send.c 31.3 KB
Newer Older
A
Andy Grover 已提交
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
/*
 * 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>
34
#include <linux/gfp.h>
A
Andy Grover 已提交
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
#include <net/sock.h>
#include <linux/in.h>
#include <linux/list.h>

#include "rds.h"

/* When transmitting messages in rds_send_xmit, we need to emerge from
 * time to time and briefly release the CPU. Otherwise the softlock watchdog
 * will kick our shin.
 * Also, it seems fairer to not let one busy connection stall all the
 * others.
 *
 * send_batch_count is the number of times we'll loop in send_xmit. Setting
 * it to 0 will restore the old behavior (where we looped until we had
 * drained the queue).
 */
static int send_batch_count = 64;
module_param(send_batch_count, int, 0444);
MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");

/*
 * Reset the send state. Caller must hold c_send_lock when calling here.
 */
void rds_send_reset(struct rds_connection *conn)
{
	struct rds_message *rm, *tmp;
	unsigned long flags;

	if (conn->c_xmit_rm) {
		/* Tell the user the RDMA op is no longer mapped by the
		 * transport. This isn't entirely true (it's flushed out
		 * independently) but as the connection is down, there's
		 * no ongoing RDMA to/from that memory */
		rds_message_unmapped(conn->c_xmit_rm);
		rds_message_put(conn->c_xmit_rm);
		conn->c_xmit_rm = NULL;
	}
	conn->c_xmit_sg = 0;
	conn->c_xmit_hdr_off = 0;
	conn->c_xmit_data_off = 0;
A
Andy Grover 已提交
75
	conn->c_xmit_atomic_sent = 0;
A
Andy Grover 已提交
76 77
	conn->c_xmit_rdma_sent = 0;
	conn->c_xmit_data_sent = 0;
A
Andy Grover 已提交
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

	conn->c_map_queued = 0;

	conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
	conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;

	/* Mark messages as retransmissions, and move them to the send q */
	spin_lock_irqsave(&conn->c_lock, flags);
	list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
		set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
		set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
	}
	list_splice_init(&conn->c_retrans, &conn->c_send_queue);
	spin_unlock_irqrestore(&conn->c_lock, flags);
}

/*
 * We're making the concious trade-off here to only send one message
 * down the connection at a time.
 *   Pro:
 *      - tx queueing is a simple fifo list
 *   	- reassembly is optional and easily done by transports per conn
 *      - no per flow rx lookup at all, straight to the socket
 *   	- less per-frag memory and wire overhead
 *   Con:
 *      - queued acks can be delayed behind large messages
 *   Depends:
 *      - small message latency is higher behind queued large messages
 *      - large message latency isn't starved by intervening small sends
 */
int rds_send_xmit(struct rds_connection *conn)
{
	struct rds_message *rm;
	unsigned long flags;
	unsigned int tmp;
	unsigned int send_quota = send_batch_count;
	struct scatterlist *sg;
	int ret = 0;
	int was_empty = 0;
	LIST_HEAD(to_be_dropped);

	/*
	 * sendmsg calls here after having queued its message on the send
	 * queue.  We only have one task feeding the connection at a time.  If
	 * another thread is already feeding the queue then we back off.  This
	 * avoids blocking the caller and trading per-connection data between
	 * caches per message.
	 *
	 * The sem holder will issue a retry if they notice that someone queued
	 * a message after they stopped walking the send queue but before they
	 * dropped the sem.
	 */
	if (!mutex_trylock(&conn->c_send_lock)) {
		rds_stats_inc(s_send_sem_contention);
		ret = -ENOMEM;
		goto out;
	}

	if (conn->c_trans->xmit_prepare)
		conn->c_trans->xmit_prepare(conn);

	/*
	 * spin trying to push headers and data down the connection until
A
Andy Grover 已提交
141
	 * the connection doesn't make forward progress.
A
Andy Grover 已提交
142 143 144 145 146
	 */
	while (--send_quota) {

		rm = conn->c_xmit_rm;

A
Andy Grover 已提交
147 148 149 150 151 152 153
		/*
		 * If between sending messages, we can send a pending congestion
		 * map update.
		 *
		 * Transports either define a special xmit_cong_map function,
		 * or we allocate a cong_map message and treat it just like any
		 * other send.
A
Andy Grover 已提交
154
		 */
155 156
		if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
			if (conn->c_trans->xmit_cong_map) {
A
Andy Grover 已提交
157 158
				unsigned long map_offset = 0;
				unsigned long map_bytes = sizeof(struct rds_header) +
A
Andy Grover 已提交
159 160
					RDS_CONG_MAP_BYTES;

