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 tinc->ti_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] = 184 local_clock(); 185 186 /* 187 * XXX * we might be able to use the __ variants when 188 * we've already serialized at a higher level. 189 */ 190 skb_queue_head_init(&tinc->ti_skb_list); 191 } 192 193 if (left && tc->t_tinc_hdr_rem) { 194 to_copy = min(tc->t_tinc_hdr_rem, left); 195 rdsdebug("copying %zu header from skb %p\n", to_copy, 196 skb); 197 skb_copy_bits(skb, offset, 198 (char *)&tinc->ti_inc.i_hdr + 199 sizeof(struct rds_header) - 200 tc->t_tinc_hdr_rem, 201 to_copy); 202 tc->t_tinc_hdr_rem -= to_copy; 203 left -= to_copy; 204 offset += to_copy; 205 206 if (tc->t_tinc_hdr_rem == 0) { 207 /* could be 0 for a 0 len message */ 208 tc->t_tinc_data_rem = 209 be32_to_cpu(tinc->ti_inc.i_hdr.h_len); 210 tinc->ti_inc.i_rx_lat_trace[RDS_MSG_RX_START] = 211 local_clock(); 212 } 213 } 214 215 if (left && tc->t_tinc_data_rem) { 216 to_copy = min(tc->t_tinc_data_rem, left); 217 218 clone = pskb_extract(skb, offset, to_copy, arg->gfp); 219 if (!clone) { 220 desc->error = -ENOMEM; 221 goto out; 222 } 223 224 skb_queue_tail(&tinc->ti_skb_list, clone); 225 226 rdsdebug("skb %p data %p len %d off %u to_copy %zu -> " 227 "clone %p data %p len %d\n", 228 skb, skb->data, skb->len, offset, to_copy, 229 clone, clone->data, clone->len); 230 231 tc->t_tinc_data_rem -= to_copy; 232 left -= to_copy; 233 offset += to_copy; 234 } 235 236 if (tc->t_tinc_hdr_rem == 0 && tc->t_tinc_data_rem == 0) { 237 struct rds_connection *conn = cp->cp_conn; 238 239 if (tinc->ti_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) 240 rds_tcp_cong_recv(conn, tinc); 241 else 242 rds_recv_incoming(conn, conn->c_faddr, 243 conn->c_laddr, &tinc->ti_inc, 244 arg->gfp); 245 246 tc->t_tinc_hdr_rem = sizeof(struct rds_header); 247 tc->t_tinc_data_rem = 0; 248 tc->t_tinc = NULL; 249 rds_inc_put(&tinc->ti_inc); 250 tinc = NULL; 251 } 252 } 253 out: 254 rdsdebug("returning len %zu left %zu skb len %d rx queue depth %d\n", 255 len, left, skb->len, 256 skb_queue_len(&tc->t_sock->sk->sk_receive_queue)); 257 return len - left; 258 } 259 260 /* the caller has to hold the sock lock */ 261 static int rds_tcp_read_sock(struct rds_conn_path *cp, gfp_t gfp) 262 { 263 struct rds_tcp_connection *tc = cp->cp_transport_data; 264 struct socket *sock = tc->t_sock; 265 read_descriptor_t desc; 266 struct rds_tcp_desc_arg arg; 267 268 /* It's like glib in the kernel! */ 269 arg.conn_path = cp; 270 arg.gfp = gfp; 271 desc.arg.data = &arg; 272 desc.error = 0; 273 desc.count = 1; /* give more than one skb per call */ 274 275 tcp_read_sock(sock->sk, &desc, rds_tcp_data_recv); 276 rdsdebug("tcp_read_sock for tc %p gfp 0x%x returned %d\n", tc, gfp, 277 desc.error); 278 279 return desc.error; 280 } 281 282 /* 283 * We hold the sock lock to serialize our rds_tcp_recv->tcp_read_sock from 284 * data_ready. 285 * 286 * if we fail to allocate we're in trouble.. blindly wait some time before 287 * trying again to see if the VM can free up something for us. 288 */ 289 int rds_tcp_recv_path(struct rds_conn_path *cp) 290 { 291 struct rds_tcp_connection *tc = cp->cp_transport_data; 292 struct socket *sock = tc->t_sock; 293 int ret = 0; 294 295 rdsdebug("recv worker path [%d] tc %p sock %p\n", 296 cp->cp_index, tc, sock); 297 298 lock_sock(sock->sk); 299 ret = rds_tcp_read_sock(cp, GFP_KERNEL); 300 release_sock(sock->sk); 301 302 return ret; 303 } 304 305 void rds_tcp_data_ready(struct sock *sk) 306 { 307 void (*ready)(struct sock *sk); 308 struct rds_conn_path *cp; 309 struct rds_tcp_connection *tc; 310 311 rdsdebug("data ready sk %p\n", sk); 312 313 read_lock_bh(&sk->sk_callback_lock); 314 cp = sk->sk_user_data; 315 if (!cp) { /* check for teardown race */ 316 ready = sk->sk_data_ready; 317 goto out; 318 } 319 320 tc = cp->cp_transport_data; 321 ready = tc->t_orig_data_ready; 322 rds_tcp_stats_inc(s_tcp_data_ready_calls); 323 324 if (rds_tcp_read_sock(cp, GFP_ATOMIC) == -ENOMEM) 325 queue_delayed_work(rds_wq, &cp->cp_recv_w, 0); 326 out: 327 read_unlock_bh(&sk->sk_callback_lock); 328 ready(sk); 329 } 330 331 int rds_tcp_recv_init(void) 332 { 333 rds_tcp_incoming_slab = kmem_cache_create("rds_tcp_incoming", 334 sizeof(struct rds_tcp_incoming), 335 0, 0, NULL); 336 if (!rds_tcp_incoming_slab) 337 return -ENOMEM; 338 return 0; 339 } 340 341 void rds_tcp_recv_exit(void) 342 { 343 kmem_cache_destroy(rds_tcp_incoming_slab); 344 } 345