1 /* 2 * Copyright (c) 2006 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/kernel.h> 34 #include <linux/slab.h> 35 #include <net/tcp.h> 36 37 #include "rds.h" 38 #include "tcp.h" 39 40 static struct kmem_cache *rds_tcp_incoming_slab; 41 42 static void rds_tcp_inc_purge(struct rds_incoming *inc) 43 { 44 struct rds_tcp_incoming *tinc; 45 tinc = container_of(inc, struct rds_tcp_incoming, ti_inc); 46 rdsdebug("purging tinc %p inc %p\n", tinc, inc); 47 skb_queue_purge(&tinc->ti_skb_list); 48 } 49 50 void rds_tcp_inc_free(struct rds_incoming *inc) 51 { 52 struct rds_tcp_incoming *tinc; 53 tinc = container_of(inc, struct rds_tcp_incoming, ti_inc); 54 rds_tcp_inc_purge(inc); 55 rdsdebug("freeing tinc %p inc %p\n", tinc, inc); 56 kmem_cache_free(rds_tcp_incoming_slab, tinc); 57 } 58 59 /* 60 * this is pretty lame, but, whatever. 61 */ 62 int rds_tcp_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to) 63 { 64 struct rds_tcp_incoming *tinc; 65 struct sk_buff *skb; 66 int ret = 0; 67 68 if (!iov_iter_count(to)) 69 goto out; 70 71 tinc = container_of(inc, struct rds_tcp_incoming, ti_inc); 72 73 skb_queue_walk(&tinc->ti_skb_list, skb) { 74 unsigned long to_copy, skb_off; 75 for (skb_off = 0; skb_off < skb->len; skb_off += to_copy) { 76 to_copy = iov_iter_count(to); 77 to_copy = min(to_copy, skb->len - skb_off); 78 79 if (skb_copy_datagram_iter(skb, skb_off, to, to_copy)) 80 return -EFAULT; 81 82 rds_stats_add(s_copy_to_user, to_copy); 83 ret += to_copy; 84 85 if (!iov_iter_count(to)) 86 goto out; 87 } 88 } 89 out: 90 return ret; 91 } 92 93 /* 94 * We have a series of skbs that have fragmented pieces of the congestion 95 * bitmap. They must add up to the exact size of the congestion bitmap. We 96 * use the skb helpers to copy those into the pages that make up the in-memory 97 * congestion bitmap for the remote address of this connection. We then tell 98 * the congestion core that the bitmap has been changed so that it can wake up 99 * sleepers. 100 * 101 * This is racing with sending paths which are using test_bit to see if the 102 * bitmap indicates that their recipient is congested. 103 */ 104 105 static void rds_tcp_cong_recv(struct rds_connection *conn, 106 struct rds_tcp_incoming *tinc) 107 { 108 struct sk_buff *skb; 109 unsigned int to_copy, skb_off; 110 unsigned int map_off; 111 unsigned int map_page; 112 struct rds_cong_map *map; 113 int ret; 114 115 /* catch completely corrupt packets */ 116 if (be32_to_cpu(tinc->ti_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) 117 return; 118 119 map_page = 0; 120 map_off = 0; 121 map = conn->c_fcong; 122 123 skb_queue_walk(&tinc->ti_skb_list, skb) { 124 skb_off = 0; 125 while (skb_off < skb->len) { 126 to_copy = min_t(unsigned int, PAGE_SIZE - map_off, 127 skb->len - skb_off); 128 129 BUG_ON(map_page >= RDS_CONG_MAP_PAGES); 130 131 /* only returns 0 or -error */ 132 ret = skb_copy_bits(skb, skb_off, 133 (void *)map->m_page_addrs[map_page] + map_off, 134 to_copy); 135 BUG_ON(ret != 0); 136 137 skb_off += to_copy; 138 map_off += to_copy; 139 if (map_off == PAGE_SIZE) { 140 map_off = 0; 141 map_page++; 142 } 143 } 144 } 145 146 rds_cong_map_updated(map, ~(u64) 0); 147 } 148 149 struct rds_tcp_desc_arg { 150 struct rds_conn_path *conn_path; 151 gfp_t gfp; 152 }; 153 154 static int rds_tcp_data_recv(read_descriptor_t *desc, struct sk_buff *skb, 155 unsigned int offset, size_t len) 156 { 157 struct rds_tcp_desc_arg *arg = desc->arg.data; 158 struct rds_conn_path *cp = arg->conn_path; 159 struct rds_tcp_connection *tc = cp->cp_transport_data; 160 struct rds_tcp_incoming *tinc = tc->t_tinc; 161 struct sk_buff *clone; 162 size_t left = len, to_copy; 163 164 rdsdebug("tcp data tc %p skb %p offset %u len %zu\n", tc, skb, offset, 165 len); 166 167 /* 168 * tcp_read_sock() interprets partial progress as an indication to stop 169 * processing. 170 */ 171 while (left) { 172 if (!tinc) { 173 tinc = kmem_cache_alloc(rds_tcp_incoming_slab, 174 arg->gfp); 175 if (!tinc) { 176 desc->error = -ENOMEM; 177 goto out; 178 } 179 tc->t_tinc = tinc; 180 rdsdebug("alloced tinc %p\n", tinc); 181 rds_inc_path_init(&tinc->ti_inc, cp, 182 cp->cp_conn->c_faddr); 183 /* 184 * XXX * we might be able to use the __ variants when 185 * we've already serialized at a higher level. 