1 /* 1
2 * Copyright (c) 2006, 2017 Oracle and/or its
3 *
4 * This software is available to you under a c
5 * licenses. You may choose to be licensed un
6 * General Public License (GPL) Version 2, ava
7 * COPYING in the main directory of this sourc
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and bi
11 * without modification, are permitted pro
12 * conditions are met:
13 *
14 * - Redistributions of source code must
15 * copyright notice, this list of condi
16 * disclaimer.
17 *
18 * - Redistributions in binary form must
19 * copyright notice, this list of condi
20 * disclaimer in the documentation and/
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT W
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMIT
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR P
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTH
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER L
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARIS
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR
30 * SOFTWARE.
31 *
32 */
33 #include <linux/kernel.h>
34 #include <linux/slab.h>
35 #include <net/tcp.h>
36 #include <trace/events/sock.h>
37
38 #include "rds.h"
39 #include "tcp.h"
40
41 static struct kmem_cache *rds_tcp_incoming_sla
42
43 static void rds_tcp_inc_purge(struct rds_incom
44 {
45 struct rds_tcp_incoming *tinc;
46 tinc = container_of(inc, struct rds_tc
47 rdsdebug("purging tinc %p inc %p\n", t
48 skb_queue_purge(&tinc->ti_skb_list);
49 }
50
51 void rds_tcp_inc_free(struct rds_incoming *inc
52 {
53 struct rds_tcp_incoming *tinc;
54 tinc = container_of(inc, struct rds_tc
55 rds_tcp_inc_purge(inc);
56 rdsdebug("freeing tinc %p inc %p\n", t
57 kmem_cache_free(rds_tcp_incoming_slab,
58 }
59
60 /*
61 * this is pretty lame, but, whatever.
62 */
63 int rds_tcp_inc_copy_to_user(struct rds_incomi
64 {
65 struct rds_tcp_incoming *tinc;
66 struct sk_buff *skb;
67 int ret = 0;
68
69 if (!iov_iter_count(to))
70 goto out;
71
72 tinc = container_of(inc, struct rds_tc
73
74 skb_queue_walk(&tinc->ti_skb_list, skb
75 unsigned long to_copy, skb_off
76 for (skb_off = 0; skb_off < sk
77 to_copy = iov_iter_cou
78 to_copy = min(to_copy,
79
80 if (skb_copy_datagram_
81 return -EFAULT
82
83 rds_stats_add(s_copy_t
84 ret += to_copy;
85
86 if (!iov_iter_count(to
87 goto out;
88 }
89 }
90 out:
91 return ret;
92 }
93
94 /*
95 * We have a series of skbs that have fragment
96 * bitmap. They must add up to the exact size
97 * use the skb helpers to copy those into the
98 * congestion bitmap for the remote address of
99 * the congestion core that the bitmap has bee
100 * sleepers.
101 *
102 * This is racing with sending paths which are
103 * bitmap indicates that their recipient is co
104 */
105
106 static void rds_tcp_cong_recv(struct rds_conne
107 struct rds_tcp_i
108 {
109 struct sk_buff *skb;
110 unsigned int to_copy, skb_off;
111 unsigned int map_off;
112 unsigned int map_page;
113 struct rds_cong_map *map;
114 int ret;
115
116 /* catch completely corrupt packets */
117 if (be32_to_cpu(tinc->ti_inc.i_hdr.h_l
118 return;
119
120 map_page = 0;
121 map_off = 0;
122 map = conn->c_fcong;
123
124 skb_queue_walk(&tinc->ti_skb_list, skb
125 skb_off = 0;
126 while (skb_off < skb->len) {
127 to_copy = min_t(unsign
128 skb->l
129
130 BUG_ON(map_page >= RDS
131
132 /* only returns 0 or -
133 ret = skb_copy_bits(sk
134 (void *)map->m
135 to_copy);
136 BUG_ON(ret != 0);
137
138 skb_off += to_copy;
139 map_off += to_copy;
140 if (map_off == PAGE_SI
141 map_off = 0;
142 map_page++;
143 }
144 }
145 }
146
147 rds_cong_map_updated(map, ~(u64) 0);
148 }
149
150 struct rds_tcp_desc_arg {
151 struct rds_conn_path *conn_path;
152 gfp_t gfp;
153 };
154
155 static int rds_tcp_data_recv(read_descriptor_t
156 unsigned int offs
157 {
158 struct rds_tcp_desc_arg *arg = desc->a
159 struct rds_conn_path *cp = arg->conn_p
160 struct rds_tcp_connection *tc = cp->cp
161 struct rds_tcp_incoming *tinc = tc->t_
162 struct sk_buff *clone;
163 size_t left = len, to_copy;
164
165 rdsdebug("tcp data tc %p skb %p offset
166 len);
167
168 /*
169 * tcp_read_sock() interprets partial
170 * processing.
