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 <[email protected]>
   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);
 177         return nfs_file_direct_write(iocb, iter);
 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  *
 429  * We use this function for direct reads instead of calling
 430  * generic_file_aio_read() in order to avoid gfar's check to see if
 431  * the request starts before the end of the file.  For that check
 432  * to work, we must generate a GETATTR before each direct read, and
 433  * even then there is a window between the GETATTR and the subsequent
 434  * READ where the file size could change.  Our preference is simply
 435  * to do all reads the application wants, and the server will take
 436  * care of managing the end of file boundary.
 437  *
 438  * This function also eliminates unnecessarily updating the file's
 439  * atime locally, as the NFS server sets the file's atime, and this
 440  * client must read the updated atime from the server back into its
 441  * cache.
 442  */
 443 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
 444 {
 445         struct file *file = iocb->ki_filp;
 446         struct address_space *mapping = file->f_mapping;
 447         struct inode *inode = mapping->host;
 448         struct nfs_direct_req *dreq;
 449         struct nfs_lock_context *l_ctx;
 450         ssize_t result, requested;
 451         size_t count = iov_iter_count(iter);
 452         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 453
 454         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 455                 file, count, (long long) iocb->ki_pos);
 456
 457         result = 0;
 458         if (!count)
 459                 goto out;
 460
 461         task_io_account_read(count);
 462
 463         result = -ENOMEM;
 464         dreq = nfs_direct_req_alloc();
 465         if (dreq == NULL)
 466                 goto out;
 467
 468         dreq->inode = inode;
 469         dreq->bytes_left = dreq->max_count = count;
 470         dreq->io_start = iocb->ki_pos;
 471         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 472         l_ctx = nfs_get_lock_context(dreq->ctx);
 473         if (IS_ERR(l_ctx)) {
 474                 result = PTR_ERR(l_ctx);
 475                 nfs_direct_req_release(dreq);
 476                 goto out_release;
 477         }
 478         dreq->l_ctx = l_ctx;
 479         if (!is_sync_kiocb(iocb))
 480                 dreq->iocb = iocb;
 481
 482         if (iter_is_iovec(iter))
 483                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 484
 485         nfs_start_io_direct(inode);
 486
 487         NFS_I(inode)->read_io += count;
 488         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 489
 490         nfs_end_io_direct(inode);
 491
 492         if (requested > 0) {
 493                 result = nfs_direct_wait(dreq);
 494                 if (result > 0) {
 495                         requested -= result;
 496                         iocb->ki_pos += result;
 497                 }
 498                 iov_iter_revert(iter, requested);
 499         } else {
 500                 result = requested;
 501         }
 502
 503 out_release:
 504         nfs_direct_req_release(dreq);
 505 out:
 506         return result;
 507 }
 508
 509 static void
 510 nfs_direct_join_group(struct list_head *list, struct inode *inode)
 511 {
 512         struct nfs_page *req, *next;
 513
 514         list_for_each_entry(req, list, wb_list) {
 515                 if (req->wb_head != req || req->wb_this_page == req)
 516                         continue;
 517                 for (next = req->wb_this_page;
 518                                 next != req->wb_head;
 519                                 next = next->wb_this_page) {
 520                         nfs_list_remove_request(next);
 521                         nfs_release_request(next);
 522                 }
 523                 nfs_join_page_group(req, inode);
 524         }
 525 }
 526
 527 static void
 528 nfs_direct_write_scan_commit_list(struct inode *inode,
 529                                   struct list_head *list,
 530                                   struct nfs_commit_info *cinfo)
 531 {
 532         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 533         pnfs_recover_commit_reqs(list, cinfo);
 534         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 535         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 536 }
 537
 538 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 539 {
 540         struct nfs_pageio_descriptor desc;
 541         struct nfs_page *req, *tmp;
 542         LIST_HEAD(reqs);
 543         struct nfs_commit_info cinfo;
 544         LIST_HEAD(failed);
 545
 546         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 547         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 548
 549         nfs_direct_join_group(&reqs, dreq->inode);
 550
 551         dreq->count = 0;
 552         dreq->max_count = 0;
 553         list_for_each_entry(req, &reqs, wb_list)
 554                 dreq->max_count += req->wb_bytes;
 555         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 556         get_dreq(dreq);
 557
 558         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 559                               &nfs_direct_write_completion_ops);
 560         desc.pg_dreq = dreq;
 561
 562         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 563                 /* Bump the transmission count */
 564                 req->wb_nio++;
 565                 if (!nfs_pageio_add_request(&desc, req)) {
 566                         nfs_list_move_request(req, &failed);
 567                         spin_lock(&cinfo.inode->i_lock);
 568                         dreq->flags = 0;
 569                         if (desc.pg_error < 0)
