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         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 263
 264         if (rw == READ)
 265                 return nfs_file_direct_read(iocb, iter, pos);
 266         return nfs_file_direct_write(iocb, iter, pos);
 267 }
 268
 269 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 270 {
 271         unsigned int i;
 272         for (i = 0; i < npages; i++)
 273                 page_cache_release(pages[i]);
 274 }
 275
 276 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 277                               struct nfs_direct_req *dreq)
 278 {
 279         cinfo->lock = &dreq->inode->i_lock;
 280         cinfo->mds = &dreq->mds_cinfo;
 281         cinfo->ds = &dreq->ds_cinfo;
 282         cinfo->dreq = dreq;
 283         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 284 }
 285
 286 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 287                                              struct nfs_pageio_descriptor *pgio,
 288                                              struct nfs_page *req)
 289 {
 290         int mirror_count = 1;
 291
 292         if (pgio->pg_ops->pg_get_mirror_count)
 293                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 294
 295         dreq->mirror_count = mirror_count;
 296 }
 297
 298 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 299 {
 300         struct nfs_direct_req *dreq;
 301
 302         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 303         if (!dreq)
 304                 return NULL;
 305
 306         kref_init(&dreq->kref);
 307         kref_get(&dreq->kref);
 308         init_completion(&dreq->completion);
 309         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 310         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 311         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 312         dreq->mirror_count = 1;
 313         spin_lock_init(&dreq->lock);
 314
 315         return dreq;
 316 }
 317
 318 static void nfs_direct_req_free(struct kref *kref)
 319 {
 320         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 321
 322         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 323         if (dreq->l_ctx != NULL)
 324                 nfs_put_lock_context(dreq->l_ctx);
 325         if (dreq->ctx != NULL)
 326                 put_nfs_open_context(dreq->ctx);
 327         kmem_cache_free(nfs_direct_cachep, dreq);
 328 }
 329
 330 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 331 {
 332         kref_put(&dreq->kref, nfs_direct_req_free);
 333 }
 334
 335 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 336 {
 337         return dreq->bytes_left;
 338 }
 339 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 340
 341 /*
 342  * Collects and returns the final error value/byte-count.
 343  */
 344 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 345 {
 346         ssize_t result = -EIOCBQUEUED;
 347
 348         /* Async requests don't wait here */
 349         if (dreq->iocb)
 350                 goto out;
 351
 352         result = wait_for_completion_killable(&dreq->completion);
 353
 354         if (!result)
 355                 result = dreq->error;
 356         if (!result)
 357                 result = dreq->count;
 358
 359 out:
 360         return (ssize_t) result;
 361 }
 362
 363 /*
 364  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 365  * the iocb is still valid here if this is a synchronous request.
 366  */
 367 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 368 {
 369         struct inode *inode = dreq->inode;
 370
 371         if (dreq->iocb && write) {
 372                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 373
 374                 spin_lock(&inode->i_lock);
 375                 if (i_size_read(inode) < pos)
 376                         i_size_write(inode, pos);
 377                 spin_unlock(&inode->i_lock);
 378         }
 379
 380         if (write)
 381                 nfs_zap_mapping(inode, inode->i_mapping);
 382
 383         inode_dio_done(inode);
 384
 385         if (dreq->iocb) {
 386                 long res = (long) dreq->error;
 387                 if (!res)
 388                         res = (long) dreq->count;
 389                 aio_complete(dreq->iocb, res, 0);
 390         }
 391
 392         complete_all(&dreq->completion);
 393
 394         nfs_direct_req_release(dreq);
 395 }
 396
 397 static void nfs_direct_readpage_release(struct nfs_page *req)
 398 {
 399         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 400                 req->wb_context->dentry->d_inode->i_sb->s_id,
 401                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 402                 req->wb_bytes,
 403                 (long long)req_offset(req));
 404         nfs_release_request(req);
 405 }
 406
 407 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 408 {
 409         unsigned long bytes = 0;
