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