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 <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  * @rw: direction (read or write)
 244  * @iocb: target I/O control block
 245  * @iov: array of vectors that define I/O buffer
 246  * @pos: offset in file to begin the operation
 247  * @nr_segs: size of iovec array
 248  *
 249  * The presence of this routine in the address space ops vector means
 250  * the NFS client supports direct I/O. However, for most direct IO, we
 251  * shunt off direct read and write requests before the VFS gets them,
 252  * so this method is only ever called for swap.
 253  */
 254 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter, loff_t pos)
 255 {
 256         struct inode *inode = iocb->ki_filp->f_mapping->host;
 257
 258         /* we only support swap file calling nfs_direct_IO */
 259         if (!IS_SWAPFILE(inode))
 260                 return 0;
 261
 262 #ifndef CONFIG_NFS_SWAP
 263         dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
 264                         iocb->ki_filp, (long long) pos, iter->nr_segs);
 265
 266         return -EINVAL;
 267 #else
 268         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 269
 270         if (rw == READ)
 271                 return nfs_file_direct_read(iocb, iter, pos);
 272         return nfs_file_direct_write(iocb, iter, pos);
 273 #endif /* CONFIG_NFS_SWAP */
 274 }
 275
 276 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 277 {
 278         unsigned int i;
 279         for (i = 0; i < npages; i++)
 280                 page_cache_release(pages[i]);
 281 }
 282
 283 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 284                               struct nfs_direct_req *dreq)
 285 {
 286         cinfo->lock = &dreq->lock;
 287         cinfo->mds = &dreq->mds_cinfo;
 288         cinfo->ds = &dreq->ds_cinfo;
 289         cinfo->dreq = dreq;
 290         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 291 }
 292
 293 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 294                                              struct nfs_pageio_descriptor *pgio,
 295                                              struct nfs_page *req)
 296 {
 297         int mirror_count = 1;
 298
 299         if (pgio->pg_ops->pg_get_mirror_count)
 300                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 301
 302         dreq->mirror_count = mirror_count;
 303 }
 304
 305 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 306 {
 307         struct nfs_direct_req *dreq;
 308
 309         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 310         if (!dreq)
 311                 return NULL;
 312
 313         kref_init(&dreq->kref);
 314         kref_get(&dreq->kref);
 315         init_completion(&dreq->completion);
 316         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 317         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 318         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 319         dreq->mirror_count = 1;
 320         spin_lock_init(&dreq->lock);
 321
 322         return dreq;
 323 }
 324
 325 static void nfs_direct_req_free(struct kref *kref)
 326 {
 327         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 328
 329         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 330         if (dreq->l_ctx != NULL)
 331                 nfs_put_lock_context(dreq->l_ctx);
 332         if (dreq->ctx != NULL)
 333                 put_nfs_open_context(dreq->ctx);
 334         kmem_cache_free(nfs_direct_cachep, dreq);
 335 }
 336
 337 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 338 {
 339         kref_put(&dreq->kref, nfs_direct_req_free);
 340 }
 341
 342 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 343 {
 344         return dreq->bytes_left;
 345 }
 346 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 347
 348 /*
 349  * Collects and returns the final error value/byte-count.
 350  */
 351 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 352 {
 353         ssize_t result = -EIOCBQUEUED;
 354
 355         /* Async requests don't wait here */
 356         if (dreq->iocb)
 357                 goto out;
 358
 359         result = wait_for_completion_killable(&dreq->completion);
 360
 361         if (!result)
 362                 result = dreq->error;
 363         if (!result)
 364                 result = dreq->count;
 365
 366 out:
 367         return (ssize_t) result;
 368 }
 369
 370 /*
 371  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 372  * the iocb is still valid here if this is a synchronous request.
