| 1 | |
| 2 | * Copyright (c) 2007, 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/pagemap.h>
|
| 34 | #include <linux/slab.h>
|
| 35 | #include <linux/rbtree.h>
|
| 36 | #include <linux/dma-mapping.h>
|
| 37 |
|
| 38 | #include "rds.h"
|
| 39 |
|
| 40 | |
| 41 | * XXX
|
| 42 | * - build with sparse
|
| 43 | * - should we detect duplicate keys on a socket? hmm.
|
| 44 | * - an rdma is an mlock, apply rlimit?
|
| 45 |
|
| 46 |
|
| 47 | |
| 48 | * get the number of pages by looking at the page indices that the start and
|
| 49 | * end addresses fall in.
|
| 50 | *
|
| 51 | * Returns 0 if the vec is invalid. It is invalid if the number of bytes
|
| 52 | * causes the address to wrap or overflows an unsigned int. This comes
|
| 53 | * from being stored in the 'length' member of 'struct scatterlist'.
|
| 54 |
|
| 55 | static unsigned int rds_pages_in_vec(struct rds_iovec *vec)
|
| 56 | {
|
| 57 | if ((vec->addr + vec->bytes <= vec->addr) ||
|
| 58 | (vec->bytes > (u64)UINT_MAX))
|
| 59 | return 0;
|
| 60 |
|
| 61 | return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) -
|
| 62 | (vec->addr >> PAGE_SHIFT);
|
| 63 | }
|
| 64 |
|
| 65 | static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key,
|
| 66 | struct rds_mr *insert)
|
| 67 | {
|
| 68 | struct rb_node **p = &root->rb_node;
|
| 69 | struct rb_node *parent = NULL;
|
| 70 | struct rds_mr *mr;
|
| 71 |
|
| 72 | while (*p) {
|
| 73 | parent = *p;
|
| 74 | mr = rb_entry(parent, struct rds_mr, r_rb_node);
|
| 75 |
|
| 76 | if (key < mr->r_key)
|
| 77 | p = &(*p)->rb_left;
|
| 78 | else if (key > mr->r_key)
|
| 79 | p = &(*p)->rb_right;
|
| 80 | else
|
| 81 | return mr;
|
| 82 | }
|
| 83 |
|
| 84 | if (insert) {
|
| 85 | rb_link_node(&insert->r_rb_node, parent, p);
|
| 86 | rb_insert_color(&insert->r_rb_node, root);
|
| 87 | kref_get(&insert->r_kref);
|
| 88 | }
|
| 89 | return NULL;
|
| 90 | }
|
| 91 |
|
| 92 | |
| 93 | * Destroy the transport-specific part of a MR.
|
| 94 |
|
| 95 | static void rds_destroy_mr(struct rds_mr *mr)
|
| 96 | {
|
| 97 | struct rds_sock *rs = mr->r_sock;
|
| 98 | void *trans_private = NULL;
|
| 99 | unsigned long flags;
|
| 100 |
|
| 101 | rdsdebug("RDS: destroy mr key is %x refcnt %u\n",
|
| 102 | mr->r_key, kref_read(&mr->r_kref));
|
| 103 |
|
| 104 | spin_lock_irqsave(&rs->rs_rdma_lock, flags);
|
| 105 | if (!RB_EMPTY_NODE(&mr->r_rb_node))
|
| 106 | rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
|
| 107 | trans_private = mr->r_trans_private;
|
| 108 | mr->r_trans_private = NULL;
|
| 109 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
|
| 110 |
|
| 111 | if (trans_private)
|
| 112 | mr->r_trans->free_mr(trans_private, mr->r_invalidate);
|
| 113 | }
|
| 114 |
|
| 115 | void __rds_put_mr_final(struct kref *kref)
|
| 116 | {
|
| 117 | struct rds_mr *mr = container_of(kref, struct rds_mr, r_kref);
|
| 118 |
|
| 119 | rds_destroy_mr(mr);
|
| 120 | sock_put(rds_rs_to_sk(mr->r_sock));
|
| 121 | kfree(mr);
|
| 122 | }
|
| 123 |
|
| 124 | |
| 125 | * By the time this is called we can't have any more ioctls called on
|
| 126 | * the socket so we don't need to worry about racing with others.
|
| 127 |
|
| 128 | void rds_rdma_drop_keys(struct rds_sock *rs)
|
| 129 | {
|
| 130 | struct rds_mr *mr;
|
| 131 | struct rb_node *node;
|
| 132 | unsigned long flags;
|
| 133 |
|
| 134 |
|
| 135 | spin_lock_irqsave(&rs->rs_rdma_lock, flags);
|
| 136 | while ((node = rb_first(&rs->rs_rdma_keys))) {
|
| 137 | mr = rb_entry(node, struct rds_mr, r_rb_node);
|
| 138 | if (mr->r_trans == rs->rs_transport)
|
| 139 | mr->r_invalidate = 0;
|
| 140 | rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
|
| 141 | RB_CLEAR_NODE(&mr->r_rb_node);
|
| 142 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
|
| 143 | kref_put(&mr->r_kref, __rds_put_mr_final);
|
| 144 | spin_lock_irqsave(&rs->rs_rdma_lock, flags);
|
| 145 | }
|
| 146 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
|
| 147 |
|
| 148 | if (rs->rs_transport && rs->rs_transport->flush_mrs)
|
| 149 | rs->rs_transport->flush_mrs();
|
| 150 | }
|
| 151 |
|
| 152 | |
| 153 | * Helper function to pin user pages.
