fs/nfs/direct.c

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