RDS/IB: get the xmit max_sge from the RDS IB device on the connection
[linux-2.6.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 unsigned int     rds_exthdr_size[__RDS_EXTHDR_MAX] = {
39 [RDS_EXTHDR_NONE]       = 0,
40 [RDS_EXTHDR_VERSION]    = sizeof(struct rds_ext_header_version),
41 [RDS_EXTHDR_RDMA]       = sizeof(struct rds_ext_header_rdma),
42 [RDS_EXTHDR_RDMA_DEST]  = sizeof(struct rds_ext_header_rdma_dest),
43 };
44
45
46 void rds_message_addref(struct rds_message *rm)
47 {
48         rdsdebug("addref rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
49         atomic_inc(&rm->m_refcount);
50 }
51 EXPORT_SYMBOL_GPL(rds_message_addref);
52
53 /*
54  * This relies on dma_map_sg() not touching sg[].page during merging.
55  */
56 static void rds_message_purge(struct rds_message *rm)
57 {
58         unsigned long i;
59
60         if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags)))
61                 return;
62
63         for (i = 0; i < rm->data.op_nents; i++) {
64                 rdsdebug("putting data page %p\n", (void *)sg_page(&rm->data.op_sg[i]));
65                 /* XXX will have to put_page for page refs */
66                 __free_page(sg_page(&rm->data.op_sg[i]));
67         }
68         rm->data.op_nents = 0;
69
70         if (rm->rdma.op_active)
71                 rds_rdma_free_op(&rm->rdma);
72         if (rm->rdma.op_rdma_mr)
73                 rds_mr_put(rm->rdma.op_rdma_mr);
74
75         if (rm->atomic.op_active)
76                 rds_atomic_free_op(&rm->atomic);
77         if (rm->atomic.op_rdma_mr)
78                 rds_mr_put(rm->atomic.op_rdma_mr);
79 }
80
81 void rds_message_put(struct rds_message *rm)
82 {
83         rdsdebug("put rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
84         if (atomic_read(&rm->m_refcount) == 0) {
85 printk(KERN_CRIT "danger refcount zero on %p\n", rm);
86 WARN_ON(1);
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_populate_header(struct rds_header *hdr, __be16 sport,
99                                  __be16 dport, u64 seq)
100 {
101         hdr->h_flags = 0;
102         hdr->h_sport = sport;
103         hdr->h_dport = dport;
104         hdr->h_sequence = cpu_to_be64(seq);
105         hdr->h_exthdr[0] = RDS_EXTHDR_NONE;
106 }
107 EXPORT_SYMBOL_GPL(rds_message_populate_header);
108
109 int rds_message_add_extension(struct rds_header *hdr,
110                 unsigned int type, const void *data, unsigned int len)
111 {
112         unsigned int ext_len = sizeof(u8) + len;
113         unsigned char *dst;
114
115         /* For now, refuse to add more than one extension header */
116         if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE)
117                 return 0;
118
119         if (type >= __RDS_EXTHDR_MAX || len != rds_exthdr_size[type])
120                 return 0;
121
122         if (ext_len >= RDS_HEADER_EXT_SPACE)
123                 return 0;
124         dst = hdr->h_exthdr;
125
126         *dst++ = type;
127         memcpy(dst, data, len);
128
129         dst[len] = RDS_EXTHDR_NONE;
130         return 1;
131 }
132 EXPORT_SYMBOL_GPL(rds_message_add_extension);
133
134 /*
135  * If a message has extension headers, retrieve them here.
136  * Call like this:
137  *
138  * unsigned int pos = 0;
139  *
140  * while (1) {
141  *      buflen = sizeof(buffer);
142  *      type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
143  *      if (type == RDS_EXTHDR_NONE)
144  *              break;
145  *      ...