 410         struct nfs_direct_req *dreq = hdr->dreq;
 411
 412         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 413                 goto out_put;
 414
 415         spin_lock(&dreq->lock);
 416         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 417                 dreq->error = hdr->error;
 418         else
 419                 nfs_direct_good_bytes(dreq, hdr);
 420
 421         spin_unlock(&dreq->lock);
 422
 423         while (!list_empty(&hdr->pages)) {
 424                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 425                 struct page *page = req->wb_page;
 426
 427                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 428                         set_page_dirty(page);
 429                 bytes += req->wb_bytes;
 430                 nfs_list_remove_request(req);
 431                 nfs_direct_readpage_release(req);
 432         }
 433 out_put:
 434         if (put_dreq(dreq))
 435                 nfs_direct_complete(dreq, false);
 436         hdr->release(hdr);
 437 }
 438
 439 static void nfs_read_sync_pgio_error(struct list_head *head)
 440 {
 441         struct nfs_page *req;
 442
 443         while (!list_empty(head)) {
 444                 req = nfs_list_entry(head->next);
 445                 nfs_list_remove_request(req);
 446                 nfs_release_request(req);
 447         }
 448 }
 449
 450 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 451 {
 452         get_dreq(hdr->dreq);
 453 }
 454
 455 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 456         .error_cleanup = nfs_read_sync_pgio_error,
 457         .init_hdr = nfs_direct_pgio_init,
 458         .completion = nfs_direct_read_completion,
 459 };
 460
 461 /*
 462  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 463  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 464  * bail and stop sending more reads.  Read length accounting is
 465  * handled automatically by nfs_direct_read_result().  Otherwise, if
 466  * no requests have been sent, just return an error.
 467  */
 468
 469 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 470                                               struct iov_iter *iter,
 471                                               loff_t pos)
 472 {
 473         struct nfs_pageio_descriptor desc;
 474         struct inode *inode = dreq->inode;
 475         ssize_t result = -EINVAL;
 476         size_t requested_bytes = 0;
 477         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 478
 479         nfs_pageio_init_read(&desc, dreq->inode, false,
 480                              &nfs_direct_read_completion_ops);
 481         get_dreq(dreq);
 482         desc.pg_dreq = dreq;
 483         atomic_inc(&inode->i_dio_count);
 484
 485         while (iov_iter_count(iter)) {
 486                 struct page **pagevec;
 487                 size_t bytes;
 488                 size_t pgbase;
 489                 unsigned npages, i;
 490
 491                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 492                                                   rsize, &pgbase);
 493                 if (result < 0)
 494                         break;
 495
 496                 bytes = result;
 497                 iov_iter_advance(iter, bytes);
 498                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 499                 for (i = 0; i < npages; i++) {
 500                         struct nfs_page *req;
 501                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 502                         /* XXX do we need to do the eof zeroing found in async_filler? */
 503                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 504                                                  pgbase, req_len);
 505                         if (IS_ERR(req)) {
 506                                 result = PTR_ERR(req);
 507                                 break;
 508                         }
 509                         req->wb_index = pos >> PAGE_SHIFT;
 510                         req->wb_offset = pos & ~PAGE_MASK;
 511                         if (!nfs_pageio_add_request(&desc, req)) {
 512                                 result = desc.pg_error;
 513                                 nfs_release_request(req);
 514                                 break;
 515                         }
 516                         pgbase = 0;
 517                         bytes -= req_len;
 518                         requested_bytes += req_len;
 519                         pos += req_len;
 520                         dreq->bytes_left -= req_len;
 521                 }
 522                 nfs_direct_release_pages(pagevec, npages);
 523                 kvfree(pagevec);
 524                 if (result < 0)
 525                         break;
 526         }
 527
 528         nfs_pageio_complete(&desc);
 529
 530         /*
 531          * If no bytes were started, return the error, and let the
 532          * generic layer handle the completion.
