Merge branch 'for-2.6.30' into for-2.6.31
[linux-2.6] / net / rds / message.c
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
35 #include "rds.h"
36 #include "rdma.h"
37
38 static DECLARE_WAIT_QUEUE_HEAD(rds_message_flush_waitq);
39
40 static unsigned int     rds_exthdr_size[__RDS_EXTHDR_MAX] = {
41 [RDS_EXTHDR_NONE]       = 0,
42 [RDS_EXTHDR_VERSION]    = sizeof(struct rds_ext_header_version),
43 [RDS_EXTHDR_RDMA]       = sizeof(struct rds_ext_header_rdma),
44 [RDS_EXTHDR_RDMA_DEST]  = sizeof(struct rds_ext_header_rdma_dest),
45 };
46
47
48 void rds_message_addref(struct rds_message *rm)
49 {
50         rdsdebug("addref rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
51         atomic_inc(&rm->m_refcount);
52 }
53
54 /*
55  * This relies on dma_map_sg() not touching sg[].page during merging.
56  */
57 static void rds_message_purge(struct rds_message *rm)
58 {
59         unsigned long i;
60
61         if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags)))
62                 return;
63
64         for (i = 0; i < rm->m_nents; i++) {
65                 rdsdebug("putting data page %p\n", (void *)sg_page(&rm->m_sg[i]));
66                 /* XXX will have to put_page for page refs */
67                 __free_page(sg_page(&rm->m_sg[i]));
68         }
69         rm->m_nents = 0;
70
71         if (rm->m_rdma_op)
72                 rds_rdma_free_op(rm->m_rdma_op);
73         if (rm->m_rdma_mr)
74                 rds_mr_put(rm->m_rdma_mr);
75 }
76
77 void rds_message_inc_purge(struct rds_incoming *inc)
78 {
79         struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
80         rds_message_purge(rm);
81 }
82
83 void rds_message_put(struct rds_message *rm)
84 {
85         rdsdebug("put rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
86
87         if (atomic_dec_and_test(&rm->m_refcount)) {
88                 BUG_ON(!list_empty(&rm->m_sock_item));
89                 BUG_ON(!list_empty(&rm->m_conn_item));
90                 rds_message_purge(rm);
91
92                 kfree(rm);
93         }
94 }
95
96 void rds_message_inc_free(struct rds_incoming *inc)
97 {
98         struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
99         rds_message_put(rm);
100 }
101
102 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
103                                  __be16 dport, u64 seq)
104 {
105         hdr->h_flags = 0;
106         hdr->h_sport = sport;
107         hdr->h_dport = dport;
108         hdr->h_sequence = cpu_to_be64(seq);
109         hdr->h_exthdr[0] = RDS_EXTHDR_NONE;
110 }
111
112 int rds_message_add_extension(struct rds_header *hdr,
113                 unsigned int type, const void *data, unsigned int len)
114 {
115         unsigned int ext_len = sizeof(u8) + len;
116         unsigned char *dst;
117
118         /* For now, refuse to add more than one extension header */
119         if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE)
120                 return 0;
121
122         if (type >= __RDS_EXTHDR_MAX
123          || len != rds_exthdr_size[type])
124                 return 0;
125
126         if (ext_len >= RDS_HEADER_EXT_SPACE)
127                 return 0;
128         dst = hdr->h_exthdr;
129
130         *dst++ = type;
131         memcpy(dst, data, len);
132
133         dst[len] = RDS_EXTHDR_NONE;
134         return 1;
135 }
136
137 /*
138  * If a message has extension headers, retrieve them here.
139  * Call like this:
140  *
141  * unsigned int pos = 0;
142  *
143  * while (1) {
144  *      buflen = sizeof(buffer);
145  *      type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
146  *      if (type == RDS_EXTHDR_NONE)
147  *              break;
148  *      ...
