1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/fs/nfs/direct.c 4 * 5 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com> 6 * 7 * High-performance uncached I/O for the Linux NFS client 8 * 9 * There are important applications whose performance or correctness 10 * depends on uncached access to file data. Database clusters 11 * (multiple copies of the same instance running on separate hosts) 12 * implement their own cache coherency protocol that subsumes file 13 * system cache protocols. Applications that process datasets 14 * considerably larger than the client's memory do not always benefit 15 * from a local cache. A streaming video server, for instance, has no 16 * need to cache the contents of a file. 17 * 18 * When an application requests uncached I/O, all read and write requests 19 * are made directly to the server; data stored or fetched via these 20 * requests is not cached in the Linux page cache. The client does not 21 * correct unaligned requests from applications. All requested bytes are 22 * held on permanent storage before a direct write system call returns to 23 * an application. 24 * 25 * Solaris implements an uncached I/O facility called directio() that 26 * is used for backups and sequential I/O to very large files. Solaris 27 * also supports uncaching whole NFS partitions with "-o forcedirectio," 28 * an undocumented mount option. 29 * 30 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with 31 * help from Andrew Morton. 32 * 33 * 18 Dec 2001 Initial implementation for 2.4 --cel 34 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy 35 * 08 Jun 2003 Port to 2.5 APIs --cel 36 * 31 Mar 2004 Handle direct I/O without VFS support --cel 37 * 15 Sep 2004 Parallel async reads --cel 38 * 04 May 2005 support O_DIRECT with aio --cel 39 * 40 */ 41 42 #include <linux/errno.h> 43 #include <linux/sched.h> 44 #include <linux/kernel.h> 45 #include <linux/file.h> 46 #include <linux/pagemap.h> 47 #include <linux/kref.h> 48 #include <linux/slab.h> 49 #include <linux/task_io_accounting_ops.h> 50 #include <linux/module.h> 51 52 #include <linux/nfs_fs.h> 53 #include <linux/nfs_page.h> 54 #include <linux/sunrpc/clnt.h> 55 56 #include <linux/uaccess.h> 57 #include <linux/atomic.h> 58 59 #include "internal.h" 60 #include "iostat.h" 61 #include "pnfs.h" 62 #include "fscache.h" 63 #include "nfstrace.h" 64 65 #define NFSDBG_FACILITY NFSDBG_VFS 66 67 static struct kmem_cache *nfs_direct_cachep; 68 69 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops; 70 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops; 71 static void nfs_direct_write_complete(struct nfs_direct_req *dreq); 72 static void nfs_direct_write_schedule_work(struct work_struct *work); 73 74 static inline void get_dreq(struct nfs_direct_req *dreq) 75 { 76 atomic_inc(&dreq->io_count); 77 } 78 79 static inline int put_dreq(struct nfs_direct_req *dreq) 80 { 81 return atomic_dec_and_test(&dreq->io_count); 82 } 83 84 static void 85 nfs_direct_handle_truncated(struct nfs_direct_req *dreq, 86 const struct nfs_pgio_header *hdr, 87 ssize_t dreq_len) 88 { 89 if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) || 90 test_bit(NFS_IOHDR_EOF, &hdr->flags))) 91 return; 92 if (dreq->max_count >= dreq_len) { 93 dreq->max_count = dreq_len; 94 if (dreq->count > dreq_len) 95 dreq->count = dreq_len; 96 97 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) 98 dreq->error = hdr->error; 99 else /* Clear outstanding error if this is EOF */ 100 dreq->error = 0; 101 } 102 } 103 104 static void 105 nfs_direct_count_bytes(struct nfs_direct_req *dreq, 106 const