fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5  *
   6  * High-performance uncached I/O for the Linux NFS client
   7  *
   8  * There are important applications whose performance or correctness
   9  * depends on uncached access to file data.  Database clusters
  10  * (multiple copies of the same instance running on separate hosts)
  11  * implement their own cache coherency protocol that subsumes file
  12  * system cache protocols.  Applications that process datasets
  13  * considerably larger than the client's memory do not always benefit
  14  * from a local cache.  A streaming video server, for instance, has no
  15  * need to cache the contents of a file.
  16  *
  17  * When an application requests uncached I/O, all read and write requests
  18  * are made directly to the server; data stored or fetched via these
  19  * requests is not cached in the Linux page cache.  The client does not
  20  * correct unaligned requests from applications.  All requested bytes are
  21  * held on permanent storage before a direct write system call returns to
  22  * an application.
  23  *
  24  * Solaris implements an uncached I/O facility called directio() that
  25  * is used for backups and sequential I/O to very large files.  Solaris
  26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27  * an undocumented mount option.
  28  *
  29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30  * help from Andrew Morton.
  31  *
  32  * 18 Dec 2001  Initial implementation for 2.4  --cel
  33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  34  * 08 Jun 2003  Port to 2.5 APIs  --cel
  35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  36  * 15 Sep 2004  Parallel async reads  --cel
  37  * 04 May 2005  support O_DIRECT with aio  --cel
  38  *
  39  */
  40
  41 #include <linux/errno.h>
  42 #include <linux/sched.h>
  43 #include <linux/kernel.h>
  44 #include <linux/file.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/kref.h>
  47 #include <linux/slab.h>
  48 #include <linux/task_io_accounting_ops.h>
  49 #include <linux/module.h>
  50
  51 #include <linux/nfs_fs.h>
  52 #include <linux/nfs_page.h>
  53 #include <linux/sunrpc/clnt.h>
  54
  55 #include <linux/uaccess.h>
  56 #include <linux/atomic.h>
  57
  58 #include "internal.h"
  59 #include "iostat.h"
  60 #include "pnfs.h"
  61
  62 #define NFSDBG_FACILITY         NFSDBG_VFS
  63
  64 static struct kmem_cache *nfs_direct_cachep;
  65
  66 /*
  67  * This represents a set of asynchronous requests that we're waiting on
  68  */
  69 struct nfs_direct_mirror {
  70         ssize_t count;
  71 };
  72
  73 struct nfs_direct_req {
  74         struct kref             kref;           /* release manager */
  75
  76         /* I/O parameters */
  77         struct nfs_open_context *ctx;           /* file open context info */
  78         struct nfs_lock_context *l_ctx;         /* Lock context info */
  79         struct kiocb *          iocb;           /* controlling i/o request */
  80         struct inode *          inode;          /* target file of i/o */
  81
  82         /* completion state */
  83         atomic_t                io_count;       /* i/os we're waiting for */
  84         spinlock_t              lock;           /* protect completion state */
  85
  86         struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
  87         int                     mirror_count;
  88
  89         loff_t                  io_start;       /* Start offset for I/O */
  90         ssize_t                 count,          /* bytes actually processed */
  91                                 max_count,      /* max expected count */
  92                                 bytes_left,     /* bytes left to be sent */
  93                                 error;          /* any reported error */
  94         struct completion       completion;     /* wait for i/o completion */
  95
  96         /* commit state */
  97         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  98         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  99         struct work_struct      work;
 100         int                     flags;
 101         /* for write */
 102 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
 103 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
 104         /* for read */
 105 #define NFS_ODIRECT_SHOULD_DIRTY        (3)     /* dirty user-space page after read */
 106         struct nfs_writeverf    verf;           /* unstable write verifier */
 107 };
 108
 109 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 110 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 111 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
 112 static void nfs_direct_write_schedule_work(struct work_struct *work);
 113
 114 static inline void get_dreq(struct nfs_direct_req *dreq)
 115 {
 116         atomic_inc(&dreq->io_count);
 117 }
 118
 119 static inline int put_dreq(struct nfs_direct_req *dreq)
 120 {
 121         return atomic_dec_and_test(&dreq->io_count);
 122 }
 123
 124 static void
 125 nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
 126                             const struct nfs_pgio_header *hdr,
 127                             ssize_t dreq_len)
 128 {
 129         struct nfs_direct_mirror *mirror = &dreq->mirrors[hdr->pgio_mirror_idx];
 130
 131         if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
 132               test_bit(NFS_IOHDR_EOF, &hdr->flags)))
 133                 return;
 134         if (dreq->max_count >= dreq_len) {
