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_alloc(iter, &pagevec, 336 rsize, &pgbase); 337 if (result < 0) 338 break; 339 340 bytes = result; 341 iov_iter_advance(iter, bytes); 342 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; 343 for (i = 0; i < npages; i++) { 344 struct nfs_page *req; 345 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); 346 /* XXX do we need to do the eof zeroing found in async_filler? */ 347 req = nfs_create_request(dreq->ctx, pagevec[i], 348 pgbase, req_len); 349 if (IS_ERR(req)) { 350 result = PTR_ERR(req); 351 break; 352 } 353 req->wb_index = pos >> PAGE_SHIFT; 354 req->wb_offset = pos & ~PAGE_MASK; 355 if (!nfs_pageio_add_request(&desc, req)) { 356 result = desc.pg_error; 357 nfs_release_request(req); 358 break; 359 } 360 pgbase = 0; 361 bytes -= req_len; 362 requested_bytes += req_len; 363 pos += req_len; 364 dreq->bytes_left -= req_len; 365 } 366 nfs_direct_release_pages(pagevec, npages); 367 kvfree(pagevec); 368 if (result < 0) 369 break; 370 } 371 372 nfs_pageio_complete(&desc); 373 374 /* 375 * If no bytes were started, return the error, and let the 376 * generic layer handle the completion. 377 */ 378 if (requested_bytes == 0) { 379 inode_dio_end(inode); 380 nfs_direct_req_release(dreq); 381 return result < 0 ? result : -EIO; 382 } 383 384 if (put_dreq(dreq)) 385 nfs_direct_complete(dreq); 386 return requested_bytes; 387 } 388 389 /** 390 * nfs_file_direct_read - file direct read operation for NFS files 391 * @iocb: target I/O control block 392 * @iter: vector of user buffers into which to read data 393 * @swap: flag indicating this is swap IO, not O_DIRECT IO 394 * 395 * We use this function for direct reads instead of calling 396 * generic_file_aio_read() in order to avoid gfar's check to see if 397 * the request starts before the end of the file. For that check 398 * to work, we must generate a GETATTR before each direct read, and 399 * even then there is a window between the GETATTR and the subsequent 400 * READ where the file size could change. Our preference is simply 401 * to do all reads the application wants, and the server will take 402 * care of managing the end of file boundary. 403 * 404 * This function also eliminates unnecessarily updating the file's 405 * atime locally, as the NFS server sets the file's atime, and this 406 * client must read the updated atime from the server back into its 407 * cache. 408 */ 409 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter, 410 bool swap) 411 { 412 struct file *file = iocb->ki_filp; 413 struct address_space *mapping = file->f_mapping; 414 struct inode *inode = mapping->host; 415 struct nfs_direct_req *dreq; 416 struct nfs_lock_context *l_ctx; 417 ssize_t result, requested; 418 size_t count = iov_iter_count(iter); 419 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); 420 421 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", 422 file, count, (long long) iocb->ki_pos); 423 424 result = 0; 425 if (!count) 426 goto out; 427 428 task_io_account_read(count); 429 430 result = -ENOMEM; 431 dreq = nfs_direct_req_alloc(); 432 if (dreq == NULL) 433 goto out; 434 435 dreq->inode = inode; 436 dreq->bytes_left = dreq->max_count = count; 437 dreq->io_start = iocb->ki_pos; 438 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); 439 l_ctx = nfs_get_lock_context(dreq->ctx); 440 if (IS_ERR(l_ctx)) { 441 result = PTR_ERR(l_ctx); 442 nfs_direct_req_release(dreq); 443 goto out_release; 444 } 445 dreq->l_ctx = l_ctx; 446 if (!is_sync_kiocb(iocb)) 447 dreq->iocb = iocb; 448 449 if (iter_is_iovec(iter)) 450 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY; 451 452 if (!swap) 453 nfs_start_io_direct(inode); 454 455 NFS_I(inode)->read_io += count; 456 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos); 457 458 if (!swap) 459 nfs_end_io_direct(inode); 460 461 if (requested > 0) { 462 result = nfs_direct_wait(dreq); 463 if (result > 0) { 464 requested -= result; 465 iocb->ki_pos += result; 466 } 467 iov_iter_revert(iter, requested); 468 } else { 469 result = requested; 470 } 471 472 out_release: 473 nfs_direct_req_release(dreq); 474 out: 475 return result; 476 } 477 478 static void 479 nfs_direct_join_group(struct list_head *list, struct inode *inode) 480 { 481 struct nfs_page *req, *next; 482 483 list_for_each_entry(req, list, wb_list) { 484 if (req->wb_head != req || req->wb_this_page == req) 485 continue; 486 for (next = req->wb_this_page; 487 next != req->wb_head; 488 next = next->wb_this_page) { 489 nfs_list_remove_request(next); 490 nfs_release_request(next); 491 } 492 nfs_join_page_group(req, inode); 493 } 494 } 495 496 static void 497 nfs_direct_write_scan_commit_list(struct inode *inode, 498 struct list_head *list, 499 struct nfs_commit_info *cinfo) 500 { 501 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); 502 pnfs_recover_commit_reqs(list, cinfo); 503 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0); 504 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex); 505 } 506 507 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) 508 { 509 struct nfs_pageio_descriptor desc; 510 struct nfs_page *req, *tmp; 511 LIST_HEAD(reqs); 512 struct nfs_commit_info cinfo; 513 LIST_HEAD(failed); 514 515 nfs_init_cinfo_from_dreq(&cinfo, dreq); 516 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); 517 518 nfs_direct_join_group(&reqs, dreq->inode); 519 520 dreq->count = 0; 521 dreq->max_count = 0; 522 list_for_each_entry(req, &reqs, wb_list) 523 dreq->max_count += req->wb_bytes; 524 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo); 525 get_dreq(dreq); 526 527 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, 528 &nfs_direct_write_completion_ops); 529 desc.pg_dreq = dreq; 530 531 list_for_each_entry_safe(req, tmp, &reqs, wb_list) { 532 /* Bump the transmission count */ 533 req->wb_nio++; 534 if (!nfs_pageio_add_request(&desc, req)) { 535 nfs_list_move_request(req, &failed); 536 spin_lock(&cinfo.inode->i_lock); 537 dreq->flags = 0; 538 if (desc.pg_error < 0) 539 dreq->error = desc.pg_error; 540 else 541 dreq->error = -EIO; 542 spin_unlock(&cinfo.inode->i_lock); 543 } 544 nfs_release_request(req); 545 } 546 nfs_pageio_complete(&desc); 547 548 while (!list_empty(&failed)) { 549 req = nfs_list_entry(failed.next); 550 nfs_list_remove_request(req); 551 nfs_unlock_and_release_request(req); 552 } 553 554 if (put_dreq(dreq)) 555 nfs_direct_write_complete(dreq); 556 } 557 558 static void nfs_direct_commit_complete(struct nfs_commit_data *data) 559 { 560 const struct nfs_writeverf *verf = data->res.verf; 561 struct nfs_direct_req *dreq = data->dreq; 562 struct nfs_commit_info cinfo; 563 struct nfs_page *req; 564 int status = data->task.tk_status; 565 566 trace_nfs_direct_commit_complete(dreq); 567 568 if (status < 0) { 569 /* Errors in commit are fatal */ 570 dreq->error = status; 571 dreq->max_count = 0; 572 dreq->count = 0; 573 dreq->flags = NFS_ODIRECT_DONE; 574 } else { 575 status = dreq->error; 576 } 577 578 nfs_init_cinfo_from_dreq(&cinfo, dreq); 579 580 while (!list_empty(&data->pages)) { 581 req = nfs_list_entry(data->pages.next); 582 nfs_list_remove_request(req); 583 if (status >= 0 && !nfs_write_match_verf(verf, req)) { 584 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; 585 /* 586 * Despite the reboot, the write was successful, 587 * so reset wb_nio. 588 */ 589 req->wb_nio = 0; 590 nfs_mark_request_commit(req, NULL, &cinfo, 0); 591 } else /* Error or