171 */
172 while (left) {
173 if (!tinc) {
174 tinc = kmem_cache_allo
175
176 if (!tinc) {
177 desc->error =
178 goto out;
179 }
180 tc->t_tinc = tinc;
181 rdsdebug("allocated ti
182 rds_inc_path_init(&tin
183 &cp-
184 tinc->ti_inc.i_rx_lat_
185 local_
186
187 /*
188 * XXX * we might be a
189 * we've already seria
190 */
191 skb_queue_head_init(&t
192 }
193
194 if (left && tc->t_tinc_hdr_rem
195 to_copy = min(tc->t_ti
196 rdsdebug("copying %zu
197 skb);
198 skb_copy_bits(skb, off
199 (char *)
200
201
202 to_copy)
203 tc->t_tinc_hdr_rem -=
204 left -= to_copy;
205 offset += to_copy;
206
207 if (tc->t_tinc_hdr_rem
208 /* could be 0
209 tc->t_tinc_dat
210 be32_t
211 tinc->ti_inc.i
212 local_
213 }
214 }
215
216 if (left && tc->t_tinc_data_re
217 to_copy = min(tc->t_ti
218
219 clone = pskb_extract(s
220 if (!clone) {
221 desc->error =
222 goto out;
223 }
224
225 skb_queue_tail(&tinc->
226
227 rdsdebug("skb %p data
228 "clone %p dat
229 skb, skb->dat
230 clone, clone-
231
232 tc->t_tinc_data_rem -=
233 left -= to_copy;
234 offset += to_copy;
235 }
236
237 if (tc->t_tinc_hdr_rem == 0 &&
238 struct rds_connection
239
240 if (tinc->ti_inc.i_hdr
241 rds_tcp_cong_r
242 else
243 rds_recv_incom
244
245
246
247
248 tc->t_tinc_hdr_rem = s
249 tc->t_tinc_data_rem =
250 tc->t_tinc = NULL;
251 rds_inc_put(&tinc->ti_
252 tinc = NULL;
253 }
254 }
255 out:
256 rdsdebug("returning len %zu left %zu s
257 len, left, skb->len,
258 skb_queue_len(&tc->t_sock->sk
259 return len - left;
260 }
261
262 /* the caller has to hold the sock lock */
263 static int rds_tcp_read_sock(struct rds_conn_p
264 {
265 struct rds_tcp_connection *tc = cp->cp
266 struct socket *sock = tc->t_sock;
267 read_descriptor_t desc;
268 struct rds_tcp_desc_arg arg;
269
270 /* It's like glib in the kernel! */
271 arg.conn_path = cp;
272 arg.gfp = gfp;
273 desc.arg.data = &arg;
274 desc.error = 0;
275 desc.count = 1; /* give more than one
276
277 tcp_read_sock(sock->sk, &desc, rds_tcp
278 rdsdebug("tcp_read_sock for tc %p gfp
279 desc.error);
280
281 return desc.error;
282 }
283
284 /*
285 * We hold the sock lock to serialize our rds_
286 * data_ready.
287 *
288 * if we fail to allocate we're in trouble.. b
289 * trying again to see if the VM can free up s
290 */
291 int rds_tcp_recv_path(struct rds_conn_path *cp
292 {
293 struct rds_tcp_connection *tc = cp->cp
294 struct socket *sock = tc->t_sock;
295 int ret = 0;
296
297 rdsdebug("recv worker path [%d] tc %p
298 cp->cp_index, tc, sock);
299
300 lock_sock(sock->sk);
301 ret = rds_tcp_read_sock(cp, GFP_KERNEL
302 release_sock(sock->sk);
303
304 return ret;
305 }
306
307 void rds_tcp_data_ready(struct sock *sk)
308 {
309 void (*ready)(struct sock *sk);
310 struct rds_conn_path *cp;
311 struct rds_tcp_connection *tc;
312
313 trace_sk_data_ready(sk);
314 rdsdebug("data ready sk %p\n", sk);
315
316 read_lock_bh(&sk->sk_callback_lock);
317 cp = sk->sk_user_data;
318 if (!cp) { /* check for teardown race
319 ready = sk->sk_data_ready;
320 goto out;
321 }
322
323 tc = cp->cp_transport_data;
324 ready = tc->t_orig_data_ready;
325 rds_tcp_stats_inc(s_tcp_data_ready_cal
326
327 if (rds_tcp_read_sock(cp, GFP_ATOMIC)
328 rcu_read_lock();
329 if (!rds_destroy_pending(cp->c
330 queue_delayed_work(rds
331 rcu_read_unlock();
332 }
333 out:
334 read_unlock_bh(&sk->sk_callback_lock);
335 ready(sk);
336 }
337
338 int rds_tcp_recv_init(void)
339 {
340 rds_tcp_incoming_slab = KMEM_CACHE(rds
341 if (!rds_tcp_incoming_slab)
342 return -ENOMEM;
343 return 0;
344 }
345
346 void rds_tcp_recv_exit(void)
347 {
348 kmem_cache_destroy(rds_tcp_incoming_sl
349 }
350
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.