1 /* 2 * Copyright(c) 2016 - 2019 Intel Corporation. 3 * 4 * This file is provided under a dual BSD/GPLv2 license. When using or 5 * redistributing this file, you may do so under either license. 6 * 7 * GPL LICENSE SUMMARY 8 * 9 * This program is free software; you can redistribute it and/or modify 10 * it under the terms of version 2 of the GNU General Public License as 11 * published by the Free Software Foundation. 12 * 13 * This program is distributed in the hope that it will be useful, but 14 * WITHOUT ANY WARRANTY; without even the implied warranty of 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 16 * General Public License for more details. 17 * 18 * BSD LICENSE 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * - Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * - Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in 28 * the documentation and/or other materials provided with the 29 * distribution. 30 * - Neither the name of Intel Corporation nor the names of its 31 * contributors may be used to endorse or promote products derived 32 * from this software without specific prior written permission. 33 * 34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 45 * 46 */ 47 48 #include <linux/hash.h> 49 #include <linux/bitops.h> 50 #include <linux/lockdep.h> 51 #include <linux/vmalloc.h> 52 #include <linux/slab.h> 53 #include <rdma/ib_verbs.h> 54 #include <rdma/ib_hdrs.h> 55 #include <rdma/opa_addr.h> 56 #include <rdma/uverbs_ioctl.h> 57 #include "qp.h" 58 #include "vt.h" 59 #include "trace.h" 60 61 #define RVT_RWQ_COUNT_THRESHOLD 16 62 63 static void rvt_rc_timeout(struct timer_list *t); 64 65 /* 66 * Convert the AETH RNR timeout code into the number of microseconds. 67 */ 68 static const u32 ib_rvt_rnr_table[32] = { 69 655360, /* 00: 655.36 */ 70 10, /* 01: .01 */ 71 20, /* 02 .02 */ 72 30, /* 03: .03 */ 73 40, /* 04: .04 */ 74 60, /* 05: .06 */ 75 80, /* 06: .08 */ 76 120, /* 07: .12 */ 77 160, /* 08: .16 */ 78 240, /* 09: .24 */ 79 320, /* 0A: .32 */ 80 480, /* 0B: .48 */ 81 640, /* 0C: .64 */ 82 960, /* 0D: .96 */ 83 1280, /* 0E: 1.28 */ 84 1920, /* 0F: 1.92 */ 85 2560, /* 10: 2.56 */ 86 3840, /* 11: 3.84 */ 87 5120, /* 12: 5.12 */ 88 7680, /* 13: 7.68 */ 89 10240, /* 14: 10.24 */ 90 15360, /* 15: 15.36 */ 91 20480, /* 16: 20.48 */ 92 30720, /* 17: 30.72 */ 93 40960, /* 18: 40.96 */ 94 61440, /* 19: 61.44 */ 95 81920, /* 1A: 81.92 */ 96 122880, /* 1B: 122.88 */ 97 163840, /* 1C: 163.84 */ 98 245760, /* 1D: 245.76 */ 99 327680, /* 1E: 327.68 */ 100 491520 /* 1F: 491.52 */ 101 }; 102 103 /* 104 * Note that it is OK to post send work requests in the SQE and ERR 105 * states; rvt_do_send() will process them and generate error 106 * completions as per IB 1.2 C10-96. 107 */ 108 const int ib_rvt_state_ops[IB_QPS_ERR + 1] = { 109 [IB_QPS_RESET] = 0, 110 [IB_QPS_INIT] = RVT_POST_RECV_OK, 111 [IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK, 112 [IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | 113 RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK | 114 RVT_PROCESS_NEXT_SEND_OK, 115 [IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | 116 RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK, 117 [IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | 118 RVT_POST_SEND_OK | RVT_FLUSH_SEND, 119 [IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV | 120 RVT_POST_SEND_OK | RVT_FLUSH_SEND, 121 }; 122 EXPORT_SYMBOL(ib_rvt_state_ops); 123 124 /* platform specific: return the last level cache (llc) size, in KiB */ 125 static int rvt_wss_llc_size(void) 126 { 127 /* assume that the boot CPU value is universal for all CPUs */ 128 return boot_cpu_data.x86_cache_size; 129 } 130 131 /* platform specific: cacheless copy */ 132 static void cacheless_memcpy(void *dst, void *src, size_t n) 133 { 134 /* 135 * Use the only available X64 cacheless copy. Add a __user cast 136 * to quiet sparse. The src agument is already in the kernel so 137 * there are no security issues. The extra fault recovery machinery 138 * is not invoked. 139 */ 140 __copy_user_nocache(dst, (void __user *)src, n, 0); 141 } 142 143 void rvt_wss_exit(struct rvt_dev_info *rdi) 144 { 145 struct rvt_wss *wss = rdi->wss; 146 147 if (!wss) 148 return; 149 150 /* coded to handle partially initialized and repeat callers */ 151 kfree(wss->entries); 152 wss->entries = NULL; 153 kfree(rdi->wss); 154 rdi->wss = NULL; 155 } 156 157 /** 158 * rvt_wss_init - Init wss data structures 159 * 160 * Return: 0 on success 161 */ 162 int rvt_wss_init(struct rvt_dev_info *rdi) 163 { 164 unsigned int sge_copy_mode = rdi->dparms.sge_copy_mode; 165 unsigned int wss_threshold = rdi->dparms.wss_threshold; 166 unsigned int wss_clean_period = rdi->dparms.wss_clean_period; 167 long llc_size; 168 long llc_bits; 169 long table_size; 170 long table_bits; 171 struct rvt_wss *wss; 172 int node = rdi->dparms.node; 173 174 if (sge_copy_mode != RVT_SGE_COPY_ADAPTIVE) { 175 rdi->wss = NULL; 176 return 0; 177 } 178 179 rdi->wss = kzalloc_node(sizeof(*rdi->wss), GFP_KERNEL, node); 180 if (!rdi->wss) 181 return -ENOMEM; 182 wss = rdi->wss; 183 184 /* check for a valid percent range - default to 80 if none or invalid */ 185 if (wss_threshold < 1 || wss_threshold > 100) 186 wss_threshold = 80; 187 188 /* reject a wildly large period */ 189 if (wss_clean_period > 1000000) 190 wss_clean_period = 256; 191 192 /* reject a zero period */ 193 if (wss_clean_period == 0) 194 wss_clean_period = 1; 195 196 /* 197 * Calculate the table size - the next power of 2 larger than the 198 * LLC size. LLC size is in KiB. 199 */ 200 llc_size = rvt_wss_llc_size() * 1024; 201 table_size = roundup_pow_of_two(llc_size); 202 203 /* one bit per page in rounded up table */ 204 llc_bits = llc_size / PAGE_SIZE; 205 table_bits = table_size / PAGE_SIZE; 206 wss->pages_mask = table_bits - 1; 207 wss->num_entries = table_bits / BITS_PER_LONG; 208 209 wss->threshold = (llc_bits * wss_threshold) / 100; 210 if (wss->threshold == 0) 211 wss->threshold = 1; 212 213 wss->clean_period = wss_clean_period; 214 atomic_set(&wss->clean_counter, wss_clean_period); 215 216 wss->entries = kcalloc_node(wss->num_entries, sizeof(*wss->entries), 217 GFP_KERNEL, node); 218 if (!wss->entries) { 219 rvt_wss_exit(rdi); 220 return -ENOMEM; 221 } 222 223 return 0; 224 } 225 226 /* 227 * Advance the clean counter. When the clean period has expired, 228 * clean an entry. 229 * 230 * This is implemented in atomics to avoid locking. Because multiple 231 * variables are involved, it can be racy which can lead to slightly 232 * inaccurate information. Since this is only a heuristic, this is 233 * OK. Any innaccuracies will clean themselves out as the counter 234 * advances. That said, it is unlikely the entry clean operation will 235 * race - the next possible racer will not start until the next clean 236 * period. 237 * 238 * The clean counter is implemented as a decrement to zero. When zero 239 * is reached an entry is cleaned. 240 */ 241 static void wss_advance_clean_counter(struct rvt_wss *wss) 242 { 243 int entry; 244 int weight; 245 unsigned long bits; 246 247 /* become the cleaner if we decrement the counter to zero */ 248 if (atomic_dec_and_test(&wss->clean_counter)) { 249 /* 250 * Set, not add, the clean period. This avoids an issue 251 * where the counter could decrement below the clean period. 252 * Doing a set can result in lost decrements, slowing the 253 * clean advance. Since this a heuristic, this possible 254 * slowdown is OK. 255 * 256 * An alternative is to loop, advancing the counter by a 257 * clean period until the result is > 0. However, this could 258 * lead to several threads keeping another in the clean loop. 259 * This could be mitigated by limiting the number of times 260 * we stay in the loop. 261 */ 262 atomic_set(&wss->clean_counter, wss->clean_period); 263 264 /* 265 * Uniquely grab the entry to clean and move to next. 266 * The current entry is always the lower bits of 267 * wss.clean_entry. The table size, wss.num_entries, 268 * is always a power-of-2. 269 */ 270 entry = (atomic_inc_return(&wss->clean_entry) - 1) 271 & (wss->num_entries - 1); 272 273 /* clear the entry and count the bits */ 274 bits = xchg(&wss->entries[entry], 0); 275 weight = hweight64((u64)bits); 276 /* only adjust the contended total count if needed */ 277 if (weight) 278 atomic_sub(weight, &wss->total_count); 279 } 280 } 281 282 /* 283 * Insert the given address into the working set array. 284 */ 285 static void wss_insert(struct rvt_wss *wss, void *address) 286 { 287 u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss->pages_mask; 288 u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */ 289 u32 nr = page & (BITS_PER_LONG - 1); 290 291 if (!test_and_set_bit(nr, &wss->entries[entry])) 292 atomic_inc(&wss->total_count); 293 294 wss_advance_clean_counter(wss); 295 } 296 297 /* 298 * Is the working set larger than the threshold? 299 */ 300 static inline bool wss_exceeds_threshold(struct rvt_wss *wss) 301 { 302 return atomic_read(&wss->total_count) >= wss->threshold; 303 } 304 305 static void get_map_page(struct rvt_qpn_table *qpt, 306 struct rvt_qpn_map *map) 307 { 308 unsigned long page = get_zeroed_page(GFP_KERNEL); 309 310 /* 311 * Free the page if someone raced with us installing it. 312 */ 313 314 spin_lock(&qpt->lock); 315 if (map->page) 316 free_page(page); 317 else 318 map->page = (void *)page; 319 spin_unlock(&qpt->lock); 320 } 321 322 /** 323 * init_qpn_table - initialize the QP number table for a device 324 * @qpt: the QPN table 325 */ 326 static int init_qpn_table(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt) 327 { 328 u32 offset, i; 329 struct rvt_qpn_map *map; 330 int ret = 0; 331 332 if (!(rdi->dparms.qpn_res_end >= rdi->dparms.qpn_res_start)) 333 return -EINVAL; 334 335 spin_lock_init(&qpt->lock); 336 337 qpt->last = rdi->dparms.qpn_start; 338 qpt->incr = rdi->dparms.qpn_inc << rdi->dparms.qos_shift; 339 340 /* 341 * Drivers may want some QPs beyond what we need for verbs let them use 342 * our qpn table. No need for two. Lets go ahead and mark the bitmaps 343 * for those. The reserved range must be *after* the range which verbs 344 * will pick from. 345 */ 346 347 /* Figure out number of bit maps needed before reserved range */ 348 qpt->nmaps = rdi->dparms.qpn_res_start / RVT_BITS_PER_PAGE; 349 350 /* This should always be zero */ 351 offset = rdi->dparms.qpn_res_start & RVT_BITS_PER_PAGE_MASK; 352 353 /* Starting with the first reserved bit map */ 354 map = &qpt->map[qpt->nmaps]; 355 356 rvt_pr_info(rdi, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n", 357 rdi->dparms.qpn_res_start, rdi->dparms.qpn_res_end); 358 for (i = rdi->dparms.qpn_res_start; i <= rdi->dparms.qpn_res_end; i++) { 359 if (!map->page) { 360 get_map_page(qpt, map); 361 if (!map->page) { 362 ret = -ENOMEM; 363 break; 364 } 365 } 366 set_bit(offset, map->page); 367 offset++; 368 if (offset == RVT_BITS_PER_PAGE) { 369 /* next page */ 370 qpt->nmaps++; 371 map++; 372 offset = 0; 373 } 374 } 375 return ret; 376 } 377 378 /** 379 * free_qpn_table - free the QP number table for a device 380 * @qpt: the QPN table 381 */ 382 static void free_qpn_table(struct rvt_qpn_table *qpt) 383 { 384 int i; 385 386 for (i = 0; i < ARRAY_SIZE(qpt->map); i++) 387 free_page((unsigned long)qpt->map[i].page); 388 } 389 390 /** 391 * rvt_driver_qp_init - Init driver qp resources 392 * @rdi: rvt dev strucutre 393 * 394 * Return: 0 on success 395 */ 396 int rvt_driver_qp_init(struct rvt_dev_info *rdi) 397 { 398 int i; 399 int ret = -ENOMEM; 400 401 if (!rdi->dparms.qp_table_size) 402 return -EINVAL; 403 404 /* 405 * If driver is not doing any QP allocation then make sure it is 406 * providing the necessary QP functions. 