 533          */
 534         if (requested_bytes == 0) {
 535                 inode_dio_done(inode);
 536                 nfs_direct_req_release(dreq);
 537                 return result < 0 ? result : -EIO;
 538         }
 539
 540         if (put_dreq(dreq))
 541                 nfs_direct_complete(dreq, false);
 542         return 0;
 543 }
 544
 545 /**
 546  * nfs_file_direct_read - file direct read operation for NFS files
 547  * @iocb: target I/O control block
 548  * @iter: vector of user buffers into which to read data
 549  * @pos: byte offset in file where reading starts
 550  *
 551  * We use this function for direct reads instead of calling
 552  * generic_file_aio_read() in order to avoid gfar's check to see if
 553  * the request starts before the end of the file.  For that check
 554  * to work, we must generate a GETATTR before each direct read, and
 555  * even then there is a window between the GETATTR and the subsequent
 556  * READ where the file size could change.  Our preference is simply
 557  * to do all reads the application wants, and the server will take
 558  * care of managing the end of file boundary.
 559  *
 560  * This function also eliminates unnecessarily updating the file's
 561  * atime locally, as the NFS server sets the file's atime, and this
 562  * client must read the updated atime from the server back into its
 563  * cache.
 564  */
 565 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 566                                 loff_t pos)
 567 {
 568         struct file *file = iocb->ki_filp;
 569         struct address_space *mapping = file->f_mapping;
 570         struct inode *inode = mapping->host;
 571         struct nfs_direct_req *dreq;
 572         struct nfs_lock_context *l_ctx;
 573         ssize_t result = -EINVAL;
 574         size_t count = iov_iter_count(iter);
 575         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 576
 577         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 578                 file, count, (long long) pos);
 579
 580         result = 0;
 581         if (!count)
 582                 goto out;
 583
 584         mutex_lock(&inode->i_mutex);
 585         result = nfs_sync_mapping(mapping);
 586         if (result)
 587                 goto out_unlock;
 588
 589         task_io_account_read(count);
 590
 591         result = -ENOMEM;
 592         dreq = nfs_direct_req_alloc();
 593         if (dreq == NULL)
 594                 goto out_unlock;
 595
 596         dreq->inode = inode;
 597         dreq->bytes_left = count;
 598         dreq->io_start = pos;
 599         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 600         l_ctx = nfs_get_lock_context(dreq->ctx);
 601         if (IS_ERR(l_ctx)) {
 602                 result = PTR_ERR(l_ctx);
 603                 goto out_release;
 604         }
 605         dreq->l_ctx = l_ctx;
 606         if (!is_sync_kiocb(iocb))
 607                 dreq->iocb = iocb;
 608
 609         NFS_I(inode)->read_io += count;
 610         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 611
 612         mutex_unlock(&inode->i_mutex);
 613
 614         if (!result) {
 615                 result = nfs_direct_wait(dreq);
 616                 if (result > 0)
 617                         iocb->ki_pos = pos + result;
 618         }
 619
 620         nfs_direct_req_release(dreq);
 621         return result;
 622
 623 out_release:
 624         nfs_direct_req_release(dreq);
 625 out_unlock:
 626         mutex_unlock(&inode->i_mutex);
 627 out:
 628         return result;
 629 }
 630
 631 static void
 632 nfs_direct_write_scan_commit_list(struct inode *inode,
 633                                   struct list_head *list,
 634                                   struct nfs_commit_info *cinfo)
 635 {
 636         spin_lock(cinfo->lock);
 637 #ifdef CONFIG_NFS_V4_1
 638         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 639                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 640 #endif
 641         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 642         spin_unlock(cinfo->lock);
 643 }
 644
 645 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 646 {
 647         struct nfs_pageio_descriptor desc;
 648         struct nfs_page *req, *tmp;
 649         LIST_HEAD(reqs);
 650         struct nfs_commit_info cinfo;
 651         LIST_HEAD(failed);
 652         int i;
 653
 654         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 655         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 656
 657         dreq->count = 0;
 658         for (i = 0; i < dreq->mirror_count; i++)
 659                 dreq->mirrors[i].count = 0;
 660         get_dreq(dreq);
 661
 662         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 663                               &nfs_direct_write_completion_ops);
 664         desc.pg_dreq = dreq;
 665
 666         req = nfs_list_entry(reqs.next);
 667         nfs_direct_setup_mirroring(dreq, &desc, req);
 668
 669         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 670                 if (!nfs_pageio_add_request(&desc, req)) {
 671                         nfs_list_remove_request(req);
 672                         nfs_list_add_request(req, &failed);
 673                         spin_lock(cinfo.lock);
 674                         dreq->flags = 0;
 675                         dreq->error = -EIO;
 676                         spin_unlock(cinfo.lock);
 677                 }
 678                 nfs_release_request(req);
 679         }
 680         nfs_pageio_complete(&desc);
 681
 682         while (!list_empty(&failed)) {
 683                 req = nfs_list_entry(failed.next);
 684                 nfs_list_remove_request(req);
 685                 nfs_unlock_and_release_request(req);
 686         }
 687
 688         if (put_dreq(dreq))
 689                 nfs_direct_write_complete(dreq, dreq->inode);
 690 }
 691
 692 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 693 {
 694         struct nfs_direct_req *dreq = data->dreq;
