Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   1/*
   2 * Copyright(c) 2016 - 2020 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
  63static void rvt_rc_timeout(struct timer_list *t);
  64static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
  65			 enum ib_qp_type type);
  66
  67/*
  68 * Convert the AETH RNR timeout code into the number of microseconds.
  69 */
  70static const u32 ib_rvt_rnr_table[32] = {
  71	655360, /* 00: 655.36 */
  72	10,     /* 01:    .01 */
  73	20,     /* 02     .02 */
  74	30,     /* 03:    .03 */
  75	40,     /* 04:    .04 */
  76	60,     /* 05:    .06 */
  77	80,     /* 06:    .08 */
  78	120,    /* 07:    .12 */
  79	160,    /* 08:    .16 */
  80	240,    /* 09:    .24 */
  81	320,    /* 0A:    .32 */
  82	480,    /* 0B:    .48 */
  83	640,    /* 0C:    .64 */
  84	960,    /* 0D:    .96 */
  85	1280,   /* 0E:   1.28 */
  86	1920,   /* 0F:   1.92 */
  87	2560,   /* 10:   2.56 */
  88	3840,   /* 11:   3.84 */
  89	5120,   /* 12:   5.12 */
  90	7680,   /* 13:   7.68 */
  91	10240,  /* 14:  10.24 */
  92	15360,  /* 15:  15.36 */
  93	20480,  /* 16:  20.48 */
  94	30720,  /* 17:  30.72 */
  95	40960,  /* 18:  40.96 */
  96	61440,  /* 19:  61.44 */
  97	81920,  /* 1A:  81.92 */
  98	122880, /* 1B: 122.88 */
  99	163840, /* 1C: 163.84 */
 100	245760, /* 1D: 245.76 */
 101	327680, /* 1E: 327.68 */
 102	491520  /* 1F: 491.52 */
 103};
 104
 105/*
 106 * Note that it is OK to post send work requests in the SQE and ERR
 107 * states; rvt_do_send() will process them and generate error
 108 * completions as per IB 1.2 C10-96.
 109 */
 110const int ib_rvt_state_ops[IB_QPS_ERR + 1] = {
 111	[IB_QPS_RESET] = 0,
 112	[IB_QPS_INIT] = RVT_POST_RECV_OK,
 113	[IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK,
 114	[IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
 115	    RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK |
 116	    RVT_PROCESS_NEXT_SEND_OK,
 117	[IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
 118	    RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK,
 119	[IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
 120	    RVT_POST_SEND_OK | RVT_FLUSH_SEND,
 121	[IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV |
 122	    RVT_POST_SEND_OK | RVT_FLUSH_SEND,
 123};
 124EXPORT_SYMBOL(ib_rvt_state_ops);
 125
 126/* platform specific: return the last level cache (llc) size, in KiB */
 127static int rvt_wss_llc_size(void)
 128{
 129	/* assume that the boot CPU value is universal for all CPUs */
 130	return boot_cpu_data.x86_cache_size;
 131}
 132
 133/* platform specific: cacheless copy */
 134static void cacheless_memcpy(void *dst, void *src, size_t n)
 135{
 136	/*
 137	 * Use the only available X64 cacheless copy.  Add a __user cast
 138	 * to quiet sparse.  The src agument is already in the kernel so
 139	 * there are no security issues.  The extra fault recovery machinery
 140	 * is not invoked.
 141	 */
 142	__copy_user_nocache(dst, (void __user *)src, n, 0);
 143}
 144
 145void rvt_wss_exit(struct rvt_dev_info *rdi)
 146{
 147	struct rvt_wss *wss = rdi->wss;
 148
 149	if (!wss)
 150		return;
 151
 152	/* coded to handle partially initialized and repeat callers */
 153	kfree(wss->entries);
 154	wss->entries = NULL;
 155	kfree(rdi->wss);
 156	rdi->wss = NULL;
 157}
 158
 159/*
 160 * rvt_wss_init - Init wss data structures
 161 *
 162 * Return: 0 on success
 163 */
 164int rvt_wss_init(struct rvt_dev_info *rdi)
 165{
 166	unsigned int sge_copy_mode = rdi->dparms.sge_copy_mode;
 167	unsigned int wss_threshold = rdi->dparms.wss_threshold;
 168	unsigned int wss_clean_period = rdi->dparms.wss_clean_period;
 169	long llc_size;
 170	long llc_bits;
 171	long table_size;
 172	long table_bits;
 173	struct rvt_wss *wss;
 174	int node = rdi->dparms.node;
 175
 176	if (sge_copy_mode != RVT_SGE_COPY_ADAPTIVE) {
 177		rdi->wss = NULL;
 178		return 0;
 179	}
 180
 181	rdi->wss = kzalloc_node(sizeof(*rdi->wss), GFP_KERNEL, node);
 182	if (!rdi->wss)
 183		return -ENOMEM;
 184	wss = rdi->wss;
 185
 186	/* check for a valid percent range - default to 80 if none or invalid */
 187	if (wss_threshold < 1 || wss_threshold > 100)
 188		wss_threshold = 80;
 189
 190	/* reject a wildly large period */
 191	if (wss_clean_period > 1000000)
 192		wss_clean_period = 256;
 193
 194	/* reject a zero period */
 195	if (wss_clean_period == 0)
 196		wss_clean_period = 1;
 197
 198	/*
 199	 * Calculate the table size - the next power of 2 larger than the
 200	 * LLC size.  LLC size is in KiB.
 201	 */
 202	llc_size = rvt_wss_llc_size() * 1024;
 203	table_size = roundup_pow_of_two(llc_size);
 204
 205	/* one bit per page in rounded up table */
 206	llc_bits = llc_size / PAGE_SIZE;
 207	table_bits = table_size / PAGE_SIZE;
 208	wss->pages_mask = table_bits - 1;
 209	wss->num_entries = table_bits / BITS_PER_LONG;
 210
 211	wss->threshold = (llc_bits * wss_threshold) / 100;
 212	if (wss->threshold == 0)
 213		wss->threshold = 1;
 214
 215	wss->clean_period = wss_clean_period;
 216	atomic_set(&wss->clean_counter, wss_clean_period);
 217
 218	wss->entries = kcalloc_node(wss->num_entries, sizeof(*wss->entries),
 219				    GFP_KERNEL, node);
 220	if (!wss->entries) {
 221		rvt_wss_exit(rdi);
 222		return -ENOMEM;
 223	}
 224
 225	return 0;
 226}
 227
 228/*
 229 * Advance the clean counter.  When the clean period has expired,
 230 * clean an entry.
 231 *
 232 * This is implemented in atomics to avoid locking.  Because multiple
 233 * variables are involved, it can be racy which can lead to slightly
 234 * inaccurate information.  Since this is only a heuristic, this is
 235 * OK.  Any innaccuracies will clean themselves out as the counter
 236 * advances.  That said, it is unlikely the entry clean operation will
 237 * race - the next possible racer will not start until the next clean
 238 * period.
 239 *
 240 * The clean counter is implemented as a decrement to zero.  When zero
 241 * is reached an entry is cleaned.
 242 */
 243static void wss_advance_clean_counter(struct rvt_wss *wss)
 244{
 245	int entry;
 246	int weight;
 247	unsigned long bits;
 248
 249	/* become the cleaner if we decrement the counter to zero */
 250	if (atomic_dec_and_test(&wss->clean_counter)) {
 251		/*
 252		 * Set, not add, the clean period.  This avoids an issue
 253		 * where the counter could decrement below the clean period.
 254		 * Doing a set can result in lost decrements, slowing the
 255		 * clean advance.  Since this a heuristic, this possible
 256		 * slowdown is OK.
 257		 *
 258		 * An alternative is to loop, advancing the counter by a
 259		 * clean period until the result is > 0. However, this could
 260		 * lead to several threads keeping another in the clean loop.
 261		 * This could be mitigated by limiting the number of times
 262		 * we stay in the loop.
 263		 */
 264		atomic_set(&wss->clean_counter, wss->clean_period);
 265
 266		/*
 267		 * Uniquely grab the entry to clean and move to next.
 268		 * The current entry is always the lower bits of
 269		 * wss.clean_entry.  The table size, wss.num_entries,
 270		 * is always a power-of-2.
 271		 */
 272		entry = (atomic_inc_return(&wss->clean_entry) - 1)
 273			& (wss->num_entries - 1);
 274
 275		/* clear the entry and count the bits */
 276		bits = xchg(&wss->entries[entry], 0);
 277		weight = hweight64((u64)bits);
 278		/* only adjust the contended total count if needed */
 279		if (weight)
 280			atomic_sub(weight, &wss->total_count);
 281	}
 282}
 283
 284/*
 285 * Insert the given address into the working set array.
 286 */
 287static void wss_insert(struct rvt_wss *wss, void *address)
 288{
 289	u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss->pages_mask;
 290	u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */
 291	u32 nr = page & (BITS_PER_LONG - 1);
 292
 293	if (!test_and_set_bit(nr, &wss->entries[entry]))
 294		atomic_inc(&wss->total_count);
 295
 296	wss_advance_clean_counter(wss);
 297}
 298
 299/*
 300 * Is the working set larger than the threshold?
 301 */
 302static inline bool wss_exceeds_threshold(struct rvt_wss *wss)
 303{
 304	return atomic_read(&wss->total_count) >= wss->threshold;
 305}
 306
 307static void get_map_page(struct rvt_qpn_table *qpt,
 308			 struct rvt_qpn_map *map)
 309{
 310	unsigned long page = get_zeroed_page(GFP_KERNEL);
 311
 312	/*
 313	 * Free the page if someone raced with us installing it.
 314	 */
 315
 316	spin_lock(&qpt->lock);
 317	if (map->page)
 318		free_page(page);
 319	else
 320		map->page = (void *)page;
 321	spin_unlock(&qpt->lock);
 322}
 323
 324/**
 325 * init_qpn_table - initialize the QP number table for a device
 326 * @rdi: rvt dev struct
 327 * @qpt: the QPN table
 328 */
 329static int init_qpn_table(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt)
 330{
 331	u32 offset, i;
 332	struct rvt_qpn_map *map;
 333	int ret = 0;
 334
 335	if (!(rdi->dparms.qpn_res_end >= rdi->dparms.qpn_res_start))
 336		return -EINVAL;
 337
 338	spin_lock_init(&qpt->lock);
 339
 340	qpt->last = rdi->dparms.qpn_start;
 341	qpt->incr = rdi->dparms.qpn_inc << rdi->dparms.qos_shift;
 342
 343	/*
 344	 * Drivers may want some QPs beyond what we need for verbs let them use
 345	 * our qpn table. No need for two. Lets go ahead and mark the bitmaps
 346	 * for those. The reserved range must be *after* the range which verbs
 347	 * will pick from.
 348	 */
 349
 350	/* Figure out number of bit maps needed before reserved range */
 351	qpt->nmaps = rdi->dparms.qpn_res_start / RVT_BITS_PER_PAGE;
 352
 353	/* This should always be zero */
 354	offset = rdi->dparms.qpn_res_start & RVT_BITS_PER_PAGE_MASK;
 355
 356	/* Starting with the first reserved bit map */
 357	map = &qpt->map[qpt->nmaps];
 358
 359	rvt_pr_info(rdi, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n",
 360		    rdi->dparms.qpn_res_start, rdi->dparms.qpn_res_end);
 361	for (i = rdi->dparms.qpn_res_start; i <= rdi->dparms.qpn_res_end; i++) {
 362		if (!map->page) {
 363			get_map_page(qpt, map);
 364			if (!map->page) {
 365				ret = -ENOMEM;
 366				break;
 367			}
 368		}
 369		set_bit(offset, map->page);
 370		offset++;
 371		if (offset == RVT_BITS_PER_PAGE) {
 372			/* next page */
 373			qpt->nmaps++;
 374			map++;
 375			offset = 0;
 376		}
 377	}
 378	return ret;
 379}
 380
 381/**
 382 * free_qpn_table - free the QP number table for a device
 383 * @qpt: the QPN table
 384 */
 385static void free_qpn_table(struct rvt_qpn_table *qpt)
 386{
 387	int i;
 388
 389	for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
 390		free_page((unsigned long)qpt->map[i].page);
 391}
 392
 393/**
 394 * rvt_driver_qp_init - Init driver qp resources
 395 * @rdi: rvt dev strucutre
 396 *
 397 * Return: 0 on success
 398 */
 399int rvt_driver_qp_init(struct rvt_dev_info *rdi)
 400{
 401	int i;
 402	int ret = -ENOMEM;
 403
 404	if (!rdi->dparms.qp_table_size)
 405		return -EINVAL;
 406
 407	/*
 408	 * If driver is not doing any QP allocation then make sure it is
 409	 * providing the necessary QP functions.
 410	 */
 411	if (!rdi->driver_f.free_all_qps ||
 412	    !rdi->driver_f.qp_priv_alloc ||
 413	    !rdi->driver_f.qp_priv_free ||
 414	    !rdi->driver_f.notify_qp_reset ||
 415	    !rdi->driver_f.notify_restart_rc)
 416		return -EINVAL;
 417
 418	/* allocate parent object */
 419	rdi->qp_dev = kzalloc_node(sizeof(*rdi->qp_dev), GFP_KERNEL,
 420				   rdi->dparms.node);
 421	if (!rdi->qp_dev)
 422		return -ENOMEM;
 423
 424	/* allocate hash table */
 425	rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size;
 426	rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size);
 427	rdi->qp_dev->qp_table =
 428		kmalloc_array_node(rdi->qp_dev->qp_table_size,
 429			     sizeof(*rdi->qp_dev->qp_table),
 430			     GFP_KERNEL, rdi->dparms.node);
 431	if (!rdi->qp_dev->qp_table)
 432		goto no_qp_table;
 433
 434	for (i = 0; i < rdi->qp_dev->qp_table_size; i++)
 435		RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL);
 436
 437	spin_lock_init(&rdi->qp_dev->qpt_lock);
 438
 439	/* initialize qpn map */
 440	if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table))
 441		goto fail_table;
 442
 443	spin_lock_init(&rdi->n_qps_lock);
 444
 445	return 0;
 446
 447fail_table:
 448	kfree(rdi->qp_dev->qp_table);
 449	free_qpn_table(&rdi->qp_dev->qpn_table);
 450
 451no_qp_table:
 452	kfree(rdi->qp_dev);
 453
 454	return ret;
 455}
 456
 457/**
 458 * rvt_free_qp_cb - callback function to reset a qp
 459 * @qp: the qp to reset
 460 * @v: a 64-bit value
 461 *
 462 * This function resets the qp and removes it from the
 463 * qp hash table.
