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 64 #define NFSDBG_FACILITY NFSDBG_VFS 65 66 static struct kmem_cache *nfs_direct_cachep; 67 68 struct nfs_direct_req { 69 struct kref kref; /* release manager */ 70 71 /* I/O parameters */ 72 struct nfs_open_context *ctx; /* file open context info */ 73 struct nfs_lock_context *l_ctx; /* Lock context info */ 74 struct kiocb * iocb; /* controlling i/o request */ 75 struct inode * inode; /* target file of i/o */ 76 77 /* completion state */ 78 atomic_t io_count; /* i/os we're waiting for */ 79 spinlock_t lock; /* protect completion state */ 80 81 loff_t io_start; /* Start offset for I/O */ 82 ssize_t count, /* bytes actually processed */ 83 max_count, /* max expected count */ 84 bytes_left, /* bytes left to be sent */ 85 error; /* any reported error */ 86 struct completion completion; /* wait for i/o completion */ 87 88 /* commit state */ 89 struct nfs_mds_commit_info mds_cinfo; /* Storage for cinfo */ 90 struct pnfs_ds_commit_info ds_cinfo; /* Storage for cinfo */ 91 struct work_struct work; 92 int flags; 93 /* for write */ 94 #define NFS_ODIRECT_DO_COMMIT (1) /* an unstable reply was received */ 95 #define NFS_ODIRECT_RESCHED_WRITES (2) /* write verification failed */ 96 /* for read */ 97 #define NFS_ODIRECT_SHOULD_DIRTY (3) /* dirty user-space page after read */ 98 #define NFS_ODIRECT_DONE INT_MAX /* write verification failed */ 99 }; 100 101 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops; 102 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops; 103 static void nfs_direct_write_complete(struct nfs_direct_req *dreq); 104 static void nfs_direct_write_schedule_work(struct work_struct *work); 105 106 static inline void get_dreq(struct nfs_direct_req *dreq) 107 { 108 atomic_inc(&dreq->io_count); 109 } 110 111 static inline int put_dreq(struct nfs_direct_req *dreq) 112 { 113 return atomic_dec_and_test(&dreq->io_count); 114 } 115 116 static void 117 nfs_direct_handle_truncated(struct nfs_direct_req *dreq, 118 const struct nfs_pgio_header *hdr, 119 ssize_t dreq_len) 120 { 121 if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) || 122 test_bit(NFS_IOHDR_EOF, &hdr->flags))) 123 return; 124 if (dreq->max_count >= dreq_len) { 125 dreq->max_count = dreq_len; 126 if (dreq->count > dreq_len) 127 dreq->count = dreq_len; 128 129 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) 130 dreq->error = hdr->error; 131 else /* Clear outstanding error if this is EOF */ 132 dreq->error = 0; 133 } 134 } 135 136 static void 137 nfs_direct_count_bytes(struct nfs_direct_req *dreq, 138 const struct nfs_pgio_header *hdr) 139 { 140 loff_t hdr_end = hdr->io_start + hdr->good_bytes; 141 ssize_t dreq_len = 0; 142 143 if (hdr_end > dreq->io_start) 144 dreq_len = hdr_end - dreq->io_start; 145 146 nfs_direct_handle_truncated(dreq, hdr, dreq_len); 147 148 if (dreq_len > dreq->max_count) 149 dreq_len = dreq->max_count; 150 151 if (dreq->count < dreq_len) 152 dreq->count = dreq_len; 153 } 154 155 /** 156 * nfs_direct_IO - NFS address space operation for direct I/O 157 * @iocb: target I/O control block 158 * @iter: I/O buffer 159 * 160 * The presence of this routine in the address space ops vector means 161 * the NFS client supports direct I/O. However, for most direct IO, we 162 * shunt off direct read and write requests before the VFS gets them, 163 * so this method is only ever called for swap. 164 */ 165 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter) 166 { 167 struct inode *inode = iocb->ki_filp->f_mapping->host; 168 169 /* we only support swap file calling nfs_direct_IO */ 170 if (!IS_SWAPFILE(inode)) 171 return 0; 172 173 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE); 174 175 if (iov_iter_rw(iter) == READ) 176 return nfs_file_direct_read(iocb, iter, true); 177 return nfs_file_direct_write(iocb, iter, true); 178 } 179 180 static void nfs_direct_release_pages(struct page **pages, unsigned int npages) 181 { 182 unsigned int i; 183 for (i = 0; i < npages; i++) 184 put_page(pages[i]); 185 } 186 187 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo, 188 struct nfs_direct_req *dreq) 189 { 190 cinfo->inode = dreq->inode; 191 cinfo->mds = &dreq->mds_cinfo; 192 cinfo->ds = &dreq->ds_cinfo; 193 cinfo->dreq = dreq; 194 cinfo->completion_ops = &nfs_direct_commit_completion_ops; 195 } 196 197 static inline struct nfs_direct_req *nfs_direct_req_alloc(void) 198 { 199 struct nfs_direct_req *dreq; 200 201 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL); 202 if (!dreq) 203 return NULL; 204 205 kref_init(&dreq->kref); 206 kref_get(&dreq->kref); 207 init_completion(&dreq->completion); 208 INIT_LIST_HEAD(&dreq->mds_cinfo.list); 209 pnfs_init_ds_commit_info(&dreq->ds_cinfo); 210 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work); 211 spin_lock_init(&dreq->lock); 212 213 return dreq; 214 } 215 216 static void nfs_direct_req_free(struct kref *kref) 217 { 218 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref); 219 220 pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode); 221 if (dreq->l_ctx != NULL) 222 nfs_put_lock_context(dreq->l_ctx); 223 if (dreq->ctx != NULL) 224 put_nfs_open_context(dreq->ctx); 225 kmem_cache_free(nfs_direct_cachep, dreq); 226 } 227 228 static void nfs_direct_req_release(struct nfs_direct_req *dreq) 229 { 230 kref_put(&dreq->kref, nfs_direct_req_free); 231 } 232 233 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq) 234 { 235 return dreq->bytes_left; 236 } 237 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left); 238 239 /* 240 * Collects and returns the final error value/byte-count. 241 */ 242 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq) 243 { 244 ssize_t result = -EIOCBQUEUED; 245 246 /* Async requests don't wait here */ 247 if (dreq->iocb) 248 goto out; 249 250 result = wait_for_completion_killable(&dreq->completion); 251 252 if (!result) { 253 result = dreq->count; 254 WARN_ON_ONCE(dreq->count < 0); 255 } 256 if (!result) 257 result = dreq->error; 258 259 out: 260 return (ssize_t) result; 261 } 262 263 /* 264 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust 265 * the iocb is still valid here if this is a synchronous request. 266 */ 267 static void nfs_direct_complete(struct nfs_direct_req *dreq) 268 { 269 struct inode *inode = dreq->inode; 270 271 inode_dio_end(inode); 272 273 if (dreq->iocb) { 274 long res = (long) dreq->error; 275 if (dreq->count != 0) { 276 res = (long) dreq->count; 277 WARN_ON_ONCE(dreq->count < 0); 278 } 279 dreq->iocb->ki_complete(dreq->iocb, res); 280 } 281 282 complete(&dreq->completion); 283 284 nfs_direct_req_release(dreq); 285 } 286 287 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr) 288 { 289 unsigned long bytes = 0; 290 struct nfs_direct_req *dreq = hdr->dreq; 291 292 spin_lock(&dreq->lock); 293 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { 294 spin_unlock(&dreq->lock); 295 goto out_put; 296 } 297 298 nfs_direct_count_bytes(dreq, hdr); 299 spin_unlock(&dreq->lock); 300 301 while (!list_empty(&hdr->pages)) { 302 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 303 struct page *page = req->wb_page; 304 305 if (!PageCompound(page) && bytes < hdr->good_bytes && 306 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY)) 307 set_page_dirty(page); 308 bytes += req->wb_bytes; 309 nfs_list_remove_request(req); 310 nfs_release_request(req); 311 } 312 out_put: 313 if (put_dreq(dreq)) 314 nfs_direct_complete(dreq); 315 hdr->release(hdr); 316 } 317 318 static void nfs_read_sync_pgio_error(struct list_head *head, int error) 319 { 320 struct nfs_page *req; 321 322 while (!list_empty(head)) { 323 req = nfs_list_entry(head->next); 324 nfs_list_remove_request(req); 325 nfs_release_request(req); 326 } 327 } 328 329 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr) 330 { 331 get_dreq(hdr->dreq); 332 } 333 334 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = { 335 .error_cleanup = nfs_read_sync_pgio_error, 336 .init_hdr = nfs_direct_pgio_init, 337 .completion = nfs_direct_read_completion, 338 }; 339 340 /* 341 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ 342 * operation. If nfs_readdata_alloc() or get_user_pages() fails, 343 * bail and stop sending more reads. Read length accounting is 344 * handled automatically by nfs_direct_read_result(). Otherwise, if 345 * no requests have been sent, just return an error. 