|
| 154 |
|
| 155 | static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
|
| 156 | struct page **pages, int write)
|
| 157 | {
|
| 158 | unsigned int gup_flags = FOLL_LONGTERM;
|
| 159 | int ret;
|
| 160 |
|
| 161 | if (write)
|
| 162 | gup_flags |= FOLL_WRITE;
|
| 163 |
|
| 164 | ret = pin_user_pages_fast(user_addr, nr_pages, gup_flags, pages);
|
| 165 | if (ret >= 0 && ret < nr_pages) {
|
| 166 | unpin_user_pages(pages, ret);
|
| 167 | ret = -EFAULT;
|
| 168 | }
|
| 169 |
|
| 170 | return ret;
|
| 171 | }
|
| 172 |
|
| 173 | static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
|
| 174 | u64 *cookie_ret, struct rds_mr **mr_ret,
|
| 175 | struct rds_conn_path *cp)
|
| 176 | {
|
| 177 | struct rds_mr *mr = NULL, *found;
|
| 178 | struct scatterlist *sg = NULL;
|
| 179 | unsigned int nr_pages;
|
| 180 | struct page **pages = NULL;
|
| 181 | void *trans_private;
|
| 182 | unsigned long flags;
|
| 183 | rds_rdma_cookie_t cookie;
|
| 184 | unsigned int nents = 0;
|
| 185 | int need_odp = 0;
|
| 186 | long i;
|
| 187 | int ret;
|
| 188 |
|
| 189 | if (ipv6_addr_any(&rs->rs_bound_addr) || !rs->rs_transport) {
|
| 190 | ret = -ENOTCONN;
|
| 191 | goto out;
|
| 192 | }
|
| 193 |
|
| 194 | if (!rs->rs_transport->get_mr) {
|
| 195 | ret = -EOPNOTSUPP;
|
| 196 | goto out;
|
| 197 | }
|
| 198 |
|
| 199 | |
| 200 | * region causes an integer overflow, return error.
|
| 201 |
|
| 202 | if (((args->vec.addr + args->vec.bytes) < args->vec.addr) ||
|
| 203 | PAGE_ALIGN(args->vec.addr + args->vec.bytes) <
|
| 204 | (args->vec.addr + args->vec.bytes)) {
|
| 205 | ret = -EINVAL;
|
| 206 | goto out;
|
| 207 | }
|
| 208 |
|
| 209 | if (!can_do_mlock()) {
|
| 210 | ret = -EPERM;
|
| 211 | goto out;
|
| 212 | }
|
| 213 |
|
| 214 | nr_pages = rds_pages_in_vec(&args->vec);
|
| 215 | if (nr_pages == 0) {
|
| 216 | ret = -EINVAL;
|
| 217 | goto out;
|
| 218 | }
|
| 219 |
|
| 220 | |
| 221 | * To account for unaligned mr regions, subtract one from nr_pages
|
| 222 |
|
| 223 | if ((nr_pages - 1) > (RDS_MAX_MSG_SIZE >> PAGE_SHIFT)) {
|
| 224 | ret = -EMSGSIZE;
|
| 225 | goto out;
|
| 226 | }
|
| 227 |
|
| 228 | rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n",
|
| 229 | args->vec.addr, args->vec.bytes, nr_pages);
|
| 230 |
|
| 231 |
|
| 232 | pages = kzalloc_objs(struct page *, nr_pages);
|
| 233 | if (!pages) {
|
| 234 | ret = -ENOMEM;
|
| 235 | goto out;
|
| 236 | }
|
| 237 |
|
| 238 | mr = kzalloc_obj(struct rds_mr);
|
| 239 | if (!mr) {
|
| 240 | ret = -ENOMEM;
|
| 241 | goto out;
|
| 242 | }
|
| 243 |
|
| 244 | kref_init(&mr->r_kref);
|
| 245 | RB_CLEAR_NODE(&mr->r_rb_node);
|
| 246 | mr->r_trans = rs->rs_transport;
|
| 247 | |
| 248 | * until the final kref is dropped in __rds_put_mr_final().
|
| 249 |
|
| 250 | mr->r_sock = rs;
|
| 251 | sock_hold(rds_rs_to_sk(rs));
|
| 252 |
|
| 253 | if (args->flags & RDS_RDMA_USE_ONCE)
|
| 254 | mr->r_use_once = 1;
|
| 255 | if (args->flags & RDS_RDMA_INVALIDATE)
|
| 256 | mr->r_invalidate = 1;
|
| 257 | if (args->flags & RDS_RDMA_READWRITE)
|
| 258 | mr->r_write = 1;
|
| 259 |
|
| 260 | |
| 261 | * Pin the pages that make up the user buffer and transfer the page
|
| 262 | * pointers to the mr's sg array. We check to see if we've mapped
|
| 263 | * the whole region after transferring the partial page references
|
| 264 | * to the sg array so that we can have one page ref cleanup path.
