RDS: fold rdma.h into rds.h
[linux-3.10.git] / 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 #include <linux/slab.h>
35
36 #include "rds.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 EXPORT_SYMBOL_GPL(rds_message_addref);
54
55 /*
56  * This relies on dma_map_sg() not touching sg[].page during merging.
57  */
58 static void rds_message_purge(struct rds_message *rm)
59 {
60         unsigned long i;
61
62         if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags)))
63                 return;
64
65         for (i = 0; i < rm->data.m_nents; i++) {
66                 rdsdebug("putting data page %p\n", (void *)sg_page(&rm->data.m_sg[i]));
67                 /* XXX will have to put_page for page refs */
68                 __free_page(sg_page(&rm->data.m_sg[i]));
69         }
70         rm->data.m_nents = 0;
71
72         if (rm->rdma.m_rdma_op)
73                 rds_rdma_free_op(rm->rdma.m_rdma_op);
74         if (rm->rdma.m_rdma_mr)
75                 rds_mr_put(rm->rdma.m_rdma_mr);
76 }
77
78 void rds_message_inc_purge(struct rds_incoming *inc)
79 {
80         struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
81         rds_message_purge(rm);
82 }
83
84 void rds_message_put(struct rds_message *rm)
85 {
86         rdsdebug("put rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
87
88         if (atomic_dec_and_test(&rm->m_refcount)) {
89                 BUG_ON(!list_empty(&rm->m_sock_item));
90                 BUG_ON(!list_empty(&rm->m_conn_item));
91                 rds_message_purge(rm);
92
93                 kfree(rm);
94         }
95 }
96 EXPORT_SYMBOL_GPL(rds_message_put);
97
98 void rds_message_inc_free(struct rds_incoming *inc)
99 {
100         struct rds_message *rm = container_of(inc, struct rds_message, m_inc);
101         rds_message_put(rm);
102 }
103
104 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
105                                  __be16 dport, u64 seq)
106 {
107         hdr->h_flags = 0;
108         hdr->h_sport = sport;
109         hdr->h_dport = dport;
110         hdr->h_sequence = cpu_to_be64(seq);
111         hdr->h_exthdr[0] = RDS_EXTHDR_NONE;
112 }
113 EXPORT_SYMBOL_GPL(rds_message_populate_header);
114
115 int rds_message_add_extension(struct rds_header *hdr,
116                 unsigned int type, const void *data, unsigned int len)
117 {
118         unsigned int ext_len = sizeof(u8) + len;
119         unsigned char *dst;
120
121         /* For now, refuse to add more than one extension header */
122         if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE)
123                 return 0;
124
125         if (type >= __RDS_EXTHDR_MAX || len != rds_exthdr_size[type])
126                 return 0;
127
128         if (ext_len >= RDS_HEADER_EXT_SPACE)
129                 return 0;
130         dst = hdr->h_exthdr;
131
132         *dst++ = type;
133         memcpy(dst, data, len);
134
135         dst[len] = RDS_EXTHDR_NONE;
136         return 1;
137 }
138 EXPORT_SYMBOL_GPL(rds_message_add_extension);
139
140 /*
141  * If a message has extension headers, retrieve them here.
142  * Call like this:
143  *
144  * unsigned int pos = 0;
145  *
146  * while (1) {
147  *      buflen = sizeof(buffer);
148  *      type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
149  *      if (type == RDS_EXTHDR_NONE)
150  *              break;
151  *      ...
152  * }
153  */
154 int rds_message_next_extension(struct rds_header *hdr,
155                 unsigned int *pos, void *buf, unsigned int *buflen)
156 {
157         unsigned int offset, ext_type, ext_len;
158         u8 *src = hdr->h_exthdr;
159
160         offset = *pos;
161         if (offset >= RDS_HEADER_EXT_SPACE)
162                 goto none;
163
164         /* Get the extension type and length. For now, the
165          * length is implied by the extension type. */
166         ext_type = src[offset++];
167
168         if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
169                 goto none;
170         ext_len = rds_exthdr_size[ext_type];
171         if (offset + ext_len > RDS_HEADER_EXT_SPACE)
172                 goto none;
173
174         *pos = offset + ext_len;
175         if (ext_len < *buflen)
176                 *buflen = ext_len;
177         memcpy(buf, src + offset, *buflen);
178         return ext_type;
179
180 none:
181         *pos = RDS_HEADER_EXT_SPACE;
182         *buflen = 0;
183         return RDS_EXTHDR_NONE;
184 }
185
186 int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version)
187 {
188         struct rds_ext_header_version ext_hdr;
189
190         ext_hdr.h_version = cpu_to_be32(version);
191         return rds_message_add_extension(hdr, RDS_EXTHDR_VERSION, &ext_hdr, sizeof(ext_hdr));
192 }
193
194 int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version)
195 {
196         struct rds_ext_header_version ext_hdr;
197         unsigned int pos = 0, len = sizeof(ext_hdr);
198
199         /* We assume the version extension is the only one present */
200         if (rds_message_next_extension(hdr, &pos, &ext_hdr, &len) != RDS_EXTHDR_VERSION)
201                 return 0;
202         *version = be32_to_cpu(ext_hdr.h_version);
203         return 1;
204 }
205
206 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
207 {
208         struct rds_ext_header_rdma_dest ext_hdr;
209
210         ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
211         ext_hdr.h_rdma_offset = cpu_to_be32(offset);
212         return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr));
213 }
214 EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
215
216 /*
217  * Each rds_message is allocated with extra space for the scatterlist entries
218  * rds ops will need. This is to minimize memory allocation count. Then, each rds op
219  * can grab SGs when initializing its part of the rds_message.