A
Andy Grover 已提交
161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
				while (map_bytes) {
					ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
									   map_offset);
					if (ret <= 0) {
						/* too far down the rabbithole! */
						mutex_unlock(&conn->c_send_lock);
						rds_conn_error(conn, "Cong map xmit failed\n");
						goto out;
					}

					map_offset += ret;
					map_bytes -= ret;
				}
			} else {
				/* send cong update like a normal rm */
				rm = rds_cong_update_alloc(conn);
				if (IS_ERR(rm)) {
					ret = PTR_ERR(rm);
					break;
				}
				rm->data.op_active = 1;
A
Andy Grover 已提交
182

A
Andy Grover 已提交
183 184
				conn->c_xmit_rm = rm;
			}
A
Andy Grover 已提交
185 186 187
		}

		/*
A
Andy Grover 已提交
188
		 * If not already working on one, grab the next message.
A
Andy Grover 已提交
189 190 191 192 193
		 *
		 * c_xmit_rm holds a ref while we're sending this message down
		 * the connction.  We can use this ref while holding the
		 * send_sem.. rds_send_reset() is serialized with it.
		 */
194
		if (!rm) {
A
Andy Grover 已提交
195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213
			unsigned int len;

			spin_lock_irqsave(&conn->c_lock, flags);

			if (!list_empty(&conn->c_send_queue)) {
				rm = list_entry(conn->c_send_queue.next,
						struct rds_message,
						m_conn_item);
				rds_message_addref(rm);

				/*
				 * Move the message from the send queue to the retransmit
				 * list right away.
				 */
				list_move_tail(&rm->m_conn_item, &conn->c_retrans);
			}

			spin_unlock_irqrestore(&conn->c_lock, flags);

214
			if (!rm) {
A
Andy Grover 已提交
215 216 217 218 219 220 221 222 223 224 225
				was_empty = 1;
				break;
			}

			/* Unfortunately, the way Infiniband deals with
			 * RDMA to a bad MR key is by moving the entire
			 * queue pair to error state. We cold possibly
			 * recover from that, but right now we drop the
			 * connection.
			 * Therefore, we never retransmit messages with RDMA ops.
			 */
A
Andy Grover 已提交
226
			if (rm->rdma.op_active &&
227
			    test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
A
Andy Grover 已提交
228 229 230 231 232 233 234 235 236 237
				spin_lock_irqsave(&conn->c_lock, flags);
				if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
					list_move(&rm->m_conn_item, &to_be_dropped);
				spin_unlock_irqrestore(&conn->c_lock, flags);
				rds_message_put(rm);
				continue;
			}

			/* Require an ACK every once in a while */
			len = ntohl(rm->m_inc.i_hdr.h_len);
238 239
			if (conn->c_unacked_packets == 0 ||
			    conn->c_unacked_bytes < len) {
A
Andy Grover 已提交
240 241 242 243 244 245 246 247 248 249 250 251 252
				__set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);

				conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
				conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
				rds_stats_inc(s_send_ack_required);
			} else {
				conn->c_unacked_bytes -= len;
				conn->c_unacked_packets--;
			}

			conn->c_xmit_rm = rm;
		}

253 254
		/* The transport either sends the whole rdma or none of it */
		if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
255
			rm->m_final_op = &rm->rdma;
256
			ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
A
Andy Grover 已提交
257 258
			if (ret)
				break;
259 260
			conn->c_xmit_rdma_sent = 1;

A
Andy Grover 已提交
261 262 263 264 265
			/* The transport owns the mapped memory for now.
			 * You can't unmap it while it's on the send queue */
			set_bit(RDS_MSG_MAPPED, &rm->m_flags);
		}

266
		if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
267 268
			rm->m_final_op = &rm->atomic;
			ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
A
Andy Grover 已提交
269 270
			if (ret)
				break;
271
			conn->c_xmit_atomic_sent = 1;
272

A
Andy Grover 已提交
273 274 275 276 277
			/* The transport owns the mapped memory for now.
			 * You can't unmap it while it's on the send queue */
			set_bit(RDS_MSG_MAPPED, &rm->m_flags);
		}

278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299
		/*
		 * A number of cases require an RDS header to be sent
		 * even if there is no data.
		 * We permit 0-byte sends; rds-ping depends on this.
		 * However, if there are exclusively attached silent ops,
		 * we skip the hdr/data send, to enable silent operation.
		 */
		if (rm->data.op_nents == 0) {
			int ops_present;
			int all_ops_are_silent = 1;

			ops_present = (rm->atomic.op_active || rm->rdma.op_active);
			if (rm->atomic.op_active && !rm->atomic.op_silent)
				all_ops_are_silent = 0;
			if (rm->rdma.op_active && !rm->rdma.op_silent)
				all_ops_are_silent = 0;

			if (ops_present && all_ops_are_silent
			    && !rm->m_rdma_cookie)
				rm->data.op_active = 0;
		}

