blob: 710e4599d76cffa7a56ad2891c127960b37cb3c0 [file] [log] [blame]
Andy Grover922cb172009-02-24 15:30:20 +00001/*
2 * Copyright (c) 2007 Oracle. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/types.h>
34#include <linux/rbtree.h>
35
David S. Millerc3b32402009-03-02 01:49:28 -080036#include <asm-generic/bitops/le.h>
37
Andy Grover922cb172009-02-24 15:30:20 +000038#include "rds.h"
39
40/*
41 * This file implements the receive side of the unconventional congestion
42 * management in RDS.
43 *
44 * Messages waiting in the receive queue on the receiving socket are accounted
45 * against the sockets SO_RCVBUF option value. Only the payload bytes in the
46 * message are accounted for. If the number of bytes queued equals or exceeds
47 * rcvbuf then the socket is congested. All sends attempted to this socket's
48 * address should return block or return -EWOULDBLOCK.
49 *
50 * Applications are expected to be reasonably tuned such that this situation
51 * very rarely occurs. An application encountering this "back-pressure" is
52 * considered a bug.
53 *
54 * This is implemented by having each node maintain bitmaps which indicate
55 * which ports on bound addresses are congested. As the bitmap changes it is
56 * sent through all the connections which terminate in the local address of the
57 * bitmap which changed.
58 *
59 * The bitmaps are allocated as connections are brought up. This avoids
60 * allocation in the interrupt handling path which queues messages on sockets.
61 * The dense bitmaps let transports send the entire bitmap on any bitmap change
62 * reasonably efficiently. This is much easier to implement than some
63 * finer-grained communication of per-port congestion. The sender does a very
64 * inexpensive bit test to test if the port it's about to send to is congested
65 * or not.
66 */
67
68/*
69 * Interaction with poll is a tad tricky. We want all processes stuck in
70 * poll to wake up and check whether a congested destination became uncongested.
71 * The really sad thing is we have no idea which destinations the application
72 * wants to send to - we don't even know which rds_connections are involved.
73 * So until we implement a more flexible rds poll interface, we have to make
74 * do with this:
75 * We maintain a global counter that is incremented each time a congestion map
76 * update is received. Each rds socket tracks this value, and if rds_poll
77 * finds that the saved generation number is smaller than the global generation
78 * number, it wakes up the process.
79 */
80static atomic_t rds_cong_generation = ATOMIC_INIT(0);
81
82/*
83 * Congestion monitoring
84 */
85static LIST_HEAD(rds_cong_monitor);
86static DEFINE_RWLOCK(rds_cong_monitor_lock);
87
88/*
89 * Yes, a global lock. It's used so infrequently that it's worth keeping it
90 * global to simplify the locking. It's only used in the following
91 * circumstances:
92 *
93 * - on connection buildup to associate a conn with its maps
94 * - on map changes to inform conns of a new map to send
95 *
96 * It's sadly ordered under the socket callback lock and the connection lock.
97 * Receive paths can mark ports congested from interrupt context so the
98 * lock masks interrupts.
99 */
100static DEFINE_SPINLOCK(rds_cong_lock);
101static struct rb_root rds_cong_tree = RB_ROOT;
102
103static struct rds_cong_map *rds_cong_tree_walk(__be32 addr,
104 struct rds_cong_map *insert)
105{
106 struct rb_node **p = &rds_cong_tree.rb_node;
107 struct rb_node *parent = NULL;
108 struct rds_cong_map *map;
109
110 while (*p) {
111 parent = *p;
112 map = rb_entry(parent, struct rds_cong_map, m_rb_node);
113
114 if (addr < map->m_addr)
115 p = &(*p)->rb_left;
116 else if (addr > map->m_addr)
117 p = &(*p)->rb_right;
118 else
119 return map;
120 }
121
122 if (insert) {
123 rb_link_node(&insert->m_rb_node, parent, p);
124 rb_insert_color(&insert->m_rb_node, &rds_cong_tree);
125 }
126 return NULL;
127}
128
129/*
130 * There is only ever one bitmap for any address. Connections try and allocate
131 * these bitmaps in the process getting pointers to them. The bitmaps are only
132 * ever freed as the module is removed after all connections have been freed.
