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                                 max_count,      /* max expected count */
  91                                 bytes_left,     /* bytes left to be sent */
  92                                 io_start,       /* start of IO */
  93                                 error;          /* any reported error */
  94         struct completion       completion;     /* wait for i/o completion */
  95
  96         /* commit state */
  97         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  98         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  99         struct work_struct      work;
 100         int                     flags;
 101 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
 102 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
 103         struct nfs_writeverf    verf;           /* unstable write verifier */
 104 };
 105
 106 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 107 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 108 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
 109 static void nfs_direct_write_schedule_work(struct work_struct *work);
 110
 111 static inline void get_dreq(struct nfs_direct_req *dreq)
 112 {
 113         atomic_inc(&dreq->io_count);
 114 }
 115
 116 static inline int put_dreq(struct nfs_direct_req *dreq)
 117 {
 118         return atomic_dec_and_test(&dreq->io_count);
 119 }
 120
 121 static void
 122 nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
 123 {
 124         int i;
 125         ssize_t count;
 126
 127         WARN_ON_ONCE(dreq->count >= dreq->max_count);
 128
 129         if (dreq->mirror_count == 1) {
 130                 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
 131                 dreq->count += hdr->good_bytes;
 132         } else {
 133                 /* mirrored writes */
 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                 /* update the dreq->count by finding the minimum agreed count from all
 140                  * mirrors */
 141                 count = dreq->mirrors[0].count;
 142
 143                 for (i = 1; i < dreq->mirror_count; i++)
 144                         count = min(count, dreq->mirrors[i].count);
 145
 146                 dreq->count = count;
 147         }
 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         /*
 167          * pNFS is in use, use the DS verf except commit_through_mds is set
 168          * for layout segment where nbuckets is zero.
 169          */
 170         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 171                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 172                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 173                 else
 174                         WARN_ON_ONCE(1);
 175         }
 176 #endif
 177         return verfp;
 178 }
 179
 180
 181 /*
 182  * nfs_direct_set_hdr_verf - set the write/commit verifier
 183  * @dreq - direct request possibly spanning multiple servers
 184  * @hdr - pageio header to validate against previously seen verfs
 185  *
 186  * Set the server's (MDS or DS) "seen" verifier
 187  */
 188 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 189                                     struct nfs_pgio_header *hdr)
 190 {
 191         struct nfs_writeverf *verfp;
 192
 193         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 194         WARN_ON_ONCE(verfp->committed >= 0);
 195         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 196         WARN_ON_ONCE(verfp->committed < 0);
 197 }
 198
 199 /*
 200  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 201  * @dreq - direct request possibly spanning multiple servers
 202  * @hdr - pageio header to validate against previously seen verf
 203  *
 204  * set the server's "seen" verf if not initialized.
 205  * returns result of comparison between @hdr->verf and the "seen"
 206  * verf of the server used by @hdr (DS or MDS)
 207  */
 208 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 209                                           struct nfs_pgio_header *hdr)
 210 {
 211         struct nfs_writeverf *verfp;
 212
 213         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 214         if (verfp->committed < 0) {
 215                 nfs_direct_set_hdr_verf(dreq, hdr);
 216                 return 0;
 217         }
 218         return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 219 }
 220
 221 /*
 222  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 223  * @dreq - direct request possibly spanning multiple servers
 224  * @data - commit data to validate against previously seen verf
 225  *
 226  * returns result of comparison between @data->verf and the verf of
 227  * the server used by @data (DS or MDS)
 228  */
 229 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 230                                            struct nfs_commit_data *data)
 231 {
 232         struct nfs_writeverf *verfp;
 233
 234         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 235                                          data->ds_commit_index);
 236
 237         /* verifier not set so always fail */
 238         if (verfp->committed < 0)
 239                 return 1;
 240
 241         return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
 242 }
 243
 244 /**
 245  * nfs_direct_IO - NFS address space operation for direct I/O
 246  * @iocb: target I/O control block
 247  * @iov: array of vectors that define I/O buffer
 248  * @pos: offset in file to begin the operation
 249  * @nr_segs: size of iovec array
 250  *
 251  * The presence of this routine in the address space ops vector means
 252  * the NFS client supports direct I/O. However, for most direct IO, we
 253  * shunt off direct read and write requests before the VFS gets them,
 254  * so this method is only ever called for swap.
