Loading...
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#include <linux/export.h>
36
37#include "rds.h"
38
39static unsigned int rds_exthdr_size[__RDS_EXTHDR_MAX] = {
40[RDS_EXTHDR_NONE] = 0,
41[RDS_EXTHDR_VERSION] = sizeof(struct rds_ext_header_version),
42[RDS_EXTHDR_RDMA] = sizeof(struct rds_ext_header_rdma),
43[RDS_EXTHDR_RDMA_DEST] = sizeof(struct rds_ext_header_rdma_dest),
44};
45
46
47void rds_message_addref(struct rds_message *rm)
48{
49 rdsdebug("addref rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
50 atomic_inc(&rm->m_refcount);
51}
52EXPORT_SYMBOL_GPL(rds_message_addref);
53
54/*
55 * This relies on dma_map_sg() not touching sg[].page during merging.
56 */
57static 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->data.op_nents; i++) {
65 rdsdebug("putting data page %p\n", (void *)sg_page(&rm->data.op_sg[i]));
66 /* XXX will have to put_page for page refs */
67 __free_page(sg_page(&rm->data.op_sg[i]));
68 }
69 rm->data.op_nents = 0;
70
71 if (rm->rdma.op_active)
72 rds_rdma_free_op(&rm->rdma);
73 if (rm->rdma.op_rdma_mr)
74 rds_mr_put(rm->rdma.op_rdma_mr);
75
76 if (rm->atomic.op_active)
77 rds_atomic_free_op(&rm->atomic);
78 if (rm->atomic.op_rdma_mr)
79 rds_mr_put(rm->atomic.op_rdma_mr);
80}
81
82void rds_message_put(struct rds_message *rm)
83{
84 rdsdebug("put rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
85 if (atomic_read(&rm->m_refcount) == 0) {
86printk(KERN_CRIT "danger refcount zero on %p\n", rm);
87WARN_ON(1);
88 }
89 if (atomic_dec_and_test(&rm->m_refcount)) {
90 BUG_ON(!list_empty(&rm->m_sock_item));
91 BUG_ON(!list_empty(&rm->m_conn_item));
92 rds_message_purge(rm);
93
94 kfree(rm);
95 }
96}
97EXPORT_SYMBOL_GPL(rds_message_put);
98
99void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
100 __be16 dport, u64 seq)
101{
102 hdr->h_flags = 0;
103 hdr->h_sport = sport;
104 hdr->h_dport = dport;
105 hdr->h_sequence = cpu_to_be64(seq);
106 hdr->h_exthdr[0] = RDS_EXTHDR_NONE;
107}
108EXPORT_SYMBOL_GPL(rds_message_populate_header);
109
110int rds_message_add_extension(struct rds_header *hdr, unsigned int type,
111 const void *data, unsigned int len)
112{
113 unsigned int ext_len = sizeof(u8) + len;
114 unsigned char *dst;
115
116 /* For now, refuse to add more than one extension header */
117 if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE)
118 return 0;
119
120 if (type >= __RDS_EXTHDR_MAX || len != rds_exthdr_size[type])
121 return 0;
122
123 if (ext_len >= RDS_HEADER_EXT_SPACE)
124 return 0;
125 dst = hdr->h_exthdr;
126
127 *dst++ = type;
128 memcpy(dst, data, len);
129
130 dst[len] = RDS_EXTHDR_NONE;
131 return 1;
132}
133EXPORT_SYMBOL_GPL(rds_message_add_extension);
134
135/*
136 * If a message has extension headers, retrieve them here.
137 * Call like this:
138 *
139 * unsigned int pos = 0;
140 *
141 * while (1) {
142 * buflen = sizeof(buffer);
143 * type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
144 * if (type == RDS_EXTHDR_NONE)
145 * break;
146 * ...
