1 /* SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause */ 2 /* 3 * Copyright(c) 2015 - 2018 Intel Corporation. 4 */ 5 6 #ifndef _HFI1_SDMA_H 7 #define _HFI1_SDMA_H 8 9 #include <linux/types.h> 10 #include <linux/list.h> 11 #include <asm/byteorder.h> 12 #include <linux/workqueue.h> 13 #include <linux/rculist.h> 14 15 #include "hfi.h" 16 #include "verbs.h" 17 #include "sdma_txreq.h" 18 19 /* Hardware limit */ 20 #define MAX_DESC 64 21 /* Hardware limit for SDMA packet size */ 22 #define MAX_SDMA_PKT_SIZE ((16 * 1024) - 1) 23 24 #define SDMA_MAP_NONE 0 25 #define SDMA_MAP_SINGLE 1 26 #define SDMA_MAP_PAGE 2 27 28 #define SDMA_AHG_VALUE_MASK 0xffff 29 #define SDMA_AHG_VALUE_SHIFT 0 30 #define SDMA_AHG_INDEX_MASK 0xf 31 #define SDMA_AHG_INDEX_SHIFT 16 32 #define SDMA_AHG_FIELD_LEN_MASK 0xf 33 #define SDMA_AHG_FIELD_LEN_SHIFT 20 34 #define SDMA_AHG_FIELD_START_MASK 0x1f 35 #define SDMA_AHG_FIELD_START_SHIFT 24 36 #define SDMA_AHG_UPDATE_ENABLE_MASK 0x1 37 #define SDMA_AHG_UPDATE_ENABLE_SHIFT 31 38 39 /* AHG modes */ 40 41 /* 42 * Be aware the ordering and values 43 * for SDMA_AHG_APPLY_UPDATE[123] 44 * are assumed in generating a skip 45 * count in submit_tx() in sdma.c 46 */ 47 #define SDMA_AHG_NO_AHG 0 48 #define SDMA_AHG_COPY 1 49 #define SDMA_AHG_APPLY_UPDATE1 2 50 #define SDMA_AHG_APPLY_UPDATE2 3 51 #define SDMA_AHG_APPLY_UPDATE3 4 52 53 /* 54 * Bits defined in the send DMA descriptor. 55 */ 56 #define SDMA_DESC0_FIRST_DESC_FLAG BIT_ULL(63) 57 #define SDMA_DESC0_LAST_DESC_FLAG BIT_ULL(62) 58 #define SDMA_DESC0_BYTE_COUNT_SHIFT 48 59 #define SDMA_DESC0_BYTE_COUNT_WIDTH 14 60 #define SDMA_DESC0_BYTE_COUNT_MASK \ 61 ((1ULL << SDMA_DESC0_BYTE_COUNT_WIDTH) - 1) 62 #define SDMA_DESC0_BYTE_COUNT_SMASK \ 63 (SDMA_DESC0_BYTE_COUNT_MASK << SDMA_DESC0_BYTE_COUNT_SHIFT) 64 #define SDMA_DESC0_PHY_ADDR_SHIFT 0 65 #define SDMA_DESC0_PHY_ADDR_WIDTH 48 66 #define SDMA_DESC0_PHY_ADDR_MASK \ 67 ((1ULL << SDMA_DESC0_PHY_ADDR_WIDTH) - 1) 68 #define SDMA_DESC0_PHY_ADDR_SMASK \ 69 (SDMA_DESC0_PHY_ADDR_MASK << SDMA_DESC0_PHY_ADDR_SHIFT) 70 71 #define SDMA_DESC1_HEADER_UPDATE1_SHIFT 32 72 #define SDMA_DESC1_HEADER_UPDATE1_WIDTH 32 73 #define SDMA_DESC1_HEADER_UPDATE1_MASK \ 74 ((1ULL << SDMA_DESC1_HEADER_UPDATE1_WIDTH) - 1) 75 #define SDMA_DESC1_HEADER_UPDATE1_SMASK \ 76 (SDMA_DESC1_HEADER_UPDATE1_MASK << SDMA_DESC1_HEADER_UPDATE1_SHIFT) 77 #define SDMA_DESC1_HEADER_MODE_SHIFT 13 78 #define SDMA_DESC1_HEADER_MODE_WIDTH 3 79 #define SDMA_DESC1_HEADER_MODE_MASK \ 80 ((1ULL << SDMA_DESC1_HEADER_MODE_WIDTH) - 1) 81 #define SDMA_DESC1_HEADER_MODE_SMASK \ 82 (SDMA_DESC1_HEADER_MODE_MASK << SDMA_DESC1_HEADER_MODE_SHIFT) 83 #define SDMA_DESC1_HEADER_INDEX_SHIFT 8 84 #define SDMA_DESC1_HEADER_INDEX_WIDTH 5 85 #define SDMA_DESC1_HEADER_INDEX_MASK \ 86 ((1ULL << SDMA_DESC1_HEADER_INDEX_WIDTH) - 1) 87 #define SDMA_DESC1_HEADER_INDEX_SMASK \ 88 (SDMA_DESC1_HEADER_INDEX_MASK << SDMA_DESC1_HEADER_INDEX_SHIFT) 89 #define SDMA_DESC1_HEADER_DWS_SHIFT 4 90 #define SDMA_DESC1_HEADER_DWS_WIDTH 4 91 #define SDMA_DESC1_HEADER_DWS_MASK \ 92 ((1ULL << SDMA_DESC1_HEADER_DWS_WIDTH) - 1) 93 #define SDMA_DESC1_HEADER_DWS_SMASK \ 94 (SDMA_DESC1_HEADER_DWS_MASK << SDMA_DESC1_HEADER_DWS_SHIFT) 95 #define SDMA_DESC1_GENERATION_SHIFT 2 96 #define SDMA_DESC1_GENERATION_WIDTH 2 97 #define SDMA_DESC1_GENERATION_MASK \ 98 ((1ULL << SDMA_DESC1_GENERATION_WIDTH) - 1) 99 #define SDMA_DESC1_GENERATION_SMASK \ 100 (SDMA_DESC1_GENERATION_MASK << SDMA_DESC1_GENERATION_SHIFT) 101 #define SDMA_DESC1_INT_REQ_FLAG BIT_ULL(1) 102 #define SDMA_DESC1_HEAD_TO_HOST_FLAG BIT_ULL(0) 103 104 enum sdma_states { 105 sdma_state_s00_hw_down, 106 sdma_state_s10_hw_start_up_halt_wait, 107 sdma_state_s15_hw_start_up_clean_wait, 108 sdma_state_s20_idle, 109 sdma_state_s30_sw_clean_up_wait, 110 sdma_state_s40_hw_clean_up_wait, 111 sdma_state_s50_hw_halt_wait, 112 sdma_state_s60_idle_halt_wait, 113 sdma_state_s80_hw_freeze, 114 sdma_state_s82_freeze_sw_clean, 115 sdma_state_s99_running, 116 }; 