• Chris Mason's avatar
    RDS/IB: Add caching of frags and incs · 33244125
    Chris Mason authored
    This patch is based heavily on an initial patch by Chris Mason.
    Instead of freeing slab memory and pages, it keeps them, and
    funnels them back to be reused.
    
    The lock minimization strategy uses xchg and cmpxchg atomic ops
    for manipulation of pointers to list heads. We anchor the lists with a
    pointer to a list_head struct instead of a static list_head struct.
    We just have to carefully use the existing primitives with
    the difference between a pointer and a static head struct.
    
    For example, 'list_empty()' means that our anchor pointer points to a list with
    a single item instead of meaning that our static head element doesn't point to
    any list items.
    
    Original patch by Chris, with significant mods and fixes by Andy and Zach.
    Signed-off-by: default avatarChris Mason <chris.mason@oracle.com>
    Signed-off-by: default avatarAndy Grover <andy.grover@oracle.com>
    Signed-off-by: default avatarZach Brown <zach.brown@oracle.com>
    33244125
ib_recv.c 29.3 KB
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064
/*
 * Copyright (c) 2006 Oracle.  All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 */
#include <linux/kernel.h>
#include <linux/slab.h>
#include <linux/pci.h>
#include <linux/dma-mapping.h>
#include <rdma/rdma_cm.h>

#include "rds.h"
#include "ib.h"

static struct kmem_cache *rds_ib_incoming_slab;
static struct kmem_cache *rds_ib_frag_slab;
static atomic_t	rds_ib_allocation = ATOMIC_INIT(0);

void rds_ib_recv_init_ring(struct rds_ib_connection *ic)
{
	struct rds_ib_recv_work *recv;
	u32 i;

	for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) {
		struct ib_sge *sge;

		recv->r_ibinc = NULL;
		recv->r_frag = NULL;

		recv->r_wr.next = NULL;
		recv->r_wr.wr_id = i;
		recv->r_wr.sg_list = recv->r_sge;
		recv->r_wr.num_sge = RDS_IB_RECV_SGE;

		sge = &recv->r_sge[0];
		sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header));
		sge->length = sizeof(struct rds_header);
		sge->lkey = ic->i_mr->lkey;

		sge = &recv->r_sge[1];
		sge->addr = 0;
		sge->length = RDS_FRAG_SIZE;
		sge->lkey = ic->i_mr->lkey;
	}
}

/*
 * The entire 'from' list, including the from element itself, is put on
 * to the tail of the 'to' list.
 */
static void list_splice_entire_tail(struct list_head *from,
				    struct list_head *to)
{
	struct list_head *from_last = from->prev;

	list_splice_tail(from_last, to);
	list_add_tail(from_last, to);
}

static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache)
{
	struct list_head *tmp;

	tmp = xchg(&cache->xfer, NULL);
	if (tmp) {
		if (cache->ready)
			list_splice_entire_tail(tmp, cache->ready);
		else
			cache->ready = tmp;
	}
}

static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache)
{
	struct rds_ib_cache_head *head;
	int cpu;

	cache->percpu = alloc_percpu(struct rds_ib_cache_head);
	if (!cache->percpu)
	       return -ENOMEM;

	for_each_possible_cpu(cpu) {
		head = per_cpu_ptr(cache->percpu, cpu);
		head->first = NULL;
		head->count = 0;
	}
	cache->xfer = NULL;
	cache->ready = NULL;

	return 0;
}

int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic)
{
	int ret;

	ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs);
	if (!ret) {
		ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags);
		if (ret)
			free_percpu(ic->i_cache_incs.percpu);
	}

	return ret;
}

static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache,
					  struct list_head *caller_list)
{
	struct rds_ib_cache_head *head;
	int cpu;

	for_each_possible_cpu(cpu) {
		head = per_cpu_ptr(cache->percpu, cpu);
		if (head->first) {
			list_splice_entire_tail(head->first, caller_list);
			head->first = NULL;
		}
	}

	if (cache->ready) {
		list_splice_entire_tail(cache->ready, caller_list);
		cache->ready = NULL;
	}
}

void rds_ib_recv_free_caches(struct rds_ib_connection *ic)
{
	struct rds_ib_incoming *inc;
	struct rds_ib_incoming *inc_tmp;
	struct rds_page_frag *frag;
	struct rds_page_frag *frag_tmp;
	LIST_HEAD(list);

	rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
	rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list);
	free_percpu(ic->i_cache_incs.percpu);

	list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) {
		list_del(&inc->ii_cache_entry);
		WARN_ON(!list_empty(&inc->ii_frags));
		kmem_cache_free(rds_ib_incoming_slab, inc);
	}

	rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);
	rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list);
	free_percpu(ic->i_cache_frags.percpu);

	list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) {
		list_del(&frag->f_cache_entry);
		WARN_ON(!list_empty(&frag->f_item));
		kmem_cache_free(rds_ib_frag_slab, frag);
	}
}

/* fwd decl */
static void rds_ib_recv_cache_put(struct list_head *new_item,
				  struct rds_ib_refill_cache *cache);
static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache);


/* Recycle frag and attached recv buffer f_sg */
static void rds_ib_frag_free(struct rds_ib_connection *ic,
			     struct rds_page_frag *frag)
{
	rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg));

	rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags);
}

/* Recycle inc after freeing attached frags */
void rds_ib_inc_free(struct rds_incoming *inc)
{
	struct rds_ib_incoming *ibinc;
	struct rds_page_frag *frag;
	struct rds_page_frag *pos;
	struct rds_ib_connection *ic = inc->i_conn->c_transport_data;

	ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);

	/* Free attached frags */
	list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) {
		list_del_init(&frag->f_item);
		rds_ib_frag_free(ic, frag);
	}
	BUG_ON(!list_empty(&ibinc->ii_frags));

	rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc);
	rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs);
}

static void rds_ib_recv_clear_one(struct rds_ib_connection *ic,
				  struct rds_ib_recv_work *recv)
{
	if (recv->r_ibinc) {
		rds_inc_put(&recv->r_ibinc->ii_inc);
		recv->r_ibinc = NULL;
	}
	if (recv->r_frag) {
		ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);
		rds_ib_frag_free(ic, recv->r_frag);
		recv->r_frag = NULL;
	}
}

void rds_ib_recv_clear_ring(struct rds_ib_connection *ic)
{
	u32 i;

	for (i = 0; i < ic->i_recv_ring.w_nr; i++)
		rds_ib_recv_clear_one(ic, &ic->i_recvs[i]);
}

static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic)
{
	struct rds_ib_incoming *ibinc;
	struct list_head *cache_item;
	int avail_allocs;

	cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs);
	if (cache_item) {
		ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry);
	} else {
		avail_allocs = atomic_add_unless(&rds_ib_allocation,
						 1, rds_ib_sysctl_max_recv_allocation);
		if (!avail_allocs) {
			rds_ib_stats_inc(s_ib_rx_alloc_limit);
			return NULL;
		}
		ibinc = kmem_cache_alloc(rds_ib_incoming_slab, GFP_NOWAIT);
		if (!ibinc) {
			atomic_dec(&rds_ib_allocation);
			return NULL;
		}
	}
	INIT_LIST_HEAD(&ibinc->ii_frags);
	rds_inc_init(&ibinc->ii_inc, ic->conn, ic->conn->c_faddr);

	return ibinc;
}

static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic)
{
	struct rds_page_frag *frag;
	struct list_head *cache_item;
	int ret;

	cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags);
	if (cache_item) {
		frag = container_of(cache_item, struct rds_page_frag, f_cache_entry);
	} else {
		frag = kmem_cache_alloc(rds_ib_frag_slab, GFP_NOWAIT);
		if (!frag)
			return NULL;

		ret = rds_page_remainder_alloc(&frag->f_sg,
					       RDS_FRAG_SIZE, GFP_NOWAIT);
		if (ret) {
			kmem_cache_free(rds_ib_frag_slab, frag);
			return NULL;
		}
	}