match */ 592 nfs_release_request(req); 593 nfs_unlock_and_release_request(req); 594 } 595 596 if (nfs_commit_end(cinfo.mds)) 597 nfs_direct_write_complete(dreq); 598 } 599 600 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo, 601 struct nfs_page *req) 602 { 603 struct nfs_direct_req *dreq = cinfo->dreq; 604 605 trace_nfs_direct_resched_write(dreq); 606 607 spin_lock(&dreq->lock); 608 if (dreq->flags != NFS_ODIRECT_DONE) 609 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; 610 spin_unlock(&dreq->lock); 611 nfs_mark_request_commit(req, NULL, cinfo, 0); 612 } 613 614 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { 615 .completion = nfs_direct_commit_complete, 616 .resched_write = nfs_direct_resched_write, 617 }; 618 619 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) 620 { 621 int res; 622 struct nfs_commit_info cinfo; 623 LIST_HEAD(mds_list); 624 625 nfs_init_cinfo_from_dreq(&cinfo, dreq); 626 nfs_scan_commit(dreq->inode, &mds_list, &cinfo); 627 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); 628 if (res < 0) /* res == -ENOMEM */ 629 nfs_direct_write_reschedule(dreq); 630 } 631 632 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq) 633 { 634 struct nfs_commit_info cinfo; 635 struct nfs_page *req; 636 LIST_HEAD(reqs); 637 638 nfs_init_cinfo_from_dreq(&cinfo, dreq); 639 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); 640 641 while (!list_empty(&reqs)) { 642 req = nfs_list_entry(reqs.next); 643 nfs_list_remove_request(req); 644 nfs_release_request(req); 645 nfs_unlock_and_release_request(req); 646 } 647 } 648 649 static void nfs_direct_write_schedule_work(struct work_struct *work) 650 { 651 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); 652 int flags = dreq->flags; 653 654 dreq->flags = 0; 655 switch (flags) { 656 case NFS_ODIRECT_DO_COMMIT: 657 nfs_direct_commit_schedule(dreq); 658 break; 659 case NFS_ODIRECT_RESCHED_WRITES: 660 nfs_direct_write_reschedule(dreq); 661 break; 662 default: 663 nfs_direct_write_clear_reqs(dreq); 664 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping); 665 nfs_direct_complete(dreq); 666 } 667 } 668 669 static void nfs_direct_write_complete(struct nfs_direct_req *dreq) 670 { 671 trace_nfs_direct_write_complete(dreq); 672 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */ 673 } 674 675 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) 676 { 677 struct nfs_direct_req *dreq = hdr->dreq; 678 struct nfs_commit_info cinfo; 679 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 680 int flags = NFS_ODIRECT_DONE; 681 682 trace_nfs_direct_write_completion(dreq); 683 684 nfs_init_cinfo_from_dreq(&cinfo, dreq); 685 686 spin_lock(&dreq->lock); 687 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { 688 spin_unlock(&dreq->lock); 689 goto out_put; 690 } 691 692 nfs_direct_count_bytes(dreq, hdr); 693 if (hdr->good_bytes != 0 && nfs_write_need_commit(hdr)) { 694 if (!dreq->flags) 695 dreq->flags = NFS_ODIRECT_DO_COMMIT; 696 flags = dreq->flags; 697 } 698 spin_unlock(&dreq->lock); 699 700 while (!list_empty(&hdr->pages)) { 701 702 req = nfs_list_entry(hdr->pages.next); 703 nfs_list_remove_request(req); 704 if (flags == NFS_ODIRECT_DO_COMMIT) { 705 kref_get(&req->wb_kref); 706 memcpy(&req->wb_verf, &hdr->verf.verifier, 707 sizeof(req->wb_verf)); 708 nfs_mark_request_commit(req, hdr->lseg, &cinfo, 709 hdr->ds_commit_idx); 710 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) { 711 kref_get(&req->wb_kref); 712 nfs_mark_request_commit(req, NULL, &cinfo, 0); 713 } 714 nfs_unlock_and_release_request(req); 715 } 716 717 out_put: 718 if (put_dreq(dreq)) 719 nfs_direct_write_complete(dreq); 720 hdr->release(hdr); 721 } 722 723 static void nfs_write_sync_pgio_error(struct