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