net/rds/message.c

   1 /*
   2  * Copyright (c) 2006, 2020 Oracle and/or its affiliates.
   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
  42 static 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
  51 void 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 }
  56 EXPORT_SYMBOL_GPL(rds_message_addref);
  57
  58 static 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
  70 static 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
  75 void 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, &copy);
  83         INIT_LIST_HEAD(&q->zcookie_head);
  84         spin_unlock_irqrestore(&q->lock, flags);
  85
  86         list_for_each_entry_safe(info, tmp, &copy, rs_zcookie_next) {
  87                 list_del(&info->rs_zcookie_next);
  88                 kfree(info);
  89         }
  90 }
  91
  92 static 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  */
 130 static 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                 kref_put(&rm->rdma.op_rdma_mr->r_kref, __rds_put_mr_final);
 166
 167         if (rm->atomic.op_active)
 168                 rds_atomic_free_op(&rm->atomic);
 169         if (rm->atomic.op_rdma_mr)
 170                 kref_put(&rm->atomic.op_rdma_mr->r_kref, __rds_put_mr_final);
 171 }
 172
 173 void 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                 rm = NULL;
 184         }
 185 }
 186 EXPORT_SYMBOL_GPL(rds_message_put);
 187
 188 void rds_message_populate_header(struct rds_header *hdr, __be16 sport,
 189                                  __be16 dport, u64 seq)
 190 {
 191         hdr->h_flags = 0;
 192         hdr->h_sport = sport;
 193         hdr->h_dport = dport;
 194         hdr->h_sequence = cpu_to_be64(seq);
 195         hdr->h_exthdr[0] = RDS_EXTHDR_NONE;
 196 }
 197 EXPORT_SYMBOL_GPL(rds_message_populate_header);
 198
 199 int rds_message_add_extension(struct rds_header *hdr, unsigned int type,
 200                               const void *data, unsigned int len)
 201 {
 202         unsigned int ext_len = sizeof(u8) + len;
 203         unsigned char *dst;
 204
 205         /* For now, refuse to add more than one extension header */
 206         if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE)
 207                 return 0;
 208
 209         if (type >= __RDS_EXTHDR_MAX || len != rds_exthdr_size[type])
 210                 return 0;
 211
 212         if (ext_len >= RDS_HEADER_EXT_SPACE)
 213                 return 0;
 214         dst = hdr->h_exthdr;
 215
 216         *dst++ = type;
 217         memcpy(dst, data, len);
 218
 219         dst[len] = RDS_EXTHDR_NONE;
 220         return 1;
 221 }
 222 EXPORT_SYMBOL_GPL(rds_message_add_extension);
 223
 224 /*
 225  * If a message has extension headers, retrieve them here.
 226  * Call like this:
 227  *
 228  * unsigned int pos = 0;
 229  *
 230  * while (1) {
 231  *      buflen = sizeof(buffer);
 232  *      type = rds_message_next_extension(hdr, &pos, buffer, &buflen);
 233  *      if (type == RDS_EXTHDR_NONE)
 234  *              break;
 235  *      ...
 236  * }
 237  */
 238 int rds_message_next_extension(struct rds_header *hdr,
 239                 unsigned int *pos, void *buf, unsigned int *buflen)
 240 {
 241         unsigned int offset, ext_type, ext_len;
 242         u8 *src = hdr->h_exthdr;
 243
 244         offset = *pos;
 245         if (offset >= RDS_HEADER_EXT_SPACE)
 246                 goto none;
 247
 248         /* Get the extension type and length. For now, the
 249          * length is implied by the extension type. */
 250         ext_type = src[offset++];
 251
 252         if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX)
 253                 goto none;
 254         ext_len = rds_exthdr_size[ext_type];
 255         if (offset + ext_len > RDS_HEADER_EXT_SPACE)
 256                 goto none;
 257
 258         *pos = offset + ext_len;
 259         if (ext_len < *buflen)
 260                 *buflen = ext_len;
 261         memcpy(buf, src + offset, *buflen);
 262         return ext_type;
 263
 264 none:
 265         *pos = RDS_HEADER_EXT_SPACE;
 266         *buflen = 0;
 267         return RDS_EXTHDR_NONE;
 268 }
 269
 270 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset)
 271 {
 272         struct rds_ext_header_rdma_dest ext_hdr;
 273
 274         ext_hdr.h_rdma_rkey = cpu_to_be32(r_key);
 275         ext_hdr.h_rdma_offset = cpu_to_be32(offset);
 276         return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr));
 277 }
 278 EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension);
 279
 280 /*
 281  * Each rds_message is allocated with extra space for the scatterlist entries
 282  * rds ops will need. This is to minimize memory allocation count. Then, each rds op
 283  * can grab SGs when initializing its part of the rds_message.