 135                 dreq->max_count = dreq_len;
 136                 if (dreq->count > dreq_len)
 137                         dreq->count = dreq_len;
 138
 139                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
 140                         dreq->error = hdr->error;
 141                 else /* Clear outstanding error if this is EOF */
 142                         dreq->error = 0;
 143         }
 144         if (mirror->count > dreq_len)
 145                 mirror->count = dreq_len;
 146 }
 147
 148 static void
 149 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
 150                        const struct nfs_pgio_header *hdr)
 151 {
 152         struct nfs_direct_mirror *mirror = &dreq->mirrors[hdr->pgio_mirror_idx];
 153         loff_t hdr_end = hdr->io_start + hdr->good_bytes;
 154         ssize_t dreq_len = 0;
 155
 156         if (hdr_end > dreq->io_start)
 157                 dreq_len = hdr_end - dreq->io_start;
 158
 159         nfs_direct_handle_truncated(dreq, hdr, dreq_len);
 160
 161         if (dreq_len > dreq->max_count)
 162                 dreq_len = dreq->max_count;
 163
 164         if (mirror->count < dreq_len)
 165                 mirror->count = dreq_len;
 166         if (dreq->count < dreq_len)
 167                 dreq->count = dreq_len;
 168 }
 169
 170 /*
 171  * nfs_direct_select_verf - select the right verifier
 172  * @dreq - direct request possibly spanning multiple servers
 173  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 174  * @commit_idx - commit bucket index for the DS
 175  *
 176  * returns the correct verifier to use given the role of the server
 177  */
 178 static struct nfs_writeverf *
 179 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 180                        struct nfs_client *ds_clp,
 181                        int commit_idx)
 182 {
 183         struct nfs_writeverf *verfp = &dreq->verf;
 184
 185 #ifdef CONFIG_NFS_V4_1
 186         /*
 187          * pNFS is in use, use the DS verf except commit_through_mds is set
 188          * for layout segment where nbuckets is zero.
 189          */
 190         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 191                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 192                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 193                 else
 194                         WARN_ON_ONCE(1);
 195         }
 196 #endif
 197         return verfp;
 198 }
 199
 200
 201 /*
 202  * nfs_direct_set_hdr_verf - set the write/commit verifier
 203  * @dreq - direct request possibly spanning multiple servers
 204  * @hdr - pageio header to validate against previously seen verfs
 205  *
 206  * Set the server's (MDS or DS) "seen" verifier
 207  */
 208 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 209                                     struct nfs_pgio_header *hdr)
 210 {
 211         struct nfs_writeverf *verfp;
 212
 213         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 214         WARN_ON_ONCE(verfp->committed >= 0);
 215         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 216         WARN_ON_ONCE(verfp->committed < 0);
 217 }
 218
 219 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
 220                 const struct nfs_writeverf *v2)
 221 {
 222         return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
 223 }
 224
 225 /*
 226  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 227  * @dreq - direct request possibly spanning multiple servers
 228  * @hdr - pageio header to validate against previously seen verf
 229  *
 230  * set the server's "seen" verf if not initialized.
 231  * returns result of comparison between @hdr->verf and the "seen"
 232  * verf of the server used by @hdr (DS or MDS)
 233  */
 234 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 235                                           struct nfs_pgio_header *hdr)
 236 {
 237         struct nfs_writeverf *verfp;
 238
 239         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 240         if (verfp->committed < 0) {
 241                 nfs_direct_set_hdr_verf(dreq, hdr);
 242                 return 0;
 243         }
 244         return nfs_direct_cmp_verf(verfp, &hdr->verf);
 245 }
 246
 247 /*
 248  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 249  * @dreq - direct request possibly spanning multiple servers
 250  * @data - commit data to validate against previously seen verf
 251  *
 252  * returns result of comparison between @data->verf and the verf of
 253  * the server used by @data (DS or MDS)
 254  */
 255 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 256                                            struct nfs_commit_data *data)
 257 {
 258         struct nfs_writeverf *verfp;
 259
 260         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 261                                          data->ds_commit_index);
 262
 263         /* verifier not set so always fail */
 264         if (verfp->committed < 0 || data->res.verf->committed <= NFS_UNSTABLE)
 265                 return 1;
 266
 267         return nfs_direct_cmp_verf(verfp, data->res.verf);
 268 }
 269
 270 /**
 271  * nfs_direct_IO - NFS address space operation for direct I/O
 272  * @iocb: target I/O control block
 273  * @iter: I/O buffer
 274  *
 275  * The presence of this routine in the address space ops vector means
 276  * the NFS client supports direct I/O. However, for most direct IO, we
 277  * shunt off direct read and write requests before the VFS gets them,
 278  * so this method is only ever called for swap.