 695         struct nfs_commit_info cinfo;
 696         struct nfs_page *req;
 697         int status = data->task.tk_status;
 698
 699         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 700         if (status < 0) {
 701                 dprintk("NFS: %5u commit failed with error %d.\n",
 702                         data->task.tk_pid, status);
 703                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 704         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 705                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 706                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 707         }
 708
 709         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 710         while (!list_empty(&data->pages)) {
 711                 req = nfs_list_entry(data->pages.next);
 712                 nfs_list_remove_request(req);
 713                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 714                         /* Note the rewrite will go through mds */
 715                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 716                 } else
 717                         nfs_release_request(req);
 718                 nfs_unlock_and_release_request(req);
 719         }
 720
 721         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 722                 nfs_direct_write_complete(dreq, data->inode);
 723 }
 724
 725 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 726 {
 727         /* There is no lock to clear */
 728 }
 729
 730 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 731         .completion = nfs_direct_commit_complete,
 732         .error_cleanup = nfs_direct_error_cleanup,
 733 };
 734
 735 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 736 {
 737         int res;
 738         struct nfs_commit_info cinfo;
 739         LIST_HEAD(mds_list);
 740
 741         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 742         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 743         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 744         if (res < 0) /* res == -ENOMEM */
 745                 nfs_direct_write_reschedule(dreq);
 746 }
 747
 748 static void nfs_direct_write_schedule_work(struct work_struct *work)
 749 {
 750         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 751         int flags = dreq->flags;
 752
 753         dreq->flags = 0;
 754         switch (flags) {
 755                 case NFS_ODIRECT_DO_COMMIT:
 756                         nfs_direct_commit_schedule(dreq);
 757                         break;
 758                 case NFS_ODIRECT_RESCHED_WRITES:
 759                         nfs_direct_write_reschedule(dreq);
 760                         break;
 761                 default:
 762                         nfs_direct_complete(dreq, true);
 763         }
 764 }
 765
 766 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 767 {
 768         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 769 }
 770
 771 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 772 {
 773         struct nfs_direct_req *dreq = hdr->dreq;
 774         struct nfs_commit_info cinfo;
 775         bool request_commit = false;
 776         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 777
 778         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 779                 goto out_put;
 780
 781         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 782
 783         spin_lock(&dreq->lock);
 784
 785         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 786                 dreq->flags = 0;
 787                 dreq->error = hdr->error;
 788         }
 789         if (dreq->error == 0) {
 790                 nfs_direct_good_bytes(dreq, hdr);
 791                 if (nfs_write_need_commit(hdr)) {
 792                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 793                                 request_commit = true;
 794                         else if (dreq->flags == 0) {
 795                                 nfs_direct_set_hdr_verf(dreq, hdr);
 796                                 request_commit = true;
 797                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 798                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 799                                 request_commit = true;
 800                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 801                                         dreq->flags =
 802                                                 NFS_ODIRECT_RESCHED_WRITES;
 803                         }
 804                 }
 805         }
 806         spin_unlock(&dreq->lock);
 807
 808         while (!list_empty(&hdr->pages)) {
 809
 810                 req = nfs_list_entry(hdr->pages.next);
 811                 nfs_list_remove_request(req);
 812                 if (request_commit) {
 813                         kref_get(&req->wb_kref);
 814                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 815                                 hdr->ds_commit_idx);
 816                 }
 817                 nfs_unlock_and_release_request(req);
 818         }
 819
 820 out_put:
 821         if (put_dreq(dreq))
 822                 nfs_direct_write_complete(dreq, hdr->inode);
 823         hdr->release(hdr);
 824 }
 825
 826 static void nfs_write_sync_pgio_error(struct list_head *head)
 827 {
 828         struct nfs_page *req;
 829
 830         while (!list_empty(head)) {
 831                 req = nfs_list_entry(head->next);
 832                 nfs_list_remove_request(req);
 833                 nfs_unlock_and_release_request(req);
 834         }
 835 }
 836
 837 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 838         .error_cleanup = nfs_write_sync_pgio_error,
 839         .init_hdr = nfs_direct_pgio_init,
 840         .completion = nfs_direct_write_completion,
 841 };
 842
 843
 844 /*
 845  * NB: Return the value of the first error return code.  Subsequent
 846  *     errors after the first one are ignored.
