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         if (atomic_read(&rm->m_refcount) == 0) {
  86 printk(KERN_CRIT "danger refcount zero on %p\n", rm);
  87 WARN_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 }
  97 EXPORT_SYMBOL_GPL(rds_message_put);
  98
  99 void 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 }
 108 EXPORT_SYMBOL_GPL(rds_message_populate_header);
 109
 110 int 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 }
 133 EXPORT_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  */
 149 int 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
 175 none:
 176         *pos = RDS_HEADER_EXT_SPACE;
 177         *buflen = 0;
 178         return RDS_EXTHDR_NONE;
 179 }
 180
 181 int 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 }
 189 EXPORT_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  */
 196 struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
 197 {
 198         struct rds_message *rm;
 199
 200         if (extra_len > KMALLOC_MAX_SIZE - sizeof(struct rds_message))
 201                 return NULL;
 202
 203         rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
 204         if (!rm)
 205                 goto out;
 206
 207         rm->m_used_sgs = 0;
 208         rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
 209
 210         atomic_set(&rm->m_refcount, 1);
 211         INIT_LIST_HEAD(&rm->m_sock_item);
 212         INIT_LIST_HEAD(&rm->m_conn_item);
 213         spin_lock_init(&rm->m_rs_lock);
 214         init_waitqueue_head(&rm->m_flush_wait);
 215
 216 out:
 217         return rm;
 218 }
 219
 220 /*
 221  * RDS ops use this to grab SG entries from the rm's sg pool.
 222  */
 223 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
 224 {
 225         struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
 226         struct scatterlist *sg_ret;
 227
 228         WARN_ON(rm->m_used_sgs + nents > rm->m_total_sgs);
 229         WARN_ON(!nents);
 230
 231         if (rm->m_used_sgs + nents > rm->m_total_sgs)
 232                 return NULL;
 233
 234         sg_ret = &sg_first[rm->m_used_sgs];
 235         sg_init_table(sg_ret, nents);
 236         rm->m_used_sgs += nents;
 237
 238         return sg_ret;
 239 }
 240
 241 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
 242 {
 243         struct rds_message *rm;
 244         unsigned int i;
 245         int num_sgs = ceil(total_len, PAGE_SIZE);
 246         int extra_bytes = num_sgs * sizeof(struct scatterlist);
 247
 248         rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
 249         if (!rm)
 250                 return ERR_PTR(-ENOMEM);
 251
 252         set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
 253         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
 254         rm->data.op_nents = ceil(total_len, PAGE_SIZE);
 255         rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
 256         if (!rm->data.op_sg) {
 257                 rds_message_put(rm);
 258                 return ERR_PTR(-ENOMEM);
 259         }
 260
 261         for (i = 0; i < rm->data.op_nents; ++i) {
 262                 sg_set_page(&rm->data.op_sg[i],
 263                                 virt_to_page(page_addrs[i]),
 264                                 PAGE_SIZE, 0);
 265         }
 266
 267         return rm;
 268 }
 269
 270 int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
 271                                                size_t total_len)
 272 {
 273         unsigned long to_copy;
 274         unsigned long iov_off;
 275         unsigned long sg_off;
 276         struct iovec *iov;
 277         struct scatterlist *sg;
 278         int ret = 0;
 279
 280         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
 281
 282         /*
 283          * now allocate and copy in the data payload.
 284          */
 285         sg = rm->data.op_sg;
 286         iov = first_iov;
 287         iov_off = 0;
 288         sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
 289
 290         while (total_len) {
 291                 if (!sg_page(sg)) {
 292                         ret = rds_page_remainder_alloc(sg, total_len,
 293                                                        GFP_HIGHUSER);
 294                         if (ret)
 295                                 goto out;
 296                         rm->data.op_nents++;
 297                         sg_off = 0;
 298                 }
 299
 300                 while (iov_off == iov->iov_len) {
 301                         iov_off = 0;
 302                         iov++;
 303                 }
 304
 305                 to_copy = min(iov->iov_len - iov_off, sg->length - sg_off);
 306                 to_copy = min_t(size_t, to_copy, total_len);
 307
 308                 rdsdebug("copying %lu bytes from user iov [%p, %zu] + %lu to "
 309                          "sg [%p, %u, %u] + %lu\n",
 310                          to_copy, iov->iov_base, iov->iov_len, iov_off,
 311                          (void *)sg_page(sg), sg->offset, sg->length, sg_off);
 312
 313                 ret = rds_page_copy_from_user(sg_page(sg), sg->offset + sg_off,
 314                                               iov->iov_base + iov_off,
 315                                               to_copy);
 316                 if (ret)
 317                         goto out;
 318
 319                 iov_off += to_copy;
 320                 total_len -= to_copy;
 321                 sg_off += to_copy;
 322
 323                 if (sg_off == sg->length)
 324                         sg++;
 325         }
 326
 327 out:
 328         return ret;
 329 }
 330
 331 int rds_message_inc_copy_to_user(struct rds_incoming *inc,
 332                                  struct iovec *first_iov, size_t size)
 333 {
 334         struct rds_message *rm;
 335         struct iovec *iov;
 336         struct scatterlist *sg;
 337         unsigned long to_copy;
 338         unsigned long iov_off;
 339         unsigned long vec_off;
 340         int copied;
 341         int ret;
 342         u32 len;
 343
 344         rm = container_of(inc, struct rds_message, m_inc);
 345         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 346
 347         iov = first_iov;
 348         iov_off = 0;
 349         sg = rm->data.op_sg;
 350         vec_off = 0;
 351         copied = 0;
 352
 353         while (copied < size && copied < len) {
 354                 while (iov_off == iov->iov_len) {
 355                         iov_off = 0;
 356                         iov++;
 357                 }
 358
 359                 to_copy = min(iov->iov_len - iov_off, sg->length - vec_off);
 360                 to_copy = min_t(size_t, to_copy, size - copied);
 361                 to_copy = min_t(unsigned long, to_copy, len - copied);
 362
 363                 rdsdebug("copying %lu bytes to user iov [%p, %zu] + %lu to "
 364                          "sg [%p, %u, %u] + %lu\n",
 365                          to_copy, iov->iov_base, iov->iov_len, iov_off,
 366                          sg_page(sg), sg->offset, sg->length, vec_off);
 367
 368                 ret = rds_page_copy_to_user(sg_page(sg), sg->offset + vec_off,
 369                                             iov->iov_base + iov_off,
 370                                             to_copy);
 371                 if (ret) {
 372                         copied = ret;
 373                         break;
 374                 }
 375
 376                 iov_off += to_copy;
 377                 vec_off += to_copy;
 378                 copied += to_copy;
 379
 380                 if (vec_off == sg->length) {
 381                         vec_off = 0;
 382                         sg++;
 383                 }
 384         }
 385
 386         return copied;
 387 }
 388
 389 /*
 390  * If the message is still on the send queue, wait until the transport
 391  * is done with it. This is particularly important for RDMA operations.
 392  */
 393 void rds_message_wait(struct rds_message *rm)
 394 {
 395         wait_event_interruptible(rm->m_flush_wait,
 396                         !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
 397 }
 398
 399 void rds_message_unmapped(struct rds_message *rm)
 400 {
 401         clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 402         wake_up_interruptible(&rm->m_flush_wait);
 403 }
 404 EXPORT_SYMBOL_GPL(rds_message_unmapped);
 405