A
Andy Grover 已提交
300
		if (rm->data.op_active && !conn->c_xmit_data_sent) {
301
			rm->m_final_op = &rm->data;
A
Andy Grover 已提交
302 303 304 305 306 307 308 309 310 311 312 313 314 315 316
			ret = conn->c_trans->xmit(conn, rm,
						  conn->c_xmit_hdr_off,
						  conn->c_xmit_sg,
						  conn->c_xmit_data_off);
			if (ret <= 0)
				break;

			if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
				tmp = min_t(int, ret,
					    sizeof(struct rds_header) -
					    conn->c_xmit_hdr_off);
				conn->c_xmit_hdr_off += tmp;
				ret -= tmp;
			}

317
			sg = &rm->data.op_sg[conn->c_xmit_sg];
A
Andy Grover 已提交
318 319 320 321 322 323 324 325 326 327
			while (ret) {
				tmp = min_t(int, ret, sg->length -
						      conn->c_xmit_data_off);
				conn->c_xmit_data_off += tmp;
				ret -= tmp;
				if (conn->c_xmit_data_off == sg->length) {
					conn->c_xmit_data_off = 0;
					sg++;
					conn->c_xmit_sg++;
					BUG_ON(ret != 0 &&
328
					       conn->c_xmit_sg == rm->data.op_nents);
A
Andy Grover 已提交
329 330
				}
			}
A
Andy Grover 已提交
331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351

			if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
			    (conn->c_xmit_sg == rm->data.op_nents))
				conn->c_xmit_data_sent = 1;
		}

		/*
		 * A rm will only take multiple times through this loop
		 * if there is a data op. Thus, if the data is sent (or there was
		 * none), then we're done with the rm.
		 */
		if (!rm->data.op_active || conn->c_xmit_data_sent) {
			conn->c_xmit_rm = NULL;
			conn->c_xmit_sg = 0;
			conn->c_xmit_hdr_off = 0;
			conn->c_xmit_data_off = 0;
			conn->c_xmit_rdma_sent = 0;
			conn->c_xmit_atomic_sent = 0;
			conn->c_xmit_data_sent = 0;

			rds_message_put(rm);
A
Andy Grover 已提交
352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375
		}
	}

	/* Nuke any messages we decided not to retransmit. */
	if (!list_empty(&to_be_dropped))
		rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);

	if (conn->c_trans->xmit_complete)
		conn->c_trans->xmit_complete(conn);

	/*
	 * We might be racing with another sender who queued a message but
	 * backed off on noticing that we held the c_send_lock.  If we check
	 * for queued messages after dropping the sem then either we'll
	 * see the queued message or the queuer will get the sem.  If we
	 * notice the queued message then we trigger an immediate retry.
	 *
	 * We need to be careful only to do this when we stopped processing
	 * the send queue because it was empty.  It's the only way we
	 * stop processing the loop when the transport hasn't taken
	 * responsibility for forward progress.
	 */
	mutex_unlock(&conn->c_send_lock);

A
Andy Grover 已提交
376
	if (send_quota == 0 && !was_empty) {
A
Andy Grover 已提交
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
		/* We exhausted the send quota, but there's work left to
		 * do. Return and (re-)schedule the send worker.
		 */
		ret = -EAGAIN;
	}

	if (ret == 0 && was_empty) {
		/* A simple bit test would be way faster than taking the
		 * spin lock */
		spin_lock_irqsave(&conn->c_lock, flags);
		if (!list_empty(&conn->c_send_queue)) {
			rds_stats_inc(s_send_sem_queue_raced);
			ret = -EAGAIN;
		}
		spin_unlock_irqrestore(&conn->c_lock, flags);
	}
out:
	return ret;
}

static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
{
	u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);

	assert_spin_locked(&rs->rs_lock);

	BUG_ON(rs->rs_snd_bytes < len);
	rs->rs_snd_bytes -= len;

	if (rs->rs_snd_bytes == 0)
		rds_stats_inc(s_send_queue_empty);
}

static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
				    is_acked_func is_acked)
{
	if (is_acked)
		return is_acked(rm, ack);
	return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
}

/*
 * Returns true if there are no messages on the send and retransmit queues
 * which have a sequence number greater than or equal to the given sequence
 * number.
 */
int rds_send_acked_before(struct rds_connection *conn, u64 seq)
{
	struct rds_message *rm, *tmp;
	int ret = 1;

	spin_lock(&conn->c_lock);

	list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
		if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
			ret = 0;
		break;
	}

	list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
		if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
			ret = 0;
		break;
	}

	spin_unlock(&conn->c_lock);

	return ret;
}

/*
 * This is pretty similar to what happens below in the ACK
 * handling code - except that we call here as soon as we get
 * the IB send completion on the RDMA op and the accompanying
 * message.
 */
void rds_rdma_send_complete(struct rds_message *rm, int status)
{
	struct rds_sock *rs = NULL;
A
Andy Grover 已提交
456
	struct rm_rdma_op *ro;
A
Andy Grover 已提交
457
	struct rds_notifier *notifier;
458
	unsigned long flags;
A
Andy Grover 已提交
459

