fs/nfs/direct.c

   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
  98         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && !dreq->error)
  99                 dreq->error = hdr->error;
 100 }
 101
 102 static void
 103 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
 104                        const struct nfs_pgio_header *hdr)
 105 {
 106         loff_t hdr_end = hdr->io_start + hdr->good_bytes;
 107         ssize_t dreq_len = 0;
 108
 109         if (hdr_end > dreq->io_start)
 110                 dreq_len = hdr_end - dreq->io_start;
 111
 112         nfs_direct_handle_truncated(dreq, hdr, dreq_len);
 113
 114         if (dreq_len > dreq->max_count)
 115                 dreq_len = dreq->max_count;
 116
 117         if (dreq->count < dreq_len)
 118                 dreq->count = dreq_len;
 119 }
 120
 121 static void nfs_direct_truncate_request(struct nfs_direct_req *dreq,
 122                                         struct nfs_page *req)
 123 {
 124         loff_t offs = req_offset(req);
 125         size_t req_start = (size_t)(offs - dreq->io_start);
 126
 127         if (req_start < dreq->max_count)
 128                 dreq->max_count = req_start;
 129         if (req_start < dreq->count)
 130                 dreq->count = req_start;
 131 }
 132
 133 /**
 134  * nfs_swap_rw - NFS address space operation for swap I/O
 135  * @iocb: target I/O control block
 136  * @iter: I/O buffer
 137  *
 138  * Perform IO to the swap-file.  This is much like direct IO.
 139  */
 140 int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter)
 141 {
 142         ssize_t ret;
 143
 144         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 145
 146         if (iov_iter_rw(iter) == READ)
 147                 ret = nfs_file_direct_read(iocb, iter, true);
 148         else
 149                 ret = nfs_file_direct_write(iocb, iter, true);
 150         if (ret < 0)
 151                 return ret;
 152         return 0;
 153 }
 154
 155 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 156 {
 157         unsigned int i;
 158         for (i = 0; i < npages; i++)
 159                 put_page(pages[i]);
 160 }
 161
 162 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 163                               struct nfs_direct_req *dreq)
 164 {
 165         cinfo->inode = dreq->inode;
 166         cinfo->mds = &dreq->mds_cinfo;
 167         cinfo->ds = &dreq->ds_cinfo;
 168         cinfo->dreq = dreq;
 169         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 170 }
 171
 172 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 173 {
 174         struct nfs_direct_req *dreq;
 175
 176         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 177         if (!dreq)
 178                 return NULL;
 179
 180         kref_init(&dreq->kref);
 181         kref_get(&dreq->kref);
 182         init_completion(&dreq->completion);
 183         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 184         pnfs_init_ds_commit_info(&dreq->ds_cinfo);
 185         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 186         spin_lock_init(&dreq->lock);
 187
 188         return dreq;
 189 }
 190
 191 static void nfs_direct_req_free(struct kref *kref)
 192 {
 193         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 194
 195         pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
 196         if (dreq->l_ctx != NULL)
 197                 nfs_put_lock_context(dreq->l_ctx);
 198         if (dreq->ctx != NULL)
 199                 put_nfs_open_context(dreq->ctx);
 200         kmem_cache_free(nfs_direct_cachep, dreq);
 201 }
 202
 203 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 204 {
 205         kref_put(&dreq->kref, nfs_direct_req_free);
 206 }
 207
 208 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 209 {
 210         return dreq->bytes_left;
 211 }
 212 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 213
 214 /*
 215  * Collects and returns the final error value/byte-count.
 216  */
 217 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 218 {
 219         ssize_t result = -EIOCBQUEUED;
 220
 221         /* Async requests don't wait here */
 222         if (dreq->iocb)
 223                 goto out;
 224
 225         result = wait_for_completion_killable(&dreq->completion);
 226
 227         if (!result) {
 228                 result = dreq->count;
 229                 WARN_ON_ONCE(dreq->count < 0);
 230         }
 231         if (!result)
 232                 result = dreq->error;
 233
 234 out:
 235         return (ssize_t) result;
 236 }
 237
 238 /*
 239  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 240  * the iocb is still valid here if this is a synchronous request.
