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 131 132 enum 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 */ 168 struct 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 */ 180 struct 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 */ 199 struct 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 */ 279 struct 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 }; 344 345 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta); 346 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta); 347 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts); 348 static int efx_phc_settime(struct ptp_clock_info *ptp, 349 const struct timespec64 *e_ts); 350 static int efx_phc_enable(struct ptp_clock_info *ptp, 351 struct ptp_clock_request *request, int on); 352 353 bool 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 */ 366 static 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 } 375 static 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) 392 static const unsigned long efx_ptp_stat_mask[] = { 393 [0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL, 394 }; 395 396 size_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 } 404 405 size_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 */ 441 static 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 } 447 448 static 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 */ 471 static 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 } 489 490 static 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 } 496 497 static 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. */ 514 static 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 } 521 522 static 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 } 538 539 struct efx_channel *efx_ptp_channel(struct efx_nic *efx) 540 { 541 return efx->ptp_data ? efx->ptp_data->channel : NULL; 542 } 543 544 static 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 */ 557 static ktime_t 558 efx_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 u32 sync_timestamp; 564 ktime_t kt = { 0 }; 565 s16 delta; 566 567 if (!(nic_major & 0x80000000)) { 568 WARN_ON_ONCE(nic_major >> 16); 569 570 /* Medford provides 48 bits of timestamp, so we must get the top 571 * 16 bits from the timesync event state. 572 * 573 * We only have the lower 16 bits of the time now, but we do 574 * have a full resolution timestamp at some point in past. As 575 * long as the difference between the (real) now and the sync 576 * is less than 2^15, then we can reconstruct the difference 577 * between those two numbers using only the lower 16 bits of 578 * each. 579 * 580 * Put another way 581 * 582 * a - b = ((a mod k) - b) mod k 583 * 584 * when -k/2 < (a-b) < k/2. In our case k is 2^16. We know 585 * (a mod k) and b, so can calculate the delta, a - b. 586 * 587 */ 588 sync_timestamp = last_sync_timestamp_major(efx); 589 590 /* Because delta is s16 this does an implicit mask down to 591 * 16 bits which is what we need, assuming 592 * MEDFORD_TX_SECS_EVENT_BITS is 16. delta is signed so that 593 * we can deal with the (unlikely) case of sync timestamps 594 * arriving from the future. 595 */ 596 delta = nic_major - sync_timestamp; 597 598 /* Recover the fully specified time now, by applying the offset 599 * to the (fully specified) sync time. 600 */ 601 nic_major = sync_timestamp + delta; 602 603 kt = ptp->nic_to_kernel_time(nic_major, nic_minor, 604 correction); 605 } 606 return kt; 607 } 608 609 ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue) 610 { 611 struct efx_nic *efx = tx_queue->efx; 612 struct efx_ptp_data *ptp = efx->ptp_data; 613 ktime_t kt; 614 615 if (efx_ptp_use_mac_tx_timestamps(efx)) 616 kt = efx_ptp_mac_nic_to_ktime_correction(efx, ptp, 617 tx_queue->completed_timestamp_major, 618 tx_queue->completed_timestamp_minor, 619 ptp->ts_corrections.general_tx); 620 else 621 kt = ptp->nic_to_kernel_time( 622 tx_queue->completed_timestamp_major, 623 tx_queue->completed_timestamp_minor, 624 ptp->ts_corrections.general_tx); 625 return kt; 626 } 627 628 /* Get PTP attributes and set up time conversions */ 629 static int efx_ptp_get_attributes(struct efx_nic *efx) 630 { 631 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN); 632 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN); 633 struct efx_ptp_data *ptp = efx->ptp_data; 634 int rc; 635 u32 fmt; 636 size_t out_len; 637 638 /* Get the PTP attributes. If the NIC doesn't support the operation we 639 * use the default format for compatibility with older NICs i.e. 640 * seconds and nanoseconds. 641 */ 642 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES); 643 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 644 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 645 outbuf, sizeof(outbuf), &out_len); 646 if (rc == 0) { 647 fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT); 648 } else if (rc == -EINVAL) { 649 fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS; 650 } else if (rc == -EPERM) { 651 netif_info(efx, probe, efx->net_dev, "no PTP support\n"); 652 return rc; 653 } else { 654 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), 655 outbuf, sizeof(outbuf), rc); 656 return rc; 657 } 658 659 switch (fmt) { 660 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION: 661 ptp->ns_to_nic_time = efx_ptp_ns_to_s27; 662 ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction; 663 ptp->nic_time.minor_max = 1 << 27; 664 ptp->nic_time.sync_event_minor_shift = 19; 665 break; 666 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS: 667 ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns; 668 ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction; 669 ptp->nic_time.minor_max = 1000000000; 670 ptp->nic_time.sync_event_minor_shift = 22; 671 break; 672 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_QTR_NANOSECONDS: 673 ptp->ns_to_nic_time = efx_ptp_ns_to_s_qns; 674 ptp->nic_to_kernel_time = efx_ptp_s_qns_to_ktime_correction; 675 ptp->nic_time.minor_max = 4000000000UL; 676 ptp->nic_time.sync_event_minor_shift = 24; 677 break; 678 default: 679 return -ERANGE; 680 } 681 682 /* Precalculate acceptable difference between the minor time in the 683 * packet prefix and the last MCDI time sync event. We expect the 684 * packet prefix timestamp to be after of sync event by up to one 685 * sync event interval (0.25s) but we allow it to exceed this by a 686 * fuzz factor of (0.1s) 687 */ 688 ptp->nic_time.sync_event_diff_min = ptp->nic_time.minor_max 689 - (ptp->nic_time.minor_max / 10); 690 ptp->nic_time.sync_event_diff_max = (ptp->nic_time.minor_max / 4) 691 + (ptp->nic_time.minor_max / 10); 692 693 /* MC_CMD_PTP_OP_GET_ATTRIBUTES has been extended twice from an older 694 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT. The function now may return 695 * a value to use for the minimum acceptable corrected synchronization 696 * window and may return further capabilities. 