struct nfs_pgio_header *hdr) 107 { 108 loff_t hdr_end = hdr->io_start + hdr->good_bytes; 109 ssize_t dreq_len = 0; 110 111 if (hdr_end > dreq->io_start) 112 dreq_len = hdr_end - dreq->io_start; 113 114 nfs_direct_handle_truncated(dreq, hdr, dreq_len); 115 116 if (dreq_len > dreq->max_count) 117 dreq_len = dreq->max_count; 118 119 if (dreq->count < dreq_len) 120 dreq->count = dreq_len; 121 } 122 123 /** 124 * nfs_swap_rw - NFS address space operation for swap I/O 125 * @iocb: target I/O control block 126 * @iter: I/O buffer 127 * 128 * Perform IO to the swap-file. This is much like direct IO. 129 */ 130 int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter) 131 { 132 ssize_t ret; 133 134 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE); 135 136 if (iov_iter_rw(iter) == READ) 137 ret = nfs_file_direct_read(iocb, iter, true); 138 else 139 ret = nfs_file_direct_write(iocb, iter, true); 140 if (ret < 0) 141 return ret; 142 return 0; 143 } 144 145 static void nfs_direct_release_pages(struct page **pages, unsigned int npages) 146 { 147 unsigned int i; 148 for (i = 0; i < npages; i++) 149 put_page(pages[i]); 150 } 151 152 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, 153 struct nfs_direct_req *dreq) 154 { 155 cinfo->inode = dreq->inode; 156 cinfo->mds = &dreq->mds_cinfo; 157 cinfo->ds = &dreq->ds_cinfo; 158 cinfo->dreq = dreq; 159 cinfo->completion_ops = &nfs_direct_commit_completion_ops; 160 } 161 162 static inline struct nfs_direct_req *nfs_direct_req_alloc(void) 163 { 164 struct nfs_direct_req *dreq; 165 166 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL); 167 if (!dreq) 168 return NULL; 169 170 kref_init(&dreq->kref); 171 kref_get(&dreq->kref); 172 init_completion(&dreq->completion); 173 INIT_LIST_HEAD(&dreq->mds_cinfo.list); 174 pnfs_init_ds_commit_info(&dreq->ds_cinfo); 175 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); 176 spin_lock_init(&dreq->lock); 177 178 return dreq; 179 } 180 181 static void nfs_direct_req_free(struct kref *kref) 182 { 183 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); 184 185 pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode); 186 if (dreq->l_ctx != NULL) 187 nfs_put_lock_context(dreq->l_ctx); 188 if (dreq->ctx != NULL) 189 put_nfs_open_context(dreq->ctx); 190 kmem_cache_free(nfs_direct_cachep, dreq); 191 } 192 193 static void nfs_direct_req_release(struct nfs_direct_req *dreq) 194 { 195 kref_put(&dreq->kref, nfs_direct_req_free); 196 } 197 198 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq) 199 { 200 return dreq->bytes_left; 201 } 202 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left); 203 204 /* 205 * Collects and returns the final error value/byte-count. 206 */ 207 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) 208 { 209 ssize_t result = -EIOCBQUEUED; 210 211 /* Async requests don't wait here */ 212 if (dreq->iocb) 213 goto out; 214 215 result = wait_for_completion_killable(&dreq->completion); 216 217 if (!result) { 218 result = dreq->count; 219 WARN_ON_ONCE(dreq->count < 0); 220 } 221 if (!result) 222 result = dreq->error; 223 224 out: 225 return (ssize_t) result; 226 } 227 228 /* 229 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust 230 * the iocb is still valid here if this is a synchronous request. 