 255  */
 256 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 257 {
 258         struct inode *inode = iocb->ki_filp->f_mapping->host;
 259
 260         /* we only support swap file calling nfs_direct_IO */
 261         if (!IS_SWAPFILE(inode))
 262                 return 0;
 263
 264         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 265
 266         if (iov_iter_rw(iter) == READ)
 267                 return nfs_file_direct_read(iocb, iter);
 268         return nfs_file_direct_write(iocb, iter);
 269 }
 270
 271 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 272 {
 273         unsigned int i;
 274         for (i = 0; i < npages; i++)
 275                 put_page(pages[i]);
 276 }
 277
 278 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 279                               struct nfs_direct_req *dreq)
 280 {
 281         cinfo->inode = dreq->inode;
 282         cinfo->mds = &dreq->mds_cinfo;
 283         cinfo->ds = &dreq->ds_cinfo;
 284         cinfo->dreq = dreq;
 285         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 286 }
 287
 288 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 289                                              struct nfs_pageio_descriptor *pgio,
 290                                              struct nfs_page *req)
 291 {
 292         int mirror_count = 1;
 293
 294         if (pgio->pg_ops->pg_get_mirror_count)
 295                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 296
 297         dreq->mirror_count = mirror_count;
 298 }
 299
 300 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 301 {
 302         struct nfs_direct_req *dreq;
 303
 304         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 305         if (!dreq)
 306                 return NULL;
 307
 308         kref_init(&dreq->kref);
 309         kref_get(&dreq->kref);
 310         init_completion(&dreq->completion);
 311         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 312         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 313         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 314         dreq->mirror_count = 1;
 315         spin_lock_init(&dreq->lock);
 316
 317         return dreq;
 318 }
 319
 320 static void nfs_direct_req_free(struct kref *kref)
 321 {
 322         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 323
 324         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 325         if (dreq->l_ctx != NULL)
 326                 nfs_put_lock_context(dreq->l_ctx);
 327         if (dreq->ctx != NULL)
 328                 put_nfs_open_context(dreq->ctx);
 329         kmem_cache_free(nfs_direct_cachep, dreq);
 330 }
 331
 332 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 333 {
 334         kref_put(&dreq->kref, nfs_direct_req_free);
 335 }
 336
 337 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 338 {
 339         return dreq->bytes_left;
 340 }
 341 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 342
 343 /*
 344  * Collects and returns the final error value/byte-count.
 345  */
 346 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 347 {
 348         ssize_t result = -EIOCBQUEUED;
 349
 350         /* Async requests don't wait here */
 351         if (dreq->iocb)
 352                 goto out;
 353
 354         result = wait_for_completion_killable(&dreq->completion);
 355
 356         if (!result) {
 357                 result = dreq->count;
 358                 WARN_ON_ONCE(dreq->count < 0);
 359         }
 360         if (!result)
 361                 result = dreq->error;
 362
 363 out:
 364         return (ssize_t) result;
 365 }
 366
 367 /*
 368  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 369  * the iocb is still valid here if this is a synchronous request.
