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 [RDS_EXTHDR_NPATHS]     = sizeof(u16),
  45 [RDS_EXTHDR_GEN_NUM]    = sizeof(u32),
  46 };
  47
  48
  49 void rds_message_addref(struct rds_message *rm)
  50 {
  51         rdsdebug("addref rm %p ref %d\n", rm, refcount_read(&rm->m_refcount));
  52         refcount_inc(&rm->m_refcount);
  53 }
  54 EXPORT_SYMBOL_GPL(rds_message_addref);
  55
  56 /*
  57  * This relies on dma_map_sg() not touching sg[].page during merging.
  58  */
  59 static void rds_message_purge(struct rds_message *rm)
  60 {
  61         unsigned long i;
  62
  63         if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags)))
  64                 return;
  65
  66         for (i = 0; i < rm->data.op_nents; i++) {
  67                 rdsdebug("putting data page %p\n", (void *)sg_page(&rm->data.op_sg[i]));
  68                 /* XXX will have to put_page for page refs */
  69                 __free_page(sg_page(&rm->data.op_sg[i]));
  70         }
  71         rm->data.op_nents = 0;
  72
  73         if (rm->rdma.op_active)
  74                 rds_rdma_free_op(&rm->rdma);
  75         if (rm->rdma.op_rdma_mr)
  76                 rds_mr_put(rm->rdma.op_rdma_mr);
  77
  78         if (rm->atomic.op_active)
  79                 rds_atomic_free_op(&rm->atomic);
  80         if (rm->atomic.op_rdma_mr)
  81                 rds_mr_put(rm->atomic.op_rdma_mr);
  82 }
  83
  84 void rds_message_put(struct rds_message *rm)
  85 {
  86         rdsdebug("put rm %p ref %d\n", rm, refcount_read(&rm->m_refcount));
  87         WARN(!refcount_read(&rm->m_refcount), "danger refcount zero on %p\n", rm);
  88         if (refcount_dec_and_test(&rm->m_refcount)) {
  89                 BUG_ON(!list_empty(&rm->m_sock_item));
  90                 BUG_ON(!list_empty(&rm->m_conn_item));
  91                 rds_message_purge(rm);
  92
  93                 kfree(rm);
  94         }
  95 }
  96 EXPORT_SYMBOL_GPL(rds_message_put);
  97
  98 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
  99                                  __be16 dport, u64 seq)
 100 {
 101         hdr->h_flags = 0;
 102         hdr->h_sport = sport;
 103         hdr->h_dport = dport;
 104         hdr->h_sequence = cpu_to_be64(seq);
 105         hdr->h_exthdr[0] = RDS_EXTHDR_NONE;
 106 }
 107 EXPORT_SYMBOL_GPL(rds_message_populate_header);
 108
 109 int rds_message_add_extension(struct rds_header *hdr, unsigned int type,
 110                               const void *data, unsigned int len)
 111 {
 112         unsigned int ext_len = sizeof(u8) + len;
 113         unsigned char *dst;
 114
 115         /* For now, refuse to add more than one extension header */
 116         if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE)
 117                 return 0;
 118
 119         if (type >= __RDS_EXTHDR_MAX || len != rds_exthdr_size[type])
 120                 return 0;
 121
 122         if (ext_len >= RDS_HEADER_EXT_SPACE)
 123                 return 0;
 124         dst = hdr->h_exthdr;
 125
 126         *dst++ = type;
 127         memcpy(dst, data, len);
 128
 129         dst[len] = RDS_EXTHDR_NONE;
 130         return 1;
 131 }
 132 EXPORT_SYMBOL_GPL(rds_message_add_extension);
 133
 134 /*
 135  * If a message has extension headers, retrieve them here.
 136  * Call like this:
 137  *
 138  * unsigned int pos = 0;
 139  *
 140  * while (1) {
 141  *      buflen = sizeof(buffer);
 142  *      type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
 143  *      if (type == RDS_EXTHDR_NONE)
 144  *              break;
 145  *      ...
 146  * }
 147  */
 148 int rds_message_next_extension(struct rds_header *hdr,
 149                 unsigned int *pos, void *buf, unsigned int *buflen)
 150 {
 151         unsigned int offset, ext_type, ext_len;
 152         u8 *src = hdr->h_exthdr;
 153
 154         offset = *pos;
 155         if (offset >= RDS_HEADER_EXT_SPACE)
 156                 goto none;
 157
 158         /* Get the extension type and length. For now, the
 159          * length is implied by the extension type. */
 160         ext_type = src[offset++];
 161
 162         if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
 163                 goto none;
 164         ext_len = rds_exthdr_size[ext_type];
 165         if (offset + ext_len > RDS_HEADER_EXT_SPACE)
 166                 goto none;
 167
 168         *pos = offset + ext_len;
 169         if (ext_len < *buflen)
 170                 *buflen = ext_len;
 171         memcpy(buf, src + offset, *buflen);
 172         return ext_type;
 173
 174 none:
 175         *pos = RDS_HEADER_EXT_SPACE;
 176         *buflen = 0;
 177         return RDS_EXTHDR_NONE;
 178 }
 179
 180 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
 181 {
 182         struct rds_ext_header_rdma_dest ext_hdr;
 183
 184         ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
 185         ext_hdr.h_rdma_offset = cpu_to_be32(offset);
 186         return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr));
 187 }
 188 EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
 189
 190 /*
 191  * Each rds_message is allocated with extra space for the scatterlist entries
 192  * rds ops will need. This is to minimize memory allocation count. Then, each rds op
 193  * can grab SGs when initializing its part of the rds_message.
