1  /*
2   * Copyright (c) 2007, 2017 Oracle and/or its affiliates. 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  #include <linux/bitops.h>
37  #include <linux/export.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   */
81  static atomic_t		rds_cong_generation = ATOMIC_INIT(0);
82  
83  /*
84   * Congestion monitoring
85   */
86  static LIST_HEAD(rds_cong_monitor);
87  static 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   */
101  static DEFINE_SPINLOCK(rds_cong_lock);
102  static struct rb_root rds_cong_tree = RB_ROOT;
103  
rds_cong_tree_walk(const struct in6_addr * addr,struct rds_cong_map * insert)104  static struct rds_cong_map *rds_cong_tree_walk(const struct in6_addr *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  		int diff;
113  
114  		parent = *p;
115  		map = rb_entry(parent, struct rds_cong_map, m_rb_node);
116  
117  		diff = rds_addr_cmp(addr, &map->m_addr);
118  		if (diff < 0)
119  			p = &(*p)->rb_left;
120  		else if (diff > 0)
121  			p = &(*p)->rb_right;
122  		else
123  			return map;
124  	}
125  
126  	if (insert) {
127  		rb_link_node(&insert->m_rb_node, parent, p);
128  		rb_insert_color(&insert->m_rb_node, &rds_cong_tree);
129  	}
130  	return NULL;
131  }
132  
133  /*
134   * There is only ever one bitmap for any address.  Connections try and allocate
135   * these bitmaps in the process getting pointers to them.  The bitmaps are only
136   * ever freed as the module is removed after all connections have been freed.
137   */
rds_cong_from_addr(const struct in6_addr * addr)138  static struct rds_cong_map *rds_cong_from_addr(const struct in6_addr *addr)
139  {
140  	struct rds_cong_map *map;
141  	struct rds_cong_map *ret = NULL;
142  	unsigned long zp;
143  	unsigned long i;
144  	unsigned long flags;
145  
146  	map = kzalloc(sizeof(struct rds_cong_map), GFP_KERNEL);
147  	if (!map)
148  		return NULL;
149  
150  	map->m_addr = *addr;
151  	init_waitqueue_head(&map->m_waitq);
152  	INIT_LIST_HEAD(&map->m_conn_list);
153  
154  	for (i = 0; i < RDS_CONG_MAP_PAGES; i++) {
155  		zp = get_zeroed_page(GFP_KERNEL);
156  		if (zp == 0)
157  			goto out;
158  		map->m_page_addrs[i] = zp;
159  	}
160  
161  	spin_lock_irqsave(&rds_cong_lock, flags);
162  	ret = rds_cong_tree_walk(addr, map);
163  	spin_unlock_irqrestore(&rds_cong_lock, flags);
164  
165  	if (!ret) {
166  		ret = map;
167  		map = NULL;
168  	}
169  
170  out:
171  	if (map) {
172  		for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
173  			free_page(map->m_page_addrs[i]);
174  		kfree(map);
175  	}
176  
177  	rdsdebug("map %p for addr %pI6c\n", ret, addr);
178  
179  	return ret;
180  }
181  
182  /*
183   * Put the conn on its local map's list.  This is called when the conn is
184   * really added to the hash.  It's nested under the rds_conn_lock, sadly.