147 * }
148 */
149int rds_message_next_extension(struct rds_header *hdr,
150 unsigned int *pos, void *buf, unsigned int *buflen)
151{
152 unsigned int offset, ext_type, ext_len;
153 u8 *src = hdr->h_exthdr;
154
155 offset = *pos;
156 if (offset >= RDS_HEADER_EXT_SPACE)
157 goto none;
158
159 /* Get the extension type and length. For now, the
160 * length is implied by the extension type. */
161 ext_type = src[offset++];
162
163 if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
164 goto none;
165 ext_len = rds_exthdr_size[ext_type];
166 if (offset + ext_len > RDS_HEADER_EXT_SPACE)
167 goto none;
168
169 *pos = offset + ext_len;
170 if (ext_len < *buflen)
171 *buflen = ext_len;
172 memcpy(buf, src + offset, *buflen);
173 return ext_type;
174
175none:
176 *pos = RDS_HEADER_EXT_SPACE;
177 *buflen = 0;
178 return RDS_EXTHDR_NONE;
179}
180
181int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
182{
183 struct rds_ext_header_rdma_dest ext_hdr;
184
185 ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
186 ext_hdr.h_rdma_offset = cpu_to_be32(offset);
187 return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr));
188}
189EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
190
191/*
192 * Each rds_message is allocated with extra space for the scatterlist entries
193 * rds ops will need. This is to minimize memory allocation count. Then, each rds op
194 * can grab SGs when initializing its part of the rds_message.
195 */
196struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
197{
198 struct rds_message *rm;
199
200 rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
201 if (!rm)
202 goto out;
203
204 rm->m_used_sgs = 0;
205 rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
206
207 atomic_set(&rm->m_refcount, 1);
208 INIT_LIST_HEAD(&rm->m_sock_item);
209 INIT_LIST_HEAD(&rm->m_conn_item);
210 spin_lock_init(&rm->m_rs_lock);
211 init_waitqueue_head(&rm->m_flush_wait);
212
213out:
214 return rm;
215}
216
217/*
218 * RDS ops use this to grab SG entries from the rm's sg pool.
219 */
220struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
221{
222 struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
223 struct scatterlist *sg_ret;
224
225 WARN_ON(rm->m_used_sgs + nents > rm->m_total_sgs);
226 WARN_ON(!nents);
227
228 if (rm->m_used_sgs + nents > rm->m_total_sgs)
229 return NULL;
230
231 sg_ret = &sg_first[rm->m_used_sgs];
232 sg_init_table(sg_ret, nents);
233 rm->m_used_sgs += nents;
234
235 return sg_ret;
236}
237
238struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
239{
240 struct rds_message *rm;
241 unsigned int i;
242 int num_sgs = ceil(total_len, PAGE_SIZE);
243 int extra_bytes = num_sgs * sizeof(struct scatterlist);
244
245 rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
246 if (!rm)
247 return ERR_PTR(-ENOMEM);
248
249 set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
250 rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
251 rm->data.op_nents = ceil(total_len, PAGE_SIZE);
252 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
253 if (!rm->data.op_sg) {
254 rds_message_put(rm);
255 return ERR_PTR(-ENOMEM);
256 }
257
258 for (i = 0; i < rm->data.op_nents; ++i) {
259 sg_set_page(&rm->data.op_sg[i],
260 virt_to_page(page_addrs[i]),
261 PAGE_SIZE, 0);
262 }
263
264 return rm;
265}
266
267int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
268 size_t total_len)
269{
270 unsigned long to_copy;
271 unsigned long iov_off;
272 unsigned long sg_off;
273 struct iovec *iov;
274 struct scatterlist *sg;
275 int ret = 0;
276
277 rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
278
279 /*
280 * now allocate and copy in the data payload.