407 */ 408 if (!rdi->driver_f.free_all_qps || 409 !rdi->driver_f.qp_priv_alloc || 410 !rdi->driver_f.qp_priv_free || 411 !rdi->driver_f.notify_qp_reset || 412 !rdi->driver_f.notify_restart_rc) 413 return -EINVAL; 414 415 /* allocate parent object */ 416 rdi->qp_dev = kzalloc_node(sizeof(*rdi->qp_dev), GFP_KERNEL, 417 rdi->dparms.node); 418 if (!rdi->qp_dev) 419 return -ENOMEM; 420 421 /* allocate hash table */ 422 rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size; 423 rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size); 424 rdi->qp_dev->qp_table = 425 kmalloc_array_node(rdi->qp_dev->qp_table_size, 426 sizeof(*rdi->qp_dev->qp_table), 427 GFP_KERNEL, rdi->dparms.node); 428 if (!rdi->qp_dev->qp_table) 429 goto no_qp_table; 430 431 for (i = 0; i < rdi->qp_dev->qp_table_size; i++) 432 RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL); 433 434 spin_lock_init(&rdi->qp_dev->qpt_lock); 435 436 /* initialize qpn map */ 437 if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table)) 438 goto fail_table; 439 440 spin_lock_init(&rdi->n_qps_lock); 441 442 return 0; 443 444 fail_table: 445 kfree(rdi->qp_dev->qp_table); 446 free_qpn_table(&rdi->qp_dev->qpn_table); 447 448 no_qp_table: 449 kfree(rdi->qp_dev); 450 451 return ret; 452 } 453 454 /** 455 * free_all_qps - check for QPs still in use 456 * @rdi: rvt device info structure 457 * 458 * There should not be any QPs still in use. 459 * Free memory for table. 460 */ 461 static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi) 462 { 463 unsigned long flags; 464 struct rvt_qp *qp; 465 unsigned n, qp_inuse = 0; 466 spinlock_t *ql; /* work around too long line below */ 467 468 if (rdi->driver_f.free_all_qps) 469 qp_inuse = rdi->driver_f.free_all_qps(rdi); 470 471 qp_inuse += rvt_mcast_tree_empty(rdi); 472 473 if (!rdi->qp_dev) 474 return qp_inuse; 475 476 ql = &rdi->qp_dev->qpt_lock; 477 spin_lock_irqsave(ql, flags); 478 for (n = 0; n < rdi->qp_dev->qp_table_size; n++) { 479 qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n], 480 lockdep_is_held(ql)); 481 RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL); 482 483 for (; qp; qp = rcu_dereference_protected(qp->next, 484 lockdep_is_held(ql))) 485 qp_inuse++; 486 } 487 spin_unlock_irqrestore(ql, flags); 488 synchronize_rcu(); 489 return qp_inuse; 490 } 491 492 /** 493 * rvt_qp_exit - clean up qps on device exit 494 * @rdi: rvt dev structure 495 * 496 * Check for qp leaks and free resources. 497 */ 498 void rvt_qp_exit(struct rvt_dev_info *rdi) 499 { 500 u32 qps_inuse = rvt_free_all_qps(rdi); 501 502 if (qps_inuse) 503 rvt_pr_err(rdi, "QP memory leak! %u still in use\n", 504 qps_inuse); 505 if (!rdi->qp_dev) 506 return; 507 508 kfree(rdi->qp_dev->qp_table); 509 free_qpn_table(&rdi->qp_dev->qpn_table); 510 kfree(rdi->qp_dev); 511 } 512 513 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt, 514 struct rvt_qpn_map *map, unsigned off) 515 { 516 return (map - qpt->map) * RVT_BITS_PER_PAGE + off; 517 } 518 519 /** 520 * alloc_qpn - Allocate the next available qpn or zero/one for QP type 521 * IB_QPT_SMI/IB_QPT_GSI 522 * @rdi: rvt device info structure 523 * @qpt: queue pair number table pointer 524 * @port_num: IB port number, 1 based, comes from core 525 * 526 * Return: The queue pair number 527 */ 528 static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt, 529 enum ib_qp_type type, u8 port_num) 530 { 531 u32 i, offset, max_scan, qpn; 532 struct rvt_qpn_map *map; 533 u32 ret; 534 535 if (rdi->driver_f.alloc_qpn) 536 return rdi->driver_f.alloc_qpn(rdi, qpt, type, port_num); 537 538 if (type == IB_QPT_SMI || type == IB_QPT_GSI) { 539 unsigned n; 540 541 ret = type == IB_QPT_GSI; 542 n = 1 << (ret + 2 * (port_num - 1)); 543 spin_lock(&qpt->lock); 544 if (qpt->flags & n) 545 ret = -EINVAL; 546 else 547 qpt->flags |= n; 548 spin_unlock(&qpt->lock); 549 goto bail; 550 } 551 552 qpn = qpt->last + qpt->incr; 553 if (qpn >= RVT_QPN_MAX) 554 qpn = qpt->incr | ((qpt->last & 1) ^ 1); 555 /* offset carries bit 0 */ 556 offset = qpn & RVT_BITS_PER_PAGE_MASK; 557 map = &qpt->map[qpn / RVT_BITS_PER_PAGE]; 558 max_scan = qpt->nmaps - !offset; 559 for (i = 0;;) { 560 if (unlikely(!map->page)) { 561 get_map_page(qpt, map); 562 if (unlikely(!map->page)) 563 break; 564 } 565 do { 566 if (!test_and_set_bit(offset, map->page)) { 567 qpt->last = qpn; 568 ret = qpn; 569 goto bail; 570 } 571 offset += qpt->incr; 572 /* 573 * This qpn might be bogus if offset >= BITS_PER_PAGE. 574 * That is OK. It gets re-assigned below 575 */ 576 qpn = mk_qpn(qpt, map, offset); 577 } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX); 578 /* 579 * In order to keep the number of pages allocated to a 580 * minimum, we scan the all existing pages before increasing 581 * the size of the bitmap table. 582 */ 583 if (++i > max_scan) { 584 if (qpt->nmaps == RVT_QPNMAP_ENTRIES) 585 break; 586 map = &qpt->map[qpt->nmaps++]; 587 /* start at incr with current bit 0 */ 588 offset = qpt->incr | (offset & 1); 589 } else if (map < &qpt->map[qpt->nmaps]) { 590 ++map; 591 /* start at incr with current bit 0 */ 592 offset = qpt->incr | (offset & 1); 593 } else { 594 map = &qpt->map[0]; 595 /* wrap to first map page, invert bit 0 */ 596 offset = qpt->incr | ((offset & 1) ^ 1); 597 } 598 /* there can be no set bits in low-order QoS bits */ 599 WARN_ON(rdi->dparms.qos_shift > 1 && 600 offset & ((BIT(rdi->dparms.qos_shift - 1) - 1) << 1)); 601 qpn = mk_qpn(qpt, map, offset); 602 } 603 604 ret = -ENOMEM; 605 606 bail: 607 return ret; 608 } 609 610 /** 611 * rvt_clear_mr_refs - Drop help mr refs 612 * @qp: rvt qp data structure 613 * @clr_sends: If shoudl clear send side or not 614 */ 615 static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends) 616 { 617 unsigned n; 618 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 619 620 if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) 621 rvt_put_ss(&qp->s_rdma_read_sge); 622 623 rvt_put_ss(&qp->r_sge); 624 625 if (clr_sends) { 626 while (qp->s_last != qp->s_head) { 627 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last); 628 629 rvt_put_qp_swqe(qp, wqe); 630 if (++qp->s_last >= qp->s_size) 631 qp->s_last = 0; 632 smp_wmb(); /* see qp_set_savail */ 633 } 634 if (qp->s_rdma_mr) { 635 rvt_put_mr(qp->s_rdma_mr); 636 qp->s_rdma_mr = NULL; 637 } 638 } 639 640 for (n = 0; qp->s_ack_queue && n < rvt_max_atomic(rdi); n++) { 641 struct rvt_ack_entry *e = &qp->s_ack_queue[n]; 642 643 if (e->rdma_sge.mr) { 644 rvt_put_mr(e->rdma_sge.mr); 645 e->rdma_sge.mr = NULL; 646 } 647 } 648 } 649 650 /** 651 * rvt_swqe_has_lkey - return true if lkey is used by swqe 652 * @wqe - the send wqe 653 * @lkey - the lkey 654 * 655 * Test the swqe for using lkey 656 */ 657 static bool rvt_swqe_has_lkey(struct rvt_swqe *wqe, u32 lkey) 658 { 659 int i; 660 661 for (i = 0; i < wqe->wr.num_sge; i++) { 662 struct rvt_sge *sge = &wqe->sg_list[i]; 663 664 if (rvt_mr_has_lkey(sge->mr, lkey)) 665 return true; 666 } 667 return false; 668 } 669 670 /** 671 * rvt_qp_sends_has_lkey - return true is qp sends use lkey 672 * @qp - the rvt_qp 673 * @lkey - the lkey 674 */ 675 static bool rvt_qp_sends_has_lkey(struct rvt_qp *qp, u32 lkey) 676 { 677 u32 s_last = qp->s_last; 678 679 while (s_last != qp->s_head) { 680 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, s_last); 681 682 if (rvt_swqe_has_lkey(wqe, lkey)) 683 return true; 684 685 if (++s_last >= qp->s_size) 686 s_last = 0; 687 } 688 if (qp->s_rdma_mr) 689 if (rvt_mr_has_lkey(qp->s_rdma_mr, lkey)) 690 return true; 691 return false; 692 } 693 694 /** 695 * rvt_qp_acks_has_lkey - return true if acks have lkey 696 * @qp - the qp 697 * @lkey - the lkey 698 */ 699 static bool rvt_qp_acks_has_lkey(struct rvt_qp *qp, u32 lkey) 700 { 701 int i; 702 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 703 704 for (i = 0; qp->s_ack_queue && i < rvt_max_atomic(rdi); i++) { 705 struct rvt_ack_entry *e = &qp->s_ack_queue[i]; 706 707 if (rvt_mr_has_lkey(e->rdma_sge.mr, lkey)) 708 return true; 709 } 710 return false; 711 } 712 713 /* 714 * rvt_qp_mr_clean - clean up remote ops for lkey 715 * @qp - the qp 716 * @lkey - the lkey that is being de-registered 717 * 718 * This routine checks if the lkey is being used by 719 * the qp. 720 * 721 * If so, the qp is put into an error state to elminate 722 * any references from the qp. 723 */ 724 void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey) 725 { 726 bool lastwqe = false; 727 728 if (qp->ibqp.qp_type == IB_QPT_SMI || 729 qp->ibqp.qp_type == IB_QPT_GSI) 730 /* avoid special QPs */ 731 return; 732 spin_lock_irq(&qp->r_lock); 733 spin_lock(&qp->s_hlock); 734 spin_lock(&qp->s_lock); 735 736 if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET) 737 goto check_lwqe; 738 739 if (rvt_ss_has_lkey(&qp->r_sge, lkey) || 740 rvt_qp_sends_has_lkey(qp, lkey) || 741 rvt_qp_acks_has_lkey(qp, lkey)) 742 lastwqe = rvt_error_qp(qp, IB_WC_LOC_PROT_ERR); 743 check_lwqe: 744 spin_unlock(&qp->s_lock); 745 spin_unlock(&qp->s_hlock); 746 spin_unlock_irq(&qp->r_lock); 747 if (lastwqe) { 748 struct ib_event ev; 749 750 ev.device = qp->ibqp.device; 751 ev.element.qp = &qp->ibqp; 752 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 753 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 754 } 755 } 756 757 /** 758 * rvt_remove_qp - remove qp form table 759 * @rdi: rvt dev struct 760 * @qp: qp to remove 761 * 762 * Remove the QP from the table so it can't be found asynchronously by 763 * the receive routine. 764 */ 765 static void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) 766 { 767 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; 768 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); 769 unsigned long flags; 770 int removed = 1; 771 772 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); 773 774 if (rcu_dereference_protected(rvp->qp[0], 775 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { 776 RCU_INIT_POINTER(rvp->qp[0], NULL); 777 } else if (rcu_dereference_protected(rvp->qp[1], 778 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { 779 RCU_INIT_POINTER(rvp->qp[1], NULL); 780 } else { 781 struct rvt_qp *q; 782 struct rvt_qp __rcu **qpp; 783 784 removed = 0; 785 qpp = &rdi->qp_dev->qp_table[n]; 786 for (; (q = rcu_dereference_protected(*qpp, 787 lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL; 788 qpp = &q->next) { 789 if (q == qp) { 790 RCU_INIT_POINTER(*qpp, 791 rcu_dereference_protected(qp->next, 792 lockdep_is_held(&rdi->qp_dev->qpt_lock))); 793 removed = 1; 794 trace_rvt_qpremove(qp, n); 795 break; 796 } 797 } 798 } 799 800 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); 801 if (removed) { 802 synchronize_rcu(); 803 rvt_put_qp(qp); 804 } 805 } 806 807 /** 808 * rvt_alloc_rq - allocate memory for user or kernel buffer 809 * @rq: receive queue data structure 810 * @size: number of request queue entries 811 * @node: The NUMA node 812 * @udata: True if user data is available or not false 813 * 814 * Return: If memory allocation failed, return -ENONEM 815 * This function is used by both shared receive 816 * queues and non-shared receive queues to allocate 817 * memory. 818 */ 819 int rvt_alloc_rq(struct rvt_rq *rq, u32 size, int node, 820 struct ib_udata *udata) 821 { 822 if (udata) { 823 rq->wq = vmalloc_user(sizeof(struct rvt_rwq) + size); 824 if (!rq->wq) 825 goto bail; 826 /* need kwq with no buffers */ 827 rq->kwq = kzalloc_node(sizeof(*rq->kwq), GFP_KERNEL, node); 828 if (!rq->kwq) 829 goto bail; 830 rq->kwq->curr_wq = rq->wq->wq; 831 } else { 832 /* need kwq with buffers */ 833 rq->kwq = 834 vzalloc_node(sizeof(struct rvt_krwq) + size, node); 835 if (!rq->kwq) 836 goto bail; 837 rq->kwq->curr_wq = rq->kwq->wq; 838 } 839 840 spin_lock_init(&rq->kwq->p_lock); 841 spin_lock_init(&rq->kwq->c_lock); 842 return 0; 843 bail: 844 rvt_free_rq(rq); 845 return -ENOMEM; 846 } 847 848 /** 849 * rvt_init_qp - initialize the QP state to the reset state 850 * @qp: the QP to init or reinit 851 * @type: the QP type 852 * 853 * This function is called from both rvt_create_qp() and 854 * rvt_reset_qp(). The difference is that the reset 855 * patch the necessary locks to protect against concurent 856 * access. 