185   */
rds_cong_add_conn(struct rds_connection * conn)186  void rds_cong_add_conn(struct rds_connection *conn)
187  {
188  	unsigned long flags;
189  
190  	rdsdebug("conn %p now on map %p\n", conn, conn->c_lcong);
191  	spin_lock_irqsave(&rds_cong_lock, flags);
192  	list_add_tail(&conn->c_map_item, &conn->c_lcong->m_conn_list);
193  	spin_unlock_irqrestore(&rds_cong_lock, flags);
194  }
195  
rds_cong_remove_conn(struct rds_connection * conn)196  void rds_cong_remove_conn(struct rds_connection *conn)
197  {
198  	unsigned long flags;
199  
200  	rdsdebug("removing conn %p from map %p\n", conn, conn->c_lcong);
201  	spin_lock_irqsave(&rds_cong_lock, flags);
202  	list_del_init(&conn->c_map_item);
203  	spin_unlock_irqrestore(&rds_cong_lock, flags);
204  }
205  
rds_cong_get_maps(struct rds_connection * conn)206  int rds_cong_get_maps(struct rds_connection *conn)
207  {
208  	conn->c_lcong = rds_cong_from_addr(&conn->c_laddr);
209  	conn->c_fcong = rds_cong_from_addr(&conn->c_faddr);
210  
211  	if (!(conn->c_lcong && conn->c_fcong))
212  		return -ENOMEM;
213  
214  	return 0;
215  }
216  
rds_cong_queue_updates(struct rds_cong_map * map)217  void rds_cong_queue_updates(struct rds_cong_map *map)
218  {
219  	struct rds_connection *conn;
220  	unsigned long flags;
221  
222  	spin_lock_irqsave(&rds_cong_lock, flags);
223  
224  	list_for_each_entry(conn, &map->m_conn_list, c_map_item) {
225  		struct rds_conn_path *cp = &conn->c_path[0];
226  
227  		rcu_read_lock();
228  		if (!test_and_set_bit(0, &conn->c_map_queued) &&
229  		    !rds_destroy_pending(cp->cp_conn)) {
230  			rds_stats_inc(s_cong_update_queued);
231  			/* We cannot inline the call to rds_send_xmit() here
232  			 * for two reasons (both pertaining to a TCP transport):
233  			 * 1. When we get here from the receive path, we
234  			 *    are already holding the sock_lock (held by
235  			 *    tcp_v4_rcv()). So inlining calls to
236  			 *    tcp_setsockopt and/or tcp_sendmsg will deadlock
237  			 *    when it tries to get the sock_lock())
238  			 * 2. Interrupts are masked so that we can mark the
239  			 *    port congested from both send and recv paths.
240  			 *    (See comment around declaration of rdc_cong_lock).
241  			 *    An attempt to get the sock_lock() here will
242  			 *    therefore trigger warnings.
243  			 * Defer the xmit to rds_send_worker() instead.
244  			 */
245  			queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
246  		}
247  		rcu_read_unlock();
248  	}
249  
250  	spin_unlock_irqrestore(&rds_cong_lock, flags);
251  }
252  
rds_cong_map_updated(struct rds_cong_map * map,uint64_t portmask)253  void rds_cong_map_updated(struct rds_cong_map *map, uint64_t portmask)
254  {
255  	rdsdebug("waking map %p for %pI4\n",
256  	  map, &map->m_addr);
257  	rds_stats_inc(s_cong_update_received);
258  	atomic_inc(&rds_cong_generation);
259  	if (waitqueue_active(&map->m_waitq))
260  		wake_up(&map->m_waitq);
261  	if (waitqueue_active(&rds_poll_waitq))
262  		wake_up_all(&rds_poll_waitq);
263  
264  	if (portmask && !list_empty(&rds_cong_monitor)) {
265  		unsigned long flags;
266  		struct rds_sock *rs;
267  
268  		read_lock_irqsave(&rds_cong_monitor_lock, flags);
269  		list_for_each_entry(rs, &rds_cong_monitor, rs_cong_list) {
270  			spin_lock(&rs->rs_lock);
271  			rs->rs_cong_notify |= (rs->rs_cong_mask & portmask);
272  			rs->rs_cong_mask &= ~portmask;
273  			spin_unlock(&rs->rs_lock);
274  			if (rs->rs_cong_notify)
275  				rds_wake_sk_sleep(rs);
276  		}
277  		read_unlock_irqrestore(&rds_cong_monitor_lock, flags);
278  	}
279  }
280  EXPORT_SYMBOL_GPL(rds_cong_map_updated);
281  
rds_cong_updated_since(unsigned long * recent)282  int rds_cong_updated_since(unsigned long *recent)
283  {
284  	unsigned long gen = atomic_read(&rds_cong_generation);
285  
286  	if (likely(*recent == gen))
287  		return 0;
288  	*recent = gen;
289  	return 1;
290  }
291  
292  /*
293   * We're called under the locking that protects the sockets receive buffer
294   * consumption.  This makes it a lot easier for the caller to only call us
295   * when it knows that an existing set bit needs to be cleared, and vice versa.
296   * We can't block and we need to deal with concurrent sockets working against
297   * the same per-address map.