|
| 265 | *
|
| 266 | * For now we have no flag that tells us whether the mapping is
|
| 267 | * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to
|
| 268 | * the zero page.
|
| 269 |
|
| 270 | ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
|
| 271 | if (ret == -EOPNOTSUPP) {
|
| 272 | need_odp = 1;
|
| 273 | } else if (ret <= 0) {
|
| 274 | goto out;
|
| 275 | } else {
|
| 276 | nents = ret;
|
| 277 | sg = kmalloc_objs(*sg, nents);
|
| 278 | if (!sg) {
|
| 279 | ret = -ENOMEM;
|
| 280 | goto out;
|
| 281 | }
|
| 282 | WARN_ON(!nents);
|
| 283 | sg_init_table(sg, nents);
|
| 284 |
|
| 285 |
|
| 286 | for (i = 0 ; i < nents; i++)
|
| 287 | sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
|
| 288 |
|
| 289 | rdsdebug("RDS: trans_private nents is %u\n", nents);
|
| 290 | }
|
| 291 | |
| 292 | * s/g list is now owned by the MR.
|
| 293 | * Note that dma_map() implies that pending writes are
|
| 294 | * flushed to RAM, so no dma_sync is needed here.
|
| 295 | trans_private = rs->rs_transport->get_mr(
|
| 296 | sg, nents, rs, &mr->r_key, cp ? cp->cp_conn : NULL,
|
| 297 | args->vec.addr, args->vec.bytes,
|
| 298 | need_odp ? ODP_ZEROBASED : ODP_NOT_NEEDED);
|
| 299 |
|
| 300 | if (IS_ERR(trans_private)) {
|
| 301 | |
| 302 | * to release anything.
|
| 303 |
|
| 304 | if (!need_odp) {
|
| 305 | unpin_user_pages(pages, nr_pages);
|
| 306 | kfree(sg);
|
| 307 | }
|
| 308 | ret = PTR_ERR(trans_private);
|
| 309 |
|
| 310 | if (ret == -ENODEV && cp)
|
| 311 | rds_conn_connect_if_down(cp->cp_conn);
|
| 312 | goto out;
|
| 313 | }
|
| 314 |
|
| 315 | mr->r_trans_private = trans_private;
|
| 316 |
|
| 317 | rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n",
|
| 318 | mr->r_key, (void *)(unsigned long) args->cookie_addr);
|
| 319 |
|
| 320 | |
| 321 | * map page aligned regions. So we keep the offset, and build
|
| 322 | * a 64bit cookie containing <R_Key, offset> and pass that
|
| 323 | * around.
|
| 324 | if (need_odp)
|
| 325 | cookie = rds_rdma_make_cookie(mr->r_key, 0);
|
| 326 | else
|
| 327 | cookie = rds_rdma_make_cookie(mr->r_key,
|
| 328 | args->vec.addr & ~PAGE_MASK);
|
| 329 | if (cookie_ret)
|
| 330 | *cookie_ret = cookie;
|
| 331 |
|
| 332 | if (args->cookie_addr &&
|
| 333 | put_user(cookie, (u64 __user *)(unsigned long)args->cookie_addr)) {
|
| 334 | ret = -EFAULT;
|
| 335 | goto out;
|
| 336 | }
|
| 337 |
|
| 338 | |
| 339 | * reference count.
|
| 340 | spin_lock_irqsave(&rs->rs_rdma_lock, flags);
|
| 341 | found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr);
|
| 342 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
|
| 343 |
|
| 344 | BUG_ON(found && found != mr);
|
| 345 |
|
| 346 | rdsdebug("RDS: get_mr key is %x\n", mr->r_key);
|
| 347 | if (mr_ret) {
|
| 348 | kref_get(&mr->r_kref);
|
| 349 | *mr_ret = mr;
|
| 350 | }
|
| 351 |
|
| 352 | ret = 0;
|
| 353 | out:
|
| 354 | kfree(pages);
|
| 355 | if (mr)
|
| 356 | kref_put(&mr->r_kref, __rds_put_mr_final);
|
| 357 | return ret;
|
| 358 | }
|
| 359 |
|
| 360 | int rds_get_mr(struct rds_sock *rs, sockptr_t optval, int optlen)
|
| 361 | {
|
| 362 | struct rds_get_mr_args args;
|
| 363 |
|
| 364 | if (optlen != sizeof(struct rds_get_mr_args))
|
| 365 | return -EINVAL;
|
| 366 |
|
| 367 | if (copy_from_sockptr(&args, optval, sizeof(struct rds_get_mr_args)))
|
| 368 | return -EFAULT;
|
| 369 |
|
| 370 | return __rds_rdma_map(rs, &args, NULL, NULL, NULL);
|
| 371 | }
|
| 372 |
|
| 373 | int rds_get_mr_for_dest(struct rds_sock *rs, sockptr_t optval, int optlen)
|
| 374 | {
|
| 375 | struct rds_get_mr_for_dest_args args;
|
| 376 | struct rds_get_mr_args new_args;
|
| 377 |
|
| 378 | if (optlen != sizeof(struct rds_get_mr_for_dest_args))
|
| 379 | return -EINVAL;
|
| 380 |
|
| 381 | if (copy_from_sockptr(&args, optval,
|
| 382 | sizeof(struct rds_get_mr_for_dest_args)))
|
| 383 | return -EFAULT;
|
| 384 |
|
| 385 | |
| 386 | * Initially, just behave like get_mr().