231 */ 232 static void nfs_direct_complete(struct nfs_direct_req *dreq) 233 { 234 struct inode *inode = dreq->inode; 235 236 inode_dio_end(inode); 237 238 if (dreq->iocb) { 239 long res = (long) dreq->error; 240 if (dreq->count != 0) { 241 res = (long) dreq->count; 242 WARN_ON_ONCE(dreq->count < 0); 243 } 244 dreq->iocb->ki_complete(dreq->iocb, res); 245 } 246 247 complete(&dreq->completion); 248 249 nfs_direct_req_release(dreq); 250 } 251 252 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) 253 { 254 unsigned long bytes = 0; 255 struct nfs_direct_req *dreq = hdr->dreq; 256 257 spin_lock(&dreq->lock); 258 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { 259 spin_unlock(&dreq->lock); 260 goto out_put; 261 } 262 263 nfs_direct_count_bytes(dreq, hdr); 264 spin_unlock(&dreq->lock); 265 266 while (!list_empty(&hdr->pages)) { 267 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 268 struct page *page = req->wb_page; 269 270 if (!PageCompound(page) && bytes < hdr->good_bytes && 271 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY)) 272 set_page_dirty(page); 273 bytes += req->wb_bytes; 274 nfs_list_remove_request(req); 275 nfs_release_request(req); 276 } 277 out_put: 278 if (put_dreq(dreq)) 279 nfs_direct_complete(dreq); 280 hdr->release(hdr); 281 } 282 283 static void nfs_read_sync_pgio_error(struct list_head *head, int error) 284 { 285 struct nfs_page *req; 286 287 while (!list_empty(head)) { 288 req = nfs_list_entry(head->next); 289 nfs_list_remove_request(req); 290 nfs_release_request(req); 291 } 292 } 293 294 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) 295 { 296 get_dreq(hdr->dreq); 297 } 298 299 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { 300 .error_cleanup = nfs_read_sync_pgio_error, 301 .init_hdr = nfs_direct_pgio_init, 302 .completion = nfs_direct_read_completion, 303 }; 304 305 /* 306 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ 307 * operation. If nfs_readdata_alloc() or get_user_pages() fails, 308 * bail and stop sending more reads. Read length accounting is 309 * handled automatically by nfs_direct_read_result(). Otherwise, if 310 * no requests have been sent, just return an error. 311 */ 312 313 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, 314 struct iov_iter *iter, 315 loff_t pos) 316 { 317 struct nfs_pageio_descriptor desc; 318 struct inode *inode = dreq->inode; 319 ssize_t result = -EINVAL; 320 size_t requested_bytes = 0; 321 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE); 322 323 nfs_pageio_init_read(&desc, dreq->inode, false, 324 &nfs_direct_read_completion_ops); 325 get_dreq(dreq); 326 desc.pg_dreq = dreq; 327 inode_dio_begin(inode); 328 329 while (iov_iter_count(iter)) { 330 struct page **pagevec; 331 size_t bytes; 332 size_t pgbase; 333 unsigned npages, i; 334 335 result = iov_iter_get_pages_alloc2(iter, &pagevec, 336 rsize, &pgbase); 337 if (result < 0) 338 break; 339 340 bytes = result; 341 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; 342 for (i = 0; i < npages; i++) { 343 struct nfs_page *req; 344 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); 345 /* XXX do we need to do the eof zeroing found in async_filler? */ 346 req = nfs_page_create_from_page(dreq->ctx, pagevec[i], 347 pgbase, pos, req_len); 348 if (IS_ERR(req)) { 349 result = PTR_ERR(req); 350 break; 351 } 352 if (!nfs_pageio_add_request(&desc, req)) { 353 result = desc.pg_error; 354 nfs_release_request(req); 355 break; 356 } 357 pgbase = 0; 358 bytes -= req_len; 359 requested_bytes += req_len; 360 pos += req_len; 361 dreq->bytes_left -= req_len; 362 } 363 nfs_direct_release_pages(pagevec, npages); 364 kvfree(pagevec); 365 if (result < 0) 366 break; 367 } 368 369 nfs_pageio_complete(&desc); 370 371 /* 372 * If no bytes were started, return the error, and let the 373 * generic layer handle the completion. 