1 /**************************************************************************** 2 * Driver for Solarflare network controllers and boards 3 * Copyright 2011-2013 Solarflare Communications Inc. 4 * 5 * This program is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 as published 7 * by the Free Software Foundation, incorporated herein by reference. 8 */ 9 10 /* Theory of operation: 11 * 12 * PTP support is assisted by firmware running on the MC, which provides 13 * the hardware timestamping capabilities. Both transmitted and received 14 * PTP event packets are queued onto internal queues for subsequent processing; 15 * this is because the MC operations are relatively long and would block 16 * block NAPI/interrupt operation. 17 * 18 * Receive event processing: 19 * The event contains the packet's UUID and sequence number, together 20 * with the hardware timestamp. The PTP receive packet queue is searched 21 * for this UUID/sequence number and, if found, put on a pending queue. 22 * Packets not matching are delivered without timestamps (MCDI events will 23 * always arrive after the actual packet). 24 * It is important for the operation of the PTP protocol that the ordering 25 * of packets between the event and general port is maintained. 26 * 27 * Work queue processing: 28 * If work waiting, synchronise host/hardware time 29 * 30 * Transmit: send packet through MC, which returns the transmission time 31 * that is converted to an appropriate timestamp. 32 * 33 * Receive: the packet's reception time is converted to an appropriate 34 * timestamp. 35 */ 36 #include <linux/ip.h> 37 #include <linux/udp.h> 38 #include <linux/time.h> 39 #include <linux/ktime.h> 40 #include <linux/module.h> 41 #include <linux/net_tstamp.h> 42 #include <linux/pps_kernel.h> 43 #include <linux/ptp_clock_kernel.h> 44 #include "net_driver.h" 45 #include "efx.h" 46 #include "mcdi.h" 47 #include "mcdi_pcol.h" 48 #include "io.h" 49 #include "farch_regs.h" 50 #include "nic.h" 51 52 /* Maximum number of events expected to make up a PTP event */ 53 #define MAX_EVENT_FRAGS 3 54 55 /* Maximum delay, ms, to begin synchronisation */ 56 #define MAX_SYNCHRONISE_WAIT_MS 2 57 58 /* How long, at most, to spend synchronising */ 59 #define SYNCHRONISE_PERIOD_NS 250000 60 61 /* How often to update the shared memory time */ 62 #define SYNCHRONISATION_GRANULARITY_NS 200 63 64 /* Minimum permitted length of a (corrected) synchronisation time */ 65 #define DEFAULT_MIN_SYNCHRONISATION_NS 120 66 67 /* Maximum permitted length of a (corrected) synchronisation time */ 68 #define MAX_SYNCHRONISATION_NS 1000 69 70 /* How many (MC) receive events that can be queued */ 71 #define MAX_RECEIVE_EVENTS 8 72 73 /* Length of (modified) moving average. */ 74 #define AVERAGE_LENGTH 16 75 76 /* How long an unmatched event or packet can be held */ 77 #define PKT_EVENT_LIFETIME_MS 10 78 79 /* Offsets into PTP packet for identification. These offsets are from the 80 * start of the IP header, not the MAC header. Note that neither PTP V1 nor 81 * PTP V2 permit the use of IPV4 options. 82 */ 83 #define PTP_DPORT_OFFSET 22 84 85 #define PTP_V1_VERSION_LENGTH 2 86 #define PTP_V1_VERSION_OFFSET 28 87 88 #define PTP_V1_UUID_LENGTH 6 89 #define PTP_V1_UUID_OFFSET 50 90 91 #define PTP_V1_SEQUENCE_LENGTH 2 92 #define PTP_V1_SEQUENCE_OFFSET 58 93 94 /* The minimum length of a PTP V1 packet for offsets, etc. to be valid: 95 * includes IP header. 96 */ 97 #define PTP_V1_MIN_LENGTH 64 98 99 #define PTP_V2_VERSION_LENGTH 1 100 #define PTP_V2_VERSION_OFFSET 29 101 102 #define PTP_V2_UUID_LENGTH 8 103 #define PTP_V2_UUID_OFFSET 48 104 105 /* Although PTP V2 UUIDs are comprised a ClockIdentity (8) and PortNumber (2), 106 * the MC only captures the last six bytes of the clock identity. These values 107 * reflect those, not the ones used in the standard. The standard permits 108 * mapping of V1 UUIDs to V2 UUIDs with these same values. 109 */ 110 #define PTP_V2_MC_UUID_LENGTH 6 111 #define PTP_V2_MC_UUID_OFFSET 50 112 113 #define PTP_V2_SEQUENCE_LENGTH 2 114 #define PTP_V2_SEQUENCE_OFFSET 58 115 116 /* The minimum length of a PTP V2 packet for offsets, etc. to be valid: 117 * includes IP header. 118 */ 119 #define PTP_V2_MIN_LENGTH 63 120 121 #define PTP_MIN_LENGTH 63 122 123 #define PTP_ADDRESS 0xe0000181 /* 224.0.1.129 */ 124 #define PTP_EVENT_PORT 319 125 #define PTP_GENERAL_PORT 320 126 127 /* Annoyingly the format of the version numbers are different between 128 * versions 1 and 2 so it isn't possible to simply look for 1 or 2. 129 */ 130 #define PTP_VERSION_V1 1 131 132 #define PTP_VERSION_V2 2 133 #define PTP_VERSION_V2_MASK 0x0f 134 135 enum ptp_packet_state { 136 PTP_PACKET_STATE_UNMATCHED = 0, 137 PTP_PACKET_STATE_MATCHED, 138 PTP_PACKET_STATE_TIMED_OUT, 139 PTP_PACKET_STATE_MATCH_UNWANTED 140 }; 141 142 /* NIC synchronised with single word of time only comprising 143 * partial seconds and full nanoseconds: 10^9 ~ 2^30 so 2 bits for seconds. 144 */ 145 #define MC_NANOSECOND_BITS 30 146 #define MC_NANOSECOND_MASK ((1 << MC_NANOSECOND_BITS) - 1) 147 #define MC_SECOND_MASK ((1 << (32 - MC_NANOSECOND_BITS)) - 1) 148 149 /* Maximum parts-per-billion adjustment that is acceptable */ 150 #define MAX_PPB 1000000 151 152 /* Number of bits required to hold the above */ 153 #define MAX_PPB_BITS 20 154 155 /* Number of extra bits allowed when calculating fractional ns. 156 * EXTRA_BITS + MC_CMD_PTP_IN_ADJUST_BITS + MAX_PPB_BITS should 157 * be less than 63. 158 */ 159 #define PPB_EXTRA_BITS 2 160 161 /* Precalculate scale word to avoid long long division at runtime */ 162 #define PPB_SCALE_WORD ((1LL << (PPB_EXTRA_BITS + MC_CMD_PTP_IN_ADJUST_BITS +\ 163 MAX_PPB_BITS)) / 1000000000LL) 164 165 #define PTP_SYNC_ATTEMPTS 4 166 167 /** 168 * struct efx_ptp_match - Matching structure, stored in sk_buff's cb area. 169 * @words: UUID and (partial) sequence number 170 * @expiry: Time after which the packet should be delivered irrespective of 171 * event arrival. 172 * @state: The state of the packet - whether it is ready for processing or 173 * whether that is of no interest. 174 */ 175 struct efx_ptp_match { 176 u32 words[DIV_ROUND_UP(PTP_V1_UUID_LENGTH, 4)]; 177 unsigned long expiry; 178 enum ptp_packet_state state; 179 }; 180 181 /** 182 * struct efx_ptp_event_rx - A PTP receive event (from MC) 183 * @seq0: First part of (PTP) UUID 184 * @seq1: Second part of (PTP) UUID and sequence number 185 * @hwtimestamp: Event timestamp 186 */ 187 struct efx_ptp_event_rx { 188 struct list_head link; 189 u32 seq0; 190 u32 seq1; 191 ktime_t hwtimestamp; 192 unsigned long expiry; 193 }; 194 195 /** 196 * struct efx_ptp_timeset - Synchronisation between host and MC 197 * @host_start: Host time immediately before hardware timestamp taken 198 * @major: Hardware timestamp, major 199 * @minor: Hardware timestamp, minor 200 * @host_end: Host time immediately after hardware timestamp taken 201 * @wait: Number of NIC clock ticks between hardware timestamp being read and 202 * host end time being seen 203 * @window: Difference of host_end and host_start 204 * @valid: Whether this timeset is valid 205 */ 206 struct efx_ptp_timeset { 207 u32 host_start; 208 u32 major; 209 u32 minor; 210 u32 host_end; 211 u32 wait; 212 u32 window; /* Derived: end - start, allowing for wrap */ 213 }; 214 215 /** 216 * struct efx_ptp_data - Precision Time Protocol (PTP) state 217 * @efx: The NIC context 218 * @channel: The PTP channel (Siena only) 219 * @rx_ts_inline: Flag for whether RX timestamps are inline (else they are 220 * separate events) 221 * @rxq: Receive queue (awaiting timestamps) 222 * @txq: Transmit queue 223 * @evt_list: List of MC receive events awaiting packets 224 * @evt_free_list: List of free events 225 * @evt_lock: Lock for manipulating evt_list and evt_free_list 226 * @evt_overflow: Boolean indicating that event list has overflowed 227 * @rx_evts: Instantiated events (on evt_list and evt_free_list) 228 * @workwq: Work queue for processing pending PTP operations 229 * @work: Work task 230 * @reset_required: A serious error has occurred and the PTP task needs to be 231 * reset (disable, enable). 