 464 */
 465static void rvt_free_qp_cb(struct rvt_qp *qp, u64 v)
 466{
 467	unsigned int *qp_inuse = (unsigned int *)v;
 468	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
 469
 470	/* Reset the qp and remove it from the qp hash list */
 471	rvt_reset_qp(rdi, qp, qp->ibqp.qp_type);
 472
 473	/* Increment the qp_inuse count */
 474	(*qp_inuse)++;
 475}
 476
 477/**
 478 * rvt_free_all_qps - check for QPs still in use
 479 * @rdi: rvt device info structure
 480 *
 481 * There should not be any QPs still in use.
 482 * Free memory for table.
 483 * Return the number of QPs still in use.
 484 */
 485static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi)
 486{
 487	unsigned int qp_inuse = 0;
 488
 489	qp_inuse += rvt_mcast_tree_empty(rdi);
 490
 491	rvt_qp_iter(rdi, (u64)&qp_inuse, rvt_free_qp_cb);
 492
 493	return qp_inuse;
 494}
 495
 496/**
 497 * rvt_qp_exit - clean up qps on device exit
 498 * @rdi: rvt dev structure
 499 *
 500 * Check for qp leaks and free resources.
 501 */
 502void rvt_qp_exit(struct rvt_dev_info *rdi)
 503{
 504	u32 qps_inuse = rvt_free_all_qps(rdi);
 505
 506	if (qps_inuse)
 507		rvt_pr_err(rdi, "QP memory leak! %u still in use\n",
 508			   qps_inuse);
 509	if (!rdi->qp_dev)
 510		return;
 511
 512	kfree(rdi->qp_dev->qp_table);
 513	free_qpn_table(&rdi->qp_dev->qpn_table);
 514	kfree(rdi->qp_dev);
 515}
 516
 517static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
 518			      struct rvt_qpn_map *map, unsigned off)
 519{
 520	return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
 521}
 522
 523/**
 524 * alloc_qpn - Allocate the next available qpn or zero/one for QP type
 525 *	       IB_QPT_SMI/IB_QPT_GSI
 526 * @rdi: rvt device info structure
 527 * @qpt: queue pair number table pointer
 528 * @type: the QP type
 529 * @port_num: IB port number, 1 based, comes from core
 530 * @exclude_prefix: prefix of special queue pair number being allocated
 531 *
 532 * Return: The queue pair number
 533 */
 534static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
 535		     enum ib_qp_type type, u8 port_num, u8 exclude_prefix)
 536{
 537	u32 i, offset, max_scan, qpn;
 538	struct rvt_qpn_map *map;
 539	u32 ret;
 540	u32 max_qpn = exclude_prefix == RVT_AIP_QP_PREFIX ?
 541		RVT_AIP_QPN_MAX : RVT_QPN_MAX;
 542
 543	if (rdi->driver_f.alloc_qpn)
 544		return rdi->driver_f.alloc_qpn(rdi, qpt, type, port_num);
 545
 546	if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
 547		unsigned n;
 548
 549		ret = type == IB_QPT_GSI;
 550		n = 1 << (ret + 2 * (port_num - 1));
 551		spin_lock(&qpt->lock);
 552		if (qpt->flags & n)
 553			ret = -EINVAL;
 554		else
 555			qpt->flags |= n;
 556		spin_unlock(&qpt->lock);
 557		goto bail;
 558	}
 559
 560	qpn = qpt->last + qpt->incr;
 561	if (qpn >= max_qpn)
 562		qpn = qpt->incr | ((qpt->last & 1) ^ 1);
 563	/* offset carries bit 0 */
 564	offset = qpn & RVT_BITS_PER_PAGE_MASK;
 565	map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
 566	max_scan = qpt->nmaps - !offset;
 567	for (i = 0;;) {
 568		if (unlikely(!map->page)) {
 569			get_map_page(qpt, map);
 570			if (unlikely(!map->page))
 571				break;
 572		}
 573		do {
 574			if (!test_and_set_bit(offset, map->page)) {
 575				qpt->last = qpn;
 576				ret = qpn;
 577				goto bail;
 578			}
 579			offset += qpt->incr;
 580			/*
 581			 * This qpn might be bogus if offset >= BITS_PER_PAGE.
 582			 * That is OK.   It gets re-assigned below
 583			 */
 584			qpn = mk_qpn(qpt, map, offset);
 585		} while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
 586		/*
 587		 * In order to keep the number of pages allocated to a
 588		 * minimum, we scan the all existing pages before increasing
 589		 * the size of the bitmap table.
 590		 */
 591		if (++i > max_scan) {
 592			if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
 593				break;
 594			map = &qpt->map[qpt->nmaps++];
 595			/* start at incr with current bit 0 */
 596			offset = qpt->incr | (offset & 1);
 597		} else if (map < &qpt->map[qpt->nmaps]) {
 598			++map;
 599			/* start at incr with current bit 0 */
 600			offset = qpt->incr | (offset & 1);
 601		} else {
 602			map = &qpt->map[0];
 603			/* wrap to first map page, invert bit 0 */
 604			offset = qpt->incr | ((offset & 1) ^ 1);
 605		}
 606		/* there can be no set bits in low-order QoS bits */
 607		WARN_ON(rdi->dparms.qos_shift > 1 &&
 608			offset & ((BIT(rdi->dparms.qos_shift - 1) - 1) << 1));
 609		qpn = mk_qpn(qpt, map, offset);
 610	}
 611
 612	ret = -ENOMEM;
 613
 614bail:
 615	return ret;
 616}
 617
 618/**
 619 * rvt_clear_mr_refs - Drop help mr refs
 620 * @qp: rvt qp data structure
 621 * @clr_sends: If shoudl clear send side or not
 622 */
 623static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends)
 624{
 625	unsigned n;
 626	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
 627
 628	if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags))
 629		rvt_put_ss(&qp->s_rdma_read_sge);
 630
 631	rvt_put_ss(&qp->r_sge);
 632
 633	if (clr_sends) {
 634		while (qp->s_last != qp->s_head) {
 635			struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last);
 636
 637			rvt_put_qp_swqe(qp, wqe);
 638			if (++qp->s_last >= qp->s_size)
 639				qp->s_last = 0;
 640			smp_wmb(); /* see qp_set_savail */
 641		}
 642		if (qp->s_rdma_mr) {
 643			rvt_put_mr(qp->s_rdma_mr);
 644			qp->s_rdma_mr = NULL;
 645		}
 646	}
 647
 648	for (n = 0; qp->s_ack_queue && n < rvt_max_atomic(rdi); n++) {
 649		struct rvt_ack_entry *e = &qp->s_ack_queue[n];
 650
 651		if (e->rdma_sge.mr) {
 652			rvt_put_mr(e->rdma_sge.mr);
 653			e->rdma_sge.mr = NULL;
 654		}
 655	}
 656}
 657
 658/**
 659 * rvt_swqe_has_lkey - return true if lkey is used by swqe
 660 * @wqe: the send wqe
 661 * @lkey: the lkey
 662 *
 663 * Test the swqe for using lkey
 664 */
 665static bool rvt_swqe_has_lkey(struct rvt_swqe *wqe, u32 lkey)
 666{
 667	int i;
 668
 669	for (i = 0; i < wqe->wr.num_sge; i++) {
 670		struct rvt_sge *sge = &wqe->sg_list[i];
 671
 672		if (rvt_mr_has_lkey(sge->mr, lkey))
 673			return true;
 674	}
 675	return false;
 676}
 677
 678/**
 679 * rvt_qp_sends_has_lkey - return true is qp sends use lkey
 680 * @qp: the rvt_qp
 681 * @lkey: the lkey
 682 */
 683static bool rvt_qp_sends_has_lkey(struct rvt_qp *qp, u32 lkey)
 684{
 685	u32 s_last = qp->s_last;
 686
 687	while (s_last != qp->s_head) {
 688		struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, s_last);
 689
 690		if (rvt_swqe_has_lkey(wqe, lkey))
 691			return true;
 692
 693		if (++s_last >= qp->s_size)
 694			s_last = 0;
 695	}
 696	if (qp->s_rdma_mr)
 697		if (rvt_mr_has_lkey(qp->s_rdma_mr, lkey))
 698			return true;
 699	return false;
 700}
 701
 702/**
 703 * rvt_qp_acks_has_lkey - return true if acks have lkey
 704 * @qp: the qp
 705 * @lkey: the lkey
 706 */
 707static bool rvt_qp_acks_has_lkey(struct rvt_qp *qp, u32 lkey)
 708{
 709	int i;
 710	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
 711
 712	for (i = 0; qp->s_ack_queue && i < rvt_max_atomic(rdi); i++) {
 713		struct rvt_ack_entry *e = &qp->s_ack_queue[i];
 714
 715		if (rvt_mr_has_lkey(e->rdma_sge.mr, lkey))
 716			return true;
 717	}
 718	return false;
 719}
 720
 721/**
 722 * rvt_qp_mr_clean - clean up remote ops for lkey
 723 * @qp: the qp
 724 * @lkey: the lkey that is being de-registered
 725 *
 726 * This routine checks if the lkey is being used by
 727 * the qp.
 728 *
 729 * If so, the qp is put into an error state to elminate
 730 * any references from the qp.
 731 */
 732void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey)
 733{
 734	bool lastwqe = false;
 735
 736	if (qp->ibqp.qp_type == IB_QPT_SMI ||
 737	    qp->ibqp.qp_type == IB_QPT_GSI)
 738		/* avoid special QPs */
 739		return;
 740	spin_lock_irq(&qp->r_lock);
 741	spin_lock(&qp->s_hlock);
 742	spin_lock(&qp->s_lock);
 743
 744	if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
 745		goto check_lwqe;
 746
 747	if (rvt_ss_has_lkey(&qp->r_sge, lkey) ||
 748	    rvt_qp_sends_has_lkey(qp, lkey) ||
 749	    rvt_qp_acks_has_lkey(qp, lkey))
 750		lastwqe = rvt_error_qp(qp, IB_WC_LOC_PROT_ERR);
 751check_lwqe:
 752	spin_unlock(&qp->s_lock);
 753	spin_unlock(&qp->s_hlock);
 754	spin_unlock_irq(&qp->r_lock);
 755	if (lastwqe) {
 756		struct ib_event ev;
 757
 758		ev.device = qp->ibqp.device;
 759		ev.element.qp = &qp->ibqp;
 760		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 761		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 762	}
 763}
 764
 765/**
 766 * rvt_remove_qp - remove qp form table
 767 * @rdi: rvt dev struct
 768 * @qp: qp to remove
 769 *
 770 * Remove the QP from the table so it can't be found asynchronously by
 771 * the receive routine.
 772 */
 773static void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 774{
 775	struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
 776	u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
 777	unsigned long flags;
 778	int removed = 1;
 779
 780	spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
 781
 782	if (rcu_dereference_protected(rvp->qp[0],
 783			lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
 784		RCU_INIT_POINTER(rvp->qp[0], NULL);
 785	} else if (rcu_dereference_protected(rvp->qp[1],
 786			lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
 787		RCU_INIT_POINTER(rvp->qp[1], NULL);
 788	} else {
 789		struct rvt_qp *q;
 790		struct rvt_qp __rcu **qpp;
 791
 792		removed = 0;
 793		qpp = &rdi->qp_dev->qp_table[n];
 794		for (; (q = rcu_dereference_protected(*qpp,
 795			lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL;
 796			qpp = &q->next) {
 797			if (q == qp) {
 798				RCU_INIT_POINTER(*qpp,
 799				     rcu_dereference_protected(qp->next,
 800				     lockdep_is_held(&rdi->qp_dev->qpt_lock)));
 801				removed = 1;
 802				trace_rvt_qpremove(qp, n);
 803				break;
 804			}
 805		}
 806	}
 807
 808	spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
 809	if (removed) {
 810		synchronize_rcu();
 811		rvt_put_qp(qp);
 812	}
 813}
 814
 815/**
 816 * rvt_alloc_rq - allocate memory for user or kernel buffer
 817 * @rq: receive queue data structure
 818 * @size: number of request queue entries
 819 * @node: The NUMA node
 820 * @udata: True if user data is available or not false
 821 *
 822 * Return: If memory allocation failed, return -ENONEM
 823 * This function is used by both shared receive
 824 * queues and non-shared receive queues to allocate
 825 * memory.
 826 */
 827int rvt_alloc_rq(struct rvt_rq *rq, u32 size, int node,
 828		 struct ib_udata *udata)
 829{
 830	if (udata) {
 831		rq->wq = vmalloc_user(sizeof(struct rvt_rwq) + size);
 832		if (!rq->wq)
 833			goto bail;
 834		/* need kwq with no buffers */
 835		rq->kwq = kzalloc_node(sizeof(*rq->kwq), GFP_KERNEL, node);
 836		if (!rq->kwq)
 837			goto bail;
 838		rq->kwq->curr_wq = rq->wq->wq;
 839	} else {
 840		/* need kwq with buffers */
 841		rq->kwq =
 842			vzalloc_node(sizeof(struct rvt_krwq) + size, node);
 843		if (!rq->kwq)
 844			goto bail;
 845		rq->kwq->curr_wq = rq->kwq->wq;
 846	}
 847
 848	spin_lock_init(&rq->kwq->p_lock);
 849	spin_lock_init(&rq->kwq->c_lock);
 850	return 0;
 851bail:
 852	rvt_free_rq(rq);
 853	return -ENOMEM;
 854}
 855
 856/**
 857 * rvt_init_qp - initialize the QP state to the reset state
 858 * @rdi: rvt dev struct
 859 * @qp: the QP to init or reinit
 860 * @type: the QP type
 861 *
 862 * This function is called from both rvt_create_qp() and
 863 * rvt_reset_qp().   The difference is that the reset
 864 * patch the necessary locks to protect against concurent
 865 * access.