186 */ 187 skb_queue_head_init(&tinc->ti_skb_list); 188 } 189 190 if (left && tc->t_tinc_hdr_rem) { 191 to_copy = min(tc->t_tinc_hdr_rem, left); 192 rdsdebug("copying %zu header from skb %p\n", to_copy, 193 skb); 194 skb_copy_bits(skb, offset, 195 (char *)&tinc->ti_inc.i_hdr + 196 sizeof(struct rds_header) - 197 tc->t_tinc_hdr_rem, 198 to_copy); 199 tc->t_tinc_hdr_rem -= to_copy; 200 left -= to_copy; 201 offset += to_copy; 202 203 if (tc->t_tinc_hdr_rem == 0) { 204 /* could be 0 for a 0 len message */ 205 tc->t_tinc_data_rem = 206 be32_to_cpu(tinc->ti_inc.i_hdr.h_len); 207 } 208 } 209 210 if (left && tc->t_tinc_data_rem) { 211 to_copy = min(tc->t_tinc_data_rem, left); 212 213 clone = pskb_extract(skb, offset, to_copy, arg->gfp); 214 if (!clone) { 215 desc->error = -ENOMEM; 216 goto out; 217 } 218 219 skb_queue_tail(&tinc->ti_skb_list, clone); 220 221 rdsdebug("skb %p data %p len %d off %u to_copy %zu -> " 222 "clone %p data %p len %d\n", 223 skb, skb->data, skb->len, offset, to_copy, 224 clone, clone->data, clone->len); 225 226 tc->t_tinc_data_rem -= to_copy; 227 left -= to_copy; 228 offset += to_copy; 229 } 230 231 if (tc->t_tinc_hdr_rem == 0 && tc->t_tinc_data_rem == 0) { 232 struct rds_connection *conn = cp->cp_conn; 233 234 if (tinc->ti_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) 235 rds_tcp_cong_recv(conn, tinc); 236 else 237 rds_recv_incoming(conn, conn->c_faddr, 238 conn->c_laddr, &tinc->ti_inc, 239 arg->gfp); 240 241 tc->t_tinc_hdr_rem = sizeof(struct rds_header); 242 tc->t_tinc_data_rem = 0; 243 tc->t_tinc = NULL; 244 rds_inc_put(&tinc->ti_inc); 245 tinc = NULL; 246 } 247 } 248 out: 249 rdsdebug("returning len %zu left %zu skb len %d rx queue depth %d\n", 250 len, left, skb->len, 251 skb_queue_len(&tc->t_sock->sk->sk_receive_queue)); 252 return len - left; 253 } 254 255 /* the caller has to hold the sock lock */ 256 static int rds_tcp_read_sock(struct rds_conn_path *cp, gfp_t gfp) 257 { 258 struct rds_tcp_connection *tc = cp->cp_transport_data; 259 struct socket *sock = tc->t_sock; 260 read_descriptor_t desc; 261 struct rds_tcp_desc_arg arg; 262 263 /* It's like glib in the kernel! */ 264 arg.conn_path = cp; 265 arg.gfp = gfp; 266 desc.arg.data = &arg; 267 desc.error = 0; 268 desc.count = 1; /* give more than one skb per call */ 269 270 tcp_read_sock(sock->sk, &desc, rds_tcp_data_recv); 271 rdsdebug("tcp_read_sock for tc %p gfp 0x%x returned %d\n", tc, gfp, 272 desc.error); 273 274 return desc.error; 275 } 276 277 /* 278 * We hold the sock lock to serialize our rds_tcp_recv->tcp_read_sock from 279 * data_ready. 280 * 281 * if we fail to allocate we're in trouble.. blindly wait some time before 282 * trying again to see if the VM can free up something for us. 283 */ 284 int rds_tcp_recv_path(struct rds_conn_path *cp) 285 { 286 struct rds_tcp_connection *tc = cp->cp_transport_data; 287 struct socket *sock = tc->t_sock; 288 int ret = 0; 289 290 rdsdebug("recv worker path [%d] tc %p sock %p\n", 291 cp->cp_index, tc, sock); 292 293 lock_sock(sock->sk); 294 ret = rds_tcp_read_sock(cp, GFP_KERNEL); 295 release_sock(sock->sk); 296 297 return ret; 298 } 299 300 void rds_tcp_data_ready(struct sock *sk) 301 { 302 void (*ready)(struct sock *sk); 303 struct rds_conn_path *cp; 304 struct rds_tcp_connection *tc; 305 306 rdsdebug("data ready sk %p\n", sk); 307 308 read_lock_bh(&sk->sk_callback_lock); 309 cp = sk->sk_user_data; 310 if (!cp) { /* check for teardown race */ 311 ready = sk->sk_data_ready; 312 goto out; 313 } 314 315 tc = cp->cp_transport_data; 316 ready = tc->t_orig_data_ready; 317 rds_tcp_stats_inc(s_tcp_data_ready_calls); 318 319 if (rds_tcp_read_sock(cp, GFP_ATOMIC) == -ENOMEM) 320 queue_delayed_work(rds_wq, &cp->cp_recv_w, 0); 321 out: 322 read_unlock_bh(&sk->sk_callback_lock); 323 ready(sk); 324 } 325 326 int rds_tcp_recv_init(void) 327 { 328 rds_tcp_incoming_slab = kmem_cache_create("rds_tcp_incoming", 329 sizeof(struct rds_tcp_incoming), 330 0, 0, NULL); 331 if (!rds_tcp_incoming_slab) 332 return -ENOMEM; 333 return 0; 334 } 335 336 void rds_tcp_recv_exit(void) 337 { 338 kmem_cache_destroy(rds_tcp_incoming_slab); 339 } 340