 373  */
 374 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 375 {
 376         struct inode *inode = dreq->inode;
 377
 378         if (dreq->iocb && write) {
 379                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 380
 381                 spin_lock(&inode->i_lock);
 382                 if (i_size_read(inode) < pos)
 383                         i_size_write(inode, pos);
 384                 spin_unlock(&inode->i_lock);
 385         }
 386
 387         if (write)
 388                 nfs_zap_mapping(inode, inode->i_mapping);
 389
 390         inode_dio_done(inode);
 391
 392         if (dreq->iocb) {
 393                 long res = (long) dreq->error;
 394                 if (!res)
 395                         res = (long) dreq->count;
 396                 aio_complete(dreq->iocb, res, 0);
 397         }
 398
 399         complete_all(&dreq->completion);
 400
 401         nfs_direct_req_release(dreq);
 402 }
 403
 404 static void nfs_direct_readpage_release(struct nfs_page *req)
 405 {
 406         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 407                 req->wb_context->dentry->d_inode->i_sb->s_id,
 408                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 409                 req->wb_bytes,
 410                 (long long)req_offset(req));
 411         nfs_release_request(req);
 412 }
 413
 414 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 415 {
 416         unsigned long bytes = 0;
 417         struct nfs_direct_req *dreq = hdr->dreq;
 418
 419         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 420                 goto out_put;
 421
 422         spin_lock(&dreq->lock);
 423         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 424                 dreq->error = hdr->error;
 425         else
 426                 nfs_direct_good_bytes(dreq, hdr);
 427
 428         spin_unlock(&dreq->lock);
 429
 430         while (!list_empty(&hdr->pages)) {
 431                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 432                 struct page *page = req->wb_page;
 433
 434                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 435                         set_page_dirty(page);
 436                 bytes += req->wb_bytes;
 437                 nfs_list_remove_request(req);
 438                 nfs_direct_readpage_release(req);
 439         }
 440 out_put:
 441         if (put_dreq(dreq))
 442                 nfs_direct_complete(dreq, false);
 443         hdr->release(hdr);
 444 }
 445
 446 static void nfs_read_sync_pgio_error(struct list_head *head)
 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         atomic_inc(&inode->i_dio_count);
 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_done(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, false);
 549         return 0;
 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  * @pos: byte offset in file where reading starts
 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                                 loff_t pos)
 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;
 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) pos);
 586
 587         result = 0;
 588         if (!count)
 589                 goto out;
 590
 591         mutex_lock(&inode->i_mutex);
 592         result = nfs_sync_mapping(mapping);
 593         if (result)
 594                 goto out_unlock;
 595
 596         task_io_account_read(count);
 597
 598         result = -ENOMEM;
 599         dreq = nfs_direct_req_alloc();
 600         if (dreq == NULL)
 601                 goto out_unlock;
 602
 603         dreq->inode = inode;
 604         dreq->bytes_left = count;
 605         dreq->io_start = pos;
 606         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 607         l_ctx = nfs_get_lock_context(dreq->ctx);
 608         if (IS_ERR(l_ctx)) {
 609                 result = PTR_ERR(l_ctx);
 610                 goto out_release;
 611         }
 612         dreq->l_ctx = l_ctx;
 613         if (!is_sync_kiocb(iocb))
 614                 dreq->iocb = iocb;
 615
 616         NFS_I(inode)->read_io += count;
 617         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 618
 619         mutex_unlock(&inode->i_mutex);
 620
 621         if (!result) {
 622                 result = nfs_direct_wait(dreq);
 623                 if (result > 0)
 624                         iocb->ki_pos = pos + result;
 625         }
 626
 627         nfs_direct_req_release(dreq);
 628         return result;
 629
 630 out_release:
 631         nfs_direct_req_release(dreq);
 632 out_unlock:
 633         mutex_unlock(&inode->i_mutex);
 634 out:
 635         return result;
 636 }
 637
 638 static void
 639 nfs_direct_write_scan_commit_list(struct inode *inode,
 640                                   struct list_head *list,
 641                                   struct nfs_commit_info *cinfo)
 642 {
 643         spin_lock(cinfo->lock);
 644 #ifdef CONFIG_NFS_V4_1
 645         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 646                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 647 #endif
 648         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 649         spin_unlock(cinfo->lock);
 650 }
 651
 652 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 653 {
 654         struct nfs_pageio_descriptor desc;
 655         struct nfs_page *req, *tmp;
 656         LIST_HEAD(reqs);
 657         struct nfs_commit_info cinfo;
 658         LIST_HEAD(failed);
 659         int i;
 660
 661         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 662         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 663
 664         dreq->count = 0;
 665         for (i = 0; i < dreq->mirror_count; i++)
 666                 dreq->mirrors[i].count = 0;
 667         get_dreq(dreq);
 668
 669         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 670                               &nfs_direct_write_completion_ops);
 671         desc.pg_dreq = dreq;
 672
 673         req = nfs_list_entry(reqs.next);
 674         nfs_direct_setup_mirroring(dreq, &desc, req);
 675
 676         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 677                 if (!nfs_pageio_add_request(&desc, req)) {
 678                         nfs_list_remove_request(req);
 679                         nfs_list_add_request(req, &failed);
 680                         spin_lock(cinfo.lock);
 681                         dreq->flags = 0;
 682                         dreq->error = -EIO;
 683                         spin_unlock(cinfo.lock);
 684                 }