857 */ 858 static void rvt_init_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, 859 enum ib_qp_type type) 860 { 861 qp->remote_qpn = 0; 862 qp->qkey = 0; 863 qp->qp_access_flags = 0; 864 qp->s_flags &= RVT_S_SIGNAL_REQ_WR; 865 qp->s_hdrwords = 0; 866 qp->s_wqe = NULL; 867 qp->s_draining = 0; 868 qp->s_next_psn = 0; 869 qp->s_last_psn = 0; 870 qp->s_sending_psn = 0; 871 qp->s_sending_hpsn = 0; 872 qp->s_psn = 0; 873 qp->r_psn = 0; 874 qp->r_msn = 0; 875 if (type == IB_QPT_RC) { 876 qp->s_state = IB_OPCODE_RC_SEND_LAST; 877 qp->r_state = IB_OPCODE_RC_SEND_LAST; 878 } else { 879 qp->s_state = IB_OPCODE_UC_SEND_LAST; 880 qp->r_state = IB_OPCODE_UC_SEND_LAST; 881 } 882 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; 883 qp->r_nak_state = 0; 884 qp->r_aflags = 0; 885 qp->r_flags = 0; 886 qp->s_head = 0; 887 qp->s_tail = 0; 888 qp->s_cur = 0; 889 qp->s_acked = 0; 890 qp->s_last = 0; 891 qp->s_ssn = 1; 892 qp->s_lsn = 0; 893 qp->s_mig_state = IB_MIG_MIGRATED; 894 qp->r_head_ack_queue = 0; 895 qp->s_tail_ack_queue = 0; 896 qp->s_acked_ack_queue = 0; 897 qp->s_num_rd_atomic = 0; 898 if (qp->r_rq.kwq) 899 qp->r_rq.kwq->count = qp->r_rq.size; 900 qp->r_sge.num_sge = 0; 901 atomic_set(&qp->s_reserved_used, 0); 902 } 903 904 /** 905 * rvt_reset_qp - initialize the QP state to the reset state 906 * @qp: the QP to reset 907 * @type: the QP type 908 * 909 * r_lock, s_hlock, and s_lock are required to be held by the caller 910 */ 911 static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, 912 enum ib_qp_type type) 913 __must_hold(&qp->s_lock) 914 __must_hold(&qp->s_hlock) 915 __must_hold(&qp->r_lock) 916 { 917 lockdep_assert_held(&qp->r_lock); 918 lockdep_assert_held(&qp->s_hlock); 919 lockdep_assert_held(&qp->s_lock); 920 if (qp->state != IB_QPS_RESET) { 921 qp->state = IB_QPS_RESET; 922 923 /* Let drivers flush their waitlist */ 924 rdi->driver_f.flush_qp_waiters(qp); 925 rvt_stop_rc_timers(qp); 926 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT); 927 spin_unlock(&qp->s_lock); 928 spin_unlock(&qp->s_hlock); 929 spin_unlock_irq(&qp->r_lock); 930 931 /* Stop the send queue and the retry timer */ 932 rdi->driver_f.stop_send_queue(qp); 933 rvt_del_timers_sync(qp); 934 /* Wait for things to stop */ 935 rdi->driver_f.quiesce_qp(qp); 936 937 /* take qp out the hash and wait for it to be unused */ 938 rvt_remove_qp(rdi, qp); 939 940 /* grab the lock b/c it was locked at call time */ 941 spin_lock_irq(&qp->r_lock); 942 spin_lock(&qp->s_hlock); 943 spin_lock(&qp->s_lock); 944 945 rvt_clear_mr_refs(qp, 1); 946 /* 947 * Let the driver do any tear down or re-init it needs to for 948 * a qp that has been reset 949 */ 950 rdi->driver_f.notify_qp_reset(qp); 951 } 952 rvt_init_qp(rdi, qp, type); 953 lockdep_assert_held(&qp->r_lock); 954 lockdep_assert_held(&qp->s_hlock); 955 lockdep_assert_held(&qp->s_lock); 956 } 957 958 /** rvt_free_qpn - Free a qpn from the bit map 959 * @qpt: QP table 960 * @qpn: queue pair number to free 961 */ 962 static void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn) 963 { 964 struct rvt_qpn_map *map; 965 966 map = qpt->map + (qpn & RVT_QPN_MASK) / RVT_BITS_PER_PAGE; 967 if (map->page) 968 clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); 969 } 970 971 /** 972 * get_allowed_ops - Given a QP type return the appropriate allowed OP 973 * @type: valid, supported, QP type 974 */ 975 static u8 get_allowed_ops(enum ib_qp_type type) 976 { 977 return type == IB_QPT_RC ? IB_OPCODE_RC : type == IB_QPT_UC ? 978 IB_OPCODE_UC : IB_OPCODE_UD; 979 } 980 981 /** 982 * free_ud_wq_attr - Clean up AH attribute cache for UD QPs 983 * @qp: Valid QP with allowed_ops set 984 * 985 * The rvt_swqe data structure being used is a union, so this is 986 * only valid for UD QPs. 987 */ 988 static void free_ud_wq_attr(struct rvt_qp *qp) 989 { 990 struct rvt_swqe *wqe; 991 int i; 992 993 for (i = 0; qp->allowed_ops == IB_OPCODE_UD && i < qp->s_size; i++) { 994 wqe = rvt_get_swqe_ptr(qp, i); 995 kfree(wqe->ud_wr.attr); 996 wqe->ud_wr.attr = NULL; 997 } 998 } 999 1000 /** 1001 * alloc_ud_wq_attr - AH attribute cache for UD QPs 1002 * @qp: Valid QP with allowed_ops set 1003 * @node: Numa node for allocation 1004 * 1005 * The rvt_swqe data structure being used is a union, so this is 1006 * only valid for UD QPs. 1007 */ 1008 static int alloc_ud_wq_attr(struct rvt_qp *qp, int node) 1009 { 1010 struct rvt_swqe *wqe; 1011 int i; 1012 1013 for (i = 0; qp->allowed_ops == IB_OPCODE_UD && i < qp->s_size; i++) { 1014 wqe = rvt_get_swqe_ptr(qp, i); 1015 wqe->ud_wr.attr = kzalloc_node(sizeof(*wqe->ud_wr.attr), 1016 GFP_KERNEL, node); 1017 if (!wqe->ud_wr.attr) { 1018 free_ud_wq_attr(qp); 1019 return -ENOMEM; 1020 } 1021 } 1022 1023 return 0; 1024 } 1025 1026 /** 1027 * rvt_create_qp - create a queue pair for a device 1028 * @ibpd: the protection domain who's device we create the queue pair for 1029 * @init_attr: the attributes of the queue pair 1030 * @udata: user data for libibverbs.so 1031 * 1032 * Queue pair creation is mostly an rvt issue. However, drivers have their own 1033 * unique idea of what queue pair numbers mean. For instance there is a reserved 1034 * range for PSM. 1035 * 1036 * Return: the queue pair on success, otherwise returns an errno. 1037 * 1038 * Called by the ib_create_qp() core verbs function. 1039 */ 1040 struct ib_qp *rvt_create_qp(struct ib_pd *ibpd, 1041 struct ib_qp_init_attr *init_attr, 1042 struct ib_udata *udata) 1043 { 1044 struct rvt_qp *qp; 1045 int err; 1046 struct rvt_swqe *swq = NULL; 1047 size_t sz; 1048 size_t sg_list_sz; 1049 struct ib_qp *ret = ERR_PTR(-ENOMEM); 1050 struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device); 1051 void *priv = NULL; 1052 size_t sqsize; 1053 1054 if (!rdi) 1055 return ERR_PTR(-EINVAL); 1056 1057 if (init_attr->cap.max_send_sge > rdi->dparms.props.max_send_sge || 1058 init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr || 1059 init_attr->create_flags) 1060 return ERR_PTR(-EINVAL); 1061 1062 /* Check receive queue parameters if no SRQ is specified. */ 1063 if (!init_attr->srq) { 1064 if (init_attr->cap.max_recv_sge > 1065 rdi->dparms.props.max_recv_sge || 1066 init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr) 1067 return ERR_PTR(-EINVAL); 1068 1069 if (init_attr->cap.max_send_sge + 1070 init_attr->cap.max_send_wr + 1071 init_attr->cap.max_recv_sge + 1072 init_attr->cap.max_recv_wr == 0) 1073 return ERR_PTR(-EINVAL); 1074 } 1075 sqsize = 1076 init_attr->cap.max_send_wr + 1 + 1077 rdi->dparms.reserved_operations; 1078 switch (init_attr->qp_type) { 1079 case IB_QPT_SMI: 1080 case IB_QPT_GSI: 1081 if (init_attr->port_num == 0 || 1082 init_attr->port_num > ibpd->device->phys_port_cnt) 1083 return ERR_PTR(-EINVAL); 1084 /* fall through */ 1085 case IB_QPT_UC: 1086 case IB_QPT_RC: 1087 case IB_QPT_UD: 1088 sz = struct_size(swq, sg_list, init_attr->cap.max_send_sge); 1089 swq = vzalloc_node(array_size(sz, sqsize), rdi->dparms.node); 1090 if (!swq) 1091 return ERR_PTR(-ENOMEM); 1092 1093 sz = sizeof(*qp); 1094 sg_list_sz = 0; 1095 if (init_attr->srq) { 1096 struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq); 1097 1098 if (srq->rq.max_sge > 1) 1099 sg_list_sz = sizeof(*qp->r_sg_list) * 1100 (srq->rq.max_sge - 1); 1101 } else if (init_attr->cap.max_recv_sge > 1) 1102 sg_list_sz = sizeof(*qp->r_sg_list) * 1103 (init_attr->cap.max_recv_sge - 1); 1104 qp = kzalloc_node(sz + sg_list_sz, GFP_KERNEL, 1105 rdi->dparms.node); 1106 if (!qp) 1107 goto bail_swq; 1108 qp->allowed_ops = get_allowed_ops(init_attr->qp_type); 1109 1110 RCU_INIT_POINTER(qp->next, NULL); 1111 if (init_attr->qp_type == IB_QPT_RC) { 1112 qp->s_ack_queue = 1113 kcalloc_node(rvt_max_atomic(rdi), 1114 sizeof(*qp->s_ack_queue), 1115 GFP_KERNEL, 1116 rdi->dparms.node); 1117 if (!qp->s_ack_queue) 1118 goto bail_qp; 1119 } 1120 /* initialize timers needed for rc qp */ 1121 timer_setup(&qp->s_timer, rvt_rc_timeout, 0); 1122 hrtimer_init(&qp->s_rnr_timer, CLOCK_MONOTONIC, 1123 HRTIMER_MODE_REL); 1124 qp->s_rnr_timer.function = rvt_rc_rnr_retry; 1125 1126 /* 1127 * Driver needs to set up it's private QP structure and do any 1128 * initialization that is needed. 1129 */ 1130 priv = rdi->driver_f.qp_priv_alloc(rdi, qp); 1131 if (IS_ERR(priv)) { 1132 ret = priv; 1133 goto bail_qp; 1134 } 1135 qp->priv = priv; 1136 qp->timeout_jiffies = 1137 usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / 1138 1000UL); 1139 if (init_attr->srq) { 1140 sz = 0; 1141 } else { 1142 qp->r_rq.size = init_attr->cap.max_recv_wr + 1; 1143 qp->r_rq.max_sge = init_attr->cap.max_recv_sge; 1144 sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) + 1145 sizeof(struct rvt_rwqe); 1146 err = rvt_alloc_rq(&qp->r_rq, qp->r_rq.size * sz, 1147 rdi->dparms.node, udata); 1148 if (err) { 1149 ret = ERR_PTR(err); 1150 goto bail_driver_priv; 1151 } 1152 } 1153 1154 /* 1155 * ib_create_qp() will initialize qp->ibqp 1156 * except for qp->ibqp.qp_num. 1157 */ 1158 spin_lock_init(&qp->r_lock); 1159 spin_lock_init(&qp->s_hlock); 1160 spin_lock_init(&qp->s_lock); 1161 atomic_set(&qp->refcount, 0); 1162 atomic_set(&qp->local_ops_pending, 0); 1163 init_waitqueue_head(&qp->wait); 1164 INIT_LIST_HEAD(&qp->rspwait); 1165 qp->state = IB_QPS_RESET; 1166 qp->s_wq = swq; 1167 qp->s_size = sqsize; 1168 qp->s_avail = init_attr->cap.max_send_wr; 1169 qp->s_max_sge = init_attr->cap.max_send_sge; 1170 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR) 1171 qp->s_flags = RVT_S_SIGNAL_REQ_WR; 1172 err = alloc_ud_wq_attr(qp, rdi->dparms.node); 1173 if (err) { 1174 ret = (ERR_PTR(err)); 1175 goto bail_driver_priv; 1176 } 1177 1178 err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table, 1179 init_attr->qp_type, 1180 init_attr->port_num); 1181 if (err < 0) { 1182 ret = ERR_PTR(err); 1183 goto bail_rq_wq; 1184 } 1185 qp->ibqp.qp_num = err; 1186 qp->port_num = init_attr->port_num; 1187 rvt_init_qp(rdi, qp, init_attr->qp_type); 1188 if (rdi->driver_f.qp_priv_init) { 1189 err = rdi->driver_f.qp_priv_init(rdi, qp, init_attr); 1190 if (err) { 1191 ret = ERR_PTR(err); 1192 goto bail_rq_wq; 1193 } 1194 } 1195 break; 1196 1197 default: 1198 /* Don't support raw QPs */ 1199 return ERR_PTR(-EINVAL); 1200 } 1201 1202 init_attr->cap.max_inline_data = 0; 1203 1204 /* 1205 * Return the address of the RWQ as the offset to mmap. 1206 * See rvt_mmap() for details. 1207 */ 1208 if (udata && udata->outlen >= sizeof(__u64)) { 1209 if (!qp->r_rq.wq) { 1210 __u64 offset = 0; 1211 1212 err = ib_copy_to_udata(udata, &offset, 1213 sizeof(offset)); 1214 if (err) { 1215 ret = ERR_PTR(err); 1216 goto bail_qpn; 1217 } 1218 } else { 1219 u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz; 1220 1221 qp->ip = rvt_create_mmap_info(rdi, s, udata, 1222 qp->r_rq.wq); 1223 if (!qp->ip) { 1224 ret = ERR_PTR(-ENOMEM); 1225 goto bail_qpn; 1226 } 1227 1228 err = ib_copy_to_udata(udata, &qp->ip->offset, 1229 sizeof(qp->ip->offset)); 1230 if (err) { 1231 ret = ERR_PTR(err); 1232 goto bail_ip; 1233 } 1234 } 1235 qp->pid = current->pid; 1236 } 1237 1238 spin_lock(&rdi->n_qps_lock); 1239 if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) { 1240 spin_unlock(&rdi->n_qps_lock); 1241 ret = ERR_PTR(-ENOMEM); 1242 goto bail_ip; 1243 } 1244 1245 rdi->n_qps_allocated++; 1246 /* 1247 * Maintain a busy_jiffies variable that will be added to the timeout 1248 * period in mod_retry_timer and add_retry_timer. This busy jiffies 1249 * is scaled by the number of rc qps created for the device to reduce 1250 * the number of timeouts occurring when there is a large number of 1251 * qps. busy_jiffies is incremented every rc qp scaling interval. 