 866 */
 867static void rvt_init_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 868			enum ib_qp_type type)
 869{
 870	qp->remote_qpn = 0;
 871	qp->qkey = 0;
 872	qp->qp_access_flags = 0;
 873	qp->s_flags &= RVT_S_SIGNAL_REQ_WR;
 874	qp->s_hdrwords = 0;
 875	qp->s_wqe = NULL;
 876	qp->s_draining = 0;
 877	qp->s_next_psn = 0;
 878	qp->s_last_psn = 0;
 879	qp->s_sending_psn = 0;
 880	qp->s_sending_hpsn = 0;
 881	qp->s_psn = 0;
 882	qp->r_psn = 0;
 883	qp->r_msn = 0;
 884	if (type == IB_QPT_RC) {
 885		qp->s_state = IB_OPCODE_RC_SEND_LAST;
 886		qp->r_state = IB_OPCODE_RC_SEND_LAST;
 887	} else {
 888		qp->s_state = IB_OPCODE_UC_SEND_LAST;
 889		qp->r_state = IB_OPCODE_UC_SEND_LAST;
 890	}
 891	qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
 892	qp->r_nak_state = 0;
 893	qp->r_aflags = 0;
 894	qp->r_flags = 0;
 895	qp->s_head = 0;
 896	qp->s_tail = 0;
 897	qp->s_cur = 0;
 898	qp->s_acked = 0;
 899	qp->s_last = 0;
 900	qp->s_ssn = 1;
 901	qp->s_lsn = 0;
 902	qp->s_mig_state = IB_MIG_MIGRATED;
 903	qp->r_head_ack_queue = 0;
 904	qp->s_tail_ack_queue = 0;
 905	qp->s_acked_ack_queue = 0;
 906	qp->s_num_rd_atomic = 0;
 907	qp->r_sge.num_sge = 0;
 908	atomic_set(&qp->s_reserved_used, 0);
 909}
 910
 911/**
 912 * _rvt_reset_qp - initialize the QP state to the reset state
 913 * @rdi: rvt dev struct
 914 * @qp: the QP to reset
 915 * @type: the QP type
 916 *
 917 * r_lock, s_hlock, and s_lock are required to be held by the caller
 918 */
 919static void _rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 920			  enum ib_qp_type type)
 921	__must_hold(&qp->s_lock)
 922	__must_hold(&qp->s_hlock)
 923	__must_hold(&qp->r_lock)
 924{
 925	lockdep_assert_held(&qp->r_lock);
 926	lockdep_assert_held(&qp->s_hlock);
 927	lockdep_assert_held(&qp->s_lock);
 928	if (qp->state != IB_QPS_RESET) {
 929		qp->state = IB_QPS_RESET;
 930
 931		/* Let drivers flush their waitlist */
 932		rdi->driver_f.flush_qp_waiters(qp);
 933		rvt_stop_rc_timers(qp);
 934		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT);
 935		spin_unlock(&qp->s_lock);
 936		spin_unlock(&qp->s_hlock);
 937		spin_unlock_irq(&qp->r_lock);
 938
 939		/* Stop the send queue and the retry timer */
 940		rdi->driver_f.stop_send_queue(qp);
 941		rvt_del_timers_sync(qp);
 942		/* Wait for things to stop */
 943		rdi->driver_f.quiesce_qp(qp);
 944
 945		/* take qp out the hash and wait for it to be unused */
 946		rvt_remove_qp(rdi, qp);
 947
 948		/* grab the lock b/c it was locked at call time */
 949		spin_lock_irq(&qp->r_lock);
 950		spin_lock(&qp->s_hlock);
 951		spin_lock(&qp->s_lock);
 952
 953		rvt_clear_mr_refs(qp, 1);
 954		/*
 955		 * Let the driver do any tear down or re-init it needs to for
 956		 * a qp that has been reset
 957		 */
 958		rdi->driver_f.notify_qp_reset(qp);
 959	}
 960	rvt_init_qp(rdi, qp, type);
 961	lockdep_assert_held(&qp->r_lock);
 962	lockdep_assert_held(&qp->s_hlock);
 963	lockdep_assert_held(&qp->s_lock);
 964}
 965
 966/**
 967 * rvt_reset_qp - initialize the QP state to the reset state
 968 * @rdi: the device info
 969 * @qp: the QP to reset
 970 * @type: the QP type
 971 *
 972 * This is the wrapper function to acquire the r_lock, s_hlock, and s_lock
 973 * before calling _rvt_reset_qp().
 974 */
 975static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 976			 enum ib_qp_type type)
 977{
 978	spin_lock_irq(&qp->r_lock);
 979	spin_lock(&qp->s_hlock);
 980	spin_lock(&qp->s_lock);
 981	_rvt_reset_qp(rdi, qp, type);
 982	spin_unlock(&qp->s_lock);
 983	spin_unlock(&qp->s_hlock);
 984	spin_unlock_irq(&qp->r_lock);
 985}
 986
 987/**
 988 * rvt_free_qpn - Free a qpn from the bit map
 989 * @qpt: QP table
 990 * @qpn: queue pair number to free
 991 */
 992static void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
 993{
 994	struct rvt_qpn_map *map;
 995
 996	if ((qpn & RVT_AIP_QP_PREFIX_MASK) == RVT_AIP_QP_BASE)
 997		qpn &= RVT_AIP_QP_SUFFIX;
 998
 999	map = qpt->map + (qpn & RVT_QPN_MASK) / RVT_BITS_PER_PAGE;
1000	if (map->page)
1001		clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
1002}
1003
1004/**
1005 * get_allowed_ops - Given a QP type return the appropriate allowed OP
1006 * @type: valid, supported, QP type
1007 */
1008static u8 get_allowed_ops(enum ib_qp_type type)
1009{
1010	return type == IB_QPT_RC ? IB_OPCODE_RC : type == IB_QPT_UC ?
1011		IB_OPCODE_UC : IB_OPCODE_UD;
1012}
1013
1014/**
1015 * free_ud_wq_attr - Clean up AH attribute cache for UD QPs
1016 * @qp: Valid QP with allowed_ops set
1017 *
1018 * The rvt_swqe data structure being used is a union, so this is
1019 * only valid for UD QPs.
1020 */
1021static void free_ud_wq_attr(struct rvt_qp *qp)
1022{
1023	struct rvt_swqe *wqe;
1024	int i;
1025
1026	for (i = 0; qp->allowed_ops == IB_OPCODE_UD && i < qp->s_size; i++) {
1027		wqe = rvt_get_swqe_ptr(qp, i);
1028		kfree(wqe->ud_wr.attr);
1029		wqe->ud_wr.attr = NULL;
1030	}
1031}
1032
1033/**
1034 * alloc_ud_wq_attr - AH attribute cache for UD QPs
1035 * @qp: Valid QP with allowed_ops set
1036 * @node: Numa node for allocation
1037 *
1038 * The rvt_swqe data structure being used is a union, so this is
1039 * only valid for UD QPs.
1040 */
1041static int alloc_ud_wq_attr(struct rvt_qp *qp, int node)
1042{
1043	struct rvt_swqe *wqe;
1044	int i;
1045
1046	for (i = 0; qp->allowed_ops == IB_OPCODE_UD && i < qp->s_size; i++) {
1047		wqe = rvt_get_swqe_ptr(qp, i);
1048		wqe->ud_wr.attr = kzalloc_node(sizeof(*wqe->ud_wr.attr),
1049					       GFP_KERNEL, node);
1050		if (!wqe->ud_wr.attr) {
1051			free_ud_wq_attr(qp);
1052			return -ENOMEM;
1053		}
1054	}
1055
1056	return 0;
1057}
1058
1059/**
1060 * rvt_create_qp - create a queue pair for a device
1061 * @ibpd: the protection domain who's device we create the queue pair for
1062 * @init_attr: the attributes of the queue pair
1063 * @udata: user data for libibverbs.so
1064 *
1065 * Queue pair creation is mostly an rvt issue. However, drivers have their own
1066 * unique idea of what queue pair numbers mean. For instance there is a reserved
1067 * range for PSM.
1068 *
1069 * Return: the queue pair on success, otherwise returns an errno.
1070 *
1071 * Called by the ib_create_qp() core verbs function.
1072 */
1073struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
1074			    struct ib_qp_init_attr *init_attr,
1075			    struct ib_udata *udata)
1076{
1077	struct rvt_qp *qp;
1078	int err;
1079	struct rvt_swqe *swq = NULL;
1080	size_t sz;
1081	size_t sg_list_sz;
1082	struct ib_qp *ret = ERR_PTR(-ENOMEM);
1083	struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device);
1084	void *priv = NULL;
1085	size_t sqsize;
1086	u8 exclude_prefix = 0;
1087
1088	if (!rdi)
1089		return ERR_PTR(-EINVAL);
1090
1091	if (init_attr->create_flags & ~IB_QP_CREATE_NETDEV_USE)
1092		return ERR_PTR(-EOPNOTSUPP);
1093
1094	if (init_attr->cap.max_send_sge > rdi->dparms.props.max_send_sge ||
1095	    init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr)
1096		return ERR_PTR(-EINVAL);
1097
1098	/* Check receive queue parameters if no SRQ is specified. */
1099	if (!init_attr->srq) {
1100		if (init_attr->cap.max_recv_sge >
1101		    rdi->dparms.props.max_recv_sge ||
1102		    init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr)
1103			return ERR_PTR(-EINVAL);
1104
1105		if (init_attr->cap.max_send_sge +
1106		    init_attr->cap.max_send_wr +
1107		    init_attr->cap.max_recv_sge +
1108		    init_attr->cap.max_recv_wr == 0)
1109			return ERR_PTR(-EINVAL);
1110	}
1111	sqsize =
1112		init_attr->cap.max_send_wr + 1 +
1113		rdi->dparms.reserved_operations;
1114	switch (init_attr->qp_type) {
1115	case IB_QPT_SMI:
1116	case IB_QPT_GSI:
1117		if (init_attr->port_num == 0 ||
1118		    init_attr->port_num > ibpd->device->phys_port_cnt)
1119			return ERR_PTR(-EINVAL);
1120		fallthrough;
1121	case IB_QPT_UC:
1122	case IB_QPT_RC:
1123	case IB_QPT_UD:
1124		sz = struct_size(swq, sg_list, init_attr->cap.max_send_sge);
1125		swq = vzalloc_node(array_size(sz, sqsize), rdi->dparms.node);
1126		if (!swq)
1127			return ERR_PTR(-ENOMEM);
1128
1129		sz = sizeof(*qp);
1130		sg_list_sz = 0;
1131		if (init_attr->srq) {
1132			struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq);
1133
1134			if (srq->rq.max_sge > 1)
1135				sg_list_sz = sizeof(*qp->r_sg_list) *
1136					(srq->rq.max_sge - 1);
1137		} else if (init_attr->cap.max_recv_sge > 1)
1138			sg_list_sz = sizeof(*qp->r_sg_list) *
1139				(init_attr->cap.max_recv_sge - 1);
1140		qp = kzalloc_node(sz + sg_list_sz, GFP_KERNEL,
1141				  rdi->dparms.node);
1142		if (!qp)
1143			goto bail_swq;
1144		qp->allowed_ops = get_allowed_ops(init_attr->qp_type);
1145
1146		RCU_INIT_POINTER(qp->next, NULL);
1147		if (init_attr->qp_type == IB_QPT_RC) {
1148			qp->s_ack_queue =
1149				kcalloc_node(rvt_max_atomic(rdi),
1150					     sizeof(*qp->s_ack_queue),
1151					     GFP_KERNEL,
1152					     rdi->dparms.node);
1153			if (!qp->s_ack_queue)
1154				goto bail_qp;
1155		}
1156		/* initialize timers needed for rc qp */
1157		timer_setup(&qp->s_timer, rvt_rc_timeout, 0);
1158		hrtimer_init(&qp->s_rnr_timer, CLOCK_MONOTONIC,
1159			     HRTIMER_MODE_REL);
1160		qp->s_rnr_timer.function = rvt_rc_rnr_retry;
1161
1162		/*
1163		 * Driver needs to set up it's private QP structure and do any
1164		 * initialization that is needed.
1165		 */
1166		priv = rdi->driver_f.qp_priv_alloc(rdi, qp);
1167		if (IS_ERR(priv)) {
1168			ret = priv;
1169			goto bail_qp;
1170		}
1171		qp->priv = priv;
1172		qp->timeout_jiffies =
1173			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1174				1000UL);
1175		if (init_attr->srq) {
1176			sz = 0;
1177		} else {
1178			qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
1179			qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
1180			sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
1181				sizeof(struct rvt_rwqe);
1182			err = rvt_alloc_rq(&qp->r_rq, qp->r_rq.size * sz,
1183					   rdi->dparms.node, udata);
1184			if (err) {
1185				ret = ERR_PTR(err);
1186				goto bail_driver_priv;
1187			}
1188		}
1189
1190		/*
1191		 * ib_create_qp() will initialize qp->ibqp
1192		 * except for qp->ibqp.qp_num.
1193		 */
1194		spin_lock_init(&qp->r_lock);
1195		spin_lock_init(&qp->s_hlock);
1196		spin_lock_init(&qp->s_lock);
1197		atomic_set(&qp->refcount, 0);
1198		atomic_set(&qp->local_ops_pending, 0);
1199		init_waitqueue_head(&qp->wait);
1200		INIT_LIST_HEAD(&qp->rspwait);
1201		qp->state = IB_QPS_RESET;
1202		qp->s_wq = swq;
1203		qp->s_size = sqsize;
1204		qp->s_avail = init_attr->cap.max_send_wr;
1205		qp->s_max_sge = init_attr->cap.max_send_sge;
1206		if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
1207			qp->s_flags = RVT_S_SIGNAL_REQ_WR;
1208		err = alloc_ud_wq_attr(qp, rdi->dparms.node);
1209		if (err) {
1210			ret = (ERR_PTR(err));
1211			goto bail_rq_rvt;
1212		}
1213
1214		if (init_attr->create_flags & IB_QP_CREATE_NETDEV_USE)
1215			exclude_prefix = RVT_AIP_QP_PREFIX;
1216
1217		err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table,
1218				init_attr->qp_type,
1219				init_attr->port_num,
1220				exclude_prefix);
1221		if (err < 0) {
1222			ret = ERR_PTR(err);
1223			goto bail_rq_wq;
1224		}
1225		qp->ibqp.qp_num = err;
1226		if (init_attr->create_flags & IB_QP_CREATE_NETDEV_USE)
1227			qp->ibqp.qp_num |= RVT_AIP_QP_BASE;
1228		qp->port_num = init_attr->port_num;
1229		rvt_init_qp(rdi, qp, init_attr->qp_type);
1230		if (rdi->driver_f.qp_priv_init) {
1231			err = rdi->driver_f.qp_priv_init(rdi, qp, init_attr);
1232			if (err) {
1233				ret = ERR_PTR(err);
1234				goto bail_rq_wq;
1235			}
1236		}
1237		break;
1238
1239	default:
1240		/* Don't support raw QPs */
1241		return ERR_PTR(-EOPNOTSUPP);
1242	}
1243
1244	init_attr->cap.max_inline_data = 0;
1245
1246	/*
1247	 * Return the address of the RWQ as the offset to mmap.
1248	 * See rvt_mmap() for details.