460
	spin_lock_irqsave(&rm->m_rs_lock, flags);
A
Andy Grover 已提交
461

A
Andy Grover 已提交
462
	ro = &rm->rdma;
463
	if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
A
Andy Grover 已提交
464 465
	    ro->op_active && ro->op_notify && ro->op_notifier) {
		notifier = ro->op_notifier;
A
Andy Grover 已提交
466 467 468 469 470 471 472 473
		rs = rm->m_rs;
		sock_hold(rds_rs_to_sk(rs));

		notifier->n_status = status;
		spin_lock(&rs->rs_lock);
		list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
		spin_unlock(&rs->rs_lock);

A
Andy Grover 已提交
474
		ro->op_notifier = NULL;
A
Andy Grover 已提交
475 476
	}

477
	spin_unlock_irqrestore(&rm->m_rs_lock, flags);
A
Andy Grover 已提交
478 479 480 481 482 483

	if (rs) {
		rds_wake_sk_sleep(rs);
		sock_put(rds_rs_to_sk(rs));
	}
}
A
Andy Grover 已提交
484
EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
A
Andy Grover 已提交
485

A
Andy Grover 已提交
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
/*
 * Just like above, except looks at atomic op
 */
void rds_atomic_send_complete(struct rds_message *rm, int status)
{
	struct rds_sock *rs = NULL;
	struct rm_atomic_op *ao;
	struct rds_notifier *notifier;

	spin_lock(&rm->m_rs_lock);

	ao = &rm->atomic;
	if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
	    && ao->op_active && ao->op_notify && ao->op_notifier) {
		notifier = ao->op_notifier;
		rs = rm->m_rs;
		sock_hold(rds_rs_to_sk(rs));

		notifier->n_status = status;
		spin_lock(&rs->rs_lock);
		list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
		spin_unlock(&rs->rs_lock);

		ao->op_notifier = NULL;
	}

	spin_unlock(&rm->m_rs_lock);

	if (rs) {
		rds_wake_sk_sleep(rs);
		sock_put(rds_rs_to_sk(rs));
	}
}
EXPORT_SYMBOL_GPL(rds_atomic_send_complete);

A
Andy Grover 已提交
521 522 523 524 525 526
/*
 * This is the same as rds_rdma_send_complete except we
 * don't do any locking - we have all the ingredients (message,
 * socket, socket lock) and can just move the notifier.
 */
static inline void
527
__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
A
Andy Grover 已提交
528
{
A
Andy Grover 已提交
529
	struct rm_rdma_op *ro;
530
	struct rm_atomic_op *ao;
A
Andy Grover 已提交
531

A
Andy Grover 已提交
532 533 534 535 536
	ro = &rm->rdma;
	if (ro->op_active && ro->op_notify && ro->op_notifier) {
		ro->op_notifier->n_status = status;
		list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
		ro->op_notifier = NULL;
A
Andy Grover 已提交
537 538
	}

539 540 541 542 543 544 545
	ao = &rm->atomic;
	if (ao->op_active && ao->op_notify && ao->op_notifier) {
		ao->op_notifier->n_status = status;
		list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
		ao->op_notifier = NULL;
	}

A
Andy Grover 已提交
546 547 548 549 550 551 552 553 554
	/* No need to wake the app - caller does this */
}

/*
 * This is called from the IB send completion when we detect
 * a RDMA operation that failed with remote access error.
 * So speed is not an issue here.
 */
struct rds_message *rds_send_get_message(struct rds_connection *conn,
A
Andy Grover 已提交
555
					 struct rm_rdma_op *op)
A
Andy Grover 已提交
556 557 558 559 560 561 562
{
	struct rds_message *rm, *tmp, *found = NULL;
	unsigned long flags;

	spin_lock_irqsave(&conn->c_lock, flags);

	list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
A
Andy Grover 已提交
563
		if (&rm->rdma == op) {
A
Andy Grover 已提交
564 565 566 567 568 569 570
			atomic_inc(&rm->m_refcount);
			found = rm;
			goto out;
		}
	}

	list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
A
Andy Grover 已提交
571
		if (&rm->rdma == op) {
A
Andy Grover 已提交
572 573 574 575 576 577 578 579 580 581 582
			atomic_inc(&rm->m_refcount);
			found = rm;
			break;
		}
	}

out:
	spin_unlock_irqrestore(&conn->c_lock, flags);

	return found;
}
A
Andy Grover 已提交
583
EXPORT_SYMBOL_GPL(rds_send_get_message);
A
Andy Grover 已提交
584 585 586 587 588 589 590 591 592 593 594