697 * If we have the extra information store it. For older firmware that 698 * does not implement the extended command use the default value. 699 */ 700 if (rc == 0 && 701 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_CAPABILITIES_OFST) 702 ptp->min_synchronisation_ns = 703 MCDI_DWORD(outbuf, 704 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN); 705 else 706 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS; 707 708 if (rc == 0 && 709 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN) 710 ptp->capabilities = MCDI_DWORD(outbuf, 711 PTP_OUT_GET_ATTRIBUTES_CAPABILITIES); 712 else 713 ptp->capabilities = 0; 714 715 /* Set up the shift for conversion between frequency 716 * adjustments in parts-per-billion and the fixed-point 717 * fractional ns format that the adapter uses. 718 */ 719 if (ptp->capabilities & (1 << MC_CMD_PTP_OUT_GET_ATTRIBUTES_FP44_FREQ_ADJ_LBN)) 720 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP44; 721 else 722 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP40; 723 724 return 0; 725 } 726 727 /* Get PTP timestamp corrections */ 728 static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx) 729 { 730 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN); 731 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN); 732 int rc; 733 size_t out_len; 734 735 /* Get the timestamp corrections from the NIC. If this operation is 736 * not supported (older NICs) then no correction is required. 737 */ 738 MCDI_SET_DWORD(inbuf, PTP_IN_OP, 739 MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS); 740 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 741 742 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 743 outbuf, sizeof(outbuf), &out_len); 744 if (rc == 0) { 745 efx->ptp_data->ts_corrections.ptp_tx = MCDI_DWORD(outbuf, 746 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT); 747 efx->ptp_data->ts_corrections.ptp_rx = MCDI_DWORD(outbuf, 748 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE); 749 efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf, 750 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT); 751 efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf, 752 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN); 753 754 if (out_len >= MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN) { 755 efx->ptp_data->ts_corrections.general_tx = MCDI_DWORD( 756 outbuf, 757 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_TX); 758 efx->ptp_data->ts_corrections.general_rx = MCDI_DWORD( 759 outbuf, 760 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_RX); 761 } else { 762 efx->ptp_data->ts_corrections.general_tx = 763 efx->ptp_data->ts_corrections.ptp_tx; 764 efx->ptp_data->ts_corrections.general_rx = 765 efx->ptp_data->ts_corrections.ptp_rx; 766 } 767 } else if (rc == -EINVAL) { 768 efx->ptp_data->ts_corrections.ptp_tx = 0; 769 efx->ptp_data->ts_corrections.ptp_rx = 0; 770 efx->ptp_data->ts_corrections.pps_out = 0; 771 efx->ptp_data->ts_corrections.pps_in = 0; 772 efx->ptp_data->ts_corrections.general_tx = 0; 773 efx->ptp_data->ts_corrections.general_rx = 0; 774 } else { 775 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), outbuf, 776 sizeof(outbuf), rc); 777 return rc; 778 } 779 780 return 0; 781 } 782 783 /* Enable MCDI PTP support. */ 784 static int efx_ptp_enable(struct efx_nic *efx) 785 { 786 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN); 787 MCDI_DECLARE_BUF_ERR(outbuf); 788 int rc; 789 790 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE); 791 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 792 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE, 793 efx->ptp_data->channel ? 794 efx->ptp_data->channel->channel : 0); 795 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode); 796 797 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 798 outbuf, sizeof(outbuf), NULL); 799 rc = (rc == -EALREADY) ? 0 : rc; 800 if (rc) 801 efx_mcdi_display_error(efx, MC_CMD_PTP, 802 MC_CMD_PTP_IN_ENABLE_LEN, 803 outbuf, sizeof(outbuf), rc); 804 return rc; 805 } 806 807 /* Disable MCDI PTP support. 808 * 809 * Note that this function should never rely on the presence of ptp_data - 810 * may be called before that exists. 811 */ 812 static int efx_ptp_disable(struct efx_nic *efx) 813 { 814 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN); 815 MCDI_DECLARE_BUF_ERR(outbuf); 816 int rc; 817 818 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE); 819 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 820 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 821 outbuf, sizeof(outbuf), NULL); 822 rc = (rc == -EALREADY) ? 0 : rc; 823 /* If we get ENOSYS, the NIC doesn't support PTP, and thus this function 824 * should only have been called during probe. 825 */ 826 if (rc == -ENOSYS || rc == -EPERM) 827 netif_info(efx, probe, efx->net_dev, "no PTP support\n"); 828 else if (rc) 829 efx_mcdi_display_error(efx, MC_CMD_PTP, 830 MC_CMD_PTP_IN_DISABLE_LEN, 831 outbuf, sizeof(outbuf), rc); 832 return rc; 833 } 834 835 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q) 836 { 837 struct sk_buff *skb; 838 839 while ((skb = skb_dequeue(q))) { 840 local_bh_disable(); 841 netif_receive_skb(skb); 842 local_bh_enable(); 843 } 844 } 845 846 static void efx_ptp_handle_no_channel(struct efx_nic *efx) 847 { 848 netif_err(efx, drv, efx->net_dev, 849 "ERROR: PTP requires MSI-X and 1 additional interrupt" 850 "vector. PTP disabled\n"); 851 } 852 853 /* Repeatedly send the host time to the MC which will capture the hardware 854 * time. 855 */ 856 static void efx_ptp_send_times(struct efx_nic *efx, 857 struct pps_event_time *last_time) 858 { 859 struct pps_event_time now; 860 struct timespec64 limit; 861 struct efx_ptp_data *ptp = efx->ptp_data; 862 int *mc_running = ptp->start.addr; 863 864 pps_get_ts(&now); 865 limit = now.ts_real; 866 timespec64_add_ns(&limit, SYNCHRONISE_PERIOD_NS); 867 868 /* Write host time for specified period or until MC is done */ 869 while ((timespec64_compare(&now.ts_real, &limit) < 0) && 870 READ_ONCE(*mc_running)) { 871 struct timespec64 update_time; 872 unsigned int host_time; 873 874 /* Don't update continuously to avoid saturating the PCIe bus */ 875 update_time = now.ts_real; 876 timespec64_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS); 877 do { 878 pps_get_ts(&now); 879 } while ((timespec64_compare(&now.ts_real, &update_time) < 0) && 880 READ_ONCE(*mc_running)); 881 882 /* Synchronise NIC with single word of time only */ 883 host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS | 884 now.ts_real.tv_nsec); 885 /* Update host time in NIC memory */ 886 efx->type->ptp_write_host_time(efx, host_time); 887 } 888 *last_time = now; 889 } 890 891 /* Read a timeset from the MC's results and partial process. */ 892 