1252 * The scaling interval is selected based on extensive performance 1253 * evaluation of targeted workloads. 1254 */ 1255 if (init_attr->qp_type == IB_QPT_RC) { 1256 rdi->n_rc_qps++; 1257 rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL; 1258 } 1259 spin_unlock(&rdi->n_qps_lock); 1260 1261 if (qp->ip) { 1262 spin_lock_irq(&rdi->pending_lock); 1263 list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps); 1264 spin_unlock_irq(&rdi->pending_lock); 1265 } 1266 1267 ret = &qp->ibqp; 1268 1269 return ret; 1270 1271 bail_ip: 1272 if (qp->ip) 1273 kref_put(&qp->ip->ref, rvt_release_mmap_info); 1274 1275 bail_qpn: 1276 rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); 1277 1278 bail_rq_wq: 1279 rvt_free_rq(&qp->r_rq); 1280 free_ud_wq_attr(qp); 1281 1282 bail_driver_priv: 1283 rdi->driver_f.qp_priv_free(rdi, qp); 1284 1285 bail_qp: 1286 kfree(qp->s_ack_queue); 1287 kfree(qp); 1288 1289 bail_swq: 1290 vfree(swq); 1291 1292 return ret; 1293 } 1294 1295 /** 1296 * rvt_error_qp - put a QP into the error state 1297 * @qp: the QP to put into the error state 1298 * @err: the receive completion error to signal if a RWQE is active 1299 * 1300 * Flushes both send and receive work queues. 1301 * 1302 * Return: true if last WQE event should be generated. 1303 * The QP r_lock and s_lock should be held and interrupts disabled. 1304 * If we are already in error state, just return. 1305 */ 1306 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err) 1307 { 1308 struct ib_wc wc; 1309 int ret = 0; 1310 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 1311 1312 lockdep_assert_held(&qp->r_lock); 1313 lockdep_assert_held(&qp->s_lock); 1314 if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET) 1315 goto bail; 1316 1317 qp->state = IB_QPS_ERR; 1318 1319 if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) { 1320 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR); 1321 del_timer(&qp->s_timer); 1322 } 1323 1324 if (qp->s_flags & RVT_S_ANY_WAIT_SEND) 1325 qp->s_flags &= ~RVT_S_ANY_WAIT_SEND; 1326 1327 rdi->driver_f.notify_error_qp(qp); 1328 1329 /* Schedule the sending tasklet to drain the send work queue. */ 1330 if (READ_ONCE(qp->s_last) != qp->s_head) 1331 rdi->driver_f.schedule_send(qp); 1332 1333 rvt_clear_mr_refs(qp, 0); 1334 1335 memset(&wc, 0, sizeof(wc)); 1336 wc.qp = &qp->ibqp; 1337 wc.opcode = IB_WC_RECV; 1338 1339 if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) { 1340 wc.wr_id = qp->r_wr_id; 1341 wc.status = err; 1342 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 1343 } 1344 wc.status = IB_WC_WR_FLUSH_ERR; 1345 1346 if (qp->r_rq.kwq) { 1347 u32 head; 1348 u32 tail; 1349 struct rvt_rwq *wq = NULL; 1350 struct rvt_krwq *kwq = NULL; 1351 1352 spin_lock(&qp->r_rq.kwq->c_lock); 1353 /* qp->ip used to validate if there is a user buffer mmaped */ 1354 if (qp->ip) { 1355 wq = qp->r_rq.wq; 1356 head = RDMA_READ_UAPI_ATOMIC(wq->head); 1357 tail = RDMA_READ_UAPI_ATOMIC(wq->tail); 1358 } else { 1359 kwq = qp->r_rq.kwq; 1360 head = kwq->head; 1361 tail = kwq->tail; 1362 } 1363 /* sanity check pointers before trusting them */ 1364 if (head >= qp->r_rq.size) 1365 head = 0; 1366 if (tail >= qp->r_rq.size) 1367 tail = 0; 1368 while (tail != head) { 1369 wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id; 1370 if (++tail >= qp->r_rq.size) 1371 tail = 0; 1372 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 1373 } 1374 if (qp->ip) 1375 RDMA_WRITE_UAPI_ATOMIC(wq->tail, tail); 1376 else 1377 kwq->tail = tail; 1378 spin_unlock(&qp->r_rq.kwq->c_lock); 1379 } else if (qp->ibqp.event_handler) { 1380 ret = 1; 1381 } 1382 1383 bail: 1384 return ret; 1385 } 1386 EXPORT_SYMBOL(rvt_error_qp); 1387 1388 /* 1389 * Put the QP into the hash table. 1390 * The hash table holds a reference to the QP. 1391 */ 1392 static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) 1393 { 1394 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; 1395 unsigned long flags; 1396 1397 rvt_get_qp(qp); 1398 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); 1399 1400 if (qp->ibqp.qp_num <= 1) { 1401 rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp); 1402 } else { 1403 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); 1404 1405 qp->next = rdi->qp_dev->qp_table[n]; 1406 rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp); 1407 trace_rvt_qpinsert(qp, n); 1408 } 1409 1410 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); 1411 } 1412 1413 /** 1414 * rvt_modify_qp - modify the attributes of a queue pair 1415 * @ibqp: the queue pair who's attributes we're modifying 1416 * @attr: the new attributes 1417 * @attr_mask: the mask of attributes to modify 1418 * @udata: user data for libibverbs.so 1419 * 1420 * Return: 0 on success, otherwise returns an errno. 1421 */ 1422 int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 1423 int attr_mask, struct ib_udata *udata) 1424 { 1425 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 1426 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1427 enum ib_qp_state cur_state, new_state; 1428 struct ib_event ev; 1429 int lastwqe = 0; 1430 int mig = 0; 1431 int pmtu = 0; /* for gcc warning only */ 1432 int opa_ah; 1433 1434 spin_lock_irq(&qp->r_lock); 1435 spin_lock(&qp->s_hlock); 1436 spin_lock(&qp->s_lock); 1437 1438 cur_state = attr_mask & IB_QP_CUR_STATE ? 1439 attr->cur_qp_state : qp->state; 1440 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 1441 opa_ah = rdma_cap_opa_ah(ibqp->device, qp->port_num); 1442 1443 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, 1444 attr_mask)) 1445 goto inval; 1446 1447 if (rdi->driver_f.check_modify_qp && 1448 rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata)) 1449 goto inval; 1450 1451 if (attr_mask & IB_QP_AV) { 1452 if (opa_ah) { 1453 if (rdma_ah_get_dlid(&attr->ah_attr) >= 1454 opa_get_mcast_base(OPA_MCAST_NR)) 1455 goto inval; 1456 } else { 1457 if (rdma_ah_get_dlid(&attr->ah_attr) >= 1458 be16_to_cpu(IB_MULTICAST_LID_BASE)) 1459 goto inval; 1460 } 1461 1462 if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr)) 1463 goto inval; 1464 } 1465 1466 if (attr_mask & IB_QP_ALT_PATH) { 1467 if (opa_ah) { 1468 if (rdma_ah_get_dlid(&attr->alt_ah_attr) >= 1469 opa_get_mcast_base(OPA_MCAST_NR)) 1470 goto inval; 1471 } else { 1472 if (rdma_ah_get_dlid(&attr->alt_ah_attr) >= 1473 be16_to_cpu(IB_MULTICAST_LID_BASE)) 1474 goto inval; 1475 } 1476 1477 if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr)) 1478 goto inval; 1479 if (attr->alt_pkey_index >= rvt_get_npkeys(rdi)) 1480 goto inval; 1481 } 1482 1483 if (attr_mask & IB_QP_PKEY_INDEX) 1484 if (attr->pkey_index >= rvt_get_npkeys(rdi)) 1485 goto inval; 1486 1487 if (attr_mask & IB_QP_MIN_RNR_TIMER) 1488 if (attr->min_rnr_timer > 31) 1489 goto inval; 1490 1491 if (attr_mask & IB_QP_PORT) 1492 if (qp->ibqp.qp_type == IB_QPT_SMI || 1493 qp->ibqp.qp_type == IB_QPT_GSI || 1494 attr->port_num == 0 || 1495 attr->port_num > ibqp->device->phys_port_cnt) 1496 goto inval; 1497 1498 if (attr_mask & IB_QP_DEST_QPN) 1499 if (attr->dest_qp_num > RVT_QPN_MASK) 1500 goto inval; 1501 1502 if (attr_mask & IB_QP_RETRY_CNT) 1503 if (attr->retry_cnt > 7) 1504 goto inval; 1505 1506 if (attr_mask & IB_QP_RNR_RETRY) 1507 if (attr->rnr_retry > 7) 1508 goto inval; 1509 1510 /* 1511 * Don't allow invalid path_mtu values. OK to set greater 1512 * than the active mtu (or even the max_cap, if we have tuned 1513 * that to a small mtu. We'll set qp->path_mtu 1514 * to the lesser of requested attribute mtu and active, 1515 * for packetizing messages. 1516 * Note that the QP port has to be set in INIT and MTU in RTR. 1517 */ 1518 if (attr_mask & IB_QP_PATH_MTU) { 1519 pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr); 1520 if (pmtu < 0) 1521 goto inval; 1522 } 1523 1524 if (attr_mask & IB_QP_PATH_MIG_STATE) { 1525 if (attr->path_mig_state == IB_MIG_REARM) { 1526 if (qp->s_mig_state == IB_MIG_ARMED) 1527 goto inval; 1528 if (new_state != IB_QPS_RTS) 1529 goto inval; 1530 } else if (attr->path_mig_state == IB_MIG_MIGRATED) { 1531 if (qp->s_mig_state == IB_MIG_REARM) 1532 goto inval; 1533 if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD) 1534 goto inval; 1535 if (qp->s_mig_state == IB_MIG_ARMED) 1536 mig = 1; 1537 } else { 1538 goto inval; 1539 } 1540 } 1541 1542 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1543 if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic) 1544 goto inval; 1545 1546 switch (new_state) { 1547 case IB_QPS_RESET: 1548 if (qp->state != IB_QPS_RESET) 1549 rvt_reset_qp(rdi, qp, ibqp->qp_type); 1550 break; 1551 1552 case IB_QPS_RTR: 1553 /* Allow event to re-trigger if QP set to RTR more than once */ 1554 qp->r_flags &= ~RVT_R_COMM_EST; 1555 qp->state = new_state; 1556 break; 1557 1558 case IB_QPS_SQD: 1559 qp->s_draining = qp->s_last != qp->s_cur; 1560 qp->state = new_state; 1561 break; 1562 1563 case IB_QPS_SQE: 1564 if (qp->ibqp.qp_type == IB_QPT_RC) 1565 goto inval; 1566 qp->state = new_state; 1567 break; 1568 1569 case IB_QPS_ERR: 1570 lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); 1571 break; 1572 1573 default: 1574 qp->state = new_state; 1575 break; 1576 } 1577 1578 if (attr_mask & IB_QP_PKEY_INDEX) 1579 qp->s_pkey_index = attr->pkey_index; 1580 1581 if (attr_mask & IB_QP_PORT) 1582 qp->port_num = attr->port_num; 1583 1584 if (attr_mask & IB_QP_DEST_QPN) 1585 qp->remote_qpn = attr->dest_qp_num; 1586 1587 if (attr_mask & IB_QP_SQ_PSN) { 1588 qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask; 1589 qp->s_psn = qp->s_next_psn; 1590 qp->s_sending_psn = qp->s_next_psn; 1591 qp->s_last_psn = qp->s_next_psn - 1; 1592 qp->s_sending_hpsn = qp->s_last_psn; 1593 } 1594 1595 if (attr_mask & IB_QP_RQ_PSN) 1596 qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask; 1597 1598 if (attr_mask & IB_QP_ACCESS_FLAGS) 1599 qp->qp_access_flags = attr->qp_access_flags; 1600 1601 if (attr_mask & IB_QP_AV) { 1602 rdma_replace_ah_attr(&qp->remote_ah_attr, &attr->ah_attr); 1603 qp->s_srate = rdma_ah_get_static_rate(&attr->ah_attr); 1604 qp->srate_mbps = ib_rate_to_mbps(qp->s_srate); 1605 } 1606 1607 if (attr_mask & IB_QP_ALT_PATH) { 1608 rdma_replace_ah_attr(&qp->alt_ah_attr, &attr->alt_ah_attr); 1609 qp->s_alt_pkey_index = attr->alt_pkey_index; 1610 } 1611 1612 if (attr_mask & IB_QP_PATH_MIG_STATE) { 1613 qp->s_mig_state = attr->path_mig_state; 1614 if (mig) { 1615 qp->remote_ah_attr = qp->alt_ah_attr; 1616 qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr); 1617 qp->s_pkey_index = qp->s_alt_pkey_index; 1618 } 1619 } 1620 1621 if (attr_mask & IB_QP_PATH_MTU) { 1622 qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu); 1623 qp->log_pmtu = ilog2(qp->pmtu); 1624 } 1625 1626 if (attr_mask & IB_QP_RETRY_CNT) { 1627 qp->s_retry_cnt = attr->retry_cnt; 1628 qp->s_retry = attr->retry_cnt; 1629 } 1630 1631 if (attr_mask & IB_QP_RNR_RETRY) { 1632 qp->s_rnr_retry_cnt = attr->rnr_retry; 1633 qp->s_rnr_retry = attr->rnr_retry; 1634 } 1635 1636 if (attr_mask & IB_QP_MIN_RNR_TIMER) 1637 qp->r_min_rnr_timer = attr->min_rnr_timer; 1638 1639 if (attr_mask & IB_QP_TIMEOUT) { 1640 qp->timeout = attr->timeout; 1641 qp->timeout_jiffies = rvt_timeout_to_jiffies(qp->timeout); 1642 } 1643 1644 if (attr_mask & IB_QP_QKEY) 1645 qp->qkey = attr->qkey; 1646 1647 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 1648 qp->r_max_rd_atomic = attr->max_dest_rd_atomic; 1649 1650 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) 1651 qp->s_max_rd_atomic = attr->max_rd_atomic; 1652 1653 if (rdi->driver_f.modify_qp) 1654 rdi->driver_f.modify_qp(qp, attr, attr_mask, udata); 1655 1656 spin_unlock(&qp->s_lock); 1657 spin_unlock(&qp->s_hlock); 1658 spin_unlock_irq(&qp->r_lock); 1659 1660 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) 1661 rvt_insert_qp(rdi, qp); 1662 1663 if (lastwqe) { 1664 ev.device = qp->ibqp.device; 1665 ev.element.qp = &qp->ibqp; 1666 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 1667 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 1668 } 1669 if (mig) { 1670 ev.device = qp->ibqp.device; 1671 ev.element.qp = &qp->ibqp; 1672 ev.event = IB_EVENT_PATH_MIG; 1673 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 1674 } 1675 return 0; 1676 1677 inval: 1678 spin_unlock(&qp->s_lock); 1679 spin_unlock(&qp->s_hlock); 1680 spin_unlock_irq(&qp->r_lock); 1681 return -EINVAL; 1682 } 1683 1684 /** 1685 * rvt_destroy_qp - destroy a queue pair 1686 * @ibqp: the queue pair to destroy 1687 * 1688 * Note that this can be called while the QP is actively sending or 1689 * receiving! 