/*
 * This removes messages from the socket's list if they're on it.  The list
 * argument must be private to the caller, we must be able to modify it
 * without locks.  The messages must have a reference held for their
 * position on the list.  This function will drop that reference after
 * removing the messages from the 'messages' list regardless of if it found
 * the messages on the socket list or not.
 */
void rds_send_remove_from_sock(struct list_head *messages, int status)
{
595
	unsigned long flags;
A
Andy Grover 已提交
596 597 598 599
	struct rds_sock *rs = NULL;
	struct rds_message *rm;

	while (!list_empty(messages)) {
600 601
		int was_on_sock = 0;

A
Andy Grover 已提交
602 603 604 605 606 607 608 609 610 611 612 613 614 615
		rm = list_entry(messages->next, struct rds_message,
				m_conn_item);
		list_del_init(&rm->m_conn_item);

		/*
		 * If we see this flag cleared then we're *sure* that someone
		 * else beat us to removing it from the sock.  If we race
		 * with their flag update we'll get the lock and then really
		 * see that the flag has been cleared.
		 *
		 * The message spinlock makes sure nobody clears rm->m_rs
		 * while we're messing with it. It does not prevent the
		 * message from being removed from the socket, though.
		 */
616
		spin_lock_irqsave(&rm->m_rs_lock, flags);
A
Andy Grover 已提交
617 618 619 620 621 622 623 624 625 626 627
		if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
			goto unlock_and_drop;

		if (rs != rm->m_rs) {
			if (rs) {
				rds_wake_sk_sleep(rs);
				sock_put(rds_rs_to_sk(rs));
			}
			rs = rm->m_rs;
			sock_hold(rds_rs_to_sk(rs));
		}
T
Tina Yang 已提交
628
		spin_lock(&rs->rs_lock);
A
Andy Grover 已提交
629 630

		if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
A
Andy Grover 已提交
631
			struct rm_rdma_op *ro = &rm->rdma;
A
Andy Grover 已提交
632 633 634 635 636
			struct rds_notifier *notifier;

			list_del_init(&rm->m_sock_item);
			rds_send_sndbuf_remove(rs, rm);

A
Andy Grover 已提交
637 638 639
			if (ro->op_active && ro->op_notifier &&
			       (ro->op_notify || (ro->op_recverr && status))) {
				notifier = ro->op_notifier;
A
Andy Grover 已提交
640 641 642 643
				list_add_tail(&notifier->n_list,
						&rs->rs_notify_queue);
				if (!notifier->n_status)
					notifier->n_status = status;
A
Andy Grover 已提交
644
				rm->rdma.op_notifier = NULL;
A
Andy Grover 已提交
645
			}
646
			was_on_sock = 1;
A
Andy Grover 已提交
647 648
			rm->m_rs = NULL;
		}
T
Tina Yang 已提交
649
		spin_unlock(&rs->rs_lock);
A
Andy Grover 已提交
650 651

unlock_and_drop:
652
		spin_unlock_irqrestore(&rm->m_rs_lock, flags);
A
Andy Grover 已提交
653
		rds_message_put(rm);
654 655
		if (was_on_sock)
			rds_message_put(rm);
A
Andy Grover 已提交
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
	}

	if (rs) {
		rds_wake_sk_sleep(rs);
		sock_put(rds_rs_to_sk(rs));
	}
}

/*
 * Transports call here when they've determined that the receiver queued
 * messages up to, and including, the given sequence number.  Messages are
 * moved to the retrans queue when rds_send_xmit picks them off the send
 * queue. This means that in the TCP case, the message may not have been
 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
 * checks the RDS_MSG_HAS_ACK_SEQ bit.
 *
 * XXX It's not clear to me how this is safely serialized with socket
 * destruction.  Maybe it should bail if it sees SOCK_DEAD.
 */
void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
			 is_acked_func is_acked)
{
	struct rds_message *rm, *tmp;
	unsigned long flags;
	LIST_HEAD(list);

	spin_lock_irqsave(&conn->c_lock, flags);

	list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
		if (!rds_send_is_acked(rm, ack, is_acked))
			break;

		list_move(&rm->m_conn_item, &list);
		clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
	}

	/* order flag updates with spin locks */
	if (!list_empty(&list))
		smp_mb__after_clear_bit();

	spin_unlock_irqrestore(&conn->c_lock, flags);