|
| 387 | * TODO: Implement get_mr as wrapper around this
|
| 388 | * and deprecate it.
|
| 389 |
|
| 390 | new_args.vec = args.vec;
|
| 391 | new_args.cookie_addr = args.cookie_addr;
|
| 392 | new_args.flags = args.flags;
|
| 393 |
|
| 394 | return __rds_rdma_map(rs, &new_args, NULL, NULL, NULL);
|
| 395 | }
|
| 396 |
|
| 397 | |
| 398 | * Free the MR indicated by the given R_Key
|
| 399 |
|
| 400 | int rds_free_mr(struct rds_sock *rs, sockptr_t optval, int optlen)
|
| 401 | {
|
| 402 | struct rds_free_mr_args args;
|
| 403 | struct rds_mr *mr;
|
| 404 | unsigned long flags;
|
| 405 |
|
| 406 | if (optlen != sizeof(struct rds_free_mr_args))
|
| 407 | return -EINVAL;
|
| 408 |
|
| 409 | if (copy_from_sockptr(&args, optval, sizeof(struct rds_free_mr_args)))
|
| 410 | return -EFAULT;
|
| 411 |
|
| 412 |
|
| 413 | if (args.cookie == 0) {
|
| 414 | if (!rs->rs_transport || !rs->rs_transport->flush_mrs)
|
| 415 | return -EINVAL;
|
| 416 | rs->rs_transport->flush_mrs();
|
| 417 | return 0;
|
| 418 | }
|
| 419 |
|
| 420 | |
| 421 | * so nobody else finds it.
|
| 422 | * This should also prevent races with rds_rdma_unuse.
|
| 423 |
|
| 424 | spin_lock_irqsave(&rs->rs_rdma_lock, flags);
|
| 425 | mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL);
|
| 426 | if (mr) {
|
| 427 | rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
|
| 428 | RB_CLEAR_NODE(&mr->r_rb_node);
|
| 429 | if (args.flags & RDS_RDMA_INVALIDATE)
|
| 430 | mr->r_invalidate = 1;
|
| 431 | }
|
| 432 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
|
| 433 |
|
| 434 | if (!mr)
|
| 435 | return -EINVAL;
|
| 436 |
|
| 437 | kref_put(&mr->r_kref, __rds_put_mr_final);
|
| 438 | return 0;
|
| 439 | }
|
| 440 |
|
| 441 | |
| 442 | * This is called when we receive an extension header that
|
| 443 | * tells us this MR was used. It allows us to implement
|
| 444 | * use_once semantics
|
| 445 |
|
| 446 | void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force)
|
| 447 | {
|
| 448 | struct rds_mr *mr;
|
| 449 | unsigned long flags;
|
| 450 | int zot_me = 0;
|
| 451 |
|
| 452 | spin_lock_irqsave(&rs->rs_rdma_lock, flags);
|
| 453 | mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
|
| 454 | if (!mr) {
|
| 455 | pr_debug("rds: trying to unuse MR with unknown r_key %u!\n",
|
| 456 | r_key);
|
| 457 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
|
| 458 | return;
|
| 459 | }
|
| 460 |
|
| 461 | |
| 462 | * sync_mr() below.
|
| 463 |
|
| 464 | kref_get(&mr->r_kref);
|
| 465 |
|
| 466 | |
| 467 | * that no other thread can find it and free it.
|
| 468 |
|
| 469 | if (mr->r_use_once || force) {
|
| 470 | rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys);
|
| 471 | RB_CLEAR_NODE(&mr->r_rb_node);
|
| 472 | zot_me = 1;
|
| 473 | }
|
| 474 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
|
| 475 |
|
| 476 | |
| 477 | * Note we could avoid this if the operation was a RDMA READ,
|
| 478 | * but at this point we can't tell.
|
| 479 | if (mr->r_trans->sync_mr)
|
| 480 | mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE);
|
| 481 |
|
| 482 |
|
| 483 | kref_put(&mr->r_kref, __rds_put_mr_final);
|
| 484 |
|
| 485 | |
| 486 | * trigger an async flush.
|
| 487 | if (zot_me)
|
| 488 | kref_put(&mr->r_kref, __rds_put_mr_final);
|
| 489 | }
|
| 490 |
|
| 491 | void rds_rdma_op_unpin_pages(struct rm_rdma_op *ro)
|
| 492 | {
|
| 493 | unsigned int i;
|
| 494 |
|
| 495 | for (i = 0; i < ro->op_nents; i++) {
|
| 496 | struct page *page = sg_page(&ro->op_sg[i]);
|
| 497 |
|
| 498 | |
| 499 | * is the case for a RDMA_READ which copies from remote
|
| 500 | * to local memory
|
| 501 |
|
| 502 | unpin_user_pages_dirty_lock(&page, 1, !ro->op_write);
|
| 503 | }
|
| 504 | }
|
| 505 |
|
| 506 | void rds_rdma_free_op(struct rm_rdma_op *ro)
|
| 507 | {
|
| 508 | if (ro->op_odp_mr) {
|
| 509 | kref_put(&ro->op_odp_mr->r_kref, __rds_put_mr_final);
|
| 510 | } else if (in_task() || ro->op_write) {
|
| 511 | |
| 512 | * side; unpinning without dirtying does not sleep.