1690 * 1691 * Return: 0 on success. 1692 */ 1693 int rvt_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata) 1694 { 1695 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1696 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 1697 1698 spin_lock_irq(&qp->r_lock); 1699 spin_lock(&qp->s_hlock); 1700 spin_lock(&qp->s_lock); 1701 rvt_reset_qp(rdi, qp, ibqp->qp_type); 1702 spin_unlock(&qp->s_lock); 1703 spin_unlock(&qp->s_hlock); 1704 spin_unlock_irq(&qp->r_lock); 1705 1706 wait_event(qp->wait, !atomic_read(&qp->refcount)); 1707 /* qpn is now available for use again */ 1708 rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); 1709 1710 spin_lock(&rdi->n_qps_lock); 1711 rdi->n_qps_allocated--; 1712 if (qp->ibqp.qp_type == IB_QPT_RC) { 1713 rdi->n_rc_qps--; 1714 rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL; 1715 } 1716 spin_unlock(&rdi->n_qps_lock); 1717 1718 if (qp->ip) 1719 kref_put(&qp->ip->ref, rvt_release_mmap_info); 1720 kvfree(qp->r_rq.kwq); 1721 rdi->driver_f.qp_priv_free(rdi, qp); 1722 kfree(qp->s_ack_queue); 1723 rdma_destroy_ah_attr(&qp->remote_ah_attr); 1724 rdma_destroy_ah_attr(&qp->alt_ah_attr); 1725 free_ud_wq_attr(qp); 1726 vfree(qp->s_wq); 1727 kfree(qp); 1728 return 0; 1729 } 1730 1731 /** 1732 * rvt_query_qp - query an ipbq 1733 * @ibqp: IB qp to query 1734 * @attr: attr struct to fill in 1735 * @attr_mask: attr mask ignored 1736 * @init_attr: struct to fill in 1737 * 1738 * Return: always 0 1739 */ 1740 int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, 1741 int attr_mask, struct ib_qp_init_attr *init_attr) 1742 { 1743 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1744 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 1745 1746 attr->qp_state = qp->state; 1747 attr->cur_qp_state = attr->qp_state; 1748 attr->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu); 1749 attr->path_mig_state = qp->s_mig_state; 1750 attr->qkey = qp->qkey; 1751 attr->rq_psn = qp->r_psn & rdi->dparms.psn_mask; 1752 attr->sq_psn = qp->s_next_psn & rdi->dparms.psn_mask; 1753 attr->dest_qp_num = qp->remote_qpn; 1754 attr->qp_access_flags = qp->qp_access_flags; 1755 attr->cap.max_send_wr = qp->s_size - 1 - 1756 rdi->dparms.reserved_operations; 1757 attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1; 1758 attr->cap.max_send_sge = qp->s_max_sge; 1759 attr->cap.max_recv_sge = qp->r_rq.max_sge; 1760 attr->cap.max_inline_data = 0; 1761 attr->ah_attr = qp->remote_ah_attr; 1762 attr->alt_ah_attr = qp->alt_ah_attr; 1763 attr->pkey_index = qp->s_pkey_index; 1764 attr->alt_pkey_index = qp->s_alt_pkey_index; 1765 attr->en_sqd_async_notify = 0; 1766 attr->sq_draining = qp->s_draining; 1767 attr->max_rd_atomic = qp->s_max_rd_atomic; 1768 attr->max_dest_rd_atomic = qp->r_max_rd_atomic; 1769 attr->min_rnr_timer = qp->r_min_rnr_timer; 1770 attr->port_num = qp->port_num; 1771 attr->timeout = qp->timeout; 1772 attr->retry_cnt = qp->s_retry_cnt; 1773 attr->rnr_retry = qp->s_rnr_retry_cnt; 1774 attr->alt_port_num = 1775 rdma_ah_get_port_num(&qp->alt_ah_attr); 1776 attr->alt_timeout = qp->alt_timeout; 1777 1778 init_attr->event_handler = qp->ibqp.event_handler; 1779 init_attr->qp_context = qp->ibqp.qp_context; 1780 init_attr->send_cq = qp->ibqp.send_cq; 1781 init_attr->recv_cq = qp->ibqp.recv_cq; 1782 init_attr->srq = qp->ibqp.srq; 1783 init_attr->cap = attr->cap; 1784 if (qp->s_flags & RVT_S_SIGNAL_REQ_WR) 1785 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR; 1786 else 1787 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR; 1788 init_attr->qp_type = qp->ibqp.qp_type; 1789 init_attr->port_num = qp->port_num; 1790 return 0; 1791 } 1792 1793 /** 1794 * rvt_post_receive - post a receive on a QP 1795 * @ibqp: the QP to post the receive on 1796 * @wr: the WR to post 1797 * @bad_wr: the first bad WR is put here 1798 * 1799 * This may be called from interrupt context. 1800 * 1801 * Return: 0 on success otherwise errno 1802 */ 1803 int rvt_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr, 1804 const struct ib_recv_wr **bad_wr) 1805 { 1806 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 1807 struct rvt_krwq *wq = qp->r_rq.kwq; 1808 unsigned long flags; 1809 int qp_err_flush = (ib_rvt_state_ops[qp->state] & RVT_FLUSH_RECV) && 1810 !qp->ibqp.srq; 1811 1812 /* Check that state is OK to post receive. */ 1813 if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) { 1814 *bad_wr = wr; 1815 return -EINVAL; 1816 } 1817 1818 for (; wr; wr = wr->next) { 1819 struct rvt_rwqe *wqe; 1820 u32 next; 1821 int i; 1822 1823 if ((unsigned)wr->num_sge > qp->r_rq.max_sge) { 1824 *bad_wr = wr; 1825 return -EINVAL; 1826 } 1827 1828 spin_lock_irqsave(&qp->r_rq.kwq->p_lock, flags); 1829 next = wq->head + 1; 1830 if (next >= qp->r_rq.size) 1831 next = 0; 1832 if (next == READ_ONCE(wq->tail)) { 1833 spin_unlock_irqrestore(&qp->r_rq.kwq->p_lock, flags); 1834 *bad_wr = wr; 1835 return -ENOMEM; 1836 } 1837 if (unlikely(qp_err_flush)) { 1838 struct ib_wc wc; 1839 1840 memset(&wc, 0, sizeof(wc)); 1841 wc.qp = &qp->ibqp; 1842 wc.opcode = IB_WC_RECV; 1843 wc.wr_id = wr->wr_id; 1844 wc.status = IB_WC_WR_FLUSH_ERR; 1845 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 1846 } else { 1847 wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head); 1848 wqe->wr_id = wr->wr_id; 1849 wqe->num_sge = wr->num_sge; 1850 for (i = 0; i < wr->num_sge; i++) { 1851 wqe->sg_list[i].addr = wr->sg_list[i].addr; 1852 wqe->sg_list[i].length = wr->sg_list[i].length; 1853 wqe->sg_list[i].lkey = wr->sg_list[i].lkey; 1854 } 1855 /* 1856 * Make sure queue entry is written 1857 * before the head index. 1858 */ 1859 smp_store_release(&wq->head, next); 1860 } 1861 spin_unlock_irqrestore(&qp->r_rq.kwq->p_lock, flags); 1862 } 1863 return 0; 1864 } 1865 1866 /** 1867 * rvt_qp_valid_operation - validate post send wr request 1868 * @qp - the qp 1869 * @post-parms - the post send table for the driver 1870 * @wr - the work request 1871 * 1872 * The routine validates the operation based on the 1873 * validation table an returns the length of the operation 1874 * which can extend beyond the ib_send_bw. Operation 1875 * dependent flags key atomic operation validation. 1876 * 1877 * There is an exception for UD qps that validates the pd and 1878 * overrides the length to include the additional UD specific 1879 * length. 1880 * 1881 * Returns a negative error or the length of the work request 1882 * for building the swqe. 1883 */ 1884 static inline int rvt_qp_valid_operation( 1885 struct rvt_qp *qp, 1886 const struct rvt_operation_params *post_parms, 1887 const struct ib_send_wr *wr) 1888 { 1889 int len; 1890 1891 if (wr->opcode >= RVT_OPERATION_MAX || !post_parms[wr->opcode].length) 1892 return -EINVAL; 1893 if (!(post_parms[wr->opcode].qpt_support & BIT(qp->ibqp.qp_type))) 1894 return -EINVAL; 1895 if ((post_parms[wr->opcode].flags & RVT_OPERATION_PRIV) && 1896 ibpd_to_rvtpd(qp->ibqp.pd)->user) 1897 return -EINVAL; 1898 if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC_SGE && 1899 (wr->num_sge == 0 || 1900 wr->sg_list[0].length < sizeof(u64) || 1901 wr->sg_list[0].addr & (sizeof(u64) - 1))) 1902 return -EINVAL; 1903 if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC && 1904 !qp->s_max_rd_atomic) 1905 return -EINVAL; 1906 len = post_parms[wr->opcode].length; 1907 /* UD specific */ 1908 if (qp->ibqp.qp_type != IB_QPT_UC && 1909 qp->ibqp.qp_type != IB_QPT_RC) { 1910 if (qp->ibqp.pd != ud_wr(wr)->ah->pd) 1911 return -EINVAL; 1912 len = sizeof(struct ib_ud_wr); 1913 } 1914 return len; 1915 } 1916 1917 /** 1918 * rvt_qp_is_avail - determine queue capacity 1919 * @qp: the qp 1920 * @rdi: the rdmavt device 1921 * @reserved_op: is reserved operation 1922 * 1923 * This assumes the s_hlock is held but the s_last 1924 * qp variable is uncontrolled. 1925 * 1926 * For non reserved operations, the qp->s_avail 1927 * may be changed. 1928 * 1929 * The return value is zero or a -ENOMEM. 1930 */ 1931 static inline int rvt_qp_is_avail( 1932 struct rvt_qp *qp, 1933 struct rvt_dev_info *rdi, 1934 bool reserved_op) 1935 { 1936 u32 slast; 1937 u32 avail; 1938 u32 reserved_used; 1939 1940 /* see rvt_qp_wqe_unreserve() */ 1941 smp_mb__before_atomic(); 1942 if (unlikely(reserved_op)) { 1943 /* see rvt_qp_wqe_unreserve() */ 1944 reserved_used = atomic_read(&qp->s_reserved_used); 1945 if (reserved_used >= rdi->dparms.reserved_operations) 1946 return -ENOMEM; 1947 return 0; 1948 } 1949 /* non-reserved operations */ 1950 if (likely(qp->s_avail)) 1951 return 0; 1952 /* See rvt_qp_complete_swqe() */ 1953 slast = smp_load_acquire(&qp->s_last); 1954 if (qp->s_head >= slast) 1955 avail = qp->s_size - (qp->s_head - slast); 1956 else 1957 avail = slast - qp->s_head; 1958 1959 reserved_used = atomic_read(&qp->s_reserved_used); 1960 avail = avail - 1 - 1961 (rdi->dparms.reserved_operations - reserved_used); 1962 /* insure we don't assign a negative s_avail */ 1963 if ((s32)avail <= 0) 1964 return -ENOMEM; 1965 qp->s_avail = avail; 1966 if (WARN_ON(qp->s_avail > 1967 (qp->s_size - 1 - rdi->dparms.reserved_operations))) 1968 rvt_pr_err(rdi, 1969 "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u", 1970 qp->ibqp.qp_num, qp->s_size, qp->s_avail, 1971 qp->s_head, qp->s_tail, qp->s_cur, 1972 qp->s_acked, qp->s_last); 1973 return 0; 1974 } 1975 1976 /** 1977 * rvt_post_one_wr - post one RC, UC, or UD send work request 1978 * @qp: the QP to post on 1979 * @wr: the work request to send 1980 */ 1981 static int rvt_post_one_wr(struct rvt_qp *qp, 1982 const struct ib_send_wr *wr, 1983 bool *call_send) 1984 { 1985 struct rvt_swqe *wqe; 1986 u32 next; 1987 int i; 1988 int j; 1989 int acc; 1990 struct rvt_lkey_table *rkt; 1991 struct rvt_pd *pd; 1992 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 1993 u8 log_pmtu; 1994 int ret; 1995 size_t cplen; 1996 bool reserved_op; 1997 int local_ops_delayed = 0; 1998 1999 BUILD_BUG_ON(IB_QPT_MAX >= (sizeof(u32) * BITS_PER_BYTE)); 2000 2001 /* IB spec says that num_sge == 0 is OK. */ 2002 if (unlikely(wr->num_sge > qp->s_max_sge)) 2003 return -EINVAL; 2004 2005 ret = rvt_qp_valid_operation(qp, rdi->post_parms, wr); 2006 if (ret < 0) 2007 return ret; 2008 cplen = ret; 2009 2010 /* 2011 * Local operations include fast register and local invalidate. 2012 * Fast register needs to be processed immediately because the 2013 * registered lkey may be used by following work requests and the 2014 * lkey needs to be valid at the time those requests are posted. 