	/* now remove the messages from the sock list as needed */
	rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
}
A
Andy Grover 已提交
701
EXPORT_SYMBOL_GPL(rds_send_drop_acked);
A
Andy Grover 已提交
702 703 704 705 706

void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
{
	struct rds_message *rm, *tmp;
	struct rds_connection *conn;
707
	unsigned long flags;
A
Andy Grover 已提交
708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723
	LIST_HEAD(list);

	/* get all the messages we're dropping under the rs lock */
	spin_lock_irqsave(&rs->rs_lock, flags);

	list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
		if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
			     dest->sin_port != rm->m_inc.i_hdr.h_dport))
			continue;

		list_move(&rm->m_sock_item, &list);
		rds_send_sndbuf_remove(rs, rm);
		clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
	}

	/* order flag updates with the rs lock */
724
	smp_mb__after_clear_bit();
A
Andy Grover 已提交
725 726 727

	spin_unlock_irqrestore(&rs->rs_lock, flags);

728 729
	if (list_empty(&list))
		return;
A
Andy Grover 已提交
730

731
	/* Remove the messages from the conn */
A
Andy Grover 已提交
732
	list_for_each_entry(rm, &list, m_sock_item) {
733 734

		conn = rm->m_inc.i_conn;
A
Andy Grover 已提交
735

736
		spin_lock_irqsave(&conn->c_lock, flags);
A
Andy Grover 已提交
737
		/*
738 739 740
		 * Maybe someone else beat us to removing rm from the conn.
		 * If we race with their flag update we'll get the lock and
		 * then really see that the flag has been cleared.
A
Andy Grover 已提交
741
		 */
742 743
		if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
			spin_unlock_irqrestore(&conn->c_lock, flags);
A
Andy Grover 已提交
744 745
			continue;
		}
746 747
		list_del_init(&rm->m_conn_item);
		spin_unlock_irqrestore(&conn->c_lock, flags);
A
Andy Grover 已提交
748

749 750 751 752
		/*
		 * Couldn't grab m_rs_lock in top loop (lock ordering),
		 * but we can now.
		 */
753
		spin_lock_irqsave(&rm->m_rs_lock, flags);
A
Andy Grover 已提交
754

755
		spin_lock(&rs->rs_lock);
756
		__rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
757 758 759
		spin_unlock(&rs->rs_lock);

		rm->m_rs = NULL;
760
		spin_unlock_irqrestore(&rm->m_rs_lock, flags);
761 762 763

		rds_message_put(rm);
	}
A
Andy Grover 已提交
764

765
	rds_wake_sk_sleep(rs);
766

A
Andy Grover 已提交
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
	while (!list_empty(&list)) {
		rm = list_entry(list.next, struct rds_message, m_sock_item);
		list_del_init(&rm->m_sock_item);

		rds_message_wait(rm);
		rds_message_put(rm);
	}
}

/*
 * we only want this to fire once so we use the callers 'queued'.  It's
 * possible that another thread can race with us and remove the
 * message from the flow with RDS_CANCEL_SENT_TO.
 */
static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
			     struct rds_message *rm, __be16 sport,
			     __be16 dport, int *queued)
{
	unsigned long flags;
	u32 len;

	if (*queued)
		goto out;

	len = be32_to_cpu(rm->m_inc.i_hdr.h_len);

	/* this is the only place which holds both the socket's rs_lock
	 * and the connection's c_lock */
	spin_lock_irqsave(&rs->rs_lock, flags);

	/*
	 * If there is a little space in sndbuf, we don't queue anything,
	 * and userspace gets -EAGAIN. But poll() indicates there's send
	 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
	 * freed up by incoming acks. So we check the *old* value of
	 * rs_snd_bytes here to allow the last msg to exceed the buffer,
	 * and poll() now knows no more data can be sent.
	 */
	if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
		rs->rs_snd_bytes += len;

		/* let recv side know we are close to send space exhaustion.
		 * This is probably not the optimal way to do it, as this
		 * means we set the flag on *all* messages as soon as our
		 * throughput hits a certain threshold.
		 */
		if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
			__set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);

		list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
		set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
		rds_message_addref(rm);
		rm->m_rs = rs;

		/* The code ordering is a little weird, but we're
		   trying to minimize the time we hold c_lock */
		rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
		rm->m_inc.i_conn = conn;
		rds_message_addref(rm);

		spin_lock(&conn->c_lock);
		rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
		list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
		set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
		spin_unlock(&conn->c_lock);

		rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
			 rm, len, rs, rs->rs_snd_bytes,
			 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));