1249	 */
1250	if (udata && udata->outlen >= sizeof(__u64)) {
1251		if (!qp->r_rq.wq) {
1252			__u64 offset = 0;
1253
1254			err = ib_copy_to_udata(udata, &offset,
1255					       sizeof(offset));
1256			if (err) {
1257				ret = ERR_PTR(err);
1258				goto bail_qpn;
1259			}
1260		} else {
1261			u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz;
1262
1263			qp->ip = rvt_create_mmap_info(rdi, s, udata,
1264						      qp->r_rq.wq);
1265			if (IS_ERR(qp->ip)) {
1266				ret = ERR_CAST(qp->ip);
1267				goto bail_qpn;
1268			}
1269
1270			err = ib_copy_to_udata(udata, &qp->ip->offset,
1271					       sizeof(qp->ip->offset));
1272			if (err) {
1273				ret = ERR_PTR(err);
1274				goto bail_ip;
1275			}
1276		}
1277		qp->pid = current->pid;
1278	}
1279
1280	spin_lock(&rdi->n_qps_lock);
1281	if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) {
1282		spin_unlock(&rdi->n_qps_lock);
1283		ret = ERR_PTR(-ENOMEM);
1284		goto bail_ip;
1285	}
1286
1287	rdi->n_qps_allocated++;
1288	/*
1289	 * Maintain a busy_jiffies variable that will be added to the timeout
1290	 * period in mod_retry_timer and add_retry_timer. This busy jiffies
1291	 * is scaled by the number of rc qps created for the device to reduce
1292	 * the number of timeouts occurring when there is a large number of
1293	 * qps. busy_jiffies is incremented every rc qp scaling interval.
1294	 * The scaling interval is selected based on extensive performance
1295	 * evaluation of targeted workloads.
1296	 */
1297	if (init_attr->qp_type == IB_QPT_RC) {
1298		rdi->n_rc_qps++;
1299		rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
1300	}
1301	spin_unlock(&rdi->n_qps_lock);
1302
1303	if (qp->ip) {
1304		spin_lock_irq(&rdi->pending_lock);
1305		list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps);
1306		spin_unlock_irq(&rdi->pending_lock);
1307	}
1308
1309	ret = &qp->ibqp;
1310
1311	return ret;
1312
1313bail_ip:
1314	if (qp->ip)
1315		kref_put(&qp->ip->ref, rvt_release_mmap_info);
1316
1317bail_qpn:
1318	rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
1319
1320bail_rq_wq:
1321	free_ud_wq_attr(qp);
1322
1323bail_rq_rvt:
1324	rvt_free_rq(&qp->r_rq);
1325
1326bail_driver_priv:
1327	rdi->driver_f.qp_priv_free(rdi, qp);
1328
1329bail_qp:
1330	kfree(qp->s_ack_queue);
1331	kfree(qp);
1332
1333bail_swq:
1334	vfree(swq);
1335
1336	return ret;
1337}
1338
1339/**
1340 * rvt_error_qp - put a QP into the error state
1341 * @qp: the QP to put into the error state
1342 * @err: the receive completion error to signal if a RWQE is active
1343 *
1344 * Flushes both send and receive work queues.
1345 *
1346 * Return: true if last WQE event should be generated.
1347 * The QP r_lock and s_lock should be held and interrupts disabled.
1348 * If we are already in error state, just return.
1349 */
1350int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
1351{
1352	struct ib_wc wc;
1353	int ret = 0;
1354	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
1355
1356	lockdep_assert_held(&qp->r_lock);
1357	lockdep_assert_held(&qp->s_lock);
1358	if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
1359		goto bail;
1360
1361	qp->state = IB_QPS_ERR;
1362
1363	if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
1364		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
1365		del_timer(&qp->s_timer);
1366	}
1367
1368	if (qp->s_flags & RVT_S_ANY_WAIT_SEND)
1369		qp->s_flags &= ~RVT_S_ANY_WAIT_SEND;
1370
1371	rdi->driver_f.notify_error_qp(qp);
1372
1373	/* Schedule the sending tasklet to drain the send work queue. */
1374	if (READ_ONCE(qp->s_last) != qp->s_head)
1375		rdi->driver_f.schedule_send(qp);
1376
1377	rvt_clear_mr_refs(qp, 0);
1378
1379	memset(&wc, 0, sizeof(wc));
1380	wc.qp = &qp->ibqp;
1381	wc.opcode = IB_WC_RECV;
1382
1383	if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) {
1384		wc.wr_id = qp->r_wr_id;
1385		wc.status = err;
1386		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1387	}
1388	wc.status = IB_WC_WR_FLUSH_ERR;
1389
1390	if (qp->r_rq.kwq) {
1391		u32 head;
1392		u32 tail;
1393		struct rvt_rwq *wq = NULL;
1394		struct rvt_krwq *kwq = NULL;
1395
1396		spin_lock(&qp->r_rq.kwq->c_lock);
1397		/* qp->ip used to validate if there is a  user buffer mmaped */
1398		if (qp->ip) {
1399			wq = qp->r_rq.wq;
1400			head = RDMA_READ_UAPI_ATOMIC(wq->head);
1401			tail = RDMA_READ_UAPI_ATOMIC(wq->tail);
1402		} else {
1403			kwq = qp->r_rq.kwq;
1404			head = kwq->head;
1405			tail = kwq->tail;
1406		}
1407		/* sanity check pointers before trusting them */
1408		if (head >= qp->r_rq.size)
1409			head = 0;
1410		if (tail >= qp->r_rq.size)
1411			tail = 0;
1412		while (tail != head) {
1413			wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
1414			if (++tail >= qp->r_rq.size)
1415				tail = 0;
1416			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1417		}
1418		if (qp->ip)
1419			RDMA_WRITE_UAPI_ATOMIC(wq->tail, tail);
1420		else
1421			kwq->tail = tail;
1422		spin_unlock(&qp->r_rq.kwq->c_lock);
1423	} else if (qp->ibqp.event_handler) {
1424		ret = 1;
1425	}
1426
1427bail:
1428	return ret;
1429}
1430EXPORT_SYMBOL(rvt_error_qp);
1431
1432/*
1433 * Put the QP into the hash table.
1434 * The hash table holds a reference to the QP.
1435 */
1436static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
1437{
1438	struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
1439	unsigned long flags;
1440
1441	rvt_get_qp(qp);
1442	spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
1443
1444	if (qp->ibqp.qp_num <= 1) {
1445		rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp);
1446	} else {
1447		u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
1448
1449		qp->next = rdi->qp_dev->qp_table[n];
1450		rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp);
1451		trace_rvt_qpinsert(qp, n);
1452	}
1453
1454	spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
1455}
1456
1457/**
1458 * rvt_modify_qp - modify the attributes of a queue pair
1459 * @ibqp: the queue pair who's attributes we're modifying
1460 * @attr: the new attributes
1461 * @attr_mask: the mask of attributes to modify
1462 * @udata: user data for libibverbs.so
1463 *
1464 * Return: 0 on success, otherwise returns an errno.
1465 */
1466int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1467		  int attr_mask, struct ib_udata *udata)
1468{
1469	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1470	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1471	enum ib_qp_state cur_state, new_state;
1472	struct ib_event ev;
1473	int lastwqe = 0;
1474	int mig = 0;
1475	int pmtu = 0; /* for gcc warning only */
1476	int opa_ah;
1477
1478	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1479		return -EOPNOTSUPP;
1480
1481	spin_lock_irq(&qp->r_lock);
1482	spin_lock(&qp->s_hlock);
1483	spin_lock(&qp->s_lock);
1484
1485	cur_state = attr_mask & IB_QP_CUR_STATE ?
1486		attr->cur_qp_state : qp->state;
1487	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1488	opa_ah = rdma_cap_opa_ah(ibqp->device, qp->port_num);
1489
1490	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1491				attr_mask))
1492		goto inval;
1493
1494	if (rdi->driver_f.check_modify_qp &&
1495	    rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata))
1496		goto inval;
1497
1498	if (attr_mask & IB_QP_AV) {
1499		if (opa_ah) {
1500			if (rdma_ah_get_dlid(&attr->ah_attr) >=
1501				opa_get_mcast_base(OPA_MCAST_NR))
1502				goto inval;
1503		} else {
1504			if (rdma_ah_get_dlid(&attr->ah_attr) >=
1505				be16_to_cpu(IB_MULTICAST_LID_BASE))
1506				goto inval;
1507		}
1508
1509		if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr))
1510			goto inval;
1511	}
1512
1513	if (attr_mask & IB_QP_ALT_PATH) {
1514		if (opa_ah) {
1515			if (rdma_ah_get_dlid(&attr->alt_ah_attr) >=
1516				opa_get_mcast_base(OPA_MCAST_NR))
1517				goto inval;
1518		} else {
1519			if (rdma_ah_get_dlid(&attr->alt_ah_attr) >=
1520				be16_to_cpu(IB_MULTICAST_LID_BASE))
1521				goto inval;
1522		}
1523
1524		if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
1525			goto inval;
1526		if (attr->alt_pkey_index >= rvt_get_npkeys(rdi))
1527			goto inval;
1528	}
1529
1530	if (attr_mask & IB_QP_PKEY_INDEX)
1531		if (attr->pkey_index >= rvt_get_npkeys(rdi))
1532			goto inval;
1533
1534	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1535		if (attr->min_rnr_timer > 31)
1536			goto inval;
1537
1538	if (attr_mask & IB_QP_PORT)
1539		if (qp->ibqp.qp_type == IB_QPT_SMI ||
1540		    qp->ibqp.qp_type == IB_QPT_GSI ||
1541		    attr->port_num == 0 ||
1542		    attr->port_num > ibqp->device->phys_port_cnt)
1543			goto inval;
1544
1545	if (attr_mask & IB_QP_DEST_QPN)
1546		if (attr->dest_qp_num > RVT_QPN_MASK)
1547			goto inval;
1548
1549	if (attr_mask & IB_QP_RETRY_CNT)
1550		if (attr->retry_cnt > 7)
1551			goto inval;
1552
1553	if (attr_mask & IB_QP_RNR_RETRY)
1554		if (attr->rnr_retry > 7)
1555			goto inval;
1556
1557	/*
1558	 * Don't allow invalid path_mtu values.  OK to set greater
1559	 * than the active mtu (or even the max_cap, if we have tuned
1560	 * that to a small mtu.  We'll set qp->path_mtu
1561	 * to the lesser of requested attribute mtu and active,
1562	 * for packetizing messages.
1563	 * Note that the QP port has to be set in INIT and MTU in RTR.
1564	 */
1565	if (attr_mask & IB_QP_PATH_MTU) {
1566		pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr);
1567		if (pmtu < 0)
1568			goto inval;
1569	}
1570
1571	if (attr_mask & IB_QP_PATH_MIG_STATE) {
1572		if (attr->path_mig_state == IB_MIG_REARM) {
1573			if (qp->s_mig_state == IB_MIG_ARMED)
1574				goto inval;
1575			if (new_state != IB_QPS_RTS)
1576				goto inval;
1577		} else if (attr->path_mig_state == IB_MIG_MIGRATED) {
1578			if (qp->s_mig_state == IB_MIG_REARM)
1579				goto inval;
1580			if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
1581				goto inval;
1582			if (qp->s_mig_state == IB_MIG_ARMED)
1583				mig = 1;
1584		} else {
1585			goto inval;
1586		}
1587	}
1588
1589	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1590		if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic)
1591			goto inval;
1592
1593	switch (new_state) {
1594	case IB_QPS_RESET:
1595		if (qp->state != IB_QPS_RESET)
1596			_rvt_reset_qp(rdi, qp, ibqp->qp_type);
1597		break;
1598
1599	case IB_QPS_RTR:
1600		/* Allow event to re-trigger if QP set to RTR more than once */
1601		qp->r_flags &= ~RVT_R_COMM_EST;
1602		qp->state = new_state;
1603		break;
1604
1605	case IB_QPS_SQD:
1606		qp->s_draining = qp->s_last != qp->s_cur;
1607		qp->state = new_state;
1608		break;
1609
1610	case IB_QPS_SQE:
1611		if (qp->ibqp.qp_type == IB_QPT_RC)
1612			goto inval;
1613		qp->state = new_state;
1614		break;
1615
1616	case IB_QPS_ERR:
1617		lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1618		break;
1619
1620	default:
1621		qp->state = new_state;
1622		break;
1623	}
1624
1625	if (attr_mask & IB_QP_PKEY_INDEX)
1626		qp->s_pkey_index = attr->pkey_index;
1627
1628	if (attr_mask & IB_QP_PORT)
1629		qp->port_num = attr->port_num;
1630
1631	if (attr_mask & IB_QP_DEST_QPN)
1632		qp->remote_qpn = attr->dest_qp_num;
1633
1634	if (attr_mask & IB_QP_SQ_PSN) {
1635		qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask;
1636		qp->s_psn = qp->s_next_psn;
1637		qp->s_sending_psn = qp->s_next_psn;
1638		qp->s_last_psn = qp->s_next_psn - 1;
1639		qp->s_sending_hpsn = qp->s_last_psn;
1640	}
1641
1642	if (attr_mask & IB_QP_RQ_PSN)
1643		qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask;
1644
1645	if (attr_mask & IB_QP_ACCESS_FLAGS)
1646		qp->qp_access_flags = attr->qp_access_flags;
1647
1648	if (attr_mask & IB_QP_AV) {
1649		rdma_replace_ah_attr(&qp->remote_ah_attr, &attr->ah_attr);
1650		qp->s_srate = rdma_ah_get_static_rate(&attr->ah_attr);
1651		qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
1652	}
1653
1654	if (attr_mask & IB_QP_ALT_PATH) {
1655		rdma_replace_ah_attr(&qp->alt_ah_attr, &attr->alt_ah_attr);
1656		qp->s_alt_pkey_index = attr->alt_pkey_index;
1657	}
1658
1659	if (attr_mask & IB_QP_PATH_MIG_STATE) {
1660		qp->s_mig_state = attr->path_mig_state;
1661		if (mig) {
1662			qp->remote_ah_attr = qp->alt_ah_attr;
1663			qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
1664			qp->s_pkey_index = qp->s_alt_pkey_index;
1665		}
1666	}
1667
1668	if (attr_mask & IB_QP_PATH_MTU) {
1669		qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu);
1670		qp->log_pmtu = ilog2(qp->pmtu);
1671	}
1672
1673	if (attr_mask & IB_QP_RETRY_CNT) {
1674		qp->s_retry_cnt = attr->retry_cnt;
1675		qp->s_retry = attr->retry_cnt;
1676	}
1677
1678	if (attr_mask & IB_QP_RNR_RETRY) {
1679		qp->s_rnr_retry_cnt = attr->rnr_retry;
1680		qp->s_rnr_retry = attr->rnr_retry;
1681	}
1682
1683	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1684		qp->r_min_rnr_timer = attr->min_rnr_timer;
1685
1686	if (attr_mask & IB_QP_TIMEOUT) {
1687		qp->timeout = attr->timeout;
1688		qp->timeout_jiffies = rvt_timeout_to_jiffies(qp->timeout);
1689	}
1690
1691	if (attr_mask & IB_QP_QKEY)
1692		qp->qkey = attr->qkey;
1693
1694	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1695		qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
1696
1697	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
1698		qp->s_max_rd_atomic = attr->max_rd_atomic;
1699
1700	if (rdi->driver_f.modify_qp)
1701		rdi->driver_f.modify_qp(qp, attr, attr_mask, udata);
1702
1703	spin_unlock(&qp->s_lock);
1704	spin_unlock(&qp->s_hlock);
1705	spin_unlock_irq(&qp->r_lock);
1706
1707	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1708		rvt_insert_qp(rdi, qp);
1709
1710	if (lastwqe) {
1711		ev.device = qp->ibqp.device;
1712		ev.element.qp = &qp->ibqp;
1713		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1714		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1715	}
1716	if (mig) {
1717		ev.device = qp->ibqp.device;
1718		ev.element.qp = &qp->ibqp;
1719		ev.event = IB_EVENT_PATH_MIG;
1720		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1721	}
1722	return 0;
1723
1724inval:
1725	spin_unlock(&qp->s_lock);
1726	spin_unlock(&qp->s_hlock);
1727	spin_unlock_irq(&qp->r_lock);
1728	return -EINVAL;
1729}
1730
1731/**
1732 * rvt_destroy_qp - destroy a queue pair
1733 * @ibqp: the queue pair to destroy
1734 * @udata: unused by the driver
1735 *
1736 * Note that this can be called while the QP is actively sending or
1737 * receiving!