374 */ 375 if (requested_bytes == 0) { 376 inode_dio_end(inode); 377 nfs_direct_req_release(dreq); 378 return result < 0 ? result : -EIO; 379 } 380 381 if (put_dreq(dreq)) 382 nfs_direct_complete(dreq); 383 return requested_bytes; 384 } 385 386 /** 387 * nfs_file_direct_read - file direct read operation for NFS files 388 * @iocb: target I/O control block 389 * @iter: vector of user buffers into which to read data 390 * @swap: flag indicating this is swap IO, not O_DIRECT IO 391 * 392 * We use this function for direct reads instead of calling 393 * generic_file_aio_read() in order to avoid gfar's check to see if 394 * the request starts before the end of the file. For that check 395 * to work, we must generate a GETATTR before each direct read, and 396 * even then there is a window between the GETATTR and the subsequent 397 * READ where the file size could change. Our preference is simply 398 * to do all reads the application wants, and the server will take 399 * care of managing the end of file boundary. 400 * 401 * This function also eliminates unnecessarily updating the file's 402 * atime locally, as the NFS server sets the file's atime, and this 403 * client must read the updated atime from the server back into its 404 * cache. 405 */ 406 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter, 407 bool swap) 408 { 409 struct file *file = iocb->ki_filp; 410 struct address_space *mapping = file->f_mapping; 411 struct inode *inode = mapping->host; 412 struct nfs_direct_req *dreq; 413 struct nfs_lock_context *l_ctx; 414 ssize_t result, requested; 415 size_t count = iov_iter_count(iter); 416 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); 417 418 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", 419 file, count, (long long) iocb->ki_pos); 420 421 result = 0; 422 if (!count) 423 goto out; 424 425 task_io_account_read(count); 426 427 result = -ENOMEM; 428 dreq = nfs_direct_req_alloc(); 429 if (dreq == NULL) 430 goto out; 431 432 dreq->inode = inode; 433 dreq->bytes_left = dreq->max_count = count; 434 dreq->io_start = iocb->ki_pos; 435 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); 436 l_ctx = nfs_get_lock_context(dreq->ctx); 437 if (IS_ERR(l_ctx)) { 438 result = PTR_ERR(l_ctx); 439 nfs_direct_req_release(dreq); 440 goto out_release; 441 } 442 dreq->l_ctx = l_ctx; 443 if (!is_sync_kiocb(iocb)) 444 dreq->iocb = iocb; 445 446 if (user_backed_iter(iter)) 447 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY; 448 449 if (!swap) 450 nfs_start_io_direct(inode); 451 452 NFS_I(inode)->read_io += count; 453 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos); 454 455 if (!swap) 456 nfs_end_io_direct(inode); 457 458 if (requested > 0) { 459 result = nfs_direct_wait(dreq); 460 if (result > 0) { 461 requested -= result; 462 iocb->ki_pos += result; 463 } 464 iov_iter_revert(iter, requested); 465 } else { 466 result = requested; 467 } 468 469 out_release: 470 nfs_direct_req_release(dreq); 471 out: 472 return result; 473 } 474 475 static void nfs_direct_add_page_head(struct list_head *list, 476 struct nfs_page *req) 477 { 478 struct nfs_page *head = req->wb_head; 479 480 if (!list_empty(&head->wb_list) || !nfs_lock_request(head)) 481 return; 482 if (!list_empty(&head->wb_list)) { 483 nfs_unlock_request(head); 484 return; 485 } 486 list_add(&head->wb_list, list); 487 kref_get(&head->wb_kref); 488 kref_get(&head->wb_kref); 489 } 490 491 static void nfs_direct_join_group(struct list_head *list, struct inode *inode) 492 { 493 struct nfs_page *req, *subreq; 494 495 list_for_each_entry(req, list, wb_list) { 496 if (req->wb_head != req) { 497 nfs_direct_add_page_head(&req->wb_list, req); 498 continue; 499 } 500 subreq = req->wb_this_page; 501 if (subreq == req) 502 continue; 503 do { 504 /* 505 * Remove subrequests from this list before freeing 506 * them in the call to nfs_join_page_group(). 