117 118 enum sdma_events { 119 sdma_event_e00_go_hw_down, 120 sdma_event_e10_go_hw_start, 121 sdma_event_e15_hw_halt_done, 122 sdma_event_e25_hw_clean_up_done, 123 sdma_event_e30_go_running, 124 sdma_event_e40_sw_cleaned, 125 sdma_event_e50_hw_cleaned, 126 sdma_event_e60_hw_halted, 127 sdma_event_e70_go_idle, 128 sdma_event_e80_hw_freeze, 129 sdma_event_e81_hw_frozen, 130 sdma_event_e82_hw_unfreeze, 131 sdma_event_e85_link_down, 132 sdma_event_e90_sw_halted, 133 }; 134 135 struct sdma_set_state_action { 136 unsigned op_enable:1; 137 unsigned op_intenable:1; 138 unsigned op_halt:1; 139 unsigned op_cleanup:1; 140 unsigned go_s99_running_tofalse:1; 141 unsigned go_s99_running_totrue:1; 142 }; 143 144 struct sdma_state { 145 struct kref kref; 146 struct completion comp; 147 enum sdma_states current_state; 148 unsigned current_op; 149 unsigned go_s99_running; 150 /* debugging/development */ 151 enum sdma_states previous_state; 152 unsigned previous_op; 153 enum sdma_events last_event; 154 }; 155 156 /** 157 * DOC: sdma exported routines 158 * 159 * These sdma routines fit into three categories: 160 * - The SDMA API for building and submitting packets 161 * to the ring 162 * 163 * - Initialization and tear down routines to buildup 164 * and tear down SDMA 165 * 166 * - ISR entrances to handle interrupts, state changes 167 * and errors 168 */ 169 170 /** 171 * DOC: sdma PSM/verbs API 172 * 173 * The sdma API is designed to be used by both PSM 174 * and verbs to supply packets to the SDMA ring. 175 * 176 * The usage of the API is as follows: 177 * 178 * Embed a struct iowait in the QP or 179 * PQ. The iowait should be initialized with a 180 * call to iowait_init(). 181 * 182 * The user of the API should create an allocation method 183 * for their version of the txreq. slabs, pre-allocated lists, 184 * and dma pools can be used. Once the user's overload of 185 * the sdma_txreq has been allocated, the sdma_txreq member 186 * must be initialized with sdma_txinit() or sdma_txinit_ahg(). 187 * 188 * The txreq must be declared with the sdma_txreq first. 189 * 190 * The tx request, once initialized, is manipulated with calls to 191 * sdma_txadd_daddr(), sdma_txadd_page(), or sdma_txadd_kvaddr() 192 * for each disjoint memory location. It is the user's responsibility 193 * to understand the packet boundaries and page boundaries to do the 194 * appropriate number of sdma_txadd_* calls.. The user 195 * must be prepared to deal with failures from these routines due to 196 * either memory allocation or dma_mapping failures. 197 * 198 * The mapping specifics for each memory location are recorded 199 * in the tx. Memory locations added with sdma_txadd_page() 200 * and sdma_txadd_kvaddr() are automatically mapped when added 201 * to the tx and nmapped as part of the progress processing in the 202 * SDMA interrupt handling. 203 * 204 * sdma_txadd_daddr() is used to add an dma_addr_t memory to the 205 * tx. An example of a use case would be a pre-allocated 206 * set of headers allocated via dma_pool_alloc() or 207 * dma_alloc_coherent(). For these memory locations, it 208 * is the responsibility of the user to handle that unmapping. 209 * (This would usually be at an unload or job termination.) 210 * 211 * The routine sdma_send_txreq() is used to submit 212 * a tx to the ring after the appropriate number of 213 * sdma_txadd_* have been done. 214 * 215 * If it is desired to send a burst of sdma_txreqs, sdma_send_txlist() 216 * can be used to submit a list of packets. 217 * 218 * The user is free to use the link overhead in the struct sdma_txreq as 219 * long as the tx isn't in flight. 220 * 221 * The extreme degenerate case of the number of descriptors 222 * exceeding the ring size is automatically handled as 223 * memory locations are added. An overflow of the descriptor 224 * array that is part of the sdma_txreq is also automatically 225 * handled. 226 * 227 */ 228 229 /** 230 * DOC: Infrastructure calls 231 * 232 * sdma_init() is used to initialize data structures and 233 * CSRs for the desired number of SDMA engines. 