1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
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
40static struct kmem_cache *rds_tcp_incoming_slab;
41
42static 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
50void 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
61
62int 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
95 if (skb_copy_datagram_iovec(skb, skb_off, &tmp,
96 to_copy)) {
97 ret = -EFAULT;
98 goto out;
99 }
100
101 rds_stats_add(s_copy_to_user, to_copy);
102 size -= to_copy;
103 ret += to_copy;
104 skb_off += to_copy;
105 if (size == 0)
106 goto out;
107 }
108 }
109out:
110 return ret;
111}
112
113
114
115
116
117
118
119
120
121
122
123
124
125static 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
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
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
169struct rds_tcp_desc_arg {
170 struct rds_connection *conn;
171 gfp_t gfp;
172};
173
174static int rds_tcp_data_recv(read_descriptor_t *desc, struct sk_buff *skb,
175 unsigned int offset, size_t len)
176{
177 struct rds_tcp_desc_arg *arg = desc->arg.data;
178 struct rds_connection *conn = arg->conn;
179 struct rds_tcp_connection *tc = conn->c_transport_data;
180 struct rds_tcp_incoming *tinc = tc->t_tinc;
181 struct sk_buff *clone;
182 size_t left = len, to_copy;
183
184 rdsdebug("tcp data tc %p skb %p offset %u len %zu\n", tc, skb, offset,
185 len);
186
187
188
189
190
191 while (left) {
192 if (!tinc) {
193 tinc = kmem_cache_alloc(rds_tcp_incoming_slab,
194 arg->gfp);
195 if (!tinc) {
196 desc->error = -ENOMEM;
197 goto out;
198 }
199 tc->t_tinc = tinc;
200 rdsdebug("alloced tinc %p\n", tinc);
201 rds_inc_init(&tinc->ti_inc, conn, conn->c_faddr);
202
203
204
205
206 skb_queue_head_init(&tinc->ti_skb_list);
207 }
208
209 if (left && tc->t_tinc_hdr_rem) {
210 to_copy = min(tc->t_tinc_hdr_rem, left);
211 rdsdebug("copying %zu header from skb %p\n", to_copy,
212 skb);
213 skb_copy_bits(skb, offset,
214 (char *)&tinc->ti_inc.i_hdr +
215 sizeof(struct rds_header) -
216 tc->t_tinc_hdr_rem,
217 to_copy);
218 tc->t_tinc_hdr_rem -= to_copy;
219 left -= to_copy;
220 offset += to_copy;
221
222 if (tc->t_tinc_hdr_rem == 0) {
223
224 tc->t_tinc_data_rem =
225 be32_to_cpu(tinc->ti_inc.i_hdr.h_len);
226 }
227 }
228
229 if (left && tc->t_tinc_data_rem) {
230 clone = skb_clone(skb, arg->gfp);
231 if (!clone) {
232 desc->error = -ENOMEM;
233 goto out;
234 }
235
236 to_copy = min(tc->t_tinc_data_rem, left);
237 pskb_pull(clone, offset);
238 pskb_trim(clone, to_copy);
239 skb_queue_tail(&tinc->ti_skb_list, clone);
240
241 rdsdebug("skb %p data %p len %d off %u to_copy %zu -> "
242 "clone %p data %p len %d\n",
243 skb, skb->data, skb->len, offset, to_copy,
244 clone, clone->data, clone->len);
245
246 tc->t_tinc_data_rem -= to_copy;
247 left -= to_copy;
248 offset += to_copy;
249 }
250
251 if (tc->t_tinc_hdr_rem == 0 && tc->t_tinc_data_rem == 0) {
252 if (tinc->ti_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP)
253 rds_tcp_cong_recv(conn, tinc);
254 else
255 rds_recv_incoming(conn, conn->c_faddr,
256 conn->c_laddr, &tinc->ti_inc,
257 arg->gfp);
258
259 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
260 tc->t_tinc_data_rem = 0;
261 tc->t_tinc = NULL;
262 rds_inc_put(&tinc->ti_inc);
263 tinc = NULL;
264 }
265 }
266out:
267 rdsdebug("returning len %zu left %zu skb len %d rx queue depth %d\n",
268 len, left, skb->len,
269 skb_queue_len(&tc->t_sock->sk->sk_receive_queue));
270 return len - left;
271}
272
273
274static int rds_tcp_read_sock(struct rds_connection *conn, gfp_t gfp)
275{
276 struct rds_tcp_connection *tc = conn->c_transport_data;
277 struct socket *sock = tc->t_sock;
278 read_descriptor_t desc;
279 struct rds_tcp_desc_arg arg;
280
281
282 arg.conn = conn;
283 arg.gfp = gfp;
284 desc.arg.data = &arg;
285 desc.error = 0;
286 desc.count = 1;
287
288 tcp_read_sock(sock->sk, &desc, rds_tcp_data_recv);
289 rdsdebug("tcp_read_sock for tc %p gfp 0x%x returned %d\n", tc, gfp,
290 desc.error);
291
292 return desc.error;
293}
294
295
296
297
298
299
300
301
302int rds_tcp_recv(struct rds_connection *conn)
303{
304 struct rds_tcp_connection *tc = conn->c_transport_data;
305 struct socket *sock = tc->t_sock;
306 int ret = 0;
307
308 rdsdebug("recv worker conn %p tc %p sock %p\n", conn, tc, sock);
309
310 lock_sock(sock->sk);
311 ret = rds_tcp_read_sock(conn, GFP_KERNEL);
312 release_sock(sock->sk);
313
314 return ret;
315}
316
317void rds_tcp_data_ready(struct sock *sk, int bytes)
318{
319 void (*ready)(struct sock *sk, int bytes);
320 struct rds_connection *conn;
321 struct rds_tcp_connection *tc;
322
323 rdsdebug("data ready sk %p bytes %d\n", sk, bytes);
324
325 read_lock(&sk->sk_callback_lock);
326 conn = sk->sk_user_data;
327 if (!conn) {
328 ready = sk->sk_data_ready;
329 goto out;
330 }
331
332 tc = conn->c_transport_data;
333 ready = tc->t_orig_data_ready;
334 rds_tcp_stats_inc(s_tcp_data_ready_calls);
335
336 if (rds_tcp_read_sock(conn, GFP_ATOMIC) == -ENOMEM)
337 queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
338out:
339 read_unlock(&sk->sk_callback_lock);
340 ready(sk, bytes);
341}
342
343int rds_tcp_recv_init(void)
344{
345 rds_tcp_incoming_slab = kmem_cache_create("rds_tcp_incoming",
346 sizeof(struct rds_tcp_incoming),
347 0, 0, NULL);
348 if (!rds_tcp_incoming_slab)
349 return -ENOMEM;
350 return 0;
351}
352
353void rds_tcp_recv_exit(void)
354{
355 kmem_cache_destroy(rds_tcp_incoming_slab);
356}
357