232 * @rxfilter_event: Receive filter when operating 233 * @rxfilter_general: Receive filter when operating 234 * @config: Current timestamp configuration 235 * @enabled: PTP operation enabled 236 * @mode: Mode in which PTP operating (PTP version) 237 * @time_format: Time format supported by this NIC 238 * @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time 239 * @nic_to_kernel_time: Function to convert from NIC to kernel time 240 * @min_synchronisation_ns: Minimum acceptable corrected sync window 241 * @ts_corrections.tx: Required driver correction of transmit timestamps 242 * @ts_corrections.rx: Required driver correction of receive timestamps 243 * @ts_corrections.pps_out: PPS output error (information only) 244 * @ts_corrections.pps_in: Required driver correction of PPS input timestamps 245 * @evt_frags: Partly assembled PTP events 246 * @evt_frag_idx: Current fragment number 247 * @evt_code: Last event code 248 * @start: Address at which MC indicates ready for synchronisation 249 * @host_time_pps: Host time at last PPS 250 * @current_adjfreq: Current ppb adjustment. 251 * @phc_clock: Pointer to registered phc device (if primary function) 252 * @phc_clock_info: Registration structure for phc device 253 * @pps_work: pps work task for handling pps events 254 * @pps_workwq: pps work queue 255 * @nic_ts_enabled: Flag indicating if NIC generated TS events are handled 256 * @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids 257 * allocations in main data path). 258 * @good_syncs: Number of successful synchronisations. 259 * @fast_syncs: Number of synchronisations requiring short delay 260 * @bad_syncs: Number of failed synchronisations. 261 * @sync_timeouts: Number of synchronisation timeouts 262 * @no_time_syncs: Number of synchronisations with no good times. 263 * @invalid_sync_windows: Number of sync windows with bad durations. 264 * @undersize_sync_windows: Number of corrected sync windows that are too small 265 * @oversize_sync_windows: Number of corrected sync windows that are too large 266 * @rx_no_timestamp: Number of packets received without a timestamp. 267 * @timeset: Last set of synchronisation statistics. 268 */ 269 struct efx_ptp_data { 270 struct efx_nic *efx; 271 struct efx_channel *channel; 272 bool rx_ts_inline; 273 struct sk_buff_head rxq; 274 struct sk_buff_head txq; 275 struct list_head evt_list; 276 struct list_head evt_free_list; 277 spinlock_t evt_lock; 278 bool evt_overflow; 279 struct efx_ptp_event_rx rx_evts[MAX_RECEIVE_EVENTS]; 280 struct workqueue_struct *workwq; 281 struct work_struct work; 282 bool reset_required; 283 u32 rxfilter_event; 284 u32 rxfilter_general; 285 bool rxfilter_installed; 286 struct hwtstamp_config config; 287 bool enabled; 288 unsigned int mode; 289 unsigned int time_format; 290 void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor); 291 ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor, 292 s32 correction); 293 unsigned int min_synchronisation_ns; 294 struct { 295 s32 tx; 296 s32 rx; 297 s32 pps_out; 298 s32 pps_in; 299 } ts_corrections; 300 efx_qword_t evt_frags[MAX_EVENT_FRAGS]; 301 int evt_frag_idx; 302 int evt_code; 303 struct efx_buffer start; 304 struct pps_event_time host_time_pps; 305 s64 current_adjfreq; 306 struct ptp_clock *phc_clock; 307 struct ptp_clock_info phc_clock_info; 308 struct work_struct pps_work; 309 struct workqueue_struct *pps_workwq; 310 bool nic_ts_enabled; 311 MCDI_DECLARE_BUF(txbuf, MC_CMD_PTP_IN_TRANSMIT_LENMAX); 312 313 unsigned int good_syncs; 314 unsigned int fast_syncs; 315 unsigned int bad_syncs; 316 unsigned int sync_timeouts; 317 unsigned int no_time_syncs; 318 unsigned int invalid_sync_windows; 319 unsigned int undersize_sync_windows; 320 unsigned int oversize_sync_windows; 321 unsigned int rx_no_timestamp; 322 struct efx_ptp_timeset 323 timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM]; 324 }; 325 326 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta); 327 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta); 328 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec *ts); 329 static int efx_phc_settime(struct ptp_clock_info *ptp, 330 const struct timespec *e_ts); 331 static int efx_phc_enable(struct ptp_clock_info *ptp, 332 struct ptp_clock_request *request, int on); 333 334 #define PTP_SW_STAT(ext_name, field_name) \ 335 { #ext_name, 0, offsetof(struct efx_ptp_data, field_name) } 336 #define PTP_MC_STAT(ext_name, mcdi_name) \ 337 { #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST } 338 static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = { 339 PTP_SW_STAT(ptp_good_syncs, good_syncs), 340 PTP_SW_STAT(ptp_fast_syncs, fast_syncs), 341 PTP_SW_STAT(ptp_bad_syncs, bad_syncs), 342 PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts), 343 PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs), 344 PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows), 345 PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows), 346 PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows), 347 PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp), 348 PTP_MC_STAT(ptp_tx_timestamp_packets, TX), 349 PTP_MC_STAT(ptp_rx_timestamp_packets, RX), 350 PTP_MC_STAT(ptp_timestamp_packets, TS), 351 PTP_MC_STAT(ptp_filter_matches, FM), 352 PTP_MC_STAT(ptp_non_filter_matches, NFM), 353 }; 354 #define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc) 355 static const unsigned long efx_ptp_stat_mask[] = { 356 [0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL, 357 }; 358 359 size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 *strings) 360 { 361 if (!efx->ptp_data) 362 return 0; 363 364 return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT, 365 efx_ptp_stat_mask, strings); 366 } 367 368 size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats) 369 { 370 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN); 371 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN); 372 size_t i; 373 int rc; 374 375 if (!efx->ptp_data) 376 return 0; 377 378 /* Copy software statistics */ 379 for (i = 0; i < PTP_STAT_COUNT; i++) { 380 if (efx_ptp_stat_desc[i].dma_width) 381 continue; 382 stats[i] = *(unsigned int *)((char *)efx->ptp_data + 383 efx_ptp_stat_desc[i].offset); 384 } 385 386 /* Fetch MC statistics. We *must* fill in all statistics or 387 * risk leaking kernel memory to userland, so if the MCDI 388 * request fails we pretend we got zeroes. 389 */ 390 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS); 391 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 392 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 393 outbuf, sizeof(outbuf), NULL); 394 if (rc) { 395 netif_err(efx, hw, efx->net_dev, 396 "MC_CMD_PTP_OP_STATUS failed (%d)\n", rc); 397 memset(outbuf, 0, sizeof(outbuf)); 398 } 399 efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT, 400 efx_ptp_stat_mask, 401 stats, _MCDI_PTR(outbuf, 0), false); 402 403 return PTP_STAT_COUNT; 404 } 405 406 /* For Siena platforms NIC time is s and ns */ 407 static void efx_ptp_ns_to_s_ns(s64 ns, u32 *nic_major, u32 *nic_minor) 408 { 409 struct timespec ts = ns_to_timespec(ns); 410 *nic_major = ts.tv_sec; 411 *nic_minor = ts.tv_nsec; 412 } 413 414 static ktime_t efx_ptp_s_ns_to_ktime_correction(u32 nic_major, u32 nic_minor, 415 s32 correction) 416 { 417 ktime_t kt = ktime_set(nic_major, nic_minor); 418 if (correction >= 0) 419 kt = ktime_add_ns(kt, (u64)correction); 420 else 421 kt = ktime_sub_ns(kt, (u64)-correction); 422 return kt; 423 } 424 425 /* To convert from s27 format to ns we multiply then divide by a power of 2. 426 * For the conversion from ns to s27, the operation is also converted to a 427 * multiply and shift. 