507 */ 508 if (!list_empty(&subreq->wb_list)) { 509 nfs_list_remove_request(subreq); 510 nfs_release_request(subreq); 511 } 512 } while ((subreq = subreq->wb_this_page) != req); 513 nfs_join_page_group(req, inode); 514 } 515 } 516 517 static void 518 nfs_direct_write_scan_commit_list(struct inode *inode, 519 struct list_head *list, 520 struct nfs_commit_info *cinfo) 521 { 522 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); 523 pnfs_recover_commit_reqs(list, cinfo); 524 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0); 525 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex); 526 } 527 528 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) 529 { 530 struct nfs_pageio_descriptor desc; 531 struct nfs_page *req, *tmp; 532 LIST_HEAD(reqs); 533 struct nfs_commit_info cinfo; 534 LIST_HEAD(failed); 535 536 nfs_init_cinfo_from_dreq(&cinfo, dreq); 537 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); 538 539 nfs_direct_join_group(&reqs, dreq->inode); 540 541 dreq->count = 0; 542 dreq->max_count = 0; 543 list_for_each_entry(req, &reqs, wb_list) 544 dreq->max_count += req->wb_bytes; 545 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo); 546 get_dreq(dreq); 547 548 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, 549 &nfs_direct_write_completion_ops); 550 desc.pg_dreq = dreq; 551 552 list_for_each_entry_safe(req, tmp, &reqs, wb_list) { 553 /* Bump the transmission count */ 554 req->wb_nio++; 555 if (!nfs_pageio_add_request(&desc, req)) { 556 nfs_list_move_request(req, &failed); 557 spin_lock(&cinfo.inode->i_lock); 558 dreq->flags = 0; 559 if (desc.pg_error < 0) 560 dreq->error = desc.pg_error; 561 else 562 dreq->error = -EIO; 563 spin_unlock(&cinfo.inode->i_lock); 564 } 565 nfs_release_request(req); 566 } 567 nfs_pageio_complete(&desc); 568 569 while (!list_empty(&failed)) { 570 req = nfs_list_entry(failed.next); 571 nfs_list_remove_request(req); 572 nfs_unlock_and_release_request(req); 573 } 574 575 if (put_dreq(dreq)) 576 nfs_direct_write_complete(dreq); 577 } 578 579 static void nfs_direct_commit_complete(struct nfs_commit_data *data) 580 { 581 const struct nfs_writeverf *verf = data->res.verf; 582 struct nfs_direct_req *dreq = data->dreq; 583 struct nfs_commit_info cinfo; 584 struct nfs_page *req; 585 int status = data->task.tk_status; 586 587 trace_nfs_direct_commit_complete(dreq); 588 589 if (status < 0) { 590 /* Errors in commit are fatal */ 591 dreq->error = status; 592 dreq->max_count = 0; 593 dreq->count = 0; 594 dreq->flags = NFS_ODIRECT_DONE; 595 } else { 596 status = dreq->error; 597 } 598 599 nfs_init_cinfo_from_dreq(&cinfo, dreq); 600 601 while (!list_empty(&data->pages)) { 602 req = nfs_list_entry(data->pages.next); 603 nfs_list_remove_request(req); 604 if (status >= 0 && !nfs_write_match_verf(verf, req)) { 605 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; 606 /* 607 * Despite the reboot, the write was successful, 608 * so reset wb_nio. 609 */ 610 req->wb_nio = 0; 611 nfs_mark_request_commit(req, NULL, &cinfo, 0); 612 } else /* Error or match */ 613 nfs_release_request(req); 614 nfs_unlock_and_release_request(req); 615 } 616 617 if (nfs_commit_end(cinfo.mds)) 618 nfs_direct_write_complete(dreq); 619 } 620 621 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo, 622 struct nfs_page *req) 623 { 624 struct nfs_direct_req *dreq = cinfo->dreq; 625 