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         if (dreq->mirror_count == 1) {
 133                 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
 134                 dreq->count += hdr->good_bytes;
 135         } else {
 136                 /* mirrored writes */
 137                 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
 138                 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
 139                         count = hdr->io_start + hdr->good_bytes - dreq->io_start;
 140                         dreq->mirrors[hdr->pgio_mirror_idx].count = count;
 141                 }
 142                 /* update the dreq->count by finding the minimum agreed count from all
 143                  * mirrors */
 144                 count = dreq->mirrors[0].count;
 145
 146                 for (i = 1; i < dreq->mirror_count; i++)
 147                         count = min(count, dreq->mirrors[i].count);
 148
 149                 dreq->count = count;
 150         }
 151 }
 152
 153 /*
 154  * nfs_direct_select_verf - select the right verifier
 155  * @dreq - direct request possibly spanning multiple servers
 156  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 157  * @commit_idx - commit bucket index for the DS
 158  *
 159  * returns the correct verifier to use given the role of the server
 160  */
 161 static struct nfs_writeverf *
 162 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 163                        struct nfs_client *ds_clp,
 164                        int commit_idx)
 165 {
 166         struct nfs_writeverf *verfp = &dreq->verf;
 167
 168 #ifdef CONFIG_NFS_V4_1
 169         /*
 170          * pNFS is in use, use the DS verf except commit_through_mds is set
 171          * for layout segment where nbuckets is zero.
 172          */
 173         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 174                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 175                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 176                 else
 177                         WARN_ON_ONCE(1);
 178         }
 179 #endif
 180         return verfp;
 181 }
 182
 183
 184 /*
 185  * nfs_direct_set_hdr_verf - set the write/commit verifier
 186  * @dreq - direct request possibly spanning multiple servers
 187  * @hdr - pageio header to validate against previously seen verfs
 188  *
 189  * Set the server's (MDS or DS) "seen" verifier
 190  */
 191 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 192                                     struct nfs_pgio_header *hdr)
 193 {
 194         struct nfs_writeverf *verfp;
 195
 196         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 197         WARN_ON_ONCE(verfp->committed >= 0);
 198         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 199         WARN_ON_ONCE(verfp->committed < 0);
 200 }
 201
 202 /*
 203  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 204  * @dreq - direct request possibly spanning multiple servers
 205  * @hdr - pageio header to validate against previously seen verf
 206  *
 207  * set the server's "seen" verf if not initialized.
 208  * returns result of comparison between @hdr->verf and the "seen"
 209  * verf of the server used by @hdr (DS or MDS)
 210  */
 211 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 212                                           struct nfs_pgio_header *hdr)
 213 {
 214         struct nfs_writeverf *verfp;
 215
 216         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 217         if (verfp->committed < 0) {
 218                 nfs_direct_set_hdr_verf(dreq, hdr);
 219                 return 0;
 220         }
 221         return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 222 }
 223
 224 /*
 225  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 226  * @dreq - direct request possibly spanning multiple servers
 227  * @data - commit data to validate against previously seen verf
 228  *
 229  * returns result of comparison between @data->verf and the verf of
 230  * the server used by @data (DS or MDS)
 231  */
 232 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 233                                            struct nfs_commit_data *data)
 234 {
 235         struct nfs_writeverf *verfp;
 236
 237         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 238                                          data->ds_commit_index);
 239
 240         /* verifier not set so always fail */
 241         if (verfp->committed < 0)
 242                 return 1;
 243
 244         return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
 245 }
 246
 247 /**
 248  * nfs_direct_IO - NFS address space operation for direct I/O
 249  * @iocb: target I/O control block
 250  * @iov: array of vectors that define I/O buffer
 251  * @pos: offset in file to begin the operation
 252  * @nr_segs: size of iovec array
 253  *
 254  * The presence of this routine in the address space ops vector means
 255  * the NFS client supports direct I/O. However, for most direct IO, we
 256  * shunt off direct read and write requests before the VFS gets them,
 257  * so this method is only ever called for swap.
