1 // SPDX-License-Identifier: GPL-2.0-only 2 /**************************************************************************** 3 * Driver for Solarflare network controllers and boards 4 * Copyright 2011-2013 Solarflare Communications Inc. 5 */ 6 7 /* Theory of operation: 8 * 9 * PTP support is assisted by firmware running on the MC, which provides 10 * the hardware timestamping capabilities. Both transmitted and received 11 * PTP event packets are queued onto internal queues for subsequent processing; 12 * this is because the MC operations are relatively long and would block 13 * block NAPI/interrupt operation. 14 * 15 * Receive event processing: 16 * The event contains the packet's UUID and sequence number, together 17 * with the hardware timestamp. The PTP receive packet queue is searched 18 * for this UUID/sequence number and, if found, put on a pending queue. 19 * Packets not matching are delivered without timestamps (MCDI events will 20 * always arrive after the actual packet). 21 * It is important for the operation of the PTP protocol that the ordering 22 * of packets between the event and general port is maintained. 23 * 24 * Work queue processing: 25 * If work waiting, synchronise host/hardware time 26 * 27 * Transmit: send packet through MC, which returns the transmission time 28 * that is converted to an appropriate timestamp. 29 * 30 * Receive: the packet's reception time is converted to an appropriate 31 * timestamp. 32 */ 33 #include <linux/ip.h> 34 #include <linux/udp.h> 35 #include <linux/time.h> 36 #include <linux/errno.h> 37 #include <linux/ktime.h> 38 #include <linux/module.h> 39 #include <linux/pps_kernel.h> 40 #include <linux/ptp_clock_kernel.h> 41 #include "net_driver.h" 42 #include "efx.h" 43 #include "mcdi.h" 44 #include "mcdi_pcol.h" 45 #include "io.h" 46 #include "tx.h" 47 #include "nic.h" /* indirectly includes ptp.h */ 48 #include "efx_channels.h" 49 50 /* Maximum number of events expected to make up a PTP event */ 51 #define MAX_EVENT_FRAGS 3 52 53 /* Maximum delay, ms, to begin synchronisation */ 54 #define MAX_SYNCHRONISE_WAIT_MS 2 55 56 /* How long, at most, to spend synchronising */ 57 #define SYNCHRONISE_PERIOD_NS 250000 58 59 /* How often to update the shared memory time */ 60 #define SYNCHRONISATION_GRANULARITY_NS 200 61 62 /* Minimum permitted length of a (corrected) synchronisation time */ 63 #define DEFAULT_MIN_SYNCHRONISATION_NS 120 64 65 /* Maximum permitted length of a (corrected) synchronisation time */ 66 #define MAX_SYNCHRONISATION_NS 1000 67 68 /* How many (MC) receive events that can be queued */ 69 #define MAX_RECEIVE_EVENTS 8 70 71 /* Length of (modified) moving average. */ 72 #define AVERAGE_LENGTH 16 73 74 /* How long an unmatched event or packet can be held */ 75 #define PKT_EVENT_LIFETIME_MS 10 76 77 /* How long unused unicast filters can be held */ 78 #define UCAST_FILTER_EXPIRY_JIFFIES msecs_to_jiffies(30000) 79 80 /* Offsets into PTP packet for identification. These offsets are from the 81 * start of the IP header, not the MAC header. Note that neither PTP V1 nor 82 * PTP V2 permit the use of IPV4 options. 83 */ 84 #define PTP_DPORT_OFFSET 22 85 86 #define PTP_V1_VERSION_LENGTH 2 87 #define PTP_V1_VERSION_OFFSET 28 88 89 #define PTP_V1_SEQUENCE_LENGTH 2 90 #define PTP_V1_SEQUENCE_OFFSET 58 91 92 /* The minimum length of a PTP V1 packet for offsets, etc. to be valid: 93 * includes IP header. 94 */ 95 #define PTP_V1_MIN_LENGTH 64 96 97 #define PTP_V2_VERSION_LENGTH 1 98 #define PTP_V2_VERSION_OFFSET 29 99 100 #define PTP_V2_SEQUENCE_LENGTH 2 101 #define PTP_V2_SEQUENCE_OFFSET 58 102 103 /* The minimum length of a PTP V2 packet for offsets, etc. to be valid: 104 * includes IP header. 105 */ 106 #define PTP_V2_MIN_LENGTH 63 107 108 #define PTP_MIN_LENGTH 63 109 110 #define PTP_ADDR_IPV4 0xe0000181 /* 224.0.1.129 */ 111 112 /* ff0e::181 */ 113 static const struct in6_addr ptp_addr_ipv6 = { { { 114 0xff, 0x0e, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0x01, 0x81 } } }; 115 116 /* 01-1B-19-00-00-00 */ 117 static const u8 ptp_addr_ether[ETH_ALEN] __aligned(2) = { 118 0x01, 0x1b, 0x19, 0x00, 0x00, 0x00 }; 119 120 #define PTP_EVENT_PORT 319 121 #define PTP_GENERAL_PORT 320 122 123 /* Annoyingly the format of the version numbers are different between 124 * versions 1 and 2 so it isn't possible to simply look for 1 or 2. 125 */ 126 #define PTP_VERSION_V1 1 127 128 #define PTP_VERSION_V2 2 129 #define PTP_VERSION_V2_MASK 0x0f 130 131 enum ptp_packet_state { 132 PTP_PACKET_STATE_UNMATCHED = 0, 133 PTP_PACKET_STATE_MATCHED, 134 PTP_PACKET_STATE_TIMED_OUT, 135 PTP_PACKET_STATE_MATCH_UNWANTED 136 }; 137 138 /* NIC synchronised with single word of time only comprising 139 * partial seconds and full nanoseconds: 10^9 ~ 2^30 so 2 bits for seconds. 140 */ 141 #define MC_NANOSECOND_BITS 30 142 #define MC_NANOSECOND_MASK ((1 << MC_NANOSECOND_BITS) - 1) 143 #define MC_SECOND_MASK ((1 << (32 - MC_NANOSECOND_BITS)) - 1) 144 145 /* Maximum parts-per-billion adjustment that is acceptable */ 146 #define MAX_PPB 1000000 147 148 /* Precalculate scale word to avoid long long division at runtime */ 149 /* This is equivalent to 2^66 / 10^9. */ 150 #define PPB_SCALE_WORD ((1LL << (57)) / 1953125LL) 151 152 /* How much to shift down after scaling to convert to FP40 */ 153 #define PPB_SHIFT_FP40 26 154 /* ... and FP44. */ 155 #define PPB_SHIFT_FP44 22 156 157 #define PTP_SYNC_ATTEMPTS 4 158 159 /** 160 * struct efx_ptp_match - Matching structure, stored in sk_buff's cb area. 161 * @expiry: Time after which the packet should be delivered irrespective of 162 * event arrival. 163 * @state: The state of the packet - whether it is ready for processing or 164 * whether that is of no interest. 165 */ 166 struct efx_ptp_match { 167 unsigned long expiry; 168 enum ptp_packet_state state; 169 }; 170 171 /** 172 * struct efx_ptp_event_rx - A PTP receive event (from MC) 173 * @link: list of events 174 * @seq0: First part of (PTP) UUID 175 * @seq1: Second part of (PTP) UUID and sequence number 176 * @hwtimestamp: Event timestamp 177 * @expiry: Time which the packet arrived 178 */ 179 struct efx_ptp_event_rx { 180 struct list_head link; 181 u32 seq0; 182 u32 seq1; 183 ktime_t hwtimestamp; 184 unsigned long expiry; 185 }; 186 187 /** 188 * struct efx_ptp_timeset - Synchronisation between host and MC 189 * @host_start: Host time immediately before hardware timestamp taken 190 * @major: Hardware timestamp, major 191 * @minor: Hardware timestamp, minor 192 * @host_end: Host time immediately after hardware timestamp taken 193 * @wait: Number of NIC clock ticks between hardware timestamp being read and 194 * host end time being seen 195 * @window: Difference of host_end and host_start 196 */ 197 struct efx_ptp_timeset { 198 u32 host_start; 199 u32 major; 200 u32 minor; 201 u32 host_end; 202 u32 wait; 203 u32 window; /* Derived: end - start, allowing for wrap */ 204 }; 205 206 /** 207 * struct efx_ptp_rxfilter - Filter for PTP packets 208 * @list: Node of the list where the filter is added 209 * @ether_type: Network protocol of the filter (ETHER_P_IP / ETHER_P_IPV6) 210 * @loc_port: UDP port of the filter (PTP_EVENT_PORT / PTP_GENERAL_PORT) 211 * @loc_host: IPv4/v6 address of the filter 212 * @expiry: time when the filter expires, in jiffies 213 * @handle: Handle ID for the MCDI filters table 214 */ 215 struct efx_ptp_rxfilter { 216 struct list_head list; 217 __be16 ether_type; 218 __be16 loc_port; 219 __be32 loc_host[4]; 220 unsigned long expiry; 221 int handle; 222 }; 223 224 /** 225 * struct efx_ptp_data - Precision Time Protocol (PTP) state 226 * @efx: The NIC context 227 * @channel: The PTP channel (for Medford and Medford2) 228 * @rxq: Receive SKB queue (awaiting timestamps) 229 * @txq: Transmit SKB queue 230 * @workwq: Work queue for processing pending PTP operations 231 * @work: Work task 232 * @cleanup_work: Work task for periodic cleanup 233 * @reset_required: A serious error has occurred and the PTP task needs to be 234 * reset (disable, enable). 235 * @rxfilters_mcast: Receive filters for multicast PTP packets 236 * @rxfilters_ucast: Receive filters for unicast PTP packets 237 * @config: Current timestamp configuration 238 * @enabled: PTP operation enabled 239 * @mode: Mode in which PTP operating (PTP version) 240 * @ns_to_nic_time: Function to convert from scalar nanoseconds to NIC time 241 * @nic_to_kernel_time: Function to convert from NIC to kernel time 242 * @nic_time: contains time details 243 * @nic_time.minor_max: Wrap point for NIC minor times 244 * @nic_time.sync_event_diff_min: Minimum acceptable difference between time 245 * in packet prefix and last MCDI time sync event i.e. how much earlier than 246 * the last sync event time a packet timestamp can be. 247 * @nic_time.sync_event_diff_max: Maximum acceptable difference between time 248 * in packet prefix and last MCDI time sync event i.e. how much later than 249 * the last sync event time a packet timestamp can be. 250 * @nic_time.sync_event_minor_shift: Shift required to make minor time from 251 * field in MCDI time sync event. 