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 #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, ©); 83 INIT_LIST_HEAD(&q->zcookie_head); 84 spin_unlock_irqrestore(&q->lock, flags); 85 86 list_for_each_entry_safe(info, tmp, ©, 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 rds_mr_put(rm->rdma.op_rdma_mr); 166 167 if (rm->atomic.op_active) 168 rds_atomic_free_op(&rm->atomic); 169 if (rm->atomic.op_rdma_mr) 170 rds_mr_put(rm->atomic.op_rdma_mr); 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 } 184 } 185 EXPORT_SYMBOL_GPL(rds_message_put); 186 187 void rds_message_populate_header(struct rds_header *hdr, __be16 sport, 188 __be16 dport, u64 seq) 189 { 190 hdr->h_flags = 0; 191 hdr->h_sport = sport; 192 hdr->h_dport = dport; 193 hdr->h_sequence = cpu_to_be64(seq); 194 hdr->h_exthdr[0] = RDS_EXTHDR_NONE; 195 } 196 EXPORT_SYMBOL_GPL(rds_message_populate_header); 197 198 int rds_message_add_extension(struct rds_header *hdr, unsigned int type, 199 const void *data, unsigned int len) 200 { 201 unsigned int ext_len = sizeof(u8) + len; 202 unsigned char *dst; 203 204 /* For now, refuse to add more than one extension header */ 205 if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE) 206 return 0; 207 208 if (type >= __RDS_EXTHDR_MAX || len != rds_exthdr_size[type]) 209 return 0; 210 211 if (ext_len >= RDS_HEADER_EXT_SPACE) 212 return 0; 213 dst = hdr->h_exthdr; 214 215 *dst++ = type; 216 memcpy(dst, data, len); 217 218 dst[len] = RDS_EXTHDR_NONE; 219 return 1; 220 } 221 EXPORT_SYMBOL_GPL(rds_message_add_extension); 222 223 /* 224 * If a message has extension headers, retrieve them here. 225 * Call like this: 226 * 227 * unsigned int pos = 0; 228 * 229 * while (1) { 230 * buflen = sizeof(buffer); 231 * type = rds_message_next_extension(hdr, &pos, buffer, &buflen); 232 * if (type == RDS_EXTHDR_NONE) 233 * break; 234 * ... 235 * } 236 */ 237 int rds_message_next_extension(struct rds_header *hdr, 238 unsigned int *pos, void *buf, unsigned int *buflen) 239 { 240 unsigned int offset, ext_type, ext_len; 241 u8 *src = hdr->h_exthdr; 242 243 offset = *pos; 244 if (offset >= RDS_HEADER_EXT_SPACE) 245 goto none; 246 247 /* Get the extension type and length. For now, the 248 * length is implied by the extension type. */ 249 ext_type = src[offset++]; 250 251 if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX) 252 goto none; 253 ext_len = rds_exthdr_size[ext_type]; 254 if (offset + ext_len > RDS_HEADER_EXT_SPACE) 255 goto none; 256 257 *pos = offset + ext_len; 258 if (ext_len < *buflen) 259 *buflen = ext_len; 260 memcpy(buf, src + offset, *buflen); 261 return ext_type; 262 263 none: 264 *pos = RDS_HEADER_EXT_SPACE; 265 *buflen = 0; 266 return RDS_EXTHDR_NONE; 267 } 268 269 int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset) 270 { 271 struct rds_ext_header_rdma_dest ext_hdr; 272 273 ext_hdr.h_rdma_rkey = cpu_to_be32(r_key); 274 ext_hdr.h_rdma_offset = cpu_to_be32(offset); 275 return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr)); 276 } 277 EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension); 278 279 /* 280 * Each rds_message is allocated with extra space for the scatterlist entries 281 * rds ops will need. This is to minimize memory allocation count. Then, each rds op 282 * can grab SGs when initializing its part of the rds_message. 