 258  */
 259 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter, loff_t pos)
 260 {
 261         struct inode *inode = iocb->ki_filp->f_mapping->host;
 262
 263         /* we only support swap file calling nfs_direct_IO */
 264         if (!IS_SWAPFILE(inode))
 265                 return 0;
 266
 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 }
 273
 274 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 275 {
 276         unsigned int i;
 277         for (i = 0; i < npages; i++)
 278                 page_cache_release(pages[i]);
 279 }
 280
 281 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 282                               struct nfs_direct_req *dreq)
 283 {
 284         cinfo->lock = &dreq->inode->i_lock;
 285         cinfo->mds = &dreq->mds_cinfo;
 286         cinfo->ds = &dreq->ds_cinfo;
 287         cinfo->dreq = dreq;
 288         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 289 }
 290
 291 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 292                                              struct nfs_pageio_descriptor *pgio,
 293                                              struct nfs_page *req)
 294 {
 295         int mirror_count = 1;
 296
 297         if (pgio->pg_ops->pg_get_mirror_count)
 298                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 299
 300         dreq->mirror_count = mirror_count;
 301 }
 302
 303 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 304 {
 305         struct nfs_direct_req *dreq;
 306
 307         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 308         if (!dreq)
 309                 return NULL;
 310
 311         kref_init(&dreq->kref);
 312         kref_get(&dreq->kref);
 313         init_completion(&dreq->completion);
 314         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 315         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 316         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 317         dreq->mirror_count = 1;
 318         spin_lock_init(&dreq->lock);
 319
 320         return dreq;
 321 }
 322
 323 static void nfs_direct_req_free(struct kref *kref)
 324 {
 325         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 326
 327         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 328         if (dreq->l_ctx != NULL)
 329                 nfs_put_lock_context(dreq->l_ctx);
 330         if (dreq->ctx != NULL)
 331                 put_nfs_open_context(dreq->ctx);
 332         kmem_cache_free(nfs_direct_cachep, dreq);
 333 }
 334
 335 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 336 {
 337         kref_put(&dreq->kref, nfs_direct_req_free);
 338 }
 339
 340 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 341 {
 342         return dreq->bytes_left;
 343 }
 344 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 345
 346 /*
 347  * Collects and returns the final error value/byte-count.
 348  */
 349 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 350 {
 351         ssize_t result = -EIOCBQUEUED;
 352
 353         /* Async requests don't wait here */
 354         if (dreq->iocb)
 355                 goto out;
 356
 357         result = wait_for_completion_killable(&dreq->completion);
 358
 359         if (!result)
 360                 result = dreq->error;
 361         if (!result)
 362                 result = dreq->count;
 363
 364 out:
 365         return (ssize_t) result;
 366 }
 367
 368 /*
 369  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 370  * the iocb is still valid here if this is a synchronous request.