626 trace_nfs_direct_resched_write(dreq); 627 628 spin_lock(&dreq->lock); 629 if (dreq->flags != NFS_ODIRECT_DONE) 630 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; 631 spin_unlock(&dreq->lock); 632 nfs_mark_request_commit(req, NULL, cinfo, 0); 633 } 634 635 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { 636 .completion = nfs_direct_commit_complete, 637 .resched_write = nfs_direct_resched_write, 638 }; 639 640 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) 641 { 642 int res; 643 struct nfs_commit_info cinfo; 644 LIST_HEAD(mds_list); 645 646 nfs_init_cinfo_from_dreq(&cinfo, dreq); 647 nfs_scan_commit(dreq->inode, &mds_list, &cinfo); 648 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); 649 if (res < 0) /* res == -ENOMEM */ 650 nfs_direct_write_reschedule(dreq); 651 } 652 653 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq) 654 { 655 struct nfs_commit_info cinfo; 656 struct nfs_page *req; 657 LIST_HEAD(reqs); 658 659 nfs_init_cinfo_from_dreq(&cinfo, dreq); 660 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); 661 662 while (!list_empty(&reqs)) { 663 req = nfs_list_entry(reqs.next); 664 nfs_list_remove_request(req); 665 nfs_release_request(req); 666 nfs_unlock_and_release_request(req); 667 } 668 } 669 670 static void nfs_direct_write_schedule_work(struct work_struct *work) 671 { 672 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); 673 int flags = dreq->flags; 674 675 dreq->flags = 0; 676 switch (flags) { 677 case NFS_ODIRECT_DO_COMMIT: 678 nfs_direct_commit_schedule(dreq); 679 break; 680 case NFS_ODIRECT_RESCHED_WRITES: 681 nfs_direct_write_reschedule(dreq); 682 break; 683 default: 684 nfs_direct_write_clear_reqs(dreq); 685 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping); 686 nfs_direct_complete(dreq); 687 } 688 } 689 690 static void nfs_direct_write_complete(struct nfs_direct_req *dreq) 691 { 692 trace_nfs_direct_write_complete(dreq); 693 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */ 694 } 695 696 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) 697 { 698 struct nfs_direct_req *dreq = hdr->dreq; 699 struct nfs_commit_info cinfo; 700 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 701 int flags = NFS_ODIRECT_DONE; 702 703 trace_nfs_direct_write_completion(dreq); 704 705 nfs_init_cinfo_from_dreq(&cinfo, dreq); 706 707 spin_lock(&dreq->lock); 708 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { 709 spin_unlock(&dreq->lock); 710 goto out_put; 711 } 712 713 nfs_direct_count_bytes(dreq, hdr); 714 if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags)) { 715 if (!dreq->flags) 716 dreq->flags = NFS_ODIRECT_DO_COMMIT; 717 flags = dreq->flags; 718 } 719 spin_unlock(&dreq->lock); 720 721 while (!list_empty(&hdr->pages)) { 722 723 req = nfs_list_entry(hdr->pages.next); 724 nfs_list_remove_request(req); 725 if (flags == NFS_ODIRECT_DO_COMMIT) { 726 kref_get(&req->wb_kref); 727 memcpy(&req->wb_verf, &hdr->verf.verifier, 728 sizeof(req->wb_verf)); 729 nfs_mark_request_commit(req, hdr->lseg, &cinfo, 730 hdr->ds_commit_idx); 731 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) { 732 kref_get(&req->wb_kref); 733 nfs_mark_request_commit(req, NULL, &cinfo, 0); 734 } 735 nfs_unlock_and_release_request(req); 736 } 737 738 out_put: 739 if (put_dreq(dreq)) 740 nfs_direct_write_complete(dreq); 741 hdr->release(hdr); 742 } 743 744 static void nfs_write_sync_pgio_error(struct list_head *head, int error) 745 { 746 struct nfs_page *req; 747 748 while (!list_empty(head)) { 749 req = nfs_list_entry(head->next); 750 nfs_list_remove_request(req); 751 