 847  */
 848 /*
 849  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 850  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 851  * bail and stop sending more writes.  Write length accounting is
 852  * handled automatically by nfs_direct_write_result().  Otherwise, if
 853  * no requests have been sent, just return an error.
 854  */
 855 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 856                                                struct iov_iter *iter,
 857                                                loff_t pos)
 858 {
 859         struct nfs_pageio_descriptor desc;
 860         struct inode *inode = dreq->inode;
 861         ssize_t result = 0;
 862         size_t requested_bytes = 0;
 863         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 864
 865         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 866                               &nfs_direct_write_completion_ops);
 867         desc.pg_dreq = dreq;
 868         get_dreq(dreq);
 869         atomic_inc(&inode->i_dio_count);
 870
 871         NFS_I(inode)->write_io += iov_iter_count(iter);
 872         while (iov_iter_count(iter)) {
 873                 struct page **pagevec;
 874                 size_t bytes;
 875                 size_t pgbase;
 876                 unsigned npages, i;
 877
 878                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 879                                                   wsize, &pgbase);
 880                 if (result < 0)
 881                         break;
 882
 883                 bytes = result;
 884                 iov_iter_advance(iter, bytes);
 885                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 886                 for (i = 0; i < npages; i++) {
 887                         struct nfs_page *req;
 888                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 889
 890                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 891                                                  pgbase, req_len);
 892                         if (IS_ERR(req)) {
 893                                 result = PTR_ERR(req);
 894                                 break;
 895                         }
 896
 897                         nfs_direct_setup_mirroring(dreq, &desc, req);
 898
 899                         nfs_lock_request(req);
 900                         req->wb_index = pos >> PAGE_SHIFT;
 901                         req->wb_offset = pos & ~PAGE_MASK;
 902                         if (!nfs_pageio_add_request(&desc, req)) {
 903                                 result = desc.pg_error;
 904                                 nfs_unlock_and_release_request(req);
 905                                 break;
 906                         }
 907                         pgbase = 0;
 908                         bytes -= req_len;
 909                         requested_bytes += req_len;
 910                         pos += req_len;
 911                         dreq->bytes_left -= req_len;
 912                 }
 913                 nfs_direct_release_pages(pagevec, npages);
 914                 kvfree(pagevec);
 915                 if (result < 0)
 916                         break;
 917         }
 918         nfs_pageio_complete(&desc);
 919
 920         /*
 921          * If no bytes were started, return the error, and let the
 922          * generic layer handle the completion.
 923          */
 924         if (requested_bytes == 0) {
 925                 inode_dio_done(inode);
 926                 nfs_direct_req_release(dreq);
 927                 return result < 0 ? result : -EIO;
 928         }
 929
 930         if (put_dreq(dreq))
 931                 nfs_direct_write_complete(dreq, dreq->inode);
 932         return 0;
 933 }
 934
 935 /**
 936  * nfs_file_direct_write - file direct write operation for NFS files
 937  * @iocb: target I/O control block
 938  * @iter: vector of user buffers from which to write data
 939  * @pos: byte offset in file where writing starts
 940  *
 941  * We use this function for direct writes instead of calling
 942  * generic_file_aio_write() in order to avoid taking the inode
 943  * semaphore and updating the i_size.  The NFS server will set
 944  * the new i_size and this client must read the updated size
 945  * back into its cache.  We let the server do generic write
 946  * parameter checking and report problems.