|
| 513 |
|
| 514 | rds_rdma_op_unpin_pages(ro);
|
| 515 | } else {
|
| 516 | |
| 517 | * pinned and have rds_message_put() finish the unpin
|
| 518 | * from process context.
|
| 519 |
|
| 520 | ro->op_unpin_deferred = 1;
|
| 521 | }
|
| 522 |
|
| 523 | kfree(ro->op_notifier);
|
| 524 | ro->op_notifier = NULL;
|
| 525 | ro->op_active = 0;
|
| 526 | ro->op_odp_mr = NULL;
|
| 527 | }
|
| 528 |
|
| 529 | void rds_atomic_op_unpin_page(struct rm_atomic_op *ao)
|
| 530 | {
|
| 531 | struct page *page = sg_page(ao->op_sg);
|
| 532 |
|
| 533 | |
| 534 | * is the case for a RDMA_READ which copies from remote
|
| 535 | * to local memory
|
| 536 | unpin_user_pages_dirty_lock(&page, 1, true);
|
| 537 | }
|
| 538 |
|
| 539 | void rds_atomic_free_op(struct rm_atomic_op *ao)
|
| 540 | {
|
| 541 | if (in_task()) {
|
| 542 | rds_atomic_op_unpin_page(ao);
|
| 543 | } else {
|
| 544 | |
| 545 | * pinned and have rds_message_put() finish the unpin
|
| 546 | * from process context.
|
| 547 |
|
| 548 | ao->op_unpin_deferred = 1;
|
| 549 | }
|
| 550 |
|
| 551 | kfree(ao->op_notifier);
|
| 552 | ao->op_notifier = NULL;
|
| 553 | ao->op_active = 0;
|
| 554 | }
|
| 555 |
|
| 556 |
|
| 557 | |
| 558 | * Count the number of pages needed to describe an incoming iovec array.
|
| 559 |
|
| 560 | static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs)
|
| 561 | {
|
| 562 | int tot_pages = 0;
|
| 563 | unsigned int nr_pages;
|
| 564 | unsigned int i;
|
| 565 |
|
| 566 |
|
| 567 | for (i = 0; i < nr_iovecs; i++) {
|
| 568 | nr_pages = rds_pages_in_vec(&iov[i]);
|
| 569 | if (nr_pages == 0)
|
| 570 | return -EINVAL;
|
| 571 |
|
| 572 | tot_pages += nr_pages;
|
| 573 |
|
| 574 | |
| 575 | * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
|
| 576 | * so tot_pages cannot overflow without first going negative.
|
| 577 |
|
| 578 | if (tot_pages < 0)
|
| 579 | return -EINVAL;
|
| 580 | }
|
| 581 |
|
| 582 | return tot_pages;
|
| 583 | }
|
| 584 |
|
| 585 | int rds_rdma_extra_size(struct rds_rdma_args *args,
|
| 586 | struct rds_iov_vector *iov)
|
| 587 | {
|
| 588 | struct rds_iovec *vec;
|
| 589 | struct rds_iovec __user *local_vec;
|
| 590 | int tot_pages = 0;
|
| 591 | unsigned int nr_pages;
|
| 592 | unsigned int i;
|
| 593 |
|
| 594 | local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr;
|
| 595 |
|
| 596 | if (args->nr_local == 0)
|
| 597 | return -EINVAL;
|
| 598 |
|
| 599 | if (args->nr_local > UIO_MAXIOV)
|
| 600 | return -EMSGSIZE;
|
| 601 |
|
| 602 | iov->iov = kzalloc_objs(struct rds_iovec, args->nr_local);
|
| 603 | if (!iov->iov)
|
| 604 | return -ENOMEM;
|
| 605 |
|
| 606 | vec = &iov->iov[0];
|
| 607 |
|
| 608 | if (copy_from_user(vec, local_vec, args->nr_local *
|
| 609 | sizeof(struct rds_iovec)))
|
| 610 | return -EFAULT;
|
| 611 | iov->len = args->nr_local;
|
| 612 |
|
| 613 |
|
| 614 | for (i = 0; i < args->nr_local; i++, vec++) {
|
| 615 |
|
| 616 | nr_pages = rds_pages_in_vec(vec);
|
| 617 | if (nr_pages == 0)
|
| 618 | return -EINVAL;
|
| 619 |
|
| 620 | tot_pages += nr_pages;
|
| 621 |
|
| 622 | |
| 623 | * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1,
|
| 624 | * so tot_pages cannot overflow without first going negative.
|
| 625 |
|
| 626 | if (tot_pages < 0)
|
| 627 | return -EINVAL;
|
| 628 | }
|
| 629 |
|
| 630 | return tot_pages * sizeof(struct scatterlist);
|
| 631 | }
|
| 632 |
|
| 633 | |
| 634 | * The application asks for a RDMA transfer.
