Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 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 | */ |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 33 | #include <linux/slab.h> |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 34 | #include <linux/types.h> |
| 35 | #include <linux/rbtree.h> |
Akinobu Mita | 12ce224 | 2011-03-23 16:41:45 -0700 | [diff] [blame] | 36 | #include <linux/bitops.h> |
Paul Gortmaker | bc3b2d7 | 2011-07-15 11:47:34 -0400 | [diff] [blame] | 37 | #include <linux/export.h> |
David S. Miller | c3b3240 | 2009-03-02 01:49:28 -0800 | [diff] [blame] | 38 | |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 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 | |
| 104 | static 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 | */ |
| 135 | static 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); |
Andy Grover | 8690bfa | 2010-01-12 11:56:44 -0800 | [diff] [blame] | 144 | if (!map) |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 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 | |
Andy Grover | 8690bfa | 2010-01-12 11:56:44 -0800 | [diff] [blame] | 162 | if (!ret) { |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 163 | ret = map; |
| 164 | map = NULL; |
| 165 | } |
| 166 | |
| 167 | out: |
| 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 | */ |
| 183 | void 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 | |
| 193 | void 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 | |
| 203 | int 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 | |
Andy Grover | 8690bfa | 2010-01-12 11:56:44 -0800 | [diff] [blame] | 208 | if (!(conn->c_lcong && conn->c_fcong)) |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 209 | return -ENOMEM; |
| 210 | |
| 211 | return 0; |
| 212 | } |
| 213 | |
| 214 | void 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 (!test_and_set_bit(0, &conn->c_map_queued)) { |
| 223 | rds_stats_inc(s_cong_update_queued); |
Andy Grover | 2fa5712 | 2010-03-29 16:45:40 -0700 | [diff] [blame] | 224 | rds_send_xmit(conn); |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 225 | } |
| 226 | } |
| 227 | |
| 228 | spin_unlock_irqrestore(&rds_cong_lock, flags); |
| 229 | } |
| 230 | |
| 231 | void rds_cong_map_updated(struct rds_cong_map *map, uint64_t portmask) |
| 232 | { |
| 233 | rdsdebug("waking map %p for %pI4\n", |
| 234 | map, &map->m_addr); |
| 235 | rds_stats_inc(s_cong_update_received); |
| 236 | atomic_inc(&rds_cong_generation); |
| 237 | if (waitqueue_active(&map->m_waitq)) |
| 238 | wake_up(&map->m_waitq); |
| 239 | if (waitqueue_active(&rds_poll_waitq)) |
| 240 | wake_up_all(&rds_poll_waitq); |
| 241 | |
| 242 | if (portmask && !list_empty(&rds_cong_monitor)) { |
| 243 | unsigned long flags; |
| 244 | struct rds_sock *rs; |
| 245 | |
| 246 | read_lock_irqsave(&rds_cong_monitor_lock, flags); |
| 247 | list_for_each_entry(rs, &rds_cong_monitor, rs_cong_list) { |
| 248 | spin_lock(&rs->rs_lock); |
| 249 | rs->rs_cong_notify |= (rs->rs_cong_mask & portmask); |
| 250 | rs->rs_cong_mask &= ~portmask; |
| 251 | spin_unlock(&rs->rs_lock); |
| 252 | if (rs->rs_cong_notify) |
| 253 | rds_wake_sk_sleep(rs); |
| 254 | } |
| 255 | read_unlock_irqrestore(&rds_cong_monitor_lock, flags); |
| 256 | } |
| 257 | } |
Andy Grover | 616b757 | 2009-08-21 12:28:32 +0000 | [diff] [blame] | 258 | EXPORT_SYMBOL_GPL(rds_cong_map_updated); |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 259 | |
| 260 | int rds_cong_updated_since(unsigned long *recent) |
| 261 | { |
| 262 | unsigned long gen = atomic_read(&rds_cong_generation); |
| 263 | |
| 264 | if (likely(*recent == gen)) |
| 265 | return 0; |
| 266 | *recent = gen; |
| 267 | return 1; |
| 268 | } |
| 269 | |
| 270 | /* |
| 271 | * We're called under the locking that protects the sockets receive buffer |
| 272 | * consumption. This makes it a lot easier for the caller to only call us |
| 273 | * when it knows that an existing set bit needs to be cleared, and vice versa. |
| 274 | * We can't block and we need to deal with concurrent sockets working against |
| 275 | * the same per-address map. |
| 276 | */ |
| 277 | void rds_cong_set_bit(struct rds_cong_map *map, __be16 port) |
| 278 | { |
| 279 | unsigned long i; |
| 280 | unsigned long off; |
| 281 | |
| 282 | rdsdebug("setting congestion for %pI4:%u in map %p\n", |
| 283 | &map->m_addr, ntohs(port), map); |
| 284 | |
| 285 | i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS; |
| 286 | off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS; |
| 287 | |