1738 *
1739 * Return: 0 on success.
1740 */
1741int rvt_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
1742{
1743	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1744	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1745
1746	rvt_reset_qp(rdi, qp, ibqp->qp_type);
1747
1748	wait_event(qp->wait, !atomic_read(&qp->refcount));
1749	/* qpn is now available for use again */
1750	rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
1751
1752	spin_lock(&rdi->n_qps_lock);
1753	rdi->n_qps_allocated--;
1754	if (qp->ibqp.qp_type == IB_QPT_RC) {
1755		rdi->n_rc_qps--;
1756		rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
1757	}
1758	spin_unlock(&rdi->n_qps_lock);
1759
1760	if (qp->ip)
1761		kref_put(&qp->ip->ref, rvt_release_mmap_info);
1762	kvfree(qp->r_rq.kwq);
1763	rdi->driver_f.qp_priv_free(rdi, qp);
1764	kfree(qp->s_ack_queue);
1765	rdma_destroy_ah_attr(&qp->remote_ah_attr);
1766	rdma_destroy_ah_attr(&qp->alt_ah_attr);
1767	free_ud_wq_attr(qp);
1768	vfree(qp->s_wq);
1769	kfree(qp);
1770	return 0;
1771}
1772
1773/**
1774 * rvt_query_qp - query an ipbq
1775 * @ibqp: IB qp to query
1776 * @attr: attr struct to fill in
1777 * @attr_mask: attr mask ignored
1778 * @init_attr: struct to fill in
1779 *
1780 * Return: always 0
1781 */
1782int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1783		 int attr_mask, struct ib_qp_init_attr *init_attr)
1784{
1785	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1786	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1787
1788	attr->qp_state = qp->state;
1789	attr->cur_qp_state = attr->qp_state;
1790	attr->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu);
1791	attr->path_mig_state = qp->s_mig_state;
1792	attr->qkey = qp->qkey;
1793	attr->rq_psn = qp->r_psn & rdi->dparms.psn_mask;
1794	attr->sq_psn = qp->s_next_psn & rdi->dparms.psn_mask;
1795	attr->dest_qp_num = qp->remote_qpn;
1796	attr->qp_access_flags = qp->qp_access_flags;
1797	attr->cap.max_send_wr = qp->s_size - 1 -
1798		rdi->dparms.reserved_operations;
1799	attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
1800	attr->cap.max_send_sge = qp->s_max_sge;
1801	attr->cap.max_recv_sge = qp->r_rq.max_sge;
1802	attr->cap.max_inline_data = 0;
1803	attr->ah_attr = qp->remote_ah_attr;
1804	attr->alt_ah_attr = qp->alt_ah_attr;
1805	attr->pkey_index = qp->s_pkey_index;
1806	attr->alt_pkey_index = qp->s_alt_pkey_index;
1807	attr->en_sqd_async_notify = 0;
1808	attr->sq_draining = qp->s_draining;
1809	attr->max_rd_atomic = qp->s_max_rd_atomic;
1810	attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
1811	attr->min_rnr_timer = qp->r_min_rnr_timer;
1812	attr->port_num = qp->port_num;
1813	attr->timeout = qp->timeout;
1814	attr->retry_cnt = qp->s_retry_cnt;
1815	attr->rnr_retry = qp->s_rnr_retry_cnt;
1816	attr->alt_port_num =
1817		rdma_ah_get_port_num(&qp->alt_ah_attr);
1818	attr->alt_timeout = qp->alt_timeout;
1819
1820	init_attr->event_handler = qp->ibqp.event_handler;
1821	init_attr->qp_context = qp->ibqp.qp_context;
1822	init_attr->send_cq = qp->ibqp.send_cq;
1823	init_attr->recv_cq = qp->ibqp.recv_cq;
1824	init_attr->srq = qp->ibqp.srq;
1825	init_attr->cap = attr->cap;
1826	if (qp->s_flags & RVT_S_SIGNAL_REQ_WR)
1827		init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
1828	else
1829		init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1830	init_attr->qp_type = qp->ibqp.qp_type;
1831	init_attr->port_num = qp->port_num;
1832	return 0;
1833}
1834
1835/**
1836 * rvt_post_recv - post a receive on a QP
1837 * @ibqp: the QP to post the receive on
1838 * @wr: the WR to post
1839 * @bad_wr: the first bad WR is put here
1840 *
1841 * This may be called from interrupt context.
1842 *
1843 * Return: 0 on success otherwise errno
1844 */
1845int rvt_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
1846		  const struct ib_recv_wr **bad_wr)
1847{
1848	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1849	struct rvt_krwq *wq = qp->r_rq.kwq;
1850	unsigned long flags;
1851	int qp_err_flush = (ib_rvt_state_ops[qp->state] & RVT_FLUSH_RECV) &&
1852				!qp->ibqp.srq;
1853
1854	/* Check that state is OK to post receive. */
1855	if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) {
1856		*bad_wr = wr;
1857		return -EINVAL;
1858	}
1859
1860	for (; wr; wr = wr->next) {
1861		struct rvt_rwqe *wqe;
1862		u32 next;
1863		int i;
1864
1865		if ((unsigned)wr->num_sge > qp->r_rq.max_sge) {
1866			*bad_wr = wr;
1867			return -EINVAL;
1868		}
1869
1870		spin_lock_irqsave(&qp->r_rq.kwq->p_lock, flags);
1871		next = wq->head + 1;
1872		if (next >= qp->r_rq.size)
1873			next = 0;
1874		if (next == READ_ONCE(wq->tail)) {
1875			spin_unlock_irqrestore(&qp->r_rq.kwq->p_lock, flags);
1876			*bad_wr = wr;
1877			return -ENOMEM;
1878		}
1879		if (unlikely(qp_err_flush)) {
1880			struct ib_wc wc;
1881
1882			memset(&wc, 0, sizeof(wc));
1883			wc.qp = &qp->ibqp;
1884			wc.opcode = IB_WC_RECV;
1885			wc.wr_id = wr->wr_id;
1886			wc.status = IB_WC_WR_FLUSH_ERR;
1887			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1888		} else {
1889			wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head);
1890			wqe->wr_id = wr->wr_id;
1891			wqe->num_sge = wr->num_sge;
1892			for (i = 0; i < wr->num_sge; i++) {
1893				wqe->sg_list[i].addr = wr->sg_list[i].addr;
1894				wqe->sg_list[i].length = wr->sg_list[i].length;
1895				wqe->sg_list[i].lkey = wr->sg_list[i].lkey;
1896			}
1897			/*
1898			 * Make sure queue entry is written
1899			 * before the head index.
1900			 */
1901			smp_store_release(&wq->head, next);
1902		}
1903		spin_unlock_irqrestore(&qp->r_rq.kwq->p_lock, flags);
1904	}
1905	return 0;
1906}
1907
1908/**
1909 * rvt_qp_valid_operation - validate post send wr request
1910 * @qp: the qp
1911 * @post_parms: the post send table for the driver
1912 * @wr: the work request
1913 *
1914 * The routine validates the operation based on the
1915 * validation table an returns the length of the operation
1916 * which can extend beyond the ib_send_bw.  Operation
1917 * dependent flags key atomic operation validation.
1918 *
1919 * There is an exception for UD qps that validates the pd and
1920 * overrides the length to include the additional UD specific
1921 * length.
1922 *
1923 * Returns a negative error or the length of the work request
1924 * for building the swqe.
1925 */
1926static inline int rvt_qp_valid_operation(
1927	struct rvt_qp *qp,
1928	const struct rvt_operation_params *post_parms,
1929	const struct ib_send_wr *wr)
1930{
1931	int len;
1932
1933	if (wr->opcode >= RVT_OPERATION_MAX || !post_parms[wr->opcode].length)
1934		return -EINVAL;
1935	if (!(post_parms[wr->opcode].qpt_support & BIT(qp->ibqp.qp_type)))
1936		return -EINVAL;
1937	if ((post_parms[wr->opcode].flags & RVT_OPERATION_PRIV) &&
1938	    ibpd_to_rvtpd(qp->ibqp.pd)->user)
1939		return -EINVAL;
1940	if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC_SGE &&
1941	    (wr->num_sge == 0 ||
1942	     wr->sg_list[0].length < sizeof(u64) ||
1943	     wr->sg_list[0].addr & (sizeof(u64) - 1)))
1944		return -EINVAL;
1945	if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC &&
1946	    !qp->s_max_rd_atomic)
1947		return -EINVAL;
1948	len = post_parms[wr->opcode].length;
1949	/* UD specific */
1950	if (qp->ibqp.qp_type != IB_QPT_UC &&
1951	    qp->ibqp.qp_type != IB_QPT_RC) {
1952		if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
1953			return -EINVAL;
1954		len = sizeof(struct ib_ud_wr);
1955	}
1956	return len;
1957}
1958
1959/**
1960 * rvt_qp_is_avail - determine queue capacity
1961 * @qp: the qp
1962 * @rdi: the rdmavt device
1963 * @reserved_op: is reserved operation
1964 *
1965 * This assumes the s_hlock is held but the s_last
1966 * qp variable is uncontrolled.
1967 *
1968 * For non reserved operations, the qp->s_avail
1969 * may be changed.
1970 *
1971 * The return value is zero or a -ENOMEM.
1972 */
1973static inline int rvt_qp_is_avail(
1974	struct rvt_qp *qp,
1975	struct rvt_dev_info *rdi,
1976	bool reserved_op)
1977{
1978	u32 slast;
1979	u32 avail;
1980	u32 reserved_used;
1981
1982	/* see rvt_qp_wqe_unreserve() */
1983	smp_mb__before_atomic();
1984	if (unlikely(reserved_op)) {
1985		/* see rvt_qp_wqe_unreserve() */
1986		reserved_used = atomic_read(&qp->s_reserved_used);
1987		if (reserved_used >= rdi->dparms.reserved_operations)
1988			return -ENOMEM;
1989		return 0;
1990	}
1991	/* non-reserved operations */
1992	if (likely(qp->s_avail))
1993		return 0;
1994	/* See rvt_qp_complete_swqe() */
1995	slast = smp_load_acquire(&qp->s_last);
1996	if (qp->s_head >= slast)
1997		avail = qp->s_size - (qp->s_head - slast);
1998	else
1999		avail = slast - qp->s_head;
2000
2001	reserved_used = atomic_read(&qp->s_reserved_used);
2002	avail =  avail - 1 -
2003		(rdi->dparms.reserved_operations - reserved_used);
2004	/* insure we don't assign a negative s_avail */
2005	if ((s32)avail <= 0)
2006		return -ENOMEM;
2007	qp->s_avail = avail;
2008	if (WARN_ON(qp->s_avail >
2009		    (qp->s_size - 1 - rdi->dparms.reserved_operations)))
2010		rvt_pr_err(rdi,
2011			   "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u",
2012			   qp->ibqp.qp_num, qp->s_size, qp->s_avail,
2013			   qp->s_head, qp->s_tail, qp->s_cur,
2014			   qp->s_acked, qp->s_last);
2015	return 0;
2016}
2017
2018/**
2019 * rvt_post_one_wr - post one RC, UC, or UD send work request
2020 * @qp: the QP to post on
2021 * @wr: the work request to send
2022 * @call_send: kick the send engine into gear
2023 */
2024static int rvt_post_one_wr(struct rvt_qp *qp,
2025			   const struct ib_send_wr *wr,
2026			   bool *call_send)
2027{
2028	struct rvt_swqe *wqe;
2029	u32 next;
2030	int i;
2031	int j;
2032	int acc;
2033	struct rvt_lkey_table *rkt;
2034	struct rvt_pd *pd;
2035	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
2036	u8 log_pmtu;
2037	int ret;
2038	size_t cplen;
2039	bool reserved_op;
2040	int local_ops_delayed = 0;
2041
2042	BUILD_BUG_ON(IB_QPT_MAX >= (sizeof(u32) * BITS_PER_BYTE));
2043
2044	/* IB spec says that num_sge == 0 is OK. */
2045	if (unlikely(wr->num_sge > qp->s_max_sge))
2046		return -EINVAL;
2047
2048	ret = rvt_qp_valid_operation(qp, rdi->post_parms, wr);
2049	if (ret < 0)
2050		return ret;
2051	cplen = ret;
2052
2053	/*
2054	 * Local operations include fast register and local invalidate.
2055	 * Fast register needs to be processed immediately because the
2056	 * registered lkey may be used by following work requests and the
2057	 * lkey needs to be valid at the time those requests are posted.
2058	 * Local invalidate can be processed immediately if fencing is
2059	 * not required and no previous local invalidate ops are pending.