428 */ 429 #define S27_TO_NS_SHIFT (27) 430 #define NS_TO_S27_MULT (((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC) 431 #define NS_TO_S27_SHIFT (63 - S27_TO_NS_SHIFT) 432 #define S27_MINOR_MAX (1 << S27_TO_NS_SHIFT) 433 434 /* For Huntington platforms NIC time is in seconds and fractions of a second 435 * where the minor register only uses 27 bits in units of 2^-27s. 436 */ 437 static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor) 438 { 439 struct timespec ts = ns_to_timespec(ns); 440 u32 maj = ts.tv_sec; 441 u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT + 442 (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT); 443 444 /* The conversion can result in the minor value exceeding the maximum. 445 * In this case, round up to the next second. 446 */ 447 if (min >= S27_MINOR_MAX) { 448 min -= S27_MINOR_MAX; 449 maj++; 450 } 451 452 *nic_major = maj; 453 *nic_minor = min; 454 } 455 456 static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor) 457 { 458 u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC + 459 (1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT); 460 return ktime_set(nic_major, ns); 461 } 462 463 static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor, 464 s32 correction) 465 { 466 /* Apply the correction and deal with carry */ 467 nic_minor += correction; 468 if ((s32)nic_minor < 0) { 469 nic_minor += S27_MINOR_MAX; 470 nic_major--; 471 } else if (nic_minor >= S27_MINOR_MAX) { 472 nic_minor -= S27_MINOR_MAX; 473 nic_major++; 474 } 475 476 return efx_ptp_s27_to_ktime(nic_major, nic_minor); 477 } 478 479 /* Get PTP attributes and set up time conversions */ 480 static int efx_ptp_get_attributes(struct efx_nic *efx) 481 { 482 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN); 483 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN); 484 struct efx_ptp_data *ptp = efx->ptp_data; 485 int rc; 486 u32 fmt; 487 size_t out_len; 488 489 /* Get the PTP attributes. If the NIC doesn't support the operation we 490 * use the default format for compatibility with older NICs i.e. 491 * seconds and nanoseconds. 492 */ 493 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES); 494 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 495 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 496 outbuf, sizeof(outbuf), &out_len); 497 if (rc == 0) 498 fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT); 499 else if (rc == -EINVAL) 500 fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS; 501 else 502 return rc; 503 504 if (fmt == MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION) { 505 ptp->ns_to_nic_time = efx_ptp_ns_to_s27; 506 ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction; 507 } else if (fmt == MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS) { 508 ptp->ns_to_nic_time = efx_ptp_ns_to_s_ns; 509 ptp->nic_to_kernel_time = efx_ptp_s_ns_to_ktime_correction; 510 } else { 511 return -ERANGE; 512 } 513 514 ptp->time_format = fmt; 515 516 /* MC_CMD_PTP_OP_GET_ATTRIBUTES is an extended version of an older 517 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT that also returns a value 518 * to use for the minimum acceptable corrected synchronization window. 519 * If we have the extra information store it. For older firmware that 520 * does not implement the extended command use the default value. 521 */ 522 if (rc == 0 && out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN) 523 ptp->min_synchronisation_ns = 524 MCDI_DWORD(outbuf, 525 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN); 526 else 527 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS; 528 529 return 0; 530 } 531 532 /* Get PTP timestamp corrections */ 533 static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx) 534 { 535 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN); 536 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_LEN); 537 int rc; 538 539 /* Get the timestamp corrections from the NIC. If this operation is 540 * not supported (older NICs) then no correction is required. 541 */ 542 MCDI_SET_DWORD(inbuf, PTP_IN_OP, 543 MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS); 544 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 545 546 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 547 outbuf, sizeof(outbuf), NULL); 548 if (rc == 0) { 549 efx->ptp_data->ts_corrections.tx = MCDI_DWORD(outbuf, 550 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT); 551 efx->ptp_data->ts_corrections.rx = MCDI_DWORD(outbuf, 552 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE); 553 efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf, 554 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT); 555 efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf, 556 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN); 557 } else if (rc == -EINVAL) { 558 efx->ptp_data->ts_corrections.tx = 0; 559 efx->ptp_data->ts_corrections.rx = 0; 560 efx->ptp_data->ts_corrections.pps_out = 0; 561 efx->ptp_data->ts_corrections.pps_in = 0; 562 } else { 563 return rc; 564 } 565 566 return 0; 567 } 568 569 /* Enable MCDI PTP support. */ 570 static int efx_ptp_enable(struct efx_nic *efx) 571 { 572 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN); 573 MCDI_DECLARE_BUF_OUT_OR_ERR(outbuf, 0); 574 int rc; 575 576 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE); 577 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 578 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE, 579 efx->ptp_data->channel ? 580 efx->ptp_data->channel->channel : 0); 581 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode); 582 583 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 584 outbuf, sizeof(outbuf), NULL); 585 rc = (rc == -EALREADY) ? 0 : rc; 586 if (rc) 587 efx_mcdi_display_error(efx, MC_CMD_PTP, 588 MC_CMD_PTP_IN_ENABLE_LEN, 589 outbuf, sizeof(outbuf), rc); 590 return rc; 591 } 592 593 /* Disable MCDI PTP support. 594 * 595 * Note that this function should never rely on the presence of ptp_data - 596 * may be called before that exists. 597 */ 598 static int efx_ptp_disable(struct efx_nic *efx) 599 { 600 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN); 601 MCDI_DECLARE_BUF_OUT_OR_ERR(outbuf, 0); 602 int rc; 603 604 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE); 605 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 606 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 607 outbuf, sizeof(outbuf), NULL); 608 rc = (rc == -EALREADY) ? 0 : rc; 609 if (rc) 610 efx_mcdi_display_error(efx, MC_CMD_PTP, 611 MC_CMD_PTP_IN_DISABLE_LEN, 612 outbuf, sizeof(outbuf), rc); 613 return rc; 614 } 615 616 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q) 617 { 618 struct sk_buff *skb; 619 620 while ((skb = skb_dequeue(q))) { 621 local_bh_disable(); 622 netif_receive_skb(skb); 623 local_bh_enable(); 624 } 625 } 626 627 static void efx_ptp_handle_no_channel(struct efx_nic *efx) 628 { 629 netif_err(efx, drv, efx->net_dev, 630 "ERROR: PTP requires MSI-X and 1 additional interrupt" 631 "vector. PTP disabled\n"); 632 } 633 634 /* Repeatedly send the host time to the MC which will capture the hardware 635 * time. 636 */ 637 static void efx_ptp_send_times(struct efx_nic *efx, 638 struct pps_event_time *last_time) 639 { 640 struct pps_event_time now; 641 struct timespec limit; 642 struct efx_ptp_data *ptp = efx->ptp_data; 643 struct timespec start; 644 int *mc_running = ptp->start.addr; 645 646 pps_get_ts(&now); 647 start = now.ts_real; 648 limit = now.ts_real; 649 timespec_add_ns(&limit, SYNCHRONISE_PERIOD_NS); 650 651 /* Write host time for specified period or until MC is done */ 652 while ((timespec_compare(&now.ts_real, &limit) < 0) && 653 ACCESS_ONCE(*mc_running)) { 654 struct timespec update_time; 655 unsigned int host_time; 656 657 /* Don't update continuously to avoid saturating the PCIe bus */ 658 update_time = now.ts_real; 659 timespec_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS); 660 do { 661 pps_get_ts(&now); 662 } while ((timespec_compare(&now.ts_real, &update_time) < 0) && 663 ACCESS_ONCE(*mc_running)); 664 665 /* Synchronise NIC with single word of time only */ 666 host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS | 667 now.ts_real.tv_nsec); 668 /* Update host time in NIC memory */ 669 efx->type->ptp_write_host_time(efx, host_time); 670 } 671 *last_time = now; 672 } 673 674 /* Read a timeset from the MC's results and partial process. */ 675 static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data), 676 struct efx_ptp_timeset *timeset) 677 { 678 unsigned start_ns, end_ns; 679 680 timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART); 681 timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR); 682 timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR); 683 timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND), 684 timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS); 685 686 /* Ignore seconds */ 687 start_ns = timeset->host_start & MC_NANOSECOND_MASK; 688 end_ns = timeset->host_end & MC_NANOSECOND_MASK; 689 /* Allow for rollover */ 690 if (end_ns < start_ns) 691 end_ns += NSEC_PER_SEC; 692 /* Determine duration of operation */ 693 timeset->window = end_ns - start_ns; 694 } 695 696 /* Process times received from MC. 697 * 698 * Extract times from returned results, and establish the minimum value 699 * seen. The minimum value represents the "best" possible time and events 700 * too much greater than this are rejected - the machine is, perhaps, too 701 * busy. A number of readings are taken so that, hopefully, at least one good 702 * synchronisation will be seen in the results. 703 */ 704 static int 705 efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf), 706 size_t response_length, 707 const struct pps_event_time *last_time) 708 { 709 unsigned number_readings = 710 MCDI_VAR_ARRAY_LEN(response_length, 711 PTP_OUT_SYNCHRONIZE_TIMESET); 712 unsigned i; 713 unsigned ngood = 0; 714 unsigned last_good = 0; 715 struct efx_ptp_data *ptp = efx->ptp_data; 716 u32 last_sec; 717 u32 start_sec; 718 struct timespec delta; 719 ktime_t mc_time; 720 721 if (number_readings == 0) 722 return -EAGAIN; 723 724 /* Read the set of results and find the last good host-MC 725 * synchronization result. The MC times when it finishes reading the 726 * host time so the corrected window time should be fairly constant 727 * for a given platform. Increment stats for any results that appear 728 * to be erroneous. 729 */ 730 for (i = 0; i < number_readings; i++) { 731 s32 window, corrected; 732 struct timespec wait; 733 734 efx_ptp_read_timeset( 735 MCDI_ARRAY_STRUCT_PTR(synch_buf, 736 PTP_OUT_SYNCHRONIZE_TIMESET, i), 737 &ptp->timeset[i]); 738 739 wait = ktime_to_timespec( 740 ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0)); 741 window = ptp->timeset[i].window; 742 corrected = window - wait.tv_nsec; 743 744 /* We expect the uncorrected synchronization window to be at 745 * least as large as the interval between host start and end 746 * times. If it is smaller than this then this is mostly likely 747 * to be a consequence of the host's time being adjusted. 748 * Check that the corrected sync window is in a reasonable 749 * range. If it is out of range it is likely to be because an 750 * interrupt or other delay occurred between reading the system 751 * time and writing it to MC memory. 752 */ 753 if (window < SYNCHRONISATION_GRANULARITY_NS) { 754 ++ptp->invalid_sync_windows; 755 } else if (corrected >= MAX_SYNCHRONISATION_NS) { 756 ++ptp->oversize_sync_windows; 757 } else if (corrected < ptp->min_synchronisation_ns) { 758 ++ptp->undersize_sync_windows; 759 } else { 760 ngood++; 761 last_good = i; 762 } 763 } 764 765 if (ngood == 0) { 766 netif_warn(efx, drv, efx->net_dev, 767 "PTP no suitable synchronisations\n"); 768 return -EAGAIN; 769 } 770 771 /* Convert the NIC time into kernel time. No correction is required- 772 * this time is the output of a firmware process. 773 */ 774 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major, 775 ptp->timeset[last_good].minor, 0); 776 777 /* Calculate delay from actual PPS to last_time */ 778 delta = ktime_to_timespec(mc_time); 779 delta.tv_nsec += 780 last_time->ts_real.tv_nsec - 781 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK); 782 783 /* It is possible that the seconds rolled over between taking 784 * the start reading and the last value written by the host. The 785 * timescales are such that a gap of more than one second is never 786 * expected. 787 */ 788 start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS; 789 last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK; 790 if (start_sec != last_sec) { 791 if (((start_sec + 1) & MC_SECOND_MASK) != last_sec) { 792 netif_warn(efx, hw, efx->net_dev, 793 "PTP bad synchronisation seconds\n"); 794 return -EAGAIN; 795 } else { 796 delta.tv_sec = 1; 797 } 798 } else { 799 delta.tv_sec = 0; 800 } 801 802 ptp->host_time_pps = *last_time; 803 pps_sub_ts(&ptp->host_time_pps, delta); 804 805 return 0; 806 } 807 808 /* Synchronize times between the host and the MC */ 809 static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings) 810 { 811 struct efx_ptp_data *ptp = efx->ptp_data; 812 MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX); 813 size_t response_length; 814 int rc; 815 unsigned long timeout; 816 struct pps_event_time last_time = {}; 817 unsigned int loops = 0; 818 int *start = ptp->start.addr; 819 820 MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE); 821 MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0); 822 MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS, 823 num_readings); 824 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR, 825 ptp->start.dma_addr); 826 827 /* Clear flag that signals MC ready */ 828 ACCESS_ONCE(*start) = 0; 829 rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf, 830 MC_CMD_PTP_IN_SYNCHRONIZE_LEN); 831 EFX_BUG_ON_PARANOID(rc); 832 833 /* Wait for start from MCDI (or timeout) */ 834 timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS); 835 while (!ACCESS_ONCE(*start) && (time_before(jiffies, timeout))) { 836 udelay(20); /* Usually start MCDI execution quickly */ 837 loops++; 838 } 839 840 if (loops <= 1) 841 ++ptp->fast_syncs; 842 if (!time_before(jiffies, timeout)) 843 ++ptp->sync_timeouts; 844 845 if (ACCESS_ONCE(*start)) 846 efx_ptp_send_times(efx, &last_time); 847 848 /* Collect results */ 849 rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP, 850 MC_CMD_PTP_IN_SYNCHRONIZE_LEN, 851 synch_buf, sizeof(synch_buf), 852 &response_length); 853 if (rc == 0) { 854 rc = efx_ptp_process_times(efx, synch_buf, response_length, 855 &last_time); 856 if (rc == 0) 857 ++ptp->good_syncs; 858 else 859 ++ptp->no_time_syncs; 860 } 861 862 /* Increment the bad syncs counter if the synchronize fails, whatever 863 * the reason. 