 370  */
 371 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 372 {
 373         struct inode *inode = dreq->inode;
 374
 375         if (dreq->iocb && write) {
 376                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 377
 378                 spin_lock(&inode->i_lock);
 379                 if (i_size_read(inode) < pos)
 380                         i_size_write(inode, pos);
 381                 spin_unlock(&inode->i_lock);
 382         }
 383
 384         if (write)
 385                 nfs_zap_mapping(inode, inode->i_mapping);
 386
 387         inode_dio_end(inode);
 388
 389         if (dreq->iocb) {
 390                 long res = (long) dreq->error;
 391                 if (dreq->count != 0) {
 392                         res = (long) dreq->count;
 393                         WARN_ON_ONCE(dreq->count < 0);
 394                 }
 395                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 396         }
 397
 398         complete_all(&dreq->completion);
 399
 400         nfs_direct_req_release(dreq);
 401 }
 402
 403 static void nfs_direct_readpage_release(struct nfs_page *req)
 404 {
 405         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 406                 req->wb_context->dentry->d_sb->s_id,
 407                 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
 408                 req->wb_bytes,
 409                 (long long)req_offset(req));
 410         nfs_release_request(req);
 411 }
 412
 413 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 414 {
 415         unsigned long bytes = 0;
 416         struct nfs_direct_req *dreq = hdr->dreq;
 417
 418         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 419                 goto out_put;
 420
 421         spin_lock(&dreq->lock);
 422         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 423                 dreq->error = hdr->error;
 424         else
 425                 nfs_direct_good_bytes(dreq, hdr);
 426
 427         spin_unlock(&dreq->lock);
 428
 429         while (!list_empty(&hdr->pages)) {
 430                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 431                 struct page *page = req->wb_page;
 432
 433                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 434                         set_page_dirty(page);
 435                 bytes += req->wb_bytes;
 436                 nfs_list_remove_request(req);
 437                 nfs_direct_readpage_release(req);
 438         }
 439 out_put:
 440         if (put_dreq(dreq))
 441                 nfs_direct_complete(dreq, false);
 442         hdr->release(hdr);
 443 }
 444
 445 static void nfs_read_sync_pgio_error(struct list_head *head)
 446 {
 447         struct nfs_page *req;
 448
 449         while (!list_empty(head)) {
 450                 req = nfs_list_entry(head->next);
 451                 nfs_list_remove_request(req);
 452                 nfs_release_request(req);
 453         }
 454 }
 455
 456 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 457 {
 458         get_dreq(hdr->dreq);
 459 }
 460
 461 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 462         .error_cleanup = nfs_read_sync_pgio_error,
 463         .init_hdr = nfs_direct_pgio_init,
 464         .completion = nfs_direct_read_completion,
 465 };
 466
 467 /*
 468  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 469  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 470  * bail and stop sending more reads.  Read length accounting is
 471  * handled automatically by nfs_direct_read_result().  Otherwise, if
 472  * no requests have been sent, just return an error.
 473  */
 474
 475 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 476                                               struct iov_iter *iter,
 477                                               loff_t pos)
 478 {
 479         struct nfs_pageio_descriptor desc;
 480         struct inode *inode = dreq->inode;
 481         ssize_t result = -EINVAL;
 482         size_t requested_bytes = 0;
 483         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 484
 485         nfs_pageio_init_read(&desc, dreq->inode, false,
 486                              &nfs_direct_read_completion_ops);
 487         get_dreq(dreq);
 488         desc.pg_dreq = dreq;
 489         inode_dio_begin(inode);
 490
 491         while (iov_iter_count(iter)) {
 492                 struct page **pagevec;
 493                 size_t bytes;
 494                 size_t pgbase;
 495                 unsigned npages, i;
 496
 497                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 498                                                   rsize, &pgbase);
 499                 if (result < 0)
 500                         break;
 501
 502                 bytes = result;
 503                 iov_iter_advance(iter, bytes);
 504                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 505                 for (i = 0; i < npages; i++) {
 506                         struct nfs_page *req;
 507                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 508                         /* XXX do we need to do the eof zeroing found in async_filler? */
 509                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 510                                                  pgbase, req_len);
 511                         if (IS_ERR(req)) {
 512                                 result = PTR_ERR(req);
 513                                 break;
 514                         }
 515                         req->wb_index = pos >> PAGE_SHIFT;
 516                         req->wb_offset = pos & ~PAGE_MASK;
 517                         if (!nfs_pageio_add_request(&desc, req)) {
 518                                 result = desc.pg_error;
 519                                 nfs_release_request(req);
 520                                 break;
 521                         }
 522                         pgbase = 0;
 523                         bytes -= req_len;
 524                         requested_bytes += req_len;
 525                         pos += req_len;
 526                         dreq->bytes_left -= req_len;
 527                 }
 528                 nfs_direct_release_pages(pagevec, npages);
 529                 kvfree(pagevec);
 530                 if (result < 0)
 531                         break;
 532         }
 533
 534         nfs_pageio_complete(&desc);
 535
 536         /*
 537          * If no bytes were started, return the error, and let the
 538          * generic layer handle the completion.