 284  */
 285 struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp)
 286 {
 287         struct rds_message *rm;
 288
 289         if (extra_len > KMALLOC_MAX_SIZE - sizeof(struct rds_message))
 290                 return NULL;
 291
 292         rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp);
 293         if (!rm)
 294                 goto out;
 295
 296         rm->m_used_sgs = 0;
 297         rm->m_total_sgs = extra_len / sizeof(struct scatterlist);
 298
 299         refcount_set(&rm->m_refcount, 1);
 300         INIT_LIST_HEAD(&rm->m_sock_item);
 301         INIT_LIST_HEAD(&rm->m_conn_item);
 302         spin_lock_init(&rm->m_rs_lock);
 303         init_waitqueue_head(&rm->m_flush_wait);
 304
 305 out:
 306         return rm;
 307 }
 308
 309 /*
 310  * RDS ops use this to grab SG entries from the rm's sg pool.
 311  */
 312 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents)
 313 {
 314         struct scatterlist *sg_first = (struct scatterlist *) &rm[1];
 315         struct scatterlist *sg_ret;
 316
 317         if (nents <= 0) {
 318                 pr_warn("rds: alloc sgs failed! nents <= 0\n");
 319                 return ERR_PTR(-EINVAL);
 320         }
 321
 322         if (rm->m_used_sgs + nents > rm->m_total_sgs) {
 323                 pr_warn("rds: alloc sgs failed! total %d used %d nents %d\n",
 324                         rm->m_total_sgs, rm->m_used_sgs, nents);
 325                 return ERR_PTR(-ENOMEM);
 326         }
 327
 328         sg_ret = &sg_first[rm->m_used_sgs];
 329         sg_init_table(sg_ret, nents);
 330         rm->m_used_sgs += nents;
 331
 332         return sg_ret;
 333 }
 334
 335 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len)
 336 {
 337         struct rds_message *rm;
 338         unsigned int i;
 339         int num_sgs = DIV_ROUND_UP(total_len, PAGE_SIZE);
 340         int extra_bytes = num_sgs * sizeof(struct scatterlist);
 341
 342         rm = rds_message_alloc(extra_bytes, GFP_NOWAIT);
 343         if (!rm)
 344                 return ERR_PTR(-ENOMEM);
 345
 346         set_bit(RDS_MSG_PAGEVEC, &rm->m_flags);
 347         rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len);
 348         rm->data.op_nents = DIV_ROUND_UP(total_len, PAGE_SIZE);
 349         rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs);
 350         if (IS_ERR(rm->data.op_sg)) {
 351                 void *err = ERR_CAST(rm->data.op_sg);
 352                 rds_message_put(rm);
 353                 return err;
 354         }
 355
 356         for (i = 0; i < rm->data.op_nents; ++i) {
 357                 sg_set_page(&rm->data.op_sg[i],
 358                                 virt_to_page(page_addrs[i]),
 359                                 PAGE_SIZE, 0);
 360         }
 361
 362         return rm;
 363 }
 364
 365 static int rds_message_zcopy_from_user(struct rds_message *rm, struct iov_iter *from)
 366 {
 367         struct scatterlist *sg;
 368         int ret = 0;
 369         int length = iov_iter_count(from);
 370         int total_copied = 0;
 371         struct rds_msg_zcopy_info *info;
 372
 373         rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from));
 374
 375         /*
 376          * now allocate and copy in the data payload.