234 * 235 * sdma_start() is used to kick the SDMA engines initialized 236 * with sdma_init(). Interrupts must be enabled at this 237 * point since aspects of the state machine are interrupt 238 * driven. 239 * 240 * sdma_engine_error() and sdma_engine_interrupt() are 241 * entrances for interrupts. 242 * 243 * sdma_map_init() is for the management of the mapping 244 * table when the number of vls is changed. 245 * 246 */ 247 248 /* 249 * struct hw_sdma_desc - raw 128 bit SDMA descriptor 250 * 251 * This is the raw descriptor in the SDMA ring 252 */ 253 struct hw_sdma_desc { 254 /* private: don't use directly */ 255 __le64 qw[2]; 256 }; 257 258 /** 259 * struct sdma_engine - Data pertaining to each SDMA engine. 260 * @dd: a back-pointer to the device data 261 * @ppd: per port back-pointer 262 * @imask: mask for irq manipulation 263 * @idle_mask: mask for determining if an interrupt is due to sdma_idle 264 * 265 * This structure has the state for each sdma_engine. 266 * 267 * Accessing to non public fields are not supported 268 * since the private members are subject to change. 269 */ 270 struct sdma_engine { 271 /* read mostly */ 272 struct hfi1_devdata *dd; 273 struct hfi1_pportdata *ppd; 274 /* private: */ 275 void __iomem *tail_csr; 276 u64 imask; /* clear interrupt mask */ 277 u64 idle_mask; 278 u64 progress_mask; 279 u64 int_mask; 280 /* private: */ 281 volatile __le64 *head_dma; /* DMA'ed by chip */ 282 /* private: */ 283 dma_addr_t head_phys; 284 /* private: */ 285 struct hw_sdma_desc *descq; 286 /* private: */ 287 unsigned descq_full_count; 288 struct sdma_txreq **tx_ring; 289 /* private: */ 290 dma_addr_t descq_phys; 291 /* private */ 292 u32 sdma_mask; 293 /* private */ 294 struct sdma_state state; 295 /* private */ 296 int cpu; 297 /* private: */ 298 u8 sdma_shift; 299 /* private: */ 300 u8 this_idx; /* zero relative engine */ 301 /* protect changes to senddmactrl shadow */ 302 spinlock_t senddmactrl_lock; 303 /* private: */ 304 u64 p_senddmactrl; /* shadow per-engine SendDmaCtrl */ 305 306 /* read/write using tail_lock */ 307 spinlock_t tail_lock ____cacheline_aligned_in_smp; 308 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER 309 /* private: */ 310 u64 tail_sn; 311 #endif 312 /* private: */ 313 u32 descq_tail; 314 /* private: */ 315 unsigned long ahg_bits; 316 /* private: */ 317 u16 desc_avail; 318 /* private: */ 319 u16 tx_tail; 320 /* private: */ 321 u16 descq_cnt; 322 323 /* read/write using head_lock */ 324 /* private: */ 325 seqlock_t head_lock ____cacheline_aligned_in_smp; 326 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER 327 /* private: */ 328 u64 head_sn; 329 #endif 330 /* private: */ 331 u32 descq_head; 332 /* private: */ 333 u16 tx_head; 334 /* private: */ 335 u64 last_status; 336 /* private */ 337 u64 err_cnt; 338 /* private */ 339 u64 sdma_int_cnt; 340 u64 idle_int_cnt; 341 u64 progress_int_cnt; 342 343 /* private: */ 344 seqlock_t waitlock; 345 struct list_head dmawait; 346 347 /* CONFIG SDMA for now, just blindly duplicate */ 348 /* private: */ 349 struct tasklet_struct sdma_hw_clean_up_task 350 ____cacheline_aligned_in_smp; 351 352 /* private: */ 353 struct tasklet_struct sdma_sw_clean_up_task 354 ____cacheline_aligned_in_smp; 355 /* private: */ 356 struct work_struct err_halt_worker; 357 /* private */ 358 struct timer_list err_progress_check_timer; 359 u32 progress_check_head; 360 /* private: */ 361 struct work_struct flush_worker; 362 /* protect flush list */ 363 spinlock_t flushlist_lock; 364 /* private: */ 365 struct list_head flushlist; 366 struct cpumask cpu_mask; 367 struct kobject kobj; 368 u32 msix_intr; 369 }; 370 371 int sdma_init(struct hfi1_devdata *dd, u8 port); 372 void sdma_start(struct hfi1_devdata *dd); 373 void sdma_exit(struct hfi1_devdata *dd); 374 void sdma_clean(struct hfi1_devdata *dd, size_t num_engines); 375 void sdma_all_running(struct hfi1_devdata *dd); 376 void sdma_all_idle(struct hfi1_devdata *dd); 377 void sdma_freeze_notify(struct hfi1_devdata *dd, int go_idle); 378 void sdma_freeze(struct hfi1_devdata *dd); 379 void sdma_unfreeze(struct hfi1_devdata *dd); 380 void sdma_wait(struct hfi1_devdata *dd); 381 382 /** 383 * sdma_empty() - idle engine test 384 * @engine: sdma engine 385 * 386 * Currently used by verbs as a latency optimization. 