nfs_unlock_and_release_request(req); 752 } 753 } 754 755 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr) 756 { 757 struct nfs_direct_req *dreq = hdr->dreq; 758 759 trace_nfs_direct_write_reschedule_io(dreq); 760 761 spin_lock(&dreq->lock); 762 if (dreq->error == 0) { 763 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; 764 /* fake unstable write to let common nfs resend pages */ 765 hdr->verf.committed = NFS_UNSTABLE; 766 hdr->good_bytes = hdr->args.offset + hdr->args.count - 767 hdr->io_start; 768 } 769 spin_unlock(&dreq->lock); 770 } 771 772 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { 773 .error_cleanup = nfs_write_sync_pgio_error, 774 .init_hdr = nfs_direct_pgio_init, 775 .completion = nfs_direct_write_completion, 776 .reschedule_io = nfs_direct_write_reschedule_io, 777 }; 778 779 780 /* 781 * NB: Return the value of the first error return code. Subsequent 782 * errors after the first one are ignored. 783 */ 784 /* 785 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE 786 * operation. If nfs_writedata_alloc() or get_user_pages() fails, 787 * bail and stop sending more writes. Write length accounting is 788 * handled automatically by nfs_direct_write_result(). Otherwise, if 789 * no requests have been sent, just return an error. 790 */ 791 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, 792 struct iov_iter *iter, 793 loff_t pos, int ioflags) 794 { 795 struct nfs_pageio_descriptor desc; 796 struct inode *inode = dreq->inode; 797 ssize_t result = 0; 798 size_t requested_bytes = 0; 799 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE); 800 801 trace_nfs_direct_write_schedule_iovec(dreq); 802 803 nfs_pageio_init_write(&desc, inode, ioflags, false, 804 &nfs_direct_write_completion_ops); 805 desc.pg_dreq = dreq; 806 get_dreq(dreq); 807 inode_dio_begin(inode); 808 809 NFS_I(inode)->write_io += iov_iter_count(iter); 810 while (iov_iter_count(iter)) { 811 struct page **pagevec; 812 size_t bytes; 813 size_t pgbase; 814 unsigned npages, i; 815 816 result = iov_iter_get_pages_alloc2(iter, &pagevec, 817 wsize, &pgbase); 818 if (result < 0) 819 break; 820 821 bytes = result; 822 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; 823 for (i = 0; i < npages; i++) { 824 struct nfs_page *req; 825 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); 826 827 req = nfs_page_create_from_page(dreq->ctx, pagevec[i], 828 pgbase, pos, req_len); 829 if (IS_ERR(req)) { 830 result = PTR_ERR(req); 831 break; 832 } 833 834 if (desc.pg_error < 0) { 835 nfs_free_request(req); 836 result = desc.pg_error; 837 break; 838 } 839 840 nfs_lock_request(req); 841 if (!nfs_pageio_add_request(&desc, req)) { 842 result = desc.pg_error; 843 nfs_unlock_and_release_request(req); 844 break; 845 } 846 pgbase = 0; 847 bytes -= req_len; 848 requested_bytes += req_len; 849 pos += req_len; 850 dreq->bytes_left -= req_len; 851 } 852 nfs_direct_release_pages(pagevec, npages); 853 kvfree(pagevec); 854 if (result < 0) 855 break; 856 } 857 nfs_pageio_complete(&desc); 858 859 /* 860 * If no bytes were started, return the error, and let the 861 * generic layer handle the completion. 862 */ 863 if (requested_bytes == 0) { 864 inode_dio_end(inode); 865 nfs_direct_req_release(dreq); 866 return result < 0 ? result : -EIO; 867 } 868 869 if (put_dreq(dreq)) 870 nfs_direct_write_complete(dreq); 871 return requested_bytes; 872 } 873 874 /** 875 * nfs_file_direct_write - file direct write operation for NFS files 876 * @iocb: target I/O control block 877 * @iter: vector of user buffers from which to write data 878 * @swap: flag indicating this is swap IO, not O_DIRECT IO 879 * 880 * We use this function for direct writes instead of calling 881 * generic_file_aio_write() in order to avoid taking the inode 882 * semaphore and updating the i_size. The NFS server will set 883 * the new i_size and this client must read the updated size 884 * back into its cache. We let the server do generic write 885 * parameter checking and report problems. 