298   */
rds_cong_set_bit(struct rds_cong_map * map,__be16 port)299  void rds_cong_set_bit(struct rds_cong_map *map, __be16 port)
300  {
301  	unsigned long i;
302  	unsigned long off;
303  
304  	rdsdebug("setting congestion for %pI4:%u in map %p\n",
305  	  &map->m_addr, ntohs(port), map);
306  
307  	i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
308  	off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
309  
310  	set_bit_le(off, (void *)map->m_page_addrs[i]);
311  }
312  
rds_cong_clear_bit(struct rds_cong_map * map,__be16 port)313  void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port)
314  {
315  	unsigned long i;
316  	unsigned long off;
317  
318  	rdsdebug("clearing congestion for %pI4:%u in map %p\n",
319  	  &map->m_addr, ntohs(port), map);
320  
321  	i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
322  	off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
323  
324  	clear_bit_le(off, (void *)map->m_page_addrs[i]);
325  }
326  
rds_cong_test_bit(struct rds_cong_map * map,__be16 port)327  static int rds_cong_test_bit(struct rds_cong_map *map, __be16 port)
328  {
329  	unsigned long i;
330  	unsigned long off;
331  
332  	i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS;
333  	off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS;
334  
335  	return test_bit_le(off, (void *)map->m_page_addrs[i]);
336  }
337  
rds_cong_add_socket(struct rds_sock * rs)338  void rds_cong_add_socket(struct rds_sock *rs)
339  {
340  	unsigned long flags;
341  
342  	write_lock_irqsave(&rds_cong_monitor_lock, flags);
343  	if (list_empty(&rs->rs_cong_list))
344  		list_add(&rs->rs_cong_list, &rds_cong_monitor);
345  	write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
346  }
347  
rds_cong_remove_socket(struct rds_sock * rs)348  void rds_cong_remove_socket(struct rds_sock *rs)
349  {
350  	unsigned long flags;
351  	struct rds_cong_map *map;
352  
353  	write_lock_irqsave(&rds_cong_monitor_lock, flags);
354  	list_del_init(&rs->rs_cong_list);
355  	write_unlock_irqrestore(&rds_cong_monitor_lock, flags);
356  
357  	/* update congestion map for now-closed port */
358  	spin_lock_irqsave(&rds_cong_lock, flags);
359  	map = rds_cong_tree_walk(&rs->rs_bound_addr, NULL);
360  	spin_unlock_irqrestore(&rds_cong_lock, flags);
361  
362  	if (map && rds_cong_test_bit(map, rs->rs_bound_port)) {
363  		rds_cong_clear_bit(map, rs->rs_bound_port);
364  		rds_cong_queue_updates(map);
365  	}
366  }
367  
rds_cong_wait(struct rds_cong_map * map,__be16 port,int nonblock,struct rds_sock * rs)368  int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock,
369  		  struct rds_sock *rs)
370  {
371  	if (!rds_cong_test_bit(map, port))
372  		return 0;
373  	if (nonblock) {
374  		if (rs && rs->rs_cong_monitor) {
375  			unsigned long flags;
376  
377  			/* It would have been nice to have an atomic set_bit on
378  			 * a uint64_t. */
379  			spin_lock_irqsave(&rs->rs_lock, flags);
380  			rs->rs_cong_mask |= RDS_CONG_MONITOR_MASK(ntohs(port));
381  			spin_unlock_irqrestore(&rs->rs_lock, flags);
382  
383  			/* Test again - a congestion update may have arrived in
384  			 * the meantime. */
385  			if (!rds_cong_test_bit(map, port))
386  				return 0;
387  		}
388  		rds_stats_inc(s_cong_send_error);
389  		return -ENOBUFS;
390  	}
391  
392  	rds_stats_inc(s_cong_send_blocked);
393  	rdsdebug("waiting on map %p for port %u\n", map, be16_to_cpu(port));
394  
395  	return wait_event_interruptible(map->m_waitq,
396  					!rds_cong_test_bit(map, port));
397  }
398  
rds_cong_exit(void)399  void rds_cong_exit(void)
400  {
401  	struct rb_node *node;
402  	struct rds_cong_map *map;
403  	unsigned long i;
404  
405  	while ((node = rb_first(&rds_cong_tree))) {
406  		map = rb_entry(node, struct rds_cong_map, m_rb_node);
407  		rdsdebug("freeing map %p\n", map);
408  		rb_erase(&map->m_rb_node, &rds_cong_tree);
409  		for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++)
410  			free_page(map->m_page_addrs[i]);
411  		kfree(map);
412  	}
413  }
414  
415  /*
416   * Allocate a RDS message containing a congestion update.
417   */
rds_cong_update_alloc(struct rds_connection * conn)418  struct rds_message *rds_cong_update_alloc(struct rds_connection *conn)
419  {
420  	struct rds_cong_map *map = conn->c_lcong;
421  	struct rds_message *rm;
422  
423  	rm = rds_message_map_pages(map->m_page_addrs, RDS_CONG_MAP_BYTES);
424  	if (!IS_ERR(rm))
425  		rm->m_inc.i_hdr.h_flags = RDS_FLAG_CONG_BITMAP;
426  
427  	return rm;
428  }
429