387 * 388 * Return: 389 * 1 - empty, 0 - non-empty 390 */ 391 static inline int sdma_empty(struct sdma_engine *sde) 392 { 393 return sde->descq_tail == sde->descq_head; 394 } 395 396 static inline u16 sdma_descq_freecnt(struct sdma_engine *sde) 397 { 398 return sde->descq_cnt - 399 (sde->descq_tail - 400 READ_ONCE(sde->descq_head)) - 1; 401 } 402 403 static inline u16 sdma_descq_inprocess(struct sdma_engine *sde) 404 { 405 return sde->descq_cnt - sdma_descq_freecnt(sde); 406 } 407 408 /* 409 * Either head_lock or tail lock required to see 410 * a steady state. 411 */ 412 static inline int __sdma_running(struct sdma_engine *engine) 413 { 414 return engine->state.current_state == sdma_state_s99_running; 415 } 416 417 /** 418 * sdma_running() - state suitability test 419 * @engine: sdma engine 420 * 421 * sdma_running probes the internal state to determine if it is suitable 422 * for submitting packets. 423 * 424 * Return: 425 * 1 - ok to submit, 0 - not ok to submit 426 * 427 */ 428 static inline int sdma_running(struct sdma_engine *engine) 429 { 430 unsigned long flags; 431 int ret; 432 433 spin_lock_irqsave(&engine->tail_lock, flags); 434 ret = __sdma_running(engine); 435 spin_unlock_irqrestore(&engine->tail_lock, flags); 436 return ret; 437 } 438 439 void _sdma_txreq_ahgadd( 440 struct sdma_txreq *tx, 441 u8 num_ahg, 442 u8 ahg_entry, 443 u32 *ahg, 444 u8 ahg_hlen); 445 446 /** 447 * sdma_txinit_ahg() - initialize an sdma_txreq struct with AHG 448 * @tx: tx request to initialize 449 * @flags: flags to key last descriptor additions 450 * @tlen: total packet length (pbc + headers + data) 451 * @ahg_entry: ahg entry to use (0 - 31) 452 * @num_ahg: ahg descriptor for first descriptor (0 - 9) 453 * @ahg: array of AHG descriptors (up to 9 entries) 454 * @ahg_hlen: number of bytes from ASIC entry to use 455 * @cb: callback 456 * 457 * The allocation of the sdma_txreq and it enclosing structure is user 458 * dependent. This routine must be called to initialize the user independent 459 * fields. 460 * 461 * The currently supported flags are SDMA_TXREQ_F_URGENT, 462 * SDMA_TXREQ_F_AHG_COPY, and SDMA_TXREQ_F_USE_AHG. 463 * 464 * SDMA_TXREQ_F_URGENT is used for latency sensitive situations where the 465 * completion is desired as soon as possible. 466 * 467 * SDMA_TXREQ_F_AHG_COPY causes the header in the first descriptor to be 468 * copied to chip entry. SDMA_TXREQ_F_USE_AHG causes the code to add in 469 * the AHG descriptors into the first 1 to 3 descriptors. 470 * 471 * Completions of submitted requests can be gotten on selected 472 * txreqs by giving a completion routine callback to sdma_txinit() or 473 * sdma_txinit_ahg(). The environment in which the callback runs 474 * can be from an ISR, a tasklet, or a thread, so no sleeping 475 * kernel routines can be used. Aspects of the sdma ring may 476 * be locked so care should be taken with locking. 477 * 478 * The callback pointer can be NULL to avoid any callback for the packet 479 * being submitted. The callback will be provided this tx, a status, and a flag. 480 * 481 * The status will be one of SDMA_TXREQ_S_OK, SDMA_TXREQ_S_SENDERROR, 482 * SDMA_TXREQ_S_ABORTED, or SDMA_TXREQ_S_SHUTDOWN. 483 * 484 * The flag, if the is the iowait had been used, indicates the iowait 485 * sdma_busy count has reached zero. 486 * 487 * user data portion of tlen should be precise. The sdma_txadd_* entrances 488 * will pad with a descriptor references 1 - 3 bytes when the number of bytes 489 * specified in tlen have been supplied to the sdma_txreq. 