list_head *head, int error) 724 { 725 struct nfs_page *req; 726 727 while (!list_empty(head)) { 728 req = nfs_list_entry(head->next); 729 nfs_list_remove_request(req); 730 nfs_unlock_and_release_request(req); 731 } 732 } 733 734 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr) 735 { 736 struct nfs_direct_req *dreq = hdr->dreq; 737 738 trace_nfs_direct_write_reschedule_io(dreq); 739 740 spin_lock(&dreq->lock); 741 if (dreq->error == 0) { 742 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; 743 /* fake unstable write to let common nfs resend pages */ 744 hdr->verf.committed = NFS_UNSTABLE; 745 hdr->good_bytes = hdr->args.offset + hdr->args.count - 746 hdr->io_start; 747 } 748 spin_unlock(&dreq->lock); 749 } 750 751 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { 752 .error_cleanup = nfs_write_sync_pgio_error, 753 .init_hdr = nfs_direct_pgio_init, 754 .completion = nfs_direct_write_completion, 755 .reschedule_io = nfs_direct_write_reschedule_io, 756 }; 757 758 759 /* 760 * NB: Return the value of the first error return code. Subsequent 761 * errors after the first one are ignored. 762 */ 763 /* 764 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE 765 * operation. If nfs_writedata_alloc() or get_user_pages() fails, 766 * bail and stop sending more writes. Write length accounting is 767 * handled automatically by nfs_direct_write_result(). Otherwise, if 768 * no requests have been sent, just return an error. 769 */ 770 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, 771 struct iov_iter *iter, 772 loff_t pos, int ioflags) 773 { 774 struct nfs_pageio_descriptor desc; 775 struct inode *inode = dreq->inode; 776 ssize_t result = 0; 777 size_t requested_bytes = 0; 778 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE); 779 780 trace_nfs_direct_write_schedule_iovec(dreq); 781 782 nfs_pageio_init_write(&desc, inode, ioflags, false, 783 &nfs_direct_write_completion_ops); 784 desc.pg_dreq = dreq; 785 get_dreq(dreq); 786 inode_dio_begin(inode); 787 788 NFS_I(inode)->write_io += iov_iter_count(iter); 789 while (iov_iter_count(iter)) { 790 struct page **pagevec; 791 size_t bytes; 792 size_t pgbase; 793 unsigned npages, i; 794 795 result = iov_iter_get_pages_alloc(iter, &pagevec, 796 wsize, &pgbase); 797 if (result < 0) 798 break; 799 800 bytes = result; 801 iov_iter_advance(iter, bytes); 802 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; 803 for (i = 0; i < npages; i++) { 804 struct nfs_page *req; 805 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); 806 807 req = nfs_create_request(dreq->ctx, pagevec[i], 808 pgbase, req_len); 809 if (IS_ERR(req)) { 810 result = PTR_ERR(req); 811 break; 812 } 813 814 if (desc.pg_error < 0) { 815 nfs_free_request(req); 816 result = desc.pg_error; 817 break; 818 } 819 820 nfs_lock_request(req); 821 req->wb_index = pos >> PAGE_SHIFT; 822 req->wb_offset = pos & ~PAGE_MASK; 823 if (!nfs_pageio_add_request(&desc, req)) { 824 result = desc.pg_error; 825 nfs_unlock_and_release_request(req); 826 break; 827 } 828 pgbase = 0; 829 bytes -= req_len; 830 requested_bytes += req_len; 831 pos += req_len; 832 dreq->bytes_left -= req_len; 833 } 834 nfs_direct_release_pages(pagevec, npages); 835 kvfree(pagevec); 836 if (result < 0) 837 break; 838 } 839 nfs_pageio_complete(&desc); 840 841 /* 842 * If no bytes were started, return the error, and let the 843 * generic layer handle the completion. 