346 */ 347 348 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq, 349 struct iov_iter *iter, 350 loff_t pos) 351 { 352 struct nfs_pageio_descriptor desc; 353 struct inode *inode = dreq->inode; 354 ssize_t result = -EINVAL; 355 size_t requested_bytes = 0; 356 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE); 357 358 nfs_pageio_init_read(&desc, dreq->inode, false, 359 &nfs_direct_read_completion_ops); 360 get_dreq(dreq); 361 desc.pg_dreq = dreq; 362 inode_dio_begin(inode); 363 364 while (iov_iter_count(iter)) { 365 struct page **pagevec; 366 size_t bytes; 367 size_t pgbase; 368 unsigned npages, i; 369 370 result = iov_iter_get_pages_alloc(iter, &pagevec, 371 rsize, &pgbase); 372 if (result < 0) 373 break; 374 375 bytes = result; 376 iov_iter_advance(iter, bytes); 377 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; 378 for (i = 0; i < npages; i++) { 379 struct nfs_page *req; 380 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); 381 /* XXX do we need to do the eof zeroing found in async_filler? */ 382 req = nfs_create_request(dreq->ctx, pagevec[i], 383 pgbase, req_len); 384 if (IS_ERR(req)) { 385 result = PTR_ERR(req); 386 break; 387 } 388 req->wb_index = pos >> PAGE_SHIFT; 389 req->wb_offset = pos & ~PAGE_MASK; 390 if (!nfs_pageio_add_request(&desc, req)) { 391 result = desc.pg_error; 392 nfs_release_request(req); 393 break; 394 } 395 pgbase = 0; 396 bytes -= req_len; 397 requested_bytes += req_len; 398 pos += req_len; 399 dreq->bytes_left -= req_len; 400 } 401 nfs_direct_release_pages(pagevec, npages); 402 kvfree(pagevec); 403 if (result < 0) 404 break; 405 } 406 407 nfs_pageio_complete(&desc); 408 409 /* 410 * If no bytes were started, return the error, and let the 411 * generic layer handle the completion. 412 */ 413 if (requested_bytes == 0) { 414 inode_dio_end(inode); 415 nfs_direct_req_release(dreq); 416 return result < 0 ? result : -EIO; 417 } 418 419 if (put_dreq(dreq)) 420 nfs_direct_complete(dreq); 421 return requested_bytes; 422 } 423 424 /** 425 * nfs_file_direct_read - file direct read operation for NFS files 426 * @iocb: target I/O control block 427 * @iter: vector of user buffers into which to read data 428 * @swap: flag indicating this is swap IO, not O_DIRECT IO 429 * 430 * We use this function for direct reads instead of calling 431 * generic_file_aio_read() in order to avoid gfar's check to see if 432 * the request starts before the end of the file. For that check 433 * to work, we must generate a GETATTR before each direct read, and 434 * even then there is a window between the GETATTR and the subsequent 435 * READ where the file size could change. Our preference is simply 436 * to do all reads the application wants, and the server will take 437 * care of managing the end of file boundary. 438 * 439 * This function also eliminates unnecessarily updating the file's 440 * atime locally, as the NFS server sets the file's atime, and this 441 * client must read the updated atime from the server back into its 442 * cache. 