490 * 491 * ahg_hlen is used to determine the number of on-chip entry bytes to 492 * use as the header. This is for cases where the stored header is 493 * larger than the header to be used in a packet. This is typical 494 * for verbs where an RDMA_WRITE_FIRST is larger than the packet in 495 * and RDMA_WRITE_MIDDLE. 496 * 497 */ 498 static inline int sdma_txinit_ahg( 499 struct sdma_txreq *tx, 500 u16 flags, 501 u16 tlen, 502 u8 ahg_entry, 503 u8 num_ahg, 504 u32 *ahg, 505 u8 ahg_hlen, 506 void (*cb)(struct sdma_txreq *, int)) 507 { 508 if (tlen == 0) 509 return -ENODATA; 510 if (tlen > MAX_SDMA_PKT_SIZE) 511 return -EMSGSIZE; 512 tx->desc_limit = ARRAY_SIZE(tx->descs); 513 tx->descp = &tx->descs[0]; 514 INIT_LIST_HEAD(&tx->list); 515 tx->num_desc = 0; 516 tx->flags = flags; 517 tx->complete = cb; 518 tx->coalesce_buf = NULL; 519 tx->wait = NULL; 520 tx->packet_len = tlen; 521 tx->tlen = tx->packet_len; 522 tx->descs[0].qw[0] = SDMA_DESC0_FIRST_DESC_FLAG; 523 tx->descs[0].qw[1] = 0; 524 if (flags & SDMA_TXREQ_F_AHG_COPY) 525 tx->descs[0].qw[1] |= 526 (((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK) 527 << SDMA_DESC1_HEADER_INDEX_SHIFT) | 528 (((u64)SDMA_AHG_COPY & SDMA_DESC1_HEADER_MODE_MASK) 529 << SDMA_DESC1_HEADER_MODE_SHIFT); 530 else if (flags & SDMA_TXREQ_F_USE_AHG && num_ahg) 531 _sdma_txreq_ahgadd(tx, num_ahg, ahg_entry, ahg, ahg_hlen); 532 return 0; 533 } 534 535 /** 536 * sdma_txinit() - initialize an sdma_txreq struct (no AHG) 537 * @tx: tx request to initialize 538 * @flags: flags to key last descriptor additions 539 * @tlen: total packet length (pbc + headers + data) 540 * @cb: callback pointer 541 * 542 * The allocation of the sdma_txreq and it enclosing structure is user 543 * dependent. This routine must be called to initialize the user 544 * independent fields. 545 * 546 * The currently supported flags is SDMA_TXREQ_F_URGENT. 547 * 548 * SDMA_TXREQ_F_URGENT is used for latency sensitive situations where the 549 * completion is desired as soon as possible. 550 * 551 * Completions of submitted requests can be gotten on selected 552 * txreqs by giving a completion routine callback to sdma_txinit() or 553 * sdma_txinit_ahg(). The environment in which the callback runs 554 * can be from an ISR, a tasklet, or a thread, so no sleeping 555 * kernel routines can be used. The head size of the sdma ring may 556 * be locked so care should be taken with locking. 557 * 558 * The callback pointer can be NULL to avoid any callback for the packet 559 * being submitted. 560 * 561 * The callback, if non-NULL, will be provided this tx and a status. The 562 * status will be one of SDMA_TXREQ_S_OK, SDMA_TXREQ_S_SENDERROR, 563 * SDMA_TXREQ_S_ABORTED, or SDMA_TXREQ_S_SHUTDOWN. 564 * 565 */ 566 static inline int sdma_txinit( 567 struct sdma_txreq *tx, 568 u16 flags, 569 u16 tlen, 570 void (*cb)(struct sdma_txreq *, int)) 571 { 572 return sdma_txinit_ahg(tx, flags, tlen, 0, 0, NULL, 0, cb); 573 } 574 575 /* helpers - don't use */ 576 static inline int sdma_mapping_type(struct sdma_desc *d) 577 { 578 return (d->qw[1] & SDMA_DESC1_GENERATION_SMASK) 579 >> SDMA_DESC1_GENERATION_SHIFT; 580 } 581 582 static inline size_t sdma_mapping_len(struct sdma_desc *d) 583 { 584 return (d->qw[0] & SDMA_DESC0_BYTE_COUNT_SMASK) 585 >> SDMA_DESC0_BYTE_COUNT_SHIFT; 586 } 587 588 static inline dma_addr_t sdma_mapping_addr(struct sdma_desc *d) 589 { 590 return (d->qw[0] & SDMA_DESC0_PHY_ADDR_SMASK) 591 >> SDMA_DESC0_PHY_ADDR_SHIFT; 592 } 593 594 static inline void make_tx_sdma_desc( 595 struct sdma_txreq *tx, 596 int type, 597 dma_addr_t addr, 598 size_t len) 599 { 600 struct sdma_desc *desc = &tx->descp[tx->num_desc]; 601 602 if (!tx->num_desc) { 603 /* qw[0] zero; qw[1] first, ahg mode already in from init */ 604 desc->qw[1] |= ((u64)type & SDMA_DESC1_GENERATION_MASK) 605 << SDMA_DESC1_GENERATION_SHIFT; 606 } else { 607 desc->qw[0] = 0; 608 desc->qw[1] = ((u64)type & SDMA_DESC1_GENERATION_MASK) 609 << SDMA_DESC1_GENERATION_SHIFT; 610 } 611 desc->qw[0] |= (((u64)addr & SDMA_DESC0_PHY_ADDR_MASK) 