|
| 635 | * Extract all arguments and set up the rdma_op
|
| 636 |
|
| 637 | int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
|
| 638 | struct cmsghdr *cmsg,
|
| 639 | struct rds_iov_vector *vec)
|
| 640 | {
|
| 641 | struct rds_rdma_args *args;
|
| 642 | struct rm_rdma_op *op = &rm->rdma;
|
| 643 | int nr_pages;
|
| 644 | unsigned int nr_bytes;
|
| 645 | struct page **pages = NULL;
|
| 646 | struct rds_iovec *iovs;
|
| 647 | unsigned int i, j;
|
| 648 | int ret = 0;
|
| 649 | bool odp_supported = true;
|
| 650 |
|
| 651 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
|
| 652 | || rm->rdma.op_active)
|
| 653 | return -EINVAL;
|
| 654 |
|
| 655 | args = CMSG_DATA(cmsg);
|
| 656 |
|
| 657 | if (ipv6_addr_any(&rs->rs_bound_addr)) {
|
| 658 | ret = -ENOTCONN;
|
| 659 | goto out_ret;
|
| 660 | }
|
| 661 |
|
| 662 | if (args->nr_local > UIO_MAXIOV) {
|
| 663 | ret = -EMSGSIZE;
|
| 664 | goto out_ret;
|
| 665 | }
|
| 666 |
|
| 667 | if (vec->len != args->nr_local) {
|
| 668 | ret = -EINVAL;
|
| 669 | goto out_ret;
|
| 670 | }
|
| 671 |
|
| 672 | if (args->nr_local != 1)
|
| 673 | odp_supported = false;
|
| 674 |
|
| 675 | iovs = vec->iov;
|
| 676 |
|
| 677 | nr_pages = rds_rdma_pages(iovs, args->nr_local);
|
| 678 | if (nr_pages < 0) {
|
| 679 | ret = -EINVAL;
|
| 680 | goto out_ret;
|
| 681 | }
|
| 682 |
|
| 683 | pages = kzalloc_objs(struct page *, nr_pages);
|
| 684 | if (!pages) {
|
| 685 | ret = -ENOMEM;
|
| 686 | goto out_ret;
|
| 687 | }
|
| 688 |
|
| 689 | op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
|
| 690 | op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
|
| 691 | op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
|
| 692 | op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
|
| 693 | op->op_active = 1;
|
| 694 | op->op_recverr = rs->rs_recverr;
|
| 695 | op->op_odp_mr = NULL;
|
| 696 |
|
| 697 | WARN_ON(!nr_pages);
|
| 698 | op->op_sg = rds_message_alloc_sgs(rm, nr_pages);
|
| 699 | if (IS_ERR(op->op_sg)) {
|
| 700 | ret = PTR_ERR(op->op_sg);
|
| 701 | goto out_pages;
|
| 702 | }
|
| 703 |
|
| 704 | if (op->op_notify || op->op_recverr) {
|
| 705 | |
| 706 | * we don't want to do that in the completion handler. We
|
| 707 | * would have to use GFP_ATOMIC there, and don't want to deal
|
| 708 | * with failed allocations.
|
| 709 |
|
| 710 | op->op_notifier = kmalloc_obj(struct rds_notifier);
|
| 711 | if (!op->op_notifier) {
|
| 712 | ret = -ENOMEM;
|
| 713 | goto out_pages;
|
| 714 | }
|
| 715 | op->op_notifier->n_user_token = args->user_token;
|
| 716 | op->op_notifier->n_status = RDS_RDMA_SUCCESS;
|
| 717 | }
|
| 718 |
|
| 719 | |
| 720 | * optionally an offset into it. This is how we implement RDMA into
|
| 721 | * unaligned memory.
|
| 722 | * When setting up the RDMA, we need to add that offset to the
|
| 723 | * destination address (which is really an offset into the MR)
|
| 724 | * FIXME: We may want to move this into ib_rdma.c
|
| 725 |
|
| 726 | op->op_rkey = rds_rdma_cookie_key(args->cookie);
|
| 727 | op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
|
| 728 |
|
| 729 | nr_bytes = 0;
|
| 730 |
|
| 731 | rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",
|
| 732 | (unsigned long long)args->nr_local,
|
| 733 | (unsigned long long)args->remote_vec.addr,
|
| 734 | op->op_rkey);
|
| 735 |
|
| 736 | for (i = 0; i < args->nr_local; i++) {
|
| 737 | struct rds_iovec *iov = &iovs[i];
|
| 738 |
|
| 739 | unsigned int nr = rds_pages_in_vec(iov);
|
| 740 |
|
| 741 | rs->rs_user_addr = iov->addr;
|
| 742 | rs->rs_user_bytes = iov->bytes;
|
| 743 |
|
| 744 | |
| 745 | * If it's a READ operation, we need to pin the pages for writing.
|
| 746 |
|
| 747 | ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
|
| 748 | if ((!odp_supported && ret <= 0) ||
|
| 749 | (odp_supported && ret <= 0 && ret != -EOPNOTSUPP))
|
| 750 | goto out_pages;
|
| 751 |
|
| 752 | if (ret == -EOPNOTSUPP) {
|
| 753 | struct rds_mr *local_odp_mr;
|
| 754 |
|
| 755 | if (!rs->rs_transport->get_mr) {
|
| 756 | ret = -EOPNOTSUPP;
|
| 757 | goto out_pages;
|
| 758 | }
|
| 759 | local_odp_mr = kzalloc_obj(*local_odp_mr);
|
| 760 | if (!local_odp_mr) {
|
| 761 | ret = -ENOMEM;
|
| 762 | goto out_pages;
|
| 763 | }
|
| 764 | RB_CLEAR_NODE(&local_odp_mr->r_rb_node);
|
| 765 | kref_init(&local_odp_mr->r_kref);
|
| 766 | local_odp_mr->r_trans = rs->rs_transport;
|
| 767 | |
| 768 | * reference is held until the final kref is
|
| 769 | * dropped in __rds_put_mr_final().