133 */
134static struct rds_cong_map *rds_cong_from_addr(__be32 addr)
135{
136 struct rds_cong_map *map;
137 struct rds_cong_map *ret = NULL;
138 unsigned long zp;
139 unsigned long i;
140 unsigned long flags;
141
142 map = kzalloc(sizeof(struct rds_cong_map), GFP_KERNEL);
143 if (map == NULL)
144 return NULL;
145
146 map->m_addr = addr;
147 init_waitqueue_head(&map->m_waitq);
148 INIT_LIST_HEAD(&map->m_conn_list);
149
150 for (i = 0; i < RDS_CONG_MAP_PAGES; i++) {
151 zp = get_zeroed_page(GFP_KERNEL);
152 if (zp == 0)
153 goto out;
154 map->m_page_addrs[i] = zp;
155 }
156
157 spin_lock_irqsave(&rds_cong_lock, flags);
158 ret = rds_cong_tree_walk(addr, map);
159 spin_unlock_irqrestore(&rds_cong_lock, flags);
160
161 if (ret == NULL) {
162 ret = map;
163 map = NULL;
164 }
165
166out:
167 if (map) {
168 for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
169 free_page(map->m_page_addrs[i]);
170 kfree(map);
171 }
172
173 rdsdebug("map %p for addr %x\n", ret, be32_to_cpu(addr));
174
175 return ret;
176}
177
178/*
179 * Put the conn on its local map's list. This is called when the conn is
180 * really added to the hash. It's nested under the rds_conn_lock, sadly.
181 */
182void rds_cong_add_conn(struct rds_connection *conn)
183{
184 unsigned long flags;
185
186 rdsdebug("conn %p now on map %p\n", conn, conn->c_lcong);
187 spin_lock_irqsave(&rds_cong_lock, flags);
188 list_add_tail(&conn->c_map_item, &conn->c_lcong->m_conn_list);
189 spin_unlock_irqrestore(&rds_cong_lock, flags);
190}
191
192void rds_cong_remove_conn(struct rds_connection *conn)
193{
194 unsigned long flags;
195
196 rdsdebug("removing conn %p from map %p\n", conn, conn->c_lcong);
197 spin_lock_irqsave(&rds_cong_lock, flags);
198 list_del_init(&conn->c_map_item);
199 spin_unlock_irqrestore(&rds_cong_lock, flags);
200}
201
202int rds_cong_get_maps(struct rds_connection *conn)
203{
204 conn->c_lcong = rds_cong_from_addr(conn->c_laddr);
205 conn->c_fcong = rds_cong_from_addr(conn->c_faddr);
206
207 if (conn->c_lcong == NULL || conn->c_fcong == NULL)
208 return -ENOMEM;
209
210 return 0;
211}
212
213void rds_cong_queue_updates(struct rds_cong_map *map)
214{
215 struct rds_connection *conn;
216 unsigned long flags;
217
218 spin_lock_irqsave(&rds_cong_lock, flags);
219
220 list_for_each_entry(conn, &map->m_conn_list, c_map_item) {
221 if (!test_and_set_bit(0, &conn->c_map_queued)) {
222 rds_stats_inc(s_cong_update_queued);
223 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
224 }
225 }
226
227 spin_unlock_irqrestore(&rds_cong_lock, flags);
228}
229
230void rds_cong_map_updated(struct rds_cong_map *map, uint64_t portmask)
231{
232 rdsdebug("waking map %p for %pI4\n",
233 map, &map->m_addr);
234 rds_stats_inc(s_cong_update_received);
235 atomic_inc(&rds_cong_generation);
236 if (waitqueue_active(&map->m_waitq))
237 wake_up(&map->m_waitq);
238 if (waitqueue_active(&rds_poll_waitq))
239 wake_up_all(&rds_poll_waitq);
240
241 if (portmask && !list_empty(&rds_cong_monitor)) {
242 unsigned long flags;
243 struct rds_sock *rs;
244
245 read_lock_irqsave(&rds_cong_monitor_lock, flags);
246 list_for_each_entry(rs, &rds_cong_monitor, rs_cong_list) {
247 spin_lock(&rs->rs_lock);
248 rs->rs_cong_notify |= (rs->rs_cong_mask & portmask);
249 rs->rs_cong_mask &= ~portmask;
250 spin_unlock(&rs->rs_lock);
251 if (rs->rs_cong_notify)
252 rds_wake_sk_sleep(rs);
253 }
254 read_unlock_irqrestore(&rds_cong_monitor_lock, flags);
255 }
256}
257
258int rds_cong_updated_since(unsigned long *recent)
259{
260 unsigned long gen = atomic_read(&rds_cong_generation);
261
262 if (likely(*recent == gen))
263 return 0;
264 *recent = gen;
265 return 1;
266}
267
268/*
269 * We're called under the locking that protects the sockets receive buffer
270 * consumption. This makes it a lot easier for the caller to only call us
271 * when it knows that an existing set bit needs to be cleared, and vice versa.