 241  */
 242 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 243 {
 244         struct inode *inode = dreq->inode;
 245
 246         inode_dio_end(inode);
 247
 248         if (dreq->iocb) {
 249                 long res = (long) dreq->error;
 250                 if (dreq->count != 0) {
 251                         res = (long) dreq->count;
 252                         WARN_ON_ONCE(dreq->count < 0);
 253                 }
 254                 dreq->iocb->ki_complete(dreq->iocb, res);
 255         }
 256
 257         complete(&dreq->completion);
 258
 259         nfs_direct_req_release(dreq);
 260 }
 261
 262 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 263 {
 264         unsigned long bytes = 0;
 265         struct nfs_direct_req *dreq = hdr->dreq;
 266
 267         spin_lock(&dreq->lock);
 268         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 269                 spin_unlock(&dreq->lock);
 270                 goto out_put;
 271         }
 272
 273         nfs_direct_count_bytes(dreq, hdr);
 274         spin_unlock(&dreq->lock);
 275
 276         while (!list_empty(&hdr->pages)) {
 277                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 278                 struct page *page = req->wb_page;
 279
 280                 if (!PageCompound(page) && bytes < hdr->good_bytes &&
 281                     (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 282                         set_page_dirty(page);
 283                 bytes += req->wb_bytes;
 284                 nfs_list_remove_request(req);
 285                 nfs_release_request(req);
 286         }
 287 out_put:
 288         if (put_dreq(dreq))
 289                 nfs_direct_complete(dreq);
 290         hdr->release(hdr);
 291 }
 292
 293 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 294 {
 295         struct nfs_page *req;
 296
 297         while (!list_empty(head)) {
 298                 req = nfs_list_entry(head->next);
 299                 nfs_list_remove_request(req);
 300                 nfs_release_request(req);
 301         }
 302 }
 303
 304 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 305 {
 306         get_dreq(hdr->dreq);
 307 }
 308
 309 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 310         .error_cleanup = nfs_read_sync_pgio_error,
 311         .init_hdr = nfs_direct_pgio_init,
 312         .completion = nfs_direct_read_completion,
 313 };
 314
 315 /*
 316  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 317  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 318  * bail and stop sending more reads.  Read length accounting is
 319  * handled automatically by nfs_direct_read_result().  Otherwise, if
 320  * no requests have been sent, just return an error.
 321  */
 322
 323 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 324                                               struct iov_iter *iter,
 325                                               loff_t pos)
 326 {
 327         struct nfs_pageio_descriptor desc;
 328         struct inode *inode = dreq->inode;
 329         ssize_t result = -EINVAL;
 330         size_t requested_bytes = 0;
 331         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 332
 333         nfs_pageio_init_read(&desc, dreq->inode, false,
 334                              &nfs_direct_read_completion_ops);
 335         get_dreq(dreq);
 336         desc.pg_dreq = dreq;
 337         inode_dio_begin(inode);
 338
 339         while (iov_iter_count(iter)) {
 340                 struct page **pagevec;
 341                 size_t bytes;
 342                 size_t pgbase;
 343                 unsigned npages, i;
 344
 345                 result = iov_iter_get_pages_alloc2(iter, &pagevec,
 346                                                   rsize, &pgbase);
 347                 if (result < 0)
 348                         break;
 349
 350                 bytes = result;
 351                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 352                 for (i = 0; i < npages; i++) {
 353                         struct nfs_page *req;
 354                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 355                         /* XXX do we need to do the eof zeroing found in async_filler? */
 356                         req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
 357                                                         pgbase, pos, req_len);
 358                         if (IS_ERR(req)) {
 359                                 result = PTR_ERR(req);
 360                                 break;
 361                         }
 362                         if (!nfs_pageio_add_request(&desc, req)) {
 363                                 result = desc.pg_error;
 364                                 nfs_release_request(req);
 365                                 break;
 366                         }
 367                         pgbase = 0;
 368                         bytes -= req_len;
 369                         requested_bytes += req_len;
 370                         pos += req_len;
 371                         dreq->bytes_left -= req_len;
 372                 }
 373                 nfs_direct_release_pages(pagevec, npages);
 374                 kvfree(pagevec);
 375                 if (result < 0)
 376                         break;
 377         }
 378
 379         nfs_pageio_complete(&desc);
 380
 381         /*
 382          * If no bytes were started, return the error, and let the
 383          * generic layer handle the completion.
