */
2014 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR);
2015 channel->sync_timestamp_minor =
2016 (MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_MS_8BITS) & 0xFC)
2017 << ptp->nic_time.sync_event_minor_shift;
2018
2019 /* if sync events have been disabled then we want to silently ignore
2020 * this event, so throw away result.
2021 */
2022 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED,
2023 SYNC_EVENTS_VALID);
2024}
2025
2026static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh)
2027{
2028#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
2029 return __le32_to_cpup(p: (const __le32 *)(eh + efx->rx_packet_ts_offset));
2030#else
2031 const u8 *data = eh + efx->rx_packet_ts_offset;
2032 return (u32)data[0] |
2033 (u32)data[1] << 8 |
2034 (u32)data[2] << 16 |
2035 (u32)data[3] << 24;
2036#endif
2037}
2038
2039void __efx_rx_skb_attach_timestamp(struct efx_channel *channel,
2040 struct sk_buff *skb)
2041{
2042 struct efx_nic *efx = channel->efx;
2043 struct efx_ptp_data *ptp = efx->ptp_data;
2044 u32 pkt_timestamp_major, pkt_timestamp_minor;
2045 u32 diff, carry;
2046 struct skb_shared_hwtstamps *timestamps;
2047
2048 if (channel->sync_events_state != SYNC_EVENTS_VALID)
2049 return;
2050
2051 pkt_timestamp_minor = efx_rx_buf_timestamp_minor(efx, eh: skb_mac_header(skb));
2052
2053 /* get the difference between the packet and sync timestamps,
2054 * modulo one second
2055 */
2056 diff = pkt_timestamp_minor - channel->sync_timestamp_minor;
2057 if (pkt_timestamp_minor < channel->sync_timestamp_minor)
2058 diff += ptp->nic_time.minor_max;
2059
2060 /* do we roll over a second boundary and need to carry the one? */
2061 carry = (channel->sync_timestamp_minor >= ptp->nic_time.minor_max - diff) ?
2062 1 : 0;
2063
2064 if (diff <= ptp->nic_time.sync_event_diff_max) {
2065 /* packet is ahead of the sync event by a quarter of a second or
2066 * less (allowing for fuzz)
2067 */
2068 pkt_timestamp_major = channel->sync_timestamp_major + carry;
2069 } else if (diff >= ptp->nic_time.sync_event_diff_min) {
2070 /* packet is behind the sync event but within the fuzz factor.
2071 * This means the RX packet and sync event crossed as they were
2072 * placed on the event queue, which can sometimes happen.
2073 */
2074 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry;
2075 } else {
2076 /* it's outside tolerance in both directions. this might be
2077 * indicative of us missing sync events for some reason, so
2078 * we'll call it an error rather than risk giving a bogus
2079 * timestamp.
2080 */
2081 netif_vdbg(efx, drv, efx->net_dev,
2082 "packet timestamp %x too far from sync event %x:%x\n",
2083 pkt_timestamp_minor, channel->sync_timestamp_major,
2084 channel->sync_timestamp_minor);
2085 return;
2086 }
2087
2088 /* attach the timestamps to the skb */
2089 timestamps = skb_hwtstamps(skb);
2090 timestamps->hwtstamp =
2091 ptp->nic_to_kernel_time(pkt_timestamp_major,
2092 pkt_timestamp_minor,
2093 ptp->ts_corrections.general_rx);
2094}
2095
2096static int efx_phc_adjfine(struct ptp_clock_info *ptp, long scaled_ppm)
2097{
2098 struct efx_ptp_data *ptp_data = container_of(ptp,
2099 struct efx_ptp_data,
2100 phc_clock_info);
2101 s32 delta = scaled_ppm_to_ppb(ppm: scaled_ppm);
2102 struct efx_nic *efx = ptp_data->efx;
2103 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN);
2104 s64 adjustment_ns;
2105 int rc;
2106
2107 if (delta > MAX_PPB)
2108 delta = MAX_PPB;
2109 else if (delta < -MAX_PPB)
2110 delta = -MAX_PPB;
2111
2112 /* Convert ppb to fixed point ns taking care to round correctly. */
2113 adjustment_ns = ((s64)delta * PPB_SCALE_WORD +
2114 (1 << (ptp_data->adjfreq_ppb_shift - 1))) >>
2115 ptp_data->adjfreq_ppb_shift;
2116
2117 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2118 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0);
2119 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns);
2120 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0);
2121 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0);
2122 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf: inadj, inlen: sizeof(inadj),
2123 NULL, outlen: 0, NULL);