	INIT_LIST_HEAD(&frag->f_item);

	return frag;
}

static int rds_ib_recv_refill_one(struct rds_connection *conn,
				  struct rds_ib_recv_work *recv)
{
	struct rds_ib_connection *ic = conn->c_transport_data;
	struct ib_sge *sge;
	int ret = -ENOMEM;

	if (!ic->i_cache_incs.ready)
		rds_ib_cache_xfer_to_ready(&ic->i_cache_incs);
	if (!ic->i_cache_frags.ready)
		rds_ib_cache_xfer_to_ready(&ic->i_cache_frags);

	/*
	 * ibinc was taken from recv if recv contained the start of a message.
	 * recvs that were continuations will still have this allocated.
	 */
	if (!recv->r_ibinc) {
		recv->r_ibinc = rds_ib_refill_one_inc(ic);
		if (!recv->r_ibinc)
			goto out;
	}

	WARN_ON(recv->r_frag); /* leak! */
	recv->r_frag = rds_ib_refill_one_frag(ic);
	if (!recv->r_frag)
		goto out;

	ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg,
			    1, DMA_FROM_DEVICE);
	WARN_ON(ret != 1);

	sge = &recv->r_sge[0];
	sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header);
	sge->length = sizeof(struct rds_header);

	sge = &recv->r_sge[1];
	sge->addr = sg_dma_address(&recv->r_frag->f_sg);
	sge->length = sg_dma_len(&recv->r_frag->f_sg);

	ret = 0;
out:
	return ret;
}

/*
 * This tries to allocate and post unused work requests after making sure that
 * they have all the allocations they need to queue received fragments into
 * sockets.
 *
 * -1 is returned if posting fails due to temporary resource exhaustion.
 */
int rds_ib_recv_refill(struct rds_connection *conn, int prefill)
{
	struct rds_ib_connection *ic = conn->c_transport_data;
	struct rds_ib_recv_work *recv;
	struct ib_recv_wr *failed_wr;
	unsigned int posted = 0;
	int ret = 0;
	u32 pos;

	while ((prefill || rds_conn_up(conn)) &&
	       rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) {
		if (pos >= ic->i_recv_ring.w_nr) {
			printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n",
					pos);
			ret = -EINVAL;
			break;
		}

		recv = &ic->i_recvs[pos];
		ret = rds_ib_recv_refill_one(conn, recv);
		if (ret) {
			ret = -1;
			break;
		}

		/* XXX when can this fail? */
		ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, &failed_wr);
		rdsdebug("recv %p ibinc %p page %p addr %lu ret %d\n", recv,
			 recv->r_ibinc, sg_page(&recv->r_frag->f_sg),
			 (long) sg_dma_address(&recv->r_frag->f_sg), ret);
		if (ret) {
			rds_ib_conn_error(conn, "recv post on "
			       "%pI4 returned %d, disconnecting and "
			       "reconnecting\n", &conn->c_faddr,
			       ret);
			ret = -1;
			break;
		}

		posted++;
	}

	/* We're doing flow control - update the window. */
	if (ic->i_flowctl && posted)
		rds_ib_advertise_credits(conn, posted);

	if (ret)
		rds_ib_ring_unalloc(&ic->i_recv_ring, 1);
	return ret;
}

/*
 * We want to recycle several types of recv allocations, like incs and frags.
 * To use this, the *_free() function passes in the ptr to a list_head within
 * the recyclee, as well as the cache to put it on.
 *
 * First, we put the memory on a percpu list. When this reaches a certain size,
 * We move it to an intermediate non-percpu list in a lockless manner, with some
 * xchg/compxchg wizardry.
 *
 * N.B. Instead of a list_head as the anchor, we use a single pointer, which can
 * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and
 * list_empty() will return true with one element is actually present.
 */
static void rds_ib_recv_cache_put(struct list_head *new_item,
				 struct rds_ib_refill_cache *cache)
{
	unsigned long flags;
	struct rds_ib_cache_head *chp;
	struct list_head *old;

	local_irq_save(flags);

	chp = per_cpu_ptr(cache->percpu, smp_processor_id());
	if (!chp->first)
		INIT_LIST_HEAD(new_item);
	else /* put on front */
		list_add_tail(new_item, chp->first);
	chp->first = new_item;
	chp->count++;

	if (chp->count < RDS_IB_RECYCLE_BATCH_COUNT)
		goto end;