864 */ 865 if (rc != 0) 866 ++ptp->bad_syncs; 867 868 return rc; 869 } 870 871 /* Transmit a PTP packet, via the MCDI interface, to the wire. */ 872 static int efx_ptp_xmit_skb(struct efx_nic *efx, struct sk_buff *skb) 873 { 874 struct efx_ptp_data *ptp_data = efx->ptp_data; 875 struct skb_shared_hwtstamps timestamps; 876 int rc = -EIO; 877 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN); 878 size_t len; 879 880 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT); 881 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0); 882 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len); 883 if (skb_shinfo(skb)->nr_frags != 0) { 884 rc = skb_linearize(skb); 885 if (rc != 0) 886 goto fail; 887 } 888 889 if (skb->ip_summed == CHECKSUM_PARTIAL) { 890 rc = skb_checksum_help(skb); 891 if (rc != 0) 892 goto fail; 893 } 894 skb_copy_from_linear_data(skb, 895 MCDI_PTR(ptp_data->txbuf, 896 PTP_IN_TRANSMIT_PACKET), 897 skb->len); 898 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, 899 ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len), 900 txtime, sizeof(txtime), &len); 901 if (rc != 0) 902 goto fail; 903 904 memset(×tamps, 0, sizeof(timestamps)); 905 timestamps.hwtstamp = ptp_data->nic_to_kernel_time( 906 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR), 907 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR), 908 ptp_data->ts_corrections.tx); 909 910 skb_tstamp_tx(skb, ×tamps); 911 912 rc = 0; 913 914 fail: 915 dev_kfree_skb(skb); 916 917 return rc; 918 } 919 920 static void efx_ptp_drop_time_expired_events(struct efx_nic *efx) 921 { 922 struct efx_ptp_data *ptp = efx->ptp_data; 923 struct list_head *cursor; 924 struct list_head *next; 925 926 if (ptp->rx_ts_inline) 927 return; 928 929 /* Drop time-expired events */ 930 spin_lock_bh(&ptp->evt_lock); 931 if (!list_empty(&ptp->evt_list)) { 932 list_for_each_safe(cursor, next, &ptp->evt_list) { 933 struct efx_ptp_event_rx *evt; 934 935 evt = list_entry(cursor, struct efx_ptp_event_rx, 936 link); 937 if (time_after(jiffies, evt->expiry)) { 938 list_move(&evt->link, &ptp->evt_free_list); 939 netif_warn(efx, hw, efx->net_dev, 940 "PTP rx event dropped\n"); 941 } 942 } 943 } 944 /* If the event overflow flag is set and the event list is now empty 945 * clear the flag to re-enable the overflow warning message. 946 */ 947 if (ptp->evt_overflow && list_empty(&ptp->evt_list)) 948 ptp->evt_overflow = false; 949 spin_unlock_bh(&ptp->evt_lock); 950 } 951 952 static enum ptp_packet_state efx_ptp_match_rx(struct efx_nic *efx, 953 struct sk_buff *skb) 954 { 955 struct efx_ptp_data *ptp = efx->ptp_data; 956 bool evts_waiting; 957 struct list_head *cursor; 958 struct list_head *next; 959 struct efx_ptp_match *match; 960 enum ptp_packet_state rc = PTP_PACKET_STATE_UNMATCHED; 961 962 WARN_ON_ONCE(ptp->rx_ts_inline); 963 964 spin_lock_bh(&ptp->evt_lock); 965 evts_waiting = !list_empty(&ptp->evt_list); 966 spin_unlock_bh(&ptp->evt_lock); 967 968 if (!evts_waiting) 969 return PTP_PACKET_STATE_UNMATCHED; 970 971 match = (struct efx_ptp_match *)skb->cb; 972 /* Look for a matching timestamp in the event queue */ 973 spin_lock_bh(&ptp->evt_lock); 974 list_for_each_safe(cursor, next, &ptp->evt_list) { 975 struct efx_ptp_event_rx *evt; 976 977 evt = list_entry(cursor, struct efx_ptp_event_rx, link); 978 if ((evt->seq0 == match->words[0]) && 979 (evt->seq1 == match->words[1])) { 980 struct skb_shared_hwtstamps *timestamps; 981 982 /* Match - add in hardware timestamp */ 983 timestamps = skb_hwtstamps(skb); 984 timestamps->hwtstamp = evt->hwtimestamp; 985 986 match->state = PTP_PACKET_STATE_MATCHED; 987 rc = PTP_PACKET_STATE_MATCHED; 988 list_move(&evt->link, &ptp->evt_free_list); 989 break; 990 } 991 } 992 /* If the event overflow flag is set and the event list is now empty 993 * clear the flag to re-enable the overflow warning message. 994 */ 995 if (ptp->evt_overflow && list_empty(&ptp->evt_list)) 996 ptp->evt_overflow = false; 997 spin_unlock_bh(&ptp->evt_lock); 998 999 return rc; 1000 } 1001 1002 /* Process any queued receive events and corresponding packets 1003 * 1004 * q is returned with all the packets that are ready for delivery. 1005 */ 1006 static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q) 1007 { 1008 struct efx_ptp_data *ptp = efx->ptp_data; 1009 struct sk_buff *skb; 1010 1011 while ((skb = skb_dequeue(&ptp->rxq))) { 1012 struct efx_ptp_match *match; 1013 1014 match = (struct efx_ptp_match *)skb->cb; 1015 if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) { 1016 __skb_queue_tail(q, skb); 1017 } else if (efx_ptp_match_rx(efx, skb) == 1018 PTP_PACKET_STATE_MATCHED) { 1019 __skb_queue_tail(q, skb); 1020 } else if (time_after(jiffies, match->expiry)) { 1021 match->state = PTP_PACKET_STATE_TIMED_OUT; 1022 ++ptp->rx_no_timestamp; 1023 __skb_queue_tail(q, skb); 1024 } else { 1025 /* Replace unprocessed entry and stop */ 1026 skb_queue_head(&ptp->rxq, skb); 1027 break; 1028 } 1029 } 1030 } 1031 1032 /* Complete processing of a received packet */ 1033 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb) 1034 { 1035 local_bh_disable(); 1036 netif_receive_skb(skb); 1037 local_bh_enable(); 1038 } 1039 1040 static void efx_ptp_remove_multicast_filters(struct efx_nic *efx) 1041 { 1042 struct efx_ptp_data *ptp = efx->ptp_data; 1043 1044 if (ptp->rxfilter_installed) { 1045 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED, 1046 ptp->rxfilter_general); 1047 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED, 1048 ptp->rxfilter_event); 1049 ptp->rxfilter_installed = false; 1050 } 1051 } 1052 1053 static int efx_ptp_insert_multicast_filters(struct efx_nic *efx) 1054 { 1055 struct efx_ptp_data *ptp = efx->ptp_data; 1056 struct efx_filter_spec rxfilter; 1057 int rc; 1058 1059 if (!ptp->channel || ptp->rxfilter_installed) 1060 return 0; 1061 1062 /* Must filter on both event and general ports to ensure 1063 * that there is no packet re-ordering. 1064 */ 1065 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0, 1066 efx_rx_queue_index( 1067 efx_channel_get_rx_queue(ptp->channel))); 1068 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP, 1069 htonl(PTP_ADDRESS), 1070 htons(PTP_EVENT_PORT)); 1071 if (rc != 0) 1072 return rc; 1073 1074 rc = efx_filter_insert_filter(efx, &rxfilter, true); 1075 if (rc < 0) 1076 return rc; 1077 ptp->rxfilter_event = rc; 1078 1079 efx_filter_init_rx(&rxfilter, EFX_FILTER_PRI_REQUIRED, 0, 1080 efx_rx_queue_index( 1081 efx_channel_get_rx_queue(ptp->channel))); 1082 rc = efx_filter_set_ipv4_local(&rxfilter, IPPROTO_UDP, 1083 htonl(PTP_ADDRESS), 1084 htons(PTP_GENERAL_PORT)); 1085 if (rc != 0) 1086 goto fail; 1087 1088 rc = efx_filter_insert_filter(efx, &rxfilter, true); 1089 if (rc < 0) 1090 goto fail; 1091 ptp->rxfilter_general = rc; 1092 1093 ptp->rxfilter_installed = true; 1094 return 0; 1095 1096 fail: 1097 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED, 1098 ptp->rxfilter_event); 1099 return rc; 1100 } 1101 1102 static int efx_ptp_start(struct efx_nic *efx) 1103 { 1104 struct efx_ptp_data *ptp = efx->ptp_data; 1105 int rc; 1106 1107 ptp->reset_required = false; 1108 1109 rc = efx_ptp_insert_multicast_filters(efx); 1110 if (rc) 1111 return rc; 1112 1113 rc = efx_ptp_enable(efx); 1114 if (rc != 0) 1115 goto fail; 1116 1117 ptp->evt_frag_idx = 0; 1118 ptp->current_adjfreq = 0; 1119 1120 return 0; 1121 1122 fail: 1123 efx_ptp_remove_multicast_filters(efx); 1124 return rc; 1125 } 1126 1127 static int efx_ptp_stop(struct efx_nic *efx) 1128 { 1129 struct efx_ptp_data *ptp = efx->ptp_data; 1130 struct list_head *cursor; 1131 struct list_head *next; 1132 int rc; 1133 1134 if (ptp == NULL) 1135 return 0; 1136 1137 rc = efx_ptp_disable(efx); 1138 1139 efx_ptp_remove_multicast_filters(efx); 1140 1141 /* Make sure RX packets are really delivered */ 1142 efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq); 1143 skb_queue_purge(&efx->ptp_data->txq); 1144 1145 /* Drop any pending receive events */ 1146 spin_lock_bh(&efx->ptp_data->evt_lock); 1147 list_for_each_safe(cursor, next, &efx->ptp_data->evt_list) { 1148 list_move(cursor, &efx->ptp_data->evt_free_list); 1149 } 1150 ptp->evt_overflow = false; 1151 spin_unlock_bh(&efx->ptp_data->evt_lock); 1152 1153 return rc; 1154 } 1155 1156 static int efx_ptp_restart(struct efx_nic *efx) 1157 { 1158 if (efx->ptp_data && efx->ptp_data->enabled) 1159 return efx_ptp_start(efx); 1160 return 0; 1161 } 1162 1163 static void efx_ptp_pps_worker(struct work_struct *work) 1164 { 1165 struct efx_ptp_data *ptp = 1166 container_of(work, struct efx_ptp_data, pps_work); 1167 struct efx_nic *efx = ptp->efx; 1168 struct ptp_clock_event ptp_evt; 1169 1170 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS)) 1171 return; 1172 1173 ptp_evt.type = PTP_CLOCK_PPSUSR; 1174 ptp_evt.pps_times = ptp->host_time_pps; 1175 ptp_clock_event(ptp->phc_clock, &ptp_evt); 1176 } 1177 1178 static void efx_ptp_worker(struct work_struct *work) 1179 { 1180 struct efx_ptp_data *ptp_data = 1181 container_of(work, struct efx_ptp_data, work); 1182 struct efx_nic *efx = ptp_data->efx; 