2124 if (rc != 0)
2125 return rc;
2126
2127 ptp_data->current_adjfreq = adjustment_ns;
2128 return 0;
2129}
2130
2131static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta)
2132{
2133 u32 nic_major, nic_minor;
2134 struct efx_ptp_data *ptp_data = container_of(ptp,
2135 struct efx_ptp_data,
2136 phc_clock_info);
2137 struct efx_nic *efx = ptp_data->efx;
2138 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN);
2139
2140 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor);
2141
2142 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST);
2143 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2144 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq);
2145 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major);
2146 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor);
2147 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
2148 NULL, outlen: 0, NULL);
2149}
2150
2151static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
2152{
2153 struct efx_ptp_data *ptp_data = container_of(ptp,
2154 struct efx_ptp_data,
2155 phc_clock_info);
2156 struct efx_nic *efx = ptp_data->efx;
2157 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN);
2158 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN);
2159 int rc;
2160 ktime_t kt;
2161
2162 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME);
2163 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0);
2164
2165 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, inlen: sizeof(inbuf),
2166 outbuf, outlen: sizeof(outbuf), NULL);
2167 if (rc != 0)
2168 return rc;
2169
2170 kt = ptp_data->nic_to_kernel_time(
2171 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR),
2172 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0);
2173 *ts = ktime_to_timespec64(kt);
2174 return 0;
2175}
2176
2177static int efx_phc_settime(struct ptp_clock_info *ptp,
2178 const struct timespec64 *e_ts)
2179{
2180 /* Get the current NIC time, efx_phc_gettime.
2181 * Subtract from the desired time to get the offset
2182 * call efx_phc_adjtime with the offset
2183 */
2184 int rc;
2185 struct timespec64 time_now;
2186 struct timespec64 delta;
2187
2188 rc = efx_phc_gettime(ptp, ts: &time_now);
2189 if (rc != 0)
2190 return rc;
2191
2192 delta = timespec64_sub(lhs: *e_ts, rhs: time_now);
2193
2194 rc = efx_phc_adjtime(ptp, delta: timespec64_to_ns(ts: &delta));
2195 if (rc != 0)
2196 return rc;
2197
2198 return 0;
2199}
2200
2201static int efx_phc_enable(struct ptp_clock_info *ptp,
2202 struct ptp_clock_request *request,
2203 int enable)
2204{
2205 struct efx_ptp_data *ptp_data = container_of(ptp,
2206 struct efx_ptp_data,
2207 phc_clock_info);
2208 if (request->type != PTP_CLK_REQ_PPS)
2209 return -EOPNOTSUPP;
2210
2211 ptp_data->nic_ts_enabled = !!enable;
2212 return 0;
2213}
2214
2215static const struct efx_channel_type efx_ptp_channel_type = {
2216 .handle_no_channel = efx_ptp_handle_no_channel,
2217 .pre_probe = efx_ptp_probe_channel,
2218 .post_remove = efx_ptp_remove_channel,
2219 .get_name = efx_ptp_get_channel_name,
2220 .copy = efx_copy_channel,
2221 .receive_skb = efx_ptp_rx,
2222 .want_txqs = efx_ptp_want_txqs,
2223 .keep_eventq = false,
2224};
2225
2226void efx_ptp_defer_probe_with_channel(struct efx_nic *efx)
2227{
2228 /* Check whether PTP is implemented on this NIC. The DISABLE
2229 * operation will succeed if and only if it is implemented.
2230 */
2231 if (efx_ptp_disable(efx) == 0)
2232 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] =
2233 &efx_ptp_channel_type;
2234}
2235
2236void efx_ptp_start_datapath(struct efx_nic *efx)
2237{
2238 if (efx_ptp_restart(efx))
2239 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n");
2240 /* re-enable timestamping if it was previously enabled */
2241 if (efx->type->ptp_set_ts_sync_events)
2242 efx->type->ptp_set_ts_sync_events(efx, true, true);
2243}
2244
2245void efx_ptp_stop_datapath(struct efx_nic *efx)
2246{
2247 /* temporarily disable timestamping */
2248 if (efx->type->ptp_set_ts_sync_events)
2249 efx->type->ptp_set_ts_sync_events(efx, false, true);
2250 efx_ptp_stop(efx);
2251}
2252

source code of linux/drivers/net/ethernet/sfc/ptp.c