 570                                 dreq->error = desc.pg_error;
 571                         else
 572                                 dreq->error = -EIO;
 573                         spin_unlock(&cinfo.inode->i_lock);
 574                 }
 575                 nfs_release_request(req);
 576         }
 577         nfs_pageio_complete(&desc);
 578
 579         while (!list_empty(&failed)) {
 580                 req = nfs_list_entry(failed.next);
 581                 nfs_list_remove_request(req);
 582                 nfs_unlock_and_release_request(req);
 583         }
 584
 585         if (put_dreq(dreq))
 586                 nfs_direct_write_complete(dreq);
 587 }
 588
 589 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 590 {
 591         const struct nfs_writeverf *verf = data->res.verf;
 592         struct nfs_direct_req *dreq = data->dreq;
 593         struct nfs_commit_info cinfo;
 594         struct nfs_page *req;
 595         int status = data->task.tk_status;
 596
 597         if (status < 0) {
 598                 /* Errors in commit are fatal */
 599                 dreq->error = status;
 600                 dreq->max_count = 0;
 601                 dreq->count = 0;
 602                 dreq->flags = NFS_ODIRECT_DONE;
 603         } else if (dreq->flags == NFS_ODIRECT_DONE)
 604                 status = dreq->error;
 605
 606         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 607
 608         while (!list_empty(&data->pages)) {
 609                 req = nfs_list_entry(data->pages.next);
 610                 nfs_list_remove_request(req);
 611                 if (status >= 0 && !nfs_write_match_verf(verf, req)) {
 612                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 613                         /*
 614                          * Despite the reboot, the write was successful,
 615                          * so reset wb_nio.
 616                          */
 617                         req->wb_nio = 0;
 618                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 619                 } else /* Error or match */
 620                         nfs_release_request(req);
 621                 nfs_unlock_and_release_request(req);
 622         }
 623
 624         if (nfs_commit_end(cinfo.mds))
 625                 nfs_direct_write_complete(dreq);
 626 }
 627
 628 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 629                 struct nfs_page *req)
 630 {
 631         struct nfs_direct_req *dreq = cinfo->dreq;
 632
 633         spin_lock(&dreq->lock);
 634         if (dreq->flags != NFS_ODIRECT_DONE)
 635                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 636         spin_unlock(&dreq->lock);
 637         nfs_mark_request_commit(req, NULL, cinfo, 0);
 638 }
 639
 640 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 641         .completion = nfs_direct_commit_complete,
 642         .resched_write = nfs_direct_resched_write,
 643 };
 644
 645 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 646 {
 647         int res;
 648         struct nfs_commit_info cinfo;
 649         LIST_HEAD(mds_list);
 650
 651         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 652         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 653         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 654         if (res < 0) /* res == -ENOMEM */
 655                 nfs_direct_write_reschedule(dreq);
 656 }
 657
 658 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
 659 {
 660         struct nfs_commit_info cinfo;
 661         struct nfs_page *req;
 662         LIST_HEAD(reqs);
 663
 664         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 665         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 666
 667         while (!list_empty(&reqs)) {
 668                 req = nfs_list_entry(reqs.next);
 669                 nfs_list_remove_request(req);
 670                 nfs_release_request(req);
 671                 nfs_unlock_and_release_request(req);
 672         }
 673 }
 674
 675 static void nfs_direct_write_schedule_work(struct work_struct *work)
 676 {
 677         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 678         int flags = dreq->flags;
 679
 680         dreq->flags = 0;
 681         switch (flags) {
 682                 case NFS_ODIRECT_DO_COMMIT:
 683                         nfs_direct_commit_schedule(dreq);
 684                         break;
 685                 case NFS_ODIRECT_RESCHED_WRITES:
 686                         nfs_direct_write_reschedule(dreq);
 687                         break;
 688                 default:
 689                         nfs_direct_write_clear_reqs(dreq);
 690                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 691                         nfs_direct_complete(dreq);
 692         }
 693 }
 694
 695 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 696 {
 697         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 698 }
 699
 700 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 701 {
 702         struct nfs_direct_req *dreq = hdr->dreq;
 703         struct nfs_commit_info cinfo;
 704         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 705         int flags = NFS_ODIRECT_DONE;
 706
 707         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 708
 709         spin_lock(&dreq->lock);
 710         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 711                 spin_unlock(&dreq->lock);
 712                 goto out_put;
 713         }
 714
 715         nfs_direct_count_bytes(dreq, hdr);
 716         if (hdr->good_bytes != 0 && nfs_write_need_commit(hdr)) {
 717                 if (!dreq->flags)
 718                         dreq->flags = NFS_ODIRECT_DO_COMMIT;
 719                 flags = dreq->flags;
 720         }
 721         spin_unlock(&dreq->lock);
 722
 723         while (!list_empty(&hdr->pages)) {