 371  */
 372 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 373 {
 374         struct inode *inode = dreq->inode;
 375
 376         if (dreq->iocb && write) {
 377                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 378
 379                 spin_lock(&inode->i_lock);
 380                 if (i_size_read(inode) < pos)
 381                         i_size_write(inode, pos);
 382                 spin_unlock(&inode->i_lock);
 383         }
 384
 385         if (write)
 386                 nfs_zap_mapping(inode, inode->i_mapping);
 387
 388         inode_dio_end(inode);
 389
 390         if (dreq->iocb) {
 391                 long res = (long) dreq->error;
 392                 if (!res)
 393                         res = (long) dreq->count;
 394                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 395         }
 396
 397         complete_all(&dreq->completion);
 398
 399         nfs_direct_req_release(dreq);
 400 }
 401
 402 static void nfs_direct_readpage_release(struct nfs_page *req)
 403 {
 404         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 405                 d_inode(req->wb_context->dentry)->i_sb->s_id,
 406                 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
 407                 req->wb_bytes,
 408                 (long long)req_offset(req));
 409         nfs_release_request(req);
 410 }
 411
 412 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 413 {
 414         unsigned long bytes = 0;
 415         struct nfs_direct_req *dreq = hdr->dreq;
 416
 417         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 418                 goto out_put;
 419
 420         spin_lock(&dreq->lock);
 421         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 422                 dreq->error = hdr->error;
 423         else
 424                 nfs_direct_good_bytes(dreq, hdr);
 425
 426         spin_unlock(&dreq->lock);
 427
 428         while (!list_empty(&hdr->pages)) {
 429                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 430                 struct page *page = req->wb_page;
 431
 432                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 433                         set_page_dirty(page);
 434                 bytes += req->wb_bytes;
 435                 nfs_list_remove_request(req);
 436                 nfs_direct_readpage_release(req);
 437         }
 438 out_put:
 439         if (put_dreq(dreq))
 440                 nfs_direct_complete(dreq, false);
 441         hdr->release(hdr);
 442 }
 443
 444 static void nfs_read_sync_pgio_error(struct list_head *head)
 445 {
 446         struct nfs_page *req;
 447
 448         while (!list_empty(head)) {
 449                 req = nfs_list_entry(head->next);
 450                 nfs_list_remove_request(req);
 451                 nfs_release_request(req);
 452         }
 453 }
 454
 455 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 456 {
 457         get_dreq(hdr->dreq);
 458 }
 459
 460 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 461         .error_cleanup = nfs_read_sync_pgio_error,
 462         .init_hdr = nfs_direct_pgio_init,
 463         .completion = nfs_direct_read_completion,
 464 };
 465
 466 /*
 467  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 468  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 469  * bail and stop sending more reads.  Read length accounting is
 470  * handled automatically by nfs_direct_read_result().  Otherwise, if
 471  * no requests have been sent, just return an error.
 472  */
 473
 474 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 475                                               struct iov_iter *iter,
 476                                               loff_t pos)
 477 {
 478         struct nfs_pageio_descriptor desc;
 479         struct inode *inode = dreq->inode;
 480         ssize_t result = -EINVAL;
 481         size_t requested_bytes = 0;
 482         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 483
 484         nfs_pageio_init_read(&desc, dreq->inode, false,
 485                              &nfs_direct_read_completion_ops);
 486         get_dreq(dreq);
 487         desc.pg_dreq = dreq;
 488         inode_dio_begin(inode);
 489
 490         while (iov_iter_count(iter)) {
 491                 struct page **pagevec;
 492                 size_t bytes;
 493                 size_t pgbase;
 494                 unsigned npages, i;
 495
 496                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 497                                                   rsize, &pgbase);
 498                 if (result < 0)
 499                         break;
 500
 501                 bytes = result;
 502                 iov_iter_advance(iter, bytes);
 503                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 504                 for (i = 0; i < npages; i++) {
 505                         struct nfs_page *req;
 506                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 507                         /* XXX do we need to do the eof zeroing found in async_filler? */
 508                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 509                                                  pgbase, req_len);
 510                         if (IS_ERR(req)) {
 511                                 result = PTR_ERR(req);
 512                                 break;
 513                         }
 514                         req->wb_index = pos >> PAGE_SHIFT;
 515                         req->wb_offset = pos & ~PAGE_MASK;
 516                         if (!nfs_pageio_add_request(&desc, req)) {
 517                                 result = desc.pg_error;
 518                                 nfs_release_request(req);
 519                                 break;
 520                         }
 521                         pgbase = 0;
 522                         bytes -= req_len;
 523                         requested_bytes += req_len;
 524                         pos += req_len;
 525                         dreq->bytes_left -= req_len;
 526                 }
 527                 nfs_direct_release_pages(pagevec, npages);
 528                 kvfree(pagevec);
 529                 if (result < 0)
 530                         break;
 531         }
 532
 533         nfs_pageio_complete(&desc);
 534
 535         /*
 536          * If no bytes were started, return the error, and let the
 537          * generic layer handle the completion.