 947  *
 948  * We eliminate local atime updates, see direct read above.
 949  *
 950  * We avoid unnecessary page cache invalidations for normal cached
 951  * readers of this file.
 952  *
 953  * Note that O_APPEND is not supported for NFS direct writes, as there
 954  * is no atomic O_APPEND write facility in the NFS protocol.
 955  */
 956 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 957                                 loff_t pos)
 958 {
 959         ssize_t result = -EINVAL;
 960         struct file *file = iocb->ki_filp;
 961         struct address_space *mapping = file->f_mapping;
 962         struct inode *inode = mapping->host;
 963         struct nfs_direct_req *dreq;
 964         struct nfs_lock_context *l_ctx;
 965         loff_t end;
 966         size_t count = iov_iter_count(iter);
 967         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 968
 969         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 970
 971         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 972                 file, count, (long long) pos);
 973
 974         result = generic_write_checks(file, &pos, &count, 0);
 975         if (result)
 976                 goto out;
 977
 978         result = -EINVAL;
 979         if ((ssize_t) count < 0)
 980                 goto out;
 981         result = 0;
 982         if (!count)
 983                 goto out;
 984
 985         mutex_lock(&inode->i_mutex);
 986
 987         result = nfs_sync_mapping(mapping);
 988         if (result)
 989                 goto out_unlock;
 990
 991         if (mapping->nrpages) {
 992                 result = invalidate_inode_pages2_range(mapping,
 993                                         pos >> PAGE_CACHE_SHIFT, end);
 994                 if (result)
 995                         goto out_unlock;
 996         }
 997
 998         task_io_account_write(count);
 999
1000         result = -ENOMEM;
1001         dreq = nfs_direct_req_alloc();
1002         if (!dreq)
1003                 goto out_unlock;
1004
1005         dreq->inode = inode;
1006         dreq->bytes_left = count;
1007         dreq->io_start = pos;
1008         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1009         l_ctx = nfs_get_lock_context(dreq->ctx);
1010         if (IS_ERR(l_ctx)) {
1011                 result = PTR_ERR(l_ctx);
1012                 goto out_release;
1013         }
1014         dreq->l_ctx = l_ctx;
1015         if (!is_sync_kiocb(iocb))
1016                 dreq->iocb = iocb;
1017
1018         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1019
1020         if (mapping->nrpages) {
1021                 invalidate_inode_pages2_range(mapping,
1022                                               pos >> PAGE_CACHE_SHIFT, end);
1023         }
1024
1025         mutex_unlock(&inode->i_mutex);
1026
1027         if (!result) {
1028                 result = nfs_direct_wait(dreq);
1029                 if (result > 0) {
1030                         struct inode *inode = mapping->host;
1031
1032                         iocb->ki_pos = pos + result;
1033                         spin_lock(&inode->i_lock);
1034                         if (i_size_read(inode) < iocb->ki_pos)
1035                                 i_size_write(inode, iocb->ki_pos);
1036                         spin_unlock(&inode->i_lock);
1037                 }
1038         }
1039         nfs_direct_req_release(dreq);
1040         return result;
1041
1042 out_release:
1043         nfs_direct_req_release(dreq);
1044 out_unlock:
1045         mutex_unlock(&inode->i_mutex);
1046 out:
1047         return result;
1048 }
1049
1050 /**
1051  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1052  *
1053  */
1054 int __init nfs_init_directcache(void)
1055 {
1056         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1057                                                 sizeof(struct nfs_direct_req),
1058                                                 0, (SLAB_RECLAIM_ACCOUNT|
1059                                                         SLAB_MEM_SPREAD),
1060                                                 NULL);
1061         if (nfs_direct_cachep == NULL)
1062                 return -ENOMEM;
1063
1064         return 0;
1065 }
1066
1067 /**
1068  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1069  *
1070  */
1071 void nfs_destroy_directcache(void)
1072 {
1073         kmem_cache_destroy(nfs_direct_cachep);
1074 }