 724
 725                 req = nfs_list_entry(hdr->pages.next);
 726                 nfs_list_remove_request(req);
 727                 if (flags == NFS_ODIRECT_DO_COMMIT) {
 728                         kref_get(&req->wb_kref);
 729                         memcpy(&req->wb_verf, &hdr->verf.verifier,
 730                                sizeof(req->wb_verf));
 731                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 732                                 hdr->ds_commit_idx);
 733                 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
 734                         kref_get(&req->wb_kref);
 735                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 736                 }
 737                 nfs_unlock_and_release_request(req);
 738         }
 739
 740 out_put:
 741         if (put_dreq(dreq))
 742                 nfs_direct_write_complete(dreq);
 743         hdr->release(hdr);
 744 }
 745
 746 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 747 {
 748         struct nfs_page *req;
 749
 750         while (!list_empty(head)) {
 751                 req = nfs_list_entry(head->next);
 752                 nfs_list_remove_request(req);
 753                 nfs_unlock_and_release_request(req);
 754         }
 755 }
 756
 757 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 758 {
 759         struct nfs_direct_req *dreq = hdr->dreq;
 760
 761         spin_lock(&dreq->lock);
 762         if (dreq->error == 0) {
 763                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 764                 /* fake unstable write to let common nfs resend pages */
 765                 hdr->verf.committed = NFS_UNSTABLE;
 766                 hdr->good_bytes = hdr->args.offset + hdr->args.count -
 767                         hdr->io_start;
 768         }
 769         spin_unlock(&dreq->lock);
 770 }
 771
 772 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 773         .error_cleanup = nfs_write_sync_pgio_error,
 774         .init_hdr = nfs_direct_pgio_init,
 775         .completion = nfs_direct_write_completion,
 776         .reschedule_io = nfs_direct_write_reschedule_io,
 777 };
 778
 779
 780 /*
 781  * NB: Return the value of the first error return code.  Subsequent
 782  *     errors after the first one are ignored.
 783  */
 784 /*
 785  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 786  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 787  * bail and stop sending more writes.  Write length accounting is
 788  * handled automatically by nfs_direct_write_result().  Otherwise, if
 789  * no requests have been sent, just return an error.
 790  */
 791 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 792                                                struct iov_iter *iter,
 793                                                loff_t pos)
 794 {
 795         struct nfs_pageio_descriptor desc;
 796         struct inode *inode = dreq->inode;
 797         ssize_t result = 0;
 798         size_t requested_bytes = 0;
 799         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 800
 801         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 802                               &nfs_direct_write_completion_ops);
 803         desc.pg_dreq = dreq;
 804         get_dreq(dreq);
 805         inode_dio_begin(inode);
 806
 807         NFS_I(inode)->write_io += iov_iter_count(iter);
 808         while (iov_iter_count(iter)) {
 809                 struct page **pagevec;
 810                 size_t bytes;
 811                 size_t pgbase;
 812                 unsigned npages, i;
 813
 814                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 815                                                   wsize, &pgbase);
 816                 if (result < 0)
 817                         break;
 818
 819                 bytes = result;
 820                 iov_iter_advance(iter, bytes);
 821                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 822                 for (i = 0; i < npages; i++) {
 823                         struct nfs_page *req;
 824                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 825
 826                         req = nfs_create_request(dreq->ctx, pagevec[i],
 827                                                  pgbase, req_len);
 828                         if (IS_ERR(req)) {
 829                                 result = PTR_ERR(req);
 830                                 break;
 831                         }
 832
 833                         if (desc.pg_error < 0) {
 834                                 nfs_free_request(req);
 835                                 result = desc.pg_error;
 836                                 break;
 837                         }
 838
 839                         nfs_lock_request(req);
 840                         req->wb_index = pos >> PAGE_SHIFT;
 841                         req->wb_offset = pos & ~PAGE_MASK;
 842                         if (!nfs_pageio_add_request(&desc, req)) {
 843                                 result = desc.pg_error;
 844                                 nfs_unlock_and_release_request(req);
 845                                 break;
 846                         }
 847                         pgbase = 0;
 848                         bytes -= req_len;
 849                         requested_bytes += req_len;
 850                         pos += req_len;
 851                         dreq->bytes_left -= req_len;
 852                 }
 853                 nfs_direct_release_pages(pagevec, npages);
 854                 kvfree(pagevec);
 855                 if (result < 0)
 856                         break;
 857         }
 858         nfs_pageio_complete(&desc);
 859
 860         /*
 861          * If no bytes were started, return the error, and let the
 862          * generic layer handle the completion.