281 */
282 sg = rm->data.op_sg;
283 iov = first_iov;
284 iov_off = 0;
285 sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
286
287 while (total_len) {
288 if (!sg_page(sg)) {
289 ret = rds_page_remainder_alloc(sg, total_len,
290 GFP_HIGHUSER);
291 if (ret)
292 goto out;
293 rm->data.op_nents++;
294 sg_off = 0;
295 }
296
297 while (iov_off == iov->iov_len) {
298 iov_off = 0;
299 iov++;
300 }
301
302 to_copy = min(iov->iov_len - iov_off, sg->length - sg_off);
303 to_copy = min_t(size_t, to_copy, total_len);
304
305 rdsdebug("copying %lu bytes from user iov [%p, %zu] + %lu to "
306 "sg [%p, %u, %u] + %lu\n",
307 to_copy, iov->iov_base, iov->iov_len, iov_off,
308 (void *)sg_page(sg), sg->offset, sg->length, sg_off);
309
310 ret = rds_page_copy_from_user(sg_page(sg), sg->offset + sg_off,
311 iov->iov_base + iov_off,
312 to_copy);
313 if (ret)
314 goto out;
315
316 iov_off += to_copy;
317 total_len -= to_copy;
318 sg_off += to_copy;
319
320 if (sg_off == sg->length)
321 sg++;
322 }
323
324out:
325 return ret;
326}
327
328int 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->data.op_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 */
390void rds_message_wait(struct rds_message *rm)
391{
392 wait_event_interruptible(rm->m_flush_wait,
393 !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
394}
395
396void rds_message_unmapped(struct rds_message *rm)
397{
398 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
399 wake_up_interruptible(&rm->m_flush_wait);
400}
401EXPORT_SYMBOL_GPL(rds_message_unmapped);
402
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#include <linux/export.h>
36#include <linux/skbuff.h>
37#include <linux/list.h>
38#include <linux/errqueue.h>
39
40#include "rds.h"
41
42static unsigned int rds_exthdr_size[__RDS_EXTHDR_MAX] = {
43[RDS_EXTHDR_NONE] = 0,
44[RDS_EXTHDR_VERSION] = sizeof(struct rds_ext_header_version),
45[RDS_EXTHDR_RDMA] = sizeof(struct rds_ext_header_rdma),
46[RDS_EXTHDR_RDMA_DEST] = sizeof(struct rds_ext_header_rdma_dest),
47[RDS_EXTHDR_NPATHS] = sizeof(u16),
48[RDS_EXTHDR_GEN_NUM] = sizeof(u32),
49};
50
51void rds_message_addref(struct rds_message *rm)
52{
53 rdsdebug("addref rm %p ref %d\n", rm, refcount_read(&rm->m_refcount));
54 refcount_inc(&rm->m_refcount);
55}
56EXPORT_SYMBOL_GPL(rds_message_addref);
57
58static inline bool rds_zcookie_add(struct rds_msg_zcopy_info *info, u32 cookie)
59{
60 struct rds_zcopy_cookies *ck = &info->zcookies;
61 int ncookies = ck->num;
62
63 if (ncookies == RDS_MAX_ZCOOKIES)
64 return false;
65 ck->cookies[ncookies] = cookie;
66 ck->num = ++ncookies;
67 return true;
68}
69
70static struct rds_msg_zcopy_info *rds_info_from_znotifier(struct rds_znotifier *znotif)
71{
72 return container_of(znotif, struct rds_msg_zcopy_info, znotif);
73}
74
75void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *q)
76{
77 unsigned long flags;
78 LIST_HEAD(copy);
79 struct rds_msg_zcopy_info *info, *tmp;
80
81 spin_lock_irqsave(&q->lock, flags);
82 list_splice(&q->zcookie_head, ©);
83 INIT_LIST_HEAD(&q->zcookie_head);
84 spin_unlock_irqrestore(&q->lock, flags);
85
86 list_for_each_entry_safe(info, tmp, ©, rs_zcookie_next) {
87 list_del(&info->rs_zcookie_next);
88 kfree(info);
89 }
90}
91
92static void rds_rm_zerocopy_callback(struct rds_sock *rs,
93 struct rds_znotifier *znotif)
94{
95 struct rds_msg_zcopy_info *info;
96 struct rds_msg_zcopy_queue *q;
97 u32 cookie = znotif->z_cookie;
98 struct rds_zcopy_cookies *ck;
99 struct list_head *head;
100 unsigned long flags;
101
102 mm_unaccount_pinned_pages(&znotif->z_mmp);
103 q = &rs->rs_zcookie_queue;
104 spin_lock_irqsave(&q->lock, flags);
105 head = &q->zcookie_head;
106 if (!list_empty(head)) {
107 info = list_entry(head, struct rds_msg_zcopy_info,
108 rs_zcookie_next);
109 if (info && rds_zcookie_add(info, cookie)) {
110 spin_unlock_irqrestore(&q->lock, flags);
111 kfree(rds_info_from_znotifier(znotif));
112 /* caller invokes rds_wake_sk_sleep() */
113 return;
114 }
115 }
116
117 info = rds_info_from_znotifier(znotif);
118 ck = &info->zcookies;
119 memset(ck, 0, sizeof(*ck));
120 WARN_ON(!rds_zcookie_add(info, cookie));
121 list_add_tail(&q->zcookie_head, &info->rs_zcookie_next);
122
123 spin_unlock_irqrestore(&q->lock, flags);
124 /* caller invokes rds_wake_sk_sleep() */
125}
126
127/*
128 * This relies on dma_map_sg() not touching sg[].page during merging.