2015 * Local invalidate can be processed immediately if fencing is 2016 * not required and no previous local invalidate ops are pending. 2017 * Signaled local operations that have been processed immediately 2018 * need to have requests with "completion only" flags set posted 2019 * to the send queue in order to generate completions. 2020 */ 2021 if ((rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL)) { 2022 switch (wr->opcode) { 2023 case IB_WR_REG_MR: 2024 ret = rvt_fast_reg_mr(qp, 2025 reg_wr(wr)->mr, 2026 reg_wr(wr)->key, 2027 reg_wr(wr)->access); 2028 if (ret || !(wr->send_flags & IB_SEND_SIGNALED)) 2029 return ret; 2030 break; 2031 case IB_WR_LOCAL_INV: 2032 if ((wr->send_flags & IB_SEND_FENCE) || 2033 atomic_read(&qp->local_ops_pending)) { 2034 local_ops_delayed = 1; 2035 } else { 2036 ret = rvt_invalidate_rkey( 2037 qp, wr->ex.invalidate_rkey); 2038 if (ret || !(wr->send_flags & IB_SEND_SIGNALED)) 2039 return ret; 2040 } 2041 break; 2042 default: 2043 return -EINVAL; 2044 } 2045 } 2046 2047 reserved_op = rdi->post_parms[wr->opcode].flags & 2048 RVT_OPERATION_USE_RESERVE; 2049 /* check for avail */ 2050 ret = rvt_qp_is_avail(qp, rdi, reserved_op); 2051 if (ret) 2052 return ret; 2053 next = qp->s_head + 1; 2054 if (next >= qp->s_size) 2055 next = 0; 2056 2057 rkt = &rdi->lkey_table; 2058 pd = ibpd_to_rvtpd(qp->ibqp.pd); 2059 wqe = rvt_get_swqe_ptr(qp, qp->s_head); 2060 2061 /* cplen has length from above */ 2062 memcpy(&wqe->wr, wr, cplen); 2063 2064 wqe->length = 0; 2065 j = 0; 2066 if (wr->num_sge) { 2067 struct rvt_sge *last_sge = NULL; 2068 2069 acc = wr->opcode >= IB_WR_RDMA_READ ? 2070 IB_ACCESS_LOCAL_WRITE : 0; 2071 for (i = 0; i < wr->num_sge; i++) { 2072 u32 length = wr->sg_list[i].length; 2073 2074 if (length == 0) 2075 continue; 2076 ret = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j], last_sge, 2077 &wr->sg_list[i], acc); 2078 if (unlikely(ret < 0)) 2079 goto bail_inval_free; 2080 wqe->length += length; 2081 if (ret) 2082 last_sge = &wqe->sg_list[j]; 2083 j += ret; 2084 } 2085 wqe->wr.num_sge = j; 2086 } 2087 2088 /* 2089 * Calculate and set SWQE PSN values prior to handing it off 2090 * to the driver's check routine. This give the driver the 2091 * opportunity to adjust PSN values based on internal checks. 2092 */ 2093 log_pmtu = qp->log_pmtu; 2094 if (qp->allowed_ops == IB_OPCODE_UD) { 2095 struct rvt_ah *ah = rvt_get_swqe_ah(wqe); 2096 2097 log_pmtu = ah->log_pmtu; 2098 rdma_copy_ah_attr(wqe->ud_wr.attr, &ah->attr); 2099 } 2100 2101 if (rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL) { 2102 if (local_ops_delayed) 2103 atomic_inc(&qp->local_ops_pending); 2104 else 2105 wqe->wr.send_flags |= RVT_SEND_COMPLETION_ONLY; 2106 wqe->ssn = 0; 2107 wqe->psn = 0; 2108 wqe->lpsn = 0; 2109 } else { 2110 wqe->ssn = qp->s_ssn++; 2111 wqe->psn = qp->s_next_psn; 2112 wqe->lpsn = wqe->psn + 2113 (wqe->length ? 2114 ((wqe->length - 1) >> log_pmtu) : 2115 0); 2116 } 2117 2118 /* general part of wqe valid - allow for driver checks */ 2119 if (rdi->driver_f.setup_wqe) { 2120 ret = rdi->driver_f.setup_wqe(qp, wqe, call_send); 2121 if (ret < 0) 2122 goto bail_inval_free_ref; 2123 } 2124 2125 if (!(rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL)) 2126 qp->s_next_psn = wqe->lpsn + 1; 2127 2128 if (unlikely(reserved_op)) { 2129 wqe->wr.send_flags |= RVT_SEND_RESERVE_USED; 2130 rvt_qp_wqe_reserve(qp, wqe); 2131 } else { 2132 wqe->wr.send_flags &= ~RVT_SEND_RESERVE_USED; 2133 qp->s_avail--; 2134 } 2135 trace_rvt_post_one_wr(qp, wqe, wr->num_sge); 2136 smp_wmb(); /* see request builders */ 2137 qp->s_head = next; 2138 2139 return 0; 2140 2141 bail_inval_free_ref: 2142 if (qp->allowed_ops == IB_OPCODE_UD) 2143 rdma_destroy_ah_attr(wqe->ud_wr.attr); 2144 bail_inval_free: 2145 /* release mr holds */ 2146 while (j) { 2147 struct rvt_sge *sge = &wqe->sg_list[--j]; 2148 2149 rvt_put_mr(sge->mr); 2150 } 2151 return ret; 2152 } 2153 2154 /** 2155 * rvt_post_send - post a send on a QP 2156 * @ibqp: the QP to post the send on 2157 * @wr: the list of work requests to post 2158 * @bad_wr: the first bad WR is put here 2159 * 2160 * This may be called from interrupt context. 2161 * 2162 * Return: 0 on success else errno 2163 */ 2164 int rvt_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr, 2165 const struct ib_send_wr **bad_wr) 2166 { 2167 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); 2168 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 2169 unsigned long flags = 0; 2170 bool call_send; 2171 unsigned nreq = 0; 2172 int err = 0; 2173 2174 spin_lock_irqsave(&qp->s_hlock, flags); 2175 2176 /* 2177 * Ensure QP state is such that we can send. If not bail out early, 2178 * there is no need to do this every time we post a send. 2179 */ 2180 if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) { 2181 spin_unlock_irqrestore(&qp->s_hlock, flags); 2182 return -EINVAL; 2183 } 2184 2185 /* 2186 * If the send queue is empty, and we only have a single WR then just go 2187 * ahead and kick the send engine into gear. Otherwise we will always 2188 * just schedule the send to happen later. 2189 */ 2190 call_send = qp->s_head == READ_ONCE(qp->s_last) && !wr->next; 2191 2192 for (; wr; wr = wr->next) { 2193 err = rvt_post_one_wr(qp, wr, &call_send); 2194 if (unlikely(err)) { 2195 *bad_wr = wr; 2196 goto bail; 2197 } 2198 nreq++; 2199 } 2200 bail: 2201 spin_unlock_irqrestore(&qp->s_hlock, flags); 2202 if (nreq) { 2203 /* 2204 * Only call do_send if there is exactly one packet, and the 2205 * driver said it was ok. 2206 */ 2207 if (nreq == 1 && call_send) 2208 rdi->driver_f.do_send(qp); 2209 else 2210 rdi->driver_f.schedule_send_no_lock(qp); 2211 } 2212 return err; 2213 } 2214 2215 /** 2216 * rvt_post_srq_receive - post a receive on a shared receive queue 2217 * @ibsrq: the SRQ to post the receive on 2218 * @wr: the list of work requests to post 2219 * @bad_wr: A pointer to the first WR to cause a problem is put here 2220 * 2221 * This may be called from interrupt context. 2222 * 2223 * Return: 0 on success else errno 2224 */ 2225 int rvt_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr, 2226 const struct ib_recv_wr **bad_wr) 2227 { 2228 struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq); 2229 struct rvt_krwq *wq; 2230 unsigned long flags; 2231 2232 for (; wr; wr = wr->next) { 2233 struct rvt_rwqe *wqe; 2234 u32 next; 2235 int i; 2236 2237 if ((unsigned)wr->num_sge > srq->rq.max_sge) { 2238 *bad_wr = wr; 2239 return -EINVAL; 2240 } 2241 2242 spin_lock_irqsave(&srq->rq.kwq->p_lock, flags); 2243 wq = srq->rq.kwq; 2244 next = wq->head + 1; 2245 if (next >= srq->rq.size) 2246 next = 0; 2247 if (next == READ_ONCE(wq->tail)) { 2248 spin_unlock_irqrestore(&srq->rq.kwq->p_lock, flags); 2249 *bad_wr = wr; 2250 return -ENOMEM; 2251 } 2252 2253 wqe = rvt_get_rwqe_ptr(&srq->rq, wq->head); 2254 wqe->wr_id = wr->wr_id; 2255 wqe->num_sge = wr->num_sge; 2256 for (i = 0; i < wr->num_sge; i++) { 2257 wqe->sg_list[i].addr = wr->sg_list[i].addr; 2258 wqe->sg_list[i].length = wr->sg_list[i].length; 2259 wqe->sg_list[i].lkey = wr->sg_list[i].lkey; 2260 } 2261 /* Make sure queue entry is written before the head index. */ 2262 smp_store_release(&wq->head, next); 2263 spin_unlock_irqrestore(&srq->rq.kwq->p_lock, flags); 2264 } 2265 return 0; 2266 } 2267 2268 /* 2269 * rvt used the internal kernel struct as part of its ABI, for now make sure 2270 * the kernel struct does not change layout. FIXME: rvt should never cast the 2271 * user struct to a kernel struct. 2272 */ 2273 static struct ib_sge *rvt_cast_sge(struct rvt_wqe_sge *sge) 2274 { 2275 BUILD_BUG_ON(offsetof(struct ib_sge, addr) != 2276 offsetof(struct rvt_wqe_sge, addr)); 2277 BUILD_BUG_ON(offsetof(struct ib_sge, length) != 2278 offsetof(struct rvt_wqe_sge, length)); 2279 BUILD_BUG_ON(offsetof(struct ib_sge, lkey) != 2280 offsetof(struct rvt_wqe_sge, lkey)); 2281 return (struct ib_sge *)sge; 2282 } 2283 2284 /* 2285 * Validate a RWQE and fill in the SGE state. 2286 * Return 1 if OK. 2287 */ 2288 static int init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe) 2289 { 2290 int i, j, ret; 2291 struct ib_wc wc; 2292 struct rvt_lkey_table *rkt; 2293 struct rvt_pd *pd; 2294 struct rvt_sge_state *ss; 2295 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 2296 2297 rkt = &rdi->lkey_table; 2298 pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd); 2299 ss = &qp->r_sge; 2300 ss->sg_list = qp->r_sg_list; 2301 qp->r_len = 0; 2302 for (i = j = 0; i < wqe->num_sge; i++) { 2303 if (wqe->sg_list[i].length == 0) 2304 continue; 2305 /* Check LKEY */ 2306 ret = rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge, 2307 NULL, rvt_cast_sge(&wqe->sg_list[i]), 2308 IB_ACCESS_LOCAL_WRITE); 2309 if (unlikely(ret <= 0)) 2310 goto bad_lkey; 2311 qp->r_len += wqe->sg_list[i].length; 2312 j++; 2313 } 2314 ss->num_sge = j; 2315 ss->total_len = qp->r_len; 2316 return 1; 2317 2318 bad_lkey: 2319 while (j) { 2320 struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge; 2321 2322 rvt_put_mr(sge->mr); 2323 } 2324 ss->num_sge = 0; 2325 memset(&wc, 0, sizeof(wc)); 2326 wc.wr_id = wqe->wr_id; 2327 wc.status = IB_WC_LOC_PROT_ERR; 2328 wc.opcode = IB_WC_RECV; 2329 wc.qp = &qp->ibqp; 2330 /* Signal solicited completion event. */ 2331 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); 2332 return 0; 2333 } 2334 2335 /** 2336 * get_count - count numbers of request work queue entries 2337 * in circular buffer 2338 * @rq: data structure for request queue entry 2339 * @tail: tail indices of the circular buffer 2340 * @head: head indices of the circular buffer 2341 * 2342 * Return - total number of entries in the circular buffer 2343 */ 2344 static u32 get_count(struct rvt_rq *rq, u32 tail, u32 head) 2345 { 2346 u32 count; 2347 2348 count = head; 2349 2350 if (count >= rq->size) 2351 count = 0; 2352 if (count < tail) 2353 count += rq->size - tail; 2354 else 2355 count -= tail; 2356 2357 return count; 2358 } 2359 2360 /** 2361 * get_rvt_head - get head indices of the circular buffer 2362 * @rq: data structure for request queue entry 2363 * @ip: the QP 2364 * 2365 * Return - head index value 2366 */ 2367 static inline u32 get_rvt_head(struct rvt_rq *rq, void *ip) 2368 { 2369 u32 head; 2370 2371 if (ip) 2372 head = RDMA_READ_UAPI_ATOMIC(rq->wq->head); 2373 else 2374 head = rq->kwq->head; 2375 2376 return head; 2377 } 2378 2379 /** 2380 * rvt_get_rwqe - copy the next RWQE into the QP's RWQE 2381 * @qp: the QP 2382 * @wr_id_only: update qp->r_wr_id only, not qp->r_sge 2383 * 2384 * Return -1 if there is a local error, 0 if no RWQE is available, 2385 * otherwise return 1. 2386 * 2387 * Can be called from interrupt level. 2388 */ 2389 int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only) 2390 { 2391 unsigned long flags; 2392 struct rvt_rq *rq; 2393 struct rvt_krwq *kwq = NULL; 2394 struct rvt_rwq *wq; 2395 struct rvt_srq *srq; 2396 struct rvt_rwqe *wqe; 2397 void (*handler)(struct ib_event *, void *); 2398 u32 tail; 2399 u32 head; 2400 int ret; 2401 void *ip = NULL; 2402 2403 if (qp->ibqp.srq) { 2404 srq = ibsrq_to_rvtsrq(qp->ibqp.srq); 2405 handler = srq->ibsrq.event_handler; 2406 rq = &srq->rq; 2407 ip = srq->ip; 2408 } else { 2409 srq = NULL; 2410 handler = NULL; 2411 rq = &qp->r_rq; 2412 ip = qp->ip; 2413 } 2414 2415 spin_lock_irqsave(&rq->kwq->c_lock, flags); 2416 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) { 2417 ret = 0; 2418 goto unlock; 2419 } 2420 kwq = rq->kwq; 2421 if (ip) { 2422 wq = rq->wq; 2423 tail = RDMA_READ_UAPI_ATOMIC(wq->tail); 2424 } else { 2425 tail = kwq->tail; 2426 } 2427 2428 /* Validate tail before using it since it is user writable. */ 2429 if (tail >= rq->size) 2430 tail = 0; 2431 2432 if (kwq->count < RVT_RWQ_COUNT_THRESHOLD) { 2433 head = get_rvt_head(rq, ip); 2434 kwq->count = get_count(rq, tail, head); 2435 } 2436 if (unlikely(kwq->count == 0)) { 2437 ret = 0; 2438 goto unlock; 2439 } 2440 /* Make sure entry is read after the count is read. */ 2441 smp_rmb(); 2442 wqe = rvt_get_rwqe_ptr(rq, tail); 2443 /* 2444 * Even though we update the tail index in memory, the verbs 2445 * consumer is not supposed to post more entries until a 2446 * completion is generated. 