 538          */
 539         if (requested_bytes == 0) {
 540                 inode_dio_end(inode);
 541                 nfs_direct_req_release(dreq);
 542                 return result < 0 ? result : -EIO;
 543         }
 544
 545         if (put_dreq(dreq))
 546                 nfs_direct_complete(dreq, false);
 547         return 0;
 548 }
 549
 550 /**
 551  * nfs_file_direct_read - file direct read operation for NFS files
 552  * @iocb: target I/O control block
 553  * @iter: vector of user buffers into which to read data
 554  * @pos: byte offset in file where reading starts
 555  *
 556  * We use this function for direct reads instead of calling
 557  * generic_file_aio_read() in order to avoid gfar's check to see if
 558  * the request starts before the end of the file.  For that check
 559  * to work, we must generate a GETATTR before each direct read, and
 560  * even then there is a window between the GETATTR and the subsequent
 561  * READ where the file size could change.  Our preference is simply
 562  * to do all reads the application wants, and the server will take
 563  * care of managing the end of file boundary.
 564  *
 565  * This function also eliminates unnecessarily updating the file's
 566  * atime locally, as the NFS server sets the file's atime, and this
 567  * client must read the updated atime from the server back into its
 568  * cache.
 569  */
 570 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 571                                 loff_t pos)
 572 {
 573         struct file *file = iocb->ki_filp;
 574         struct address_space *mapping = file->f_mapping;
 575         struct inode *inode = mapping->host;
 576         struct nfs_direct_req *dreq;
 577         struct nfs_lock_context *l_ctx;
 578         ssize_t result = -EINVAL;
 579         size_t count = iov_iter_count(iter);
 580         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 581
 582         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 583                 file, count, (long long) pos);
 584
 585         result = 0;
 586         if (!count)
 587                 goto out;
 588
 589         mutex_lock(&inode->i_mutex);
 590         result = nfs_sync_mapping(mapping);
 591         if (result)
 592                 goto out_unlock;
 593
 594         task_io_account_read(count);
 595
 596         result = -ENOMEM;
 597         dreq = nfs_direct_req_alloc();
 598         if (dreq == NULL)
 599                 goto out_unlock;
 600
 601         dreq->inode = inode;
 602         dreq->bytes_left = count;
 603         dreq->io_start = pos;
 604         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 605         l_ctx = nfs_get_lock_context(dreq->ctx);
 606         if (IS_ERR(l_ctx)) {
 607                 result = PTR_ERR(l_ctx);
 608                 goto out_release;
 609         }
 610         dreq->l_ctx = l_ctx;
 611         if (!is_sync_kiocb(iocb))
 612                 dreq->iocb = iocb;
 613
 614         NFS_I(inode)->read_io += count;
 615         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 616
 617         mutex_unlock(&inode->i_mutex);
 618
 619         if (!result) {
 620                 result = nfs_direct_wait(dreq);
 621                 if (result > 0)
 622                         iocb->ki_pos = pos + result;
 623         }
 624
 625         nfs_direct_req_release(dreq);
 626         return result;
 627
 628 out_release:
 629         nfs_direct_req_release(dreq);
 630 out_unlock:
 631         mutex_unlock(&inode->i_mutex);
 632 out:
 633         return result;
 634 }
 635
 636 static void
 637 nfs_direct_write_scan_commit_list(struct inode *inode,
 638                                   struct list_head *list,
 639                                   struct nfs_commit_info *cinfo)
 640 {
 641         spin_lock(cinfo->lock);
 642 #ifdef CONFIG_NFS_V4_1
 643         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 644                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 645 #endif
 646         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 647         spin_unlock(cinfo->lock);
 648 }
 649
 650 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 651 {
 652         struct nfs_pageio_descriptor desc;
 653         struct nfs_page *req, *tmp;
 654         LIST_HEAD(reqs);
 655         struct nfs_commit_info cinfo;
 656         LIST_HEAD(failed);
 657         int i;
 658
 659         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 660         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 661
 662         dreq->count = 0;
 663         for (i = 0; i < dreq->mirror_count; i++)
 664                 dreq->mirrors[i].count = 0;
 665         get_dreq(dreq);
 666
 667         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 668                               &nfs_direct_write_completion_ops);
 669         desc.pg_dreq = dreq;
 670
 671         req = nfs_list_entry(reqs.next);
 672         nfs_direct_setup_mirroring(dreq, &desc, req);
 673
 674         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 675                 if (!nfs_pageio_add_request(&desc, req)) {
 676                         nfs_list_remove_request(req);