 685                 nfs_release_request(req);
 686         }
 687         nfs_pageio_complete(&desc);
 688
 689         while (!list_empty(&failed)) {
 690                 req = nfs_list_entry(failed.next);
 691                 nfs_list_remove_request(req);
 692                 nfs_unlock_and_release_request(req);
 693         }
 694
 695         if (put_dreq(dreq))
 696                 nfs_direct_write_complete(dreq, dreq->inode);
 697 }
 698
 699 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 700 {
 701         struct nfs_direct_req *dreq = data->dreq;
 702         struct nfs_commit_info cinfo;
 703         struct nfs_page *req;
 704         int status = data->task.tk_status;
 705
 706         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 707         if (status < 0) {
 708                 dprintk("NFS: %5u commit failed with error %d.\n",
 709                         data->task.tk_pid, status);
 710                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 711         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 712                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 713                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 714         }
 715
 716         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 717         while (!list_empty(&data->pages)) {
 718                 req = nfs_list_entry(data->pages.next);
 719                 nfs_list_remove_request(req);
 720                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 721                         /* Note the rewrite will go through mds */
 722                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 723                 } else
 724                         nfs_release_request(req);
 725                 nfs_unlock_and_release_request(req);
 726         }
 727
 728         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 729                 nfs_direct_write_complete(dreq, data->inode);
 730 }
 731
 732 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 733 {
 734         /* There is no lock to clear */
 735 }
 736
 737 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 738         .completion = nfs_direct_commit_complete,
 739         .error_cleanup = nfs_direct_error_cleanup,
 740 };
 741
 742 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 743 {
 744         int res;
 745         struct nfs_commit_info cinfo;
 746         LIST_HEAD(mds_list);
 747
 748         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 749         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 750         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 751         if (res < 0) /* res == -ENOMEM */
 752                 nfs_direct_write_reschedule(dreq);
 753 }
 754
 755 static void nfs_direct_write_schedule_work(struct work_struct *work)
 756 {
 757         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 758         int flags = dreq->flags;
 759
 760         dreq->flags = 0;
 761         switch (flags) {
 762                 case NFS_ODIRECT_DO_COMMIT:
 763                         nfs_direct_commit_schedule(dreq);
 764                         break;
 765                 case NFS_ODIRECT_RESCHED_WRITES:
 766                         nfs_direct_write_reschedule(dreq);
 767                         break;
 768                 default:
 769                         nfs_direct_complete(dreq, true);
 770         }
 771 }
 772
 773 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 774 {
 775         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 776 }
 777
 778 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 779 {
 780         struct nfs_direct_req *dreq = hdr->dreq;
 781         struct nfs_commit_info cinfo;
 782         bool request_commit = false;
 783         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 784
 785         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 786                 goto out_put;
 787
 788         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 789
 790         spin_lock(&dreq->lock);
 791
 792         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 793                 dreq->flags = 0;
 794                 dreq->error = hdr->error;
 795         }
 796         if (dreq->error == 0) {
 797                 nfs_direct_good_bytes(dreq, hdr);
 798                 if (nfs_write_need_commit(hdr)) {
 799                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 800                                 request_commit = true;
 801                         else if (dreq->flags == 0) {
 802                                 nfs_direct_set_hdr_verf(dreq, hdr);
 803                                 request_commit = true;
 804                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 805                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 806                                 request_commit = true;
 807                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 808                                         dreq->flags =
 809                                                 NFS_ODIRECT_RESCHED_WRITES;
 810                         }
 811                 }
 812         }
 813         spin_unlock(&dreq->lock);
 814
 815         while (!list_empty(&hdr->pages)) {
 816
 817                 req = nfs_list_entry(hdr->pages.next);
 818                 nfs_list_remove_request(req);
 819                 if (request_commit) {
 820                         kref_get(&req->wb_kref);
 821                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 822                                 hdr->ds_commit_idx);
 823                 }
 824                 nfs_unlock_and_release_request(req);
 825         }
 826
 827 out_put:
 828         if (put_dreq(dreq))
 829                 nfs_direct_write_complete(dreq, hdr->inode);
 830         hdr->release(hdr);
 831 }
 832
 833 static void nfs_write_sync_pgio_error(struct list_head *head)
 834 {
 835         struct nfs_page *req;
 836
 837         while (!list_empty(head)) {
 838                 req = nfs_list_entry(head->next);
 839                 nfs_list_remove_request(req);
 840                 nfs_unlock_and_release_request(req);
 841         }
 842 }
 843
 844 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 845         .error_cleanup = nfs_write_sync_pgio_error,
 846         .init_hdr = nfs_direct_pgio_init,
 847         .completion = nfs_direct_write_completion,
 848 };
 849
 850
 851 /*
 852  * NB: Return the value of the first error return code.  Subsequent
 853  *     errors after the first one are ignored.
