Andy Grover | 7004108 | 2009-08-21 12:28:31 +0000 | [diff] [blame] | 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> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 34 | #include <linux/slab.h> |
Andy Grover | 7004108 | 2009-08-21 12:28:31 +0000 | [diff] [blame] | 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 | 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 iovec *first_iov, |
| 63 | size_t size) |
| 64 | { |
| 65 | struct rds_tcp_incoming *tinc; |
| 66 | struct iovec *iov, tmp; |
| 67 | struct sk_buff *skb; |
| 68 | unsigned long to_copy, skb_off; |
| 69 | int ret = 0; |
| 70 | |
| 71 | if (size == 0) |
| 72 | goto out; |
| 73 | |
| 74 | tinc = container_of(inc, struct rds_tcp_incoming, ti_inc); |
| 75 | iov = first_iov; |
| 76 | tmp = *iov; |
| 77 | |
| 78 | skb_queue_walk(&tinc->ti_skb_list, skb) { |
| 79 | skb_off = 0; |
| 80 | while (skb_off < skb->len) { |
| 81 | while (tmp.iov_len == 0) { |
| 82 | iov++; |
| 83 | tmp = *iov; |
| 84 | } |
| 85 | |
| 86 | to_copy = min(tmp.iov_len, size); |
| 87 | to_copy = min(to_copy, skb->len - skb_off); |
| 88 | |
| 89 | rdsdebug("ret %d size %zu skb %p skb_off %lu " |
| 90 | "skblen %d iov_base %p iov_len %zu cpy %lu\n", |
| 91 | ret, size, skb, skb_off, skb->len, |
| 92 | tmp.iov_base, tmp.iov_len, to_copy); |
| 93 | |
| 94 | /* modifies tmp as it copies */ |
| 95 | if (skb_copy_datagram_iovec(skb, skb_off, &tmp, |
| 96 | to_copy)) { |
| 97 | ret = -EFAULT; |
| 98 | goto out; |
| 99 | } |
| 100 | |
Andy Grover | b075cfd | 2010-03-11 13:49:57 +0000 | [diff] [blame] | 101 | rds_stats_add(s_copy_to_user, to_copy); |
Andy Grover | 7004108 | 2009-08-21 12:28:31 +0000 | [diff] [blame] | 102 | size -= to_copy; |
| 103 | ret += to_copy; |
| 104 | skb_off += to_copy; |
| 105 | if (size == 0) |
| 106 | goto out; |
| 107 | } |
| 108 | } |
| 109 | out: |
| 110 | return ret; |
| 111 | } |
| 112 | |
| 113 | /* |
| 114 | * We have a series of skbs that have fragmented pieces of the congestion |
| 115 | * bitmap. They must add up to the exact size of the congestion bitmap. We |
| 116 | * use the skb helpers to copy those into the pages that make up the in-memory |
| 117 | * congestion bitmap for the remote address of this connection. We then tell |
| 118 | * the congestion core that the bitmap has been changed so that it can wake up |
| 119 | * sleepers. |
| 120 | * |
| 121 | * This is racing with sending paths which are using test_bit to see if the |
| 122 | * bitmap indicates that their recipient is congested. |
| 123 | */ |
| 124 | |
| 125 | static void rds_tcp_cong_recv(struct rds_connection *conn, |
| 126 | struct rds_tcp_incoming *tinc) |
| 127 | { |
| 128 | struct sk_buff *skb; |
| 129 | unsigned int to_copy, skb_off; |
| 130 | unsigned int map_off; |
| 131 | unsigned int map_page; |
| 132 | struct rds_cong_map *map; |
| 133 | int ret; |
| 134 | |
| 135 | /* catch completely corrupt packets */ |
| 136 | if (be32_to_cpu(tinc->ti_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) |
| 137 | return; |
| 138 | |
| 139 | map_page = 0; |
| 140 | map_off = 0; |
| 141 | map = conn->c_fcong; |
| 142 | |
| 143 | skb_queue_walk(&tinc->ti_skb_list, skb) { |
| 144 | skb_off = 0; |
| 145 | while (skb_off < skb->len) { |
| 146 | to_copy = min_t(unsigned int, PAGE_SIZE - map_off, |
| 147 | skb->len - skb_off); |
| 148 | |
| 149 | BUG_ON(map_page >= RDS_CONG_MAP_PAGES); |
| 150 | |
| 151 | /* only returns 0 or -error */ |
| 152 | ret = skb_copy_bits(skb, skb_off, |
| 153 | (void *)map->m_page_addrs[map_page] + map_off, |