149  * }
150  */
151 int rds_message_next_extension(struct rds_header *hdr,
152                 unsigned int *pos, void *buf, unsigned int *buflen)
153 {
154         unsigned int offset, ext_type, ext_len;
155         u8 *src = hdr->h_exthdr;
156
157         offset = *pos;
158         if (offset >= RDS_HEADER_EXT_SPACE)
159                 goto none;
160
161         /* Get the extension type and length. For now, the
162          * length is implied by the extension type. */
163         ext_type = src[offset++];
164
165         if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
166                 goto none;
167         ext_len = rds_exthdr_size[ext_type];
168         if (offset + ext_len > RDS_HEADER_EXT_SPACE)
169                 goto none;
170
171         *pos = offset + ext_len;
172         if (ext_len < *buflen)
173                 *buflen = ext_len;
174         memcpy(buf, src + offset, *buflen);
175         return ext_type;
176
177 none:
178         *pos = RDS_HEADER_EXT_SPACE;
179         *buflen = 0;
180         return RDS_EXTHDR_NONE;
181 }
182
183 int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version)
184 {
185         struct rds_ext_header_version ext_hdr;
186
187         ext_hdr.h_version = cpu_to_be32(version);
188         return rds_message_add_extension(hdr, RDS_EXTHDR_VERSION, &ext_hdr, sizeof(ext_hdr));
189 }
190
191 int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version)
192 {
193         struct rds_ext_header_version ext_hdr;
194         unsigned int pos = 0, len = sizeof(ext_hdr);
195
196         /* We assume the version extension is the only one present */
197         if (rds_message_next_extension(hdr, &pos, &ext_hdr, &len) != RDS_EXTHDR_VERSION)
198                 return 0;
199         *version = be32_to_cpu(ext_hdr.h_version);
200         return 1;
201 }
202
203 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
204 {
205         struct rds_ext_header_rdma_dest ext_hdr;
206
207         ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
208         ext_hdr.h_rdma_offset = cpu_to_be32(offset);
209         return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr));
210 }
211
212 struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp)
213 {
214         struct rds_message *rm;
215
216         rm = kzalloc(sizeof(struct rds_message) +
217                      (nents * sizeof(struct scatterlist)), gfp);
218         if (!rm)
219                 goto out;
220
221         if (nents)
222                 sg_init_table(rm->m_sg, nents);
223         atomic_set(&rm->m_refcount, 1);
224         INIT_LIST_HEAD(&rm->m_sock_item);
225         INIT_LIST_HEAD(&rm->m_conn_item);
226         spin_lock_init(&rm->m_rs_lock);
227
228 out:
229         return rm;
230 }
231
232 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
233 {
234         struct rds_message *rm;
235         unsigned int i;
236
237         rm = rds_message_alloc(ceil(total_len, PAGE_SIZE), GFP_KERNEL);
238         if (rm == NULL)
239                 return ERR_PTR(-ENOMEM);
240
241         set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
242         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
243         rm->m_nents = ceil(total_len, PAGE_SIZE);
244
245         for (i = 0; i < rm->m_nents; ++i) {
246                 sg_set_page(&rm->m_sg[i],
247                                 virt_to_page(page_addrs[i]),
248                                 PAGE_SIZE, 0);
249         }
250
251         return rm;
252 }
253
254 struct rds_message *rds_message_copy_from_user(struct iovec *first_iov,
255                                                size_t total_len)
256 {
257         unsigned long to_copy;
258         unsigned long iov_off;
259         unsigned long sg_off;
260         struct rds_message *rm;
261         struct iovec *iov;
262         struct scatterlist *sg;
263         int ret;
264
265         rm = rds_message_alloc(ceil(total_len, PAGE_SIZE), GFP_KERNEL);
266         if (rm == NULL) {
267                 ret = -ENOMEM;
268                 goto out;
269         }
270
271         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
272
273         /*
274          * now allocate and copy in the data payload.