146  * }
147  */
148 int rds_message_next_extension(struct rds_header *hdr,
149                 unsigned int *pos, void *buf, unsigned int *buflen)
150 {
151         unsigned int offset, ext_type, ext_len;
152         u8 *src = hdr->h_exthdr;
153
154         offset = *pos;
155         if (offset >= RDS_HEADER_EXT_SPACE)
156                 goto none;
157
158         /* Get the extension type and length. For now, the
159          * length is implied by the extension type. */
160         ext_type = src[offset++];
161
162         if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
163                 goto none;
164         ext_len = rds_exthdr_size[ext_type];
165         if (offset + ext_len > RDS_HEADER_EXT_SPACE)
166                 goto none;
167
168         *pos = offset + ext_len;
169         if (ext_len < *buflen)
170                 *buflen = ext_len;
171         memcpy(buf, src + offset, *buflen);
172         return ext_type;
173
174 none:
175         *pos = RDS_HEADER_EXT_SPACE;
176         *buflen = 0;
177         return RDS_EXTHDR_NONE;
178 }
179
180 int rds_message_add_version_extension(struct rds_header *hdr, unsigned int version)
181 {
182         struct rds_ext_header_version ext_hdr;
183
184         ext_hdr.h_version = cpu_to_be32(version);
185         return rds_message_add_extension(hdr, RDS_EXTHDR_VERSION, &ext_hdr, sizeof(ext_hdr));
186 }
187
188 int rds_message_get_version_extension(struct rds_header *hdr, unsigned int *version)
189 {
190         struct rds_ext_header_version ext_hdr;
191         unsigned int pos = 0, len = sizeof(ext_hdr);
192
193         /* We assume the version extension is the only one present */
194         if (rds_message_next_extension(hdr, &pos, &ext_hdr, &len) != RDS_EXTHDR_VERSION)
195                 return 0;
196         *version = be32_to_cpu(ext_hdr.h_version);
197         return 1;
198 }
199
200 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
201 {
202         struct rds_ext_header_rdma_dest ext_hdr;
203
204         ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
205         ext_hdr.h_rdma_offset = cpu_to_be32(offset);
206         return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr));
207 }
208 EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
209
210 /*
211  * Each rds_message is allocated with extra space for the scatterlist entries
212  * rds ops will need. This is to minimize memory allocation count. Then, each rds op
213  * can grab SGs when initializing its part of the rds_message.
214  */
215 struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
216 {
217         struct rds_message *rm;
218
219         rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
220         if (!rm)
221                 goto out;
222
223         rm->m_used_sgs = 0;
224         rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
225
226         atomic_set(&rm->m_refcount, 1);
227         INIT_LIST_HEAD(&rm->m_sock_item);
228         INIT_LIST_HEAD(&rm->m_conn_item);
229         spin_lock_init(&rm->m_rs_lock);
230         init_waitqueue_head(&rm->m_flush_wait);
231
232 out:
233         return rm;
234 }
235
236 /*
237  * RDS ops use this to grab SG entries from the rm's sg pool.
238  */
239 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
240 {
241         struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
242         struct scatterlist *sg_ret;
243
244         WARN_ON(rm->m_used_sgs + nents > rm->m_total_sgs);
245         WARN_ON(!nents);
246
247         sg_ret = &sg_first[rm->m_used_sgs];
248         sg_init_table(sg_ret, nents);
249         rm->m_used_sgs += nents;
250
251         return sg_ret;
252 }
253
254 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
255 {
256         struct rds_message *rm;
257         unsigned int i;
258         int num_sgs = ceil(total_len, PAGE_SIZE);
259         int extra_bytes = num_sgs * sizeof(struct scatterlist);
260
261         rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
262         if (!rm)
263                 return ERR_PTR(-ENOMEM);
264
265         set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
266         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
267         rm->data.op_nents = ceil(total_len, PAGE_SIZE);
268         rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
269
270         for (i = 0; i < rm->data.op_nents; ++i) {
271                 sg_set_page(&rm->data.op_sg[i],
272                                 virt_to_page(page_addrs[i]),
273                                 PAGE_SIZE, 0);
274         }
275
276         return rm;
277 }
278
279 int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
280                                                size_t total_len)
281 {
282         unsigned long to_copy;
283         unsigned long iov_off;
284         unsigned long sg_off;
285         struct iovec *iov;
286         struct scatterlist *sg;
287         int ret = 0;
288
289         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
290
291         /*
292          * now allocate and copy in the data payload.