static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data), 893 struct efx_ptp_timeset *timeset) 894 { 895 unsigned start_ns, end_ns; 896 897 timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART); 898 timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR); 899 timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR); 900 timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND), 901 timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS); 902 903 /* Ignore seconds */ 904 start_ns = timeset->host_start & MC_NANOSECOND_MASK; 905 end_ns = timeset->host_end & MC_NANOSECOND_MASK; 906 /* Allow for rollover */ 907 if (end_ns < start_ns) 908 end_ns += NSEC_PER_SEC; 909 /* Determine duration of operation */ 910 timeset->window = end_ns - start_ns; 911 } 912 913 /* Process times received from MC. 914 * 915 * Extract times from returned results, and establish the minimum value 916 * seen. The minimum value represents the "best" possible time and events 917 * too much greater than this are rejected - the machine is, perhaps, too 918 * busy. A number of readings are taken so that, hopefully, at least one good 919 * synchronisation will be seen in the results. 920 */ 921 static int 922 efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf), 923 size_t response_length, 924 const struct pps_event_time *last_time) 925 { 926 unsigned number_readings = 927 MCDI_VAR_ARRAY_LEN(response_length, 928 PTP_OUT_SYNCHRONIZE_TIMESET); 929 unsigned i; 930 unsigned ngood = 0; 931 unsigned last_good = 0; 932 struct efx_ptp_data *ptp = efx->ptp_data; 933 u32 last_sec; 934 u32 start_sec; 935 struct timespec64 delta; 936 ktime_t mc_time; 937 938 if (number_readings == 0) 939 return -EAGAIN; 940 941 /* Read the set of results and find the last good host-MC 942 * synchronization result. The MC times when it finishes reading the 943 * host time so the corrected window time should be fairly constant 944 * for a given platform. Increment stats for any results that appear 945 * to be erroneous. 946 */ 947 for (i = 0; i < number_readings; i++) { 948 s32 window, corrected; 949 struct timespec64 wait; 950 951 efx_ptp_read_timeset( 952 MCDI_ARRAY_STRUCT_PTR(synch_buf, 953 PTP_OUT_SYNCHRONIZE_TIMESET, i), 954 &ptp->timeset[i]); 955 956 wait = ktime_to_timespec64( 957 ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0)); 958 window = ptp->timeset[i].window; 959 corrected = window - wait.tv_nsec; 960 961 /* We expect the uncorrected synchronization window to be at 962 * least as large as the interval between host start and end 963 * times. If it is smaller than this then this is mostly likely 964 * to be a consequence of the host's time being adjusted. 965 * Check that the corrected sync window is in a reasonable 966 * range. If it is out of range it is likely to be because an 967 * interrupt or other delay occurred between reading the system 968 * time and writing it to MC memory. 969 */ 970 if (window < SYNCHRONISATION_GRANULARITY_NS) { 971 ++ptp->invalid_sync_windows; 972 } else if (corrected >= MAX_SYNCHRONISATION_NS) { 973 ++ptp->oversize_sync_windows; 974 } else if (corrected < ptp->min_synchronisation_ns) { 975 ++ptp->undersize_sync_windows; 976 } else { 977 ngood++; 978 last_good = i; 979 } 980 } 981 982 if (ngood == 0) { 983 netif_warn(efx, drv, efx->net_dev, 984 "PTP no suitable synchronisations\n"); 985 return -EAGAIN; 986 } 987 988 /* Calculate delay from last good sync (host time) to last_time. 989 * It is possible that the seconds rolled over between taking 990 * the start reading and the last value written by the host. The 991 * timescales are such that a gap of more than one second is never 992 * expected. delta is *not* normalised. 993 */ 994 start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS; 995 last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK; 996 if (start_sec != last_sec && 997 ((start_sec + 1) & MC_SECOND_MASK) != last_sec) { 998 netif_warn(efx, hw, efx->net_dev, 999 "PTP bad synchronisation seconds\n"); 1000 return -EAGAIN; 1001 } 1002 delta.tv_sec = (last_sec - start_sec) & 1; 1003 delta.tv_nsec = 1004 last_time->ts_real.tv_nsec - 1005 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK); 1006 1007 /* Convert the NIC time at last good sync into kernel time. 1008 * No correction is required - this time is the output of a 1009 * firmware process. 1010 */ 1011 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major, 1012 ptp->timeset[last_good].minor, 0); 1013 1014 /* Calculate delay from NIC top of second to last_time */ 1015 delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec; 1016 1017 /* Set PPS timestamp to match NIC top of second */ 1018 ptp->host_time_pps = *last_time; 1019 pps_sub_ts(&ptp->host_time_pps, delta); 1020 1021 return 0; 1022 } 1023 1024 /* Synchronize times between the host and the MC */ 1025 static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings) 1026 { 1027 struct efx_ptp_data *ptp = efx->ptp_data; 1028 MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX); 1029 size_t response_length; 1030 int rc; 1031 unsigned long timeout; 1032 struct pps_event_time last_time = {}; 1033 unsigned int loops = 0; 1034 int *start = ptp->start.addr; 1035 1036 MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE); 1037 MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0); 1038 MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS, 1039 num_readings); 1040 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR, 1041 ptp->start.dma_addr); 1042 1043 /* Clear flag that signals MC ready */ 1044 WRITE_ONCE(*start, 0); 1045 rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf, 1046 MC_CMD_PTP_IN_SYNCHRONIZE_LEN); 1047 EFX_WARN_ON_ONCE_PARANOID(rc); 1048 1049 /* Wait for start from MCDI (or timeout) */ 1050 timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS); 1051 while (!READ_ONCE(*start) && (time_before(jiffies, timeout))) { 1052 udelay(20); /* Usually start MCDI execution quickly */ 1053 loops++; 1054 } 1055 1056 if (loops <= 1) 1057 ++ptp->fast_syncs; 1058 if (!time_before(jiffies, timeout)) 1059 ++ptp->sync_timeouts; 1060 1061 if (READ_ONCE(*start)) 1062 efx_ptp_send_times(efx, &last_time); 1063 1064 /* Collect results */ 1065 rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP, 1066 MC_CMD_PTP_IN_SYNCHRONIZE_LEN, 1067 synch_buf, sizeof(synch_buf), 1068 &response_length); 1069 if (rc == 0) { 1070 rc = efx_ptp_process_times(efx, synch_buf, response_length, 1071 &last_time); 1072 if (rc == 0) 1073 ++ptp->good_syncs; 1074 else 1075 ++ptp->no_time_syncs; 1076 } 1077 1078 /* Increment the bad syncs counter if the synchronize fails, whatever 1079 * the reason. 