2447 */ 2448 if (++tail >= rq->size) 2449 tail = 0; 2450 if (ip) 2451 RDMA_WRITE_UAPI_ATOMIC(wq->tail, tail); 2452 else 2453 kwq->tail = tail; 2454 if (!wr_id_only && !init_sge(qp, wqe)) { 2455 ret = -1; 2456 goto unlock; 2457 } 2458 qp->r_wr_id = wqe->wr_id; 2459 2460 kwq->count--; 2461 ret = 1; 2462 set_bit(RVT_R_WRID_VALID, &qp->r_aflags); 2463 if (handler) { 2464 /* 2465 * Validate head pointer value and compute 2466 * the number of remaining WQEs. 2467 */ 2468 if (kwq->count < srq->limit) { 2469 kwq->count = get_count(rq, tail, get_rvt_head(rq, ip)); 2470 if (kwq->count < srq->limit) { 2471 struct ib_event ev; 2472 2473 srq->limit = 0; 2474 spin_unlock_irqrestore(&rq->kwq->c_lock, flags); 2475 ev.device = qp->ibqp.device; 2476 ev.element.srq = qp->ibqp.srq; 2477 ev.event = IB_EVENT_SRQ_LIMIT_REACHED; 2478 handler(&ev, srq->ibsrq.srq_context); 2479 goto bail; 2480 } 2481 } 2482 } 2483 unlock: 2484 spin_unlock_irqrestore(&rq->kwq->c_lock, flags); 2485 bail: 2486 return ret; 2487 } 2488 EXPORT_SYMBOL(rvt_get_rwqe); 2489 2490 /** 2491 * qp_comm_est - handle trap with QP established 2492 * @qp: the QP 2493 */ 2494 void rvt_comm_est(struct rvt_qp *qp) 2495 { 2496 qp->r_flags |= RVT_R_COMM_EST; 2497 if (qp->ibqp.event_handler) { 2498 struct ib_event ev; 2499 2500 ev.device = qp->ibqp.device; 2501 ev.element.qp = &qp->ibqp; 2502 ev.event = IB_EVENT_COMM_EST; 2503 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 2504 } 2505 } 2506 EXPORT_SYMBOL(rvt_comm_est); 2507 2508 void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err) 2509 { 2510 unsigned long flags; 2511 int lastwqe; 2512 2513 spin_lock_irqsave(&qp->s_lock, flags); 2514 lastwqe = rvt_error_qp(qp, err); 2515 spin_unlock_irqrestore(&qp->s_lock, flags); 2516 2517 if (lastwqe) { 2518 struct ib_event ev; 2519 2520 ev.device = qp->ibqp.device; 2521 ev.element.qp = &qp->ibqp; 2522 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 2523 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); 2524 } 2525 } 2526 EXPORT_SYMBOL(rvt_rc_error); 2527 2528 /* 2529 * rvt_rnr_tbl_to_usec - return index into ib_rvt_rnr_table 2530 * @index - the index 2531 * return usec from an index into ib_rvt_rnr_table 2532 */ 2533 unsigned long rvt_rnr_tbl_to_usec(u32 index) 2534 { 2535 return ib_rvt_rnr_table[(index & IB_AETH_CREDIT_MASK)]; 2536 } 2537 EXPORT_SYMBOL(rvt_rnr_tbl_to_usec); 2538 2539 static inline unsigned long rvt_aeth_to_usec(u32 aeth) 2540 { 2541 return ib_rvt_rnr_table[(aeth >> IB_AETH_CREDIT_SHIFT) & 2542 IB_AETH_CREDIT_MASK]; 2543 } 2544 2545 /* 2546 * rvt_add_retry_timer_ext - add/start a retry timer 2547 * @qp - the QP 2548 * @shift - timeout shift to wait for multiple packets 2549 * add a retry timer on the QP 2550 */ 2551 void rvt_add_retry_timer_ext(struct rvt_qp *qp, u8 shift) 2552 { 2553 struct ib_qp *ibqp = &qp->ibqp; 2554 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); 2555 2556 lockdep_assert_held(&qp->s_lock); 2557 qp->s_flags |= RVT_S_TIMER; 2558 /* 4.096 usec. * (1 << qp->timeout) */ 2559 qp->s_timer.expires = jiffies + rdi->busy_jiffies + 2560 (qp->timeout_jiffies << shift); 2561 add_timer(&qp->s_timer); 2562 } 2563 EXPORT_SYMBOL(rvt_add_retry_timer_ext); 2564 2565 /** 2566 * rvt_add_rnr_timer - add/start an rnr timer 2567 * @qp - the QP 2568 * @aeth - aeth of RNR timeout, simulated aeth for loopback 2569 * add an rnr timer on the QP 2570 */ 2571 void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth) 2572 { 2573 u32 to; 2574 2575 lockdep_assert_held(&qp->s_lock); 2576 qp->s_flags |= RVT_S_WAIT_RNR; 2577 to = rvt_aeth_to_usec(aeth); 2578 trace_rvt_rnrnak_add(qp, to); 2579 hrtimer_start(&qp->s_rnr_timer, 2580 ns_to_ktime(1000 * to), HRTIMER_MODE_REL_PINNED); 2581 } 2582 EXPORT_SYMBOL(rvt_add_rnr_timer); 2583 2584 /** 2585 * rvt_stop_rc_timers - stop all timers 2586 * @qp - the QP 2587 * stop any pending timers 2588 */ 2589 void rvt_stop_rc_timers(struct rvt_qp *qp) 2590 { 2591 lockdep_assert_held(&qp->s_lock); 2592 /* Remove QP from all timers */ 2593 if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) { 2594 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR); 2595 del_timer(&qp->s_timer); 2596 hrtimer_try_to_cancel(&qp->s_rnr_timer); 2597 } 2598 } 2599 EXPORT_SYMBOL(rvt_stop_rc_timers); 2600 2601 /** 2602 * rvt_stop_rnr_timer - stop an rnr timer 2603 * @qp - the QP 2604 * 2605 * stop an rnr timer and return if the timer 2606 * had been pending. 2607 */ 2608 static void rvt_stop_rnr_timer(struct rvt_qp *qp) 2609 { 2610 lockdep_assert_held(&qp->s_lock); 2611 /* Remove QP from rnr timer */ 2612 if (qp->s_flags & RVT_S_WAIT_RNR) { 2613 qp->s_flags &= ~RVT_S_WAIT_RNR; 2614 trace_rvt_rnrnak_stop(qp, 0); 2615 } 2616 } 2617 2618 /** 2619 * rvt_del_timers_sync - wait for any timeout routines to exit 2620 * @qp - the QP 2621 */ 2622 void rvt_del_timers_sync(struct rvt_qp *qp) 2623 { 2624 del_timer_sync(&qp->s_timer); 2625 hrtimer_cancel(&qp->s_rnr_timer); 2626 } 2627 EXPORT_SYMBOL(rvt_del_timers_sync); 2628 2629 /** 2630 * This is called from s_timer for missing responses. 2631 */ 2632 static void rvt_rc_timeout(struct timer_list *t) 2633 { 2634 struct rvt_qp *qp = from_timer(qp, t, s_timer); 2635 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 2636 unsigned long flags; 2637 2638 spin_lock_irqsave(&qp->r_lock, flags); 2639 spin_lock(&qp->s_lock); 2640 if (qp->s_flags & RVT_S_TIMER) { 2641 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; 2642 2643 qp->s_flags &= ~RVT_S_TIMER; 2644 rvp->n_rc_timeouts++; 2645 del_timer(&qp->s_timer); 2646 trace_rvt_rc_timeout(qp, qp->s_last_psn + 1); 2647 if (rdi->driver_f.notify_restart_rc) 2648 rdi->driver_f.notify_restart_rc(qp, 2649 qp->s_last_psn + 1, 2650 1); 2651 rdi->driver_f.schedule_send(qp); 2652 } 2653 spin_unlock(&qp->s_lock); 2654 spin_unlock_irqrestore(&qp->r_lock, flags); 2655 } 2656 2657 /* 2658 * This is called from s_timer for RNR timeouts. 2659 */ 2660 enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t) 2661 { 2662 struct rvt_qp *qp = container_of(t, struct rvt_qp, s_rnr_timer); 2663 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 2664 unsigned long flags; 2665 2666 spin_lock_irqsave(&qp->s_lock, flags); 2667 rvt_stop_rnr_timer(qp); 2668 trace_rvt_rnrnak_timeout(qp, 0); 2669 rdi->driver_f.schedule_send(qp); 2670 spin_unlock_irqrestore(&qp->s_lock, flags); 2671 return HRTIMER_NORESTART; 2672 } 2673 EXPORT_SYMBOL(rvt_rc_rnr_retry); 2674 2675 /** 2676 * rvt_qp_iter_init - initial for QP iteration 2677 * @rdi: rvt devinfo 2678 * @v: u64 value 2679 * 2680 * This returns an iterator suitable for iterating QPs 2681 * in the system. 2682 * 2683 * The @cb is a user defined callback and @v is a 64 2684 * bit value passed to and relevant for processing in the 2685 * @cb. An example use case would be to alter QP processing 2686 * based on criteria not part of the rvt_qp. 2687 * 2688 * Use cases that require memory allocation to succeed 2689 * must preallocate appropriately. 2690 * 2691 * Return: a pointer to an rvt_qp_iter or NULL 2692 */ 2693 struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi, 2694 u64 v, 2695 void (*cb)(struct rvt_qp *qp, u64 v)) 2696 { 2697 struct rvt_qp_iter *i; 2698 2699 i = kzalloc(sizeof(*i), GFP_KERNEL); 2700 if (!i) 2701 return NULL; 2702 2703 i->rdi = rdi; 2704 /* number of special QPs (SMI/GSI) for device */ 2705 i->specials = rdi->ibdev.phys_port_cnt * 2; 2706 i->v = v; 2707 i->cb = cb; 2708 2709 return i; 2710 } 2711 EXPORT_SYMBOL(rvt_qp_iter_init); 2712 2713 /** 2714 * rvt_qp_iter_next - return the next QP in iter 2715 * @iter - the iterator 2716 * 2717 * Fine grained QP iterator suitable for use 2718 * with debugfs seq_file mechanisms. 2719 * 2720 * Updates iter->qp with the current QP when the return 2721 * value is 0. 2722 * 2723 * Return: 0 - iter->qp is valid 1 - no more QPs 2724 */ 2725 int rvt_qp_iter_next(struct rvt_qp_iter *iter) 2726 __must_hold(RCU) 2727 { 2728 int n = iter->n; 2729 int ret = 1; 2730 struct rvt_qp *pqp = iter->qp; 2731 struct rvt_qp *qp; 2732 struct rvt_dev_info *rdi = iter->rdi; 2733 2734 /* 2735 * The approach is to consider the special qps 2736 * as additional table entries before the 2737 * real hash table. Since the qp code sets 2738 * the qp->next hash link to NULL, this works just fine. 2739 * 2740 * iter->specials is 2 * # ports 2741 * 2742 * n = 0..iter->specials is the special qp indices 2743 * 2744 * n = iter->specials..rdi->qp_dev->qp_table_size+iter->specials are 2745 * the potential hash bucket entries 2746 * 2747 */ 2748 for (; n < rdi->qp_dev->qp_table_size + iter->specials; n++) { 2749 if (pqp) { 2750 qp = rcu_dereference(pqp->next); 2751 } else { 2752 if (n < iter->specials) { 2753 struct rvt_ibport *rvp; 2754 int pidx; 2755 2756 pidx = n % rdi->ibdev.phys_port_cnt; 2757 rvp = rdi->ports[pidx]; 2758 qp = rcu_dereference(rvp->qp[n & 1]); 2759 } else { 2760 qp = rcu_dereference( 2761 rdi->qp_dev->qp_table[ 2762 (n - iter->specials)]); 2763 } 2764 } 2765 pqp = qp; 2766 if (qp) { 2767 iter->qp = qp; 2768 iter->n = n; 2769 return 0; 2770 } 2771 } 2772 return ret; 2773 } 2774 EXPORT_SYMBOL(rvt_qp_iter_next); 2775 2776 /** 2777 * rvt_qp_iter - iterate all QPs 2778 * @rdi - rvt devinfo 2779 * @v - a 64 bit value 2780 * @cb - a callback 2781 * 2782 * This provides a way for iterating all QPs. 2783 * 2784 * The @cb is a user defined callback and @v is a 64 2785 * bit value passed to and relevant for processing in the 2786 * cb. An example use case would be to alter QP processing 2787 * based on criteria not part of the rvt_qp. 2788 * 2789 * The code has an internal iterator to simplify 2790 * non seq_file use cases. 2791 */ 2792 void rvt_qp_iter(struct rvt_dev_info *rdi, 2793 u64 v, 2794 void (*cb)(struct rvt_qp *qp, u64 v)) 2795 { 2796 int ret; 2797 struct rvt_qp_iter i = { 2798 .rdi = rdi, 2799 .specials = rdi->ibdev.phys_port_cnt * 2, 2800 .v = v, 2801 .cb = cb 2802 }; 2803 2804 rcu_read_lock(); 2805 do { 2806 ret = rvt_qp_iter_next(&i); 2807 if (!ret) { 2808 rvt_get_qp(i.qp); 2809 rcu_read_unlock(); 2810 i.cb(i.qp, i.v); 2811 rcu_read_lock(); 2812 rvt_put_qp(i.qp); 2813 } 2814 } while (!ret); 2815 rcu_read_unlock(); 2816 } 2817 EXPORT_SYMBOL(rvt_qp_iter); 2818 2819 /* 2820 * This should be called with s_lock held. 2821 */ 2822 void rvt_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe, 2823 enum ib_wc_status status) 2824 { 2825 u32 old_last, last; 2826 struct rvt_dev_info *rdi; 2827 2828 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND)) 2829 return; 2830 rdi = ib_to_rvt(qp->ibqp.device); 2831 2832 old_last = qp->s_last; 2833 trace_rvt_qp_send_completion(qp, wqe, old_last); 2834 last = rvt_qp_complete_swqe(qp, wqe, rdi->wc_opcode[wqe->wr.opcode], 2835 status); 2836 if (qp->s_acked == old_last) 2837 qp->s_acked = last; 2838 if (qp->s_cur == old_last) 2839 qp->s_cur = last; 2840 if (qp->s_tail == old_last) 2841 qp->s_tail = last; 2842 if (qp->state == IB_QPS_SQD && last == qp->s_cur) 2843 qp->s_draining = 0; 2844 } 2845 EXPORT_SYMBOL(rvt_send_complete); 2846 2847 /** 2848 * rvt_copy_sge - copy data to SGE memory 2849 * @qp: associated QP 2850 * @ss: the SGE state 2851 * @data: the data to copy 2852 * @length: the length of the data 2853 * @release: boolean to release MR 2854 * @copy_last: do a separate copy of the last 8 bytes 2855 */ 2856 void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss, 2857 void *data, u32 length, 2858 bool release, bool copy_last) 2859 { 2860 struct rvt_sge *sge = &ss->sge; 2861 int i; 2862 bool in_last = false; 2863 bool cacheless_copy = false; 2864 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); 2865 struct rvt_wss *wss = rdi->wss; 2866 unsigned int sge_copy_mode = rdi->dparms.sge_copy_mode; 2867 2868 if (sge_copy_mode == RVT_SGE_COPY_CACHELESS) { 2869 cacheless_copy = length >= PAGE_SIZE; 2870 } else if (sge_copy_mode == RVT_SGE_COPY_ADAPTIVE) { 2871 if (length >= PAGE_SIZE) { 2872 /* 2873 * NOTE: this *assumes*: 2874 * o The first vaddr is the dest. 2875 * o If multiple pages, then vaddr is sequential. 