252 * @min_synchronisation_ns: Minimum acceptable corrected sync window 253 * @capabilities: Capabilities flags from the NIC 254 * @ts_corrections: contains corrections details 255 * @ts_corrections.ptp_tx: Required driver correction of PTP packet transmit 256 * timestamps 257 * @ts_corrections.ptp_rx: Required driver correction of PTP packet receive 258 * timestamps 259 * @ts_corrections.pps_out: PPS output error (information only) 260 * @ts_corrections.pps_in: Required driver correction of PPS input timestamps 261 * @ts_corrections.general_tx: Required driver correction of general packet 262 * transmit timestamps 263 * @ts_corrections.general_rx: Required driver correction of general packet 264 * receive timestamps 265 * @evt_frags: Partly assembled PTP events 266 * @evt_frag_idx: Current fragment number 267 * @evt_code: Last event code 268 * @start: Address at which MC indicates ready for synchronisation 269 * @host_time_pps: Host time at last PPS 270 * @adjfreq_ppb_shift: Shift required to convert scaled parts-per-billion 271 * frequency adjustment into a fixed point fractional nanosecond format. 272 * @current_adjfreq: Current ppb adjustment. 273 * @phc_clock: Pointer to registered phc device (if primary function) 274 * @phc_clock_info: Registration structure for phc device 275 * @pps_work: pps work task for handling pps events 276 * @pps_workwq: pps work queue 277 * @nic_ts_enabled: Flag indicating if NIC generated TS events are handled 278 * @txbuf: Buffer for use when transmitting (PTP) packets to MC (avoids 279 * allocations in main data path). 280 * @good_syncs: Number of successful synchronisations. 281 * @fast_syncs: Number of synchronisations requiring short delay 282 * @bad_syncs: Number of failed synchronisations. 283 * @sync_timeouts: Number of synchronisation timeouts 284 * @no_time_syncs: Number of synchronisations with no good times. 285 * @invalid_sync_windows: Number of sync windows with bad durations. 286 * @undersize_sync_windows: Number of corrected sync windows that are too small 287 * @oversize_sync_windows: Number of corrected sync windows that are too large 288 * @rx_no_timestamp: Number of packets received without a timestamp. 289 * @timeset: Last set of synchronisation statistics. 290 * @xmit_skb: Transmit SKB function. 291 */ 292 struct efx_ptp_data { 293 struct efx_nic *efx; 294 struct efx_channel *channel; 295 struct sk_buff_head rxq; 296 struct sk_buff_head txq; 297 struct workqueue_struct *workwq; 298 struct work_struct work; 299 struct delayed_work cleanup_work; 300 bool reset_required; 301 struct list_head rxfilters_mcast; 302 struct list_head rxfilters_ucast; 303 struct kernel_hwtstamp_config config; 304 bool enabled; 305 unsigned int mode; 306 void (*ns_to_nic_time)(s64 ns, u32 *nic_major, u32 *nic_minor); 307 ktime_t (*nic_to_kernel_time)(u32 nic_major, u32 nic_minor, 308 s32 correction); 309 struct { 310 u32 minor_max; 311 u32 sync_event_diff_min; 312 u32 sync_event_diff_max; 313 unsigned int sync_event_minor_shift; 314 } nic_time; 315 unsigned int min_synchronisation_ns; 316 unsigned int capabilities; 317 struct { 318 s32 ptp_tx; 319 s32 ptp_rx; 320 s32 pps_out; 321 s32 pps_in; 322 s32 general_tx; 323 s32 general_rx; 324 } ts_corrections; 325 efx_qword_t evt_frags[MAX_EVENT_FRAGS]; 326 int evt_frag_idx; 327 int evt_code; 328 struct efx_buffer start; 329 struct pps_event_time host_time_pps; 330 unsigned int adjfreq_ppb_shift; 331 s64 current_adjfreq; 332 struct ptp_clock *phc_clock; 333 struct ptp_clock_info phc_clock_info; 334 struct work_struct pps_work; 335 struct workqueue_struct *pps_workwq; 336 bool nic_ts_enabled; 337 efx_dword_t txbuf[MCDI_TX_BUF_LEN(MC_CMD_PTP_IN_TRANSMIT_LENMAX)]; 338 339 unsigned int good_syncs; 340 unsigned int fast_syncs; 341 unsigned int bad_syncs; 342 unsigned int sync_timeouts; 343 unsigned int no_time_syncs; 344 unsigned int invalid_sync_windows; 345 unsigned int undersize_sync_windows; 346 unsigned int oversize_sync_windows; 347 unsigned int rx_no_timestamp; 348 struct efx_ptp_timeset 349 timeset[MC_CMD_PTP_OUT_SYNCHRONIZE_TIMESET_MAXNUM]; 350 void (*xmit_skb)(struct efx_nic *efx, struct sk_buff *skb); 351 }; 352 353 static int efx_phc_adjfine(struct ptp_clock_info *ptp, long scaled_ppm); 354 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta); 355 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts); 356 static int efx_phc_settime(struct ptp_clock_info *ptp, 357 const struct timespec64 *e_ts); 358 static int efx_phc_enable(struct ptp_clock_info *ptp, 359 struct ptp_clock_request *request, int on); 360 static int efx_ptp_insert_unicast_filter(struct efx_nic *efx, 361 struct sk_buff *skb); 362 363 bool efx_ptp_use_mac_tx_timestamps(struct efx_nic *efx) 364 { 365 return efx_has_cap(efx, TX_MAC_TIMESTAMPING); 366 } 367 368 /* PTP 'extra' channel is still a traffic channel, but we only create TX queues 369 * if PTP uses MAC TX timestamps, not if PTP uses the MC directly to transmit. 370 */ 371 static bool efx_ptp_want_txqs(struct efx_channel *channel) 372 { 373 return efx_ptp_use_mac_tx_timestamps(channel->efx); 374 } 375 376 #define PTP_SW_STAT(ext_name, field_name) \ 377 { #ext_name, 0, offsetof(struct efx_ptp_data, field_name) } 378 #define PTP_MC_STAT(ext_name, mcdi_name) \ 379 { #ext_name, 32, MC_CMD_PTP_OUT_STATUS_STATS_ ## mcdi_name ## _OFST } 380 static const struct efx_hw_stat_desc efx_ptp_stat_desc[] = { 381 PTP_SW_STAT(ptp_good_syncs, good_syncs), 382 PTP_SW_STAT(ptp_fast_syncs, fast_syncs), 383 PTP_SW_STAT(ptp_bad_syncs, bad_syncs), 384 PTP_SW_STAT(ptp_sync_timeouts, sync_timeouts), 385 PTP_SW_STAT(ptp_no_time_syncs, no_time_syncs), 386 PTP_SW_STAT(ptp_invalid_sync_windows, invalid_sync_windows), 387 PTP_SW_STAT(ptp_undersize_sync_windows, undersize_sync_windows), 388 PTP_SW_STAT(ptp_oversize_sync_windows, oversize_sync_windows), 389 PTP_SW_STAT(ptp_rx_no_timestamp, rx_no_timestamp), 390 PTP_MC_STAT(ptp_tx_timestamp_packets, TX), 391 PTP_MC_STAT(ptp_rx_timestamp_packets, RX), 392 PTP_MC_STAT(ptp_timestamp_packets, TS), 393 PTP_MC_STAT(ptp_filter_matches, FM), 394 PTP_MC_STAT(ptp_non_filter_matches, NFM), 395 }; 396 #define PTP_STAT_COUNT ARRAY_SIZE(efx_ptp_stat_desc) 397 static const unsigned long efx_ptp_stat_mask[] = { 398 [0 ... BITS_TO_LONGS(PTP_STAT_COUNT) - 1] = ~0UL, 399 }; 400 401 size_t efx_ptp_describe_stats(struct efx_nic *efx, u8 **strings) 402 { 403 if (!efx->ptp_data) 404 return 0; 405 406 return efx_nic_describe_stats(efx_ptp_stat_desc, PTP_STAT_COUNT, 407 efx_ptp_stat_mask, strings); 408 } 409 410 size_t efx_ptp_update_stats(struct efx_nic *efx, u64 *stats) 411 { 412 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_STATUS_LEN); 413 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_STATUS_LEN); 414 size_t i; 415 int rc; 416 417 if (!efx->ptp_data) 418 return 0; 419 420 /* Copy software statistics */ 421 for (i = 0; i < PTP_STAT_COUNT; i++) { 422 if (efx_ptp_stat_desc[i].dma_width) 423 continue; 424 stats[i] = *(unsigned int *)((char *)efx->ptp_data + 425 efx_ptp_stat_desc[i].offset); 426 } 427 428 /* Fetch MC statistics. We *must* fill in all statistics or 429 * risk leaking kernel memory to userland, so if the MCDI 430 * request fails we pretend we got zeroes. 431 */ 432 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_STATUS); 433 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 434 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 435 outbuf, sizeof(outbuf), NULL); 436 if (rc) 437 memset(outbuf, 0, sizeof(outbuf)); 438 efx_nic_update_stats(efx_ptp_stat_desc, PTP_STAT_COUNT, 439 efx_ptp_stat_mask, 440 stats, _MCDI_PTR(outbuf, 0), false); 441 442 return PTP_STAT_COUNT; 443 } 444 445 /* To convert from s27 format to ns we multiply then divide by a power of 2. 446 * For the conversion from ns to s27, the operation is also converted to a 447 * multiply and shift. 448 */ 449 #define S27_TO_NS_SHIFT (27) 450 #define NS_TO_S27_MULT (((1ULL << 63) + NSEC_PER_SEC / 2) / NSEC_PER_SEC) 451 #define NS_TO_S27_SHIFT (63 - S27_TO_NS_SHIFT) 452 #define S27_MINOR_MAX (1 << S27_TO_NS_SHIFT) 453 454 /* For Huntington platforms NIC time is in seconds and fractions of a second 455 * where the minor register only uses 27 bits in units of 2^-27s. 456 */ 457 static void efx_ptp_ns_to_s27(s64 ns, u32 *nic_major, u32 *nic_minor) 458 { 459 struct timespec64 ts = ns_to_timespec64(ns); 460 u32 maj = (u32)ts.tv_sec; 461 u32 min = (u32)(((u64)ts.tv_nsec * NS_TO_S27_MULT + 462 (1ULL << (NS_TO_S27_SHIFT - 1))) >> NS_TO_S27_SHIFT); 463 464 /* The conversion can result in the minor value exceeding the maximum. 465 * In this case, round up to the next second. 