1080 */ 1081 if (rc != 0) 1082 ++ptp->bad_syncs; 1083 1084 return rc; 1085 } 1086 1087 /* Transmit a PTP packet via the dedicated hardware timestamped queue. */ 1088 static void efx_ptp_xmit_skb_queue(struct efx_nic *efx, struct sk_buff *skb) 1089 { 1090 struct efx_ptp_data *ptp_data = efx->ptp_data; 1091 struct efx_tx_queue *tx_queue; 1092 u8 type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0; 1093 1094 tx_queue = &ptp_data->channel->tx_queue[type]; 1095 if (tx_queue && tx_queue->timestamping) { 1096 efx_enqueue_skb(tx_queue, skb); 1097 } else { 1098 WARN_ONCE(1, "PTP channel has no timestamped tx queue\n"); 1099 dev_kfree_skb_any(skb); 1100 } 1101 } 1102 1103 /* Transmit a PTP packet, via the MCDI interface, to the wire. */ 1104 static void efx_ptp_xmit_skb_mc(struct efx_nic *efx, struct sk_buff *skb) 1105 { 1106 struct efx_ptp_data *ptp_data = efx->ptp_data; 1107 struct skb_shared_hwtstamps timestamps; 1108 int rc = -EIO; 1109 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN); 1110 size_t len; 1111 1112 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT); 1113 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0); 1114 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len); 1115 if (skb_shinfo(skb)->nr_frags != 0) { 1116 rc = skb_linearize(skb); 1117 if (rc != 0) 1118 goto fail; 1119 } 1120 1121 if (skb->ip_summed == CHECKSUM_PARTIAL) { 1122 rc = skb_checksum_help(skb); 1123 if (rc != 0) 1124 goto fail; 1125 } 1126 skb_copy_from_linear_data(skb, 1127 MCDI_PTR(ptp_data->txbuf, 1128 PTP_IN_TRANSMIT_PACKET), 1129 skb->len); 1130 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, 1131 ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len), 1132 txtime, sizeof(txtime), &len); 1133 if (rc != 0) 1134 goto fail; 1135 1136 memset(×tamps, 0, sizeof(timestamps)); 1137 timestamps.hwtstamp = ptp_data->nic_to_kernel_time( 1138 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR), 1139 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR), 1140 ptp_data->ts_corrections.ptp_tx); 1141 1142 skb_tstamp_tx(skb, ×tamps); 1143 1144 rc = 0; 1145 1146 fail: 1147 dev_kfree_skb_any(skb); 1148 1149 return; 1150 } 1151 1152 static void efx_ptp_drop_time_expired_events(struct efx_nic *efx) 1153 { 1154 struct efx_ptp_data *ptp = efx->ptp_data; 1155 struct list_head *cursor; 1156 struct list_head *next; 1157 1158 if (ptp->rx_ts_inline) 1159 return; 1160 1161 /* Drop time-expired events */ 1162 spin_lock_bh(&ptp->evt_lock); 1163 if (!list_empty(&ptp->evt_list)) { 1164 list_for_each_safe(cursor, next, &ptp->evt_list) { 1165 struct efx_ptp_event_rx *evt; 1166 1167 evt = list_entry(cursor, struct efx_ptp_event_rx, 1168 link); 1169 if (time_after(jiffies, evt->expiry)) { 1170 list_move(&evt->link, &ptp->evt_free_list); 1171 netif_warn(efx, hw, efx->net_dev, 1172 "PTP rx event dropped\n"); 1173 } 1174 } 1175 } 1176 spin_unlock_bh(&ptp->evt_lock); 1177 } 1178 1179 static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx, 1180 struct sk_buff *skb) 1181 { 1182 struct efx_ptp_data *ptp = efx->ptp_data; 1183 bool evts_waiting; 1184 struct list_head *cursor; 1185 struct list_head *next; 1186 struct efx_ptp_match *match; 1187 enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED; 1188 1189 WARN_ON_ONCE(ptp->rx_ts_inline); 1190 1191 spin_lock_bh(&ptp->evt_lock); 1192 evts_waiting = !list_empty(&ptp->evt_list); 1193 spin_unlock_bh(&ptp->evt_lock); 1194 1195 if (!evts_waiting) 1196 return PTP_PACKET_STATE_UNMATCHED; 1197 1198 match = (struct efx_ptp_match *)skb->cb; 1199 /* Look for a matching timestamp in the event queue */ 1200 spin_lock_bh(&ptp->evt_lock); 1201 list_for_each_safe(cursor, next, &ptp->evt_list) { 1202 struct efx_ptp_event_rx *evt; 1203 1204 evt = list_entry(cursor, struct efx_ptp_event_rx, link); 1205 if ((evt->seq0 == match->words[0]) && 1206 (evt->seq1 == match->words[1])) { 1207 struct skb_shared_hwtstamps *timestamps; 1208 1209 /* Match - add in hardware timestamp */ 1210 timestamps = skb_hwtstamps(skb); 1211 timestamps->hwtstamp = evt->hwtimestamp; 1212 1213 match->state = PTP_PACKET_STATE_MATCHED; 1214 rc = PTP_PACKET_STATE_MATCHED; 1215 list_move(&evt->link, &ptp->evt_free_list); 1216 break; 1217 } 1218 } 1219 spin_unlock_bh(&ptp->evt_lock); 1220 1221 return rc; 1222 } 1223 1224 /* Process any queued receive events and corresponding packets 1225 * 1226 * q is returned with all the packets that are ready for delivery. 1227 */ 1228 static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q) 1229 { 1230 struct efx_ptp_data *ptp = efx->ptp_data; 1231 struct sk_buff *skb; 1232 1233 while ((skb = skb_dequeue(&ptp->rxq))) { 1234 struct efx_ptp_match *match; 1235 1236 match = (struct efx_ptp_match *)skb->cb; 1237 if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) { 1238 __skb_queue_tail(q, skb); 1239 } else if (efx_ptp_match_rx(efx, skb) == 1240 PTP_PACKET_STATE_MATCHED) { 1241 __skb_queue_tail(q, skb); 1242 } else if (time_after(jiffies, match->expiry)) { 1243 match->state = PTP_PACKET_STATE_TIMED_OUT; 1244 ++ptp->rx_no_timestamp; 1245 __skb_queue_tail(q, skb); 1246 } else { 1247 /* Replace unprocessed entry and stop */ 1248 skb_queue_head(&ptp->rxq, skb); 1249 break; 1250 } 1251 } 1252 } 1253 1254 /* Complete processing of a received packet */ 1255 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb) 1256 { 1257 local_bh_disable(); 1258 netif_receive_skb(skb); 1259 local_bh_enable(); 1260 } 1261 1262 static void efx_ptp_remove_multicast_filters(struct efx_nic *efx) 1263 { 1264 struct efx_ptp_data *ptp = efx->ptp_data; 1265 1266 if (ptp->rxfilter_installed) { 1267 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED, 1268 ptp->rxfilter_general); 1269 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED, 1270 ptp->rxfilter_event); 1271 ptp->rxfilter_installed = false; 1272 } 1273 } 1274 1275 static int efx_ptp_insert_multicast_filters(struct efx_nic *efx) 1276 { 1277 struct efx_ptp_data *ptp = efx->ptp_data; 1278 struct efx_filter_spec rxfilter; 1279 int rc; 1280 1281 if (!ptp->channel || ptp->rxfilter_installed) 1282 return 0; 1283 1284 /* Must filter on both event and general ports to ensure 1285 * that there is no packet re-ordering. 