220  */
221 struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
222 {
223         struct rds_message *rm;
224
225         rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
226         if (!rm)
227                 goto out;
228
229         rm->m_used_sgs = 0;
230         rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
231
232         atomic_set(&rm->m_refcount, 1);
233         INIT_LIST_HEAD(&rm->m_sock_item);
234         INIT_LIST_HEAD(&rm->m_conn_item);
235         spin_lock_init(&rm->m_rs_lock);
236
237 out:
238         return rm;
239 }
240
241 /*
242  * RDS ops use this to grab SG entries from the rm's sg pool.
243  */
244 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
245 {
246         struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
247         struct scatterlist *sg_ret;
248
249         WARN_ON(rm->m_used_sgs + nents > rm->m_total_sgs);
250
251         sg_ret = &sg_first[rm->m_used_sgs];
252
253         rm->m_used_sgs += nents;
254
255         return sg_ret;
256 }
257
258 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
259 {
260         struct rds_message *rm;
261         unsigned int i;
262
263         rm = rds_message_alloc(ceil(total_len, PAGE_SIZE), GFP_KERNEL);
264         if (!rm)
265                 return ERR_PTR(-ENOMEM);
266
267         set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
268         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
269         rm->data.m_nents = ceil(total_len, PAGE_SIZE);
270
271         for (i = 0; i < rm->data.m_nents; ++i) {
272                 sg_set_page(&rm->data.m_sg[i],
273                                 virt_to_page(page_addrs[i]),
274                                 PAGE_SIZE, 0);
275         }
276
277         return rm;
278 }
279
280 int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
281                                                size_t total_len)
282 {
283         unsigned long to_copy;
284         unsigned long iov_off;
285         unsigned long sg_off;
286         struct iovec *iov;
287         struct scatterlist *sg;
288         int ret = 0;
289
290         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
291
292         /*
293          * now allocate and copy in the data payload.
294          */
295         sg = rm->data.m_sg;
296         iov = first_iov;
297         iov_off = 0;
298         sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
299
300         while (total_len) {
301                 if (!sg_page(sg)) {
302                         ret = rds_page_remainder_alloc(sg, total_len,
303                                                        GFP_HIGHUSER);
304                         if (ret)
305                                 goto out;
306                         rm->data.m_nents++;
307                         sg_off = 0;
308                 }
309
310                 while (iov_off == iov->iov_len) {
311                         iov_off = 0;
312                         iov++;
313                 }
314
315                 to_copy = min(iov->iov_len - iov_off, sg->length - sg_off);
316                 to_copy = min_t(size_t, to_copy, total_len);
317
318                 rdsdebug("copying %lu bytes from user iov [%p, %zu] + %lu to "
319                          "sg [%p, %u, %u] + %lu\n",
320                          to_copy, iov->iov_base, iov->iov_len, iov_off,
321                          (void *)sg_page(sg), sg->offset, sg->length, sg_off);
322
323                 ret = rds_page_copy_from_user(sg_page(sg), sg->offset + sg_off,
324                                               iov->iov_base + iov_off,
325                                               to_copy);
326                 if (ret)
327                         goto out;
328
329                 iov_off += to_copy;
330                 total_len -= to_copy;
331                 sg_off += to_copy;
332
333                 if (sg_off == sg->length)
334                         sg++;
335         }
336
337 out:
338         return ret;
339 }
340
341 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
342                                  struct iovec *first_iov, size_t size)
343 {
344         struct rds_message *rm;
345         struct iovec *iov;
346         struct scatterlist *sg;
347         unsigned long to_copy;
348         unsigned long iov_off;
349         unsigned long vec_off;
350         int copied;
351         int ret;
352         u32 len;
353
354         rm = container_of(inc, struct rds_message, m_inc);
355         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
356
357         iov = first_iov;
358         iov_off = 0;
359         sg = rm->data.m_sg;
360         vec_off = 0;
361         copied = 0;
362
363         while (copied < size && copied < len) {
364                 while (iov_off == iov->iov_len) {
365                         iov_off = 0;
366                         iov++;
367                 }
368
369                 to_copy = min(iov->iov_len - iov_off, sg->length - vec_off);
370                 to_copy = min_t(size_t, to_copy, size - copied);
371                 to_copy = min_t(unsigned long, to_copy, len - copied);
372
373                 rdsdebug("copying %lu bytes to user iov [%p, %zu] + %lu to "
374                          "sg [%p, %u, %u] + %lu\n",
375                          to_copy, iov->iov_base, iov->iov_len, iov_off,
376                          sg_page(sg), sg->offset, sg->length, vec_off);
377
378                 ret = rds_page_copy_to_user(sg_page(sg), sg->offset + vec_off,
379                                             iov->iov_base + iov_off,
380                                             to_copy);
381                 if (ret) {
382                         copied = ret;
383                         break;
384                 }
385
386                 iov_off += to_copy;
387                 vec_off += to_copy;
388                 copied += to_copy;
389
390                 if (vec_off == sg->length) {
391                         vec_off = 0;
392                         sg++;
393                 }
394         }
395
396         return copied;
397 }
398
399 /*
400  * If the message is still on the send queue, wait until the transport
401  * is done with it. This is particularly important for RDMA operations.
402  */
403 void rds_message_wait(struct rds_message *rm)
404 {
405         wait_event(rds_message_flush_waitq,
406                         !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
407 }
408
409 void rds_message_unmapped(struct rds_message *rm)
410 {
411         clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
412         if (waitqueue_active(&rds_message_flush_waitq))
413                 wake_up(&rds_message_flush_waitq);
414 }
415 EXPORT_SYMBOL_GPL(rds_message_unmapped);
416