 279  */
 280 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 281 {
 282         struct inode *inode = iocb->ki_filp->f_mapping->host;
 283
 284         /* we only support swap file calling nfs_direct_IO */
 285         if (!IS_SWAPFILE(inode))
 286                 return 0;
 287
 288         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 289
 290         if (iov_iter_rw(iter) == READ)
 291                 return nfs_file_direct_read(iocb, iter, true);
 292         return nfs_file_direct_write(iocb, iter, true);
 293 }
 294
 295 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 296 {
 297         unsigned int i;
 298         for (i = 0; i < npages; i++)
 299                 put_page(pages[i]);
 300 }
 301
 302 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 303                               struct nfs_direct_req *dreq)
 304 {
 305         cinfo->inode = dreq->inode;
 306         cinfo->mds = &dreq->mds_cinfo;
 307         cinfo->ds = &dreq->ds_cinfo;
 308         cinfo->dreq = dreq;
 309         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 310 }
 311
 312 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 313                                              struct nfs_pageio_descriptor *pgio,
 314                                              struct nfs_page *req)
 315 {
 316         int mirror_count = 1;
 317
 318         if (pgio->pg_ops->pg_get_mirror_count)
 319                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 320
 321         dreq->mirror_count = mirror_count;
 322 }
 323
 324 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 325 {
 326         struct nfs_direct_req *dreq;
 327
 328         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 329         if (!dreq)
 330                 return NULL;
 331
 332         kref_init(&dreq->kref);
 333         kref_get(&dreq->kref);
 334         init_completion(&dreq->completion);
 335         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 336         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 337         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 338         dreq->mirror_count = 1;
 339         spin_lock_init(&dreq->lock);
 340
 341         return dreq;
 342 }
 343
 344 static void nfs_direct_req_free(struct kref *kref)
 345 {
 346         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 347
 348         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 349         if (dreq->l_ctx != NULL)
 350                 nfs_put_lock_context(dreq->l_ctx);
 351         if (dreq->ctx != NULL)
 352                 put_nfs_open_context(dreq->ctx);
 353         kmem_cache_free(nfs_direct_cachep, dreq);
 354 }
 355
 356 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 357 {
 358         kref_put(&dreq->kref, nfs_direct_req_free);
 359 }
 360
 361 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 362 {
 363         return dreq->bytes_left;
 364 }
 365 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 366
 367 /*
 368  * Collects and returns the final error value/byte-count.
 369  */
 370 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 371 {
 372         ssize_t result = -EIOCBQUEUED;
 373
 374         /* Async requests don't wait here */
 375         if (dreq->iocb)
 376                 goto out;
 377
 378         result = wait_for_completion_killable(&dreq->completion);
 379
 380         if (!result) {
 381                 result = dreq->count;
 382                 WARN_ON_ONCE(dreq->count < 0);
 383         }
 384         if (!result)
 385                 result = dreq->error;
 386
 387 out:
 388         return (ssize_t) result;
 389 }
 390
 391 /*
 392  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 393  * the iocb is still valid here if this is a synchronous request.