| 154 | to_copy); |
| 155 | BUG_ON(ret != 0); |
| 156 | |
| 157 | skb_off += to_copy; |
| 158 | map_off += to_copy; |
| 159 | if (map_off == PAGE_SIZE) { |
| 160 | map_off = 0; |
| 161 | map_page++; |
| 162 | } |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | rds_cong_map_updated(map, ~(u64) 0); |
| 167 | } |
| 168 | |
| 169 | struct rds_tcp_desc_arg { |
| 170 | struct rds_connection *conn; |
| 171 | gfp_t gfp; |
| 172 | enum km_type km; |
| 173 | }; |
| 174 | |
| 175 | static int rds_tcp_data_recv(read_descriptor_t *desc, struct sk_buff *skb, |
| 176 | unsigned int offset, size_t len) |
| 177 | { |
| 178 | struct rds_tcp_desc_arg *arg = desc->arg.data; |
| 179 | struct rds_connection *conn = arg->conn; |
| 180 | struct rds_tcp_connection *tc = conn->c_transport_data; |
| 181 | struct rds_tcp_incoming *tinc = tc->t_tinc; |
| 182 | struct sk_buff *clone; |
| 183 | size_t left = len, to_copy; |
| 184 | |
| 185 | rdsdebug("tcp data tc %p skb %p offset %u len %zu\n", tc, skb, offset, |
| 186 | len); |
| 187 | |
| 188 | /* |
| 189 | * tcp_read_sock() interprets partial progress as an indication to stop |
| 190 | * processing. |
| 191 | */ |
| 192 | while (left) { |
| 193 | if (tinc == NULL) { |
| 194 | tinc = kmem_cache_alloc(rds_tcp_incoming_slab, |
| 195 | arg->gfp); |
| 196 | if (tinc == NULL) { |
| 197 | desc->error = -ENOMEM; |
| 198 | goto out; |
| 199 | } |
| 200 | tc->t_tinc = tinc; |
| 201 | rdsdebug("alloced tinc %p\n", tinc); |
| 202 | rds_inc_init(&tinc->ti_inc, conn, conn->c_faddr); |
| 203 | /* |
| 204 | * XXX * we might be able to use the __ variants when |
| 205 | * we've already serialized at a higher level. |
| 206 | */ |
| 207 | skb_queue_head_init(&tinc->ti_skb_list); |
| 208 | } |
| 209 | |
| 210 | if (left && tc->t_tinc_hdr_rem) { |
| 211 | to_copy = min(tc->t_tinc_hdr_rem, left); |
| 212 | rdsdebug("copying %zu header from skb %p\n", to_copy, |
| 213 | skb); |
| 214 | skb_copy_bits(skb, offset, |
| 215 | (char *)&tinc->ti_inc.i_hdr + |
| 216 | sizeof(struct rds_header) - |
| 217 | tc->t_tinc_hdr_rem, |
| 218 | to_copy); |
| 219 | tc->t_tinc_hdr_rem -= to_copy; |
| 220 | left -= to_copy; |
| 221 | offset += to_copy; |
| 222 | |
| 223 | if (tc->t_tinc_hdr_rem == 0) { |
| 224 | /* could be 0 for a 0 len message */ |
| 225 | tc->t_tinc_data_rem = |
| 226 | be32_to_cpu(tinc->ti_inc.i_hdr.h_len); |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | if (left && tc->t_tinc_data_rem) { |
| 231 | clone = skb_clone(skb, arg->gfp); |
| 232 | if (clone == NULL) { |
| 233 | desc->error = -ENOMEM; |
| 234 | goto out; |
| 235 | } |
| 236 | |
| 237 | to_copy = min(tc->t_tinc_data_rem, left); |
| 238 | pskb_pull(clone, offset); |
| 239 | pskb_trim(clone, to_copy); |
| 240 | skb_queue_tail(&tinc->ti_skb_list, clone); |
| 241 | |
| 242 | rdsdebug("skb %p data %p len %d off %u to_copy %zu -> " |
| 243 | "clone %p data %p len %d\n", |
| 244 | skb, skb->data, skb->len, offset, to_copy, |
| 245 | clone, clone->data, clone->len); |
| 246 | |
| 247 | tc->t_tinc_data_rem -= to_copy; |
| 248 | left -= to_copy; |
| 249 | offset += to_copy; |
| 250 | } |
| 251 | |
| 252 | if (tc->t_tinc_hdr_rem == 0 && tc->t_tinc_data_rem == 0) { |
| 253 | if (tinc->ti_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) |
| 254 | rds_tcp_cong_recv(conn, tinc); |
| 255 | else |
| 256 | rds_recv_incoming(conn, conn->c_faddr, |
| 257 | conn->c_laddr, &tinc->ti_inc, |
| 258 | arg->gfp, arg->km); |
| 259 | |
| 260 | tc->t_tinc_hdr_rem = sizeof(struct rds_header); |
| 261 | tc->t_tinc_data_rem = 0; |