293          */
294         sg = rm->data.op_sg;
295         iov = first_iov;
296         iov_off = 0;
297         sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
298
299         while (total_len) {
300                 if (!sg_page(sg)) {
301                         ret = rds_page_remainder_alloc(sg, total_len,
302                                                        GFP_HIGHUSER);
303                         if (ret)
304                                 goto out;
305                         rm->data.op_nents++;
306                         sg_off = 0;
307                 }
308
309                 while (iov_off == iov->iov_len) {
310                         iov_off = 0;
311                         iov++;
312                 }
313
314                 to_copy = min(iov->iov_len - iov_off, sg->length - sg_off);
315                 to_copy = min_t(size_t, to_copy, total_len);
316
317                 rdsdebug("copying %lu bytes from user iov [%p, %zu] + %lu to "
318                          "sg [%p, %u, %u] + %lu\n",
319                          to_copy, iov->iov_base, iov->iov_len, iov_off,
320                          (void *)sg_page(sg), sg->offset, sg->length, sg_off);
321
322                 ret = rds_page_copy_from_user(sg_page(sg), sg->offset + sg_off,
323                                               iov->iov_base + iov_off,
324                                               to_copy);
325                 if (ret)
326                         goto out;
327
328                 iov_off += to_copy;
329                 total_len -= to_copy;
330                 sg_off += to_copy;
331
332                 if (sg_off == sg->length)
333                         sg++;
334         }
335
336 out:
337         return ret;
338 }
339
340 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
341                                  struct iovec *first_iov, size_t size)
342 {
343         struct rds_message *rm;
344         struct iovec *iov;
345         struct scatterlist *sg;
346         unsigned long to_copy;
347         unsigned long iov_off;
348         unsigned long vec_off;
349         int copied;
350         int ret;
351         u32 len;
352
353         rm = container_of(inc, struct rds_message, m_inc);
354         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
355
356         iov = first_iov;
357         iov_off = 0;
358         sg = rm->data.op_sg;
359         vec_off = 0;
360         copied = 0;
361
362         while (copied < size && copied < len) {
363                 while (iov_off == iov->iov_len) {
364                         iov_off = 0;
365                         iov++;
366                 }
367
368                 to_copy = min(iov->iov_len - iov_off, sg->length - vec_off);
369                 to_copy = min_t(size_t, to_copy, size - copied);
370                 to_copy = min_t(unsigned long, to_copy, len - copied);
371
372                 rdsdebug("copying %lu bytes to user iov [%p, %zu] + %lu to "
373                          "sg [%p, %u, %u] + %lu\n",
374                          to_copy, iov->iov_base, iov->iov_len, iov_off,
375                          sg_page(sg), sg->offset, sg->length, vec_off);
376
377                 ret = rds_page_copy_to_user(sg_page(sg), sg->offset + vec_off,
378                                             iov->iov_base + iov_off,
379                                             to_copy);
380                 if (ret) {
381                         copied = ret;
382                         break;
383                 }
384
385                 iov_off += to_copy;
386                 vec_off += to_copy;
387                 copied += to_copy;
388
389                 if (vec_off == sg->length) {
390                         vec_off = 0;
391                         sg++;
392                 }
393         }
394
395         return copied;
396 }
397
398 /*
399  * If the message is still on the send queue, wait until the transport
400  * is done with it. This is particularly important for RDMA operations.
401  */
402 void rds_message_wait(struct rds_message *rm)
403 {
404         wait_event_interruptible(rm->m_flush_wait,
405                         !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
406 }
407
408 void rds_message_unmapped(struct rds_message *rm)
409 {
410         clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
411         wake_up_interruptible(&rm->m_flush_wait);
412 }
413 EXPORT_SYMBOL_GPL(rds_message_unmapped);
414