|
| 770 |
|
| 771 | local_odp_mr->r_sock = rs;
|
| 772 | sock_hold(rds_rs_to_sk(rs));
|
| 773 | local_odp_mr->r_trans_private =
|
| 774 | rs->rs_transport->get_mr(
|
| 775 | NULL, 0, rs, &local_odp_mr->r_key, NULL,
|
| 776 | iov->addr, iov->bytes, ODP_VIRTUAL);
|
| 777 | if (IS_ERR(local_odp_mr->r_trans_private)) {
|
| 778 | ret = PTR_ERR(local_odp_mr->r_trans_private);
|
| 779 | rdsdebug("get_mr ret %d %p\"", ret,
|
| 780 | local_odp_mr->r_trans_private);
|
| 781 | local_odp_mr->r_trans_private = NULL;
|
| 782 | kref_put(&local_odp_mr->r_kref,
|
| 783 | __rds_put_mr_final);
|
| 784 | ret = -EOPNOTSUPP;
|
| 785 | goto out_pages;
|
| 786 | }
|
| 787 | rdsdebug("Need odp; local_odp_mr %p trans_private %p\n",
|
| 788 | local_odp_mr, local_odp_mr->r_trans_private);
|
| 789 | op->op_odp_mr = local_odp_mr;
|
| 790 | op->op_odp_addr = iov->addr;
|
| 791 | }
|
| 792 |
|
| 793 | rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
|
| 794 | nr_bytes, nr, iov->bytes, iov->addr);
|
| 795 |
|
| 796 | nr_bytes += iov->bytes;
|
| 797 |
|
| 798 | for (j = 0; j < nr; j++) {
|
| 799 | unsigned int offset = iov->addr & ~PAGE_MASK;
|
| 800 | struct scatterlist *sg;
|
| 801 |
|
| 802 | sg = &op->op_sg[op->op_nents + j];
|
| 803 | sg_set_page(sg, pages[j],
|
| 804 | min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
|
| 805 | offset);
|
| 806 |
|
| 807 | sg_dma_len(sg) = sg->length;
|
| 808 | rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
|
| 809 | sg->offset, sg->length, iov->addr, iov->bytes);
|
| 810 |
|
| 811 | iov->addr += sg->length;
|
| 812 | iov->bytes -= sg->length;
|
| 813 | }
|
| 814 |
|
| 815 | op->op_nents += nr;
|
| 816 | }
|
| 817 |
|
| 818 | if (nr_bytes > args->remote_vec.bytes) {
|
| 819 | rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
|
| 820 | nr_bytes,
|
| 821 | (unsigned int) args->remote_vec.bytes);
|
| 822 | ret = -EINVAL;
|
| 823 | goto out_pages;
|
| 824 | }
|
| 825 | op->op_bytes = nr_bytes;
|
| 826 | ret = 0;
|
| 827 |
|
| 828 | out_pages:
|
| 829 | kfree(pages);
|
| 830 | out_ret:
|
| 831 | if (ret)
|
| 832 | rds_rdma_free_op(op);
|
| 833 | else
|
| 834 | rds_stats_inc(s_send_rdma);
|
| 835 |
|
| 836 | return ret;
|
| 837 | }
|
| 838 |
|
| 839 | |
| 840 | * The application wants us to pass an RDMA destination (aka MR)
|
| 841 | * to the remote
|
| 842 |
|
| 843 | int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm,
|
| 844 | struct cmsghdr *cmsg)
|
| 845 | {
|
| 846 | unsigned long flags;
|
| 847 | struct rds_mr *mr;
|
| 848 | u32 r_key;
|
| 849 | int err = 0;
|
| 850 |
|
| 851 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) ||
|
| 852 | rm->m_rdma_cookie != 0)
|
| 853 | return -EINVAL;
|
| 854 |
|
| 855 | memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie));
|
| 856 |
|
| 857 | |
| 858 | * application has written to the buffer, so we need to explicitly
|
| 859 | * flush those writes to RAM. Otherwise the HCA may not see them
|
| 860 | * when doing a DMA from that buffer.
|
| 861 |
|
| 862 | r_key = rds_rdma_cookie_key(rm->m_rdma_cookie);
|
| 863 |
|
| 864 | spin_lock_irqsave(&rs->rs_rdma_lock, flags);
|
| 865 | mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL);
|
| 866 | if (!mr)
|
| 867 | err = -EINVAL;
|
| 868 | else
|
| 869 | kref_get(&mr->r_kref);
|
| 870 | spin_unlock_irqrestore(&rs->rs_rdma_lock, flags);
|
| 871 |
|
| 872 | if (mr) {
|
| 873 | mr->r_trans->sync_mr(mr->r_trans_private,
|
| 874 | DMA_TO_DEVICE);
|
| 875 | rm->rdma.op_rdma_mr = mr;
|
| 876 | }
|
| 877 | return err;
|
| 878 | }
|
| 879 |
|
| 880 | |
| 881 | * The application passes us an address range it wants to enable RDMA
|
| 882 | * to/from. We map the area, and save the <R_Key,offset> pair
|
| 883 | * in rm->m_rdma_cookie. This causes it to be sent along to the peer
|
| 884 | * in an extension header.