 677                         nfs_list_add_request(req, &failed);
 678                         spin_lock(cinfo.lock);
 679                         dreq->flags = 0;
 680                         dreq->error = -EIO;
 681                         spin_unlock(cinfo.lock);
 682                 }
 683                 nfs_release_request(req);
 684         }
 685         nfs_pageio_complete(&desc);
 686
 687         while (!list_empty(&failed)) {
 688                 req = nfs_list_entry(failed.next);
 689                 nfs_list_remove_request(req);
 690                 nfs_unlock_and_release_request(req);
 691         }
 692
 693         if (put_dreq(dreq))
 694                 nfs_direct_write_complete(dreq, dreq->inode);
 695 }
 696
 697 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 698 {
 699         struct nfs_direct_req *dreq = data->dreq;
 700         struct nfs_commit_info cinfo;
 701         struct nfs_page *req;
 702         int status = data->task.tk_status;
 703
 704         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 705         if (status < 0) {
 706                 dprintk("NFS: %5u commit failed with error %d.\n",
 707                         data->task.tk_pid, status);
 708                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 709         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 710                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 711                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 712         }
 713
 714         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 715         while (!list_empty(&data->pages)) {
 716                 req = nfs_list_entry(data->pages.next);
 717                 nfs_list_remove_request(req);
 718                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 719                         /* Note the rewrite will go through mds */
 720                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 721                 } else
 722                         nfs_release_request(req);
 723                 nfs_unlock_and_release_request(req);
 724         }
 725
 726         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 727                 nfs_direct_write_complete(dreq, data->inode);
 728 }
 729
 730 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 731 {
 732         /* There is no lock to clear */
 733 }
 734
 735 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 736         .completion = nfs_direct_commit_complete,
 737         .error_cleanup = nfs_direct_error_cleanup,
 738 };
 739
 740 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 741 {
 742         int res;
 743         struct nfs_commit_info cinfo;
 744         LIST_HEAD(mds_list);
 745
 746         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 747         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 748         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 749         if (res < 0) /* res == -ENOMEM */
 750                 nfs_direct_write_reschedule(dreq);
 751 }
 752
 753 static void nfs_direct_write_schedule_work(struct work_struct *work)
 754 {
 755         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 756         int flags = dreq->flags;
 757
 758         dreq->flags = 0;
 759         switch (flags) {
 760                 case NFS_ODIRECT_DO_COMMIT:
 761                         nfs_direct_commit_schedule(dreq);
 762                         break;
 763                 case NFS_ODIRECT_RESCHED_WRITES:
 764                         nfs_direct_write_reschedule(dreq);
 765                         break;
 766                 default:
 767                         nfs_direct_complete(dreq, true);
 768         }
 769 }
 770
 771 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 772 {
 773         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 774 }
 775
 776 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 777 {
 778         struct nfs_direct_req *dreq = hdr->dreq;
 779         struct nfs_commit_info cinfo;
 780         bool request_commit = false;
 781         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 782
 783         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 784                 goto out_put;
 785
 786         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 787
 788         spin_lock(&dreq->lock);
 789
 790         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 791                 dreq->flags = 0;
 792                 dreq->error = hdr->error;
 793         }
 794         if (dreq->error == 0) {
 795                 nfs_direct_good_bytes(dreq, hdr);
 796                 if (nfs_write_need_commit(hdr)) {
 797                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 798                                 request_commit = true;
 799                         else if (dreq->flags == 0) {
 800                                 nfs_direct_set_hdr_verf(dreq, hdr);
 801                                 request_commit = true;
 802                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 803                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 804                                 request_commit = true;