		*queued = 1;
	}

	spin_unlock_irqrestore(&rs->rs_lock, flags);
out:
	return *queued;
}

845 846 847 848 849 850
/*
 * rds_message is getting to be quite complicated, and we'd like to allocate
 * it all in one go. This figures out how big it needs to be up front.
 */
static int rds_rm_size(struct msghdr *msg, int data_len)
{
851
	struct cmsghdr *cmsg;
852
	int size = 0;
853
	int cmsg_groups = 0;
854 855 856 857 858 859 860 861 862 863 864
	int retval;

	for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
		if (!CMSG_OK(msg, cmsg))
			return -EINVAL;

		if (cmsg->cmsg_level != SOL_RDS)
			continue;

		switch (cmsg->cmsg_type) {
		case RDS_CMSG_RDMA_ARGS:
865
			cmsg_groups |= 1;
866 867 868 869
			retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
			if (retval < 0)
				return retval;
			size += retval;
870

871 872 873 874
			break;

		case RDS_CMSG_RDMA_DEST:
		case RDS_CMSG_RDMA_MAP:
875
			cmsg_groups |= 2;
876 877 878
			/* these are valid but do no add any size */
			break;

A
Andy Grover 已提交
879 880
		case RDS_CMSG_ATOMIC_CSWP:
		case RDS_CMSG_ATOMIC_FADD:
881
			cmsg_groups |= 1;
A
Andy Grover 已提交
882 883 884
			size += sizeof(struct scatterlist);
			break;

885 886 887 888 889
		default:
			return -EINVAL;
		}

	}
890

891
	size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
892

893 894 895 896
	/* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
	if (cmsg_groups == 3)
		return -EINVAL;

897 898 899
	return size;
}

A
Andy Grover 已提交
900 901 902 903 904 905 906 907 908 909 910 911 912 913
static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
			 struct msghdr *msg, int *allocated_mr)
{
	struct cmsghdr *cmsg;
	int ret = 0;

	for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
		if (!CMSG_OK(msg, cmsg))
			return -EINVAL;

		if (cmsg->cmsg_level != SOL_RDS)
			continue;

		/* As a side effect, RDMA_DEST and RDMA_MAP will set
A
Andy Grover 已提交
914
		 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
A
Andy Grover 已提交
915 916 917 918 919 920 921 922 923 924 925 926 927 928 929
		 */
		switch (cmsg->cmsg_type) {
		case RDS_CMSG_RDMA_ARGS:
			ret = rds_cmsg_rdma_args(rs, rm, cmsg);
			break;

		case RDS_CMSG_RDMA_DEST:
			ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
			break;

		case RDS_CMSG_RDMA_MAP:
			ret = rds_cmsg_rdma_map(rs, rm, cmsg);
			if (!ret)
				*allocated_mr = 1;
			break;
A
Andy Grover 已提交
930 931 932 933
		case RDS_CMSG_ATOMIC_CSWP:
		case RDS_CMSG_ATOMIC_FADD:
			ret = rds_cmsg_atomic(rs, rm, cmsg);
			break;
A
Andy Grover 已提交
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

		default:
			return -EINVAL;
		}

		if (ret)
			break;
	}

	return ret;
}

int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
		size_t payload_len)
{
	struct sock *sk = sock->sk;
	struct rds_sock *rs = rds_sk_to_rs(sk);
	struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
	__be32 daddr;
	__be16 dport;
	struct rds_message *rm = NULL;
	struct rds_connection *conn;
	int ret = 0;
	int queued = 0, allocated_mr = 0;
	int nonblock = msg->msg_flags & MSG_DONTWAIT;
959
	long timeo = sock_sndtimeo(sk, nonblock);
A
Andy Grover 已提交
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

	/* Mirror Linux UDP mirror of BSD error message compatibility */
	/* XXX: Perhaps MSG_MORE someday */
	if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
		printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
		ret = -EOPNOTSUPP;
		goto out;
	}

	if (msg->msg_namelen) {
		/* XXX fail non-unicast destination IPs? */
		if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
			ret = -EINVAL;
			goto out;
		}
		daddr = usin->sin_addr.s_addr;
		dport = usin->sin_port;
	} else {
		/* We only care about consistency with ->connect() */
		lock_sock(sk);
		daddr = rs->rs_conn_addr;
		dport = rs->rs_conn_port;
		release_sock(sk);
	}

	/* racing with another thread binding seems ok here */
	if (daddr == 0 || rs->rs_bound_addr == 0) {
		ret = -ENOTCONN; /* XXX not a great errno */
		goto out;
	}

991 992 993 994 995 996 997 998
	/* size of rm including all sgs */
	ret = rds_rm_size(msg, payload_len);
	if (ret < 0)
		goto out;

	rm = rds_message_alloc(ret, GFP_KERNEL);
	if (!rm) {
		ret = -ENOMEM;
A
Andy Grover 已提交
999 1000 1001
		goto out;
	}