886 * 887 * We eliminate local atime updates, see direct read above. 888 * 889 * We avoid unnecessary page cache invalidations for normal cached 890 * readers of this file. 891 * 892 * Note that O_APPEND is not supported for NFS direct writes, as there 893 * is no atomic O_APPEND write facility in the NFS protocol. 894 */ 895 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter, 896 bool swap) 897 { 898 ssize_t result, requested; 899 size_t count; 900 struct file *file = iocb->ki_filp; 901 struct address_space *mapping = file->f_mapping; 902 struct inode *inode = mapping->host; 903 struct nfs_direct_req *dreq; 904 struct nfs_lock_context *l_ctx; 905 loff_t pos, end; 906 907 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", 908 file, iov_iter_count(iter), (long long) iocb->ki_pos); 909 910 if (swap) 911 /* bypass generic checks */ 912 result = iov_iter_count(iter); 913 else 914 result = generic_write_checks(iocb, iter); 915 if (result <= 0) 916 return result; 917 count = result; 918 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); 919 920 pos = iocb->ki_pos; 921 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT; 922 923 task_io_account_write(count); 924 925 result = -ENOMEM; 926 dreq = nfs_direct_req_alloc(); 927 if (!dreq) 928 goto out; 929 930 dreq->inode = inode; 931 dreq->bytes_left = dreq->max_count = count; 932 dreq->io_start = pos; 933 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); 934 l_ctx = nfs_get_lock_context(dreq->ctx); 935 if (IS_ERR(l_ctx)) { 936 result = PTR_ERR(l_ctx); 937 nfs_direct_req_release(dreq); 938 goto out_release; 939 } 940 dreq->l_ctx = l_ctx; 941 if (!is_sync_kiocb(iocb)) 942 dreq->iocb = iocb; 943 pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode); 944 945 if (swap) { 946 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos, 947 FLUSH_STABLE); 948 } else { 949 nfs_start_io_direct(inode); 950 951 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos, 952 FLUSH_COND_STABLE); 953 954 if (mapping->nrpages) { 955 invalidate_inode_pages2_range(mapping, 956 pos >> PAGE_SHIFT, end); 957 } 958 959 nfs_end_io_direct(inode); 960 } 961 962 if (requested > 0) { 963 result = nfs_direct_wait(dreq); 964 if (result > 0) { 965 requested -= result; 966 iocb->ki_pos = pos + result; 967 /* XXX: should check the generic_write_sync retval */ 968 generic_write_sync(iocb, result); 969 } 970 iov_iter_revert(iter, requested); 971 } else { 972 result = requested; 973 } 974 nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE); 975 out_release: 976 nfs_direct_req_release(dreq); 977 out: 978 return result; 979 } 980 981 /** 982 * nfs_init_directcache - create a slab cache for nfs_direct_req structures 983 * 984 */ 985 int __init nfs_init_directcache(void) 986 { 987 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", 988 sizeof(struct nfs_direct_req), 989 0, (SLAB_RECLAIM_ACCOUNT| 990 SLAB_MEM_SPREAD), 991 NULL); 992 if (nfs_direct_cachep == NULL) 993 return -ENOMEM; 994 995 return 0; 996 } 997 998 /** 999 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures 1000 * 1001 */ 1002 void nfs_destroy_directcache(void) 1003 { 1004 kmem_cache_destroy(nfs_direct_cachep); 1005 } 1006