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 #include <linux/export.h>
  36
  37 #include "rds.h"
  38
  39 static 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
  47 void 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 }
  52 EXPORT_SYMBOL_GPL(rds_message_addref);
  53
  54 /*
  55  * This relies on dma_map_sg() not touching sg[].page during merging.
  56  */
  57 static 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
  82 void rds_message_put(struct rds_message *rm)
  83 {
  84         rdsdebug("put rm %p ref %d\n", rm, atomic_read(&rm->m_refcount));
  85         WARN(!atomic_read(&rm->m_refcount), "danger refcount zero on %p\n", rm);
  86         if (atomic_dec_and_test(&rm->m_refcount)) {
  87                 BUG_ON(!list_empty(&rm->m_sock_item));
  88                 BUG_ON(!list_empty(&rm->m_conn_item));
  89                 rds_message_purge(rm);
  90
  91                 kfree(rm);
  92         }
  93 }
  94 EXPORT_SYMBOL_GPL(rds_message_put);
  95
  96 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
  97                                  __be16 dport, u64 seq)
  98 {
  99         hdr->h_flags = 0;
 100         hdr->h_sport = sport;
 101         hdr->h_dport = dport;
 102         hdr->h_sequence = cpu_to_be64(seq);
 103         hdr->h_exthdr[0] = RDS_EXTHDR_NONE;
 104 }
 105 EXPORT_SYMBOL_GPL(rds_message_populate_header);
 106
 107 int rds_message_add_extension(struct rds_header *hdr, unsigned int type,
 108                               const void *data, unsigned int len)
 109 {
 110         unsigned int ext_len = sizeof(u8) + len;
 111         unsigned char *dst;
 112
 113         /* For now, refuse to add more than one extension header */
 114         if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE)
 115                 return 0;
 116
 117         if (type >= __RDS_EXTHDR_MAX || len != rds_exthdr_size[type])
 118                 return 0;
 119
 120         if (ext_len >= RDS_HEADER_EXT_SPACE)
 121                 return 0;
 122         dst = hdr->h_exthdr;
 123
 124         *dst++ = type;
 125         memcpy(dst, data, len);
 126
 127         dst[len] = RDS_EXTHDR_NONE;
 128         return 1;
 129 }
 130 EXPORT_SYMBOL_GPL(rds_message_add_extension);
 131
 132 /*
 133  * If a message has extension headers, retrieve them here.
 134  * Call like this:
 135  *
 136  * unsigned int pos = 0;
 137  *
 138  * while (1) {
 139  *      buflen = sizeof(buffer);
 140  *      type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
 141  *      if (type == RDS_EXTHDR_NONE)
 142  *              break;
 143  *      ...
 144  * }
 145  */
 146 int rds_message_next_extension(struct rds_header *hdr,
 147                 unsigned int *pos, void *buf, unsigned int *buflen)
 148 {
 149         unsigned int offset, ext_type, ext_len;
 150         u8 *src = hdr->h_exthdr;
 151
 152         offset = *pos;
 153         if (offset >= RDS_HEADER_EXT_SPACE)
 154                 goto none;
 155
 156         /* Get the extension type and length. For now, the
 157          * length is implied by the extension type. */
 158         ext_type = src[offset++];
 159
 160         if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
 161                 goto none;
 162         ext_len = rds_exthdr_size[ext_type];
 163         if (offset + ext_len > RDS_HEADER_EXT_SPACE)
 164                 goto none;
 165
 166         *pos = offset + ext_len;
 167         if (ext_len < *buflen)
 168                 *buflen = ext_len;
 169         memcpy(buf, src + offset, *buflen);
 170         return ext_type;
 171
 172 none:
 173         *pos = RDS_HEADER_EXT_SPACE;
 174         *buflen = 0;
 175         return RDS_EXTHDR_NONE;
 176 }
 177
 178 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
 179 {
 180         struct rds_ext_header_rdma_dest ext_hdr;
 181
 182         ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
 183         ext_hdr.h_rdma_offset = cpu_to_be32(offset);
 184         return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr));
 185 }
 186 EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
 187
 188 /*
 189  * Each rds_message is allocated with extra space for the scatterlist entries
 190  * rds ops will need. This is to minimize memory allocation count. Then, each rds op
 191  * can grab SGs when initializing its part of the rds_message.
 192  */
 193 struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
 194 {
 195         struct rds_message *rm;
 196
 197         if (extra_len > KMALLOC_MAX_SIZE - sizeof(struct rds_message))
 198                 return NULL;
 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
 213 out:
 214         return rm;
 215 }
 216
 217 /*
 218  * RDS ops use this to grab SG entries from the rm's sg pool.
 219  */
 220 struct 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
 238 struct 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
 267 int 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
 324 out:
 325         return ret;
 326 }
 327
 328 int 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  */
 390 void 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
 396 void 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 }
 401 EXPORT_SYMBOL_GPL(rds_message_unmapped);
 402