1183 struct sk_buff *skb; 1184 struct sk_buff_head tempq; 1185 1186 if (ptp_data->reset_required) { 1187 efx_ptp_stop(efx); 1188 efx_ptp_start(efx); 1189 return; 1190 } 1191 1192 efx_ptp_drop_time_expired_events(efx); 1193 1194 __skb_queue_head_init(&tempq); 1195 efx_ptp_process_events(efx, &tempq); 1196 1197 while ((skb = skb_dequeue(&ptp_data->txq))) 1198 efx_ptp_xmit_skb(efx, skb); 1199 1200 while ((skb = __skb_dequeue(&tempq))) 1201 efx_ptp_process_rx(efx, skb); 1202 } 1203 1204 static const struct ptp_clock_info efx_phc_clock_info = { 1205 .owner = THIS_MODULE, 1206 .name = "sfc", 1207 .max_adj = MAX_PPB, 1208 .n_alarm = 0, 1209 .n_ext_ts = 0, 1210 .n_per_out = 0, 1211 .pps = 1, 1212 .adjfreq = efx_phc_adjfreq, 1213 .adjtime = efx_phc_adjtime, 1214 .gettime = efx_phc_gettime, 1215 .settime = efx_phc_settime, 1216 .enable = efx_phc_enable, 1217 }; 1218 1219 /* Initialise PTP state. */ 1220 int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel) 1221 { 1222 struct efx_ptp_data *ptp; 1223 int rc = 0; 1224 unsigned int pos; 1225 1226 ptp = kzalloc(sizeof(struct efx_ptp_data), GFP_KERNEL); 1227 efx->ptp_data = ptp; 1228 if (!efx->ptp_data) 1229 return -ENOMEM; 1230 1231 ptp->efx = efx; 1232 ptp->channel = channel; 1233 ptp->rx_ts_inline = efx_nic_rev(efx) >= EFX_REV_HUNT_A0; 1234 1235 rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL); 1236 if (rc != 0) 1237 goto fail1; 1238 1239 skb_queue_head_init(&ptp->rxq); 1240 skb_queue_head_init(&ptp->txq); 1241 ptp->workwq = create_singlethread_workqueue("sfc_ptp"); 1242 if (!ptp->workwq) { 1243 rc = -ENOMEM; 1244 goto fail2; 1245 } 1246 1247 INIT_WORK(&ptp->work, efx_ptp_worker); 1248 ptp->config.flags = 0; 1249 ptp->config.tx_type = HWTSTAMP_TX_OFF; 1250 ptp->config.rx_filter = HWTSTAMP_FILTER_NONE; 1251 INIT_LIST_HEAD(&ptp->evt_list); 1252 INIT_LIST_HEAD(&ptp->evt_free_list); 1253 spin_lock_init(&ptp->evt_lock); 1254 for (pos = 0; pos < MAX_RECEIVE_EVENTS; pos++) 1255 list_add(&ptp->rx_evts[pos].link, &ptp->evt_free_list); 1256 ptp->evt_overflow = false; 1257 1258 /* Get the NIC PTP attributes and set up time conversions */ 1259 rc = efx_ptp_get_attributes(efx); 1260 if (rc < 0) 1261 goto fail3; 1262 1263 /* Get the timestamp corrections */ 1264 rc = efx_ptp_get_timestamp_corrections(efx); 1265 if (rc < 0) 1266 goto fail3; 1267 1268 if (efx->mcdi->fn_flags & 1269 (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) { 1270 ptp->phc_clock_info = efx_phc_clock_info; 1271 ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info, 1272 &efx->pci_dev->dev); 1273 if (IS_ERR(ptp->phc_clock)) { 1274 rc = PTR_ERR(ptp->phc_clock); 1275 goto fail3; 1276 } 1277 1278 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker); 1279 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps"); 1280 if (!ptp->pps_workwq) { 1281 rc = -ENOMEM; 1282 goto fail4; 1283 } 1284 } 1285 ptp->nic_ts_enabled = false; 1286 1287 return 0; 1288 fail4: 1289 ptp_clock_unregister(efx->ptp_data->phc_clock); 1290 1291 fail3: 1292 destroy_workqueue(efx->ptp_data->workwq); 1293 1294 fail2: 1295 efx_nic_free_buffer(efx, &ptp->start); 1296 1297 fail1: 1298 kfree(efx->ptp_data); 1299 efx->ptp_data = NULL; 1300 1301 return rc; 1302 } 1303 1304 /* Initialise PTP channel. 1305 * 1306 * Setting core_index to zero causes the queue to be initialised and doesn't 1307 * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue. 1308 */ 1309 static int efx_ptp_probe_channel(struct efx_channel *channel) 1310 { 1311 struct efx_nic *efx = channel->efx; 1312 1313 channel->irq_moderation = 0; 1314 channel->rx_queue.core_index = 0; 1315 1316 return efx_ptp_probe(efx, channel); 1317 } 1318 1319 void efx_ptp_remove(struct efx_nic *efx) 1320 { 1321 if (!efx->ptp_data) 1322 return; 1323 1324 (void)efx_ptp_disable(efx); 1325 1326 cancel_work_sync(&efx->ptp_data->work); 1327 cancel_work_sync(&efx->ptp_data->pps_work); 1328 1329 skb_queue_purge(&efx->ptp_data->rxq); 1330 skb_queue_purge(&efx->ptp_data->txq); 1331 1332 if (efx->ptp_data->phc_clock) { 1333 destroy_workqueue(efx->ptp_data->pps_workwq); 1334 ptp_clock_unregister(efx->ptp_data->phc_clock); 1335 } 1336 1337 destroy_workqueue(efx->ptp_data->workwq); 1338 1339 efx_nic_free_buffer(efx, &efx->ptp_data->start); 1340 kfree(efx->ptp_data); 1341 } 1342 1343 static void efx_ptp_remove_channel(struct efx_channel *channel) 1344 { 1345 efx_ptp_remove(channel->efx); 1346 } 1347 1348 static void efx_ptp_get_channel_name(struct efx_channel *channel, 1349 char *buf, size_t len) 1350 { 1351 snprintf(buf, len, "%s-ptp", channel->efx->name); 1352 } 1353 1354 /* Determine whether this packet should be processed by the PTP module 1355 * or transmitted conventionally. 1356 */ 1357 bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb) 1358 { 1359 return efx->ptp_data && 1360 efx->ptp_data->enabled && 1361 skb->len >= PTP_MIN_LENGTH && 1362 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM && 1363 likely(skb->protocol == htons(ETH_P_IP)) && 1364 skb_transport_header_was_set(skb) && 1365 skb_network_header_len(skb) >= sizeof(struct iphdr) && 1366 ip_hdr(skb)->protocol == IPPROTO_UDP && 1367 skb_headlen(skb) >= 1368 skb_transport_offset(skb) + sizeof(struct udphdr) && 1369 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT); 1370 } 1371 1372 /* Receive a PTP packet. Packets are queued until the arrival of 1373 * the receive timestamp from the MC - this will probably occur after the 1374 * packet arrival because of the processing in the MC. 1375 */ 1376 static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb) 1377 { 1378 struct efx_nic *efx = channel->efx; 1379 struct efx_ptp_data *ptp = efx->ptp_data; 1380 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb; 1381 u8 *match_data_012, *match_data_345; 1382 unsigned int version; 1383 1384 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS); 1385 1386 /* Correct version? */ 1387 if (ptp->mode == MC_CMD_PTP_MODE_V1) { 1388 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) { 1389 return false; 1390 } 1391 version = ntohs(*(__be16 *)&skb->data[PTP_V1_VERSION_OFFSET]); 1392 if (version != PTP_VERSION_V1) { 1393 return false; 1394 } 1395 1396 /* PTP V1 uses all six bytes of the UUID to match the packet 1397 * to the timestamp 1398 */ 1399 match_data_012 = skb->data + PTP_V1_UUID_OFFSET; 1400 match_data_345 = skb->data + PTP_V1_UUID_OFFSET + 3; 1401 } else { 1402 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) { 1403 return false; 1404 } 1405 version = skb->data[PTP_V2_VERSION_OFFSET]; 1406 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) { 1407 return false; 1408 } 1409 1410 /* The original V2 implementation uses bytes 2-7 of 1411 * the UUID to match the packet to the timestamp. This 1412 * discards two of the bytes of the MAC address used 1413 * to create the UUID (SF bug 33070). The PTP V2 1414 * enhanced mode fixes this issue and uses bytes 0-2 1415 * and byte 5-7 of the UUID. 1416 */ 1417 match_data_345 = skb->data + PTP_V2_UUID_OFFSET + 5; 1418 if (ptp->mode == MC_CMD_PTP_MODE_V2) { 1419 match_data_012 = skb->data + PTP_V2_UUID_OFFSET + 2; 1420 } else { 1421 match_data_012 = skb->data + PTP_V2_UUID_OFFSET + 0; 1422 BUG_ON(ptp->mode != MC_CMD_PTP_MODE_V2_ENHANCED); 1423 } 1424 } 1425 1426 /* Does this packet require timestamping? */ 1427 if (ntohs(*(__be16 *)&skb->data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) { 1428 match->state = PTP_PACKET_STATE_UNMATCHED; 1429 1430 /* We expect the sequence number to be in the same position in 1431 * the packet for PTP V1 and V2 1432 */ 1433 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET); 1434 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH); 1435 1436 /* Extract UUID/Sequence information */ 1437 match->words[0] = (match_data_012[0] | 1438 (match_data_012[1] << 8) | 1439 (match_data_012[2] << 16) | 1440 (match_data_345[0] << 24)); 1441 match->words[1] = (match_data_345[1] | 1442 (match_data_345[2] << 8) | 1443 (skb->data[PTP_V1_SEQUENCE_OFFSET + 1444 PTP_V1_SEQUENCE_LENGTH - 1] << 1445 16)); 1446 } else { 1447 match->state = PTP_PACKET_STATE_MATCH_UNWANTED; 1448 } 1449 1450 skb_queue_tail(&ptp->rxq, skb); 1451 queue_work(ptp->workwq, &ptp->work); 1452 1453 return true; 1454 } 1455 1456 /* Transmit a PTP packet. This has to be transmitted by the MC 1457 * itself, through an MCDI call. MCDI calls aren't permitted 1458 * in the transmit path so defer the actual transmission to a suitable worker. 