 394  */
 395 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 396 {
 397         struct inode *inode = dreq->inode;
 398
 399         inode_dio_end(inode);
 400
 401         if (dreq->iocb) {
 402                 long res = (long) dreq->error;
 403                 if (dreq->count != 0) {
 404                         res = (long) dreq->count;
 405                         WARN_ON_ONCE(dreq->count < 0);
 406                 }
 407                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 408         }
 409
 410         complete(&dreq->completion);
 411
 412         nfs_direct_req_release(dreq);
 413 }
 414
 415 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 416 {
 417         unsigned long bytes = 0;
 418         struct nfs_direct_req *dreq = hdr->dreq;
 419
 420         spin_lock(&dreq->lock);
 421         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 422                 spin_unlock(&dreq->lock);
 423                 goto out_put;
 424         }
 425
 426         nfs_direct_count_bytes(dreq, hdr);
 427         spin_unlock(&dreq->lock);
 428
 429         while (!list_empty(&hdr->pages)) {
 430                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 431                 struct page *page = req->wb_page;
 432
 433                 if (!PageCompound(page) && bytes < hdr->good_bytes &&
 434                     (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 435                         set_page_dirty(page);
 436                 bytes += req->wb_bytes;
 437                 nfs_list_remove_request(req);
 438                 nfs_release_request(req);
 439         }
 440 out_put:
 441         if (put_dreq(dreq))
 442                 nfs_direct_complete(dreq);
 443         hdr->release(hdr);
 444 }
 445
 446 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 447 {
 448         struct nfs_page *req;
 449
 450         while (!list_empty(head)) {
 451                 req = nfs_list_entry(head->next);
 452                 nfs_list_remove_request(req);
 453                 nfs_release_request(req);
 454         }
 455 }
 456
 457 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 458 {
 459         get_dreq(hdr->dreq);
 460 }
 461
 462 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 463         .error_cleanup = nfs_read_sync_pgio_error,
 464         .init_hdr = nfs_direct_pgio_init,
 465         .completion = nfs_direct_read_completion,
 466 };
 467
 468 /*
 469  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 470  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 471  * bail and stop sending more reads.  Read length accounting is
 472  * handled automatically by nfs_direct_read_result().  Otherwise, if
 473  * no requests have been sent, just return an error.
 474  */
 475
 476 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 477                                               struct iov_iter *iter,
 478                                               loff_t pos)
 479 {
 480         struct nfs_pageio_descriptor desc;
 481         struct inode *inode = dreq->inode;
 482         ssize_t result = -EINVAL;
 483         size_t requested_bytes = 0;
 484         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 485
 486         nfs_pageio_init_read(&desc, dreq->inode, false,
 487                              &nfs_direct_read_completion_ops);
 488         get_dreq(dreq);
 489         desc.pg_dreq = dreq;
 490         inode_dio_begin(inode);
 491
 492         while (iov_iter_count(iter)) {
 493                 struct page **pagevec;
 494                 size_t bytes;
 495                 size_t pgbase;
 496                 unsigned npages, i;
 497
 498                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 499                                                   rsize, &pgbase);
 500                 if (result < 0)
 501                         break;
 502
 503                 bytes = result;
 504                 iov_iter_advance(iter, bytes);
 505                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 506                 for (i = 0; i < npages; i++) {
 507                         struct nfs_page *req;
 508                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 509                         /* XXX do we need to do the eof zeroing found in async_filler? */
 510                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 511                                                  pgbase, req_len);
 512                         if (IS_ERR(req)) {
 513                                 result = PTR_ERR(req);
 514                                 break;
 515                         }
 516                         req->wb_index = pos >> PAGE_SHIFT;
 517                         req->wb_offset = pos & ~PAGE_MASK;
 518                         if (!nfs_pageio_add_request(&desc, req)) {
 519                                 result = desc.pg_error;
 520                                 nfs_release_request(req);
 521                                 break;
 522                         }
 523                         pgbase = 0;
 524                         bytes -= req_len;
 525                         requested_bytes += req_len;
 526                         pos += req_len;
 527                         dreq->bytes_left -= req_len;
 528                 }
 529                 nfs_direct_release_pages(pagevec, npages);
 530                 kvfree(pagevec);
 531                 if (result < 0)
 532                         break;
 533         }
 534
 535         nfs_pageio_complete(&desc);
 536
 537         /*
 538          * If no bytes were started, return the error, and let the
 539          * generic layer handle the completion.