443 */ 444 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter, 445 bool swap) 446 { 447 struct file *file = iocb->ki_filp; 448 struct address_space *mapping = file->f_mapping; 449 struct inode *inode = mapping->host; 450 struct nfs_direct_req *dreq; 451 struct nfs_lock_context *l_ctx; 452 ssize_t result, requested; 453 size_t count = iov_iter_count(iter); 454 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count); 455 456 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n", 457 file, count, (long long) iocb->ki_pos); 458 459 result = 0; 460 if (!count) 461 goto out; 462 463 task_io_account_read(count); 464 465 result = -ENOMEM; 466 dreq = nfs_direct_req_alloc(); 467 if (dreq == NULL) 468 goto out; 469 470 dreq->inode = inode; 471 dreq->bytes_left = dreq->max_count = count; 472 dreq->io_start = iocb->ki_pos; 473 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); 474 l_ctx = nfs_get_lock_context(dreq->ctx); 475 if (IS_ERR(l_ctx)) { 476 result = PTR_ERR(l_ctx); 477 nfs_direct_req_release(dreq); 478 goto out_release; 479 } 480 dreq->l_ctx = l_ctx; 481 if (!is_sync_kiocb(iocb)) 482 dreq->iocb = iocb; 483 484 if (iter_is_iovec(iter)) 485 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY; 486 487 if (!swap) 488 nfs_start_io_direct(inode); 489 490 NFS_I(inode)->read_io += count; 491 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos); 492 493 if (!swap) 494 nfs_end_io_direct(inode); 495 496 if (requested > 0) { 497 result = nfs_direct_wait(dreq); 498 if (result > 0) { 499 requested -= result; 500 iocb->ki_pos += result; 501 } 502 iov_iter_revert(iter, requested); 503 } else { 504 result = requested; 505 } 506 507 out_release: 508 nfs_direct_req_release(dreq); 509 out: 510 return result; 511 } 512 513 static void 514 nfs_direct_join_group(struct list_head *list, struct inode *inode) 515 { 516 struct nfs_page *req, *next; 517 518 list_for_each_entry(req, list, wb_list) { 519 if (req->wb_head != req || req->wb_this_page == req) 520 continue; 521 for (next = req->wb_this_page; 522 next != req->wb_head; 523 next = next->wb_this_page) { 524 nfs_list_remove_request(next); 525 nfs_release_request(next); 526 } 527 nfs_join_page_group(req, inode); 528 } 529 } 530 531 static void 532 nfs_direct_write_scan_commit_list(struct inode *inode, 533 struct list_head *list, 534 struct nfs_commit_info *cinfo) 535 { 536 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex); 537 pnfs_recover_commit_reqs(list, cinfo); 538 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0); 539 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex); 540 } 541 542 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq) 543 { 544 struct nfs_pageio_descriptor desc; 545 struct nfs_page *req, *tmp; 546 LIST_HEAD(reqs); 547 struct nfs_commit_info cinfo; 548 LIST_HEAD(failed); 549 550 nfs_init_cinfo_from_dreq(&cinfo, dreq); 551 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); 552 553 nfs_direct_join_group(&reqs, dreq->inode); 554 555 dreq->count = 0; 556 dreq->max_count = 0; 557 list_for_each_entry(req, &reqs, wb_list) 558 dreq->max_count += req->wb_bytes; 559 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo); 560 get_dreq(dreq); 561 562 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false, 563 &nfs_direct_write_completion_ops); 564 desc.pg_dreq = dreq; 565 566 list_for_each_entry_safe(req, tmp, &reqs, wb_list) { 567 /* Bump the transmission count */ 568 req->wb_nio++; 569 if (!nfs_pageio_add_request(&desc, req)) { 570 nfs_list_move_request(req, &failed); 571 spin_lock(&cinfo.inode->i_lock); 572 dreq->flags = 0; 573 if (desc.pg_error < 0) 574 dreq->error = desc.pg_error; 575 else 576 dreq->error = -EIO; 577 spin_unlock(&cinfo.inode->i_lock); 578 } 579 nfs_release_request(req); 580 } 581 nfs_pageio_complete(&desc); 582 583 while (!list_empty(&failed)) { 584 req = nfs_list_entry(failed.next); 585 nfs_list_remove_request(req); 586 nfs_unlock_and_release_request(req); 587 } 588 589 if (put_dreq(dreq)) 590 nfs_direct_write_complete(dreq); 591 } 592 593 static void nfs_direct_commit_complete(struct nfs_commit_data *data) 594 { 595 const struct nfs_writeverf *verf = data->res.verf; 596 struct nfs_direct_req *dreq = data->dreq; 597 struct nfs_commit_info cinfo; 598 struct nfs_page *req; 599 int status = data->task.tk_status; 600 601 if (status < 0) { 602 /* Errors in commit are fatal */ 603 dreq->error = status; 604 dreq->max_count = 0; 605 dreq->count = 0; 606 dreq->flags = NFS_ODIRECT_DONE; 607 } else if (dreq->flags == NFS_ODIRECT_DONE) 608 status = dreq->error; 609 610 nfs_init_cinfo_from_dreq(&cinfo, dreq); 611 612 while (!list_empty(&data->pages)) { 613 req = nfs_list_entry(data->pages.next); 614 nfs_list_remove_request(req); 615 if (status >= 0 && !nfs_write_match_verf(verf, req)) { 616 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; 617 /* 618 * Despite the reboot, the write was successful, 619 * so reset wb_nio. 620 */ 621 req->wb_nio = 0; 622 nfs_mark_request_commit(req, NULL, &cinfo, 0); 623 } else /* Error or match */ 624 nfs_release_request(req); 625 nfs_unlock_and_release_request(req); 626 } 627 628 if (nfs_commit_end(cinfo.mds)) 629 nfs_direct_write_complete(dreq); 630 } 631 632 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo, 633 struct nfs_page *req) 634 { 635 struct nfs_direct_req *dreq = cinfo->dreq; 636 637 spin_lock(&dreq->lock); 638 if (dreq->flags != NFS_ODIRECT_DONE) 639 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; 640 spin_unlock(&dreq->lock); 641 nfs_mark_request_commit(req, NULL, cinfo, 0); 642 } 643 644 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = { 645 .completion = nfs_direct_commit_complete, 646 .resched_write = nfs_direct_resched_write, 647 }; 648 649 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq) 650 { 651 int res; 652 struct nfs_commit_info cinfo; 653 LIST_HEAD(mds_list); 654 655 nfs_init_cinfo_from_dreq(&cinfo, dreq); 656 nfs_scan_commit(dreq->inode, &mds_list, &cinfo); 657 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo); 658 if (res < 0) /* res == -ENOMEM */ 659 nfs_direct_write_reschedule(dreq); 660 } 661 662 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq) 663 { 664 struct nfs_commit_info cinfo; 665 struct nfs_page *req; 666 LIST_HEAD(reqs); 667 668 nfs_init_cinfo_from_dreq(&cinfo, dreq); 669 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo); 670 671 while (!list_empty(&reqs)) { 672 req = nfs_list_entry(reqs.next); 673 nfs_list_remove_request(req); 674 nfs_release_request(req); 675 nfs_unlock_and_release_request(req); 676 } 677 } 678 679 static void nfs_direct_write_schedule_work(struct work_struct *work) 680 { 681 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work); 682 int flags = dreq->flags; 683 684 dreq->flags = 0; 685 switch (flags) { 686 case NFS_ODIRECT_DO_COMMIT: 687 nfs_direct_commit_schedule(dreq); 688 break; 689 case NFS_ODIRECT_RESCHED_WRITES: 690 nfs_direct_write_reschedule(dreq); 691 break; 692 default: 693 nfs_direct_write_clear_reqs(dreq); 694 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping); 695 nfs_direct_complete(dreq); 696 } 697 } 698 699 static void nfs_direct_write_complete(struct nfs_direct_req *dreq) 700 { 701 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */ 702 } 703 704 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr) 705 { 706 struct nfs_direct_req *dreq = hdr->dreq; 707 struct nfs_commit_info cinfo; 708 struct nfs_page *req = nfs_list_entry(hdr->pages.next); 709 int flags = NFS_ODIRECT_DONE; 710 711 nfs_init_cinfo_from_dreq(&cinfo, dreq); 712 713 spin_lock(&dreq->lock); 714 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) { 715 spin_unlock(&dreq->lock); 716 goto out_put; 717 } 718 719 nfs_direct_count_bytes(dreq, hdr); 720 if (hdr->good_bytes != 0 && nfs_write_need_commit(hdr)) { 721 if (!dreq->flags) 722 dreq->flags = NFS_ODIRECT_DO_COMMIT; 723 flags = dreq->flags; 724 } 725 spin_unlock(&dreq->lock); 726 727 while (!list_empty(&hdr->pages)) { 728 729 req = nfs_list_entry(hdr->pages.next); 730 nfs_list_remove_request(req); 731 if (flags == NFS_ODIRECT_DO_COMMIT) { 732 kref_get(&req->wb_kref); 733 memcpy(&req->wb_verf, &hdr->verf.verifier, 734 