612 << SDMA_DESC0_PHY_ADDR_SHIFT) | 613 (((u64)len & SDMA_DESC0_BYTE_COUNT_MASK) 614 << SDMA_DESC0_BYTE_COUNT_SHIFT); 615 } 616 617 /* helper to extend txreq */ 618 int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx, 619 int type, void *kvaddr, struct page *page, 620 unsigned long offset, u16 len); 621 int _pad_sdma_tx_descs(struct hfi1_devdata *, struct sdma_txreq *); 622 void __sdma_txclean(struct hfi1_devdata *, struct sdma_txreq *); 623 624 static inline void sdma_txclean(struct hfi1_devdata *dd, struct sdma_txreq *tx) 625 { 626 if (tx->num_desc) 627 __sdma_txclean(dd, tx); 628 } 629 630 /* helpers used by public routines */ 631 static inline void _sdma_close_tx(struct hfi1_devdata *dd, 632 struct sdma_txreq *tx) 633 { 634 u16 last_desc = tx->num_desc - 1; 635 636 tx->descp[last_desc].qw[0] |= SDMA_DESC0_LAST_DESC_FLAG; 637 tx->descp[last_desc].qw[1] |= dd->default_desc1; 638 if (tx->flags & SDMA_TXREQ_F_URGENT) 639 tx->descp[last_desc].qw[1] |= (SDMA_DESC1_HEAD_TO_HOST_FLAG | 640 SDMA_DESC1_INT_REQ_FLAG); 641 } 642 643 static inline int _sdma_txadd_daddr( 644 struct hfi1_devdata *dd, 645 int type, 646 struct sdma_txreq *tx, 647 dma_addr_t addr, 648 u16 len) 649 { 650 int rval = 0; 651 652 make_tx_sdma_desc( 653 tx, 654 type, 655 addr, len); 656 WARN_ON(len > tx->tlen); 657 tx->num_desc++; 658 tx->tlen -= len; 659 /* special cases for last */ 660 if (!tx->tlen) { 661 if (tx->packet_len & (sizeof(u32) - 1)) { 662 rval = _pad_sdma_tx_descs(dd, tx); 663 if (rval) 664 return rval; 665 } else { 666 _sdma_close_tx(dd, tx); 667 } 668 } 669 return rval; 670 } 671 672 /** 673 * sdma_txadd_page() - add a page to the sdma_txreq 674 * @dd: the device to use for mapping 675 * @tx: tx request to which the page is added 676 * @page: page to map 677 * @offset: offset within the page 678 * @len: length in bytes 679 * 680 * This is used to add a page/offset/length descriptor. 681 * 682 * The mapping/unmapping of the page/offset/len is automatically handled. 683 * 684 * Return: 685 * 0 - success, -ENOSPC - mapping fail, -ENOMEM - couldn't 686 * extend/coalesce descriptor array 687 */ 688 static inline int sdma_txadd_page( 689 struct hfi1_devdata *dd, 690 struct sdma_txreq *tx, 691 struct page *page, 692 unsigned long offset, 693 u16 len) 694 { 695 dma_addr_t addr; 696 int rval; 697 698 if ((unlikely(tx->num_desc == tx->desc_limit))) { 699 rval = ext_coal_sdma_tx_descs(dd, tx, SDMA_MAP_PAGE, 700 NULL, page, offset, len); 701 if (rval <= 0) 702 return rval; 703 } 704 705 addr = dma_map_page( 706 &dd->pcidev->dev, 707 page, 708 offset, 709 len, 710 DMA_TO_DEVICE); 711 712 if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) { 713 __sdma_txclean(dd, tx); 714 return -ENOSPC; 715 } 716 717 return _sdma_txadd_daddr( 718 dd, SDMA_MAP_PAGE, tx, addr, len); 719 } 720 721 /** 722 * sdma_txadd_daddr() - add a dma address to the sdma_txreq 723 * @dd: the device to use for mapping 724 * @tx: sdma_txreq to which the page is added 725 * @addr: dma address mapped by caller 726 * @len: length in bytes 727 * 728 * This is used to add a descriptor for memory that is already dma mapped. 729 * 730 * In this case, there is no unmapping as part of the progress processing for 731 * this memory location. 732 * 733 * Return: 734 * 0 - success, -ENOMEM - couldn't extend descriptor array 735 */ 736 737 static inline int sdma_txadd_daddr( 738 struct hfi1_devdata *dd, 739 struct sdma_txreq *tx, 740 dma_addr_t addr, 741 u16 len) 742 { 743 int rval; 744 745 if ((unlikely(tx->num_desc == tx->desc_limit))) { 746 rval = ext_coal_sdma_tx_descs(dd, tx, SDMA_MAP_NONE, 747 NULL, NULL, 0, 0); 748 if (rval <= 0) 749 return rval; 750 } 751 752 return _sdma_txadd_daddr(dd, SDMA_MAP_NONE, tx, addr, len); 753 } 754 755 /** 756 * sdma_txadd_kvaddr() - add a kernel virtual address to sdma_txreq 757 * @dd: the device to use for mapping 758 * @tx: sdma_txreq to which the page is added 759 * @kvaddr: the kernel virtual address 760 * @len: length in bytes 761 * 762 * This is used to add a descriptor referenced by the indicated kvaddr and 763 * len. 764 * 765 * The mapping/unmapping of the kvaddr and len is automatically handled. 