Akinobu Mita | e1dc1c8 | 2011-03-23 16:42:05 -0700 | [diff] [blame] | 288 | __set_bit_le(off, (void *)map->m_page_addrs[i]); |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 289 | } |
| 290 | |
| 291 | void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port) |
| 292 | { |
| 293 | unsigned long i; |
| 294 | unsigned long off; |
| 295 | |
| 296 | rdsdebug("clearing congestion for %pI4:%u in map %p\n", |
| 297 | &map->m_addr, ntohs(port), map); |
| 298 | |
| 299 | i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS; |
| 300 | off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS; |
| 301 | |
Akinobu Mita | e1dc1c8 | 2011-03-23 16:42:05 -0700 | [diff] [blame] | 302 | __clear_bit_le(off, (void *)map->m_page_addrs[i]); |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 303 | } |
| 304 | |
| 305 | static int rds_cong_test_bit(struct rds_cong_map *map, __be16 port) |
| 306 | { |
| 307 | unsigned long i; |
| 308 | unsigned long off; |
| 309 | |
| 310 | i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS; |
| 311 | off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS; |
| 312 | |
Akinobu Mita | e1dc1c8 | 2011-03-23 16:42:05 -0700 | [diff] [blame] | 313 | return test_bit_le(off, (void *)map->m_page_addrs[i]); |
Andy Grover | 922cb17 | 2009-02-24 15:30:20 +0000 | [diff] [blame] | 314 | } |
| 315 | |
| 316 | void rds_cong_add_socket(struct rds_sock *rs) |
| 317 | { |
| 318 | unsigned long flags; |
| 319 | |
| 320 | write_lock_irqsave(&rds_cong_monitor_lock, flags); |
| 321 | if (list_empty(&rs->rs_cong_list)) |
| 322 | list_add(&rs->rs_cong_list, &rds_cong_monitor); |
| 323 | write_unlock_irqrestore(&rds_cong_monitor_lock, flags); |
| 324 | } |
| 325 | |
| 326 | void rds_cong_remove_socket(struct rds_sock *rs) |
| 327 | { |
| 328 | unsigned long flags; |
| 329 | struct rds_cong_map *map; |
| 330 | |
| 331 | write_lock_irqsave(&rds_cong_monitor_lock, flags); |
| 332 | list_del_init(&rs->rs_cong_list); |
| 333 | write_unlock_irqrestore(&rds_cong_monitor_lock, flags); |
| 334 | |
| 335 | /* update congestion map for now-closed port */ |
| 336 | spin_lock_irqsave(&rds_cong_lock, flags); |
| 337 | map = rds_cong_tree_walk(rs->rs_bound_addr, NULL); |
| 338 | spin_unlock_irqrestore(&rds_cong_lock, flags); |
| 339 | |
| 340 | if (map && rds_cong_test_bit(map, rs->rs_bound_port)) { |
| 341 | rds_cong_clear_bit(map, rs->rs_bound_port); |
| 342 | rds_cong_queue_updates(map); |
| 343 | } |
| 344 | } |
| 345 | |
| 346 | int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, |
| 347 | struct rds_sock *rs) |
| 348 | { |
| 349 | if (!rds_cong_test_bit(map, port)) |
| 350 | return 0; |
| 351 | if (nonblock) { |
| 352 | if (rs && rs->rs_cong_monitor) { |
| 353 | unsigned long flags; |
| 354 | |
| 355 | /* It would have been nice to have an atomic set_bit on |
| 356 | * a uint64_t. */ |
| 357 | spin_lock_irqsave(&rs->rs_lock, flags); |
| 358 | rs->rs_cong_mask |= RDS_CONG_MONITOR_MASK(ntohs(port)); |
| 359 | spin_unlock_irqrestore(&rs->rs_lock, flags); |
| 360 | |
| 361 | /* Test again - a congestion update may have arrived in |
| 362 | * the meantime. */ |
| 363 | if (!rds_cong_test_bit(map, port)) |
| 364 | return 0; |
| 365 | } |
| 366 | rds_stats_inc(s_cong_send_error); |
| 367 | return -ENOBUFS; |
| 368 | } |
| 369 | |
| 370 | rds_stats_inc(s_cong_send_blocked); |
| 371 | rdsdebug("waiting on map %p for port %u\n", map, be16_to_cpu(port)); |
| 372 | |
| 373 | return wait_event_interruptible(map->m_waitq, |
| 374 | !rds_cong_test_bit(map, port)); |
| 375 | } |
| 376 | |
| 377 | void rds_cong_exit(void) |
| 378 | { |
| 379 | struct rb_node *node; |
| 380 | struct rds_cong_map *map; |
| 381 | unsigned long i; |
| 382 | |
| 383 | while ((node = rb_first(&rds_cong_tree))) { |
| 384 | map = rb_entry(node, struct rds_cong_map, m_rb_node); |
| 385 | rdsdebug("freeing map %p\n", map); |
| 386 | rb_erase(&map->m_rb_node, &rds_cong_tree); |
| 387 | for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++) |
| 388 | free_page(map->m_page_addrs[i]); |
| 389 | kfree(map); |
| 390 | } |
| 391 | } |
| 392 | |
| 393 | /* |
| 394 | * Allocate a RDS message containing a congestion update. |
| 395 | */ |
| 396 | struct rds_message *rds_cong_update_alloc(struct rds_connection *conn) |
| 397 | { |
| 398 | struct rds_cong_map *map = conn->c_lcong; |
| 399 | struct rds_message *rm; |
| 400 | |
| 401 | rm = rds_message_map_pages(map->m_page_addrs, RDS_CONG_MAP_BYTES); |
| 402 | if (!IS_ERR(rm)) |
| 403 | rm->m_inc.i_hdr.h_flags = RDS_FLAG_CONG_BITMAP; |
| 404 | |
| 405 | return rm; |
| 406 | } |