1286 */ 1287 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0, 1288 efx_rx_queue_index( 1289 efx_channel_get_rx_queue(ptp->channel))); 1290 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP, 1291 htonl(PTP_ADDRESS), 1292 htons(PTP_EVENT_PORT)); 1293 if (rc != 0) 1294 return rc; 1295 1296 rc = efx_filter_insert_filter(efx, &rxfilter, true); 1297 if (rc < 0) 1298 return rc; 1299 ptp->rxfilter_event = rc; 1300 1301 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0, 1302 efx_rx_queue_index( 1303 efx_channel_get_rx_queue(ptp->channel))); 1304 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP, 1305 htonl(PTP_ADDRESS), 1306 htons(PTP_GENERAL_PORT)); 1307 if (rc != 0) 1308 goto fail; 1309 1310 rc = efx_filter_insert_filter(efx, &rxfilter, true); 1311 if (rc < 0) 1312 goto fail; 1313 ptp->rxfilter_general = rc; 1314 1315 ptp->rxfilter_installed = true; 1316 return 0; 1317 1318 fail: 1319 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED, 1320 ptp->rxfilter_event); 1321 return rc; 1322 } 1323 1324 static int efx_ptp_start(struct efx_nic *efx) 1325 { 1326 struct efx_ptp_data *ptp = efx->ptp_data; 1327 int rc; 1328 1329 ptp->reset_required = false; 1330 1331 rc = efx_ptp_insert_multicast_filters(efx); 1332 if (rc) 1333 return rc; 1334 1335 rc = efx_ptp_enable(efx); 1336 if (rc != 0) 1337 goto fail; 1338 1339 ptp->evt_frag_idx = 0; 1340 ptp->current_adjfreq = 0; 1341 1342 return 0; 1343 1344 fail: 1345 efx_ptp_remove_multicast_filters(efx); 1346 return rc; 1347 } 1348 1349 static int efx_ptp_stop(struct efx_nic *efx) 1350 { 1351 struct efx_ptp_data *ptp = efx->ptp_data; 1352 struct list_head *cursor; 1353 struct list_head *next; 1354 int rc; 1355 1356 if (ptp == NULL) 1357 return 0; 1358 1359 rc = efx_ptp_disable(efx); 1360 1361 efx_ptp_remove_multicast_filters(efx); 1362 1363 /* Make sure RX packets are really delivered */ 1364 efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq); 1365 skb_queue_purge(&efx->ptp_data->txq); 1366 1367 /* Drop any pending receive events */ 1368 spin_lock_bh(&efx->ptp_data->evt_lock); 1369 list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) { 1370 list_move(cursor, &efx->ptp_data->evt_free_list); 1371 } 1372 spin_unlock_bh(&efx->ptp_data->evt_lock); 1373 1374 return rc; 1375 } 1376 1377 static int efx_ptp_restart(struct efx_nic *efx) 1378 { 1379 if (efx->ptp_data && efx->ptp_data->enabled) 1380 return efx_ptp_start(efx); 1381 return 0; 1382 } 1383 1384 static void efx_ptp_pps_worker(struct work_struct *work) 1385 { 1386 struct efx_ptp_data *ptp = 1387 container_of(work, struct efx_ptp_data, pps_work); 1388 struct efx_nic *efx = ptp->efx; 1389 struct ptp_clock_event ptp_evt; 1390 1391 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS)) 1392 return; 1393 1394 ptp_evt.type = PTP_CLOCK_PPSUSR; 1395 ptp_evt.pps_times = ptp->host_time_pps; 1396 ptp_clock_event(ptp->phc_clock, &ptp_evt); 1397 } 1398 1399 static void efx_ptp_worker(struct work_struct *work) 1400 { 1401 struct efx_ptp_data *ptp_data = 1402 container_of(work, struct efx_ptp_data, work); 1403 struct efx_nic *efx = ptp_data->efx; 1404 struct sk_buff *skb; 1405 struct sk_buff_head tempq; 1406 1407 if (ptp_data->reset_required) { 1408 efx_ptp_stop(efx); 1409 efx_ptp_start(efx); 1410 return; 1411 } 1412 1413 efx_ptp_drop_time_expired_events(efx); 1414 1415 __skb_queue_head_init(&tempq); 1416 efx_ptp_process_events(efx, &tempq); 1417 1418 while ((skb = skb_dequeue(&ptp_data->txq))) 1419 ptp_data->xmit_skb(efx, skb); 1420 1421 while ((skb = __skb_dequeue(&tempq))) 1422 efx_ptp_process_rx(efx, skb); 1423 } 1424 1425 static const struct ptp_clock_info efx_phc_clock_info = { 1426 .owner = THIS_MODULE, 1427 .name = "sfc", 1428 .max_adj = MAX_PPB, 1429 .n_alarm = 0, 1430 .n_ext_ts = 0, 1431 .n_per_out = 0, 1432 .n_pins = 0, 1433 .pps = 1, 1434 .adjfreq = efx_phc_adjfreq, 1435 .adjtime = efx_phc_adjtime, 1436 .gettime64 = efx_phc_gettime, 1437 .settime64 = efx_phc_settime, 1438 .enable = efx_phc_enable, 1439 }; 1440 1441 /* Initialise PTP state. */ 1442 int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel) 1443 { 1444 struct efx_ptp_data *ptp; 1445 int rc = 0; 1446 unsigned int pos; 1447 1448 ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL); 1449 efx->ptp_data = ptp; 1450 if (!efx->ptp_data) 1451 return -ENOMEM; 1452 1453 ptp->efx = efx; 1454 ptp->channel = channel; 1455 ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0; 1456 1457 rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL); 1458 if (rc != 0) 1459 goto fail1; 1460 1461 skb_queue_head_init(&ptp->rxq); 1462 skb_queue_head_init(&ptp->txq); 1463 ptp->workwq = create_singlethread_workqueue("sfc_ptp"); 1464 if (!ptp->workwq) { 1465 rc = -ENOMEM; 1466 goto fail2; 1467 } 1468 1469 if (efx_ptp_use_mac_tx_timestamps(efx)) { 1470 ptp->xmit_skb = efx_ptp_xmit_skb_queue; 1471 /* Request sync events on this channel. */ 1472 channel->sync_events_state = SYNC_EVENTS_QUIESCENT; 1473 } else { 1474 ptp->xmit_skb = efx_ptp_xmit_skb_mc; 1475 } 1476 1477 INIT_WORK(&ptp->work, efx_ptp_worker); 1478 ptp->config.flags = 0; 1479 ptp->config.tx_type = HWTSTAMP_TX_OFF; 1480 ptp->config.rx_filter = HWTSTAMP_FILTER_NONE; 1481 INIT_LIST_HEAD(&ptp->evt_list); 1482 INIT_LIST_HEAD(&ptp->evt_free_list); 1483 spin_lock_init(&ptp->evt_lock); 1484 for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++) 1485 list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list); 1486 1487 /* Get the NIC PTP attributes and set up time conversions */ 1488 rc = efx_ptp_get_attributes(efx); 1489 if (rc < 0) 1490 goto fail3; 1491 1492 /* Get the timestamp corrections */ 1493 rc = efx_ptp_get_timestamp_corrections(efx); 1494 if (rc < 0) 1495 goto fail3; 1496 1497 if (efx->mcdi->fn_flags & 1498 (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) { 1499 ptp->phc_clock_info = efx_phc_clock_info; 1500 ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info, 1501 &efx->pci_dev->dev); 1502 if (IS_ERR(ptp->phc_clock)) { 1503 rc = PTR_ERR(ptp->phc_clock); 1504 goto fail3; 1505 } else if (ptp->phc_clock) { 1506 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker); 1507 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps"); 1508 if (!ptp->pps_workwq) { 1509 rc = -ENOMEM; 1510 goto fail4; 1511 } 1512 } 1513 } 1514 ptp->nic_ts_enabled = false; 1515 1516 return 0; 1517 fail4: 1518 ptp_clock_unregister(efx->ptp_data->phc_clock); 1519 1520 fail3: 1521 destroy_workqueue(efx->ptp_data->workwq); 1522 1523 fail2: 1524 efx_nic_free_buffer(efx, &ptp->start); 1525 1526 fail1: 1527 kfree(efx->ptp_data); 1528 efx->ptp_data = NULL; 1529 1530 return rc; 1531 } 1532 1533 /* Initialise PTP channel. 1534 * 1535 * Setting core_index to zero causes the queue to be initialised and doesn't 1536 * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue. 1537 */ 1538 static int efx_ptp_probe_channel(struct efx_channel *channel) 1539 { 1540 struct efx_nic *efx = channel->efx; 1541 int rc; 1542 1543 channel->irq_moderation_us = 0; 1544 channel->rx_queue.core_index = 0; 1545 1546 rc = efx_ptp_probe(efx, channel); 1547 /* Failure to probe PTP is not fatal; this channel will just not be 1548 * used for anything. 