 854  */
 855 /*
 856  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 857  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 858  * bail and stop sending more writes.  Write length accounting is
 859  * handled automatically by nfs_direct_write_result().  Otherwise, if
 860  * no requests have been sent, just return an error.
 861  */
 862 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 863                                                struct iov_iter *iter,
 864                                                loff_t pos)
 865 {
 866         struct nfs_pageio_descriptor desc;
 867         struct inode *inode = dreq->inode;
 868         ssize_t result = 0;
 869         size_t requested_bytes = 0;
 870         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 871
 872         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 873                               &nfs_direct_write_completion_ops);
 874         desc.pg_dreq = dreq;
 875         get_dreq(dreq);
 876         atomic_inc(&inode->i_dio_count);
 877
 878         NFS_I(inode)->write_io += iov_iter_count(iter);
 879         while (iov_iter_count(iter)) {
 880                 struct page **pagevec;
 881                 size_t bytes;
 882                 size_t pgbase;
 883                 unsigned npages, i;
 884
 885                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 886                                                   wsize, &pgbase);
 887                 if (result < 0)
 888                         break;
 889
 890                 bytes = result;
 891                 iov_iter_advance(iter, bytes);
 892                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 893                 for (i = 0; i < npages; i++) {
 894                         struct nfs_page *req;
 895                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 896
 897                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 898                                                  pgbase, req_len);
 899                         if (IS_ERR(req)) {
 900                                 result = PTR_ERR(req);
 901                                 break;
 902                         }
 903
 904                         nfs_direct_setup_mirroring(dreq, &desc, req);
 905
 906                         nfs_lock_request(req);
 907                         req->wb_index = pos >> PAGE_SHIFT;
 908                         req->wb_offset = pos & ~PAGE_MASK;
 909                         if (!nfs_pageio_add_request(&desc, req)) {
 910                                 result = desc.pg_error;
 911                                 nfs_unlock_and_release_request(req);
 912                                 break;
 913                         }
 914                         pgbase = 0;
 915                         bytes -= req_len;
 916                         requested_bytes += req_len;
 917                         pos += req_len;
 918                         dreq->bytes_left -= req_len;
 919                 }
 920                 nfs_direct_release_pages(pagevec, npages);
 921                 kvfree(pagevec);
 922                 if (result < 0)
 923                         break;
 924         }
 925         nfs_pageio_complete(&desc);
 926
 927         /*
 928          * If no bytes were started, return the error, and let the
 929          * generic layer handle the completion.
 930          */
 931         if (requested_bytes == 0) {
 932                 inode_dio_done(inode);
 933                 nfs_direct_req_release(dreq);
 934                 return result < 0 ? result : -EIO;
 935         }
 936
 937         if (put_dreq(dreq))
 938                 nfs_direct_write_complete(dreq, dreq->inode);
 939         return 0;
 940 }
 941
 942 /**
 943  * nfs_file_direct_write - file direct write operation for NFS files
 944  * @iocb: target I/O control block
 945  * @iter: vector of user buffers from which to write data
 946  * @pos: byte offset in file where writing starts
 947  *
 948  * We use this function for direct writes instead of calling
 949  * generic_file_aio_write() in order to avoid taking the inode
 950  * semaphore and updating the i_size.  The NFS server will set
 951  * the new i_size and this client must read the updated size
 952  * back into its cache.  We let the server do generic write
 953  * parameter checking and report problems.