766 * 767 * Return: 768 * 0 - success, -ENOSPC - mapping fail, -ENOMEM - couldn't extend/coalesce 769 * descriptor array 770 */ 771 static inline int sdma_txadd_kvaddr( 772 struct hfi1_devdata *dd, 773 struct sdma_txreq *tx, 774 void *kvaddr, 775 u16 len) 776 { 777 dma_addr_t addr; 778 int rval; 779 780 if ((unlikely(tx->num_desc == tx->desc_limit))) { 781 rval = ext_coal_sdma_tx_descs(dd, tx, SDMA_MAP_SINGLE, 782 kvaddr, NULL, 0, len); 783 if (rval <= 0) 784 return rval; 785 } 786 787 addr = dma_map_single( 788 &dd->pcidev->dev, 789 kvaddr, 790 len, 791 DMA_TO_DEVICE); 792 793 if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) { 794 __sdma_txclean(dd, tx); 795 return -ENOSPC; 796 } 797 798 return _sdma_txadd_daddr( 799 dd, SDMA_MAP_SINGLE, tx, addr, len); 800 } 801 802 struct iowait_work; 803 804 int sdma_send_txreq(struct sdma_engine *sde, 805 struct iowait_work *wait, 806 struct sdma_txreq *tx, 807 bool pkts_sent); 808 int sdma_send_txlist(struct sdma_engine *sde, 809 struct iowait_work *wait, 810 struct list_head *tx_list, 811 u16 *count_out); 812 813 int sdma_ahg_alloc(struct sdma_engine *sde); 814 void sdma_ahg_free(struct sdma_engine *sde, int ahg_index); 815 816 /** 817 * sdma_build_ahg - build ahg descriptor 818 * @data 819 * @dwindex 820 * @startbit 821 * @bits 822 * 823 * Build and return a 32 bit descriptor. 824 */ 825 static inline u32 sdma_build_ahg_descriptor( 826 u16 data, 827 u8 dwindex, 828 u8 startbit, 829 u8 bits) 830 { 831 return (u32)(1UL << SDMA_AHG_UPDATE_ENABLE_SHIFT | 832 ((startbit & SDMA_AHG_FIELD_START_MASK) << 833 SDMA_AHG_FIELD_START_SHIFT) | 834 ((bits & SDMA_AHG_FIELD_LEN_MASK) << 835 SDMA_AHG_FIELD_LEN_SHIFT) | 836 ((dwindex & SDMA_AHG_INDEX_MASK) << 837 SDMA_AHG_INDEX_SHIFT) | 838 ((data & SDMA_AHG_VALUE_MASK) << 839 SDMA_AHG_VALUE_SHIFT)); 840 } 841 842 /** 843 * sdma_progress - use seq number of detect head progress 844 * @sde: sdma_engine to check 845 * @seq: base seq count 846 * @tx: txreq for which we need to check descriptor availability 847 * 848 * This is used in the appropriate spot in the sleep routine 849 * to check for potential ring progress. This routine gets the 850 * seqcount before queuing the iowait structure for progress. 851 * 852 * If the seqcount indicates that progress needs to be checked, 853 * re-submission is detected by checking whether the descriptor 854 * queue has enough descriptor for the txreq. 855 */ 856 static inline unsigned sdma_progress(struct sdma_engine *sde, unsigned seq, 857 struct sdma_txreq *tx) 858 { 859 if (read_seqretry(&sde->head_lock, seq)) { 860 sde->desc_avail = sdma_descq_freecnt(sde); 861 if (tx->num_desc > sde->desc_avail) 862 return 0; 863 return 1; 864 } 865 return 0; 866 } 867 868 /* for use by interrupt handling */ 869 void sdma_engine_error(struct sdma_engine *sde, u64 status); 870 void sdma_engine_interrupt(struct sdma_engine *sde, u64 status); 871 872 /* 873 * 874 * The diagram below details the relationship of the mapping structures 875 * 876 * Since the mapping now allows for non-uniform engines per vl, the 877 * number of engines for a vl is either the vl_engines[vl] or 878 * a computation based on num_sdma/num_vls: 879 * 880 * For example: 881 * nactual = vl_engines ? vl_engines[vl] : num_sdma/num_vls 882 * 883 * n = roundup to next highest power of 2 using nactual 884 * 885 * In the case where there are num_sdma/num_vls doesn't divide 886 * evenly, the extras are added from the last vl downward. 