 539          */
 540         if (requested_bytes == 0) {
 541                 inode_dio_end(inode);
 542                 nfs_direct_req_release(dreq);
 543                 return result < 0 ? result : -EIO;
 544         }
 545
 546         if (put_dreq(dreq))
 547                 nfs_direct_complete(dreq, false);
 548         return 0;
 549 }
 550
 551 /**
 552  * nfs_file_direct_read - file direct read operation for NFS files
 553  * @iocb: target I/O control block
 554  * @iter: vector of user buffers into which to read data
 555  *
 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 {
 572         struct file *file = iocb->ki_filp;
 573         struct address_space *mapping = file->f_mapping;
 574         struct inode *inode = mapping->host;
 575         struct nfs_direct_req *dreq;
 576         struct nfs_lock_context *l_ctx;
 577         ssize_t result = -EINVAL;
 578         size_t count = iov_iter_count(iter);
 579         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 580
 581         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 582                 file, count, (long long) iocb->ki_pos);
 583
 584         result = 0;
 585         if (!count)
 586                 goto out;
 587
 588         inode_lock(inode);
 589         result = nfs_sync_mapping(mapping);
 590         if (result)
 591                 goto out_unlock;
 592
 593         task_io_account_read(count);
 594
 595         result = -ENOMEM;
 596         dreq = nfs_direct_req_alloc();
 597         if (dreq == NULL)
 598                 goto out_unlock;
 599
 600         dreq->inode = inode;
 601         dreq->bytes_left = dreq->max_count = count;
 602         dreq->io_start = iocb->ki_pos;
 603         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 604         l_ctx = nfs_get_lock_context(dreq->ctx);
 605         if (IS_ERR(l_ctx)) {
 606                 result = PTR_ERR(l_ctx);
 607                 goto out_release;
 608         }
 609         dreq->l_ctx = l_ctx;
 610         if (!is_sync_kiocb(iocb))
 611                 dreq->iocb = iocb;
 612
 613         NFS_I(inode)->read_io += count;
 614         result = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 615
 616         inode_unlock(inode);
 617
 618         if (!result) {
 619                 result = nfs_direct_wait(dreq);
 620                 if (result > 0)
 621                         iocb->ki_pos += result;
 622         }
 623
 624         nfs_direct_req_release(dreq);
 625         return result;
 626
 627 out_release:
 628         nfs_direct_req_release(dreq);
 629 out_unlock:
 630         inode_unlock(inode);
 631 out:
 632         return result;
 633 }
 634
 635 static void
 636 nfs_direct_write_scan_commit_list(struct inode *inode,
 637                                   struct list_head *list,
 638                                   struct nfs_commit_info *cinfo)
 639 {
 640         spin_lock(&cinfo->inode->i_lock);
 641 #ifdef CONFIG_NFS_V4_1
 642         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 643                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 644 #endif
 645         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 646         spin_unlock(&cinfo->inode->i_lock);
 647 }
 648
 649 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 650 {
 651         struct nfs_pageio_descriptor desc;
 652         struct nfs_page *req, *tmp;
 653         LIST_HEAD(reqs);
 654         struct nfs_commit_info cinfo;
 655         LIST_HEAD(failed);
 656         int i;
 657
 658         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 659         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 660
 661         dreq->count = 0;
 662         for (i = 0; i < dreq->mirror_count; i++)
 663                 dreq->mirrors[i].count = 0;
 664         get_dreq(dreq);
 665
 666         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 667                               &nfs_direct_write_completion_ops);
 668         desc.pg_dreq = dreq;
 669
 670         req = nfs_list_entry(reqs.next);
 671         nfs_direct_setup_mirroring(dreq, &desc, req);
 672         if (desc.pg_error < 0) {
 673                 list_splice_init(&reqs, &failed);
 674                 goto out_failed;
 675         }
 676
 677         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 678                 if (!nfs_pageio_add_request(&desc, req)) {
 679                         nfs_list_remove_request(req);
 680                         nfs_list_add_request(req, &failed);
 681                         spin_lock(&cinfo.inode->i_lock);
 682                         dreq->flags = 0;
 683                         if (desc.pg_error < 0)
 684                                 dreq->error = desc.pg_error;
 685                         else
 686                                 dreq->error = -EIO;
 687                         spin_unlock(&cinfo.inode->i_lock);
 688                 }
 689                 nfs_release_request(req);
 690         }
 691         nfs_pageio_complete(&desc);
 692
 693 out_failed:
 694         while (!list_empty(&failed)) {
 695                 req = nfs_list_entry(failed.next);
 696                 nfs_list_remove_request(req);
 697                 nfs_unlock_and_release_request(req);
 698         }
 699
 700         if (put_dreq(dreq))
 701                 nfs_direct_write_complete(dreq, dreq->inode);
 702 }
 703
 704 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 705 {
 706         struct nfs_direct_req *dreq = data->dreq;
 707         struct nfs_commit_info cinfo;
 708         struct nfs_page *req;
 709         int status = data->task.tk_status;