 384          */
 385         if (requested_bytes == 0) {
 386                 inode_dio_end(inode);
 387                 nfs_direct_req_release(dreq);
 388                 return result < 0 ? result : -EIO;
 389         }
 390
 391         if (put_dreq(dreq))
 392                 nfs_direct_complete(dreq);
 393         return requested_bytes;
 394 }
 395
 396 /**
 397  * nfs_file_direct_read - file direct read operation for NFS files
 398  * @iocb: target I/O control block
 399  * @iter: vector of user buffers into which to read data
 400  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 401  *
 402  * We use this function for direct reads instead of calling
 403  * generic_file_aio_read() in order to avoid gfar's check to see if
 404  * the request starts before the end of the file.  For that check
 405  * to work, we must generate a GETATTR before each direct read, and
 406  * even then there is a window between the GETATTR and the subsequent
 407  * READ where the file size could change.  Our preference is simply
 408  * to do all reads the application wants, and the server will take
 409  * care of managing the end of file boundary.
 410  *
 411  * This function also eliminates unnecessarily updating the file's
 412  * atime locally, as the NFS server sets the file's atime, and this
 413  * client must read the updated atime from the server back into its
 414  * cache.
 415  */
 416 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 417                              bool swap)
 418 {
 419         struct file *file = iocb->ki_filp;
 420         struct address_space *mapping = file->f_mapping;
 421         struct inode *inode = mapping->host;
 422         struct nfs_direct_req *dreq;
 423         struct nfs_lock_context *l_ctx;
 424         ssize_t result, requested;
 425         size_t count = iov_iter_count(iter);
 426         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 427
 428         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 429                 file, count, (long long) iocb->ki_pos);
 430
 431         result = 0;
 432         if (!count)
 433                 goto out;
 434
 435         task_io_account_read(count);
 436
 437         result = -ENOMEM;
 438         dreq = nfs_direct_req_alloc();
 439         if (dreq == NULL)
 440                 goto out;
 441
 442         dreq->inode = inode;
 443         dreq->bytes_left = dreq->max_count = count;
 444         dreq->io_start = iocb->ki_pos;
 445         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 446         l_ctx = nfs_get_lock_context(dreq->ctx);
 447         if (IS_ERR(l_ctx)) {
 448                 result = PTR_ERR(l_ctx);
 449                 nfs_direct_req_release(dreq);
 450                 goto out_release;
 451         }
 452         dreq->l_ctx = l_ctx;
 453         if (!is_sync_kiocb(iocb))
 454                 dreq->iocb = iocb;
 455
 456         if (user_backed_iter(iter))
 457                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 458
 459         if (!swap)
 460                 nfs_start_io_direct(inode);
 461
 462         NFS_I(inode)->read_io += count;
 463         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 464
 465         if (!swap)
 466                 nfs_end_io_direct(inode);
 467
 468         if (requested > 0) {
 469                 result = nfs_direct_wait(dreq);
 470                 if (result > 0) {
 471                         requested -= result;
 472                         iocb->ki_pos += result;
 473                 }
 474                 iov_iter_revert(iter, requested);
 475         } else {
 476                 result = requested;
 477         }
 478
 479 out_release:
 480         nfs_direct_req_release(dreq);
 481 out:
 482         return result;
 483 }
 484
 485 static void nfs_direct_add_page_head(struct list_head *list,
 486                                      struct nfs_page *req)
 487 {
 488         struct nfs_page *head = req->wb_head;
 489
 490         if (!list_empty(&head->wb_list) || !nfs_lock_request(head))
 491                 return;
 492         if (!list_empty(&head->wb_list)) {
 493                 nfs_unlock_request(head);
 494                 return;
 495         }
 496         list_add(&head->wb_list, list);
 497         kref_get(&head->wb_kref);
 498         kref_get(&head->wb_kref);
 499 }
 500
 501 static void nfs_direct_join_group(struct list_head *list,
 502                                   struct nfs_commit_info *cinfo,
 503                                   struct inode *inode)
 504 {
 505         struct nfs_page *req, *subreq;
 506
 507         list_for_each_entry(req, list, wb_list) {
 508                 if (req->wb_head != req) {
 509                         nfs_direct_add_page_head(&req->wb_list, req);
 510                         continue;
 511                 }
 512                 subreq = req->wb_this_page;
 513                 if (subreq == req)
 514                         continue;
 515                 do {
 516                         /*
 517                          * Remove subrequests from this list before freeing
 518                          * them in the call to nfs_join_page_group().