466 */ 467 if (min >= S27_MINOR_MAX) { 468 min -= S27_MINOR_MAX; 469 maj++; 470 } 471 472 *nic_major = maj; 473 *nic_minor = min; 474 } 475 476 static inline ktime_t efx_ptp_s27_to_ktime(u32 nic_major, u32 nic_minor) 477 { 478 u32 ns = (u32)(((u64)nic_minor * NSEC_PER_SEC + 479 (1ULL << (S27_TO_NS_SHIFT - 1))) >> S27_TO_NS_SHIFT); 480 return ktime_set(nic_major, ns); 481 } 482 483 static ktime_t efx_ptp_s27_to_ktime_correction(u32 nic_major, u32 nic_minor, 484 s32 correction) 485 { 486 /* Apply the correction and deal with carry */ 487 nic_minor += correction; 488 if ((s32)nic_minor < 0) { 489 nic_minor += S27_MINOR_MAX; 490 nic_major--; 491 } else if (nic_minor >= S27_MINOR_MAX) { 492 nic_minor -= S27_MINOR_MAX; 493 nic_major++; 494 } 495 496 return efx_ptp_s27_to_ktime(nic_major, nic_minor); 497 } 498 499 /* For Medford2 platforms the time is in seconds and quarter nanoseconds. */ 500 static void efx_ptp_ns_to_s_qns(s64 ns, u32 *nic_major, u32 *nic_minor) 501 { 502 struct timespec64 ts = ns_to_timespec64(ns); 503 504 *nic_major = (u32)ts.tv_sec; 505 *nic_minor = ts.tv_nsec * 4; 506 } 507 508 static ktime_t efx_ptp_s_qns_to_ktime_correction(u32 nic_major, u32 nic_minor, 509 s32 correction) 510 { 511 ktime_t kt; 512 513 nic_minor = DIV_ROUND_CLOSEST(nic_minor, 4); 514 correction = DIV_ROUND_CLOSEST(correction, 4); 515 516 kt = ktime_set(nic_major, nic_minor); 517 518 if (correction >= 0) 519 kt = ktime_add_ns(kt, (u64)correction); 520 else 521 kt = ktime_sub_ns(kt, (u64)-correction); 522 return kt; 523 } 524 525 struct efx_channel *efx_ptp_channel(struct efx_nic *efx) 526 { 527 return efx->ptp_data ? efx->ptp_data->channel : NULL; 528 } 529 530 void efx_ptp_update_channel(struct efx_nic *efx, struct efx_channel *channel) 531 { 532 if (efx->ptp_data) 533 efx->ptp_data->channel = channel; 534 } 535 536 static u32 last_sync_timestamp_major(struct efx_nic *efx) 537 { 538 struct efx_channel *channel = efx_ptp_channel(efx); 539 u32 major = 0; 540 541 if (channel) 542 major = channel->sync_timestamp_major; 543 return major; 544 } 545 546 /* The 8000 series and later can provide the time from the MAC, which is only 547 * 48 bits long and provides meta-information in the top 2 bits. 548 */ 549 static ktime_t 550 efx_ptp_mac_nic_to_ktime_correction(struct efx_nic *efx, 551 struct efx_ptp_data *ptp, 552 u32 nic_major, u32 nic_minor, 553 s32 correction) 554 { 555 u32 sync_timestamp; 556 ktime_t kt = { 0 }; 557 s16 delta; 558 559 if (!(nic_major & 0x80000000)) { 560 WARN_ON_ONCE(nic_major >> 16); 561 562 /* Medford provides 48 bits of timestamp, so we must get the top 563 * 16 bits from the timesync event state. 564 * 565 * We only have the lower 16 bits of the time now, but we do 566 * have a full resolution timestamp at some point in past. As 567 * long as the difference between the (real) now and the sync 568 * is less than 2^15, then we can reconstruct the difference 569 * between those two numbers using only the lower 16 bits of 570 * each. 571 * 572 * Put another way 573 * 574 * a - b = ((a mod k) - b) mod k 575 * 576 * when -k/2 < (a-b) < k/2. In our case k is 2^16. We know 577 * (a mod k) and b, so can calculate the delta, a - b. 578 * 579 */ 580 sync_timestamp = last_sync_timestamp_major(efx); 581 582 /* Because delta is s16 this does an implicit mask down to 583 * 16 bits which is what we need, assuming 584 * MEDFORD_TX_SECS_EVENT_BITS is 16. delta is signed so that 585 * we can deal with the (unlikely) case of sync timestamps 586 * arriving from the future. 587 */ 588 delta = nic_major - sync_timestamp; 589 590 /* Recover the fully specified time now, by applying the offset 591 * to the (fully specified) sync time. 592 */ 593 nic_major = sync_timestamp + delta; 594 595 kt = ptp->nic_to_kernel_time(nic_major, nic_minor, 596 correction); 597 } 598 return kt; 599 } 600 601 ktime_t efx_ptp_nic_to_kernel_time(struct efx_tx_queue *tx_queue) 602 { 603 struct efx_nic *efx = tx_queue->efx; 604 struct efx_ptp_data *ptp = efx->ptp_data; 605 ktime_t kt; 606 607 if (efx_ptp_use_mac_tx_timestamps(efx)) 608 kt = efx_ptp_mac_nic_to_ktime_correction(efx, ptp, 609 tx_queue->completed_timestamp_major, 610 tx_queue->completed_timestamp_minor, 611 ptp->ts_corrections.general_tx); 612 else 613 kt = ptp->nic_to_kernel_time( 614 tx_queue->completed_timestamp_major, 615 tx_queue->completed_timestamp_minor, 616 ptp->ts_corrections.general_tx); 617 return kt; 618 } 619 620 /* Get PTP attributes and set up time conversions */ 621 static int efx_ptp_get_attributes(struct efx_nic *efx) 622 { 623 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_ATTRIBUTES_LEN); 624 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN); 625 struct efx_ptp_data *ptp = efx->ptp_data; 626 int rc; 627 u32 fmt; 628 size_t out_len; 629 630 /* Get the PTP attributes. If the NIC doesn't support the operation we 631 * use the default format for compatibility with older NICs i.e. 632 * seconds and nanoseconds. 633 */ 634 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_GET_ATTRIBUTES); 635 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 636 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 637 outbuf, sizeof(outbuf), &out_len); 638 if (rc == 0) { 639 fmt = MCDI_DWORD(outbuf, PTP_OUT_GET_ATTRIBUTES_TIME_FORMAT); 640 } else if (rc == -EINVAL) { 641 fmt = MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_NANOSECONDS; 642 } else if (rc == -EPERM) { 643 pci_info(efx->pci_dev, "no PTP support\n"); 644 return rc; 645 } else { 646 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), 647 outbuf, sizeof(outbuf), rc); 648 return rc; 649 } 650 651 switch (fmt) { 652 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_27FRACTION: 653 ptp->ns_to_nic_time = efx_ptp_ns_to_s27; 654 ptp->nic_to_kernel_time = efx_ptp_s27_to_ktime_correction; 655 ptp->nic_time.minor_max = 1 << 27; 656 ptp->nic_time.sync_event_minor_shift = 19; 657 break; 658 case MC_CMD_PTP_OUT_GET_ATTRIBUTES_SECONDS_QTR_NANOSECONDS: 659 ptp->ns_to_nic_time = efx_ptp_ns_to_s_qns; 660 ptp->nic_to_kernel_time = efx_ptp_s_qns_to_ktime_correction; 661 ptp->nic_time.minor_max = 4000000000UL; 662 ptp->nic_time.sync_event_minor_shift = 24; 663 break; 664 default: 665 return -ERANGE; 666 } 667 668 /* Precalculate acceptable difference between the minor time in the 669 * packet prefix and the last MCDI time sync event. We expect the 670 * packet prefix timestamp to be after of sync event by up to one 671 * sync event interval (0.25s) but we allow it to exceed this by a 672 * fuzz factor of (0.1s) 673 */ 674 ptp->nic_time.sync_event_diff_min = ptp->nic_time.minor_max 675 - (ptp->nic_time.minor_max / 10); 676 ptp->nic_time.sync_event_diff_max = (ptp->nic_time.minor_max / 4) 677 + (ptp->nic_time.minor_max / 10); 678 679 /* MC_CMD_PTP_OP_GET_ATTRIBUTES has been extended twice from an older 680 * operation MC_CMD_PTP_OP_GET_TIME_FORMAT. The function now may return 681 * a value to use for the minimum acceptable corrected synchronization 682 * window and may return further capabilities. 683 * If we have the extra information store it. For older firmware that 684 * does not implement the extended command use the default value. 685 */ 686 if (rc == 0 && 687 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_CAPABILITIES_OFST) 688 ptp->min_synchronisation_ns = 689 MCDI_DWORD(outbuf, 690 PTP_OUT_GET_ATTRIBUTES_SYNC_WINDOW_MIN); 691 else 692 ptp->min_synchronisation_ns = DEFAULT_MIN_SYNCHRONISATION_NS; 693 694 if (rc == 0 && 695 out_len >= MC_CMD_PTP_OUT_GET_ATTRIBUTES_LEN) 696 ptp->capabilities = MCDI_DWORD(outbuf, 697 PTP_OUT_GET_ATTRIBUTES_CAPABILITIES); 698 else 699 ptp->capabilities = 0; 700 701 /* Set up the shift for conversion between frequency 702 * adjustments in parts-per-billion and the fixed-point 703 * fractional ns format that the adapter uses. 704 */ 705 if (ptp->capabilities & (1 << MC_CMD_PTP_OUT_GET_ATTRIBUTES_FP44_FREQ_ADJ_LBN)) 706 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP44; 707 else 708 ptp->adjfreq_ppb_shift = PPB_SHIFT_FP40; 709 710 return 0; 711 } 712 713 /* Get PTP timestamp corrections */ 714 static int efx_ptp_get_timestamp_corrections(struct efx_nic *efx) 715 { 716 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_GET_TIMESTAMP_CORRECTIONS_LEN); 717 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN); 718 int rc; 719 size_t out_len; 720 721 /* Get the timestamp corrections from the NIC. If this operation is 722 * not supported (older NICs) then no correction is required. 723 */ 724 MCDI_SET_DWORD(inbuf, PTP_IN_OP, 725 MC_CMD_PTP_OP_GET_TIMESTAMP_CORRECTIONS); 726 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 727 728 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 729 outbuf, sizeof(outbuf), &out_len); 730 if (rc == 0) { 731 efx->ptp_data->ts_corrections.ptp_tx = MCDI_DWORD(outbuf, 732 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_TRANSMIT); 733 efx->ptp_data->ts_corrections.ptp_rx = MCDI_DWORD(outbuf, 734 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_RECEIVE); 735 efx->ptp_data->ts_corrections.pps_out = MCDI_DWORD(outbuf, 736 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_OUT); 737 efx->ptp_data->ts_corrections.pps_in = MCDI_DWORD(outbuf, 738 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_PPS_IN); 739 740 if (out_len >= MC_CMD_PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_LEN) { 741 efx->ptp_data->ts_corrections.general_tx = MCDI_DWORD( 742 outbuf, 743 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_TX); 744 efx->ptp_data->ts_corrections.general_rx = MCDI_DWORD( 745 outbuf, 746 PTP_OUT_GET_TIMESTAMP_CORRECTIONS_V2_GENERAL_RX); 747 } else { 748 efx->ptp_data->ts_corrections.general_tx = 749 efx->ptp_data->ts_corrections.ptp_tx; 750 efx->ptp_data->ts_corrections.general_rx = 751 efx->ptp_data->ts_corrections.ptp_rx; 752 } 753 } else if (rc == -EINVAL) { 754 efx->ptp_data->ts_corrections.ptp_tx = 0; 755 efx->ptp_data->ts_corrections.ptp_rx = 0; 756 efx->ptp_data->ts_corrections.pps_out = 0; 757 efx->ptp_data->ts_corrections.pps_in = 0; 758 efx->ptp_data->ts_corrections.general_tx = 0; 759 efx->ptp_data->ts_corrections.general_rx = 0; 760 } else { 761 efx_mcdi_display_error(efx, MC_CMD_PTP, sizeof(inbuf), outbuf, 762 sizeof(outbuf), rc); 763 return rc; 764 } 765 766 return 0; 767 } 768 769 /* Enable MCDI PTP support. */ 770 static int efx_ptp_enable(struct efx_nic *efx) 771 { 772 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ENABLE_LEN); 773 MCDI_DECLARE_BUF_ERR(outbuf); 774 int rc; 775 776 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ENABLE); 777 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 778 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_QUEUE, 779 efx->ptp_data->channel ? 