 805                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 806                                         dreq->flags =
 807                                                 NFS_ODIRECT_RESCHED_WRITES;
 808                         }
 809                 }
 810         }
 811         spin_unlock(&dreq->lock);
 812
 813         while (!list_empty(&hdr->pages)) {
 814
 815                 req = nfs_list_entry(hdr->pages.next);
 816                 nfs_list_remove_request(req);
 817                 if (request_commit) {
 818                         kref_get(&req->wb_kref);
 819                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 820                                 hdr->ds_commit_idx);
 821                 }
 822                 nfs_unlock_and_release_request(req);
 823         }
 824
 825 out_put:
 826         if (put_dreq(dreq))
 827                 nfs_direct_write_complete(dreq, hdr->inode);
 828         hdr->release(hdr);
 829 }
 830
 831 static void nfs_write_sync_pgio_error(struct list_head *head)
 832 {
 833         struct nfs_page *req;
 834
 835         while (!list_empty(head)) {
 836                 req = nfs_list_entry(head->next);
 837                 nfs_list_remove_request(req);
 838                 nfs_unlock_and_release_request(req);
 839         }
 840 }
 841
 842 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 843         .error_cleanup = nfs_write_sync_pgio_error,
 844         .init_hdr = nfs_direct_pgio_init,
 845         .completion = nfs_direct_write_completion,
 846 };
 847
 848
 849 /*
 850  * NB: Return the value of the first error return code.  Subsequent
 851  *     errors after the first one are ignored.
 852  */
 853 /*
 854  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 855  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 856  * bail and stop sending more writes.  Write length accounting is
 857  * handled automatically by nfs_direct_write_result().  Otherwise, if
 858  * no requests have been sent, just return an error.
 859  */
 860 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 861                                                struct iov_iter *iter,
 862                                                loff_t pos)
 863 {
 864         struct nfs_pageio_descriptor desc;
 865         struct inode *inode = dreq->inode;
 866         ssize_t result = 0;
 867         size_t requested_bytes = 0;
 868         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 869
 870         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 871                               &nfs_direct_write_completion_ops);
 872         desc.pg_dreq = dreq;
 873         get_dreq(dreq);
 874         inode_dio_begin(inode);
 875
 876         NFS_I(inode)->write_io += iov_iter_count(iter);
 877         while (iov_iter_count(iter)) {
 878                 struct page **pagevec;
 879                 size_t bytes;
 880                 size_t pgbase;
 881                 unsigned npages, i;
 882
 883                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 884                                                   wsize, &pgbase);
 885                 if (result < 0)
 886                         break;
 887
 888                 bytes = result;
 889                 iov_iter_advance(iter, bytes);
 890                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 891                 for (i = 0; i < npages; i++) {
 892                         struct nfs_page *req;
 893                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 894
 895                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 896                                                  pgbase, req_len);
 897                         if (IS_ERR(req)) {
 898                                 result = PTR_ERR(req);
 899                                 break;
 900                         }
 901
 902                         nfs_direct_setup_mirroring(dreq, &desc, req);
 903
 904                         nfs_lock_request(req);
 905                         req->wb_index = pos >> PAGE_SHIFT;
 906                         req->wb_offset = pos & ~PAGE_MASK;
 907                         if (!nfs_pageio_add_request(&desc, req)) {
 908                                 result = desc.pg_error;
 909                                 nfs_unlock_and_release_request(req);
 910                                 break;
 911                         }
 912                         pgbase = 0;
 913                         bytes -= req_len;
 914                         requested_bytes += req_len;
 915                         pos += req_len;
 916                         dreq->bytes_left -= req_len;
 917                 }
 918                 nfs_direct_release_pages(pagevec, npages);
 919                 kvfree(pagevec);
 920                 if (result < 0)
 921                         break;
 922         }
 923         nfs_pageio_complete(&desc);
 924
 925         /*
 926          * If no bytes were started, return the error, and let the
 927          * generic layer handle the completion.