| 262 | tc->t_tinc = NULL; |
| 263 | rds_inc_put(&tinc->ti_inc); |
| 264 | tinc = NULL; |
| 265 | } |
| 266 | } |
| 267 | out: |
| 268 | rdsdebug("returning len %zu left %zu skb len %d rx queue depth %d\n", |
| 269 | len, left, skb->len, |
| 270 | skb_queue_len(&tc->t_sock->sk->sk_receive_queue)); |
| 271 | return len - left; |
| 272 | } |
| 273 | |
| 274 | /* the caller has to hold the sock lock */ |
| 275 | int rds_tcp_read_sock(struct rds_connection *conn, gfp_t gfp, enum km_type km) |
| 276 | { |
| 277 | struct rds_tcp_connection *tc = conn->c_transport_data; |
| 278 | struct socket *sock = tc->t_sock; |
| 279 | read_descriptor_t desc; |
| 280 | struct rds_tcp_desc_arg arg; |
| 281 | |
| 282 | /* It's like glib in the kernel! */ |
| 283 | arg.conn = conn; |
| 284 | arg.gfp = gfp; |
| 285 | arg.km = km; |
| 286 | desc.arg.data = &arg; |
| 287 | desc.error = 0; |
| 288 | desc.count = 1; /* give more than one skb per call */ |
| 289 | |
| 290 | tcp_read_sock(sock->sk, &desc, rds_tcp_data_recv); |
| 291 | rdsdebug("tcp_read_sock for tc %p gfp 0x%x returned %d\n", tc, gfp, |
| 292 | desc.error); |
| 293 | |
| 294 | return desc.error; |
| 295 | } |
| 296 | |
| 297 | /* |
| 298 | * We hold the sock lock to serialize our rds_tcp_recv->tcp_read_sock from |
| 299 | * data_ready. |
| 300 | * |
| 301 | * if we fail to allocate we're in trouble.. blindly wait some time before |
| 302 | * trying again to see if the VM can free up something for us. |
| 303 | */ |
| 304 | int rds_tcp_recv(struct rds_connection *conn) |
| 305 | { |
| 306 | struct rds_tcp_connection *tc = conn->c_transport_data; |
| 307 | struct socket *sock = tc->t_sock; |
| 308 | int ret = 0; |
| 309 | |
| 310 | rdsdebug("recv worker conn %p tc %p sock %p\n", conn, tc, sock); |
| 311 | |
| 312 | lock_sock(sock->sk); |
| 313 | ret = rds_tcp_read_sock(conn, GFP_KERNEL, KM_USER0); |
| 314 | release_sock(sock->sk); |
| 315 | |
| 316 | return ret; |
| 317 | } |
| 318 | |
| 319 | void rds_tcp_data_ready(struct sock *sk, int bytes) |
| 320 | { |
| 321 | void (*ready)(struct sock *sk, int bytes); |
| 322 | struct rds_connection *conn; |
| 323 | struct rds_tcp_connection *tc; |
| 324 | |
| 325 | rdsdebug("data ready sk %p bytes %d\n", sk, bytes); |
| 326 | |
Eric Dumazet | f064af1 | 2010-09-22 12:43:39 +0000 | [diff] [blame^] | 327 | read_lock_bh(&sk->sk_callback_lock); |
Andy Grover | 7004108 | 2009-08-21 12:28:31 +0000 | [diff] [blame] | 328 | conn = sk->sk_user_data; |
| 329 | if (conn == NULL) { /* check for teardown race */ |
| 330 | ready = sk->sk_data_ready; |
| 331 | goto out; |
| 332 | } |
| 333 | |
| 334 | tc = conn->c_transport_data; |
| 335 | ready = tc->t_orig_data_ready; |
| 336 | rds_tcp_stats_inc(s_tcp_data_ready_calls); |
| 337 | |
| 338 | if (rds_tcp_read_sock(conn, GFP_ATOMIC, KM_SOFTIRQ0) == -ENOMEM) |
| 339 | queue_delayed_work(rds_wq, &conn->c_recv_w, 0); |
| 340 | out: |
Eric Dumazet | f064af1 | 2010-09-22 12:43:39 +0000 | [diff] [blame^] | 341 | read_unlock_bh(&sk->sk_callback_lock); |
Andy Grover | 7004108 | 2009-08-21 12:28:31 +0000 | [diff] [blame] | 342 | ready(sk, bytes); |
| 343 | } |
| 344 | |
| 345 | int __init rds_tcp_recv_init(void) |
| 346 | { |
| 347 | rds_tcp_incoming_slab = kmem_cache_create("rds_tcp_incoming", |
| 348 | sizeof(struct rds_tcp_incoming), |
| 349 | 0, 0, NULL); |
| 350 | if (rds_tcp_incoming_slab == NULL) |
| 351 | return -ENOMEM; |
| 352 | return 0; |
| 353 | } |
| 354 | |
| 355 | void rds_tcp_recv_exit(void) |
| 356 | { |
| 357 | kmem_cache_destroy(rds_tcp_incoming_slab); |
| 358 | } |