sizeof(req->wb_verf)); 735 nfs_mark_request_commit(req, hdr->lseg, &cinfo, 736 hdr->ds_commit_idx); 737 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) { 738 kref_get(&req->wb_kref); 739 nfs_mark_request_commit(req, NULL, &cinfo, 0); 740 } 741 nfs_unlock_and_release_request(req); 742 } 743 744 out_put: 745 if (put_dreq(dreq)) 746 nfs_direct_write_complete(dreq); 747 hdr->release(hdr); 748 } 749 750 static void nfs_write_sync_pgio_error(struct list_head *head, int error) 751 { 752 struct nfs_page *req; 753 754 while (!list_empty(head)) { 755 req = nfs_list_entry(head->next); 756 nfs_list_remove_request(req); 757 nfs_unlock_and_release_request(req); 758 } 759 } 760 761 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr) 762 { 763 struct nfs_direct_req *dreq = hdr->dreq; 764 765 spin_lock(&dreq->lock); 766 if (dreq->error == 0) { 767 dreq->flags = NFS_ODIRECT_RESCHED_WRITES; 768 /* fake unstable write to let common nfs resend pages */ 769 hdr->verf.committed = NFS_UNSTABLE; 770 hdr->good_bytes = hdr->args.offset + hdr->args.count - 771 hdr->io_start; 772 } 773 spin_unlock(&dreq->lock); 774 } 775 776 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = { 777 .error_cleanup = nfs_write_sync_pgio_error, 778 .init_hdr = nfs_direct_pgio_init, 779 .completion = nfs_direct_write_completion, 780 .reschedule_io = nfs_direct_write_reschedule_io, 781 }; 782 783 784 /* 785 * NB: Return the value of the first error return code. Subsequent 786 * errors after the first one are ignored. 787 */ 788 /* 789 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE 790 * operation. If nfs_writedata_alloc() or get_user_pages() fails, 791 * bail and stop sending more writes. Write length accounting is 792 * handled automatically by nfs_direct_write_result(). Otherwise, if 793 * no requests have been sent, just return an error. 794 */ 795 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq, 796 struct iov_iter *iter, 797 loff_t pos, int ioflags) 798 { 799 struct nfs_pageio_descriptor desc; 800 struct inode *inode = dreq->inode; 801 ssize_t result = 0; 802 size_t requested_bytes = 0; 803 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE); 804 805 nfs_pageio_init_write(&desc, inode, ioflags, false, 806 &nfs_direct_write_completion_ops); 807 desc.pg_dreq = dreq; 808 get_dreq(dreq); 809 inode_dio_begin(inode); 810 811 NFS_I(inode)->write_io += iov_iter_count(iter); 812 while (iov_iter_count(iter)) { 813 struct page **pagevec; 814 size_t bytes; 815 size_t pgbase; 816 unsigned npages, i; 817 818 result = iov_iter_get_pages_alloc(iter, &pagevec, 819 wsize, &pgbase); 820 if (result < 0) 821 break; 822 823 bytes = result; 824 iov_iter_advance(iter, bytes); 825 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE; 826 for (i = 0; i < npages; i++) { 827 struct nfs_page *req; 828 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase); 829 830 req = nfs_create_request(dreq->ctx, pagevec[i], 831 pgbase, req_len); 832 if (IS_ERR(req)) { 833 result = PTR_ERR(req); 834 break; 835 } 836 837 if (desc.pg_error < 0) { 838 nfs_free_request(req); 839 result = desc.pg_error; 840 break; 841 } 842 843 nfs_lock_request(req); 844 req->wb_index = pos >> PAGE_SHIFT; 845 req->wb_offset = pos & ~PAGE_MASK; 846 if (!nfs_pageio_add_request(&desc, req)) { 847 result = desc.pg_error; 848 nfs_unlock_and_release_request(req); 849 break; 850 } 851 pgbase = 0; 852 bytes -= req_len; 853 requested_bytes += req_len; 854 pos += req_len; 855 dreq->bytes_left -= req_len; 856 } 857 nfs_direct_release_pages(pagevec, npages); 858 kvfree(pagevec); 859 if (result < 0) 860 break; 861 } 862 nfs_pageio_complete(&desc); 863 864 /* 865 * If no bytes were started, return the error, and let the 866 * generic layer handle the completion. 