780 efx->ptp_data->channel->channel : 0); 781 MCDI_SET_DWORD(inbuf, PTP_IN_ENABLE_MODE, efx->ptp_data->mode); 782 783 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 784 outbuf, sizeof(outbuf), NULL); 785 rc = (rc == -EALREADY) ? 0 : rc; 786 if (rc) 787 efx_mcdi_display_error(efx, MC_CMD_PTP, 788 MC_CMD_PTP_IN_ENABLE_LEN, 789 outbuf, sizeof(outbuf), rc); 790 return rc; 791 } 792 793 /* Disable MCDI PTP support. 794 * 795 * Note that this function should never rely on the presence of ptp_data - 796 * may be called before that exists. 797 */ 798 static int efx_ptp_disable(struct efx_nic *efx) 799 { 800 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_DISABLE_LEN); 801 MCDI_DECLARE_BUF_ERR(outbuf); 802 int rc; 803 804 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_DISABLE); 805 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 806 rc = efx_mcdi_rpc_quiet(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 807 outbuf, sizeof(outbuf), NULL); 808 rc = (rc == -EALREADY) ? 0 : rc; 809 /* If we get ENOSYS, the NIC doesn't support PTP, and thus this function 810 * should only have been called during probe. 811 */ 812 if (rc == -ENOSYS || rc == -EPERM) 813 pci_info(efx->pci_dev, "no PTP support\n"); 814 else if (rc) 815 efx_mcdi_display_error(efx, MC_CMD_PTP, 816 MC_CMD_PTP_IN_DISABLE_LEN, 817 outbuf, sizeof(outbuf), rc); 818 return rc; 819 } 820 821 static void efx_ptp_deliver_rx_queue(struct sk_buff_head *q) 822 { 823 struct sk_buff *skb; 824 825 while ((skb = skb_dequeue(q))) { 826 local_bh_disable(); 827 netif_receive_skb(skb); 828 local_bh_enable(); 829 } 830 } 831 832 static void efx_ptp_handle_no_channel(struct efx_nic *efx) 833 { 834 netif_err(efx, drv, efx->net_dev, 835 "ERROR: PTP requires MSI-X and 1 additional interrupt" 836 "vector. PTP disabled\n"); 837 } 838 839 /* Repeatedly send the host time to the MC which will capture the hardware 840 * time. 841 */ 842 static void efx_ptp_send_times(struct efx_nic *efx, 843 struct pps_event_time *last_time) 844 { 845 struct pps_event_time now; 846 struct timespec64 limit; 847 struct efx_ptp_data *ptp = efx->ptp_data; 848 int *mc_running = ptp->start.addr; 849 850 pps_get_ts(&now); 851 limit = now.ts_real; 852 timespec64_add_ns(&limit, SYNCHRONISE_PERIOD_NS); 853 854 /* Write host time for specified period or until MC is done */ 855 while ((timespec64_compare(&now.ts_real, &limit) < 0) && 856 READ_ONCE(*mc_running)) { 857 struct timespec64 update_time; 858 unsigned int host_time; 859 860 /* Don't update continuously to avoid saturating the PCIe bus */ 861 update_time = now.ts_real; 862 timespec64_add_ns(&update_time, SYNCHRONISATION_GRANULARITY_NS); 863 do { 864 pps_get_ts(&now); 865 } while ((timespec64_compare(&now.ts_real, &update_time) < 0) && 866 READ_ONCE(*mc_running)); 867 868 /* Synchronise NIC with single word of time only */ 869 host_time = (now.ts_real.tv_sec << MC_NANOSECOND_BITS | 870 now.ts_real.tv_nsec); 871 /* Update host time in NIC memory */ 872 efx->type->ptp_write_host_time(efx, host_time); 873 } 874 *last_time = now; 875 } 876 877 /* Read a timeset from the MC's results and partial process. */ 878 static void efx_ptp_read_timeset(MCDI_DECLARE_STRUCT_PTR(data), 879 struct efx_ptp_timeset *timeset) 880 { 881 unsigned start_ns, end_ns; 882 883 timeset->host_start = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTSTART); 884 timeset->major = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MAJOR); 885 timeset->minor = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_MINOR); 886 timeset->host_end = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_HOSTEND); 887 timeset->wait = MCDI_DWORD(data, PTP_OUT_SYNCHRONIZE_WAITNS); 888 889 /* Ignore seconds */ 890 start_ns = timeset->host_start & MC_NANOSECOND_MASK; 891 end_ns = timeset->host_end & MC_NANOSECOND_MASK; 892 /* Allow for rollover */ 893 if (end_ns < start_ns) 894 end_ns += NSEC_PER_SEC; 895 /* Determine duration of operation */ 896 timeset->window = end_ns - start_ns; 897 } 898 899 /* Process times received from MC. 900 * 901 * Extract times from returned results, and establish the minimum value 902 * seen. The minimum value represents the "best" possible time and events 903 * too much greater than this are rejected - the machine is, perhaps, too 904 * busy. A number of readings are taken so that, hopefully, at least one good 905 * synchronisation will be seen in the results. 906 */ 907 static int 908 efx_ptp_process_times(struct efx_nic *efx, MCDI_DECLARE_STRUCT_PTR(synch_buf), 909 size_t response_length, 910 const struct pps_event_time *last_time) 911 { 912 unsigned number_readings = 913 MCDI_VAR_ARRAY_LEN(response_length, 914 PTP_OUT_SYNCHRONIZE_TIMESET); 915 unsigned i; 916 unsigned ngood = 0; 917 unsigned last_good = 0; 918 struct efx_ptp_data *ptp = efx->ptp_data; 919 u32 last_sec; 920 u32 start_sec; 921 struct timespec64 delta; 922 ktime_t mc_time; 923 924 if (number_readings == 0) 925 return -EAGAIN; 926 927 /* Read the set of results and find the last good host-MC 928 * synchronization result. The MC times when it finishes reading the 929 * host time so the corrected window time should be fairly constant 930 * for a given platform. Increment stats for any results that appear 931 * to be erroneous. 932 */ 933 for (i = 0; i < number_readings; i++) { 934 s32 window, corrected; 935 struct timespec64 wait; 936 937 efx_ptp_read_timeset( 938 MCDI_ARRAY_STRUCT_PTR(synch_buf, 939 PTP_OUT_SYNCHRONIZE_TIMESET, i), 940 &ptp->timeset[i]); 941 942 wait = ktime_to_timespec64( 943 ptp->nic_to_kernel_time(0, ptp->timeset[i].wait, 0)); 944 window = ptp->timeset[i].window; 945 corrected = window - wait.tv_nsec; 946 947 /* We expect the uncorrected synchronization window to be at 948 * least as large as the interval between host start and end 949 * times. If it is smaller than this then this is mostly likely 950 * to be a consequence of the host's time being adjusted. 951 * Check that the corrected sync window is in a reasonable 952 * range. If it is out of range it is likely to be because an 953 * interrupt or other delay occurred between reading the system 954 * time and writing it to MC memory. 955 */ 956 if (window < SYNCHRONISATION_GRANULARITY_NS) { 957 ++ptp->invalid_sync_windows; 958 } else if (corrected >= MAX_SYNCHRONISATION_NS) { 959 ++ptp->oversize_sync_windows; 960 } else if (corrected < ptp->min_synchronisation_ns) { 961 ++ptp->undersize_sync_windows; 962 } else { 963 ngood++; 964 last_good = i; 965 } 966 } 967 968 if (ngood == 0) { 969 netif_warn(efx, drv, efx->net_dev, 970 "PTP no suitable synchronisations\n"); 971 return -EAGAIN; 972 } 973 974 /* Calculate delay from last good sync (host time) to last_time. 975 * It is possible that the seconds rolled over between taking 976 * the start reading and the last value written by the host. The 977 * timescales are such that a gap of more than one second is never 978 * expected. delta is *not* normalised. 979 */ 980 start_sec = ptp->timeset[last_good].host_start >> MC_NANOSECOND_BITS; 981 last_sec = last_time->ts_real.tv_sec & MC_SECOND_MASK; 982 if (start_sec != last_sec && 983 ((start_sec + 1) & MC_SECOND_MASK) != last_sec) { 984 netif_warn(efx, hw, efx->net_dev, 985 "PTP bad synchronisation seconds\n"); 986 return -EAGAIN; 987 } 988 delta.tv_sec = (last_sec - start_sec) & 1; 989 delta.tv_nsec = 990 last_time->ts_real.tv_nsec - 991 (ptp->timeset[last_good].host_start & MC_NANOSECOND_MASK); 992 993 /* Convert the NIC time at last good sync into kernel time. 994 * No correction is required - this time is the output of a 995 * firmware process. 996 */ 997 mc_time = ptp->nic_to_kernel_time(ptp->timeset[last_good].major, 998 ptp->timeset[last_good].minor, 0); 999 1000 /* Calculate delay from NIC top of second to last_time */ 1001 delta.tv_nsec += ktime_to_timespec64(mc_time).tv_nsec; 1002 1003 /* Set PPS timestamp to match NIC top of second */ 1004 ptp->host_time_pps = *last_time; 1005 pps_sub_ts(&ptp->host_time_pps, delta); 1006 1007 return 0; 1008 } 1009 1010 /* Synchronize times between the host and the MC */ 1011 static int efx_ptp_synchronize(struct efx_nic *efx, unsigned int num_readings) 1012 { 1013 struct efx_ptp_data *ptp = efx->ptp_data; 1014 MCDI_DECLARE_BUF(synch_buf, MC_CMD_PTP_OUT_SYNCHRONIZE_LENMAX); 1015 size_t response_length; 1016 int rc; 1017 unsigned long timeout; 1018 struct pps_event_time last_time = {}; 1019 unsigned int loops = 0; 1020 int *start = ptp->start.addr; 1021 1022 MCDI_SET_DWORD(synch_buf, PTP_IN_OP, MC_CMD_PTP_OP_SYNCHRONIZE); 1023 MCDI_SET_DWORD(synch_buf, PTP_IN_PERIPH_ID, 0); 1024 MCDI_SET_DWORD(synch_buf, PTP_IN_SYNCHRONIZE_NUMTIMESETS, 1025 num_readings); 1026 MCDI_SET_QWORD(synch_buf, PTP_IN_SYNCHRONIZE_START_ADDR, 1027 ptp->start.dma_addr); 1028 1029 /* Clear flag that signals MC ready */ 1030 WRITE_ONCE(*start, 0); 1031 rc = efx_mcdi_rpc_start(efx, MC_CMD_PTP, synch_buf, 1032 MC_CMD_PTP_IN_SYNCHRONIZE_LEN); 1033 EFX_WARN_ON_ONCE_PARANOID(rc); 1034 1035 /* Wait for start from MCDI (or timeout) */ 1036 timeout = jiffies + msecs_to_jiffies(MAX_SYNCHRONISE_WAIT_MS); 1037 while (!READ_ONCE(*start) && (time_before(jiffies, timeout))) { 1038 udelay(20); /* Usually start MCDI execution quickly */ 1039 loops++; 1040 } 1041 1042 if (loops <= 1) 1043 ++ptp->fast_syncs; 1044 if (!time_before(jiffies, timeout)) 1045 ++ptp->sync_timeouts; 1046 1047 if (READ_ONCE(*start)) 1048 efx_ptp_send_times(efx, &last_time); 1049 1050 /* Collect results */ 1051 rc = efx_mcdi_rpc_finish(efx, MC_CMD_PTP, 1052 MC_CMD_PTP_IN_SYNCHRONIZE_LEN, 1053 synch_buf, sizeof(synch_buf), 1054 &response_length); 1055 if (rc == 0) { 1056 rc = efx_ptp_process_times(efx, synch_buf, response_length, 1057 &last_time); 1058 if (rc == 0) 1059 ++ptp->good_syncs; 1060 else 1061 ++ptp->no_time_syncs; 1062 } 1063 1064 /* Increment the bad syncs counter if the synchronize fails, whatever 1065 * the reason. 