 540          */
 541         if (requested_bytes == 0) {
 542                 inode_dio_end(inode);
 543                 nfs_direct_req_release(dreq);
 544                 return result < 0 ? result : -EIO;
 545         }
 546
 547         if (put_dreq(dreq))
 548                 nfs_direct_complete(dreq);
 549         return requested_bytes;
 550 }
 551
 552 /**
 553  * nfs_file_direct_read - file direct read operation for NFS files
 554  * @iocb: target I/O control block
 555  * @iter: vector of user buffers into which to read data
 556  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 557  *
 558  * We use this function for direct reads instead of calling
 559  * generic_file_aio_read() in order to avoid gfar's check to see if
 560  * the request starts before the end of the file.  For that check
 561  * to work, we must generate a GETATTR before each direct read, and
 562  * even then there is a window between the GETATTR and the subsequent
 563  * READ where the file size could change.  Our preference is simply
 564  * to do all reads the application wants, and the server will take
 565  * care of managing the end of file boundary.
 566  *
 567  * This function also eliminates unnecessarily updating the file's
 568  * atime locally, as the NFS server sets the file's atime, and this
 569  * client must read the updated atime from the server back into its
 570  * cache.
 571  */
 572 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 573                              bool swap)
 574 {
 575         struct file *file = iocb->ki_filp;
 576         struct address_space *mapping = file->f_mapping;
 577         struct inode *inode = mapping->host;
 578         struct nfs_direct_req *dreq;
 579         struct nfs_lock_context *l_ctx;
 580         ssize_t result = -EINVAL, requested;
 581         size_t count = iov_iter_count(iter);
 582         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 583
 584         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 585                 file, count, (long long) iocb->ki_pos);
 586
 587         result = 0;
 588         if (!count)
 589                 goto out;
 590
 591         task_io_account_read(count);
 592
 593         result = -ENOMEM;
 594         dreq = nfs_direct_req_alloc();
 595         if (dreq == NULL)
 596                 goto out;
 597
 598         dreq->inode = inode;
 599         dreq->bytes_left = dreq->max_count = count;
 600         dreq->io_start = iocb->ki_pos;
 601         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 602         l_ctx = nfs_get_lock_context(dreq->ctx);
 603         if (IS_ERR(l_ctx)) {
 604                 result = PTR_ERR(l_ctx);
 605                 nfs_direct_req_release(dreq);
 606                 goto out_release;
 607         }
 608         dreq->l_ctx = l_ctx;
 609         if (!is_sync_kiocb(iocb))
 610                 dreq->iocb = iocb;
 611
 612         if (iter_is_iovec(iter))
 613                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 614
 615         if (!swap)
 616                 nfs_start_io_direct(inode);
 617
 618         NFS_I(inode)->read_io += count;
 619         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 620
 621         if (!swap)
 622                 nfs_end_io_direct(inode);
 623
 624         if (requested > 0) {
 625                 result = nfs_direct_wait(dreq);
 626                 if (result > 0) {
 627                         requested -= result;
 628                         iocb->ki_pos += result;
 629                 }
 630                 iov_iter_revert(iter, requested);
 631         } else {
 632                 result = requested;
 633         }
 634
 635 out_release:
 636         nfs_direct_req_release(dreq);
 637 out:
 638         return result;
 639 }
 640
 641 static void
 642 nfs_direct_write_scan_commit_list(struct inode *inode,
 643                                   struct list_head *list,
 644                                   struct nfs_commit_info *cinfo)
 645 {
 646         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 647 #ifdef CONFIG_NFS_V4_1
 648         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 649                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 650 #endif
 651         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 652         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 653 }
 654
 655 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 656 {
 657         struct nfs_pageio_descriptor desc;
 658         struct nfs_page *req, *tmp;
 659         LIST_HEAD(reqs);
 660         struct nfs_commit_info cinfo;
 661         LIST_HEAD(failed);
 662         int i;
 663
 664         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 665         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 666
 667         dreq->count = 0;
 668         dreq->max_count = 0;
 669         list_for_each_entry(req, &reqs, wb_list)
 670                 dreq->max_count += req->wb_bytes;
 671         dreq->verf.committed = NFS_INVALID_STABLE_HOW;
 672         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 673         for (i = 0; i < dreq->mirror_count; i++)
 674                 dreq->mirrors[i].count = 0;
 675         get_dreq(dreq);
 676
 677         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 678                               &nfs_direct_write_completion_ops);
 679         desc.pg_dreq = dreq;
 680
 681         req = nfs_list_entry(reqs.next);
 682         nfs_direct_setup_mirroring(dreq, &desc, req);
 683         if (desc.pg_error < 0) {
 684                 list_splice_init(&reqs, &failed);