 954  *
 955  * We eliminate local atime updates, see direct read above.
 956  *
 957  * We avoid unnecessary page cache invalidations for normal cached
 958  * readers of this file.
 959  *
 960  * Note that O_APPEND is not supported for NFS direct writes, as there
 961  * is no atomic O_APPEND write facility in the NFS protocol.
 962  */
 963 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 964                                 loff_t pos)
 965 {
 966         ssize_t result = -EINVAL;
 967         struct file *file = iocb->ki_filp;
 968         struct address_space *mapping = file->f_mapping;
 969         struct inode *inode = mapping->host;
 970         struct nfs_direct_req *dreq;
 971         struct nfs_lock_context *l_ctx;
 972         loff_t end;
 973         size_t count = iov_iter_count(iter);
 974         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 975
 976         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 977
 978         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 979                 file, count, (long long) pos);
 980
 981         result = generic_write_checks(file, &pos, &count, 0);
 982         if (result)
 983                 goto out;
 984
 985         result = -EINVAL;
 986         if ((ssize_t) count < 0)
 987                 goto out;
 988         result = 0;
 989         if (!count)
 990                 goto out;
 991
 992         mutex_lock(&inode->i_mutex);
 993
 994         result = nfs_sync_mapping(mapping);
 995         if (result)
 996                 goto out_unlock;
 997
 998         if (mapping->nrpages) {
 999                 result = invalidate_inode_pages2_range(mapping,
1000                                         pos >> PAGE_CACHE_SHIFT, end);
1001                 if (result)
1002                         goto out_unlock;
1003         }
1004
1005         task_io_account_write(count);
1006
1007         result = -ENOMEM;
1008         dreq = nfs_direct_req_alloc();
1009         if (!dreq)
1010                 goto out_unlock;
1011
1012         dreq->inode = inode;
1013         dreq->bytes_left = count;
1014         dreq->io_start = pos;
1015         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1016         l_ctx = nfs_get_lock_context(dreq->ctx);
1017         if (IS_ERR(l_ctx)) {
1018                 result = PTR_ERR(l_ctx);
1019                 goto out_release;
1020         }
1021         dreq->l_ctx = l_ctx;
1022         if (!is_sync_kiocb(iocb))
1023                 dreq->iocb = iocb;
1024
1025         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1026
1027         if (mapping->nrpages) {
1028                 invalidate_inode_pages2_range(mapping,
1029                                               pos >> PAGE_CACHE_SHIFT, end);
1030         }
1031
1032         mutex_unlock(&inode->i_mutex);
1033
1034         if (!result) {
1035                 result = nfs_direct_wait(dreq);
1036                 if (result > 0) {
1037                         struct inode *inode = mapping->host;
1038
1039                         iocb->ki_pos = pos + result;
1040                         spin_lock(&inode->i_lock);
1041                         if (i_size_read(inode) < iocb->ki_pos)
1042                                 i_size_write(inode, iocb->ki_pos);
1043                         spin_unlock(&inode->i_lock);
1044                 }
1045         }
1046         nfs_direct_req_release(dreq);
1047         return result;
1048
1049 out_release:
1050         nfs_direct_req_release(dreq);
1051 out_unlock:
1052         mutex_unlock(&inode->i_mutex);
1053 out:
1054         return result;
1055 }
1056
1057 /**
1058  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1059  *
1060  */
1061 int __init nfs_init_directcache(void)
1062 {
1063         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1064                                                 sizeof(struct nfs_direct_req),
1065                                                 0, (SLAB_RECLAIM_ACCOUNT|
1066                                                         SLAB_MEM_SPREAD),
1067                                                 NULL);
1068         if (nfs_direct_cachep == NULL)
1069                 return -ENOMEM;
1070
1071         return 0;
1072 }
1073
1074 /**
1075  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1076  *
1077  */
1078 void nfs_destroy_directcache(void)
1079 {
1080         kmem_cache_destroy(nfs_direct_cachep);
1081 }