1066 */ 1067 if (rc != 0) 1068 ++ptp->bad_syncs; 1069 1070 return rc; 1071 } 1072 1073 /* Transmit a PTP packet via the dedicated hardware timestamped queue. */ 1074 static void efx_ptp_xmit_skb_queue(struct efx_nic *efx, struct sk_buff *skb) 1075 { 1076 struct efx_ptp_data *ptp_data = efx->ptp_data; 1077 u8 type = efx_tx_csum_type_skb(skb); 1078 struct efx_tx_queue *tx_queue; 1079 1080 tx_queue = efx_channel_get_tx_queue(ptp_data->channel, type); 1081 if (tx_queue && tx_queue->timestamping) { 1082 skb_get(skb); 1083 1084 /* This code invokes normal driver TX code which is always 1085 * protected from softirqs when called from generic TX code, 1086 * which in turn disables preemption. Look at __dev_queue_xmit 1087 * which uses rcu_read_lock_bh disabling preemption for RCU 1088 * plus disabling softirqs. We do not need RCU reader 1089 * protection here. 1090 * 1091 * Although it is theoretically safe for current PTP TX/RX code 1092 * running without disabling softirqs, there are three good 1093 * reasond for doing so: 1094 * 1095 * 1) The code invoked is mainly implemented for non-PTP 1096 * packets and it is always executed with softirqs 1097 * disabled. 1098 * 2) This being a single PTP packet, better to not 1099 * interrupt its processing by softirqs which can lead 1100 * to high latencies. 1101 * 3) netdev_xmit_more checks preemption is disabled and 1102 * triggers a BUG_ON if not. 1103 */ 1104 local_bh_disable(); 1105 efx_enqueue_skb(tx_queue, skb); 1106 local_bh_enable(); 1107 1108 /* We need to add the filters after enqueuing the packet. 1109 * Otherwise, there's high latency in sending back the 1110 * timestamp, causing ptp4l timeouts 1111 */ 1112 efx_ptp_insert_unicast_filter(efx, skb); 1113 dev_consume_skb_any(skb); 1114 } else { 1115 WARN_ONCE(1, "PTP channel has no timestamped tx queue\n"); 1116 dev_kfree_skb_any(skb); 1117 } 1118 } 1119 1120 /* Transmit a PTP packet, via the MCDI interface, to the wire. */ 1121 static void efx_ptp_xmit_skb_mc(struct efx_nic *efx, struct sk_buff *skb) 1122 { 1123 MCDI_DECLARE_BUF(txtime, MC_CMD_PTP_OUT_TRANSMIT_LEN); 1124 struct efx_ptp_data *ptp_data = efx->ptp_data; 1125 struct skb_shared_hwtstamps timestamps; 1126 size_t len; 1127 int rc; 1128 1129 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_OP, MC_CMD_PTP_OP_TRANSMIT); 1130 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_PERIPH_ID, 0); 1131 MCDI_SET_DWORD(ptp_data->txbuf, PTP_IN_TRANSMIT_LENGTH, skb->len); 1132 if (skb_shinfo(skb)->nr_frags != 0) { 1133 rc = skb_linearize(skb); 1134 if (rc != 0) 1135 goto fail; 1136 } 1137 1138 if (skb->ip_summed == CHECKSUM_PARTIAL) { 1139 rc = skb_checksum_help(skb); 1140 if (rc != 0) 1141 goto fail; 1142 } 1143 skb_copy_from_linear_data(skb, 1144 MCDI_PTR(ptp_data->txbuf, 1145 PTP_IN_TRANSMIT_PACKET), 1146 skb->len); 1147 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, 1148 ptp_data->txbuf, MC_CMD_PTP_IN_TRANSMIT_LEN(skb->len), 1149 txtime, sizeof(txtime), &len); 1150 if (rc != 0) 1151 goto fail; 1152 1153 memset(×tamps, 0, sizeof(timestamps)); 1154 timestamps.hwtstamp = ptp_data->nic_to_kernel_time( 1155 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MAJOR), 1156 MCDI_DWORD(txtime, PTP_OUT_TRANSMIT_MINOR), 1157 ptp_data->ts_corrections.ptp_tx); 1158 1159 skb_tstamp_tx(skb, ×tamps); 1160 1161 /* Add the filters after sending back the timestamp to avoid delaying it 1162 * or ptp4l may timeout. 1163 */ 1164 efx_ptp_insert_unicast_filter(efx, skb); 1165 1166 fail: 1167 dev_kfree_skb_any(skb); 1168 1169 return; 1170 } 1171 1172 /* Process any queued receive events and corresponding packets 1173 * 1174 * q is returned with all the packets that are ready for delivery. 1175 */ 1176 static void efx_ptp_process_events(struct efx_nic *efx, struct sk_buff_head *q) 1177 { 1178 struct efx_ptp_data *ptp = efx->ptp_data; 1179 struct sk_buff *skb; 1180 1181 while ((skb = skb_dequeue(&ptp->rxq))) { 1182 struct efx_ptp_match *match; 1183 1184 match = (struct efx_ptp_match *)skb->cb; 1185 if (match->state == PTP_PACKET_STATE_MATCH_UNWANTED) { 1186 __skb_queue_tail(q, skb); 1187 } else if (time_after(jiffies, match->expiry)) { 1188 match->state = PTP_PACKET_STATE_TIMED_OUT; 1189 ++ptp->rx_no_timestamp; 1190 __skb_queue_tail(q, skb); 1191 } else { 1192 /* Replace unprocessed entry and stop */ 1193 skb_queue_head(&ptp->rxq, skb); 1194 break; 1195 } 1196 } 1197 } 1198 1199 /* Complete processing of a received packet */ 1200 static inline void efx_ptp_process_rx(struct efx_nic *efx, struct sk_buff *skb) 1201 { 1202 local_bh_disable(); 1203 netif_receive_skb(skb); 1204 local_bh_enable(); 1205 } 1206 1207 static struct efx_ptp_rxfilter * 1208 efx_ptp_find_filter(struct list_head *filter_list, struct efx_filter_spec *spec) 1209 { 1210 struct efx_ptp_rxfilter *rxfilter; 1211 1212 list_for_each_entry(rxfilter, filter_list, list) { 1213 if (rxfilter->ether_type == spec->ether_type && 1214 rxfilter->loc_port == spec->loc_port && 1215 !memcmp(rxfilter->loc_host, spec->loc_host, sizeof(spec->loc_host))) 1216 return rxfilter; 1217 } 1218 1219 return NULL; 1220 } 1221 1222 static void efx_ptp_remove_one_filter(struct efx_nic *efx, 1223 struct efx_ptp_rxfilter *rxfilter) 1224 { 1225 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED, 1226 rxfilter->handle); 1227 list_del(&rxfilter->list); 1228 kfree(rxfilter); 1229 } 1230 1231 static void efx_ptp_remove_filters(struct efx_nic *efx, 1232 struct list_head *filter_list) 1233 { 1234 struct efx_ptp_rxfilter *rxfilter, *tmp; 1235 1236 list_for_each_entry_safe(rxfilter, tmp, filter_list, list) 1237 efx_ptp_remove_one_filter(efx, rxfilter); 1238 } 1239 1240 static void efx_ptp_init_filter(struct efx_nic *efx, 1241 struct efx_filter_spec *rxfilter) 1242 { 1243 struct efx_channel *channel = efx->ptp_data->channel; 1244 struct efx_rx_queue *queue = efx_channel_get_rx_queue(channel); 1245 1246 efx_filter_init_rx(rxfilter, EFX_FILTER_PRI_REQUIRED, 0, 1247 efx_rx_queue_index(queue)); 1248 } 1249 1250 static int efx_ptp_insert_filter(struct efx_nic *efx, 1251 struct list_head *filter_list, 1252 struct efx_filter_spec *spec, 1253 unsigned long expiry) 1254 { 1255 struct efx_ptp_data *ptp = efx->ptp_data; 1256 struct efx_ptp_rxfilter *rxfilter; 1257 int rc; 1258 1259 rxfilter = efx_ptp_find_filter(filter_list, spec); 1260 if (rxfilter) { 1261 rxfilter->expiry = expiry; 1262 return 0; 1263 } 1264 1265 rxfilter = kzalloc_obj(*rxfilter); 1266 if (!rxfilter) 1267 return -ENOMEM; 1268 1269 rc = efx_filter_insert_filter(efx, spec, true); 1270 if (rc < 0) 1271 goto fail; 1272 1273 rxfilter->handle = rc; 1274 rxfilter->ether_type = spec->ether_type; 1275 rxfilter->loc_port = spec->loc_port; 1276 memcpy(rxfilter->loc_host, spec->loc_host, sizeof(spec->loc_host)); 1277 rxfilter->expiry = expiry; 1278 list_add(&rxfilter->list, filter_list); 1279 1280 queue_delayed_work(ptp->workwq, &ptp->cleanup_work, 1281 UCAST_FILTER_EXPIRY_JIFFIES + 1); 1282 1283 return 0; 1284 1285 fail: 1286 kfree(rxfilter); 1287 return rc; 1288 } 1289 1290 static int efx_ptp_insert_ipv4_filter(struct efx_nic *efx, 1291 struct list_head *filter_list, 1292 __be32 addr, u16 port, 1293 unsigned long expiry) 1294 { 1295 struct efx_filter_spec spec; 1296 1297 efx_ptp_init_filter(efx, &spec); 1298 efx_filter_set_ipv4_local(&spec, IPPROTO_UDP, addr, htons(port)); 1299 return efx_ptp_insert_filter(efx, filter_list, &spec, expiry); 1300 } 1301 1302 static int efx_ptp_insert_ipv6_filter(struct efx_nic *efx, 1303 struct list_head *filter_list, 1304 const struct in6_addr *addr, u16 port, 1305 unsigned long expiry) 1306 { 1307 struct efx_filter_spec spec; 1308 1309 efx_ptp_init_filter(efx, &spec); 1310 efx_filter_set_ipv6_local(&spec, IPPROTO_UDP, addr, htons(port)); 1311 return efx_ptp_insert_filter(efx, filter_list, &spec, expiry); 1312 } 1313 1314 static int efx_ptp_insert_eth_multicast_filter(struct efx_nic *efx) 1315 { 1316 struct efx_ptp_data *ptp = efx->ptp_data; 1317 struct efx_filter_spec spec; 1318 1319 efx_ptp_init_filter(efx, &spec); 1320 efx_filter_set_eth_local(&spec, EFX_FILTER_VID_UNSPEC, ptp_addr_ether); 1321 spec.match_flags |= EFX_FILTER_MATCH_ETHER_TYPE; 1322 spec.ether_type = htons(ETH_P_1588); 1323 return efx_ptp_insert_filter(efx, &ptp->rxfilters_mcast, &spec, 0); 1324 } 1325 1326 static int efx_ptp_insert_multicast_filters(struct efx_nic *efx) 1327 { 1328 struct efx_ptp_data *ptp = efx->ptp_data; 1329 int rc; 1330 1331 if (!ptp->channel || !list_empty(&ptp->rxfilters_mcast)) 1332 return 0; 1333 1334 /* Must filter on both event and general ports to ensure 1335 * that there is no packet re-ordering. 