|
| 885 |
|
| 886 | int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm,
|
| 887 | struct cmsghdr *cmsg)
|
| 888 | {
|
| 889 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) ||
|
| 890 | rm->m_rdma_cookie != 0)
|
| 891 | return -EINVAL;
|
| 892 |
|
| 893 | return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie,
|
| 894 | &rm->rdma.op_rdma_mr, rm->m_conn_path);
|
| 895 | }
|
| 896 |
|
| 897 | |
| 898 | * Fill in rds_message for an atomic request.
|
| 899 |
|
| 900 | int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm,
|
| 901 | struct cmsghdr *cmsg)
|
| 902 | {
|
| 903 | struct page *page = NULL;
|
| 904 | struct rds_atomic_args *args;
|
| 905 | int ret = 0;
|
| 906 |
|
| 907 | if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args))
|
| 908 | || rm->atomic.op_active)
|
| 909 | return -EINVAL;
|
| 910 |
|
| 911 | args = CMSG_DATA(cmsg);
|
| 912 |
|
| 913 |
|
| 914 | switch (cmsg->cmsg_type) {
|
| 915 | case RDS_CMSG_ATOMIC_FADD:
|
| 916 | rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
|
| 917 | rm->atomic.op_m_fadd.add = args->fadd.add;
|
| 918 | rm->atomic.op_m_fadd.nocarry_mask = 0;
|
| 919 | break;
|
| 920 | case RDS_CMSG_MASKED_ATOMIC_FADD:
|
| 921 | rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD;
|
| 922 | rm->atomic.op_m_fadd.add = args->m_fadd.add;
|
| 923 | rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask;
|
| 924 | break;
|
| 925 | case RDS_CMSG_ATOMIC_CSWP:
|
| 926 | rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
|
| 927 | rm->atomic.op_m_cswp.compare = args->cswp.compare;
|
| 928 | rm->atomic.op_m_cswp.swap = args->cswp.swap;
|
| 929 | rm->atomic.op_m_cswp.compare_mask = ~0;
|
| 930 | rm->atomic.op_m_cswp.swap_mask = ~0;
|
| 931 | break;
|
| 932 | case RDS_CMSG_MASKED_ATOMIC_CSWP:
|
| 933 | rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP;
|
| 934 | rm->atomic.op_m_cswp.compare = args->m_cswp.compare;
|
| 935 | rm->atomic.op_m_cswp.swap = args->m_cswp.swap;
|
| 936 | rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask;
|
| 937 | rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask;
|
| 938 | break;
|
| 939 | default:
|
| 940 | BUG();
|
| 941 | }
|
| 942 |
|
| 943 | rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
|
| 944 | rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT);
|
| 945 | rm->atomic.op_active = 1;
|
| 946 | rm->atomic.op_recverr = rs->rs_recverr;
|
| 947 | rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1);
|
| 948 | if (IS_ERR(rm->atomic.op_sg)) {
|
| 949 | ret = PTR_ERR(rm->atomic.op_sg);
|
| 950 | goto err;
|
| 951 | }
|
| 952 |
|
| 953 |
|
| 954 | if (args->local_addr & 0x7) {
|
| 955 | ret = -EFAULT;
|
| 956 | goto err;
|
| 957 | }
|
| 958 |
|
| 959 | ret = rds_pin_pages(args->local_addr, 1, &page, 1);
|
| 960 | if (ret != 1)
|
| 961 | goto err;
|
| 962 | ret = 0;
|
| 963 |
|
| 964 | sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr));
|
| 965 |
|
| 966 | if (rm->atomic.op_notify || rm->atomic.op_recverr) {
|
| 967 | |
| 968 | * we don't want to do that in the completion handler. We
|
| 969 | * would have to use GFP_ATOMIC there, and don't want to deal
|
| 970 | * with failed allocations.
|
| 971 |
|
| 972 | rm->atomic.op_notifier = kmalloc_obj(*rm->atomic.op_notifier);
|
| 973 | if (!rm->atomic.op_notifier) {
|
| 974 | ret = -ENOMEM;
|
| 975 | goto err;
|
| 976 | }
|
| 977 |
|
| 978 | rm->atomic.op_notifier->n_user_token = args->user_token;
|
| 979 | rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS;
|
| 980 | }
|
| 981 |
|
| 982 | rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie);
|
| 983 | rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie);
|
| 984 |
|
| 985 | return ret;
|
| 986 | err:
|
| 987 | if (page)
|
| 988 | unpin_user_page(page);
|
| 989 | rm->atomic.op_active = 0;
|
| 990 | kfree(rm->atomic.op_notifier);
|
| 991 |
|
| 992 | return ret;
|
| 993 | }
|