Root/net/rds/cong.c

1/*
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/slab.h>
34#include <linux/types.h>
35#include <linux/rbtree.h>
36
37#include <asm-generic/bitops/le.h>
38
39#include "rds.h"
40
41/*
42 * This file implements the receive side of the unconventional congestion
43 * management in RDS.
44 *
45 * Messages waiting in the receive queue on the receiving socket are accounted
46 * against the sockets SO_RCVBUF option value. Only the payload bytes in the
47 * message are accounted for. If the number of bytes queued equals or exceeds
48 * rcvbuf then the socket is congested. All sends attempted to this socket's
49 * address should return block or return -EWOULDBLOCK.
50 *
51 * Applications are expected to be reasonably tuned such that this situation
52 * very rarely occurs. An application encountering this "back-pressure" is
53 * considered a bug.
54 *
55 * This is implemented by having each node maintain bitmaps which indicate
56 * which ports on bound addresses are congested. As the bitmap changes it is
57 * sent through all the connections which terminate in the local address of the
58 * bitmap which changed.
59 *
60 * The bitmaps are allocated as connections are brought up. This avoids
61 * allocation in the interrupt handling path which queues messages on sockets.
62 * The dense bitmaps let transports send the entire bitmap on any bitmap change
63 * reasonably efficiently. This is much easier to implement than some
64 * finer-grained communication of per-port congestion. The sender does a very
65 * inexpensive bit test to test if the port it's about to send to is congested
66 * or not.
67 */
68
69/*
70 * Interaction with poll is a tad tricky. We want all processes stuck in
71 * poll to wake up and check whether a congested destination became uncongested.
72 * The really sad thing is we have no idea which destinations the application
73 * wants to send to - we don't even know which rds_connections are involved.
74 * So until we implement a more flexible rds poll interface, we have to make
75 * do with this:
76 * We maintain a global counter that is incremented each time a congestion map
77 * update is received. Each rds socket tracks this value, and if rds_poll
78 * finds that the saved generation number is smaller than the global generation
79 * number, it wakes up the process.
80 */
81static atomic_t rds_cong_generation = ATOMIC_INIT(0);
82
83/*
84 * Congestion monitoring
85 */
86static LIST_HEAD(rds_cong_monitor);
87static DEFINE_RWLOCK(rds_cong_monitor_lock);
88
89/*
90 * Yes, a global lock. It's used so infrequently that it's worth keeping it
91 * global to simplify the locking. It's only used in the following
92 * circumstances:
93 *
94 * - on connection buildup to associate a conn with its maps
95 * - on map changes to inform conns of a new map to send
96 *
97 * It's sadly ordered under the socket callback lock and the connection lock.
98 * Receive paths can mark ports congested from interrupt context so the
99 * lock masks interrupts.
100 */
101static DEFINE_SPINLOCK(rds_cong_lock);
102static struct rb_root rds_cong_tree = RB_ROOT;
103
104static struct rds_cong_map *rds_cong_tree_walk(__be32 addr,
105                           struct rds_cong_map *insert)
106{
107    struct rb_node **p = &rds_cong_tree.rb_node;
108    struct rb_node *parent = NULL;
109    struct rds_cong_map *map;
110
111    while (*p) {
112        parent = *p;
113        map = rb_entry(parent, struct rds_cong_map, m_rb_node);
114
115        if (addr < map->m_addr)
116            p = &(*p)->rb_left;
117        else if (addr > map->m_addr)
118            p = &(*p)->rb_right;
119        else
120            return map;
121    }
122
123    if (insert) {
124        rb_link_node(&insert->m_rb_node, parent, p);
125        rb_insert_color(&insert->m_rb_node, &rds_cong_tree);
126    }
127    return NULL;
128}
129
130/*
131 * There is only ever one bitmap for any address. Connections try and allocate
132 * these bitmaps in the process getting pointers to them. The bitmaps are only
133 * ever freed as the module is removed after all connections have been freed.