1336 */ 1337 rc = efx_ptp_insert_ipv4_filter(efx, &ptp->rxfilters_mcast, 1338 htonl(PTP_ADDR_IPV4), PTP_EVENT_PORT, 1339 0); 1340 if (rc < 0) 1341 goto fail; 1342 1343 rc = efx_ptp_insert_ipv4_filter(efx, &ptp->rxfilters_mcast, 1344 htonl(PTP_ADDR_IPV4), PTP_GENERAL_PORT, 1345 0); 1346 if (rc < 0) 1347 goto fail; 1348 1349 /* if the NIC supports hw timestamps by the MAC, we can support 1350 * PTP over IPv6 and Ethernet 1351 */ 1352 if (efx_ptp_use_mac_tx_timestamps(efx)) { 1353 rc = efx_ptp_insert_ipv6_filter(efx, &ptp->rxfilters_mcast, 1354 &ptp_addr_ipv6, PTP_EVENT_PORT, 0); 1355 if (rc < 0) 1356 goto fail; 1357 1358 rc = efx_ptp_insert_ipv6_filter(efx, &ptp->rxfilters_mcast, 1359 &ptp_addr_ipv6, PTP_GENERAL_PORT, 0); 1360 if (rc < 0) 1361 goto fail; 1362 1363 rc = efx_ptp_insert_eth_multicast_filter(efx); 1364 1365 /* Not all firmware variants support this filter */ 1366 if (rc < 0 && rc != -EPROTONOSUPPORT) 1367 goto fail; 1368 } 1369 1370 return 0; 1371 1372 fail: 1373 efx_ptp_remove_filters(efx, &ptp->rxfilters_mcast); 1374 return rc; 1375 } 1376 1377 static bool efx_ptp_valid_unicast_event_pkt(struct sk_buff *skb) 1378 { 1379 if (skb->protocol == htons(ETH_P_IP)) { 1380 return ip_hdr(skb)->daddr != htonl(PTP_ADDR_IPV4) && 1381 ip_hdr(skb)->protocol == IPPROTO_UDP && 1382 udp_hdr(skb)->source == htons(PTP_EVENT_PORT); 1383 } else if (skb->protocol == htons(ETH_P_IPV6)) { 1384 return !ipv6_addr_equal(&ipv6_hdr(skb)->daddr, &ptp_addr_ipv6) && 1385 ipv6_hdr(skb)->nexthdr == IPPROTO_UDP && 1386 udp_hdr(skb)->source == htons(PTP_EVENT_PORT); 1387 } 1388 return false; 1389 } 1390 1391 static int efx_ptp_insert_unicast_filter(struct efx_nic *efx, 1392 struct sk_buff *skb) 1393 { 1394 struct efx_ptp_data *ptp = efx->ptp_data; 1395 unsigned long expiry; 1396 int rc; 1397 1398 if (!efx_ptp_valid_unicast_event_pkt(skb)) 1399 return -EINVAL; 1400 1401 expiry = jiffies + UCAST_FILTER_EXPIRY_JIFFIES; 1402 1403 if (skb->protocol == htons(ETH_P_IP)) { 1404 __be32 addr = ip_hdr(skb)->saddr; 1405 1406 rc = efx_ptp_insert_ipv4_filter(efx, &ptp->rxfilters_ucast, 1407 addr, PTP_EVENT_PORT, expiry); 1408 if (rc < 0) 1409 goto out; 1410 1411 rc = efx_ptp_insert_ipv4_filter(efx, &ptp->rxfilters_ucast, 1412 addr, PTP_GENERAL_PORT, expiry); 1413 } else if (efx_ptp_use_mac_tx_timestamps(efx)) { 1414 /* IPv6 PTP only supported by devices with MAC hw timestamp */ 1415 struct in6_addr *addr = &ipv6_hdr(skb)->saddr; 1416 1417 rc = efx_ptp_insert_ipv6_filter(efx, &ptp->rxfilters_ucast, 1418 addr, PTP_EVENT_PORT, expiry); 1419 if (rc < 0) 1420 goto out; 1421 1422 rc = efx_ptp_insert_ipv6_filter(efx, &ptp->rxfilters_ucast, 1423 addr, PTP_GENERAL_PORT, expiry); 1424 } else { 1425 return -EOPNOTSUPP; 1426 } 1427 1428 out: 1429 return rc; 1430 } 1431 1432 static int efx_ptp_start(struct efx_nic *efx) 1433 { 1434 struct efx_ptp_data *ptp = efx->ptp_data; 1435 int rc; 1436 1437 ptp->reset_required = false; 1438 1439 rc = efx_ptp_insert_multicast_filters(efx); 1440 if (rc) 1441 return rc; 1442 1443 rc = efx_ptp_enable(efx); 1444 if (rc != 0) 1445 goto fail; 1446 1447 ptp->evt_frag_idx = 0; 1448 ptp->current_adjfreq = 0; 1449 1450 return 0; 1451 1452 fail: 1453 efx_ptp_remove_filters(efx, &ptp->rxfilters_mcast); 1454 return rc; 1455 } 1456 1457 static int efx_ptp_stop(struct efx_nic *efx) 1458 { 1459 struct efx_ptp_data *ptp = efx->ptp_data; 1460 int rc; 1461 1462 if (ptp == NULL) 1463 return 0; 1464 1465 rc = efx_ptp_disable(efx); 1466 1467 efx_ptp_remove_filters(efx, &ptp->rxfilters_mcast); 1468 efx_ptp_remove_filters(efx, &ptp->rxfilters_ucast); 1469 1470 /* Make sure RX packets are really delivered */ 1471 efx_ptp_deliver_rx_queue(&efx->ptp_data->rxq); 1472 skb_queue_purge(&efx->ptp_data->txq); 1473 1474 return rc; 1475 } 1476 1477 static int efx_ptp_restart(struct efx_nic *efx) 1478 { 1479 if (efx->ptp_data && efx->ptp_data->enabled) 1480 return efx_ptp_start(efx); 1481 return 0; 1482 } 1483 1484 static void efx_ptp_pps_worker(struct work_struct *work) 1485 { 1486 struct efx_ptp_data *ptp = 1487 container_of(work, struct efx_ptp_data, pps_work); 1488 struct efx_nic *efx = ptp->efx; 1489 struct ptp_clock_event ptp_evt; 1490 1491 if (efx_ptp_synchronize(efx, PTP_SYNC_ATTEMPTS)) 1492 return; 1493 1494 ptp_evt.type = PTP_CLOCK_PPSUSR; 1495 ptp_evt.pps_times = ptp->host_time_pps; 1496 ptp_clock_event(ptp->phc_clock, &ptp_evt); 1497 } 1498 1499 static void efx_ptp_worker(struct work_struct *work) 1500 { 1501 struct efx_ptp_data *ptp_data = 1502 container_of(work, struct efx_ptp_data, work); 1503 struct efx_nic *efx = ptp_data->efx; 1504 struct sk_buff *skb; 1505 struct sk_buff_head tempq; 1506 1507 if (ptp_data->reset_required) { 1508 efx_ptp_stop(efx); 1509 efx_ptp_start(efx); 1510 return; 1511 } 1512 1513 __skb_queue_head_init(&tempq); 1514 efx_ptp_process_events(efx, &tempq); 1515 1516 while ((skb = skb_dequeue(&ptp_data->txq))) 1517 ptp_data->xmit_skb(efx, skb); 1518 1519 while ((skb = __skb_dequeue(&tempq))) 1520 efx_ptp_process_rx(efx, skb); 1521 } 1522 1523 static void efx_ptp_cleanup_worker(struct work_struct *work) 1524 { 1525 struct efx_ptp_data *ptp = 1526 container_of(work, struct efx_ptp_data, cleanup_work.work); 1527 struct efx_ptp_rxfilter *rxfilter, *tmp; 1528 1529 list_for_each_entry_safe(rxfilter, tmp, &ptp->rxfilters_ucast, list) { 1530 if (time_is_before_jiffies(rxfilter->expiry)) 1531 efx_ptp_remove_one_filter(ptp->efx, rxfilter); 1532 } 1533 1534 if (!list_empty(&ptp->rxfilters_ucast)) { 1535 queue_delayed_work(ptp->workwq, &ptp->cleanup_work, 1536 UCAST_FILTER_EXPIRY_JIFFIES + 1); 1537 } 1538 } 1539 1540 static const struct ptp_clock_info efx_phc_clock_info = { 1541 .owner = THIS_MODULE, 1542 .name = "sfc", 1543 .max_adj = MAX_PPB, 1544 .n_alarm = 0, 1545 .n_ext_ts = 0, 1546 .n_per_out = 0, 1547 .n_pins = 0, 1548 .pps = 1, 1549 .adjfine = efx_phc_adjfine, 1550 .adjtime = efx_phc_adjtime, 1551 .gettime64 = efx_phc_gettime, 1552 .settime64 = efx_phc_settime, 1553 .enable = efx_phc_enable, 1554 }; 1555 1556 /* Initialise PTP state. */ 1557 int efx_ptp_probe(struct efx_nic *efx, struct efx_channel *channel) 1558 { 1559 struct efx_ptp_data *ptp; 1560 int rc = 0; 1561 1562 if (efx->ptp_data) { 1563 efx->ptp_data->channel = channel; 1564 return 0; 1565 } 1566 1567 ptp = kzalloc_obj(struct efx_ptp_data); 1568 efx->ptp_data = ptp; 1569 if (!efx->ptp_data) 1570 return -ENOMEM; 1571 1572 ptp->efx = efx; 1573 ptp->channel = channel; 1574 1575 rc = efx_nic_alloc_buffer(efx, &ptp->start, sizeof(int), GFP_KERNEL); 1576 if (rc != 0) 1577 goto fail1; 1578 1579 skb_queue_head_init(&ptp->rxq); 1580 skb_queue_head_init(&ptp->txq); 1581 ptp->workwq = create_singlethread_workqueue("sfc_ptp"); 1582 if (!ptp->workwq) { 1583 rc = -ENOMEM; 1584 goto fail2; 1585 } 1586 1587 if (efx_ptp_use_mac_tx_timestamps(efx)) { 1588 ptp->xmit_skb = efx_ptp_xmit_skb_queue; 1589 /* Request sync events on this channel. */ 1590 channel->sync_events_state = SYNC_EVENTS_QUIESCENT; 1591 } else { 1592 ptp->xmit_skb = efx_ptp_xmit_skb_mc; 1593 } 1594 1595 INIT_WORK(&ptp->work, efx_ptp_worker); 1596 INIT_DELAYED_WORK(&ptp->cleanup_work, efx_ptp_cleanup_worker); 1597 ptp->config.flags = 0; 1598 ptp->config.tx_type = HWTSTAMP_TX_OFF; 1599 ptp->config.rx_filter = HWTSTAMP_FILTER_NONE; 1600 INIT_LIST_HEAD(&ptp->rxfilters_mcast); 1601 INIT_LIST_HEAD(&ptp->rxfilters_ucast); 1602 1603 /* Get the NIC PTP attributes and set up time conversions */ 1604 rc = efx_ptp_get_attributes(efx); 1605 if (rc < 0) 1606 goto fail3; 1607 1608 /* Get the timestamp corrections */ 1609 rc = efx_ptp_get_timestamp_corrections(efx); 1610 if (rc < 0) 1611 goto fail3; 1612 1613 if (efx->mcdi->fn_flags & 1614 (1 << MC_CMD_DRV_ATTACH_EXT_OUT_FLAG_PRIMARY)) { 1615 ptp->phc_clock_info = efx_phc_clock_info; 1616 ptp->phc_clock = ptp_clock_register(&ptp->phc_clock_info, 1617 &efx->pci_dev->dev); 1618 if (IS_ERR(ptp->phc_clock)) { 1619 rc = PTR_ERR(ptp->phc_clock); 1620 goto fail3; 1621 } else if (ptp->phc_clock) { 1622 INIT_WORK(&ptp->pps_work, efx_ptp_pps_worker); 1623 ptp->pps_workwq = create_singlethread_workqueue("sfc_pps"); 1624 if (!ptp->pps_workwq) { 1625 rc = -ENOMEM; 1626 goto fail4; 1627 } 1628 } 1629 } 1630 ptp->nic_ts_enabled = false; 1631 1632 return 0; 1633 fail4: 1634 ptp_clock_unregister(efx->ptp_data->phc_clock); 1635 1636 fail3: 1637 destroy_workqueue(efx->ptp_data->workwq); 1638 1639 fail2: 1640 efx_nic_free_buffer(efx, &ptp->start); 1641 1642 fail1: 1643 kfree(efx->ptp_data); 1644 efx->ptp_data = NULL; 1645 1646 return rc; 1647 } 1648 1649 /* Initialise PTP channel. 1650 * 1651 * Setting core_index to zero causes the queue to be initialised and doesn't 1652 * overlap with 'rxq0' because ptp.c doesn't use skb_record_rx_queue. 