1549 * In the case of EPERM, efx_ptp_probe will print its own message (in 1550 * efx_ptp_get_attributes()), so we don't need to. 1551 */ 1552 if (rc && rc != -EPERM) 1553 netif_warn(efx, drv, efx->net_dev, 1554 "Failed to probe PTP, rc=%d\n", rc); 1555 return 0; 1556 } 1557 1558 void efx_ptp_remove(struct efx_nic *efx) 1559 { 1560 if (!efx->ptp_data) 1561 return; 1562 1563 (void)efx_ptp_disable(efx); 1564 1565 cancel_work_sync(&efx->ptp_data->work); 1566 if (efx->ptp_data->pps_workwq) 1567 cancel_work_sync(&efx->ptp_data->pps_work); 1568 1569 skb_queue_purge(&efx->ptp_data->rxq); 1570 skb_queue_purge(&efx->ptp_data->txq); 1571 1572 if (efx->ptp_data->phc_clock) { 1573 destroy_workqueue(efx->ptp_data->pps_workwq); 1574 ptp_clock_unregister(efx->ptp_data->phc_clock); 1575 } 1576 1577 destroy_workqueue(efx->ptp_data->workwq); 1578 1579 efx_nic_free_buffer(efx, &efx->ptp_data->start); 1580 kfree(efx->ptp_data); 1581 efx->ptp_data = NULL; 1582 } 1583 1584 static void efx_ptp_remove_channel(struct efx_channel *channel) 1585 { 1586 efx_ptp_remove(channel->efx); 1587 } 1588 1589 static void efx_ptp_get_channel_name(struct efx_channel *channel, 1590 char *buf, size_t len) 1591 { 1592 snprintf(buf, len, "%s-ptp", channel->efx->name); 1593 } 1594 1595 /* Determine whether this packet should be processed by the PTP module 1596 * or transmitted conventionally. 1597 */ 1598 bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb) 1599 { 1600 return efx->ptp_data && 1601 efx->ptp_data->enabled && 1602 skb->len >= PTP_MIN_LENGTH && 1603 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM && 1604 likely(skb->protocol == htons(ETH_P_IP)) && 1605 skb_transport_header_was_set(skb) && 1606 skb_network_header_len(skb) >= sizeof(struct iphdr) && 1607 ip_hdr(skb)->protocol == IPPROTO_UDP && 1608 skb_headlen(skb) >= 1609 skb_transport_offset(skb) + sizeof(struct udphdr) && 1610 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT); 1611 } 1612 1613 /* Receive a PTP packet. Packets are queued until the arrival of 1614 * the receive timestamp from the MC - this will probably occur after the 1615 * packet arrival because of the processing in the MC. 1616 */ 1617 static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb) 1618 { 1619 struct efx_nic *efx = channel->efx; 1620 struct efx_ptp_data *ptp = efx->ptp_data; 1621 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb; 1622 u8 *match_data_012, *match_data_345; 1623 unsigned int version; 1624 u8 *data; 1625 1626 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS); 1627 1628 /* Correct version? */ 1629 if (ptp->mode == MC_CMD_PTP_MODE_V1) { 1630 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) { 1631 return false; 1632 } 1633 data = skb->data; 1634 version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]); 1635 if (version != PTP_VERSION_V1) { 1636 return false; 1637 } 1638 1639 /* PTP V1 uses all six bytes of the UUID to match the packet 1640 * to the timestamp 1641 */ 1642 match_data_012 = data + PTP_V1_UUID_OFFSET; 1643 match_data_345 = data + PTP_V1_UUID_OFFSET + 3; 1644 } else { 1645 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) { 1646 return false; 1647 } 1648 data = skb->data; 1649 version = data[PTP_V2_VERSION_OFFSET]; 1650 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) { 1651 return false; 1652 } 1653 1654 /* The original V2 implementation uses bytes 2-7 of 1655 * the UUID to match the packet to the timestamp. This 1656 * discards two of the bytes of the MAC address used 1657 * to create the UUID (SF bug 33070). The PTP V2 1658 * enhanced mode fixes this issue and uses bytes 0-2 1659 * and byte 5-7 of the UUID. 1660 */ 1661 match_data_345 = data + PTP_V2_UUID_OFFSET + 5; 1662 if (ptp->mode == MC_CMD_PTP_MODE_V2) { 1663 match_data_012 = data + PTP_V2_UUID_OFFSET + 2; 1664 } else { 1665 match_data_012 = data + PTP_V2_UUID_OFFSET + 0; 1666 BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED); 1667 } 1668 } 1669 1670 /* Does this packet require timestamping? */ 1671 if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) { 1672 match->state = PTP_PACKET_STATE_UNMATCHED; 1673 1674 /* We expect the sequence number to be in the same position in 1675 * the packet for PTP V1 and V2 1676 */ 1677 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET); 1678 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH); 1679 1680 /* Extract UUID/Sequence information */ 1681 match->words[0] = (match_data_012[0] | 1682 (match_data_012[1] << 8) | 1683 (match_data_012[2] << 16) | 1684 (match_data_345[0] << 24)); 1685 match->words[1] = (match_data_345[1] | 1686 (match_data_345[2] << 8) | 1687 (data[PTP_V1_SEQUENCE_OFFSET + 1688 PTP_V1_SEQUENCE_LENGTH - 1] << 1689 16)); 1690 } else { 1691 match->state = PTP_PACKET_STATE_MATCH_UNWANTED; 1692 } 1693 1694 skb_queue_tail(&ptp->rxq, skb); 1695 queue_work(ptp->workwq, &ptp->work); 1696 1697 return true; 1698 } 1699 1700 /* Transmit a PTP packet. This has to be transmitted by the MC 1701 * itself, through an MCDI call. MCDI calls aren't permitted 1702 * in the transmit path so defer the actual transmission to a suitable worker. 1703 */ 1704 int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb) 1705 { 1706 struct efx_ptp_data *ptp = efx->ptp_data; 1707 1708 skb_queue_tail(&ptp->txq, skb); 1709 1710 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) && 1711 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM)) 1712 efx_xmit_hwtstamp_pending(skb); 1713 queue_work(ptp->workwq, &ptp->work); 1714 1715 return NETDEV_TX_OK; 1716 } 1717 1718 int efx_ptp_get_mode(struct efx_nic *efx) 1719 { 1720 return efx->ptp_data->mode; 1721 } 1722 1723 int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted, 1724 unsigned int new_mode) 1725 { 1726 if ((enable_wanted != efx->ptp_data->enabled) || 1727 (enable_wanted && (efx->ptp_data->mode != new_mode))) { 1728 int rc = 0; 1729 1730 if (enable_wanted) { 1731 /* Change of mode requires disable */ 1732 if (efx->ptp_data->enabled && 1733 (efx->ptp_data->mode != new_mode)) { 1734 efx->ptp_data->enabled = false; 1735 rc = efx_ptp_stop(efx); 1736 if (rc != 0) 1737 return rc; 1738 } 1739 1740 /* Set new operating mode and establish 1741 * baseline synchronisation, which must 1742 * succeed. 1743 */ 1744 efx->ptp_data->mode = new_mode; 1745 if (netif_running(efx->net_dev)) 1746 rc = efx_ptp_start(efx); 1747 if (rc == 0) { 1748 rc = efx_ptp_synchronize(efx, 1749 PTP_SYNC_ATTEMPTS * 2); 1750 if (rc != 0) 1751 efx_ptp_stop(efx); 1752 } 1753 } else { 1754 rc = efx_ptp_stop(efx); 1755 } 1756 1757 if (rc != 0) 1758 return rc; 1759 1760 efx->ptp_data->enabled = enable_wanted; 1761 } 1762 1763 return 0; 1764 } 1765 1766 static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init) 1767 { 1768 int rc; 1769 1770 if (init->flags) 1771 return -EINVAL; 1772 1773 if ((init->tx_type != HWTSTAMP_TX_OFF) && 1774 (init->tx_type != HWTSTAMP_TX_ON)) 1775 return -ERANGE; 1776 1777 rc = efx->type->ptp_set_ts_config(efx, init); 1778 if (rc) 1779 return rc; 1780 1781 efx->ptp_data->config = *init; 1782 return 0; 1783 } 1784 1785 void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info) 1786 { 1787 struct efx_ptp_data *ptp = efx->ptp_data; 1788 struct efx_nic *primary = efx->primary; 1789 1790 ASSERT_RTNL(); 1791 1792 if (!ptp) 1793 return; 1794 1795 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE | 1796 SOF_TIMESTAMPING_RX_HARDWARE | 1797 SOF_TIMESTAMPING_RAW_HARDWARE); 1798 /* Check licensed features. If we don't have the license for TX 1799 * timestamps, the NIC will not support them. 1800 */ 1801 if (efx_ptp_use_mac_tx_timestamps(efx)) { 1802 struct efx_ef10_nic_data *nic_data = efx->nic_data; 1803 1804 if (!