275          */
276         sg = rm->m_sg;
277         iov = first_iov;
278         iov_off = 0;
279         sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
280
281         while (total_len) {
282                 if (sg_page(sg) == NULL) {
283                         ret = rds_page_remainder_alloc(sg, total_len,
284                                                        GFP_HIGHUSER);
285                         if (ret)
286                                 goto out;
287                         rm->m_nents++;
288                         sg_off = 0;
289                 }
290
291                 while (iov_off == iov->iov_len) {
292                         iov_off = 0;
293                         iov++;
294                 }
295
296                 to_copy = min(iov->iov_len - iov_off, sg->length - sg_off);
297                 to_copy = min_t(size_t, to_copy, total_len);
298
299                 rdsdebug("copying %lu bytes from user iov [%p, %zu] + %lu to "
300                          "sg [%p, %u, %u] + %lu\n",
301                          to_copy, iov->iov_base, iov->iov_len, iov_off,
302                          (void *)sg_page(sg), sg->offset, sg->length, sg_off);
303
304                 ret = rds_page_copy_from_user(sg_page(sg), sg->offset + sg_off,
305                                               iov->iov_base + iov_off,
306                                               to_copy);
307                 if (ret)
308                         goto out;
309
310                 iov_off += to_copy;
311                 total_len -= to_copy;
312                 sg_off += to_copy;
313
314                 if (sg_off == sg->length)
315                         sg++;
316         }
317
318         ret = 0;
319 out:
320         if (ret) {
321                 if (rm)
322                         rds_message_put(rm);
323                 rm = ERR_PTR(ret);
324         }
325         return rm;
326 }
327
328 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
329                                  struct iovec *first_iov, size_t size)
330 {
331         struct rds_message *rm;
332         struct iovec *iov;
333         struct scatterlist *sg;
334         unsigned long to_copy;
335         unsigned long iov_off;
336         unsigned long vec_off;
337         int copied;
338         int ret;
339         u32 len;
340
341         rm = container_of(inc, struct rds_message, m_inc);
342         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
343
344         iov = first_iov;
345         iov_off = 0;
346         sg = rm->m_sg;
347         vec_off = 0;
348         copied = 0;
349
350         while (copied < size && copied < len) {
351                 while (iov_off == iov->iov_len) {
352                         iov_off = 0;
353                         iov++;
354                 }
355
356                 to_copy = min(iov->iov_len - iov_off, sg->length - vec_off);
357                 to_copy = min_t(size_t, to_copy, size - copied);
358                 to_copy = min_t(unsigned long, to_copy, len - copied);
359
360                 rdsdebug("copying %lu bytes to user iov [%p, %zu] + %lu to "
361                          "sg [%p, %u, %u] + %lu\n",
362                          to_copy, iov->iov_base, iov->iov_len, iov_off,
363                          sg_page(sg), sg->offset, sg->length, vec_off);
364
365                 ret = rds_page_copy_to_user(sg_page(sg), sg->offset + vec_off,
366                                             iov->iov_base + iov_off,
367                                             to_copy);
368                 if (ret) {
369                         copied = ret;
370                         break;
371                 }
372
373                 iov_off += to_copy;
374                 vec_off += to_copy;
375                 copied += to_copy;
376
377                 if (vec_off == sg->length) {
378                         vec_off = 0;
379                         sg++;
380                 }
381         }
382
383         return copied;
384 }
385
386 /*
387  * If the message is still on the send queue, wait until the transport
388  * is done with it. This is particularly important for RDMA operations.
389  */
390 void rds_message_wait(struct rds_message *rm)
391 {
392         wait_event(rds_message_flush_waitq,
393                         !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
394 }
395
396 void rds_message_unmapped(struct rds_message *rm)
397 {
398         clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
399         if (waitqueue_active(&rds_message_flush_waitq))
400                 wake_up(&rds_message_flush_waitq);
401 }
402