	/*
	 * Return our per-cpu first list to the cache's xfer by atomically
	 * grabbing the current xfer list, appending it to our per-cpu list,
	 * and then atomically returning that entire list back to the
	 * cache's xfer list as long as it's still empty.
	 */
	do {
		old = xchg(&cache->xfer, NULL);
		if (old)
			list_splice_entire_tail(old, chp->first);
		old = cmpxchg(&cache->xfer, NULL, chp->first);
	} while (old);

	chp->first = NULL;
	chp->count = 0;
end:
	local_irq_restore(flags);
}

static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache)
{
	struct list_head *head = cache->ready;

	if (head) {
		if (!list_empty(head)) {
			cache->ready = head->next;
			list_del_init(head);
		} else
			cache->ready = NULL;
	}

	return head;
}

int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *first_iov,
			    size_t size)
{
	struct rds_ib_incoming *ibinc;
	struct rds_page_frag *frag;
	struct iovec *iov = first_iov;
	unsigned long to_copy;
	unsigned long frag_off = 0;
	unsigned long iov_off = 0;
	int copied = 0;
	int ret;
	u32 len;

	ibinc = container_of(inc, struct rds_ib_incoming, ii_inc);
	frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
	len = be32_to_cpu(inc->i_hdr.h_len);

	while (copied < size && copied < len) {
		if (frag_off == RDS_FRAG_SIZE) {
			frag = list_entry(frag->f_item.next,
					  struct rds_page_frag, f_item);
			frag_off = 0;
		}
		while (iov_off == iov->iov_len) {
			iov_off = 0;
			iov++;
		}

		to_copy = min(iov->iov_len - iov_off, RDS_FRAG_SIZE - frag_off);
		to_copy = min_t(size_t, to_copy, size - copied);
		to_copy = min_t(unsigned long, to_copy, len - copied);

		rdsdebug("%lu bytes to user [%p, %zu] + %lu from frag "
			 "[%p, %u] + %lu\n",
			 to_copy, iov->iov_base, iov->iov_len, iov_off,
			 sg_page(&frag->f_sg), frag->f_sg.offset, frag_off);

		/* XXX needs + offset for multiple recvs per page */
		ret = rds_page_copy_to_user(sg_page(&frag->f_sg),
					    frag->f_sg.offset + frag_off,
					    iov->iov_base + iov_off,
					    to_copy);
		if (ret) {
			copied = ret;
			break;
		}

		iov_off += to_copy;
		frag_off += to_copy;
		copied += to_copy;
	}

	return copied;
}

/* ic starts out kzalloc()ed */
void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
{
	struct ib_send_wr *wr = &ic->i_ack_wr;
	struct ib_sge *sge = &ic->i_ack_sge;

	sge->addr = ic->i_ack_dma;
	sge->length = sizeof(struct rds_header);
	sge->lkey = ic->i_mr->lkey;

	wr->sg_list = sge;
	wr->num_sge = 1;
	wr->opcode = IB_WR_SEND;
	wr->wr_id = RDS_IB_ACK_WR_ID;
	wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED;
}