867 */ 868 if (requested_bytes == 0) { 869 inode_dio_end(inode); 870 nfs_direct_req_release(dreq); 871 return result < 0 ? result : -EIO; 872 } 873 874 if (put_dreq(dreq)) 875 nfs_direct_write_complete(dreq); 876 return requested_bytes; 877 } 878 879 /** 880 * nfs_file_direct_write - file direct write operation for NFS files 881 * @iocb: target I/O control block 882 * @iter: vector of user buffers from which to write data 883 * @swap: flag indicating this is swap IO, not O_DIRECT IO 884 * 885 * We use this function for direct writes instead of calling 886 * generic_file_aio_write() in order to avoid taking the inode 887 * semaphore and updating the i_size. The NFS server will set 888 * the new i_size and this client must read the updated size 889 * back into its cache. We let the server do generic write 890 * parameter checking and report problems. 891 * 892 * We eliminate local atime updates, see direct read above. 893 * 894 * We avoid unnecessary page cache invalidations for normal cached 895 * readers of this file. 896 * 897 * Note that O_APPEND is not supported for NFS direct writes, as there 898 * is no atomic O_APPEND write facility in the NFS protocol. 899 */ 900 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter, 901 bool swap) 902 { 903 ssize_t result, requested; 904 size_t count; 905 struct file *file = iocb->ki_filp; 906 struct address_space *mapping = file->f_mapping; 907 struct inode *inode = mapping->host; 908 struct nfs_direct_req *dreq; 909 struct nfs_lock_context *l_ctx; 910 loff_t pos, end; 911 912 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n", 913 file, iov_iter_count(iter), (long long) iocb->ki_pos); 914 915 if (swap) 916 /* bypass generic checks */ 917 result = iov_iter_count(iter); 918 else 919 result = generic_write_checks(iocb, iter); 920 if (result <= 0) 921 return result; 922 count = result; 923 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count); 924 925 pos = iocb->ki_pos; 926 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT; 927 928 task_io_account_write(count); 929 930 result = -ENOMEM; 931 dreq = nfs_direct_req_alloc(); 932 if (!dreq) 933 goto out; 934 935 dreq->inode = inode; 936 dreq->bytes_left = dreq->max_count = count; 937 dreq->io_start = pos; 938 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp)); 939 l_ctx = nfs_get_lock_context(dreq->ctx); 940 if (IS_ERR(l_ctx)) { 941 result = PTR_ERR(l_ctx); 942 nfs_direct_req_release(dreq); 943 goto out_release; 944 } 945 dreq->l_ctx = l_ctx; 946 if (!is_sync_kiocb(iocb)) 947 dreq->iocb = iocb; 948 pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode); 949 950 if (swap) { 951 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos, 952 FLUSH_STABLE); 953 } else { 954 nfs_start_io_direct(inode); 955 956 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos, 957 FLUSH_COND_STABLE); 958 959 if (mapping->nrpages) { 960 invalidate_inode_pages2_range(mapping, 961 pos >> PAGE_SHIFT, end); 962 } 963 964 nfs_end_io_direct(inode); 965 } 966 967 if (requested > 0) { 968 result = nfs_direct_wait(dreq); 969 if (result > 0) { 970 requested -= result; 971 iocb->ki_pos = pos + result; 972 /* XXX: should check the generic_write_sync retval */ 973 generic_write_sync(iocb, result); 974 } 975 iov_iter_revert(iter, requested); 976 } else { 977 result = requested; 978 } 979 nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE); 980 out_release: 981 nfs_direct_req_release(dreq); 982 out: 983 return result; 984 } 985 986 /** 987 * nfs_init_directcache - create a slab cache for nfs_direct_req structures 988 * 989 */ 990 int __init nfs_init_directcache(void) 991 { 992 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache", 993 sizeof(struct nfs_direct_req), 994 0, (SLAB_RECLAIM_ACCOUNT| 995 SLAB_MEM_SPREAD), 996 NULL); 997 if (nfs_direct_cachep == NULL) 998 return -ENOMEM; 999 1000 return 0; 1001 } 1002 1003 /** 1004 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures 1005 * 1006 */ 1007 void nfs_destroy_directcache(void) 1008 { 1009 kmem_cache_destroy(nfs_direct_cachep); 1010 } 1011