134 */
135static struct rds_cong_map *rds_cong_from_addr(__be32 addr)
136{
137    struct rds_cong_map *map;
138    struct rds_cong_map *ret = NULL;
139    unsigned long zp;
140    unsigned long i;
141    unsigned long flags;
142
143    map = kzalloc(sizeof(struct rds_cong_map), GFP_KERNEL);
144    if (map == NULL)
145        return NULL;
146
147    map->m_addr = addr;
148    init_waitqueue_head(&map->m_waitq);
149    INIT_LIST_HEAD(&map->m_conn_list);
150
151    for (i = 0; i < RDS_CONG_MAP_PAGES; i++) {
152        zp = get_zeroed_page(GFP_KERNEL);
153        if (zp == 0)
154            goto out;
155        map->m_page_addrs[i] = zp;
156    }
157
158    spin_lock_irqsave(&rds_cong_lock, flags);
159    ret = rds_cong_tree_walk(addr, map);
160    spin_unlock_irqrestore(&rds_cong_lock, flags);
161
162    if (ret == NULL) {
163        ret = map;
164        map = NULL;
165    }
166
167out:
168    if (map) {
169        for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
170            free_page(map->m_page_addrs[i]);
171        kfree(map);
172    }
173
174    rdsdebug("map %p for addr %x\n", ret, be32_to_cpu(addr));
175
176    return ret;
177}
178
179/*
180 * Put the conn on its local map's list. This is called when the conn is
181 * really added to the hash. It's nested under the rds_conn_lock, sadly.
182 */
183void rds_cong_add_conn(struct rds_connection *conn)
184{
185    unsigned long flags;
186
187    rdsdebug("conn %p now on map %p\n", conn, conn->c_lcong);
188    spin_lock_irqsave(&rds_cong_lock, flags);
189    list_add_tail(&conn->c_map_item, &conn->c_lcong->m_conn_list);
190    spin_unlock_irqrestore(&rds_cong_lock, flags);
191}
192
193void rds_cong_remove_conn(struct rds_connection *conn)
194{
195    unsigned long flags;
196
197    rdsdebug("removing conn %p from map %p\n", conn, conn->c_lcong);
198    spin_lock_irqsave(&rds_cong_lock, flags);
199    list_del_init(&conn->c_map_item);
200    spin_unlock_irqrestore(&rds_cong_lock, flags);
201}
202
203int rds_cong_get_maps(struct rds_connection *conn)
204{
205    conn->c_lcong = rds_cong_from_addr(conn->c_laddr);
206    conn->c_fcong = rds_cong_from_addr(conn->c_faddr);
207
208    if (conn->c_lcong == NULL || conn->c_fcong == NULL)
209        return -ENOMEM;
210
211    return 0;
212}
213
214void rds_cong_queue_updates(struct rds_cong_map *map)
215{
216    struct rds_connection *conn;
217    unsigned long flags;
218
219    spin_lock_irqsave(&rds_cong_lock, flags);
220
221    list_for_each_entry(conn, &map->m_conn_list, c_map_item) {
222        if (conn->c_loopback)
223            continue;
224        if (!test_and_set_bit(0, &conn->c_map_queued)) {
225            rds_stats_inc(s_cong_update_queued);
226            queue_delayed_work(rds_wq, &conn->c_send_w, 0);
227        }
228    }
229
230    spin_unlock_irqrestore(&rds_cong_lock, flags);
231}
232
233void rds_cong_map_updated(struct rds_cong_map *map, uint64_t portmask)
234{
235    rdsdebug("waking map %p for %pI4\n",
236      map, &map->m_addr);
237    rds_stats_inc(s_cong_update_received);
238    atomic_inc(&rds_cong_generation);
239    if (waitqueue_active(&map->m_waitq))
240        wake_up(&map->m_waitq);
241    if (waitqueue_active(&rds_poll_waitq))
242        wake_up_all(&rds_poll_waitq);
243
244    if (portmask && !list_empty(&rds_cong_monitor)) {
245        unsigned long flags;
246        struct rds_sock *rs;
247
248        read_lock_irqsave(&rds_cong_monitor_lock, flags);
249        list_for_each_entry(rs, &rds_cong_monitor, rs_cong_list) {
250            spin_lock(&rs->rs_lock);
251            rs->rs_cong_notify |= (rs->rs_cong_mask & portmask);
252            rs->rs_cong_mask &= ~portmask;
253            spin_unlock(&rs->rs_lock);
254            if (rs->rs_cong_notify)
255                rds_wake_sk_sleep(rs);
256        }
257        read_unlock_irqrestore(&rds_cong_monitor_lock, flags);
258    }
259}
260EXPORT_SYMBOL_GPL(rds_cong_map_updated);
261
262int rds_cong_updated_since(unsigned long *recent)
263{
264    unsigned long gen = atomic_read(&rds_cong_generation);
265
266    if (likely(*recent == gen))
267        return 0;
268    *recent = gen;
269    return 1;
270}
271
272/*
273 * We're called under the locking that protects the sockets receive buffer
274 * consumption. This makes it a lot easier for the caller to only call us
275 * when it knows that an existing set bit needs to be cleared, and vice versa.