(nic_data->licensed_features & 1805 (1 << LICENSED_V3_FEATURES_TX_TIMESTAMPS_LBN))) 1806 ts_info->so_timestamping &= 1807 ~SOF_TIMESTAMPING_TX_HARDWARE; 1808 } 1809 if (primary && primary->ptp_data && primary->ptp_data->phc_clock) 1810 ts_info->phc_index = 1811 ptp_clock_index(primary->ptp_data->phc_clock); 1812 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON; 1813 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters; 1814 } 1815 1816 int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr) 1817 { 1818 struct hwtstamp_config config; 1819 int rc; 1820 1821 /* Not a PTP enabled port */ 1822 if (!efx->ptp_data) 1823 return -EOPNOTSUPP; 1824 1825 if (copy_from_user(&config, ifr->ifr_data, sizeof(config))) 1826 return -EFAULT; 1827 1828 rc = efx_ptp_ts_init(efx, &config); 1829 if (rc != 0) 1830 return rc; 1831 1832 return copy_to_user(ifr->ifr_data, &config, sizeof(config)) 1833 ? -EFAULT : 0; 1834 } 1835 1836 int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr) 1837 { 1838 if (!efx->ptp_data) 1839 return -EOPNOTSUPP; 1840 1841 return copy_to_user(ifr->ifr_data, &efx->ptp_data->config, 1842 sizeof(efx->ptp_data->config)) ? -EFAULT : 0; 1843 } 1844 1845 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len) 1846 { 1847 struct efx_ptp_data *ptp = efx->ptp_data; 1848 1849 netif_err(efx, hw, efx->net_dev, 1850 "PTP unexpected event length: got %d expected %d\n", 1851 ptp->evt_frag_idx, expected_frag_len); 1852 ptp->reset_required = true; 1853 queue_work(ptp->workwq, &ptp->work); 1854 } 1855 1856 /* Process a completed receive event. Put it on the event queue and 1857 * start worker thread. This is required because event and their 1858 * correspoding packets may come in either order. 1859 */ 1860 static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp) 1861 { 1862 struct efx_ptp_event_rx *evt = NULL; 1863 1864 if (WARN_ON_ONCE(ptp->rx_ts_inline)) 1865 return; 1866 1867 if (ptp->evt_frag_idx != 3) { 1868 ptp_event_failure(efx, 3); 1869 return; 1870 } 1871 1872 spin_lock_bh(&ptp->evt_lock); 1873 if (!list_empty(&ptp->evt_free_list)) { 1874 evt = list_first_entry(&ptp->evt_free_list, 1875 struct efx_ptp_event_rx, link); 1876 list_del(&evt->link); 1877 1878 evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA); 1879 evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2], 1880 MCDI_EVENT_SRC) | 1881 (EFX_QWORD_FIELD(ptp->evt_frags[1], 1882 MCDI_EVENT_SRC) << 8) | 1883 (EFX_QWORD_FIELD(ptp->evt_frags[0], 1884 MCDI_EVENT_SRC) << 16)); 1885 evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time( 1886 EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA), 1887 EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA), 1888 ptp->ts_corrections.ptp_rx); 1889 evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS); 1890 list_add_tail(&evt->link, &ptp->evt_list); 1891 1892 queue_work(ptp->workwq, &ptp->work); 1893 } else if (net_ratelimit()) { 1894 /* Log a rate-limited warning message. */ 1895 netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n"); 1896 } 1897 spin_unlock_bh(&ptp->evt_lock); 1898 } 1899 1900 static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp) 1901 { 1902 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA); 1903 if (ptp->evt_frag_idx != 1) { 1904 ptp_event_failure(efx, 1); 1905 return; 1906 } 1907 1908 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code); 1909 } 1910 1911 static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp) 1912 { 1913 if (ptp->nic_ts_enabled) 1914 queue_work(ptp->pps_workwq, &ptp->pps_work); 1915 } 1916 1917 void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev) 1918 { 1919 struct efx_ptp_data *ptp = efx->ptp_data; 1920 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE); 1921 1922 if (!ptp) { 1923 if (!efx->ptp_warned) { 1924 netif_warn(efx, drv, efx->net_dev, 1925 "Received PTP event but PTP not set up\n"); 1926 efx->ptp_warned = true; 1927 } 1928 return; 1929 } 1930 1931 if (!ptp->enabled) 1932 return; 1933 1934 if (ptp->evt_frag_idx == 0) { 1935 ptp->evt_code = code; 1936 } else if (ptp->evt_code != code) { 1937 netif_err(efx, hw, efx->net_dev, 1938 "PTP out of sequence event %d\n", code); 1939 ptp->evt_frag_idx = 0; 1940 } 1941 1942 ptp->evt_frags[ptp->evt_frag_idx++] = *ev; 1943 if (!MCDI_EVENT_FIELD(*ev, CONT)) { 1944 /* Process resulting event */ 1945 switch (code) { 1946 case MCDI_EVENT_CODE_PTP_RX: 1947 ptp_event_rx(efx, ptp); 1948 break; 1949 case MCDI_EVENT_CODE_PTP_FAULT: 1950 ptp_event_fault(efx, ptp); 1951 break; 1952 case MCDI_EVENT_CODE_PTP_PPS: 1953 ptp_event_pps(efx, ptp); 1954 break; 1955 default: 1956 netif_err(efx, hw, efx->net_dev, 1957 "PTP unknown event %d\n", code); 1958 break; 1959 } 1960 ptp->evt_frag_idx = 0; 1961 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) { 1962 netif_err(efx, hw, efx->net_dev, 1963 "PTP too many event fragments\n"); 1964 ptp->evt_frag_idx = 0; 1965 } 1966 } 1967 1968 void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev) 1969 { 1970 struct efx_nic *efx = channel->efx; 1971 struct efx_ptp_data *ptp = efx->ptp_data; 1972 1973 /* When extracting the sync timestamp minor value, we should discard 1974 * the least significant two bits. These are not required in order 1975 * to reconstruct full-range timestamps and they are optionally used 1976 * to report status depending on the options supplied when subscribing 1977 * for sync events. 1978 */ 1979 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR); 1980 channel->sync_timestamp_minor = 1981 (MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_MS_8BITS) & 0xFC) 1982 << ptp->nic_time.sync_event_minor_shift; 1983 1984 /* if sync events have been disabled then we want to silently ignore 1985 * this event, so throw away result. 1986 */ 1987 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED, 1988 SYNC_EVENTS_VALID); 1989 } 1990 1991 static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh) 1992 { 1993 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) 1994 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset)); 1995 #else 1996 const u8 *data = eh + efx->rx_packet_ts_offset; 1997 return (u32)data[0] | 1998 (u32)data[1] << 8 | 1999 (u32)data[2] << 16 | 2000 (u32)data[3] << 24; 2001 #endif 2002 } 2003 2004 void __efx_rx_skb_attach_timestamp(struct efx_channel *channel, 2005 struct sk_buff *skb) 2006 { 2007 struct efx_nic *efx = channel->efx; 2008 struct efx_ptp_data *ptp = efx->ptp_data; 2009 u32 pkt_timestamp_major, pkt_timestamp_minor; 2010 u32 diff, carry; 2011 struct skb_shared_hwtstamps *timestamps; 2012 2013 if (channel->sync_events_state != SYNC_EVENTS_VALID) 2014 return; 2015 2016 pkt_timestamp_minor = efx_rx_buf_timestamp_minor(efx, skb_mac_header(skb)); 2017 2018 /* get the difference between the packet and sync timestamps, 2019 * modulo one second 2020 */ 2021 diff = pkt_timestamp_minor - channel->sync_timestamp_minor; 2022 if (pkt_timestamp_minor < channel->sync_timestamp_minor) 2023 diff += ptp->nic_time.minor_max; 2024 2025 /* do we roll over a second boundary and need to carry the one? */ 2026 carry = (channel->sync_timestamp_minor >= ptp->nic_time.minor_max - diff) ? 