283 */ 284 struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp) 285 { 286 struct rds_message *rm; 287 288 if (extra_len > KMALLOC_MAX_SIZE - sizeof(struct rds_message)) 289 return NULL; 290 291 rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp); 292 if (!rm) 293 goto out; 294 295 rm->m_used_sgs = 0; 296 rm->m_total_sgs = extra_len / sizeof(struct scatterlist); 297 298 refcount_set(&rm->m_refcount, 1); 299 INIT_LIST_HEAD(&rm->m_sock_item); 300 INIT_LIST_HEAD(&rm->m_conn_item); 301 spin_lock_init(&rm->m_rs_lock); 302 init_waitqueue_head(&rm->m_flush_wait); 303 304 out: 305 return rm; 306 } 307 308 /* 309 * RDS ops use this to grab SG entries from the rm's sg pool. 310 */ 311 struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents, 312 int *ret) 313 { 314 struct scatterlist *sg_first = (struct scatterlist *) &rm[1]; 315 struct scatterlist *sg_ret; 316 317 if (WARN_ON(!ret)) 318 return NULL; 319 320 if (nents <= 0) { 321 pr_warn("rds: alloc sgs failed! nents <= 0\n"); 322 *ret = -EINVAL; 323 return NULL; 324 } 325 326 if (rm->m_used_sgs + nents > rm->m_total_sgs) { 327 pr_warn("rds: alloc sgs failed! total %d used %d nents %d\n", 328 rm->m_total_sgs, rm->m_used_sgs, nents); 329 *ret = -ENOMEM; 330 return NULL; 331 } 332 333 sg_ret = &sg_first[rm->m_used_sgs]; 334 sg_init_table(sg_ret, nents); 335 rm->m_used_sgs += nents; 336 337 return sg_ret; 338 } 339 340 struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len) 341 { 342 struct rds_message *rm; 343 unsigned int i; 344 int num_sgs = ceil(total_len, PAGE_SIZE); 345 int extra_bytes = num_sgs * sizeof(struct scatterlist); 346 int ret; 347 348 rm = rds_message_alloc(extra_bytes, GFP_NOWAIT); 349 if (!rm) 350 return ERR_PTR(-ENOMEM); 351 352 set_bit(RDS_MSG_PAGEVEC, &rm->m_flags); 353 rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len); 354 rm->data.op_nents = ceil(total_len, PAGE_SIZE); 355 rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs, &ret); 356 if (!rm->data.op_sg) { 357 rds_message_put(rm); 358 return ERR_PTR(ret); 359 } 360 361 for (i = 0; i < rm->data.op_nents; ++i) { 362 sg_set_page(&rm->data.op_sg[i], 363 virt_to_page(page_addrs[i]), 364 PAGE_SIZE, 0); 365 } 366 367 return rm; 368 } 369 370 static int rds_message_zcopy_from_user(struct rds_message *rm, struct iov_iter *from) 371 { 372 struct scatterlist *sg; 373 int ret = 0; 374 int length = iov_iter_count(from); 375 int total_copied = 0; 376 struct rds_msg_zcopy_info *info; 377 378 rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from)); 379 380 /* 381 * now allocate and copy in the data payload. 