1653 */ 1654 static int efx_ptp_probe_channel(struct efx_channel *channel) 1655 { 1656 struct efx_nic *efx = channel->efx; 1657 int rc; 1658 1659 channel->irq_moderation_us = 0; 1660 channel->rx_queue.core_index = 0; 1661 1662 rc = efx_ptp_probe(efx, channel); 1663 /* Failure to probe PTP is not fatal; this channel will just not be 1664 * used for anything. 1665 * In the case of EPERM, efx_ptp_probe will print its own message (in 1666 * efx_ptp_get_attributes()), so we don't need to. 1667 */ 1668 if (rc && rc != -EPERM) 1669 netif_warn(efx, drv, efx->net_dev, 1670 "Failed to probe PTP, rc=%d\n", rc); 1671 return 0; 1672 } 1673 1674 void efx_ptp_remove(struct efx_nic *efx) 1675 { 1676 if (!efx->ptp_data) 1677 return; 1678 1679 (void)efx_ptp_disable(efx); 1680 1681 cancel_work_sync(&efx->ptp_data->work); 1682 cancel_delayed_work_sync(&efx->ptp_data->cleanup_work); 1683 if (efx->ptp_data->pps_workwq) 1684 cancel_work_sync(&efx->ptp_data->pps_work); 1685 1686 skb_queue_purge(&efx->ptp_data->rxq); 1687 skb_queue_purge(&efx->ptp_data->txq); 1688 1689 if (efx->ptp_data->phc_clock) { 1690 destroy_workqueue(efx->ptp_data->pps_workwq); 1691 ptp_clock_unregister(efx->ptp_data->phc_clock); 1692 } 1693 1694 destroy_workqueue(efx->ptp_data->workwq); 1695 1696 efx_nic_free_buffer(efx, &efx->ptp_data->start); 1697 kfree(efx->ptp_data); 1698 efx->ptp_data = NULL; 1699 } 1700 1701 static void efx_ptp_remove_channel(struct efx_channel *channel) 1702 { 1703 efx_ptp_remove(channel->efx); 1704 } 1705 1706 static void efx_ptp_get_channel_name(struct efx_channel *channel, 1707 char *buf, size_t len) 1708 { 1709 snprintf(buf, len, "%s-ptp", channel->efx->name); 1710 } 1711 1712 /* Determine whether this packet should be processed by the PTP module 1713 * or transmitted conventionally. 1714 */ 1715 bool efx_ptp_is_ptp_tx(struct efx_nic *efx, struct sk_buff *skb) 1716 { 1717 return efx->ptp_data && 1718 efx->ptp_data->enabled && 1719 skb->len >= PTP_MIN_LENGTH && 1720 skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM && 1721 likely(skb->protocol == htons(ETH_P_IP)) && 1722 skb_transport_header_was_set(skb) && 1723 skb_network_header_len(skb) >= sizeof(struct iphdr) && 1724 ip_hdr(skb)->protocol == IPPROTO_UDP && 1725 skb_headlen(skb) >= 1726 skb_transport_offset(skb) + sizeof(struct udphdr) && 1727 udp_hdr(skb)->dest == htons(PTP_EVENT_PORT); 1728 } 1729 1730 /* Receive a PTP packet. Packets are queued until the arrival of 1731 * the receive timestamp from the MC - this will probably occur after the 1732 * packet arrival because of the processing in the MC. 1733 */ 1734 static bool efx_ptp_rx(struct efx_channel *channel, struct sk_buff *skb) 1735 { 1736 struct efx_nic *efx = channel->efx; 1737 struct efx_ptp_data *ptp = efx->ptp_data; 1738 struct efx_ptp_match *match = (struct efx_ptp_match *)skb->cb; 1739 unsigned int version; 1740 u8 *data; 1741 1742 match->expiry = jiffies + msecs_to_jiffies(PKT_EVENT_LIFETIME_MS); 1743 1744 /* Correct version? */ 1745 if (ptp->mode == MC_CMD_PTP_MODE_V1) { 1746 if (!pskb_may_pull(skb, PTP_V1_MIN_LENGTH)) { 1747 return false; 1748 } 1749 data = skb->data; 1750 version = ntohs(*(__be16 *)&data[PTP_V1_VERSION_OFFSET]); 1751 if (version != PTP_VERSION_V1) { 1752 return false; 1753 } 1754 } else { 1755 if (!pskb_may_pull(skb, PTP_V2_MIN_LENGTH)) { 1756 return false; 1757 } 1758 data = skb->data; 1759 version = data[PTP_V2_VERSION_OFFSET]; 1760 if ((version & PTP_VERSION_V2_MASK) != PTP_VERSION_V2) { 1761 return false; 1762 } 1763 } 1764 1765 /* Does this packet require timestamping? */ 1766 if (ntohs(*(__be16 *)&data[PTP_DPORT_OFFSET]) == PTP_EVENT_PORT) { 1767 match->state = PTP_PACKET_STATE_UNMATCHED; 1768 1769 /* We expect the sequence number to be in the same position in 1770 * the packet for PTP V1 and V2 1771 */ 1772 BUILD_BUG_ON(PTP_V1_SEQUENCE_OFFSET != PTP_V2_SEQUENCE_OFFSET); 1773 BUILD_BUG_ON(PTP_V1_SEQUENCE_LENGTH != PTP_V2_SEQUENCE_LENGTH); 1774 } else { 1775 match->state = PTP_PACKET_STATE_MATCH_UNWANTED; 1776 } 1777 1778 skb_queue_tail(&ptp->rxq, skb); 1779 queue_work(ptp->workwq, &ptp->work); 1780 1781 return true; 1782 } 1783 1784 /* Transmit a PTP packet. This has to be transmitted by the MC 1785 * itself, through an MCDI call. MCDI calls aren't permitted 1786 * in the transmit path so defer the actual transmission to a suitable worker. 1787 */ 1788 int efx_ptp_tx(struct efx_nic *efx, struct sk_buff *skb) 1789 { 1790 struct efx_ptp_data *ptp = efx->ptp_data; 1791 1792 skb_queue_tail(&ptp->txq, skb); 1793 1794 if ((udp_hdr(skb)->dest == htons(PTP_EVENT_PORT)) && 1795 (skb->len <= MC_CMD_PTP_IN_TRANSMIT_PACKET_MAXNUM)) 1796 efx_xmit_hwtstamp_pending(skb); 1797 queue_work(ptp->workwq, &ptp->work); 1798 1799 return NETDEV_TX_OK; 1800 } 1801 1802 int efx_ptp_change_mode(struct efx_nic *efx, bool enable_wanted, 1803 unsigned int new_mode) 1804 { 1805 if ((enable_wanted != efx->ptp_data->enabled) || 1806 (enable_wanted && (efx->ptp_data->mode != new_mode))) { 1807 int rc = 0; 1808 1809 if (enable_wanted) { 1810 /* Change of mode requires disable */ 1811 if (efx->ptp_data->enabled && 1812 (efx->ptp_data->mode != new_mode)) { 1813 efx->ptp_data->enabled = false; 1814 rc = efx_ptp_stop(efx); 1815 if (rc != 0) 1816 return rc; 1817 } 1818 1819 /* Set new operating mode and establish 1820 * baseline synchronisation, which must 1821 * succeed. 1822 */ 1823 efx->ptp_data->mode = new_mode; 1824 if (netif_running(efx->net_dev)) 1825 rc = efx_ptp_start(efx); 1826 if (rc == 0) { 1827 rc = efx_ptp_synchronize(efx, 1828 PTP_SYNC_ATTEMPTS * 2); 1829 if (rc != 0) 1830 efx_ptp_stop(efx); 1831 } 1832 } else { 1833 rc = efx_ptp_stop(efx); 1834 } 1835 1836 if (rc != 0) 1837 return rc; 1838 1839 efx->ptp_data->enabled = enable_wanted; 1840 } 1841 1842 return 0; 1843 } 1844 1845 static int efx_ptp_ts_init(struct efx_nic *efx, struct kernel_hwtstamp_config *init) 1846 { 1847 int rc; 1848 1849 if ((init->tx_type != HWTSTAMP_TX_OFF) && 1850 (init->tx_type != HWTSTAMP_TX_ON)) 1851 return -ERANGE; 1852 1853 rc = efx->type->ptp_set_ts_config(efx, init); 1854 if (rc) 1855 return rc; 1856 1857 efx->ptp_data->config = *init; 1858 return 0; 1859 } 1860 1861 void efx_ptp_get_ts_info(struct efx_nic *efx, struct kernel_ethtool_ts_info *ts_info) 1862 { 1863 struct efx_ptp_data *ptp = efx->ptp_data; 1864 struct efx_nic *primary = efx->primary; 1865 1866 ASSERT_RTNL(); 1867 1868 if (!ptp) 1869 return; 1870 1871 ts_info->so_timestamping |= (SOF_TIMESTAMPING_TX_HARDWARE | 1872 SOF_TIMESTAMPING_RX_HARDWARE | 1873 SOF_TIMESTAMPING_RAW_HARDWARE); 1874 /* Check licensed features. If we don't have the license for TX 1875 * timestamps, the NIC will not support them. 1876 */ 1877 if (efx_ptp_use_mac_tx_timestamps(efx)) { 1878 struct efx_ef10_nic_data *nic_data = efx->nic_data; 1879 1880 if (!(nic_data->licensed_features & 1881 (1 << LICENSED_V3_FEATURES_TX_TIMESTAMPS_LBN))) 1882 ts_info->so_timestamping &= 1883 ~SOF_TIMESTAMPING_TX_HARDWARE; 1884 } 1885 if (primary && primary->ptp_data && primary->ptp_data->phc_clock) 1886 ts_info->phc_index = 1887 ptp_clock_index(primary->ptp_data->phc_clock); 1888 ts_info->tx_types = 1 << HWTSTAMP_TX_OFF | 1 << HWTSTAMP_TX_ON; 1889 ts_info->rx_filters = ptp->efx->type->hwtstamp_filters; 1890 } 1891 1892 int efx_ptp_set_ts_config(struct efx_nic *efx, 1893 struct kernel_hwtstamp_config *config, 1894 struct netlink_ext_ack __always_unused *extack) 1895 { 1896 /* Not a PTP enabled port */ 1897 if (!efx->ptp_data) 1898 return -EOPNOTSUPP; 1899 1900 return efx_ptp_ts_init(efx, config); 1901 } 1902 1903 int efx_ptp_get_ts_config(struct efx_nic *efx, 1904 struct kernel_hwtstamp_config *config) 1905 { 1906 /* Not a PTP enabled port */ 1907 if (!efx->ptp_data) 1908 return -EOPNOTSUPP; 1909 *config = efx->ptp_data->config; 1910 return 0; 1911 } 1912 1913 static void ptp_event_failure(struct efx_nic *efx, int expected_frag_len) 1914 { 1915 struct efx_ptp_data *ptp = efx->ptp_data; 1916 1917 netif_err(efx, hw, efx->net_dev, 1918 "PTP unexpected event length: got %d expected %d\n", 1919 ptp->evt_frag_idx, expected_frag_len); 1920 ptp->reset_required = true; 1921 queue_work(ptp->workwq, &ptp->work); 1922 } 1923 1924 static void ptp_event_fault(struct efx_nic *efx, struct efx_ptp_data *ptp) 1925 { 1926 int code = EFX_QWORD_FIELD(ptp->evt_frags[0], MCDI_EVENT_DATA); 1927 if (ptp->evt_frag_idx != 1) { 1928 ptp_event_failure(efx, 1); 1929 return; 1930 } 1931 1932 netif_err(efx, hw, efx->net_dev, "PTP error %d\n", code); 1933 } 1934 1935 static void ptp_event_pps(struct efx_nic *efx, struct efx_ptp_data *ptp) 1936 { 1937 if (ptp->nic_ts_enabled) 1938 queue_work(ptp->pps_workwq, &ptp->pps_work); 1939 } 1940 1941 void efx_ptp_event(struct efx_nic *efx, efx_qword_t *ev) 1942 { 1943 struct efx_ptp_data *ptp = efx->ptp_data; 1944 int code = EFX_QWORD_FIELD(*ev, MCDI_EVENT_CODE); 1945 1946 if (!ptp) { 1947 if (!efx->ptp_warned) { 1948 netif_warn(efx, drv, efx->net_dev, 1949 "Received PTP event but PTP not set up\n"); 1950 efx->ptp_warned = true; 1951 } 1952 return; 1953 } 1954 1955 if (!ptp->enabled) 1956 return; 1957 1958 if (ptp->evt_frag_idx == 0) { 1959 ptp->evt_code = code; 1960 } else if (ptp->evt_code != code) { 1961 netif_err(efx, hw, efx->net_dev, 1962 "PTP out of sequence event %d\n", code); 1963 ptp->evt_frag_idx = 0; 1964 } 1965 1966 ptp->evt_frags[ptp->evt_frag_idx++] = *ev; 1967 if (!MCDI_EVENT_FIELD(*ev, CONT)) { 1968 /* Process resulting event */ 1969 switch (code) { 1970 case MCDI_EVENT_CODE_PTP_FAULT: 1971 ptp_event_fault(efx, ptp); 1972 break; 1973 case MCDI_EVENT_CODE_PTP_PPS: 1974 ptp_event_pps(efx, ptp); 1975 break; 1976 default: 1977 netif_err(efx, hw, efx->net_dev, 1978 "PTP unknown event %d\n", code); 1979 break; 1980 } 1981 ptp->evt_frag_idx = 0; 1982 } else if (MAX_EVENT_FRAGS == ptp->evt_frag_idx) { 1983 netif_err(efx, hw, efx->net_dev, 1984 "PTP too many event fragments\n"); 1985 ptp->evt_frag_idx = 0; 1986 } 1987 } 1988 1989 void efx_time_sync_event(struct efx_channel *channel, efx_qword_t *ev) 1990 { 1991 struct efx_nic *efx = channel->efx; 1992 struct efx_ptp_data *ptp = efx->ptp_data; 1993 1994 /* When extracting the sync timestamp minor value, we should discard 1995 * the least significant two bits. These are not required in order 1996 * to reconstruct full-range timestamps and they are optionally used 1997 * to report status depending on the options supplied when subscribing 1998 * for sync events. 