 377          */
 378         sg = rm->data.op_sg;
 379
 380         info = kzalloc(sizeof(*info), GFP_KERNEL);
 381         if (!info)
 382                 return -ENOMEM;
 383         INIT_LIST_HEAD(&info->rs_zcookie_next);
 384         rm->data.op_mmp_znotifier = &info->znotif;
 385         if (mm_account_pinned_pages(&rm->data.op_mmp_znotifier->z_mmp,
 386                                     length)) {
 387                 ret = -ENOMEM;
 388                 goto err;
 389         }
 390         while (iov_iter_count(from)) {
 391                 struct page *pages;
 392                 size_t start;
 393                 ssize_t copied;
 394
 395                 copied = iov_iter_get_pages(from, &pages, PAGE_SIZE,
 396                                             1, &start);
 397                 if (copied < 0) {
 398                         struct mmpin *mmp;
 399                         int i;
 400
 401                         for (i = 0; i < rm->data.op_nents; i++)
 402                                 put_page(sg_page(&rm->data.op_sg[i]));
 403                         mmp = &rm->data.op_mmp_znotifier->z_mmp;
 404                         mm_unaccount_pinned_pages(mmp);
 405                         ret = -EFAULT;
 406                         goto err;
 407                 }
 408                 total_copied += copied;
 409                 iov_iter_advance(from, copied);
 410                 length -= copied;
 411                 sg_set_page(sg, pages, copied, start);
 412                 rm->data.op_nents++;
 413                 sg++;
 414         }
 415         WARN_ON_ONCE(length != 0);
 416         return ret;
 417 err:
 418         kfree(info);
 419         rm->data.op_mmp_znotifier = NULL;
 420         return ret;
 421 }
 422
 423 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from,
 424                                bool zcopy)
 425 {
 426         unsigned long to_copy, nbytes;
 427         unsigned long sg_off;
 428         struct scatterlist *sg;
 429         int ret = 0;
 430
 431         rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from));
 432
 433         /* now allocate and copy in the data payload.  */
 434         sg = rm->data.op_sg;
 435         sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */
 436
 437         if (zcopy)
 438                 return rds_message_zcopy_from_user(rm, from);
 439
 440         while (iov_iter_count(from)) {
 441                 if (!sg_page(sg)) {
 442                         ret = rds_page_remainder_alloc(sg, iov_iter_count(from),
 443                                                        GFP_HIGHUSER);
 444                         if (ret)
 445                                 return ret;
 446                         rm->data.op_nents++;
 447                         sg_off = 0;
 448                 }
 449
 450                 to_copy = min_t(unsigned long, iov_iter_count(from),
 451                                 sg->length - sg_off);
 452
 453                 rds_stats_add(s_copy_from_user, to_copy);
 454                 nbytes = copy_page_from_iter(sg_page(sg), sg->offset + sg_off,
 455                                              to_copy, from);
 456                 if (nbytes != to_copy)
 457                         return -EFAULT;
 458
 459                 sg_off += to_copy;
 460
 461                 if (sg_off == sg->length)
 462                         sg++;
 463         }
 464
 465         return ret;
 466 }
 467
 468 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to)
 469 {
 470         struct rds_message *rm;
 471         struct scatterlist *sg;
 472         unsigned long to_copy;
 473         unsigned long vec_off;
 474         int copied;
 475         int ret;
 476         u32 len;
 477
 478         rm = container_of(inc, struct rds_message, m_inc);
 479         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
 480
 481         sg = rm->data.op_sg;
 482         vec_off = 0;
 483         copied = 0;
 484
 485         while (iov_iter_count(to) && copied < len) {
 486                 to_copy = min_t(unsigned long, iov_iter_count(to),
 487                                 sg->length - vec_off);
 488                 to_copy = min_t(unsigned long, to_copy, len - copied);
 489
 490                 rds_stats_add(s_copy_to_user, to_copy);
 491                 ret = copy_page_to_iter(sg_page(sg), sg->offset + vec_off,
 492                                         to_copy, to);
 493                 if (ret != to_copy)
 494                         return -EFAULT;
 495
 496                 vec_off += to_copy;
 497                 copied += to_copy;
 498
 499                 if (vec_off == sg->length) {
 500                         vec_off = 0;
 501                         sg++;
 502                 }
 503         }
 504
 505         return copied;
 506 }
 507
 508 /*
 509  * If the message is still on the send queue, wait until the transport
 510  * is done with it. This is particularly important for RDMA operations.
 511  */
 512 void rds_message_wait(struct rds_message *rm)
 513 {
 514         wait_event_interruptible(rm->m_flush_wait,
 515                         !test_bit(RDS_MSG_MAPPED, &rm->m_flags));
 516 }
 517
 518 void rds_message_unmapped(struct rds_message *rm)
 519 {
 520         clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
 521         wake_up_interruptible(&rm->m_flush_wait);
 522 }
 523 EXPORT_SYMBOL_GPL(rds_message_unmapped);