1459 */ 1460 int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb) 1461 { 1462 struct efx_ptp_data *ptp = efx->ptp_data; 1463 1464 skb_queue_tail(&ptp->txq, skb); 1465 1466 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) && 1467 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM)) 1468 efx_xmit_hwtstamp_pending(skb); 1469 queue_work(ptp->workwq, &ptp->work); 1470 1471 return NETDEV_TX_OK; 1472 } 1473 1474 int efx_ptp_get_mode(struct efx_nic *efx) 1475 { 1476 return efx->ptp_data->mode; 1477 } 1478 1479 int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted, 1480 unsigned int new_mode) 1481 { 1482 if ((enable_wanted != efx->ptp_data->enabled) || 1483 (enable_wanted && (efx->ptp_data->mode != new_mode))) { 1484 int rc = 0; 1485 1486 if (enable_wanted) { 1487 /* Change of mode requires disable */ 1488 if (efx->ptp_data->enabled && 1489 (efx->ptp_data->mode != new_mode)) { 1490 efx->ptp_data->enabled = false; 1491 rc = efx_ptp_stop(efx); 1492 if (rc != 0) 1493 return rc; 1494 } 1495 1496 /* Set new operating mode and establish 1497 * baseline synchronisation, which must 1498 * succeed. 1499 */ 1500 efx->ptp_data->mode = new_mode; 1501 if (netif_running(efx->net_dev)) 1502 rc = efx_ptp_start(efx); 1503 if (rc == 0) { 1504 rc = efx_ptp_synchronize(efx, 1505 PTP_SYNC_ATTEMPTS * 2); 1506 if (rc != 0) 1507 efx_ptp_stop(efx); 1508 } 1509 } else { 1510 rc = efx_ptp_stop(efx); 1511 } 1512 1513 if (rc != 0) 1514 return rc; 1515 1516 efx->ptp_data->enabled = enable_wanted; 1517 } 1518 1519 return 0; 1520 } 1521 1522 static int efx_ptp_ts_init(struct efx_nic *efx, struct hwtstamp_config *init) 1523 { 1524 int rc; 1525 1526 if (init->flags) 1527 return -EINVAL; 1528 1529 if ((init->tx_type != HWTSTAMP_TX_OFF) && 1530 (init->tx_type != HWTSTAMP_TX_ON)) 1531 return -ERANGE; 1532 1533 rc = efx->type->ptp_set_ts_config(efx, init); 1534 if (rc) 1535 return rc; 1536 1537 efx->ptp_data->config = *init; 1538 return 0; 1539 } 1540 1541 void efx_ptp_get_ts_info(struct efx_nic *efx, struct ethtool_ts_info *ts_info) 1542 { 1543 struct efx_ptp_data *ptp = efx->ptp_data; 1544 struct efx_nic *primary = efx->primary; 1545 1546 ASSERT_RTNL(); 1547 1548 if (!ptp) 1549 return; 1550 1551 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE | 1552 SOF_TIMESTAMPING_RX_HARDWARE | 1553 SOF_TIMESTAMPING_RAW_HARDWARE); 1554 if (primary && primary->ptp_data && primary->ptp_data->phc_clock) 1555 ts_info->phc_index = 1556 ptp_clock_index(primary->ptp_data->phc_clock); 1557 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON; 1558 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters; 1559 } 1560 1561 int efx_ptp_set_ts_config(struct efx_nic *efx, struct ifreq *ifr) 1562 { 1563 struct hwtstamp_config config; 1564 int rc; 1565 1566 /* Not a PTP enabled port */ 1567 if (!efx->ptp_data) 1568 return -EOPNOTSUPP; 1569 1570 if (copy_from_user(&config, ifr->ifr_data, sizeof(config))) 1571 return -EFAULT; 1572 1573 rc = efx_ptp_ts_init(efx, &config); 1574 if (rc != 0) 1575 return rc; 1576 1577 return copy_to_user(ifr->ifr_data, &config, sizeof(config)) 1578 ? -EFAULT : 0; 1579 } 1580 1581 int efx_ptp_get_ts_config(struct efx_nic *efx, struct ifreq *ifr) 1582 { 1583 if (!efx->ptp_data) 1584 return -EOPNOTSUPP; 1585 1586 return copy_to_user(ifr->ifr_data, &efx->ptp_data->config, 1587 sizeof(efx->ptp_data->config)) ? -EFAULT : 0; 1588 } 1589 1590 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len) 1591 { 1592 struct efx_ptp_data *ptp = efx->ptp_data; 1593 1594 netif_err(efx, hw, efx->net_dev, 1595 "PTP unexpected event length: got %d expected %d\n", 1596 ptp->evt_frag_idx, expected_frag_len); 1597 ptp->reset_required = true; 1598 queue_work(ptp->workwq, &ptp->work); 1599 } 1600 1601 /* Process a completed receive event. Put it on the event queue and 1602 * start worker thread. This is required because event and their 1603 * correspoding packets may come in either order. 1604 */ 1605 static void ptp_event_rx(struct efx_nic *efx, struct efx_ptp_data *ptp) 1606 { 1607 struct efx_ptp_event_rx *evt = NULL; 1608 1609 if (WARN_ON_ONCE(ptp->rx_ts_inline)) 1610 return; 1611 1612 if (ptp->evt_frag_idx != 3) { 1613 ptp_event_failure(efx, 3); 1614 return; 1615 } 1616 1617 spin_lock_bh(&ptp->evt_lock); 1618 if (!list_empty(&ptp->evt_free_list)) { 1619 evt = list_first_entry(&ptp->evt_free_list, 1620 struct efx_ptp_event_rx, link); 1621 list_del(&evt->link); 1622 1623 evt->seq0 = EFX_QWORD_FIELD(ptp->evt_frags[2], MCDI_EVENT_DATA); 1624 evt->seq1 = (EFX_QWORD_FIELD(ptp->evt_frags[2], 1625 MCDI_EVENT_SRC) | 1626 (EFX_QWORD_FIELD(ptp->evt_frags[1], 1627 MCDI_EVENT_SRC) << 8) | 1628 (EFX_QWORD_FIELD(ptp->evt_frags[0], 1629 MCDI_EVENT_SRC) << 16)); 1630 evt->hwtimestamp = efx->ptp_data->nic_to_kernel_time( 1631 EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA), 1632 EFX_QWORD_FIELD(ptp->evt_frags[1], MCDI_EVENT_DATA), 1633 ptp->ts_corrections.rx); 1634 evt->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS); 1635 list_add_tail(&evt->link, &ptp->evt_list); 1636 1637 queue_work(ptp->workwq, &ptp->work); 1638 } else if (!ptp->evt_overflow) { 1639 /* Log a warning message and set the event overflow flag. 1640 * The message won't be logged again until the event queue 1641 * becomes empty. 1642 */ 1643 netif_err(efx, rx_err, efx->net_dev, "PTP event queue overflow\n"); 1644 ptp->evt_overflow = true; 1645 } 1646 spin_unlock_bh(&ptp->evt_lock); 1647 } 1648 1649 static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp) 1650 { 1651 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA); 1652 if (ptp->evt_frag_idx != 1) { 1653 ptp_event_failure(efx, 1); 1654 return; 1655 } 1656 1657 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code); 1658 } 1659 1660 static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp) 1661 { 1662 if (ptp->nic_ts_enabled) 1663 queue_work(ptp->pps_workwq, &ptp->pps_work); 1664 } 1665 1666 void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev) 1667 { 1668 struct efx_ptp_data *ptp = efx->ptp_data; 1669 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE); 1670 1671 if (!ptp->enabled) 1672 return; 1673 1674 if (ptp->evt_frag_idx == 0) { 1675 ptp->evt_code = code; 1676 } else if (ptp->evt_code != code) { 1677 netif_err(efx, hw, efx->net_dev, 1678 "PTP out of sequence event %d\n", code); 1679 ptp->evt_frag_idx = 0; 1680 } 1681 1682 ptp->evt_frags[ptp->evt_frag_idx++] = *ev; 1683 if (!MCDI_EVENT_FIELD(*ev, CONT)) { 1684 /* Process resulting event */ 1685 switch (code) { 1686 case MCDI_EVENT_CODE_PTP_RX: 1687 ptp_event_rx(efx, ptp); 1688 break; 1689 case MCDI_EVENT_CODE_PTP_FAULT: 1690 ptp_event_fault(efx, ptp); 1691 break; 1692 case MCDI_EVENT_CODE_PTP_PPS: 1693 ptp_event_pps(efx, ptp); 1694 break; 1695 default: 1696 netif_err(efx, hw, efx->net_dev, 1697 "PTP unknown event %d\n", code); 1698 break; 1699 } 1700 ptp->evt_frag_idx = 0; 1701 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) { 1702 netif_err(efx, hw, efx->net_dev, 1703 "PTP too many event fragments\n"); 1704 ptp->evt_frag_idx = 0; 1705 } 1706 } 1707 1708 void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev) 1709 { 1710 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR); 1711 channel->sync_timestamp_minor = 1712 MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_26_19) << 19; 1713 /* if sync events have been disabled then we want to silently ignore 1714 * this event, so throw away result. 