/*
 * You'd think that with reliable IB connections you wouldn't need to ack
 * messages that have been received.  The problem is that IB hardware generates
 * an ack message before it has DMAed the message into memory.  This creates a
 * potential message loss if the HCA is disabled for any reason between when it
 * sends the ack and before the message is DMAed and processed.  This is only a
 * potential issue if another HCA is available for fail-over.
 *
 * When the remote host receives our ack they'll free the sent message from
 * their send queue.  To decrease the latency of this we always send an ack
 * immediately after we've received messages.
 *
 * For simplicity, we only have one ack in flight at a time.  This puts
 * pressure on senders to have deep enough send queues to absorb the latency of
 * a single ack frame being in flight.  This might not be good enough.
 *
 * This is implemented by have a long-lived send_wr and sge which point to a
 * statically allocated ack frame.  This ack wr does not fall under the ring
 * accounting that the tx and rx wrs do.  The QP attribute specifically makes
 * room for it beyond the ring size.  Send completion notices its special
 * wr_id and avoids working with the ring in that case.
 */
#ifndef KERNEL_HAS_ATOMIC64
static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
				int ack_required)
{
	unsigned long flags;

	spin_lock_irqsave(&ic->i_ack_lock, flags);
	ic->i_ack_next = seq;
	if (ack_required)
		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
	spin_unlock_irqrestore(&ic->i_ack_lock, flags);
}

static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
{
	unsigned long flags;
	u64 seq;

	clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);

	spin_lock_irqsave(&ic->i_ack_lock, flags);
	seq = ic->i_ack_next;
	spin_unlock_irqrestore(&ic->i_ack_lock, flags);

	return seq;
}
#else
static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
				int ack_required)
{
	atomic64_set(&ic->i_ack_next, seq);
	if (ack_required) {
		smp_mb__before_clear_bit();
		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
	}
}

static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
{
	clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
	smp_mb__after_clear_bit();

	return atomic64_read(&ic->i_ack_next);
}
#endif


static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
{
	struct rds_header *hdr = ic->i_ack;
	struct ib_send_wr *failed_wr;
	u64 seq;
	int ret;

	seq = rds_ib_get_ack(ic);

	rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq);
	rds_message_populate_header(hdr, 0, 0, 0);
	hdr->h_ack = cpu_to_be64(seq);
	hdr->h_credit = adv_credits;
	rds_message_make_checksum(hdr);
	ic->i_ack_queued = jiffies;

	ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, &failed_wr);
	if (unlikely(ret)) {
		/* Failed to send. Release the WR, and
		 * force another ACK.
		 */
		clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);

		rds_ib_stats_inc(s_ib_ack_send_failure);

		rds_ib_conn_error(ic->conn, "sending ack failed\n");
	} else
		rds_ib_stats_inc(s_ib_ack_sent);
}

/*
 * There are 3 ways of getting acknowledgements to the peer:
 *  1.	We call rds_ib_attempt_ack from the recv completion handler
 *	to send an ACK-only frame.
 *	However, there can be only one such frame in the send queue
 *	at any time, so we may have to postpone it.
 *  2.	When another (data) packet is transmitted while there's
 *	an ACK in the queue, we piggyback the ACK sequence number
 *	on the data packet.
 *  3.	If the ACK WR is done sending, we get called from the
 *	send queue completion handler, and check whether there's
 *	another ACK pending (postponed because the WR was on the
 *	queue). If so, we transmit it.
 *
 * We maintain 2 variables:
 *  -	i_ack_flags, which keeps track of whether the ACK WR
 *	is currently in the send queue or not (IB_ACK_IN_FLIGHT)
 *  -	i_ack_next, which is the last sequence number we received
 *
 * Potentially, send queue and receive queue handlers can run concurrently.
 * It would be nice to not have to use a spinlock to synchronize things,
 * but the one problem that rules this out is that 64bit updates are
 * not atomic on all platforms. Things would be a lot simpler if
 * we had atomic64 or maybe cmpxchg64 everywhere.
 *
 * Reconnecting complicates this picture just slightly. When we
 * reconnect, we may be seeing duplicate packets. The peer
 * is retransmitting them, because it hasn't seen an ACK for
 * them. It is important that we ACK these.
 *
 * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with
 * this flag set *MUST* be acknowledged immediately.
 */

/*
 * When we get here, we're called from the recv queue handler.
 * Check whether we ought to transmit an ACK.
 */
void rds_ib_attempt_ack(struct rds_ib_connection *ic)
{
	unsigned int adv_credits;

	if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
		return;

	if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) {
		rds_ib_stats_inc(s_ib_ack_send_delayed);
		return;
	}