844 */ 845 if (requested_bytes == 0) { 846 inode_dio_end(inode); 847 nfs_direct_req_release(dreq); 848 return result < 0 ? result : -EIO; 849 } 850 851 if (put_dreq(dreq)) 852 nfs_direct_write_complete(dreq); 853 return requested_bytes; 854 } 855 856 /** 857 * nfs_file_direct_write - file direct write operation for NFS files 858 * @iocb: target I/O control block 859 * @iter: vector of user buffers from which to write data 860 * @swap: flag indicating this is swap IO, not O_DIRECT IO 861 * 862 * We use this function for direct writes instead of calling 863 * generic_file_aio_write() in order to avoid taking the inode 864 * semaphore and updating the i_size. The NFS server will set 865 * the new i_size and this client must read the updated size 866 * back into its cache. We let the server do generic write 867 * parameter checking and report problems. 868 * 869 * We eliminate local atime updates, see direct read above. 870 * 871 * We avoid unnecessary page cache invalidations for normal cached 872 * readers of this file. 873 * 874 * Note that O_APPEND is not supported for NFS direct writes, as there 875 * is no atomic O_APPEND write facility in the NFS protocol. 876 */ 877 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter, 878 bool swap) 879 { 880 ssize_t result, requested; 881 size_t count; 882 struct file *file = iocb->ki_filp; 883 struct address_space *mapping = file->f_mapping; 884 struct inode *inode = mapping->host; 885 struct nfs_direct_req *dreq; 886 struct nfs_lock_context *l_ctx; 887 loff_t pos, end; 888 889 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", 890 file, iov_iter_count(iter), (long long) iocb->ki_pos); 891 892 if (swap) 893 /* bypass generic checks */ 894 result = iov_iter_count(iter); 895 else 896 result = generic_write_checks(iocb, iter); 897 if (result <= 0) 898 return result; 899 count = result; 900 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); 901 902 pos = iocb->ki_pos; 903 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT; 904 905 task_io_account_write(count); 906 907 result = -ENOMEM; 908 dreq = nfs_direct_req_alloc(); 909 if (!dreq) 910 goto out; 911 912 dreq->inode = inode; 913 dreq->bytes_left = dreq->max_count = count; 914 dreq->io_start = pos; 915 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); 916 l_ctx = nfs_get_lock_context(dreq->ctx); 917 if (IS_ERR(l_ctx)) { 918 result = PTR_ERR(l_ctx); 919 nfs_direct_req_release(dreq); 920 goto out_release; 921 } 922 dreq->l_ctx = l_ctx; 923 if (!is_sync_kiocb(iocb)) 924 dreq->iocb = iocb; 925 pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode); 926 927 if (swap) { 928 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos, 929 FLUSH_STABLE); 930 } else { 931 nfs_start_io_direct(inode); 932 933 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos, 934 FLUSH_COND_STABLE); 935 936 if (mapping->nrpages) { 937 invalidate_inode_pages2_range(mapping, 938 pos >> PAGE_SHIFT, end); 939 } 940 941 nfs_end_io_direct(inode); 942 } 943 944 if (requested > 0) { 945 result = nfs_direct_wait(dreq); 946 if (result > 0) { 947 requested -= result; 948 iocb->ki_pos = pos + result; 949 /* XXX: should check the generic_write_sync retval */ 950 generic_write_sync(iocb, result); 951 } 952 iov_iter_revert(iter, requested); 953 } else { 954 result = requested; 955 } 956 nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE); 957 out_release: 958 nfs_direct_req_release(dreq); 959 out: 960 return result; 961 } 962 963 /** 964 * nfs_init_directcache - create a slab cache for nfs_direct_req structures 965 * 966 */ 967 int __init nfs_init_directcache(void) 968 { 969 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", 970 sizeof(struct nfs_direct_req), 971 0, (SLAB_RECLAIM_ACCOUNT| 972 SLAB_MEM_SPREAD), 973 NULL); 974 if (nfs_direct_cachep == NULL) 975 return -ENOMEM; 976 977 return 0; 978 } 979 980 /** 981 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures 982 * 983 */ 984 void nfs_destroy_directcache(void) 985 { 986 kmem_cache_destroy(nfs_direct_cachep); 987 } 988