 685                 goto out_failed;
 686         }
 687
 688         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 689                 if (!nfs_pageio_add_request(&desc, req)) {
 690                         nfs_list_move_request(req, &failed);
 691                         spin_lock(&cinfo.inode->i_lock);
 692                         dreq->flags = 0;
 693                         if (desc.pg_error < 0)
 694                                 dreq->error = desc.pg_error;
 695                         else
 696                                 dreq->error = -EIO;
 697                         spin_unlock(&cinfo.inode->i_lock);
 698                 }
 699                 nfs_release_request(req);
 700         }
 701         nfs_pageio_complete(&desc);
 702
 703 out_failed:
 704         while (!list_empty(&failed)) {
 705                 req = nfs_list_entry(failed.next);
 706                 nfs_list_remove_request(req);
 707                 nfs_unlock_and_release_request(req);
 708         }
 709
 710         if (put_dreq(dreq))
 711                 nfs_direct_write_complete(dreq);
 712 }
 713
 714 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 715 {
 716         struct nfs_direct_req *dreq = data->dreq;
 717         struct nfs_commit_info cinfo;
 718         struct nfs_page *req;
 719         int status = data->task.tk_status;
 720
 721         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 722         if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
 723                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 724
 725         while (!list_empty(&data->pages)) {
 726                 req = nfs_list_entry(data->pages.next);
 727                 nfs_list_remove_request(req);
 728                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 729                         /* Note the rewrite will go through mds */
 730                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 731                 } else
 732                         nfs_release_request(req);
 733                 nfs_unlock_and_release_request(req);
 734         }
 735
 736         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 737                 nfs_direct_write_complete(dreq);
 738 }
 739
 740 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 741                 struct nfs_page *req)
 742 {
 743         struct nfs_direct_req *dreq = cinfo->dreq;
 744
 745         spin_lock(&dreq->lock);
 746         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 747         spin_unlock(&dreq->lock);
 748         nfs_mark_request_commit(req, NULL, cinfo, 0);
 749 }
 750
 751 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 752         .completion = nfs_direct_commit_complete,
 753         .resched_write = nfs_direct_resched_write,
 754 };
 755
 756 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 757 {
 758         int res;
 759         struct nfs_commit_info cinfo;
 760         LIST_HEAD(mds_list);
 761
 762         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 763         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 764         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 765         if (res < 0) /* res == -ENOMEM */
 766                 nfs_direct_write_reschedule(dreq);
 767 }
 768
 769 static void nfs_direct_write_schedule_work(struct work_struct *work)
 770 {
 771         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 772         int flags = dreq->flags;
 773
 774         dreq->flags = 0;
 775         switch (flags) {
 776                 case NFS_ODIRECT_DO_COMMIT:
 777                         nfs_direct_commit_schedule(dreq);
 778                         break;
 779                 case NFS_ODIRECT_RESCHED_WRITES:
 780                         nfs_direct_write_reschedule(dreq);
 781                         break;
 782                 default:
 783                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 784                         nfs_direct_complete(dreq);
 785         }
 786 }
 787
 788 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 789 {
 790         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 791 }
 792
 793 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 794 {
 795         struct nfs_direct_req *dreq = hdr->dreq;
 796         struct nfs_commit_info cinfo;
 797         bool request_commit = false;
 798         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 799
 800         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 801
 802         spin_lock(&dreq->lock);
 803         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 804                 spin_unlock(&dreq->lock);
 805                 goto out_put;
 806         }
 807
 808         nfs_direct_count_bytes(dreq, hdr);
 809         if (hdr->good_bytes != 0) {
 810                 if (nfs_write_need_commit(hdr)) {
 811                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 812                                 request_commit = true;
 813                         else if (dreq->flags == 0) {
 814                                 nfs_direct_set_hdr_verf(dreq, hdr);
 815                                 request_commit = true;
 816                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 817                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 818                                 request_commit = true;
 819                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 820                                         dreq->flags =
 821                                                 NFS_ODIRECT_RESCHED_WRITES;
 822                         }
 823                 }