1002 1003 1004 1005 1006 1007 1008 1009
	/* Attach data to the rm */
	if (payload_len) {
		rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
		ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
		if (ret)
			goto out;
	}
	rm->data.op_active = 1;
1010

A
Andy Grover 已提交
1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027
	rm->m_daddr = daddr;

	/* rds_conn_create has a spinlock that runs with IRQ off.
	 * Caching the conn in the socket helps a lot. */
	if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
		conn = rs->rs_conn;
	else {
		conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
					rs->rs_transport,
					sock->sk->sk_allocation);
		if (IS_ERR(conn)) {
			ret = PTR_ERR(conn);
			goto out;
		}
		rs->rs_conn = conn;
	}

1028 1029 1030 1031 1032
	/* Parse any control messages the user may have included. */
	ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
	if (ret)
		goto out;

1033
	if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
A
Andy Grover 已提交
1034 1035
		if (printk_ratelimit())
			printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
A
Andy Grover 已提交
1036
			       &rm->rdma, conn->c_trans->xmit_rdma);
A
Andy Grover 已提交
1037 1038 1039 1040 1041 1042 1043 1044
		ret = -EOPNOTSUPP;
		goto out;
	}

	if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
		if (printk_ratelimit())
			printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
			       &rm->atomic, conn->c_trans->xmit_atomic);
A
Andy Grover 已提交
1045 1046 1047 1048 1049 1050 1051 1052
		ret = -EOPNOTSUPP;
		goto out;
	}

	/* If the connection is down, trigger a connect. We may
	 * have scheduled a delayed reconnect however - in this case
	 * we should not interfere.
	 */
1053 1054
	if (rds_conn_state(conn) == RDS_CONN_DOWN &&
	    !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
A
Andy Grover 已提交
1055 1056 1057
		queue_delayed_work(rds_wq, &conn->c_conn_w, 0);

	ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1058 1059
	if (ret) {
		rs->rs_seen_congestion = 1;
A
Andy Grover 已提交
1060
		goto out;
1061
	}
A
Andy Grover 已提交
1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075

	while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
				  dport, &queued)) {
		rds_stats_inc(s_send_queue_full);
		/* XXX make sure this is reasonable */
		if (payload_len > rds_sk_sndbuf(rs)) {
			ret = -EMSGSIZE;
			goto out;
		}
		if (nonblock) {
			ret = -EAGAIN;
			goto out;
		}

E
Eric Dumazet 已提交
1076
		timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
A
Andy Grover 已提交
1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126
					rds_send_queue_rm(rs, conn, rm,
							  rs->rs_bound_port,
							  dport,
							  &queued),
					timeo);
		rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
		if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
			continue;

		ret = timeo;
		if (ret == 0)
			ret = -ETIMEDOUT;
		goto out;
	}

	/*
	 * By now we've committed to the send.  We reuse rds_send_worker()
	 * to retry sends in the rds thread if the transport asks us to.
	 */
	rds_stats_inc(s_send_queued);

	if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
		rds_send_worker(&conn->c_send_w.work);

	rds_message_put(rm);
	return payload_len;

out:
	/* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
	 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
	 * or in any other way, we need to destroy the MR again */
	if (allocated_mr)
		rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);

	if (rm)
		rds_message_put(rm);
	return ret;
}

/*
 * Reply to a ping packet.
 */
int
rds_send_pong(struct rds_connection *conn, __be16 dport)
{
	struct rds_message *rm;
	unsigned long flags;
	int ret = 0;

	rm = rds_message_alloc(0, GFP_ATOMIC);
1127
	if (!rm) {
A
Andy Grover 已提交
1128 1129 1130 1131 1132 1133 1134 1135 1136 1137
		ret = -ENOMEM;
		goto out;
	}

	rm->m_daddr = conn->c_faddr;

	/* If the connection is down, trigger a connect. We may
	 * have scheduled a delayed reconnect however - in this case
	 * we should not interfere.
	 */
1138 1139
	if (rds_conn_state(conn) == RDS_CONN_DOWN &&
	    !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
A
Andy Grover 已提交
1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168
		queue_delayed_work(rds_wq, &conn->c_conn_w, 0);

	ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
	if (ret)
		goto out;

	spin_lock_irqsave(&conn->c_lock, flags);
	list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
	set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
	rds_message_addref(rm);
	rm->m_inc.i_conn = conn;

	rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
				    conn->c_next_tx_seq);
	conn->c_next_tx_seq++;
	spin_unlock_irqrestore(&conn->c_lock, flags);

	rds_stats_inc(s_send_queued);
	rds_stats_inc(s_send_pong);

	queue_delayed_work(rds_wq, &conn->c_send_w, 0);
	rds_message_put(rm);
	return 0;

out:
	if (rm)
		rds_message_put(rm);
	return ret;
}