	/* Can we get a send credit? */
	if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) {
		rds_ib_stats_inc(s_ib_tx_throttle);
		clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
		return;
	}

	clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
	rds_ib_send_ack(ic, adv_credits);
}

/*
 * We get here from the send completion handler, when the
 * adapter tells us the ACK frame was sent.
 */
void rds_ib_ack_send_complete(struct rds_ib_connection *ic)
{
	clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
	rds_ib_attempt_ack(ic);
}

/*
 * This is called by the regular xmit code when it wants to piggyback
 * an ACK on an outgoing frame.
 */
u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic)
{
	if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags))
		rds_ib_stats_inc(s_ib_ack_send_piggybacked);
	return rds_ib_get_ack(ic);
}

/*
 * It's kind of lame that we're copying from the posted receive pages into
 * long-lived bitmaps.  We could have posted the bitmaps and rdma written into
 * them.  But receiving new congestion bitmaps should be a *rare* event, so
 * hopefully we won't need to invest that complexity in making it more
 * efficient.  By copying we can share a simpler core with TCP which has to
 * copy.
 */
static void rds_ib_cong_recv(struct rds_connection *conn,
			      struct rds_ib_incoming *ibinc)
{
	struct rds_cong_map *map;
	unsigned int map_off;
	unsigned int map_page;
	struct rds_page_frag *frag;
	unsigned long frag_off;
	unsigned long to_copy;
	unsigned long copied;
	uint64_t uncongested = 0;
	void *addr;

	/* catch completely corrupt packets */
	if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES)
		return;

	map = conn->c_fcong;
	map_page = 0;
	map_off = 0;

	frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item);
	frag_off = 0;

	copied = 0;

	while (copied < RDS_CONG_MAP_BYTES) {
		uint64_t *src, *dst;
		unsigned int k;

		to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off);
		BUG_ON(to_copy & 7); /* Must be 64bit aligned. */

		addr = kmap_atomic(sg_page(&frag->f_sg), KM_SOFTIRQ0);

		src = addr + frag_off;
		dst = (void *)map->m_page_addrs[map_page] + map_off;
		for (k = 0; k < to_copy; k += 8) {
			/* Record ports that became uncongested, ie
			 * bits that changed from 0 to 1. */
			uncongested |= ~(*src) & *dst;
			*dst++ = *src++;
		}
		kunmap_atomic(addr, KM_SOFTIRQ0);

		copied += to_copy;

		map_off += to_copy;
		if (map_off == PAGE_SIZE) {
			map_off = 0;
			map_page++;
		}

		frag_off += to_copy;
		if (frag_off == RDS_FRAG_SIZE) {
			frag = list_entry(frag->f_item.next,
					  struct rds_page_frag, f_item);
			frag_off = 0;
		}
	}

	/* the congestion map is in little endian order */
	uncongested = le64_to_cpu(uncongested);

	rds_cong_map_updated(map, uncongested);
}

/*
 * Rings are posted with all the allocations they'll need to queue the
 * incoming message to the receiving socket so this can't fail.
 * All fragments start with a header, so we can make sure we're not receiving
 * garbage, and we can tell a small 8 byte fragment from an ACK frame.
 */
struct rds_ib_ack_state {
	u64		ack_next;
	u64		ack_recv;
	unsigned int	ack_required:1;
	unsigned int	ack_next_valid:1;
	unsigned int	ack_recv_valid:1;
};

static void rds_ib_process_recv(struct rds_connection *conn,
				struct rds_ib_recv_work *recv, u32 data_len,
				struct rds_ib_ack_state *state)
{
	struct rds_ib_connection *ic = conn->c_transport_data;
	struct rds_ib_incoming *ibinc = ic->i_ibinc;
	struct rds_header *ihdr, *hdr;