 824         }
 825         spin_unlock(&dreq->lock);
 826
 827         while (!list_empty(&hdr->pages)) {
 828
 829                 req = nfs_list_entry(hdr->pages.next);
 830                 nfs_list_remove_request(req);
 831                 if (request_commit) {
 832                         kref_get(&req->wb_kref);
 833                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 834                                 hdr->ds_commit_idx);
 835                 }
 836                 nfs_unlock_and_release_request(req);
 837         }
 838
 839 out_put:
 840         if (put_dreq(dreq))
 841                 nfs_direct_write_complete(dreq);
 842         hdr->release(hdr);
 843 }
 844
 845 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 846 {
 847         struct nfs_page *req;
 848
 849         while (!list_empty(head)) {
 850                 req = nfs_list_entry(head->next);
 851                 nfs_list_remove_request(req);
 852                 nfs_unlock_and_release_request(req);
 853         }
 854 }
 855
 856 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 857 {
 858         struct nfs_direct_req *dreq = hdr->dreq;
 859
 860         spin_lock(&dreq->lock);
 861         if (dreq->error == 0) {
 862                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 863                 /* fake unstable write to let common nfs resend pages */
 864                 hdr->verf.committed = NFS_UNSTABLE;
 865                 hdr->good_bytes = hdr->args.count;
 866         }
 867         spin_unlock(&dreq->lock);
 868 }
 869
 870 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 871         .error_cleanup = nfs_write_sync_pgio_error,
 872         .init_hdr = nfs_direct_pgio_init,
 873         .completion = nfs_direct_write_completion,
 874         .reschedule_io = nfs_direct_write_reschedule_io,
 875 };
 876
 877
 878 /*
 879  * NB: Return the value of the first error return code.  Subsequent
 880  *     errors after the first one are ignored.
 881  */
 882 /*
 883  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 884  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 885  * bail and stop sending more writes.  Write length accounting is
 886  * handled automatically by nfs_direct_write_result().  Otherwise, if
 887  * no requests have been sent, just return an error.
 888  */
 889 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 890                                                struct iov_iter *iter,
 891                                                loff_t pos, int ioflags)
 892 {
 893         struct nfs_pageio_descriptor desc;
 894         struct inode *inode = dreq->inode;
 895         ssize_t result = 0;
 896         size_t requested_bytes = 0;
 897         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 898
 899         nfs_pageio_init_write(&desc, inode, ioflags, false,
 900                               &nfs_direct_write_completion_ops);
 901         desc.pg_dreq = dreq;
 902         get_dreq(dreq);
 903         inode_dio_begin(inode);
 904
 905         NFS_I(inode)->write_io += iov_iter_count(iter);
 906         while (iov_iter_count(iter)) {
 907                 struct page **pagevec;
 908                 size_t bytes;
 909                 size_t pgbase;
 910                 unsigned npages, i;
 911
 912                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 913                                                   wsize, &pgbase);
 914                 if (result < 0)
 915                         break;
 916
 917                 bytes = result;
 918                 iov_iter_advance(iter, bytes);
 919                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 920                 for (i = 0; i < npages; i++) {
 921                         struct nfs_page *req;
 922                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 923
 924                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 925                                                  pgbase, req_len);
 926                         if (IS_ERR(req)) {
 927                                 result = PTR_ERR(req);
 928                                 break;
 929                         }
 930
 931                         nfs_direct_setup_mirroring(dreq, &desc, req);
 932                         if (desc.pg_error < 0) {
 933                                 nfs_free_request(req);
 934                                 result = desc.pg_error;
 935                                 break;
 936                         }
 937
 938                         nfs_lock_request(req);
 939                         req->wb_index = pos >> PAGE_SHIFT;
 940                         req->wb_offset = pos & ~PAGE_MASK;
 941                         if (!nfs_pageio_add_request(&desc, req)) {
 942                                 result = desc.pg_error;
 943                                 nfs_unlock_and_release_request(req);
 944                                 break;
 945                         }
 946                         pgbase = 0;
 947                         bytes -= req_len;
 948                         requested_bytes += req_len;
 949                         pos += req_len;
 950                         dreq->bytes_left -= req_len;
 951                 }
 952                 nfs_direct_release_pages(pagevec, npages);
 953                 kvfree(pagevec);
 954                 if (result < 0)
 955                         break;
 956         }
 957         nfs_pageio_complete(&desc);
 958
 959         /*
 960          * If no bytes were started, return the error, and let the
 961          * generic layer handle the completion.