2876 */ 2877 wss_insert(wss, sge->vaddr); 2878 if (length >= (2 * PAGE_SIZE)) 2879 wss_insert(wss, (sge->vaddr + PAGE_SIZE)); 2880 2881 cacheless_copy = wss_exceeds_threshold(wss); 2882 } else { 2883 wss_advance_clean_counter(wss); 2884 } 2885 } 2886 2887 if (copy_last) { 2888 if (length > 8) { 2889 length -= 8; 2890 } else { 2891 copy_last = false; 2892 in_last = true; 2893 } 2894 } 2895 2896 again: 2897 while (length) { 2898 u32 len = rvt_get_sge_length(sge, length); 2899 2900 WARN_ON_ONCE(len == 0); 2901 if (unlikely(in_last)) { 2902 /* enforce byte transfer ordering */ 2903 for (i = 0; i < len; i++) 2904 ((u8 *)sge->vaddr)[i] = ((u8 *)data)[i]; 2905 } else if (cacheless_copy) { 2906 cacheless_memcpy(sge->vaddr, data, len); 2907 } else { 2908 memcpy(sge->vaddr, data, len); 2909 } 2910 rvt_update_sge(ss, len, release); 2911 data += len; 2912 length -= len; 2913 } 2914 2915 if (copy_last) { 2916 copy_last = false; 2917 in_last = true; 2918 length = 8; 2919 goto again; 2920 } 2921 } 2922 EXPORT_SYMBOL(rvt_copy_sge); 2923 2924 static enum ib_wc_status loopback_qp_drop(struct rvt_ibport *rvp, 2925 struct rvt_qp *sqp) 2926 { 2927 rvp->n_pkt_drops++; 2928 /* 2929 * For RC, the requester would timeout and retry so 2930 * shortcut the timeouts and just signal too many retries. 2931 */ 2932 return sqp->ibqp.qp_type == IB_QPT_RC ? 2933 IB_WC_RETRY_EXC_ERR : IB_WC_SUCCESS; 2934 } 2935 2936 /** 2937 * ruc_loopback - handle UC and RC loopback requests 2938 * @sqp: the sending QP 2939 * 2940 * This is called from rvt_do_send() to forward a WQE addressed to the same HFI 2941 * Note that although we are single threaded due to the send engine, we still 2942 * have to protect against post_send(). We don't have to worry about 2943 * receive interrupts since this is a connected protocol and all packets 2944 * will pass through here. 2945 */ 2946 void rvt_ruc_loopback(struct rvt_qp *sqp) 2947 { 2948 struct rvt_ibport *rvp = NULL; 2949 struct rvt_dev_info *rdi = ib_to_rvt(sqp->ibqp.device); 2950 struct rvt_qp *qp; 2951 struct rvt_swqe *wqe; 2952 struct rvt_sge *sge; 2953 unsigned long flags; 2954 struct ib_wc wc; 2955 u64 sdata; 2956 atomic64_t *maddr; 2957 enum ib_wc_status send_status; 2958 bool release; 2959 int ret; 2960 bool copy_last = false; 2961 int local_ops = 0; 2962 2963 rcu_read_lock(); 2964 rvp = rdi->ports[sqp->port_num - 1]; 2965 2966 /* 2967 * Note that we check the responder QP state after 2968 * checking the requester's state. 2969 */ 2970 2971 qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), rvp, 2972 sqp->remote_qpn); 2973 2974 spin_lock_irqsave(&sqp->s_lock, flags); 2975 2976 /* Return if we are already busy processing a work request. */ 2977 if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) || 2978 !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND)) 2979 goto unlock; 2980 2981 sqp->s_flags |= RVT_S_BUSY; 2982 2983 again: 2984 if (sqp->s_last == READ_ONCE(sqp->s_head)) 2985 goto clr_busy; 2986 wqe = rvt_get_swqe_ptr(sqp, sqp->s_last); 2987 2988 /* Return if it is not OK to start a new work request. */ 2989 if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) { 2990 if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND)) 2991 goto clr_busy; 2992 /* We are in the error state, flush the work request. */ 2993 send_status = IB_WC_WR_FLUSH_ERR; 2994 goto flush_send; 2995 } 2996 2997 /* 2998 * We can rely on the entry not changing without the s_lock 2999 * being held until we update s_last. 3000 * We increment s_cur to indicate s_last is in progress. 3001 */ 3002 if (sqp->s_last == sqp->s_cur) { 3003 if (++sqp->s_cur >= sqp->s_size) 3004 sqp->s_cur = 0; 3005 } 3006 spin_unlock_irqrestore(&sqp->s_lock, flags); 3007 3008 if (!qp) { 3009 send_status = loopback_qp_drop(rvp, sqp); 3010 goto serr_no_r_lock; 3011 } 3012 spin_lock_irqsave(&qp->r_lock, flags); 3013 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) || 3014 qp->ibqp.qp_type != sqp->ibqp.qp_type) { 3015 send_status = loopback_qp_drop(rvp, sqp); 3016 goto serr; 3017 } 3018 3019 memset(&wc, 0, sizeof(wc)); 3020 send_status = IB_WC_SUCCESS; 3021 3022 release = true; 3023 sqp->s_sge.sge = wqe->sg_list[0]; 3024 sqp->s_sge.sg_list = wqe->sg_list + 1; 3025 sqp->s_sge.num_sge = wqe->wr.num_sge; 3026 sqp->s_len = wqe->length; 3027 switch (wqe->wr.opcode) { 3028 case IB_WR_REG_MR: 3029 goto send_comp; 3030 3031 case IB_WR_LOCAL_INV: 3032 if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) { 3033 if (rvt_invalidate_rkey(sqp, 3034 wqe->wr.ex.invalidate_rkey)) 3035 send_status = IB_WC_LOC_PROT_ERR; 3036 local_ops = 1; 3037 } 3038 goto send_comp; 3039 3040 case IB_WR_SEND_WITH_INV: 3041 case IB_WR_SEND_WITH_IMM: 3042 case IB_WR_SEND: 3043 ret = rvt_get_rwqe(qp, false); 3044 if (ret < 0) 3045 goto op_err; 3046 if (!ret) 3047 goto rnr_nak; 3048 if (wqe->length > qp->r_len) 3049 goto inv_err; 3050 switch (wqe->wr.opcode) { 3051 case IB_WR_SEND_WITH_INV: 3052 if (!rvt_invalidate_rkey(qp, 3053 wqe->wr.ex.invalidate_rkey)) { 3054 wc.wc_flags = IB_WC_WITH_INVALIDATE; 3055 wc.ex.invalidate_rkey = 3056 wqe->wr.ex.invalidate_rkey; 3057 } 3058 break; 3059 case IB_WR_SEND_WITH_IMM: 3060 wc.wc_flags = IB_WC_WITH_IMM; 3061 wc.ex.imm_data = wqe->wr.ex.imm_data; 3062 break; 3063 default: 3064 break; 3065 } 3066 break; 3067 3068 case IB_WR_RDMA_WRITE_WITH_IMM: 3069 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) 3070 goto inv_err; 3071 wc.wc_flags = IB_WC_WITH_IMM; 3072 wc.ex.imm_data = wqe->wr.ex.imm_data; 3073 ret = rvt_get_rwqe(qp, true); 3074 if (ret < 0) 3075 goto op_err; 3076 if (!ret) 3077 goto rnr_nak; 3078 /* skip copy_last set and qp_access_flags recheck */ 3079 goto do_write; 3080 case IB_WR_RDMA_WRITE: 3081 copy_last = rvt_is_user_qp(qp); 3082 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) 3083 goto inv_err; 3084 do_write: 3085 if (wqe->length == 0) 3086 break; 3087 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length, 3088 wqe->rdma_wr.remote_addr, 3089 wqe->rdma_wr.rkey, 3090 IB_ACCESS_REMOTE_WRITE))) 3091 goto acc_err; 3092 qp->r_sge.sg_list = NULL; 3093 qp->r_sge.num_sge = 1; 3094 qp->r_sge.total_len = wqe->length; 3095 break; 3096 3097 case IB_WR_RDMA_READ: 3098 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ))) 3099 goto inv_err; 3100 if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length, 3101 wqe->rdma_wr.remote_addr, 3102 wqe->rdma_wr.rkey, 3103 IB_ACCESS_REMOTE_READ))) 3104 goto acc_err; 3105 release = false; 3106 sqp->s_sge.sg_list = NULL; 3107 sqp->s_sge.num_sge = 1; 3108 qp->r_sge.sge = wqe->sg_list[0]; 3109 qp->r_sge.sg_list = wqe->sg_list + 1; 3110 qp->r_sge.num_sge = wqe->wr.num_sge; 3111 qp->r_sge.total_len = wqe->length; 3112 break; 3113 3114 case IB_WR_ATOMIC_CMP_AND_SWP: 3115 case IB_WR_ATOMIC_FETCH_AND_ADD: 3116 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC))) 3117 goto inv_err; 3118 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64), 3119 wqe->atomic_wr.remote_addr, 3120 wqe->atomic_wr.rkey, 3121 IB_ACCESS_REMOTE_ATOMIC))) 3122 goto acc_err; 3123 /* Perform atomic OP and save result. */ 3124 maddr = (atomic64_t *)qp->r_sge.sge.vaddr; 3125 sdata = wqe->atomic_wr.compare_add; 3126 *(u64 *)sqp->s_sge.sge.vaddr = 3127 (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ? 3128 (u64)atomic64_add_return(sdata, maddr) - sdata : 3129 (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr, 3130 sdata, wqe->atomic_wr.swap); 3131 rvt_put_mr(qp->r_sge.sge.mr); 3132 qp->r_sge.num_sge = 0; 3133 goto send_comp; 3134 3135 default: 3136 send_status = IB_WC_LOC_QP_OP_ERR; 3137 goto serr; 3138 } 3139 3140 sge = &sqp->s_sge.sge; 3141 while (sqp->s_len) { 3142 u32 len = rvt_get_sge_length(sge, sqp->s_len); 3143 3144 WARN_ON_ONCE(len == 0); 3145 rvt_copy_sge(qp, &qp->r_sge, sge->vaddr, 3146 len, release, copy_last); 3147 rvt_update_sge(&sqp->s_sge, len, !release); 3148 sqp->s_len -= len; 3149 } 3150 if (release) 3151 rvt_put_ss(&qp->r_sge); 3152 3153 if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) 3154 goto send_comp; 3155 3156 if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM) 3157 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM; 3158 else 3159 wc.opcode = IB_WC_RECV; 3160 wc.wr_id = qp->r_wr_id; 3161 wc.status = IB_WC_SUCCESS; 3162 wc.byte_len = wqe->length; 3163 wc.qp = &qp->ibqp; 3164 wc.src_qp = qp->remote_qpn; 3165 wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX; 3166 wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr); 3167 wc.port_num = 1; 3168 /* Signal completion event if the solicited bit is set. */ 3169 rvt_recv_cq(qp, &wc, wqe->wr.send_flags & IB_SEND_SOLICITED); 3170 3171 send_comp: 3172 spin_unlock_irqrestore(&qp->r_lock, flags); 3173 spin_lock_irqsave(&sqp->s_lock, flags); 3174 rvp->n_loop_pkts++; 3175 flush_send: 3176 sqp->s_rnr_retry = sqp->s_rnr_retry_cnt; 3177 rvt_send_complete(sqp, wqe, send_status); 3178 if (local_ops) { 3179 atomic_dec(&sqp->local_ops_pending); 3180 local_ops = 0; 3181 } 3182 goto again; 3183 3184 rnr_nak: 3185 /* Handle RNR NAK */ 3186 if (qp->ibqp.qp_type == IB_QPT_UC) 3187 goto send_comp; 3188 rvp->n_rnr_naks++; 3189 /* 3190 * Note: we don't need the s_lock held since the BUSY flag 3191 * makes this single threaded. 3192 */ 3193 if (sqp->s_rnr_retry == 0) { 3194 send_status = IB_WC_RNR_RETRY_EXC_ERR; 3195 goto serr; 3196 } 3197 if (sqp->s_rnr_retry_cnt < 7) 3198 sqp->s_rnr_retry--; 3199 spin_unlock_irqrestore(&qp->r_lock, flags); 3200 spin_lock_irqsave(&sqp->s_lock, flags); 3201 if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK)) 3202 goto clr_busy; 3203 rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer << 3204 IB_AETH_CREDIT_SHIFT); 3205 goto clr_busy; 3206 3207 op_err: 3208 send_status = IB_WC_REM_OP_ERR; 3209 wc.status = IB_WC_LOC_QP_OP_ERR; 3210 goto err; 3211 3212 inv_err: 3213 send_status = 3214 sqp->ibqp.qp_type == IB_QPT_RC ? 3215 IB_WC_REM_INV_REQ_ERR : 3216 IB_WC_SUCCESS; 3217 wc.status = IB_WC_LOC_QP_OP_ERR; 3218 goto err; 3219 3220 acc_err: 3221 send_status = IB_WC_REM_ACCESS_ERR; 3222 wc.status = IB_WC_LOC_PROT_ERR; 3223 err: 3224 /* responder goes to error state */ 3225 rvt_rc_error(qp, wc.status); 3226 3227 serr: 3228 spin_unlock_irqrestore(&qp->r_lock, flags); 3229 serr_no_r_lock: 3230 spin_lock_irqsave(&sqp->s_lock, flags); 3231 rvt_send_complete(sqp, wqe, send_status); 3232 if (sqp->ibqp.qp_type == IB_QPT_RC) { 3233 int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR); 3234 3235 sqp->s_flags &= ~RVT_S_BUSY; 3236 spin_unlock_irqrestore(&sqp->s_lock, flags); 3237 if (lastwqe) { 3238 struct ib_event ev; 3239 3240 ev.device = sqp->ibqp.device; 3241 ev.element.qp = &sqp->ibqp; 3242 ev.event = IB_EVENT_QP_LAST_WQE_REACHED; 3243 sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context); 3244 } 3245 goto done; 3246 } 3247 clr_busy: 3248 sqp->s_flags &= ~RVT_S_BUSY; 3249 unlock: 3250 spin_unlock_irqrestore(&sqp->s_lock, flags); 3251 done: 3252 rcu_read_unlock(); 3253 } 3254 EXPORT_SYMBOL(rvt_ruc_loopback); 3255