 519                          */
 520                         if (!list_empty(&subreq->wb_list)) {
 521                                 nfs_list_remove_request(subreq);
 522                                 nfs_release_request(subreq);
 523                         }
 524                 } while ((subreq = subreq->wb_this_page) != req);
 525                 nfs_join_page_group(req, cinfo, inode);
 526         }
 527 }
 528
 529 static void
 530 nfs_direct_write_scan_commit_list(struct inode *inode,
 531                                   struct list_head *list,
 532                                   struct nfs_commit_info *cinfo)
 533 {
 534         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 535         pnfs_recover_commit_reqs(list, cinfo);
 536         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 537         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 538 }
 539
 540 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 541 {
 542         struct nfs_pageio_descriptor desc;
 543         struct nfs_page *req;
 544         LIST_HEAD(reqs);
 545         struct nfs_commit_info cinfo;
 546
 547         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 548         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 549
 550         nfs_direct_join_group(&reqs, &cinfo, dreq->inode);
 551
 552         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 553         get_dreq(dreq);
 554
 555         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 556                               &nfs_direct_write_completion_ops);
 557         desc.pg_dreq = dreq;
 558
 559         while (!list_empty(&reqs)) {
 560                 req = nfs_list_entry(reqs.next);
 561                 /* Bump the transmission count */
 562                 req->wb_nio++;
 563                 if (!nfs_pageio_add_request(&desc, req)) {
 564                         spin_lock(&dreq->lock);
 565                         if (dreq->error < 0) {
 566                                 desc.pg_error = dreq->error;
 567                         } else if (desc.pg_error != -EAGAIN) {
 568                                 dreq->flags = 0;
 569                                 if (!desc.pg_error)
 570                                         desc.pg_error = -EIO;
 571                                 dreq->error = desc.pg_error;
 572                         } else
 573                                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 574                         spin_unlock(&dreq->lock);
 575                         break;
 576                 }
 577                 nfs_release_request(req);
 578         }
 579         nfs_pageio_complete(&desc);
 580
 581         while (!list_empty(&reqs)) {
 582                 req = nfs_list_entry(reqs.next);
 583                 nfs_list_remove_request(req);
 584                 nfs_unlock_and_release_request(req);
 585                 if (desc.pg_error == -EAGAIN) {
 586                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 587                 } else {
 588                         spin_lock(&dreq->lock);
 589                         nfs_direct_truncate_request(dreq, req);
 590                         spin_unlock(&dreq->lock);
 591                         nfs_release_request(req);
 592                 }
 593         }
 594
 595         if (put_dreq(dreq))
 596                 nfs_direct_write_complete(dreq);
 597 }
 598
 599 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 600 {
 601         const struct nfs_writeverf *verf = data->res.verf;
 602         struct nfs_direct_req *dreq = data->dreq;
 603         struct nfs_commit_info cinfo;
 604         struct nfs_page *req;
 605         int status = data->task.tk_status;
 606
 607         trace_nfs_direct_commit_complete(dreq);
 608
 609         if (status < 0) {
 610                 /* Errors in commit are fatal */
 611                 dreq->error = status;
 612                 dreq->flags = NFS_ODIRECT_DONE;
 613         } else {
 614                 status = dreq->error;
 615         }
 616
 617         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 618
 619         while (!list_empty(&data->pages)) {
 620                 req = nfs_list_entry(data->pages.next);
 621                 nfs_list_remove_request(req);
 622                 if (status < 0) {
 623                         spin_lock(&dreq->lock);
 624                         nfs_direct_truncate_request(dreq, req);
 625                         spin_unlock(&dreq->lock);
 626                         nfs_release_request(req);
 627                 } else if (!nfs_write_match_verf(verf, req)) {
 628                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 629                         /*
 630                          * Despite the reboot, the write was successful,
 631                          * so reset wb_nio.