129 */
130static void rds_message_purge(struct rds_message *rm)
131{
132 unsigned long i, flags;
133 bool zcopy = false;
134
135 if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags)))
136 return;
137
138 spin_lock_irqsave(&rm->m_rs_lock, flags);
139 if (rm->m_rs) {
140 struct rds_sock *rs = rm->m_rs;
141
142 if (rm->data.op_mmp_znotifier) {
143 zcopy = true;
144 rds_rm_zerocopy_callback(rs, rm->data.op_mmp_znotifier);
145 rds_wake_sk_sleep(rs);
146 rm->data.op_mmp_znotifier = NULL;
147 }
148 sock_put(rds_rs_to_sk(rs));
149 rm->m_rs = NULL;
150 }
151 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
152
153 for (i = 0; i < rm->data.op_nents; i++) {
154 /* XXX will have to put_page for page refs */
155 if (!zcopy)
156 __free_page(sg_page(&rm->data.op_sg[i]));
157 else
158 put_page(sg_page(&rm->data.op_sg[i]));
159 }
160 rm->data.op_nents = 0;
161
162 if (rm->rdma.op_active)
163 rds_rdma_free_op(&rm->rdma);
164 if (rm->rdma.op_rdma_mr)
165 rds_mr_put(rm->rdma.op_rdma_mr);
166
167 if (rm->atomic.op_active)
168 rds_atomic_free_op(&rm->atomic);
169 if (rm->atomic.op_rdma_mr)
170 rds_mr_put(rm->atomic.op_rdma_mr);
171}
172
173void rds_message_put(struct rds_message *rm)
174{
175 rdsdebug("put rm %p ref %d\n", rm, refcount_read(&rm->m_refcount));
176 WARN(!refcount_read(&rm->m_refcount), "danger refcount zero on %p\n", rm);
177 if (refcount_dec_and_test(&rm->m_refcount)) {
178 BUG_ON(!list_empty(&rm->m_sock_item));
179 BUG_ON(!list_empty(&rm->m_conn_item));
180 rds_message_purge(rm);
181
182 kfree(rm);
183 }
184}
185EXPORT_SYMBOL_GPL(rds_message_put);
186
187void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
188 __be16 dport, u64 seq)
189{
190 hdr->h_flags = 0;
191 hdr->h_sport = sport;
192 hdr->h_dport = dport;
193 hdr->h_sequence = cpu_to_be64(seq);
194 hdr->h_exthdr[0] = RDS_EXTHDR_NONE;
195}
196EXPORT_SYMBOL_GPL(rds_message_populate_header);
197
198int rds_message_add_extension(struct rds_header *hdr, unsigned int type,
199 const void *data, unsigned int len)
200{
201 unsigned int ext_len = sizeof(u8) + len;
202 unsigned char *dst;
203
204 /* For now, refuse to add more than one extension header */
205 if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE)
206 return 0;
207
208 if (type >= __RDS_EXTHDR_MAX || len != rds_exthdr_size[type])
209 return 0;
210
211 if (ext_len >= RDS_HEADER_EXT_SPACE)
212 return 0;
213 dst = hdr->h_exthdr;
214
215 *dst++ = type;
216 memcpy(dst, data, len);
217
218 dst[len] = RDS_EXTHDR_NONE;
219 return 1;
220}
221EXPORT_SYMBOL_GPL(rds_message_add_extension);
222
223/*
224 * If a message has extension headers, retrieve them here.
225 * Call like this:
226 *
227 * unsigned int pos = 0;
228 *
229 * while (1) {
230 * buflen = sizeof(buffer);
231 * type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
232 * if (type == RDS_EXTHDR_NONE)
233 * break;
234 * ...