	/* XXX shut down the connection if port 0,0 are seen? */

	rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv,
		 data_len);

	if (data_len < sizeof(struct rds_header)) {
		rds_ib_conn_error(conn, "incoming message "
		       "from %pI4 didn't inclue a "
		       "header, disconnecting and "
		       "reconnecting\n",
		       &conn->c_faddr);
		return;
	}
	data_len -= sizeof(struct rds_header);

	ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs];

	/* Validate the checksum. */
	if (!rds_message_verify_checksum(ihdr)) {
		rds_ib_conn_error(conn, "incoming message "
		       "from %pI4 has corrupted header - "
		       "forcing a reconnect\n",
		       &conn->c_faddr);
		rds_stats_inc(s_recv_drop_bad_checksum);
		return;
	}

	/* Process the ACK sequence which comes with every packet */
	state->ack_recv = be64_to_cpu(ihdr->h_ack);
	state->ack_recv_valid = 1;

	/* Process the credits update if there was one */
	if (ihdr->h_credit)
		rds_ib_send_add_credits(conn, ihdr->h_credit);

	if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) {
		/* This is an ACK-only packet. The fact that it gets
		 * special treatment here is that historically, ACKs
		 * were rather special beasts.
		 */
		rds_ib_stats_inc(s_ib_ack_received);

		/*
		 * Usually the frags make their way on to incs and are then freed as
		 * the inc is freed.  We don't go that route, so we have to drop the
		 * page ref ourselves.  We can't just leave the page on the recv
		 * because that confuses the dma mapping of pages and each recv's use
		 * of a partial page.
		 *
		 * FIXME: Fold this into the code path below.
		 */
		rds_ib_frag_free(ic, recv->r_frag);
		recv->r_frag = NULL;
		return;
	}

	/*
	 * If we don't already have an inc on the connection then this
	 * fragment has a header and starts a message.. copy its header
	 * into the inc and save the inc so we can hang upcoming fragments
	 * off its list.
	 */
	if (!ibinc) {
		ibinc = recv->r_ibinc;
		recv->r_ibinc = NULL;
		ic->i_ibinc = ibinc;

		hdr = &ibinc->ii_inc.i_hdr;
		memcpy(hdr, ihdr, sizeof(*hdr));
		ic->i_recv_data_rem = be32_to_cpu(hdr->h_len);

		rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc,
			 ic->i_recv_data_rem, hdr->h_flags);
	} else {
		hdr = &ibinc->ii_inc.i_hdr;
		/* We can't just use memcmp here; fragments of a
		 * single message may carry different ACKs */
		if (hdr->h_sequence != ihdr->h_sequence ||
		    hdr->h_len != ihdr->h_len ||
		    hdr->h_sport != ihdr->h_sport ||
		    hdr->h_dport != ihdr->h_dport) {
			rds_ib_conn_error(conn,
				"fragment header mismatch; forcing reconnect\n");
			return;
		}
	}

	list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags);
	recv->r_frag = NULL;

	if (ic->i_recv_data_rem > RDS_FRAG_SIZE)
		ic->i_recv_data_rem -= RDS_FRAG_SIZE;
	else {
		ic->i_recv_data_rem = 0;
		ic->i_ibinc = NULL;

		if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
			rds_ib_cong_recv(conn, ibinc);
		else {
			rds_recv_incoming(conn, conn->c_faddr, conn->c_laddr,
					  &ibinc->ii_inc, GFP_ATOMIC,
					  KM_SOFTIRQ0);
			state->ack_next = be64_to_cpu(hdr->h_sequence);
			state->ack_next_valid = 1;
		}