 632                          */
 633                         req->wb_nio = 0;
 634                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 635                 } else
 636                         nfs_release_request(req);
 637                 nfs_unlock_and_release_request(req);
 638         }
 639
 640         if (nfs_commit_end(cinfo.mds))
 641                 nfs_direct_write_complete(dreq);
 642 }
 643
 644 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 645                 struct nfs_page *req)
 646 {
 647         struct nfs_direct_req *dreq = cinfo->dreq;
 648
 649         trace_nfs_direct_resched_write(dreq);
 650
 651         spin_lock(&dreq->lock);
 652         if (dreq->flags != NFS_ODIRECT_DONE)
 653                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 654         spin_unlock(&dreq->lock);
 655         nfs_mark_request_commit(req, NULL, cinfo, 0);
 656 }
 657
 658 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 659         .completion = nfs_direct_commit_complete,
 660         .resched_write = nfs_direct_resched_write,
 661 };
 662
 663 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 664 {
 665         int res;
 666         struct nfs_commit_info cinfo;
 667         LIST_HEAD(mds_list);
 668
 669         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 670         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 671         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 672         if (res < 0) /* res == -ENOMEM */
 673                 nfs_direct_write_reschedule(dreq);
 674 }
 675
 676 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
 677 {
 678         struct nfs_commit_info cinfo;
 679         struct nfs_page *req;
 680         LIST_HEAD(reqs);
 681
 682         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 683         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 684
 685         while (!list_empty(&reqs)) {
 686                 req = nfs_list_entry(reqs.next);
 687                 nfs_list_remove_request(req);
 688                 nfs_direct_truncate_request(dreq, req);
 689                 nfs_release_request(req);
 690                 nfs_unlock_and_release_request(req);
 691         }
 692 }
 693
 694 static void nfs_direct_write_schedule_work(struct work_struct *work)
 695 {
 696         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 697         int flags = dreq->flags;
 698
 699         dreq->flags = 0;
 700         switch (flags) {
 701                 case NFS_ODIRECT_DO_COMMIT:
 702                         nfs_direct_commit_schedule(dreq);
 703                         break;
 704                 case NFS_ODIRECT_RESCHED_WRITES:
 705                         nfs_direct_write_reschedule(dreq);
 706                         break;
 707                 default:
 708                         nfs_direct_write_clear_reqs(dreq);
 709                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 710                         nfs_direct_complete(dreq);
 711         }
 712 }
 713
 714 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 715 {
 716         trace_nfs_direct_write_complete(dreq);
 717         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 718 }
 719
 720 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 721 {
 722         struct nfs_direct_req *dreq = hdr->dreq;
 723         struct nfs_commit_info cinfo;
 724         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 725         int flags = NFS_ODIRECT_DONE;
 726
 727         trace_nfs_direct_write_completion(dreq);
 728
 729         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 730
 731         spin_lock(&dreq->lock);
 732         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 733                 spin_unlock(&dreq->lock);
 734                 goto out_put;
 735         }
 736
 737         nfs_direct_count_bytes(dreq, hdr);
 738         if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags) &&
 739             !test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 740                 if (!dreq->flags)
 741                         dreq->flags = NFS_ODIRECT_DO_COMMIT;
 742                 flags = dreq->flags;
 743         }
 744         spin_unlock(&dreq->lock);
 745
 746         while (!list_empty(&hdr->pages)) {
 747
 748                 req = nfs_list_entry(hdr->pages.next);
 749                 nfs_list_remove_request(req);
 750                 if (flags == NFS_ODIRECT_DO_COMMIT) {
 751                         kref_get(&req->wb_kref);
 752                         memcpy(&req->wb_verf, &hdr->verf.verifier,
 753                                sizeof(req->wb_verf));
 754                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 755                                 hdr->ds_commit_idx);
 756                 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
 757                         kref_get(&req->wb_kref);
 758                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 759                 }
 760                 nfs_unlock_and_release_request(req);
 761         }
 762
 763 out_put:
 764         if (put_dreq(dreq))
 765                 nfs_direct_write_complete(dreq);
 766         hdr->release(hdr);
 767 }
 768
 769 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 770 {
 771         struct nfs_page *req;
 772
 773         while (!list_empty(head)) {
 774                 req = nfs_list_entry(head->next);
 775                 nfs_list_remove_request(req);
 776                 nfs_unlock_and_release_request(req);
 777         }
 778 }
 779
 780 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 781 {
 782         struct nfs_direct_req *dreq = hdr->dreq;
 783         struct nfs_page *req;
 784         struct nfs_commit_info cinfo;
 785
 786         trace_nfs_direct_write_reschedule_io(dreq);
 787
 788         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 789         spin_lock(&dreq->lock);
 790         if (dreq->error == 0)
 791                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 792         set_bit(NFS_IOHDR_REDO, &hdr->flags);
 793         spin_unlock(&dreq->lock);
 794         while (!list_empty(&hdr->pages)) {
 795                 req = nfs_list_entry(hdr->pages.next);
 796                 nfs_list_remove_request(req);
 797                 nfs_unlock_request(req);
 798                 nfs_mark_request_commit(req, NULL, &cinfo, 0);
 799         }
 800 }
 801
 802 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 803         .error_cleanup = nfs_write_sync_pgio_error,
 804         .init_hdr = nfs_direct_pgio_init,
 805         .completion = nfs_direct_write_completion,
 806         .reschedule_io = nfs_direct_write_reschedule_io,
 807 };
 808
 809
 810 /*
 811  * NB: Return the value of the first error return code.  Subsequent
 812  *     errors after the first one are ignored.