235 * }
236 */
237int rds_message_next_extension(struct rds_header *hdr,
238 unsigned int *pos, void *buf, unsigned int *buflen)
239{
240 unsigned int offset, ext_type, ext_len;
241 u8 *src = hdr->h_exthdr;
242
243 offset = *pos;
244 if (offset >= RDS_HEADER_EXT_SPACE)
245 goto none;
246
247 /* Get the extension type and length. For now, the
248 * length is implied by the extension type. */
249 ext_type = src[offset++];
250
251 if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
252 goto none;
253 ext_len = rds_exthdr_size[ext_type];
254 if (offset + ext_len > RDS_HEADER_EXT_SPACE)
255 goto none;
256
257 *pos = offset + ext_len;
258 if (ext_len < *buflen)
259 *buflen = ext_len;
260 memcpy(buf, src + offset, *buflen);
261 return ext_type;
262
263none:
264 *pos = RDS_HEADER_EXT_SPACE;
265 *buflen = 0;
266 return RDS_EXTHDR_NONE;
267}
268
269int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
270{
271 struct rds_ext_header_rdma_dest ext_hdr;
272
273 ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
274 ext_hdr.h_rdma_offset = cpu_to_be32(offset);
275 return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr));
276}
277EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
278
279/*
280 * Each rds_message is allocated with extra space for the scatterlist entries
281 * rds ops will need. This is to minimize memory allocation count. Then, each rds op
282 * can grab SGs when initializing its part of the rds_message.
283 */
284struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
285{
286 struct rds_message *rm;
287
288 if (extra_len > KMALLOC_MAX_SIZE - sizeof(struct rds_message))
289 return NULL;
290
291 rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
292 if (!rm)
293 goto out;
294
295 rm->m_used_sgs = 0;
296 rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
297
298 refcount_set(&rm->m_refcount, 1);
299 INIT_LIST_HEAD(&rm->m_sock_item);
300 INIT_LIST_HEAD(&rm->m_conn_item);
301 spin_lock_init(&rm->m_rs_lock);
302 init_waitqueue_head(&rm->m_flush_wait);
303
304out:
305 return rm;
306}
307
308/*
309 * RDS ops use this to grab SG entries from the rm's sg pool.
310 */
311struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
312{
313 struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
314 struct scatterlist *sg_ret;
315
316 WARN_ON(rm->m_used_sgs + nents > rm->m_total_sgs);
317 WARN_ON(!nents);
318
319 if (rm->m_used_sgs + nents > rm->m_total_sgs)
320 return NULL;
321
322 sg_ret = &sg_first[rm->m_used_sgs];
323 sg_init_table(sg_ret, nents);
324 rm->m_used_sgs += nents;
325
326 return sg_ret;
327}
328
329struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
330{
331 struct rds_message *rm;
332 unsigned int i;
333 int num_sgs = ceil(total_len, PAGE_SIZE);
334 int extra_bytes = num_sgs * sizeof(struct scatterlist);
335
336 rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
337 if (!rm)
338 return ERR_PTR(-ENOMEM);
339
340 set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
341 rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
342 rm->data.op_nents = ceil(total_len, PAGE_SIZE);
343 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
344 if (!rm->data.op_sg) {
345 rds_message_put(rm);
346 return ERR_PTR(-ENOMEM);
347 }
348
349 for (i = 0; i < rm->data.op_nents; ++i) {
350 sg_set_page(&rm->data.op_sg[i],
351 virt_to_page(page_addrs[i]),
352 PAGE_SIZE, 0);
353 }
354
355 return rm;
356}
357
358static int rds_message_zcopy_from_user(struct rds_message *rm, struct iov_iter *from)
359{
360 struct scatterlist *sg;
361 int ret = 0;
362 int length = iov_iter_count(from);
363 int total_copied = 0;
364 struct rds_msg_zcopy_info *info;
365
366 rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from));
367
368 /*
369 * now allocate and copy in the data payload.
