2 * Copyright (c) 2007 Oracle. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/types.h>
34 #include <linux/rbtree.h>
39 * This file implements the receive side of the unconventional congestion
42 * Messages waiting in the receive queue on the receiving socket are accounted
43 * against the sockets SO_RCVBUF option value. Only the payload bytes in the
44 * message are accounted for. If the number of bytes queued equals or exceeds
45 * rcvbuf then the socket is congested. All sends attempted to this socket's
46 * address should return block or return -EWOULDBLOCK.
48 * Applications are expected to be reasonably tuned such that this situation
49 * very rarely occurs. An application encountering this "back-pressure" is
52 * This is implemented by having each node maintain bitmaps which indicate
53 * which ports on bound addresses are congested. As the bitmap changes it is
54 * sent through all the connections which terminate in the local address of the
55 * bitmap which changed.
57 * The bitmaps are allocated as connections are brought up. This avoids
58 * allocation in the interrupt handling path which queues messages on sockets.
59 * The dense bitmaps let transports send the entire bitmap on any bitmap change
60 * reasonably efficiently. This is much easier to implement than some
61 * finer-grained communication of per-port congestion. The sender does a very
62 * inexpensive bit test to test if the port it's about to send to is congested
67 * Interaction with poll is a tad tricky. We want all processes stuck in
68 * poll to wake up and check whether a congested destination became uncongested.
69 * The really sad thing is we have no idea which destinations the application
70 * wants to send to - we don't even know which rds_connections are involved.
71 * So until we implement a more flexible rds poll interface, we have to make
73 * We maintain a global counter that is incremented each time a congestion map
74 * update is received. Each rds socket tracks this value, and if rds_poll
75 * finds that the saved generation number is smaller than the global generation
76 * number, it wakes up the process.
78 static atomic_t rds_cong_generation = ATOMIC_INIT(0);
81 * Congestion monitoring
83 static LIST_HEAD(rds_cong_monitor);
84 static DEFINE_RWLOCK(rds_cong_monitor_lock);
87 * Yes, a global lock. It's used so infrequently that it's worth keeping it
88 * global to simplify the locking. It's only used in the following
91 * - on connection buildup to associate a conn with its maps
92 * - on map changes to inform conns of a new map to send
94 * It's sadly ordered under the socket callback lock and the connection lock.
95 * Receive paths can mark ports congested from interrupt context so the
96 * lock masks interrupts.
98 static DEFINE_SPINLOCK(rds_cong_lock);
99 static struct rb_root rds_cong_tree = RB_ROOT;
101 static struct rds_cong_map *rds_cong_tree_walk(__be32 addr,
102 struct rds_cong_map *insert)
104 struct rb_node **p = &rds_cong_tree.rb_node;
105 struct rb_node *parent = NULL;
106 struct rds_cong_map *map;
110 map = rb_entry(parent, struct rds_cong_map, m_rb_node);
112 if (addr < map->m_addr)
114 else if (addr > map->m_addr)
121 rb_link_node(&insert->m_rb_node, parent, p);
122 rb_insert_color(&insert->m_rb_node, &rds_cong_tree);
128 * There is only ever one bitmap for any address. Connections try and allocate
129 * these bitmaps in the process getting pointers to them. The bitmaps are only
130 * ever freed as the module is removed after all connections have been freed.
132 static struct rds_cong_map *rds_cong_from_addr(__be32 addr)
134 struct rds_cong_map *map;
135 struct rds_cong_map *ret = NULL;
140 map = kzalloc(sizeof(struct rds_cong_map), GFP_KERNEL);
145 init_waitqueue_head(&map->m_waitq);
146 INIT_LIST_HEAD(&map->m_conn_list);
148 for (i = 0; i < RDS_CONG_MAP_PAGES; i++) {
149 zp = get_zeroed_page(GFP_KERNEL);
152 map->m_page_addrs[i] = zp;
155 spin_lock_irqsave(&rds_cong_lock, flags);
156 ret = rds_cong_tree_walk(addr, map);
157 spin_unlock_irqrestore(&rds_cong_lock, flags);
166 for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
167 free_page(map->m_page_addrs[i]);
171 rdsdebug("map %p for addr %x\n", ret, be32_to_cpu(addr));
177 * Put the conn on its local map's list. This is called when the conn is
178 * really added to the hash. It's nested under the rds_conn_lock, sadly.