 813  */
 814 /*
 815  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 816  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 817  * bail and stop sending more writes.  Write length accounting is
 818  * handled automatically by nfs_direct_write_result().  Otherwise, if
 819  * no requests have been sent, just return an error.
 820  */
 821 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 822                                                struct iov_iter *iter,
 823                                                loff_t pos, int ioflags)
 824 {
 825         struct nfs_pageio_descriptor desc;
 826         struct inode *inode = dreq->inode;
 827         struct nfs_commit_info cinfo;
 828         ssize_t result = 0;
 829         size_t requested_bytes = 0;
 830         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 831         bool defer = false;
 832
 833         trace_nfs_direct_write_schedule_iovec(dreq);
 834
 835         nfs_pageio_init_write(&desc, inode, ioflags, false,
 836                               &nfs_direct_write_completion_ops);
 837         desc.pg_dreq = dreq;
 838         get_dreq(dreq);
 839         inode_dio_begin(inode);
 840
 841         NFS_I(inode)->write_io += iov_iter_count(iter);
 842         while (iov_iter_count(iter)) {
 843                 struct page **pagevec;
 844                 size_t bytes;
 845                 size_t pgbase;
 846                 unsigned npages, i;
 847
 848                 result = iov_iter_get_pages_alloc2(iter, &pagevec,
 849                                                   wsize, &pgbase);
 850                 if (result < 0)
 851                         break;
 852
 853                 bytes = result;
 854                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 855                 for (i = 0; i < npages; i++) {
 856                         struct nfs_page *req;
 857                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 858
 859                         req = nfs_page_create_from_page(dreq->ctx, pagevec[i],
 860                                                         pgbase, pos, req_len);
 861                         if (IS_ERR(req)) {
 862                                 result = PTR_ERR(req);
 863                                 break;
 864                         }
 865
 866                         if (desc.pg_error < 0) {
 867                                 nfs_free_request(req);
 868                                 result = desc.pg_error;
 869                                 break;
 870                         }
 871
 872                         pgbase = 0;
 873                         bytes -= req_len;
 874                         requested_bytes += req_len;
 875                         pos += req_len;
 876                         dreq->bytes_left -= req_len;
 877
 878                         if (defer) {
 879                                 nfs_mark_request_commit(req, NULL, &cinfo, 0);
 880                                 continue;
 881                         }
 882
 883                         nfs_lock_request(req);
 884                         if (nfs_pageio_add_request(&desc, req))
 885                                 continue;
 886
 887                         /* Exit on hard errors */
 888                         if (desc.pg_error < 0 && desc.pg_error != -EAGAIN) {
 889                                 result = desc.pg_error;
 890                                 nfs_unlock_and_release_request(req);
 891                                 break;
 892                         }
 893
 894                         /* If the error is soft, defer remaining requests */
 895                         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 896                         spin_lock(&dreq->lock);
 897                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 898                         spin_unlock(&dreq->lock);
 899                         nfs_unlock_request(req);
 900                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 901                         desc.pg_error = 0;
 902                         defer = true;
 903                 }
 904                 nfs_direct_release_pages(pagevec, npages);
 905                 kvfree(pagevec);
 906                 if (result < 0)
 907                         break;
 908         }
 909         nfs_pageio_complete(&desc);
 910
 911         /*
 912          * If no bytes were started, return the error, and let the
 913          * generic layer handle the completion.