2027 1 : 0; 2028 2029 if (diff <= ptp->nic_time.sync_event_diff_max) { 2030 /* packet is ahead of the sync event by a quarter of a second or 2031 * less (allowing for fuzz) 2032 */ 2033 pkt_timestamp_major = channel->sync_timestamp_major + carry; 2034 } else if (diff >= ptp->nic_time.sync_event_diff_min) { 2035 /* packet is behind the sync event but within the fuzz factor. 2036 * This means the RX packet and sync event crossed as they were 2037 * placed on the event queue, which can sometimes happen. 2038 */ 2039 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry; 2040 } else { 2041 /* it's outside tolerance in both directions. this might be 2042 * indicative of us missing sync events for some reason, so 2043 * we'll call it an error rather than risk giving a bogus 2044 * timestamp. 2045 */ 2046 netif_vdbg(efx, drv, efx->net_dev, 2047 "packet timestamp %x too far from sync event %x:%x\n", 2048 pkt_timestamp_minor, channel->sync_timestamp_major, 2049 channel->sync_timestamp_minor); 2050 return; 2051 } 2052 2053 /* attach the timestamps to the skb */ 2054 timestamps = skb_hwtstamps(skb); 2055 timestamps->hwtstamp = 2056 ptp->nic_to_kernel_time(pkt_timestamp_major, 2057 pkt_timestamp_minor, 2058 ptp->ts_corrections.general_rx); 2059 } 2060 2061 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta) 2062 { 2063 struct efx_ptp_data *ptp_data = container_of(ptp, 2064 struct efx_ptp_data, 2065 phc_clock_info); 2066 struct efx_nic *efx = ptp_data->efx; 2067 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN); 2068 s64 adjustment_ns; 2069 int rc; 2070 2071 if (delta > MAX_PPB) 2072 delta = MAX_PPB; 2073 else if (delta < -MAX_PPB) 2074 delta = -MAX_PPB; 2075 2076 /* Convert ppb to fixed point ns taking care to round correctly. */ 2077 adjustment_ns = ((s64)delta * PPB_SCALE_WORD + 2078 (1 << (ptp_data->adjfreq_ppb_shift - 1))) >> 2079 ptp_data->adjfreq_ppb_shift; 2080 2081 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST); 2082 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0); 2083 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns); 2084 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0); 2085 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0); 2086 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj), 2087 NULL, 0, NULL); 2088 if (rc != 0) 2089 return rc; 2090 2091 ptp_data->current_adjfreq = adjustment_ns; 2092 return 0; 2093 } 2094 2095 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta) 2096 { 2097 u32 nic_major, nic_minor; 2098 struct efx_ptp_data *ptp_data = container_of(ptp, 2099 struct efx_ptp_data, 2100 phc_clock_info); 2101 struct efx_nic *efx = ptp_data->efx; 2102 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN); 2103 2104 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor); 2105 2106 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST); 2107 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 2108 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq); 2109 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major); 2110 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor); 2111 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 2112 NULL, 0, NULL); 2113 } 2114 2115 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts) 2116 { 2117 struct efx_ptp_data *ptp_data = container_of(ptp, 2118 struct efx_ptp_data, 2119 phc_clock_info); 2120 struct efx_nic *efx = ptp_data->efx; 2121 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN); 2122 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN); 2123 int rc; 2124 ktime_t kt; 2125 2126 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME); 2127 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 2128 2129 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 2130 outbuf, sizeof(outbuf), NULL); 2131 if (rc != 0) 2132 return rc; 2133 2134 kt = ptp_data->nic_to_kernel_time( 2135 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR), 2136 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0); 2137 *ts = ktime_to_timespec64(kt); 2138 return 0; 2139 } 2140 2141 static int efx_phc_settime(struct ptp_clock_info *ptp, 2142 const struct timespec64 *e_ts) 2143 { 2144 /* Get the current NIC time, efx_phc_gettime. 2145 * Subtract from the desired time to get the offset 2146 * call efx_phc_adjtime with the offset 2147 */ 2148 int rc; 2149 struct timespec64 time_now; 2150 struct timespec64 delta; 2151 2152 rc = efx_phc_gettime(ptp, &time_now); 2153 if (rc != 0) 2154 return rc; 2155 2156 delta = timespec64_sub(*e_ts, time_now); 2157 2158 rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta)); 2159 if (rc != 0) 2160 return rc; 2161 2162 return 0; 2163 } 2164 2165 static int efx_phc_enable(struct ptp_clock_info *ptp, 2166 struct ptp_clock_request *request, 2167 int enable) 2168 { 2169 struct efx_ptp_data *ptp_data = container_of(ptp, 2170 struct efx_ptp_data, 2171 phc_clock_info); 2172 if (request->type != PTP_CLK_REQ_PPS) 2173 return -EOPNOTSUPP; 2174 2175 ptp_data->nic_ts_enabled = !!enable; 2176 return 0; 2177 } 2178 2179 static const struct efx_channel_type efx_ptp_channel_type = { 2180 .handle_no_channel = efx_ptp_handle_no_channel, 2181 .pre_probe = efx_ptp_probe_channel, 2182 .post_remove = efx_ptp_remove_channel, 2183 .get_name = efx_ptp_get_channel_name, 2184 /* no copy operation; there is no need to reallocate this channel */ 2185 .receive_skb = efx_ptp_rx, 2186 .want_txqs = efx_ptp_want_txqs, 2187 .keep_eventq = false, 2188 }; 2189 2190 void efx_ptp_defer_probe_with_channel(struct efx_nic *efx) 2191 { 2192 /* Check whether PTP is implemented on this NIC. The DISABLE 2193 * operation will succeed if and only if it is implemented. 2194 */ 2195 if (efx_ptp_disable(efx) == 0) 2196 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] = 2197 &efx_ptp_channel_type; 2198 } 2199 2200 void efx_ptp_start_datapath(struct efx_nic *efx) 2201 { 2202 if (efx_ptp_restart(efx)) 2203 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n"); 2204 /* re-enable timestamping if it was previously enabled */ 2205 if (efx->type->ptp_set_ts_sync_events) 2206 efx->type->ptp_set_ts_sync_events(efx, true, true); 2207 } 2208 2209 void efx_ptp_stop_datapath(struct efx_nic *efx) 2210 { 2211 /* temporarily disable timestamping */ 2212 if (efx->type->ptp_set_ts_sync_events) 2213 efx->type->ptp_set_ts_sync_events(efx, false, true); 2214 efx_ptp_stop(efx); 2215 } 2216