180 void rds_cong_add_conn(struct rds_connection *conn)
184 rdsdebug("conn %p now on map %p\n", conn, conn->c_lcong);
185 spin_lock_irqsave(&rds_cong_lock, flags);
186 list_add_tail(&conn->c_map_item, &conn->c_lcong->m_conn_list);
187 spin_unlock_irqrestore(&rds_cong_lock, flags);
190 void rds_cong_remove_conn(struct rds_connection *conn)
194 rdsdebug("removing conn %p from map %p\n", conn, conn->c_lcong);
195 spin_lock_irqsave(&rds_cong_lock, flags);
196 list_del_init(&conn->c_map_item);
197 spin_unlock_irqrestore(&rds_cong_lock, flags);
200 int rds_cong_get_maps(struct rds_connection *conn)
202 conn->c_lcong = rds_cong_from_addr(conn->c_laddr);
203 conn->c_fcong = rds_cong_from_addr(conn->c_faddr);
205 if (conn->c_lcong == NULL || conn->c_fcong == NULL)
211 void rds_cong_queue_updates(struct rds_cong_map *map)
213 struct rds_connection *conn;
216 spin_lock_irqsave(&rds_cong_lock, flags);
218 list_for_each_entry(conn, &map->m_conn_list, c_map_item) {
219 if (!test_and_set_bit(0, &conn->c_map_queued)) {
220 rds_stats_inc(s_cong_update_queued);
221 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
225 spin_unlock_irqrestore(&rds_cong_lock, flags);
228 void rds_cong_map_updated(struct rds_cong_map *map, uint64_t portmask)
230 rdsdebug("waking map %p for %pI4\n",
232 rds_stats_inc(s_cong_update_received);
233 atomic_inc(&rds_cong_generation);
234 if (waitqueue_active(&map->m_waitq))
235 wake_up(&map->m_waitq);
236 if (waitqueue_active(&rds_poll_waitq))
237 wake_up_all(&rds_poll_waitq);
239 if (portmask && !list_empty(&rds_cong_monitor)) {
243 read_lock_irqsave(&rds_cong_monitor_lock, flags);
244 list_for_each_entry(rs, &rds_cong_monitor, rs_cong_list) {
245 spin_lock(&rs->rs_lock);
246 rs->rs_cong_notify |= (rs->rs_cong_mask & portmask);
247 rs->rs_cong_mask &= ~portmask;
248 spin_unlock(&rs->rs_lock);
249 if (rs->rs_cong_notify)
250 rds_wake_sk_sleep(rs);
252 read_unlock_irqrestore(&rds_cong_monitor_lock, flags);
256 int rds_cong_updated_since(unsigned long *recent)
258 unsigned long gen = atomic_read(&rds_cong_generation);
260 if (likely(*recent == gen))
267 * We're called under the locking that protects the sockets receive buffer
268 * consumption. This makes it a lot easier for the caller to only call us
269 * when it knows that an existing set bit needs to be cleared, and vice versa.
270 * We can't block and we need to deal with concurrent sockets working against
271 * the same per-address map.
273 void rds_cong_set_bit(struct rds_cong_map *map, __be16 port)
278 rdsdebug("setting congestion for %pI4:%u in map %p\n",
279 &map->m_addr, ntohs(port), map);
281 i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
282 off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
284 generic___set_le_bit(off, (void *)map->m_page_addrs[i]);
287 void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port)
292 rdsdebug("clearing congestion for %pI4:%u in map %p\n",
293 &map->m_addr, ntohs(port), map);
295 i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
296 off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
298 generic___clear_le_bit(off, (void *)map->m_page_addrs[i]);
301 static int rds_cong_test_bit(struct rds_cong_map *map, __be16 port)
306 i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
307 off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
309 return generic_test_le_bit(off, (void *)map->m_page_addrs[i]);
312 void rds_cong_add_socket(struct rds_sock *rs)
316 write_lock_irqsave(&rds_cong_monitor_lock, flags);
317 if (list_empty(&rs->rs_cong_list))
318 list_add(&rs->rs_cong_list, &rds_cong_monitor);
319 write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
322 void rds_cong_remove_socket(struct rds_sock *rs)
325 struct rds_cong_map *map;
327 write_lock_irqsave(&rds_cong_monitor_lock, flags);
328 list_del_init(&rs->rs_cong_list);
329 write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
331 /* update congestion map for now-closed port */
332 spin_lock_irqsave(&rds_cong_lock, flags);
333 map = rds_cong_tree_walk(rs->rs_bound_addr, NULL);
334 spin_unlock_irqrestore(&rds_cong_lock, flags);
336 if (map && rds_cong_test_bit(map, rs->rs_bound_port)) {
337 rds_cong_clear_bit(map, rs->rs_bound_port);
338 rds_cong_queue_updates(map);
342 int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock,
345 if (!rds_cong_test_bit(map, port))
348 if (rs && rs->rs_cong_monitor) {
351 /* It would have been nice to have an atomic set_bit on
353 spin_lock_irqsave(&rs->rs_lock, flags);
354 rs->rs_cong_mask |= RDS_CONG_MONITOR_MASK(ntohs(port));
355 spin_unlock_irqrestore(&rs->rs_lock, flags);
357 /* Test again - a congestion update may have arrived in
359 if (!rds_cong_test_bit(map, port))
362 rds_stats_inc(s_cong_send_error);
366 rds_stats_inc(s_cong_send_blocked);
367 rdsdebug("waiting on map %p for port %u\n", map, be16_to_cpu(port));
369 return wait_event_interruptible(map->m_waitq,
370 !rds_cong_test_bit(map, port));
373 void rds_cong_exit(void)
375 struct rb_node *node;
376 struct rds_cong_map *map;
379 while ((node = rb_first(&rds_cong_tree))) {
380 map = rb_entry(node, struct rds_cong_map, m_rb_node);
381 rdsdebug("freeing map %p\n", map);
382 rb_erase(&map->m_rb_node, &rds_cong_tree);
383 for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
384 free_page(map->m_page_addrs[i]);
390 * Allocate a RDS message containing a congestion update.
392 struct rds_message *rds_cong_update_alloc(struct rds_connection *conn)
394 struct rds_cong_map *map = conn->c_lcong;
395 struct rds_message *rm;
397 rm = rds_message_map_pages(map->m_page_addrs, RDS_CONG_MAP_BYTES);
399 rm->m_inc.i_hdr.h_flags = RDS_FLAG_CONG_BITMAP;