 194  */
 195 struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
 196 {
 197         struct rds_message *rm;
 198
 199         if (extra_len > KMALLOC_MAX_SIZE - sizeof(struct rds_message))
 200                 return NULL;
 201
 202         rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
 203         if (!rm)
 204                 goto out;
 205
 206         rm->m_used_sgs = 0;
 207         rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
 208
 209         refcount_set(&rm->m_refcount, 1);
 210         INIT_LIST_HEAD(&rm->m_sock_item);
 211         INIT_LIST_HEAD(&rm->m_conn_item);
 212         spin_lock_init(&rm->m_rs_lock);
 213         init_waitqueue_head(&rm->m_flush_wait);
 214
 215 out:
 216         return rm;
 217 }
 218
 219 /*
 220  * RDS ops use this to grab SG entries from the rm's sg pool.
 221  */
 222 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
 223 {
 224         struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
 225         struct scatterlist *sg_ret;
 226
 227         WARN_ON(rm->m_used_sgs + nents > rm->m_total_sgs);
 228         WARN_ON(!nents);
 229
 230         if (rm->m_used_sgs + nents > rm->m_total_sgs)
 231                 return NULL;
 232
 233         sg_ret = &sg_first[rm->m_used_sgs];
 234         sg_init_table(sg_ret, nents);
 235         rm->m_used_sgs += nents;
 236
 237         return sg_ret;
 238 }
 239
 240 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
 241 {
 242         struct rds_message *rm;
 243         unsigned int i;
 244         int num_sgs = ceil(total_len, PAGE_SIZE);
 245         int extra_bytes = num_sgs * sizeof(struct scatterlist);
 246
 247         rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
 248         if (!rm)
 249                 return ERR_PTR(-ENOMEM);
 250
 251         set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
 252         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
 253         rm->data.op_nents = ceil(total_len, PAGE_SIZE);
 254         rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
 255         if (!rm->data.op_sg) {
 256                 rds_message_put(rm);
 257                 return ERR_PTR(-ENOMEM);
 258         }
 259
 260         for (i = 0; i < rm->data.op_nents; ++i) {
 261                 sg_set_page(&rm->data.op_sg[i],
 262                                 virt_to_page(page_addrs[i]),
 263                                 PAGE_SIZE, 0);
 264         }
 265
 266         return rm;
 267 }
 268
 269 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from)
 270 {
 271         unsigned long to_copy, nbytes;
 272         unsigned long sg_off;
 273         struct scatterlist *sg;
 274         int ret = 0;
 275
 276         rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from));
 277
 278         /*
 279          * now allocate and copy in the data payload.
 280          */
 281         sg = rm->data.op_sg;
 282         sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
 283
 284         while (iov_iter_count(from)) {
 285                 if (!sg_page(sg)) {
 286                         ret = rds_page_remainder_alloc(sg, iov_iter_count(from),
 287                                                        GFP_HIGHUSER);
 288                         if (ret)
 289                                 return ret;
 290                         rm->data.op_nents++;
 291                         sg_off = 0;
 292                 }
 293
 294                 to_copy = min_t(unsigned long, iov_iter_count(from),
 295                                 sg->length - sg_off);
 296
 297                 rds_stats_add(s_copy_from_user, to_copy);
 298                 nbytes = copy_page_from_iter(sg_page(sg), sg->offset + sg_off,
 299                                              to_copy, from);
 300                 if (nbytes != to_copy)
 301                         return -EFAULT;
 302
 303                 sg_off += to_copy;
 304
 305                 if (sg_off == sg->length)
 306                         sg++;
 307         }
 308
 309         return ret;
 310 }
 311
 312 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
 313 {
 314         struct rds_message *rm;
 315         struct scatterlist *sg;
 316         unsigned long to_copy;
 317         unsigned long vec_off;
 318         int copied;
 319         int ret;
 320         u32 len;
 321
 322         rm = container_of(inc, struct rds_message, m_inc);
 323         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 324
 325         sg = rm->data.op_sg;
 326         vec_off = 0;
 327         copied = 0;
 328
 329         while (iov_iter_count(to) && copied < len) {
 330                 to_copy = min_t(unsigned long, iov_iter_count(to),
 331                                 sg->length - vec_off);
 332                 to_copy = min_t(unsigned long, to_copy, len - copied);
 333
 334                 rds_stats_add(s_copy_to_user, to_copy);
 335                 ret = copy_page_to_iter(sg_page(sg), sg->offset + vec_off,
 336                                         to_copy, to);
 337                 if (ret != to_copy)
 338                         return -EFAULT;
 339
 340                 vec_off += to_copy;
 341                 copied += to_copy;
 342
 343                 if (vec_off == sg->length) {
 344                         vec_off = 0;
 345                         sg++;
 346                 }
 347         }
 348
 349         return copied;
 350 }
 351
 352 /*
 353  * If the message is still on the send queue, wait until the transport
 354  * is done with it. This is particularly important for RDMA operations.
 355  */
 356 void rds_message_wait(struct rds_message *rm)
 357 {
 358         wait_event_interruptible(rm->m_flush_wait,
 359                         !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
 360 }
 361
 362 void rds_message_unmapped(struct rds_message *rm)
 363 {
 364         clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 365         wake_up_interruptible(&rm->m_flush_wait);
 366 }
 367 EXPORT_SYMBOL_GPL(rds_message_unmapped);
 368