 863          */
 864         if (requested_bytes == 0) {
 865                 inode_dio_end(inode);
 866                 nfs_direct_req_release(dreq);
 867                 return result < 0 ? result : -EIO;
 868         }
 869
 870         if (put_dreq(dreq))
 871                 nfs_direct_write_complete(dreq);
 872         return requested_bytes;
 873 }
 874
 875 /**
 876  * nfs_file_direct_write - file direct write operation for NFS files
 877  * @iocb: target I/O control block
 878  * @iter: vector of user buffers from which to write data
 879  *
 880  * We use this function for direct writes instead of calling
 881  * generic_file_aio_write() in order to avoid taking the inode
 882  * semaphore and updating the i_size.  The NFS server will set
 883  * the new i_size and this client must read the updated size
 884  * back into its cache.  We let the server do generic write
 885  * parameter checking and report problems.
 886  *
 887  * We eliminate local atime updates, see direct read above.
 888  *
 889  * We avoid unnecessary page cache invalidations for normal cached
 890  * readers of this file.
 891  *
 892  * Note that O_APPEND is not supported for NFS direct writes, as there
 893  * is no atomic O_APPEND write facility in the NFS protocol.
 894  */
 895 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 896 {
 897         ssize_t result, requested;
 898         size_t count;
 899         struct file *file = iocb->ki_filp;
 900         struct address_space *mapping = file->f_mapping;
 901         struct inode *inode = mapping->host;
 902         struct nfs_direct_req *dreq;
 903         struct nfs_lock_context *l_ctx;
 904         loff_t pos, end;
 905
 906         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 907                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
 908
 909         result = generic_write_checks(iocb, iter);
 910         if (result <= 0)
 911                 return result;
 912         count = result;
 913         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 914
 915         pos = iocb->ki_pos;
 916         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
 917
 918         task_io_account_write(count);
 919
 920         result = -ENOMEM;
 921         dreq = nfs_direct_req_alloc();
 922         if (!dreq)
 923                 goto out;
 924
 925         dreq->inode = inode;
 926         dreq->bytes_left = dreq->max_count = count;
 927         dreq->io_start = pos;
 928         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 929         l_ctx = nfs_get_lock_context(dreq->ctx);
 930         if (IS_ERR(l_ctx)) {
 931                 result = PTR_ERR(l_ctx);
 932                 nfs_direct_req_release(dreq);
 933                 goto out_release;
 934         }
 935         dreq->l_ctx = l_ctx;
 936         if (!is_sync_kiocb(iocb))
 937                 dreq->iocb = iocb;
 938         pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
 939
 940         nfs_start_io_direct(inode);
 941
 942         requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
 943
 944         if (mapping->nrpages) {
 945                 invalidate_inode_pages2_range(mapping,
 946                                               pos >> PAGE_SHIFT, end);
 947         }
 948
 949         nfs_end_io_direct(inode);
 950
 951         if (requested > 0) {
 952                 result = nfs_direct_wait(dreq);
 953                 if (result > 0) {
 954                         requested -= result;
 955                         iocb->ki_pos = pos + result;
 956                         /* XXX: should check the generic_write_sync retval */
 957                         generic_write_sync(iocb, result);
 958                 }
 959                 iov_iter_revert(iter, requested);
 960         } else {
 961                 result = requested;
 962         }
 963         nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
 964 out_release:
 965         nfs_direct_req_release(dreq);
 966 out:
 967         return result;
 968 }
 969
 970 /**
 971  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 972  *
 973  */
 974 int __init nfs_init_directcache(void)
 975 {
 976         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 977                                                 sizeof(struct nfs_direct_req),
 978                                                 0, (SLAB_RECLAIM_ACCOUNT|
 979                                                         SLAB_MEM_SPREAD),
 980                                                 NULL);
 981         if (nfs_direct_cachep == NULL)
 982                 return -ENOMEM;
 983
 984         return 0;
 985 }
 986
 987 /**
 988  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 989  *
 990  */
 991 void nfs_destroy_directcache(void)
 992 {
 993         kmem_cache_destroy(nfs_direct_cachep);
 994 }