887 * 888 * For the case where n > nactual, the engines are assigned 889 * in a round robin fashion wrapping back to the first engine 890 * for a particular vl. 891 * 892 * dd->sdma_map 893 * | sdma_map_elem[0] 894 * | +--------------------+ 895 * v | mask | 896 * sdma_vl_map |--------------------| 897 * +--------------------------+ | sde[0] -> eng 1 | 898 * | list (RCU) | |--------------------| 899 * |--------------------------| ->| sde[1] -> eng 2 | 900 * | mask | --/ |--------------------| 901 * |--------------------------| -/ | * | 902 * | actual_vls (max 8) | -/ |--------------------| 903 * |--------------------------| --/ | sde[n-1] -> eng n | 904 * | vls (max 8) | -/ +--------------------+ 905 * |--------------------------| --/ 906 * | map[0] |-/ 907 * |--------------------------| +---------------------+ 908 * | map[1] |--- | mask | 909 * |--------------------------| \---- |---------------------| 910 * | * | \-- | sde[0] -> eng 1+n | 911 * | * | \---- |---------------------| 912 * | * | \->| sde[1] -> eng 2+n | 913 * |--------------------------| |---------------------| 914 * | map[vls - 1] |- | * | 915 * +--------------------------+ \- |---------------------| 916 * \- | sde[m-1] -> eng m+n | 917 * \ +---------------------+ 918 * \- 919 * \ 920 * \- +----------------------+ 921 * \- | mask | 922 * \ |----------------------| 923 * \- | sde[0] -> eng 1+m+n | 924 * \- |----------------------| 925 * >| sde[1] -> eng 2+m+n | 926 * |----------------------| 927 * | * | 928 * |----------------------| 929 * | sde[o-1] -> eng o+m+n| 930 * +----------------------+ 931 * 932 */ 933 934 /** 935 * struct sdma_map_elem - mapping for a vl 936 * @mask - selector mask 937 * @sde - array of engines for this vl 938 * 939 * The mask is used to "mod" the selector 940 * to produce index into the trailing 941 * array of sdes. 942 */ 943 struct sdma_map_elem { 944 u32 mask; 945 struct sdma_engine *sde[]; 946 }; 947 948 /** 949 * struct sdma_map_el - mapping for a vl 950 * @engine_to_vl - map of an engine to a vl 951 * @list - rcu head for free callback 952 * @mask - vl mask to "mod" the vl to produce an index to map array 953 * @actual_vls - number of vls 954 * @vls - number of vls rounded to next power of 2 955 * @map - array of sdma_map_elem entries 956 * 957 * This is the parent mapping structure. The trailing 958 * members of the struct point to sdma_map_elem entries, which 959 * in turn point to an array of sde's for that vl. 960 */ 961 struct sdma_vl_map { 962 s8 engine_to_vl[TXE_NUM_SDMA_ENGINES]; 963 struct rcu_head list; 964 u32 mask; 965 u8 actual_vls; 966 u8 vls; 967 struct sdma_map_elem *map[]; 968 }; 969 970 int sdma_map_init( 971 struct hfi1_devdata *dd, 972 u8 port, 973 u8 num_vls, 974 u8 *vl_engines); 975 976 /* slow path */ 977 void _sdma_engine_progress_schedule(struct sdma_engine *sde); 978 979 /** 980 * sdma_engine_progress_schedule() - schedule progress on engine 981 * @sde: sdma_engine to schedule progress 982 * 983 * This is the fast path. 984 * 985 */ 986 static inline void sdma_engine_progress_schedule( 987 struct sdma_engine *sde) 988 { 989 if (!sde || sdma_descq_inprocess(sde) < (sde->descq_cnt / 8)) 990 return; 991 _sdma_engine_progress_schedule(sde); 992 } 993 994 struct sdma_engine *sdma_select_engine_sc( 995 struct hfi1_devdata *dd, 996 u32 selector, 997 u8 sc5); 998 999 struct sdma_engine *sdma_select_engine_vl( 1000 struct hfi1_devdata *dd, 1001 u32 selector, 1002 u8 vl); 1003 1004 struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd, 1005 u32 selector, u8 vl); 1006 ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf); 1007 ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf, 1008 size_t count); 1009 int sdma_engine_get_vl(struct sdma_engine *sde); 1010 void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *); 1011 void sdma_seqfile_dump_cpu_list(struct seq_file *s, struct hfi1_devdata *dd, 1012 unsigned long cpuid); 1013 1014 #ifdef CONFIG_SDMA_VERBOSITY 1015 void sdma_dumpstate(struct sdma_engine *); 1016 #endif 1017 static inline char *slashstrip(char *s) 1018 { 1019 char *r = s; 1020 1021 while (*s) 1022 if (*s++ == '/') 1023 r = s; 1024 return r; 1025 } 1026 1027 u16 sdma_get_descq_cnt(void); 1028 1029 extern uint mod_num_sdma; 1030 1031 void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid); 1032 1033 #endif 1034