 928          */
 929         if (requested_bytes == 0) {
 930                 inode_dio_end(inode);
 931                 nfs_direct_req_release(dreq);
 932                 return result < 0 ? result : -EIO;
 933         }
 934
 935         if (put_dreq(dreq))
 936                 nfs_direct_write_complete(dreq, dreq->inode);
 937         return 0;
 938 }
 939
 940 /**
 941  * nfs_file_direct_write - file direct write operation for NFS files
 942  * @iocb: target I/O control block
 943  * @iter: vector of user buffers from which to write data
 944  * @pos: byte offset in file where writing starts
 945  *
 946  * We use this function for direct writes instead of calling
 947  * generic_file_aio_write() in order to avoid taking the inode
 948  * semaphore and updating the i_size.  The NFS server will set
 949  * the new i_size and this client must read the updated size
 950  * back into its cache.  We let the server do generic write
 951  * parameter checking and report problems.
 952  *
 953  * We eliminate local atime updates, see direct read above.
 954  *
 955  * We avoid unnecessary page cache invalidations for normal cached
 956  * readers of this file.
 957  *
 958  * Note that O_APPEND is not supported for NFS direct writes, as there
 959  * is no atomic O_APPEND write facility in the NFS protocol.
 960  */
 961 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 962 {
 963         ssize_t result = -EINVAL;
 964         struct file *file = iocb->ki_filp;
 965         struct address_space *mapping = file->f_mapping;
 966         struct inode *inode = mapping->host;
 967         struct nfs_direct_req *dreq;
 968         struct nfs_lock_context *l_ctx;
 969         loff_t pos, end;
 970
 971         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 972                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
 973
 974         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES,
 975                       iov_iter_count(iter));
 976
 977         pos = iocb->ki_pos;
 978         end = (pos + iov_iter_count(iter) - 1) >> PAGE_CACHE_SHIFT;
 979
 980         mutex_lock(&inode->i_mutex);
 981
 982         result = nfs_sync_mapping(mapping);
 983         if (result)
 984                 goto out_unlock;
 985
 986         if (mapping->nrpages) {
 987                 result = invalidate_inode_pages2_range(mapping,
 988                                         pos >> PAGE_CACHE_SHIFT, end);
 989                 if (result)
 990                         goto out_unlock;
 991         }
 992
 993         task_io_account_write(iov_iter_count(iter));
 994
 995         result = -ENOMEM;
 996         dreq = nfs_direct_req_alloc();
 997         if (!dreq)
 998                 goto out_unlock;
 999
1000         dreq->inode = inode;
1001         dreq->bytes_left = iov_iter_count(iter);
1002         dreq->io_start = pos;
1003         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1004         l_ctx = nfs_get_lock_context(dreq->ctx);
1005         if (IS_ERR(l_ctx)) {
1006                 result = PTR_ERR(l_ctx);
1007                 goto out_release;
1008         }
1009         dreq->l_ctx = l_ctx;
1010         if (!is_sync_kiocb(iocb))
1011                 dreq->iocb = iocb;
1012
1013         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1014
1015         if (mapping->nrpages) {
1016                 invalidate_inode_pages2_range(mapping,
1017                                               pos >> PAGE_CACHE_SHIFT, end);
1018         }
1019
1020         mutex_unlock(&inode->i_mutex);
1021
1022         if (!result) {
1023                 result = nfs_direct_wait(dreq);
1024                 if (result > 0) {
1025                         struct inode *inode = mapping->host;
1026
1027                         iocb->ki_pos = pos + result;
1028                         spin_lock(&inode->i_lock);
1029                         if (i_size_read(inode) < iocb->ki_pos)
1030                                 i_size_write(inode, iocb->ki_pos);
1031                         spin_unlock(&inode->i_lock);
1032                         generic_write_sync(file, pos, result);
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 }