		/* Evaluate the ACK_REQUIRED flag *after* we received
		 * the complete frame, and after bumping the next_rx
		 * sequence. */
		if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) {
			rds_stats_inc(s_recv_ack_required);
			state->ack_required = 1;
		}

		rds_inc_put(&ibinc->ii_inc);
	}
}

/*
 * Plucking the oldest entry from the ring can be done concurrently with
 * the thread refilling the ring.  Each ring operation is protected by
 * spinlocks and the transient state of refilling doesn't change the
 * recording of which entry is oldest.
 *
 * This relies on IB only calling one cq comp_handler for each cq so that
 * there will only be one caller of rds_recv_incoming() per RDS connection.
 */
void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context)
{
	struct rds_connection *conn = context;
	struct rds_ib_connection *ic = conn->c_transport_data;

	rdsdebug("conn %p cq %p\n", conn, cq);

	rds_ib_stats_inc(s_ib_rx_cq_call);

	tasklet_schedule(&ic->i_recv_tasklet);
}

static inline void rds_poll_cq(struct rds_ib_connection *ic,
			       struct rds_ib_ack_state *state)
{
	struct rds_connection *conn = ic->conn;
	struct ib_wc wc;
	struct rds_ib_recv_work *recv;

	while (ib_poll_cq(ic->i_recv_cq, 1, &wc) > 0) {
		rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
			 (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
			 be32_to_cpu(wc.ex.imm_data));
		rds_ib_stats_inc(s_ib_rx_cq_event);

		recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)];

		ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE);

		/*
		 * Also process recvs in connecting state because it is possible
		 * to get a recv completion _before_ the rdmacm ESTABLISHED
		 * event is processed.
		 */
		if (rds_conn_up(conn) || rds_conn_connecting(conn)) {
			/* We expect errors as the qp is drained during shutdown */
			if (wc.status == IB_WC_SUCCESS) {
				rds_ib_process_recv(conn, recv, wc.byte_len, state);
			} else {
				rds_ib_conn_error(conn, "recv completion on "
				       "%pI4 had status %u, disconnecting and "
				       "reconnecting\n", &conn->c_faddr,
				       wc.status);
			}
		}

		rds_ib_ring_free(&ic->i_recv_ring, 1);
	}
}

void rds_ib_recv_tasklet_fn(unsigned long data)
{
	struct rds_ib_connection *ic = (struct rds_ib_connection *) data;
	struct rds_connection *conn = ic->conn;
	struct rds_ib_ack_state state = { 0, };

	rds_poll_cq(ic, &state);
	ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
	rds_poll_cq(ic, &state);

	if (state.ack_next_valid)
		rds_ib_set_ack(ic, state.ack_next, state.ack_required);
	if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
		rds_send_drop_acked(conn, state.ack_recv, NULL);
		ic->i_ack_recv = state.ack_recv;
	}
	if (rds_conn_up(conn))
		rds_ib_attempt_ack(ic);

	/* If we ever end up with a really empty receive ring, we're
	 * in deep trouble, as the sender will definitely see RNR
	 * timeouts. */
	if (rds_ib_ring_empty(&ic->i_recv_ring))
		rds_ib_stats_inc(s_ib_rx_ring_empty);

	if (rds_ib_ring_low(&ic->i_recv_ring))
		rds_ib_recv_refill(conn, 0);
}

int rds_ib_recv(struct rds_connection *conn)
{
	struct rds_ib_connection *ic = conn->c_transport_data;
	int ret = 0;

	rdsdebug("conn %p\n", conn);
	if (rds_conn_up(conn))
		rds_ib_attempt_ack(ic);

	return ret;
}

int __init rds_ib_recv_init(void)
{
	struct sysinfo si;
	int ret = -ENOMEM;

	/* Default to 30% of all available RAM for recv memory */
	si_meminfo(&si);
	rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE;

	rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming",
					sizeof(struct rds_ib_incoming),
					0, 0, NULL);
	if (!rds_ib_incoming_slab)
		goto out;

	rds_ib_frag_slab = kmem_cache_create("rds_ib_frag",
					sizeof(struct rds_page_frag),
					0, 0, NULL);
	if (!rds_ib_frag_slab)
		kmem_cache_destroy(rds_ib_incoming_slab);
	else
		ret = 0;
out:
	return ret;
}

void rds_ib_recv_exit(void)
{
	kmem_cache_destroy(rds_ib_incoming_slab);
	kmem_cache_destroy(rds_ib_frag_slab);
}