2060	 * Signaled local operations that have been processed immediately
2061	 * need to have requests with "completion only" flags set posted
2062	 * to the send queue in order to generate completions.
2063	 */
2064	if ((rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL)) {
2065		switch (wr->opcode) {
2066		case IB_WR_REG_MR:
2067			ret = rvt_fast_reg_mr(qp,
2068					      reg_wr(wr)->mr,
2069					      reg_wr(wr)->key,
2070					      reg_wr(wr)->access);
2071			if (ret || !(wr->send_flags & IB_SEND_SIGNALED))
2072				return ret;
2073			break;
2074		case IB_WR_LOCAL_INV:
2075			if ((wr->send_flags & IB_SEND_FENCE) ||
2076			    atomic_read(&qp->local_ops_pending)) {
2077				local_ops_delayed = 1;
2078			} else {
2079				ret = rvt_invalidate_rkey(
2080					qp, wr->ex.invalidate_rkey);
2081				if (ret || !(wr->send_flags & IB_SEND_SIGNALED))
2082					return ret;
2083			}
2084			break;
2085		default:
2086			return -EINVAL;
2087		}
2088	}
2089
2090	reserved_op = rdi->post_parms[wr->opcode].flags &
2091			RVT_OPERATION_USE_RESERVE;
2092	/* check for avail */
2093	ret = rvt_qp_is_avail(qp, rdi, reserved_op);
2094	if (ret)
2095		return ret;
2096	next = qp->s_head + 1;
2097	if (next >= qp->s_size)
2098		next = 0;
2099
2100	rkt = &rdi->lkey_table;
2101	pd = ibpd_to_rvtpd(qp->ibqp.pd);
2102	wqe = rvt_get_swqe_ptr(qp, qp->s_head);
2103
2104	/* cplen has length from above */
2105	memcpy(&wqe->wr, wr, cplen);
2106
2107	wqe->length = 0;
2108	j = 0;
2109	if (wr->num_sge) {
2110		struct rvt_sge *last_sge = NULL;
2111
2112		acc = wr->opcode >= IB_WR_RDMA_READ ?
2113			IB_ACCESS_LOCAL_WRITE : 0;
2114		for (i = 0; i < wr->num_sge; i++) {
2115			u32 length = wr->sg_list[i].length;
2116
2117			if (length == 0)
2118				continue;
2119			ret = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j], last_sge,
2120					  &wr->sg_list[i], acc);
2121			if (unlikely(ret < 0))
2122				goto bail_inval_free;
2123			wqe->length += length;
2124			if (ret)
2125				last_sge = &wqe->sg_list[j];
2126			j += ret;
2127		}
2128		wqe->wr.num_sge = j;
2129	}
2130
2131	/*
2132	 * Calculate and set SWQE PSN values prior to handing it off
2133	 * to the driver's check routine. This give the driver the
2134	 * opportunity to adjust PSN values based on internal checks.
2135	 */
2136	log_pmtu = qp->log_pmtu;
2137	if (qp->allowed_ops == IB_OPCODE_UD) {
2138		struct rvt_ah *ah = rvt_get_swqe_ah(wqe);
2139
2140		log_pmtu = ah->log_pmtu;
2141		rdma_copy_ah_attr(wqe->ud_wr.attr, &ah->attr);
2142	}
2143
2144	if (rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL) {
2145		if (local_ops_delayed)
2146			atomic_inc(&qp->local_ops_pending);
2147		else
2148			wqe->wr.send_flags |= RVT_SEND_COMPLETION_ONLY;
2149		wqe->ssn = 0;
2150		wqe->psn = 0;
2151		wqe->lpsn = 0;
2152	} else {
2153		wqe->ssn = qp->s_ssn++;
2154		wqe->psn = qp->s_next_psn;
2155		wqe->lpsn = wqe->psn +
2156				(wqe->length ?
2157					((wqe->length - 1) >> log_pmtu) :
2158					0);
2159	}
2160
2161	/* general part of wqe valid - allow for driver checks */
2162	if (rdi->driver_f.setup_wqe) {
2163		ret = rdi->driver_f.setup_wqe(qp, wqe, call_send);
2164		if (ret < 0)
2165			goto bail_inval_free_ref;
2166	}
2167
2168	if (!(rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL))
2169		qp->s_next_psn = wqe->lpsn + 1;
2170
2171	if (unlikely(reserved_op)) {
2172		wqe->wr.send_flags |= RVT_SEND_RESERVE_USED;
2173		rvt_qp_wqe_reserve(qp, wqe);
2174	} else {
2175		wqe->wr.send_flags &= ~RVT_SEND_RESERVE_USED;
2176		qp->s_avail--;
2177	}
2178	trace_rvt_post_one_wr(qp, wqe, wr->num_sge);
2179	smp_wmb(); /* see request builders */
2180	qp->s_head = next;
2181
2182	return 0;
2183
2184bail_inval_free_ref:
2185	if (qp->allowed_ops == IB_OPCODE_UD)
2186		rdma_destroy_ah_attr(wqe->ud_wr.attr);
2187bail_inval_free:
2188	/* release mr holds */
2189	while (j) {
2190		struct rvt_sge *sge = &wqe->sg_list[--j];
2191
2192		rvt_put_mr(sge->mr);
2193	}
2194	return ret;
2195}
2196
2197/**
2198 * rvt_post_send - post a send on a QP
2199 * @ibqp: the QP to post the send on
2200 * @wr: the list of work requests to post
2201 * @bad_wr: the first bad WR is put here
2202 *
2203 * This may be called from interrupt context.
2204 *
2205 * Return: 0 on success else errno
2206 */
2207int rvt_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
2208		  const struct ib_send_wr **bad_wr)
2209{
2210	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
2211	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
2212	unsigned long flags = 0;
2213	bool call_send;
2214	unsigned nreq = 0;
2215	int err = 0;
2216
2217	spin_lock_irqsave(&qp->s_hlock, flags);
2218
2219	/*
2220	 * Ensure QP state is such that we can send. If not bail out early,
2221	 * there is no need to do this every time we post a send.
2222	 */
2223	if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) {
2224		spin_unlock_irqrestore(&qp->s_hlock, flags);
2225		return -EINVAL;
2226	}
2227
2228	/*
2229	 * If the send queue is empty, and we only have a single WR then just go
2230	 * ahead and kick the send engine into gear. Otherwise we will always
2231	 * just schedule the send to happen later.
2232	 */
2233	call_send = qp->s_head == READ_ONCE(qp->s_last) && !wr->next;
2234
2235	for (; wr; wr = wr->next) {
2236		err = rvt_post_one_wr(qp, wr, &call_send);
2237		if (unlikely(err)) {
2238			*bad_wr = wr;
2239			goto bail;
2240		}
2241		nreq++;
2242	}
2243bail:
2244	spin_unlock_irqrestore(&qp->s_hlock, flags);
2245	if (nreq) {
2246		/*
2247		 * Only call do_send if there is exactly one packet, and the
2248		 * driver said it was ok.
2249		 */
2250		if (nreq == 1 && call_send)
2251			rdi->driver_f.do_send(qp);
2252		else
2253			rdi->driver_f.schedule_send_no_lock(qp);
2254	}
2255	return err;
2256}
2257
2258/**
2259 * rvt_post_srq_recv - post a receive on a shared receive queue
2260 * @ibsrq: the SRQ to post the receive on
2261 * @wr: the list of work requests to post
2262 * @bad_wr: A pointer to the first WR to cause a problem is put here
2263 *
2264 * This may be called from interrupt context.
2265 *
2266 * Return: 0 on success else errno
2267 */
2268int rvt_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
2269		      const struct ib_recv_wr **bad_wr)
2270{
2271	struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
2272	struct rvt_krwq *wq;
2273	unsigned long flags;
2274
2275	for (; wr; wr = wr->next) {
2276		struct rvt_rwqe *wqe;
2277		u32 next;
2278		int i;
2279
2280		if ((unsigned)wr->num_sge > srq->rq.max_sge) {
2281			*bad_wr = wr;
2282			return -EINVAL;
2283		}
2284
2285		spin_lock_irqsave(&srq->rq.kwq->p_lock, flags);
2286		wq = srq->rq.kwq;
2287		next = wq->head + 1;
2288		if (next >= srq->rq.size)
2289			next = 0;
2290		if (next == READ_ONCE(wq->tail)) {
2291			spin_unlock_irqrestore(&srq->rq.kwq->p_lock, flags);
2292			*bad_wr = wr;
2293			return -ENOMEM;
2294		}
2295
2296		wqe = rvt_get_rwqe_ptr(&srq->rq, wq->head);
2297		wqe->wr_id = wr->wr_id;
2298		wqe->num_sge = wr->num_sge;
2299		for (i = 0; i < wr->num_sge; i++) {
2300			wqe->sg_list[i].addr = wr->sg_list[i].addr;
2301			wqe->sg_list[i].length = wr->sg_list[i].length;
2302			wqe->sg_list[i].lkey = wr->sg_list[i].lkey;
2303		}
2304		/* Make sure queue entry is written before the head index. */
2305		smp_store_release(&wq->head, next);
2306		spin_unlock_irqrestore(&srq->rq.kwq->p_lock, flags);
2307	}
2308	return 0;
2309}
2310
2311/*
2312 * rvt used the internal kernel struct as part of its ABI, for now make sure
2313 * the kernel struct does not change layout. FIXME: rvt should never cast the
2314 * user struct to a kernel struct.
2315 */
2316static struct ib_sge *rvt_cast_sge(struct rvt_wqe_sge *sge)
2317{
2318	BUILD_BUG_ON(offsetof(struct ib_sge, addr) !=
2319		     offsetof(struct rvt_wqe_sge, addr));
2320	BUILD_BUG_ON(offsetof(struct ib_sge, length) !=
2321		     offsetof(struct rvt_wqe_sge, length));
2322	BUILD_BUG_ON(offsetof(struct ib_sge, lkey) !=
2323		     offsetof(struct rvt_wqe_sge, lkey));
2324	return (struct ib_sge *)sge;
2325}
2326
2327/*
2328 * Validate a RWQE and fill in the SGE state.
2329 * Return 1 if OK.
2330 */
2331static int init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe)
2332{
2333	int i, j, ret;
2334	struct ib_wc wc;
2335	struct rvt_lkey_table *rkt;
2336	struct rvt_pd *pd;
2337	struct rvt_sge_state *ss;
2338	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
2339
2340	rkt = &rdi->lkey_table;
2341	pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
2342	ss = &qp->r_sge;
2343	ss->sg_list = qp->r_sg_list;
2344	qp->r_len = 0;
2345	for (i = j = 0; i < wqe->num_sge; i++) {
2346		if (wqe->sg_list[i].length == 0)
2347			continue;
2348		/* Check LKEY */
2349		ret = rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
2350				  NULL, rvt_cast_sge(&wqe->sg_list[i]),
2351				  IB_ACCESS_LOCAL_WRITE);
2352		if (unlikely(ret <= 0))
2353			goto bad_lkey;
2354		qp->r_len += wqe->sg_list[i].length;
2355		j++;
2356	}
2357	ss->num_sge = j;
2358	ss->total_len = qp->r_len;
2359	return 1;
2360
2361bad_lkey:
2362	while (j) {
2363		struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
2364
2365		rvt_put_mr(sge->mr);
2366	}
2367	ss->num_sge = 0;
2368	memset(&wc, 0, sizeof(wc));
2369	wc.wr_id = wqe->wr_id;
2370	wc.status = IB_WC_LOC_PROT_ERR;
2371	wc.opcode = IB_WC_RECV;
2372	wc.qp = &qp->ibqp;
2373	/* Signal solicited completion event. */
2374	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
2375	return 0;
2376}
2377
2378/**
2379 * get_rvt_head - get head indices of the circular buffer
2380 * @rq: data structure for request queue entry
2381 * @ip: the QP
2382 *
2383 * Return - head index value
2384 */
2385static inline u32 get_rvt_head(struct rvt_rq *rq, void *ip)
2386{
2387	u32 head;
2388
2389	if (ip)
2390		head = RDMA_READ_UAPI_ATOMIC(rq->wq->head);
2391	else
2392		head = rq->kwq->head;
2393
2394	return head;
2395}
2396
2397/**
2398 * rvt_get_rwqe - copy the next RWQE into the QP's RWQE
2399 * @qp: the QP
2400 * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
2401 *
2402 * Return -1 if there is a local error, 0 if no RWQE is available,
2403 * otherwise return 1.
2404 *
2405 * Can be called from interrupt level.
2406 */
2407int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only)
2408{
2409	unsigned long flags;
2410	struct rvt_rq *rq;
2411	struct rvt_krwq *kwq = NULL;
2412	struct rvt_rwq *wq;
2413	struct rvt_srq *srq;
2414	struct rvt_rwqe *wqe;
2415	void (*handler)(struct ib_event *, void *);
2416	u32 tail;
2417	u32 head;
2418	int ret;
2419	void *ip = NULL;
2420
2421	if (qp->ibqp.srq) {
2422		srq = ibsrq_to_rvtsrq(qp->ibqp.srq);
2423		handler = srq->ibsrq.event_handler;
2424		rq = &srq->rq;
2425		ip = srq->ip;
2426	} else {
2427		srq = NULL;
2428		handler = NULL;
2429		rq = &qp->r_rq;
2430		ip = qp->ip;
2431	}
2432
2433	spin_lock_irqsave(&rq->kwq->c_lock, flags);
2434	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
2435		ret = 0;
2436		goto unlock;
2437	}
2438	kwq = rq->kwq;
2439	if (ip) {
2440		wq = rq->wq;
2441		tail = RDMA_READ_UAPI_ATOMIC(wq->tail);
2442	} else {
2443		tail = kwq->tail;
2444	}
2445
2446	/* Validate tail before using it since it is user writable. */
2447	if (tail >= rq->size)
2448		tail = 0;
2449
2450	if (kwq->count < RVT_RWQ_COUNT_THRESHOLD) {
2451		head = get_rvt_head(rq, ip);
2452		kwq->count = rvt_get_rq_count(rq, head, tail);
2453	}
2454	if (unlikely(kwq->count == 0)) {
2455		ret = 0;
2456		goto unlock;
2457	}
2458	/* Make sure entry is read after the count is read. */
2459	smp_rmb();
2460	wqe = rvt_get_rwqe_ptr(rq, tail);
2461	/*
2462	 * Even though we update the tail index in memory, the verbs
2463	 * consumer is not supposed to post more entries until a
2464	 * completion is generated.