 914          */
 915         if (requested_bytes == 0) {
 916                 inode_dio_end(inode);
 917                 nfs_direct_req_release(dreq);
 918                 return result < 0 ? result : -EIO;
 919         }
 920
 921         if (put_dreq(dreq))
 922                 nfs_direct_write_complete(dreq);
 923         return requested_bytes;
 924 }
 925
 926 /**
 927  * nfs_file_direct_write - file direct write operation for NFS files
 928  * @iocb: target I/O control block
 929  * @iter: vector of user buffers from which to write data
 930  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 931  *
 932  * We use this function for direct writes instead of calling
 933  * generic_file_aio_write() in order to avoid taking the inode
 934  * semaphore and updating the i_size.  The NFS server will set
 935  * the new i_size and this client must read the updated size
 936  * back into its cache.  We let the server do generic write
 937  * parameter checking and report problems.
 938  *
 939  * We eliminate local atime updates, see direct read above.
 940  *
 941  * We avoid unnecessary page cache invalidations for normal cached
 942  * readers of this file.
 943  *
 944  * Note that O_APPEND is not supported for NFS direct writes, as there
 945  * is no atomic O_APPEND write facility in the NFS protocol.
 946  */
 947 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 948                               bool swap)
 949 {
 950         ssize_t result, requested;
 951         size_t count;
 952         struct file *file = iocb->ki_filp;
 953         struct address_space *mapping = file->f_mapping;
 954         struct inode *inode = mapping->host;
 955         struct nfs_direct_req *dreq;
 956         struct nfs_lock_context *l_ctx;
 957         loff_t pos, end;
 958
 959         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 960                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
 961
 962         if (swap)
 963                 /* bypass generic checks */
 964                 result =  iov_iter_count(iter);
 965         else
 966                 result = generic_write_checks(iocb, iter);
 967         if (result <= 0)
 968                 return result;
 969         count = result;
 970         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 971
 972         pos = iocb->ki_pos;
 973         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
 974
 975         task_io_account_write(count);
 976
 977         result = -ENOMEM;
 978         dreq = nfs_direct_req_alloc();
 979         if (!dreq)
 980                 goto out;
 981
 982         dreq->inode = inode;
 983         dreq->bytes_left = dreq->max_count = count;
 984         dreq->io_start = pos;
 985         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 986         l_ctx = nfs_get_lock_context(dreq->ctx);
 987         if (IS_ERR(l_ctx)) {
 988                 result = PTR_ERR(l_ctx);
 989                 nfs_direct_req_release(dreq);
 990                 goto out_release;
 991         }
 992         dreq->l_ctx = l_ctx;
 993         if (!is_sync_kiocb(iocb))
 994                 dreq->iocb = iocb;
 995         pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
 996
 997         if (swap) {
 998                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
 999                                                             FLUSH_STABLE);
1000         } else {
1001                 nfs_start_io_direct(inode);
1002
1003                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1004                                                             FLUSH_COND_STABLE);
1005
1006                 if (mapping->nrpages) {
1007                         invalidate_inode_pages2_range(mapping,
1008                                                       pos >> PAGE_SHIFT, end);
1009                 }
1010
1011                 nfs_end_io_direct(inode);
1012         }
1013
1014         if (requested > 0) {
1015                 result = nfs_direct_wait(dreq);
1016                 if (result > 0) {
1017                         requested -= result;
1018                         iocb->ki_pos = pos + result;
1019                         /* XXX: should check the generic_write_sync retval */
1020                         generic_write_sync(iocb, result);
1021                 }
1022                 iov_iter_revert(iter, requested);
1023         } else {
1024                 result = requested;
1025         }
1026         nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
1027 out_release:
1028         nfs_direct_req_release(dreq);
1029 out:
1030         return result;
1031 }
1032
1033 /**
1034  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1035  *
1036  */
1037 int __init nfs_init_directcache(void)
1038 {
1039         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1040                                                 sizeof(struct nfs_direct_req),
1041                                                 0, (SLAB_RECLAIM_ACCOUNT|
1042                                                         SLAB_MEM_SPREAD),
1043                                                 NULL);
1044         if (nfs_direct_cachep == NULL)
1045                 return -ENOMEM;
1046
1047         return 0;
1048 }
1049
1050 /**
1051  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1052  *
1053  */
1054 void nfs_destroy_directcache(void)
1055 {
1056         kmem_cache_destroy(nfs_direct_cachep);
1057 }