 962          */
 963         if (requested_bytes == 0) {
 964                 inode_dio_end(inode);
 965                 nfs_direct_req_release(dreq);
 966                 return result < 0 ? result : -EIO;
 967         }
 968
 969         if (put_dreq(dreq))
 970                 nfs_direct_write_complete(dreq);
 971         return requested_bytes;
 972 }
 973
 974 /**
 975  * nfs_file_direct_write - file direct write operation for NFS files
 976  * @iocb: target I/O control block
 977  * @iter: vector of user buffers from which to write data
 978  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 979  *
 980  * We use this function for direct writes instead of calling
 981  * generic_file_aio_write() in order to avoid taking the inode
 982  * semaphore and updating the i_size.  The NFS server will set
 983  * the new i_size and this client must read the updated size
 984  * back into its cache.  We let the server do generic write
 985  * parameter checking and report problems.
 986  *
 987  * We eliminate local atime updates, see direct read above.
 988  *
 989  * We avoid unnecessary page cache invalidations for normal cached
 990  * readers of this file.
 991  *
 992  * Note that O_APPEND is not supported for NFS direct writes, as there
 993  * is no atomic O_APPEND write facility in the NFS protocol.
 994  */
 995 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 996                               bool swap)
 997 {
 998         ssize_t result = -EINVAL, requested;
 999         size_t count;
1000         struct file *file = iocb->ki_filp;
1001         struct address_space *mapping = file->f_mapping;
1002         struct inode *inode = mapping->host;
1003         struct nfs_direct_req *dreq;
1004         struct nfs_lock_context *l_ctx;
1005         loff_t pos, end;
1006
1007         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1008                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
1009
1010         if (swap)
1011                 /* bypass generic checks */
1012                 result =  iov_iter_count(iter);
1013         else
1014                 result = generic_write_checks(iocb, iter);
1015         if (result <= 0)
1016                 return result;
1017         count = result;
1018         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1019
1020         pos = iocb->ki_pos;
1021         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1022
1023         task_io_account_write(count);
1024
1025         result = -ENOMEM;
1026         dreq = nfs_direct_req_alloc();
1027         if (!dreq)
1028                 goto out;
1029
1030         dreq->inode = inode;
1031         dreq->bytes_left = dreq->max_count = count;
1032         dreq->io_start = pos;
1033         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1034         l_ctx = nfs_get_lock_context(dreq->ctx);
1035         if (IS_ERR(l_ctx)) {
1036                 result = PTR_ERR(l_ctx);
1037                 nfs_direct_req_release(dreq);
1038                 goto out_release;
1039         }
1040         dreq->l_ctx = l_ctx;
1041         if (!is_sync_kiocb(iocb))
1042                 dreq->iocb = iocb;
1043
1044         if (swap) {
1045                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1046                                                             FLUSH_STABLE);
1047         } else {
1048                 nfs_start_io_direct(inode);
1049
1050                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1051                                                             FLUSH_COND_STABLE);
1052
1053                 if (mapping->nrpages) {
1054                         invalidate_inode_pages2_range(mapping,
1055                                                       pos >> PAGE_SHIFT, end);
1056                 }
1057
1058                 nfs_end_io_direct(inode);
1059         }
1060
1061         if (requested > 0) {
1062                 result = nfs_direct_wait(dreq);
1063                 if (result > 0) {
1064                         requested -= result;
1065                         iocb->ki_pos = pos + result;
1066                         /* XXX: should check the generic_write_sync retval */
1067                         generic_write_sync(iocb, result);
1068                 }
1069                 iov_iter_revert(iter, requested);
1070         } else {
1071                 result = requested;
1072         }
1073 out_release:
1074         nfs_direct_req_release(dreq);
1075 out:
1076         return result;
1077 }
1078
1079 /**
1080  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1081  *
1082  */
1083 int __init nfs_init_directcache(void)
1084 {
1085         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1086                                                 sizeof(struct nfs_direct_req),
1087                                                 0, (SLAB_RECLAIM_ACCOUNT|
1088                                                         SLAB_MEM_SPREAD),
1089                                                 NULL);
1090         if (nfs_direct_cachep == NULL)
1091                 return -ENOMEM;
1092
1093         return 0;
1094 }
1095
1096 /**
1097  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1098  *
1099  */
1100 void nfs_destroy_directcache(void)
1101 {
1102         kmem_cache_destroy(nfs_direct_cachep);
1103 }