370 */
371 sg = rm->data.op_sg;
372
373 info = kzalloc(sizeof(*info), GFP_KERNEL);
374 if (!info)
375 return -ENOMEM;
376 INIT_LIST_HEAD(&info->rs_zcookie_next);
377 rm->data.op_mmp_znotifier = &info->znotif;
378 if (mm_account_pinned_pages(&rm->data.op_mmp_znotifier->z_mmp,
379 length)) {
380 ret = -ENOMEM;
381 goto err;
382 }
383 while (iov_iter_count(from)) {
384 struct page *pages;
385 size_t start;
386 ssize_t copied;
387
388 copied = iov_iter_get_pages(from, &pages, PAGE_SIZE,
389 1, &start);
390 if (copied < 0) {
391 struct mmpin *mmp;
392 int i;
393
394 for (i = 0; i < rm->data.op_nents; i++)
395 put_page(sg_page(&rm->data.op_sg[i]));
396 mmp = &rm->data.op_mmp_znotifier->z_mmp;
397 mm_unaccount_pinned_pages(mmp);
398 ret = -EFAULT;
399 goto err;
400 }
401 total_copied += copied;
402 iov_iter_advance(from, copied);
403 length -= copied;
404 sg_set_page(sg, pages, copied, start);
405 rm->data.op_nents++;
406 sg++;
407 }
408 WARN_ON_ONCE(length != 0);
409 return ret;
410err:
411 kfree(info);
412 rm->data.op_mmp_znotifier = NULL;
413 return ret;
414}
415
416int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
417 bool zcopy)
418{
419 unsigned long to_copy, nbytes;
420 unsigned long sg_off;
421 struct scatterlist *sg;
422 int ret = 0;
423
424 rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from));
425
426 /* now allocate and copy in the data payload. */
427 sg = rm->data.op_sg;
428 sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
429
430 if (zcopy)
431 return rds_message_zcopy_from_user(rm, from);
432
433 while (iov_iter_count(from)) {
434 if (!sg_page(sg)) {
435 ret = rds_page_remainder_alloc(sg, iov_iter_count(from),
436 GFP_HIGHUSER);
437 if (ret)
438 return ret;
439 rm->data.op_nents++;
440 sg_off = 0;
441 }
442
443 to_copy = min_t(unsigned long, iov_iter_count(from),
444 sg->length - sg_off);
445
446 rds_stats_add(s_copy_from_user, to_copy);
447 nbytes = copy_page_from_iter(sg_page(sg), sg->offset + sg_off,
448 to_copy, from);
449 if (nbytes != to_copy)
450 return -EFAULT;
451
452 sg_off += to_copy;
453
454 if (sg_off == sg->length)
455 sg++;
456 }
457
458 return ret;
459}
460
461int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
462{
463 struct rds_message *rm;
464 struct scatterlist *sg;
465 unsigned long to_copy;
466 unsigned long vec_off;
467 int copied;
468 int ret;
469 u32 len;
470
471 rm = container_of(inc, struct rds_message, m_inc);
472 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
473
474 sg = rm->data.op_sg;
475 vec_off = 0;
476 copied = 0;
477
478 while (iov_iter_count(to) && copied < len) {
479 to_copy = min_t(unsigned long, iov_iter_count(to),
480 sg->length - vec_off);
481 to_copy = min_t(unsigned long, to_copy, len - copied);
482
483 rds_stats_add(s_copy_to_user, to_copy);
484 ret = copy_page_to_iter(sg_page(sg), sg->offset + vec_off,
485 to_copy, to);
486 if (ret != to_copy)
487 return -EFAULT;
488
489 vec_off += to_copy;
490 copied += to_copy;
491
492 if (vec_off == sg->length) {
493 vec_off = 0;
494 sg++;
495 }
496 }
497
498 return copied;
499}
500
501/*
502 * If the message is still on the send queue, wait until the transport
503 * is done with it. This is particularly important for RDMA operations.
504 */
505void rds_message_wait(struct rds_message *rm)
506{
507 wait_event_interruptible(rm->m_flush_wait,
508 !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
509}
510
511void rds_message_unmapped(struct rds_message *rm)
512{
513 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
514 wake_up_interruptible(&rm->m_flush_wait);
515}
516EXPORT_SYMBOL_GPL(rds_message_unmapped);
517