2465	 */
2466	if (++tail >= rq->size)
2467		tail = 0;
2468	if (ip)
2469		RDMA_WRITE_UAPI_ATOMIC(wq->tail, tail);
2470	else
2471		kwq->tail = tail;
2472	if (!wr_id_only && !init_sge(qp, wqe)) {
2473		ret = -1;
2474		goto unlock;
2475	}
2476	qp->r_wr_id = wqe->wr_id;
2477
2478	kwq->count--;
2479	ret = 1;
2480	set_bit(RVT_R_WRID_VALID, &qp->r_aflags);
2481	if (handler) {
2482		/*
2483		 * Validate head pointer value and compute
2484		 * the number of remaining WQEs.
2485		 */
2486		if (kwq->count < srq->limit) {
2487			kwq->count =
2488				rvt_get_rq_count(rq,
2489						 get_rvt_head(rq, ip), tail);
2490			if (kwq->count < srq->limit) {
2491				struct ib_event ev;
2492
2493				srq->limit = 0;
2494				spin_unlock_irqrestore(&rq->kwq->c_lock, flags);
2495				ev.device = qp->ibqp.device;
2496				ev.element.srq = qp->ibqp.srq;
2497				ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
2498				handler(&ev, srq->ibsrq.srq_context);
2499				goto bail;
2500			}
2501		}
2502	}
2503unlock:
2504	spin_unlock_irqrestore(&rq->kwq->c_lock, flags);
2505bail:
2506	return ret;
2507}
2508EXPORT_SYMBOL(rvt_get_rwqe);
2509
2510/**
2511 * rvt_comm_est - handle trap with QP established
2512 * @qp: the QP
2513 */
2514void rvt_comm_est(struct rvt_qp *qp)
2515{
2516	qp->r_flags |= RVT_R_COMM_EST;
2517	if (qp->ibqp.event_handler) {
2518		struct ib_event ev;
2519
2520		ev.device = qp->ibqp.device;
2521		ev.element.qp = &qp->ibqp;
2522		ev.event = IB_EVENT_COMM_EST;
2523		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
2524	}
2525}
2526EXPORT_SYMBOL(rvt_comm_est);
2527
2528void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err)
2529{
2530	unsigned long flags;
2531	int lastwqe;
2532
2533	spin_lock_irqsave(&qp->s_lock, flags);
2534	lastwqe = rvt_error_qp(qp, err);
2535	spin_unlock_irqrestore(&qp->s_lock, flags);
2536
2537	if (lastwqe) {
2538		struct ib_event ev;
2539
2540		ev.device = qp->ibqp.device;
2541		ev.element.qp = &qp->ibqp;
2542		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
2543		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
2544	}
2545}
2546EXPORT_SYMBOL(rvt_rc_error);
2547
2548/*
2549 *  rvt_rnr_tbl_to_usec - return index into ib_rvt_rnr_table
2550 *  @index - the index
2551 *  return usec from an index into ib_rvt_rnr_table
2552 */
2553unsigned long rvt_rnr_tbl_to_usec(u32 index)
2554{
2555	return ib_rvt_rnr_table[(index & IB_AETH_CREDIT_MASK)];
2556}
2557EXPORT_SYMBOL(rvt_rnr_tbl_to_usec);
2558
2559static inline unsigned long rvt_aeth_to_usec(u32 aeth)
2560{
2561	return ib_rvt_rnr_table[(aeth >> IB_AETH_CREDIT_SHIFT) &
2562				  IB_AETH_CREDIT_MASK];
2563}
2564
2565/*
2566 *  rvt_add_retry_timer_ext - add/start a retry timer
2567 *  @qp - the QP
2568 *  @shift - timeout shift to wait for multiple packets
2569 *  add a retry timer on the QP
2570 */
2571void rvt_add_retry_timer_ext(struct rvt_qp *qp, u8 shift)
2572{
2573	struct ib_qp *ibqp = &qp->ibqp;
2574	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
2575
2576	lockdep_assert_held(&qp->s_lock);
2577	qp->s_flags |= RVT_S_TIMER;
2578       /* 4.096 usec. * (1 << qp->timeout) */
2579	qp->s_timer.expires = jiffies + rdi->busy_jiffies +
2580			      (qp->timeout_jiffies << shift);
2581	add_timer(&qp->s_timer);
2582}
2583EXPORT_SYMBOL(rvt_add_retry_timer_ext);
2584
2585/**
2586 * rvt_add_rnr_timer - add/start an rnr timer on the QP
2587 * @qp: the QP
2588 * @aeth: aeth of RNR timeout, simulated aeth for loopback
2589 */
2590void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth)
2591{
2592	u32 to;
2593
2594	lockdep_assert_held(&qp->s_lock);
2595	qp->s_flags |= RVT_S_WAIT_RNR;
2596	to = rvt_aeth_to_usec(aeth);
2597	trace_rvt_rnrnak_add(qp, to);
2598	hrtimer_start(&qp->s_rnr_timer,
2599		      ns_to_ktime(1000 * to), HRTIMER_MODE_REL_PINNED);
2600}
2601EXPORT_SYMBOL(rvt_add_rnr_timer);
2602
2603/**
2604 * rvt_stop_rc_timers - stop all timers
2605 * @qp: the QP
2606 * stop any pending timers
2607 */
2608void rvt_stop_rc_timers(struct rvt_qp *qp)
2609{
2610	lockdep_assert_held(&qp->s_lock);
2611	/* Remove QP from all timers */
2612	if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
2613		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
2614		del_timer(&qp->s_timer);
2615		hrtimer_try_to_cancel(&qp->s_rnr_timer);
2616	}
2617}
2618EXPORT_SYMBOL(rvt_stop_rc_timers);
2619
2620/**
2621 * rvt_stop_rnr_timer - stop an rnr timer
2622 * @qp: the QP
2623 *
2624 * stop an rnr timer and return if the timer
2625 * had been pending.
2626 */
2627static void rvt_stop_rnr_timer(struct rvt_qp *qp)
2628{
2629	lockdep_assert_held(&qp->s_lock);
2630	/* Remove QP from rnr timer */
2631	if (qp->s_flags & RVT_S_WAIT_RNR) {
2632		qp->s_flags &= ~RVT_S_WAIT_RNR;
2633		trace_rvt_rnrnak_stop(qp, 0);
2634	}
2635}
2636
2637/**
2638 * rvt_del_timers_sync - wait for any timeout routines to exit
2639 * @qp: the QP
2640 */
2641void rvt_del_timers_sync(struct rvt_qp *qp)
2642{
2643	del_timer_sync(&qp->s_timer);
2644	hrtimer_cancel(&qp->s_rnr_timer);
2645}
2646EXPORT_SYMBOL(rvt_del_timers_sync);
2647
2648/*
2649 * This is called from s_timer for missing responses.
2650 */
2651static void rvt_rc_timeout(struct timer_list *t)
2652{
2653	struct rvt_qp *qp = from_timer(qp, t, s_timer);
2654	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
2655	unsigned long flags;
2656
2657	spin_lock_irqsave(&qp->r_lock, flags);
2658	spin_lock(&qp->s_lock);
2659	if (qp->s_flags & RVT_S_TIMER) {
2660		struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
2661
2662		qp->s_flags &= ~RVT_S_TIMER;
2663		rvp->n_rc_timeouts++;
2664		del_timer(&qp->s_timer);
2665		trace_rvt_rc_timeout(qp, qp->s_last_psn + 1);
2666		if (rdi->driver_f.notify_restart_rc)
2667			rdi->driver_f.notify_restart_rc(qp,
2668							qp->s_last_psn + 1,
2669							1);
2670		rdi->driver_f.schedule_send(qp);
2671	}
2672	spin_unlock(&qp->s_lock);
2673	spin_unlock_irqrestore(&qp->r_lock, flags);
2674}
2675
2676/*
2677 * This is called from s_timer for RNR timeouts.
2678 */
2679enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t)
2680{
2681	struct rvt_qp *qp = container_of(t, struct rvt_qp, s_rnr_timer);
2682	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
2683	unsigned long flags;
2684
2685	spin_lock_irqsave(&qp->s_lock, flags);
2686	rvt_stop_rnr_timer(qp);
2687	trace_rvt_rnrnak_timeout(qp, 0);
2688	rdi->driver_f.schedule_send(qp);
2689	spin_unlock_irqrestore(&qp->s_lock, flags);
2690	return HRTIMER_NORESTART;
2691}
2692EXPORT_SYMBOL(rvt_rc_rnr_retry);
2693
2694/**
2695 * rvt_qp_iter_init - initial for QP iteration
2696 * @rdi: rvt devinfo
2697 * @v: u64 value
2698 * @cb: user-defined callback
2699 *
2700 * This returns an iterator suitable for iterating QPs
2701 * in the system.
2702 *
2703 * The @cb is a user-defined callback and @v is a 64-bit
2704 * value passed to and relevant for processing in the
2705 * @cb.  An example use case would be to alter QP processing
2706 * based on criteria not part of the rvt_qp.
2707 *
2708 * Use cases that require memory allocation to succeed
2709 * must preallocate appropriately.
2710 *
2711 * Return: a pointer to an rvt_qp_iter or NULL
2712 */
2713struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi,
2714				     u64 v,
2715				     void (*cb)(struct rvt_qp *qp, u64 v))
2716{
2717	struct rvt_qp_iter *i;
2718
2719	i = kzalloc(sizeof(*i), GFP_KERNEL);
2720	if (!i)
2721		return NULL;
2722
2723	i->rdi = rdi;
2724	/* number of special QPs (SMI/GSI) for device */
2725	i->specials = rdi->ibdev.phys_port_cnt * 2;
2726	i->v = v;
2727	i->cb = cb;
2728
2729	return i;
2730}
2731EXPORT_SYMBOL(rvt_qp_iter_init);
2732
2733/**
2734 * rvt_qp_iter_next - return the next QP in iter
2735 * @iter: the iterator
2736 *
2737 * Fine grained QP iterator suitable for use
2738 * with debugfs seq_file mechanisms.
2739 *
2740 * Updates iter->qp with the current QP when the return
2741 * value is 0.
2742 *
2743 * Return: 0 - iter->qp is valid 1 - no more QPs
2744 */
2745int rvt_qp_iter_next(struct rvt_qp_iter *iter)
2746	__must_hold(RCU)
2747{
2748	int n = iter->n;
2749	int ret = 1;
2750	struct rvt_qp *pqp = iter->qp;
2751	struct rvt_qp *qp;
2752	struct rvt_dev_info *rdi = iter->rdi;
2753
2754	/*
2755	 * The approach is to consider the special qps
2756	 * as additional table entries before the
2757	 * real hash table.  Since the qp code sets
2758	 * the qp->next hash link to NULL, this works just fine.
2759	 *
2760	 * iter->specials is 2 * # ports
2761	 *
2762	 * n = 0..iter->specials is the special qp indices
2763	 *
2764	 * n = iter->specials..rdi->qp_dev->qp_table_size+iter->specials are
2765	 * the potential hash bucket entries
2766	 *
2767	 */
2768	for (; n <  rdi->qp_dev->qp_table_size + iter->specials; n++) {
2769		if (pqp) {
2770			qp = rcu_dereference(pqp->next);
2771		} else {
2772			if (n < iter->specials) {
2773				struct rvt_ibport *rvp;
2774				int pidx;
2775
2776				pidx = n % rdi->ibdev.phys_port_cnt;
2777				rvp = rdi->ports[pidx];
2778				qp = rcu_dereference(rvp->qp[n & 1]);
2779			} else {
2780				qp = rcu_dereference(
2781					rdi->qp_dev->qp_table[
2782						(n - iter->specials)]);
2783			}
2784		}
2785		pqp = qp;
2786		if (qp) {
2787			iter->qp = qp;
2788			iter->n = n;
2789			return 0;
2790		}
2791	}
2792	return ret;
2793}
2794EXPORT_SYMBOL(rvt_qp_iter_next);
2795
2796/**
2797 * rvt_qp_iter - iterate all QPs
2798 * @rdi: rvt devinfo
2799 * @v: a 64-bit value
2800 * @cb: a callback
2801 *
2802 * This provides a way for iterating all QPs.
2803 *
2804 * The @cb is a user-defined callback and @v is a 64-bit
2805 * value passed to and relevant for processing in the
2806 * cb.  An example use case would be to alter QP processing
2807 * based on criteria not part of the rvt_qp.
2808 *
2809 * The code has an internal iterator to simplify
2810 * non seq_file use cases.
2811 */
2812void rvt_qp_iter(struct rvt_dev_info *rdi,
2813		 u64 v,
2814		 void (*cb)(struct rvt_qp *qp, u64 v))
2815{
2816	int ret;
2817	struct rvt_qp_iter i = {
2818		.rdi = rdi,
2819		.specials = rdi->ibdev.phys_port_cnt * 2,
2820		.v = v,
2821		.cb = cb
2822	};
2823
2824	rcu_read_lock();
2825	do {
2826		ret = rvt_qp_iter_next(&i);
2827		if (!ret) {
2828			rvt_get_qp(i.qp);
2829			rcu_read_unlock();
2830			i.cb(i.qp, i.v);
2831			rcu_read_lock();
2832			rvt_put_qp(i.qp);
2833		}
2834	} while (!ret);
2835	rcu_read_unlock();
2836}
2837EXPORT_SYMBOL(rvt_qp_iter);
2838
2839/*
2840 * This should be called with s_lock held.
2841 */
2842void rvt_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
2843		       enum ib_wc_status status)
2844{
2845	u32 old_last, last;
2846	struct rvt_dev_info *rdi;
2847
2848	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
2849		return;
2850	rdi = ib_to_rvt(qp->ibqp.device);
2851
2852	old_last = qp->s_last;
2853	trace_rvt_qp_send_completion(qp, wqe, old_last);
2854	last = rvt_qp_complete_swqe(qp, wqe, rdi->wc_opcode[wqe->wr.opcode],
2855				    status);
2856	if (qp->s_acked == old_last)
2857		qp->s_acked = last;
2858	if (qp->s_cur == old_last)
2859		qp->s_cur = last;
2860	if (qp->s_tail == old_last)
2861		qp->s_tail = last;
2862	if (qp->state == IB_QPS_SQD && last == qp->s_cur)
2863		qp->s_draining = 0;
2864}
2865EXPORT_SYMBOL(rvt_send_complete);
2866
2867/**
2868 * rvt_copy_sge - copy data to SGE memory
2869 * @qp: associated QP
2870 * @ss: the SGE state
2871 * @data: the data to copy
2872 * @length: the length of the data
2873 * @release: boolean to release MR
2874 * @copy_last: do a separate copy of the last 8 bytes
2875 */
2876void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss,
2877		  void *data, u32 length,
2878		  bool release, bool copy_last)
2879{
2880	struct rvt_sge *sge = &ss->sge;
2881	int i;
2882	bool in_last = false;
2883	bool cacheless_copy = false;
2884	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
2885	struct rvt_wss *wss = rdi->wss;
2886	unsigned int sge_copy_mode = rdi->dparms.sge_copy_mode;
2887
2888	if (sge_copy_mode == RVT_SGE_COPY_CACHELESS) {
2889		cacheless_copy = length >= PAGE_SIZE;
2890	} else if (sge_copy_mode == RVT_SGE_COPY_ADAPTIVE) {
2891		if (length >= PAGE_SIZE) {
2892			/*
2893			 * NOTE: this *assumes*:
2894			 * o The first vaddr is the dest.