1999 */ 2000 channel->sync_timestamp_major = MCDI_EVENT_FIELD(*ev, PTP_TIME_MAJOR); 2001 channel->sync_timestamp_minor = 2002 (MCDI_EVENT_FIELD(*ev, PTP_TIME_MINOR_MS_8BITS) & 0xFC) 2003 << ptp->nic_time.sync_event_minor_shift; 2004 2005 /* if sync events have been disabled then we want to silently ignore 2006 * this event, so throw away result. 2007 */ 2008 (void) cmpxchg(&channel->sync_events_state, SYNC_EVENTS_REQUESTED, 2009 SYNC_EVENTS_VALID); 2010 } 2011 2012 static inline u32 efx_rx_buf_timestamp_minor(struct efx_nic *efx, const u8 *eh) 2013 { 2014 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) 2015 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_ts_offset)); 2016 #else 2017 const u8 *data = eh + efx->rx_packet_ts_offset; 2018 return (u32)data[0] | 2019 (u32)data[1] << 8 | 2020 (u32)data[2] << 16 | 2021 (u32)data[3] << 24; 2022 #endif 2023 } 2024 2025 void __efx_rx_skb_attach_timestamp(struct efx_channel *channel, 2026 struct sk_buff *skb) 2027 { 2028 struct efx_nic *efx = channel->efx; 2029 struct efx_ptp_data *ptp = efx->ptp_data; 2030 u32 pkt_timestamp_major, pkt_timestamp_minor; 2031 u32 diff, carry; 2032 struct skb_shared_hwtstamps *timestamps; 2033 2034 if (channel->sync_events_state != SYNC_EVENTS_VALID) 2035 return; 2036 2037 pkt_timestamp_minor = efx_rx_buf_timestamp_minor(efx, skb_mac_header(skb)); 2038 2039 /* get the difference between the packet and sync timestamps, 2040 * modulo one second 2041 */ 2042 diff = pkt_timestamp_minor - channel->sync_timestamp_minor; 2043 if (pkt_timestamp_minor < channel->sync_timestamp_minor) 2044 diff += ptp->nic_time.minor_max; 2045 2046 /* do we roll over a second boundary and need to carry the one? */ 2047 carry = (channel->sync_timestamp_minor >= ptp->nic_time.minor_max - diff) ? 2048 1 : 0; 2049 2050 if (diff <= ptp->nic_time.sync_event_diff_max) { 2051 /* packet is ahead of the sync event by a quarter of a second or 2052 * less (allowing for fuzz) 2053 */ 2054 pkt_timestamp_major = channel->sync_timestamp_major + carry; 2055 } else if (diff >= ptp->nic_time.sync_event_diff_min) { 2056 /* packet is behind the sync event but within the fuzz factor. 2057 * This means the RX packet and sync event crossed as they were 2058 * placed on the event queue, which can sometimes happen. 2059 */ 2060 pkt_timestamp_major = channel->sync_timestamp_major - 1 + carry; 2061 } else { 2062 /* it's outside tolerance in both directions. this might be 2063 * indicative of us missing sync events for some reason, so 2064 * we'll call it an error rather than risk giving a bogus 2065 * timestamp. 2066 */ 2067 netif_vdbg(efx, drv, efx->net_dev, 2068 "packet timestamp %x too far from sync event %x:%x\n", 2069 pkt_timestamp_minor, channel->sync_timestamp_major, 2070 channel->sync_timestamp_minor); 2071 return; 2072 } 2073 2074 /* attach the timestamps to the skb */ 2075 timestamps = skb_hwtstamps(skb); 2076 timestamps->hwtstamp = 2077 ptp->nic_to_kernel_time(pkt_timestamp_major, 2078 pkt_timestamp_minor, 2079 ptp->ts_corrections.general_rx); 2080 } 2081 2082 static int efx_phc_adjfine(struct ptp_clock_info *ptp, long scaled_ppm) 2083 { 2084 struct efx_ptp_data *ptp_data = container_of(ptp, 2085 struct efx_ptp_data, 2086 phc_clock_info); 2087 s32 delta = scaled_ppm_to_ppb(scaled_ppm); 2088 struct efx_nic *efx = ptp_data->efx; 2089 MCDI_DECLARE_BUF(inadj, MC_CMD_PTP_IN_ADJUST_LEN); 2090 s64 adjustment_ns; 2091 int rc; 2092 2093 if (delta > MAX_PPB) 2094 delta = MAX_PPB; 2095 else if (delta < -MAX_PPB) 2096 delta = -MAX_PPB; 2097 2098 /* Convert ppb to fixed point ns taking care to round correctly. */ 2099 adjustment_ns = ((s64)delta * PPB_SCALE_WORD + 2100 (1 << (ptp_data->adjfreq_ppb_shift - 1))) >> 2101 ptp_data->adjfreq_ppb_shift; 2102 2103 MCDI_SET_DWORD(inadj, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST); 2104 MCDI_SET_DWORD(inadj, PTP_IN_PERIPH_ID, 0); 2105 MCDI_SET_QWORD(inadj, PTP_IN_ADJUST_FREQ, adjustment_ns); 2106 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_SECONDS, 0); 2107 MCDI_SET_DWORD(inadj, PTP_IN_ADJUST_NANOSECONDS, 0); 2108 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inadj, sizeof(inadj), 2109 NULL, 0, NULL); 2110 if (rc != 0) 2111 return rc; 2112 2113 ptp_data->current_adjfreq = adjustment_ns; 2114 return 0; 2115 } 2116 2117 static int efx_phc_adjtime(struct ptp_clock_info *ptp, s64 delta) 2118 { 2119 u32 nic_major, nic_minor; 2120 struct efx_ptp_data *ptp_data = container_of(ptp, 2121 struct efx_ptp_data, 2122 phc_clock_info); 2123 struct efx_nic *efx = ptp_data->efx; 2124 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_ADJUST_LEN); 2125 2126 efx->ptp_data->ns_to_nic_time(delta, &nic_major, &nic_minor); 2127 2128 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_ADJUST); 2129 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 2130 MCDI_SET_QWORD(inbuf, PTP_IN_ADJUST_FREQ, ptp_data->current_adjfreq); 2131 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MAJOR, nic_major); 2132 MCDI_SET_DWORD(inbuf, PTP_IN_ADJUST_MINOR, nic_minor); 2133 return efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 2134 NULL, 0, NULL); 2135 } 2136 2137 static int efx_phc_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts) 2138 { 2139 struct efx_ptp_data *ptp_data = container_of(ptp, 2140 struct efx_ptp_data, 2141 phc_clock_info); 2142 struct efx_nic *efx = ptp_data->efx; 2143 MCDI_DECLARE_BUF(inbuf, MC_CMD_PTP_IN_READ_NIC_TIME_LEN); 2144 MCDI_DECLARE_BUF(outbuf, MC_CMD_PTP_OUT_READ_NIC_TIME_LEN); 2145 int rc; 2146 ktime_t kt; 2147 2148 MCDI_SET_DWORD(inbuf, PTP_IN_OP, MC_CMD_PTP_OP_READ_NIC_TIME); 2149 MCDI_SET_DWORD(inbuf, PTP_IN_PERIPH_ID, 0); 2150 2151 rc = efx_mcdi_rpc(efx, MC_CMD_PTP, inbuf, sizeof(inbuf), 2152 outbuf, sizeof(outbuf), NULL); 2153 if (rc != 0) 2154 return rc; 2155 2156 kt = ptp_data->nic_to_kernel_time( 2157 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MAJOR), 2158 MCDI_DWORD(outbuf, PTP_OUT_READ_NIC_TIME_MINOR), 0); 2159 *ts = ktime_to_timespec64(kt); 2160 return 0; 2161 } 2162 2163 static int efx_phc_settime(struct ptp_clock_info *ptp, 2164 const struct timespec64 *e_ts) 2165 { 2166 /* Get the current NIC time, efx_phc_gettime. 2167 * Subtract from the desired time to get the offset 2168 * call efx_phc_adjtime with the offset 2169 */ 2170 int rc; 2171 struct timespec64 time_now; 2172 struct timespec64 delta; 2173 2174 rc = efx_phc_gettime(ptp, &time_now); 2175 if (rc != 0) 2176 return rc; 2177 2178 delta = timespec64_sub(*e_ts, time_now); 2179 2180 rc = efx_phc_adjtime(ptp, timespec64_to_ns(&delta)); 2181 if (rc != 0) 2182 return rc; 2183 2184 return 0; 2185 } 2186 2187 static int efx_phc_enable(struct ptp_clock_info *ptp, 2188 struct ptp_clock_request *request, 2189 int enable) 2190 { 2191 struct efx_ptp_data *ptp_data = container_of(ptp, 2192 struct efx_ptp_data, 2193 phc_clock_info); 2194 if (request->type != PTP_CLK_REQ_PPS) 2195 return -EOPNOTSUPP; 2196 2197 ptp_data->nic_ts_enabled = !!enable; 2198 return 0; 2199 } 2200 2201 static const struct efx_channel_type efx_ptp_channel_type = { 2202 .handle_no_channel = efx_ptp_handle_no_channel, 2203 .pre_probe = efx_ptp_probe_channel, 2204 .post_remove = efx_ptp_remove_channel, 2205 .get_name = efx_ptp_get_channel_name, 2206 .copy = efx_copy_channel, 2207 .receive_skb = efx_ptp_rx, 2208 .want_txqs = efx_ptp_want_txqs, 2209 .keep_eventq = false, 2210 }; 2211 2212 void efx_ptp_defer_probe_with_channel(struct efx_nic *efx) 2213 { 2214 /* Check whether PTP is implemented on this NIC. The DISABLE 2215 * operation will succeed if and only if it is implemented. 2216 */ 2217 if (efx_ptp_disable(efx) == 0) 2218 efx->extra_channel_type[EFX_EXTRA_CHANNEL_PTP] = 2219 &efx_ptp_channel_type; 2220 } 2221 2222 void efx_ptp_start_datapath(struct efx_nic *efx) 2223 { 2224 if (efx_ptp_restart(efx)) 2225 netif_err(efx, drv, efx->net_dev, "Failed to restart PTP.\n"); 2226 /* re-enable timestamping if it was previously enabled */ 2227 if (efx->type->ptp_set_ts_sync_events) 2228 efx->type->ptp_set_ts_sync_events(efx, true, true); 2229 } 2230 2231 void efx_ptp_stop_datapath(struct efx_nic *efx) 2232 { 2233 /* temporarily disable timestamping */ 2234 if (efx->type->ptp_set_ts_sync_events) 2235 efx->type->ptp_set_ts_sync_events(efx, false, true); 2236 efx_ptp_stop(efx); 2237 } 2238