382 */ 383 sg = rm->data.op_sg; 384 385 info = kzalloc(sizeof(*info), GFP_KERNEL); 386 if (!info) 387 return -ENOMEM; 388 INIT_LIST_HEAD(&info->rs_zcookie_next); 389 rm->data.op_mmp_znotifier = &info->znotif; 390 if (mm_account_pinned_pages(&rm->data.op_mmp_znotifier->z_mmp, 391 length)) { 392 ret = -ENOMEM; 393 goto err; 394 } 395 while (iov_iter_count(from)) { 396 struct page *pages; 397 size_t start; 398 ssize_t copied; 399 400 copied = iov_iter_get_pages(from, &pages, PAGE_SIZE, 401 1, &start); 402 if (copied < 0) { 403 struct mmpin *mmp; 404 int i; 405 406 for (i = 0; i < rm->data.op_nents; i++) 407 put_page(sg_page(&rm->data.op_sg[i])); 408 mmp = &rm->data.op_mmp_znotifier->z_mmp; 409 mm_unaccount_pinned_pages(mmp); 410 ret = -EFAULT; 411 goto err; 412 } 413 total_copied += copied; 414 iov_iter_advance(from, copied); 415 length -= copied; 416 sg_set_page(sg, pages, copied, start); 417 rm->data.op_nents++; 418 sg++; 419 } 420 WARN_ON_ONCE(length != 0); 421 return ret; 422 err: 423 kfree(info); 424 rm->data.op_mmp_znotifier = NULL; 425 return ret; 426 } 427 428 int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from, 429 bool zcopy) 430 { 431 unsigned long to_copy, nbytes; 432 unsigned long sg_off; 433 struct scatterlist *sg; 434 int ret = 0; 435 436 rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from)); 437 438 /* now allocate and copy in the data payload. */ 439 sg = rm->data.op_sg; 440 sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */ 441 442 if (zcopy) 443 return rds_message_zcopy_from_user(rm, from); 444 445 while (iov_iter_count(from)) { 446 if (!sg_page(sg)) { 447 ret = rds_page_remainder_alloc(sg, iov_iter_count(from), 448 GFP_HIGHUSER); 449 if (ret) 450 return ret; 451 rm->data.op_nents++; 452 sg_off = 0; 453 } 454 455 to_copy = min_t(unsigned long, iov_iter_count(from), 456 sg->length - sg_off); 457 458 rds_stats_add(s_copy_from_user, to_copy); 459 nbytes = copy_page_from_iter(sg_page(sg), sg->offset + sg_off, 460 to_copy, from); 461 if (nbytes != to_copy) 462 return -EFAULT; 463 464 sg_off += to_copy; 465 466 if (sg_off == sg->length) 467 sg++; 468 } 469 470 return ret; 471 } 472 473 int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to) 474 { 475 struct rds_message *rm; 476 struct scatterlist *sg; 477 unsigned long to_copy; 478 unsigned long vec_off; 479 int copied; 480 int ret; 481 u32 len; 482 483 rm = container_of(inc, struct rds_message, m_inc); 484 len = be32_to_cpu(rm->m_inc.i_hdr.h_len); 485 486 sg = rm->data.op_sg; 487 vec_off = 0; 488 copied = 0; 489 490 while (iov_iter_count(to) && copied < len) { 491 to_copy = min_t(unsigned long, iov_iter_count(to), 492 sg->length - vec_off); 493 to_copy = min_t(unsigned long, to_copy, len - copied); 494 495 rds_stats_add(s_copy_to_user, to_copy); 496 ret = copy_page_to_iter(sg_page(sg), sg->offset + vec_off, 497 to_copy, to); 498 if (ret != to_copy) 499 return -EFAULT; 500 501 vec_off += to_copy; 502 copied += to_copy; 503 504 if (vec_off == sg->length) { 505 vec_off = 0; 506 sg++; 507 } 508 } 509 510 return copied; 511 } 512 513 /* 514 * If the message is still on the send queue, wait until the transport 515 * is done with it. This is particularly important for RDMA operations. 516 */ 517 void rds_message_wait(struct rds_message *rm) 518 { 519 wait_event_interruptible(rm->m_flush_wait, 520 !test_bit(RDS_MSG_MAPPED, &rm->m_flags)); 521 } 522 523 void rds_message_unmapped(struct rds_message *rm) 524 { 525 clear_bit(RDS_MSG_MAPPED, &rm->m_flags); 526 wake_up_interruptible(&rm->m_flush_wait); 527 } 528 EXPORT_SYMBOL_GPL(rds_message_unmapped); 529