2895			 * o If multiple pages, then vaddr is sequential.
2896			 */
2897			wss_insert(wss, sge->vaddr);
2898			if (length >= (2 * PAGE_SIZE))
2899				wss_insert(wss, (sge->vaddr + PAGE_SIZE));
2900
2901			cacheless_copy = wss_exceeds_threshold(wss);
2902		} else {
2903			wss_advance_clean_counter(wss);
2904		}
2905	}
2906
2907	if (copy_last) {
2908		if (length > 8) {
2909			length -= 8;
2910		} else {
2911			copy_last = false;
2912			in_last = true;
2913		}
2914	}
2915
2916again:
2917	while (length) {
2918		u32 len = rvt_get_sge_length(sge, length);
2919
2920		WARN_ON_ONCE(len == 0);
2921		if (unlikely(in_last)) {
2922			/* enforce byte transfer ordering */
2923			for (i = 0; i < len; i++)
2924				((u8 *)sge->vaddr)[i] = ((u8 *)data)[i];
2925		} else if (cacheless_copy) {
2926			cacheless_memcpy(sge->vaddr, data, len);
2927		} else {
2928			memcpy(sge->vaddr, data, len);
2929		}
2930		rvt_update_sge(ss, len, release);
2931		data += len;
2932		length -= len;
2933	}
2934
2935	if (copy_last) {
2936		copy_last = false;
2937		in_last = true;
2938		length = 8;
2939		goto again;
2940	}
2941}
2942EXPORT_SYMBOL(rvt_copy_sge);
2943
2944static enum ib_wc_status loopback_qp_drop(struct rvt_ibport *rvp,
2945					  struct rvt_qp *sqp)
2946{
2947	rvp->n_pkt_drops++;
2948	/*
2949	 * For RC, the requester would timeout and retry so
2950	 * shortcut the timeouts and just signal too many retries.
2951	 */
2952	return sqp->ibqp.qp_type == IB_QPT_RC ?
2953		IB_WC_RETRY_EXC_ERR : IB_WC_SUCCESS;
2954}
2955
2956/**
2957 * rvt_ruc_loopback - handle UC and RC loopback requests
2958 * @sqp: the sending QP
2959 *
2960 * This is called from rvt_do_send() to forward a WQE addressed to the same HFI
2961 * Note that although we are single threaded due to the send engine, we still
2962 * have to protect against post_send().  We don't have to worry about
2963 * receive interrupts since this is a connected protocol and all packets
2964 * will pass through here.
2965 */
2966void rvt_ruc_loopback(struct rvt_qp *sqp)
2967{
2968	struct rvt_ibport *rvp =  NULL;
2969	struct rvt_dev_info *rdi = ib_to_rvt(sqp->ibqp.device);
2970	struct rvt_qp *qp;
2971	struct rvt_swqe *wqe;
2972	struct rvt_sge *sge;
2973	unsigned long flags;
2974	struct ib_wc wc;
2975	u64 sdata;
2976	atomic64_t *maddr;
2977	enum ib_wc_status send_status;
2978	bool release;
2979	int ret;
2980	bool copy_last = false;
2981	int local_ops = 0;
2982
2983	rcu_read_lock();
2984	rvp = rdi->ports[sqp->port_num - 1];
2985
2986	/*
2987	 * Note that we check the responder QP state after
2988	 * checking the requester's state.
2989	 */
2990
2991	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), rvp,
2992			    sqp->remote_qpn);
2993
2994	spin_lock_irqsave(&sqp->s_lock, flags);
2995
2996	/* Return if we are already busy processing a work request. */
2997	if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
2998	    !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
2999		goto unlock;
3000
3001	sqp->s_flags |= RVT_S_BUSY;
3002
3003again:
3004	if (sqp->s_last == READ_ONCE(sqp->s_head))
3005		goto clr_busy;
3006	wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
3007
3008	/* Return if it is not OK to start a new work request. */
3009	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
3010		if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
3011			goto clr_busy;
3012		/* We are in the error state, flush the work request. */
3013		send_status = IB_WC_WR_FLUSH_ERR;
3014		goto flush_send;
3015	}
3016
3017	/*
3018	 * We can rely on the entry not changing without the s_lock
3019	 * being held until we update s_last.
3020	 * We increment s_cur to indicate s_last is in progress.
3021	 */
3022	if (sqp->s_last == sqp->s_cur) {
3023		if (++sqp->s_cur >= sqp->s_size)
3024			sqp->s_cur = 0;
3025	}
3026	spin_unlock_irqrestore(&sqp->s_lock, flags);
3027
3028	if (!qp) {
3029		send_status = loopback_qp_drop(rvp, sqp);
3030		goto serr_no_r_lock;
3031	}
3032	spin_lock_irqsave(&qp->r_lock, flags);
3033	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
3034	    qp->ibqp.qp_type != sqp->ibqp.qp_type) {
3035		send_status = loopback_qp_drop(rvp, sqp);
3036		goto serr;
3037	}
3038
3039	memset(&wc, 0, sizeof(wc));
3040	send_status = IB_WC_SUCCESS;
3041
3042	release = true;
3043	sqp->s_sge.sge = wqe->sg_list[0];
3044	sqp->s_sge.sg_list = wqe->sg_list + 1;
3045	sqp->s_sge.num_sge = wqe->wr.num_sge;
3046	sqp->s_len = wqe->length;
3047	switch (wqe->wr.opcode) {
3048	case IB_WR_REG_MR:
3049		goto send_comp;
3050
3051	case IB_WR_LOCAL_INV:
3052		if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
3053			if (rvt_invalidate_rkey(sqp,
3054						wqe->wr.ex.invalidate_rkey))
3055				send_status = IB_WC_LOC_PROT_ERR;
3056			local_ops = 1;
3057		}
3058		goto send_comp;
3059
3060	case IB_WR_SEND_WITH_INV:
3061	case IB_WR_SEND_WITH_IMM:
3062	case IB_WR_SEND:
3063		ret = rvt_get_rwqe(qp, false);
3064		if (ret < 0)
3065			goto op_err;
3066		if (!ret)
3067			goto rnr_nak;
3068		if (wqe->length > qp->r_len)
3069			goto inv_err;
3070		switch (wqe->wr.opcode) {
3071		case IB_WR_SEND_WITH_INV:
3072			if (!rvt_invalidate_rkey(qp,
3073						 wqe->wr.ex.invalidate_rkey)) {
3074				wc.wc_flags = IB_WC_WITH_INVALIDATE;
3075				wc.ex.invalidate_rkey =
3076					wqe->wr.ex.invalidate_rkey;
3077			}
3078			break;
3079		case IB_WR_SEND_WITH_IMM:
3080			wc.wc_flags = IB_WC_WITH_IMM;
3081			wc.ex.imm_data = wqe->wr.ex.imm_data;
3082			break;
3083		default:
3084			break;
3085		}
3086		break;
3087
3088	case IB_WR_RDMA_WRITE_WITH_IMM:
3089		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
3090			goto inv_err;
3091		wc.wc_flags = IB_WC_WITH_IMM;
3092		wc.ex.imm_data = wqe->wr.ex.imm_data;
3093		ret = rvt_get_rwqe(qp, true);
3094		if (ret < 0)
3095			goto op_err;
3096		if (!ret)
3097			goto rnr_nak;
3098		/* skip copy_last set and qp_access_flags recheck */
3099		goto do_write;
3100	case IB_WR_RDMA_WRITE:
3101		copy_last = rvt_is_user_qp(qp);
3102		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
3103			goto inv_err;
3104do_write:
3105		if (wqe->length == 0)
3106			break;
3107		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
3108					  wqe->rdma_wr.remote_addr,
3109					  wqe->rdma_wr.rkey,
3110					  IB_ACCESS_REMOTE_WRITE)))
3111			goto acc_err;
3112		qp->r_sge.sg_list = NULL;
3113		qp->r_sge.num_sge = 1;
3114		qp->r_sge.total_len = wqe->length;
3115		break;
3116
3117	case IB_WR_RDMA_READ:
3118		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
3119			goto inv_err;
3120		if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
3121					  wqe->rdma_wr.remote_addr,
3122					  wqe->rdma_wr.rkey,
3123					  IB_ACCESS_REMOTE_READ)))
3124			goto acc_err;
3125		release = false;
3126		sqp->s_sge.sg_list = NULL;
3127		sqp->s_sge.num_sge = 1;
3128		qp->r_sge.sge = wqe->sg_list[0];
3129		qp->r_sge.sg_list = wqe->sg_list + 1;
3130		qp->r_sge.num_sge = wqe->wr.num_sge;
3131		qp->r_sge.total_len = wqe->length;
3132		break;
3133
3134	case IB_WR_ATOMIC_CMP_AND_SWP:
3135	case IB_WR_ATOMIC_FETCH_AND_ADD:
3136		if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
3137			goto inv_err;
3138		if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
3139					  wqe->atomic_wr.remote_addr,
3140					  wqe->atomic_wr.rkey,
3141					  IB_ACCESS_REMOTE_ATOMIC)))
3142			goto acc_err;
3143		/* Perform atomic OP and save result. */
3144		maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
3145		sdata = wqe->atomic_wr.compare_add;
3146		*(u64 *)sqp->s_sge.sge.vaddr =
3147			(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
3148			(u64)atomic64_add_return(sdata, maddr) - sdata :
3149			(u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
3150				      sdata, wqe->atomic_wr.swap);
3151		rvt_put_mr(qp->r_sge.sge.mr);
3152		qp->r_sge.num_sge = 0;
3153		goto send_comp;
3154
3155	default:
3156		send_status = IB_WC_LOC_QP_OP_ERR;
3157		goto serr;
3158	}
3159
3160	sge = &sqp->s_sge.sge;
3161	while (sqp->s_len) {
3162		u32 len = rvt_get_sge_length(sge, sqp->s_len);
3163
3164		WARN_ON_ONCE(len == 0);
3165		rvt_copy_sge(qp, &qp->r_sge, sge->vaddr,
3166			     len, release, copy_last);
3167		rvt_update_sge(&sqp->s_sge, len, !release);
3168		sqp->s_len -= len;
3169	}
3170	if (release)
3171		rvt_put_ss(&qp->r_sge);
3172
3173	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
3174		goto send_comp;
3175
3176	if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
3177		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
3178	else
3179		wc.opcode = IB_WC_RECV;
3180	wc.wr_id = qp->r_wr_id;
3181	wc.status = IB_WC_SUCCESS;
3182	wc.byte_len = wqe->length;
3183	wc.qp = &qp->ibqp;
3184	wc.src_qp = qp->remote_qpn;
3185	wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX;
3186	wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
3187	wc.port_num = 1;
3188	/* Signal completion event if the solicited bit is set. */
3189	rvt_recv_cq(qp, &wc, wqe->wr.send_flags & IB_SEND_SOLICITED);
3190
3191send_comp:
3192	spin_unlock_irqrestore(&qp->r_lock, flags);
3193	spin_lock_irqsave(&sqp->s_lock, flags);
3194	rvp->n_loop_pkts++;
3195flush_send:
3196	sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
3197	rvt_send_complete(sqp, wqe, send_status);
3198	if (local_ops) {
3199		atomic_dec(&sqp->local_ops_pending);
3200		local_ops = 0;
3201	}
3202	goto again;
3203
3204rnr_nak:
3205	/* Handle RNR NAK */
3206	if (qp->ibqp.qp_type == IB_QPT_UC)
3207		goto send_comp;
3208	rvp->n_rnr_naks++;
3209	/*
3210	 * Note: we don't need the s_lock held since the BUSY flag
3211	 * makes this single threaded.
3212	 */
3213	if (sqp->s_rnr_retry == 0) {
3214		send_status = IB_WC_RNR_RETRY_EXC_ERR;
3215		goto serr;
3216	}
3217	if (sqp->s_rnr_retry_cnt < 7)
3218		sqp->s_rnr_retry--;
3219	spin_unlock_irqrestore(&qp->r_lock, flags);
3220	spin_lock_irqsave(&sqp->s_lock, flags);
3221	if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
3222		goto clr_busy;
3223	rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer <<
3224				IB_AETH_CREDIT_SHIFT);
3225	goto clr_busy;
3226
3227op_err:
3228	send_status = IB_WC_REM_OP_ERR;
3229	wc.status = IB_WC_LOC_QP_OP_ERR;
3230	goto err;
3231
3232inv_err:
3233	send_status =
3234		sqp->ibqp.qp_type == IB_QPT_RC ?
3235			IB_WC_REM_INV_REQ_ERR :
3236			IB_WC_SUCCESS;
3237	wc.status = IB_WC_LOC_QP_OP_ERR;
3238	goto err;
3239
3240acc_err:
3241	send_status = IB_WC_REM_ACCESS_ERR;
3242	wc.status = IB_WC_LOC_PROT_ERR;
3243err:
3244	/* responder goes to error state */
3245	rvt_rc_error(qp, wc.status);
3246
3247serr:
3248	spin_unlock_irqrestore(&qp->r_lock, flags);
3249serr_no_r_lock:
3250	spin_lock_irqsave(&sqp->s_lock, flags);
3251	rvt_send_complete(sqp, wqe, send_status);
3252	if (sqp->ibqp.qp_type == IB_QPT_RC) {
3253		int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
3254
3255		sqp->s_flags &= ~RVT_S_BUSY;
3256		spin_unlock_irqrestore(&sqp->s_lock, flags);
3257		if (lastwqe) {
3258			struct ib_event ev;
3259
3260			ev.device = sqp->ibqp.device;
3261			ev.element.qp = &sqp->ibqp;
3262			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
3263			sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
3264		}
3265		goto done;
3266	}
3267clr_busy:
3268	sqp->s_flags &= ~RVT_S_BUSY;
3269unlock:
3270	spin_unlock_irqrestore(&sqp->s_lock, flags);
3271done:
3272	rcu_read_unlock();
3273}
3274EXPORT_SYMBOL(rvt_ruc_loopback);