272 * We can't block and we need to deal with concurrent sockets working against
273 * the same per-address map.
274 */
275void rds_cong_set_bit(struct rds_cong_map *map, __be16 port)
276{
277 unsigned long i;
278 unsigned long off;
279
280 rdsdebug("setting congestion for %pI4:%u in map %p\n",
281 &map->m_addr, ntohs(port), map);
282
283 i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
284 off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
285
286 generic___set_le_bit(off, (void *)map->m_page_addrs[i]);
287}
288
289void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port)
290{
291 unsigned long i;
292 unsigned long off;
293
294 rdsdebug("clearing congestion for %pI4:%u in map %p\n",
295 &map->m_addr, ntohs(port), map);
296
297 i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
298 off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
299
300 generic___clear_le_bit(off, (void *)map->m_page_addrs[i]);
301}
302
303static int rds_cong_test_bit(struct rds_cong_map *map, __be16 port)
304{
305 unsigned long i;
306 unsigned long off;
307
308 i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
309 off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
310
311 return generic_test_le_bit(off, (void *)map->m_page_addrs[i]);
312}
313
314void rds_cong_add_socket(struct rds_sock *rs)
315{
316 unsigned long flags;
317
318 write_lock_irqsave(&rds_cong_monitor_lock, flags);
319 if (list_empty(&rs->rs_cong_list))
320 list_add(&rs->rs_cong_list, &rds_cong_monitor);
321 write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
322}
323
324void rds_cong_remove_socket(struct rds_sock *rs)
325{
326 unsigned long flags;
327 struct rds_cong_map *map;
328
329 write_lock_irqsave(&rds_cong_monitor_lock, flags);
330 list_del_init(&rs->rs_cong_list);
331 write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
332
333 /* update congestion map for now-closed port */
334 spin_lock_irqsave(&rds_cong_lock, flags);
335 map = rds_cong_tree_walk(rs->rs_bound_addr, NULL);
336 spin_unlock_irqrestore(&rds_cong_lock, flags);
337
338 if (map && rds_cong_test_bit(map, rs->rs_bound_port)) {
339 rds_cong_clear_bit(map, rs->rs_bound_port);
340 rds_cong_queue_updates(map);
341 }
342}
343
344int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock,
345 struct rds_sock *rs)
346{
347 if (!rds_cong_test_bit(map, port))
348 return 0;
349 if (nonblock) {
350 if (rs && rs->rs_cong_monitor) {
351 unsigned long flags;
352
353 /* It would have been nice to have an atomic set_bit on
354 * a uint64_t. */
355 spin_lock_irqsave(&rs->rs_lock, flags);
356 rs->rs_cong_mask |= RDS_CONG_MONITOR_MASK(ntohs(port));
357 spin_unlock_irqrestore(&rs->rs_lock, flags);
358
359 /* Test again - a congestion update may have arrived in
360 * the meantime. */
361 if (!rds_cong_test_bit(map, port))
362 return 0;
363 }
364 rds_stats_inc(s_cong_send_error);
365 return -ENOBUFS;
366 }
367
368 rds_stats_inc(s_cong_send_blocked);
369 rdsdebug("waiting on map %p for port %u\n", map, be16_to_cpu(port));
370
371 return wait_event_interruptible(map->m_waitq,
372 !rds_cong_test_bit(map, port));
373}
374
375void rds_cong_exit(void)
376{
377 struct rb_node *node;
378 struct rds_cong_map *map;
379 unsigned long i;
380
381 while ((node = rb_first(&rds_cong_tree))) {
382 map = rb_entry(node, struct rds_cong_map, m_rb_node);
383 rdsdebug("freeing map %p\n", map);
384 rb_erase(&map->m_rb_node, &rds_cong_tree);
385 for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
386 free_page(map->m_page_addrs[i]);
387 kfree(map);
388 }
389}
390
391/*
392 * Allocate a RDS message containing a congestion update.
393 */
394struct rds_message *rds_cong_update_alloc(struct rds_connection *conn)
395{
396 struct rds_cong_map *map = conn->c_lcong;
397 struct rds_message *rm;
398
399 rm = rds_message_map_pages(map->m_page_addrs, RDS_CONG_MAP_BYTES);
400 if (!IS_ERR(rm))
401 rm->m_inc.i_hdr.h_flags = RDS_FLAG_CONG_BITMAP;
402
403 return rm;
404}