276 * We can't block and we need to deal with concurrent sockets working against
277 * the same per-address map.
278 */
279void rds_cong_set_bit(struct rds_cong_map *map, __be16 port)
280{
281    unsigned long i;
282    unsigned long off;
283
284    rdsdebug("setting congestion for %pI4:%u in map %p\n",
285      &map->m_addr, ntohs(port), map);
286
287    i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
288    off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
289
290    generic___set_le_bit(off, (void *)map->m_page_addrs[i]);
291}
292
293void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port)
294{
295    unsigned long i;
296    unsigned long off;
297
298    rdsdebug("clearing congestion for %pI4:%u in map %p\n",
299      &map->m_addr, ntohs(port), map);
300
301    i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
302    off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
303
304    generic___clear_le_bit(off, (void *)map->m_page_addrs[i]);
305}
306
307static int rds_cong_test_bit(struct rds_cong_map *map, __be16 port)
308{
309    unsigned long i;
310    unsigned long off;
311
312    i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
313    off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
314
315    return generic_test_le_bit(off, (void *)map->m_page_addrs[i]);
316}
317
318void rds_cong_add_socket(struct rds_sock *rs)
319{
320    unsigned long flags;
321
322    write_lock_irqsave(&rds_cong_monitor_lock, flags);
323    if (list_empty(&rs->rs_cong_list))
324        list_add(&rs->rs_cong_list, &rds_cong_monitor);
325    write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
326}
327
328void rds_cong_remove_socket(struct rds_sock *rs)
329{
330    unsigned long flags;
331    struct rds_cong_map *map;
332
333    write_lock_irqsave(&rds_cong_monitor_lock, flags);
334    list_del_init(&rs->rs_cong_list);
335    write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
336
337    /* update congestion map for now-closed port */
338    spin_lock_irqsave(&rds_cong_lock, flags);
339    map = rds_cong_tree_walk(rs->rs_bound_addr, NULL);
340    spin_unlock_irqrestore(&rds_cong_lock, flags);
341
342    if (map && rds_cong_test_bit(map, rs->rs_bound_port)) {
343        rds_cong_clear_bit(map, rs->rs_bound_port);
344        rds_cong_queue_updates(map);
345    }
346}
347
348int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock,
349          struct rds_sock *rs)
350{
351    if (!rds_cong_test_bit(map, port))
352        return 0;
353    if (nonblock) {
354        if (rs && rs->rs_cong_monitor) {
355            unsigned long flags;
356
357            /* It would have been nice to have an atomic set_bit on
358             * a uint64_t. */
359            spin_lock_irqsave(&rs->rs_lock, flags);
360            rs->rs_cong_mask |= RDS_CONG_MONITOR_MASK(ntohs(port));
361            spin_unlock_irqrestore(&rs->rs_lock, flags);
362
363            /* Test again - a congestion update may have arrived in
364             * the meantime. */
365            if (!rds_cong_test_bit(map, port))
366                return 0;
367        }
368        rds_stats_inc(s_cong_send_error);
369        return -ENOBUFS;
370    }
371
372    rds_stats_inc(s_cong_send_blocked);
373    rdsdebug("waiting on map %p for port %u\n", map, be16_to_cpu(port));
374
375    return wait_event_interruptible(map->m_waitq,
376                    !rds_cong_test_bit(map, port));
377}
378
379void rds_cong_exit(void)
380{
381    struct rb_node *node;
382    struct rds_cong_map *map;
383    unsigned long i;
384
385    while ((node = rb_first(&rds_cong_tree))) {
386        map = rb_entry(node, struct rds_cong_map, m_rb_node);
387        rdsdebug("freeing map %p\n", map);
388        rb_erase(&map->m_rb_node, &rds_cong_tree);
389        for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
390            free_page(map->m_page_addrs[i]);
391        kfree(map);
392    }
393}
394
395/*
396 * Allocate a RDS message containing a congestion update.
397 */
398struct rds_message *rds_cong_update_alloc(struct rds_connection *conn)
399{
400    struct rds_cong_map *map = conn->c_lcong;
401    struct rds_message *rm;
402
403    rm = rds_message_map_pages(map->m_page_addrs, RDS_CONG_MAP_BYTES);
404    if (!IS_ERR(rm))
405        rm->m_inc.i_hdr.h_flags = RDS_FLAG_CONG_BITMAP;
406
407    return rm;
408}
409

Archive Download this file



interactive