1715 */ 1716 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED, 1717 SYNC_EVENTS_VALID); 1718 } 1719 1720 /* make some assumptions about the time representation rather than abstract it, 1721 * since we currently only support one type of inline timestamping and only on 1722 * EF10. 1723 */ 1724 #define MINOR_TICKS_PER_SECOND 0x8000000 1725 /* Fuzz factor for sync events to be out of order with RX events */ 1726 #define FUZZ (MINOR_TICKS_PER_SECOND / 10) 1727 #define EXPECTED_SYNC_EVENTS_PER_SECOND 4 1728 1729 static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh) 1730 { 1731 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) 1732 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset)); 1733 #else 1734 const u8 *data = eh + efx->rx_packet_ts_offset; 1735 return (u32)data[0] | 1736 (u32)data[1] << 8 | 1737 (u32)data[2] << 16 | 1738 (u32)data[3] << 24; 1739 #endif 1740 } 1741 1742 void __efx_rx_skb_attach_timestamp(struct efx_channel *channel, 1743 struct sk_buff *skb) 1744 { 1745 struct efx_nic *efx = channel->efx; 1746 u32 pkt_timestamp_major, pkt_timestamp_minor; 1747 u32 diff, carry; 1748 struct skb_shared_hwtstamps *timestamps; 1749 1750 pkt_timestamp_minor = (efx_rx_buf_timestamp_minor(efx, 1751 skb_mac_header(skb)) + 1752 (u32) efx->ptp_data->ts_corrections.rx) & 1753 (MINOR_TICKS_PER_SECOND - 1); 1754 1755 /* get the difference between the packet and sync timestamps, 1756 * modulo one second 1757 */ 1758 diff = (pkt_timestamp_minor - channel->sync_timestamp_minor) & 1759 (MINOR_TICKS_PER_SECOND - 1); 1760 /* do we roll over a second boundary and need to carry the one? */ 1761 carry = channel->sync_timestamp_minor + diff > MINOR_TICKS_PER_SECOND ? 1762 1 : 0; 1763 1764 if (diff <= MINOR_TICKS_PER_SECOND / EXPECTED_SYNC_EVENTS_PER_SECOND + 1765 FUZZ) { 1766 /* packet is ahead of the sync event by a quarter of a second or 1767 * less (allowing for fuzz) 1768 */ 1769 pkt_timestamp_major = channel->sync_timestamp_major + carry; 1770 } else if (diff >= MINOR_TICKS_PER_SECOND - FUZZ) { 1771 /* packet is behind the sync event but within the fuzz factor. 1772 * This means the RX packet and sync event crossed as they were 1773 * placed on the event queue, which can sometimes happen. 1774 */ 1775 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry; 1776 } else { 1777 /* it's outside tolerance in both directions. this might be 1778 * indicative of us missing sync events for some reason, so 1779 * we'll call it an error rather than risk giving a bogus 1780 * timestamp. 1781 */ 1782 netif_vdbg(efx, drv, efx->net_dev, 1783 "packet timestamp %x too far from sync event %x:%x\n", 1784 pkt_timestamp_minor, channel->sync_timestamp_major, 1785 channel->sync_timestamp_minor); 1786 return; 1787 } 1788 1789 /* attach the timestamps to the skb */ 1790 timestamps = skb_hwtstamps(skb); 1791 timestamps->hwtstamp = 1792 efx_ptp_s27_to_ktime(pkt_timestamp_major, pkt_timestamp_minor); 1793 } 1794 1795 static int efx_phc_adjfreq(struct ptp_clock_info *ptp, s32 delta) 1796 { 1797 struct efx_ptp_data *ptp_data = container_of(ptp, 1798 struct efx_ptp_data, 1799 phc_clock_info); 1800 struct efx_nic *efx = ptp_data->efx; 1801 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN); 1802 s64 adjustment_ns; 1803 int rc; 1804 1805 if (delta > MAX_PPB) 1806 delta = MAX_PPB; 1807 else if (delta < -MAX_PPB) 1808 delta = -MAX_PPB; 1809 1810 /* Convert ppb to fixed point ns. */ 1811 adjustment_ns = (((s64)delta * PPB_SCALE_WORD) >> 1812 (PPB_EXTRA_BITS + MAX_PPB_BITS)); 1813 1814 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST); 1815 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0); 1816 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns); 1817 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0); 1818 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0); 1819 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj), 1820 NULL, 0, NULL); 1821 if (rc != 0) 1822 return rc; 1823 1824 ptp_data->current_adjfreq = adjustment_ns; 1825 return 0; 1826 } 1827 1828 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta) 1829 { 1830 u32 nic_major, nic_minor; 1831 struct efx_ptp_data *ptp_data = container_of(ptp, 1832 struct efx_ptp_data, 1833 phc_clock_info); 1834 struct efx_nic *efx = ptp_data->efx; 1835 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN); 1836 1837 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor); 1838 1839 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST); 1840 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 1841 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq); 1842 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major); 1843 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor); 1844 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 1845 NULL, 0, NULL); 1846 } 1847 1848 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec *ts) 1849 { 1850 struct efx_ptp_data *ptp_data = container_of(ptp, 1851 struct efx_ptp_data, 1852 phc_clock_info); 1853 struct efx_nic *efx = ptp_data->efx; 1854 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN); 1855 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN); 1856 int rc; 1857 ktime_t kt; 1858 1859 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME); 1860 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 1861 1862 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 1863 outbuf, sizeof(outbuf), NULL); 1864 if (rc != 0) 1865 return rc; 1866 1867 kt = ptp_data->nic_to_kernel_time( 1868 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR), 1869 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0); 1870 *ts = ktime_to_timespec(kt); 1871 return 0; 1872 } 1873 1874 static int efx_phc_settime(struct ptp_clock_info *ptp, 1875 const struct timespec *e_ts) 1876 { 1877 /* Get the current NIC time, efx_phc_gettime. 1878 * Subtract from the desired time to get the offset 1879 * call efx_phc_adjtime with the offset 1880 */ 1881 int rc; 1882 struct timespec time_now; 1883 struct timespec delta; 1884 1885 rc = efx_phc_gettime(ptp, &time_now); 1886 if (rc != 0) 1887 return rc; 1888 1889 delta = timespec_sub(*e_ts, time_now); 1890 1891 rc = efx_phc_adjtime(ptp, timespec_to_ns(&delta)); 1892 if (rc != 0) 1893 return rc; 1894 1895 return 0; 1896 } 1897 1898 static int efx_phc_enable(struct ptp_clock_info *ptp, 1899 struct ptp_clock_request *request, 1900 int enable) 1901 { 1902 struct efx_ptp_data *ptp_data = container_of(ptp, 1903 struct efx_ptp_data, 1904 phc_clock_info); 1905 if (request->type != PTP_CLK_REQ_PPS) 1906 return -EOPNOTSUPP; 1907 1908 ptp_data->nic_ts_enabled = !!enable; 1909 return 0; 1910 } 1911 1912 static const struct efx_channel_type efx_ptp_channel_type = { 1913 .handle_no_channel = efx_ptp_handle_no_channel, 1914 .pre_probe = efx_ptp_probe_channel, 1915 .post_remove = efx_ptp_remove_channel, 1916 .get_name = efx_ptp_get_channel_name, 1917 /* no copy operation; there is no need to reallocate this channel */ 1918 .receive_skb = efx_ptp_rx, 1919 .keep_eventq = false, 1920 }; 1921 1922 void efx_ptp_defer_probe_with_channel(struct efx_nic *efx) 1923 { 1924 /* Check whether PTP is implemented on this NIC. The DISABLE 1925 * operation will succeed if and only if it is implemented. 1926 */ 1927 if (efx_ptp_disable(efx) == 0) 1928 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] = 1929 &efx_ptp_channel_type; 1930 } 1931 1932 void efx_ptp_start_datapath(struct efx_nic *efx) 1933 { 1934 if (efx_ptp_restart(efx)) 1935 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n"); 1936 /* re-enable timestamping if it was previously enabled */ 1937 if (efx->type->ptp_set_ts_sync_events) 1938 efx->type->ptp_set_ts_sync_events(efx, true, true); 1939 } 1940 1941 void efx_ptp_stop_datapath(struct efx_nic *efx) 1942 { 1943 /* temporarily disable timestamping */ 1944 if (efx->type->ptp_set_ts_sync_events) 1945 efx->type->ptp_set_ts_sync_events(efx, false, true); 1946 efx_ptp_stop(efx); 1947 } 1948