 710
 711         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 712         if (status < 0) {
 713                 dprintk("NFS: %5u commit failed with error %d.\n",
 714                         data->task.tk_pid, status);
 715                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 716         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 717                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 718                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 719         }
 720
 721         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 722         while (!list_empty(&data->pages)) {
 723                 req = nfs_list_entry(data->pages.next);
 724                 nfs_list_remove_request(req);
 725                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 726                         /* Note the rewrite will go through mds */
 727                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 728                 } else
 729                         nfs_release_request(req);
 730                 nfs_unlock_and_release_request(req);
 731         }
 732
 733         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 734                 nfs_direct_write_complete(dreq, data->inode);
 735 }
 736
 737 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 738                 struct nfs_page *req)
 739 {
 740         struct nfs_direct_req *dreq = cinfo->dreq;
 741
 742         spin_lock(&dreq->lock);
 743         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 744         spin_unlock(&dreq->lock);
 745         nfs_mark_request_commit(req, NULL, cinfo, 0);
 746 }
 747
 748 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 749         .completion = nfs_direct_commit_complete,
 750         .resched_write = nfs_direct_resched_write,
 751 };
 752
 753 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 754 {
 755         int res;
 756         struct nfs_commit_info cinfo;
 757         LIST_HEAD(mds_list);
 758
 759         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 760         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 761         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 762         if (res < 0) /* res == -ENOMEM */
 763                 nfs_direct_write_reschedule(dreq);
 764 }
 765
 766 static void nfs_direct_write_schedule_work(struct work_struct *work)
 767 {
 768         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 769         int flags = dreq->flags;
 770
 771         dreq->flags = 0;
 772         switch (flags) {
 773                 case NFS_ODIRECT_DO_COMMIT:
 774                         nfs_direct_commit_schedule(dreq);
 775                         break;
 776                 case NFS_ODIRECT_RESCHED_WRITES:
 777                         nfs_direct_write_reschedule(dreq);
 778                         break;
 779                 default:
 780                         nfs_direct_complete(dreq, true);
 781         }
 782 }
 783
 784 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 785 {
 786         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 787 }
 788
 789 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 790 {
 791         struct nfs_direct_req *dreq = hdr->dreq;
 792         struct nfs_commit_info cinfo;
 793         bool request_commit = false;
 794         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 795
 796         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 797                 goto out_put;
 798
 799         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 800
 801         spin_lock(&dreq->lock);
 802
 803         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 804                 dreq->flags = 0;
 805                 dreq->error = hdr->error;
 806         }
 807         if (dreq->error == 0) {
 808                 nfs_direct_good_bytes(dreq, hdr);
 809                 if (nfs_write_need_commit(hdr)) {
 810                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 811                                 request_commit = true;
 812                         else if (dreq->flags == 0) {
 813                                 nfs_direct_set_hdr_verf(dreq, hdr);
 814                                 request_commit = true;
 815                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 816                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 817                                 request_commit = true;
 818                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 819                                         dreq->flags =
 820                                                 NFS_ODIRECT_RESCHED_WRITES;
 821                         }
 822                 }
 823         }
 824         spin_unlock(&dreq->lock);
 825
 826         while (!list_empty(&hdr->pages)) {
 827
 828                 req = nfs_list_entry(hdr->pages.next);
 829                 nfs_list_remove_request(req);
 830                 if (request_commit) {
 831                         kref_get(&req->wb_kref);
 832                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 833                                 hdr->ds_commit_idx);
 834                 }
 835                 nfs_unlock_and_release_request(req);
 836         }
 837
 838 out_put:
 839         if (put_dreq(dreq))
 840                 nfs_direct_write_complete(dreq, hdr->inode);
 841         hdr->release(hdr);
 842 }
 843
 844 static void nfs_write_sync_pgio_error(struct list_head *head)
 845 {
 846         struct nfs_page *req;
 847
 848         while (!list_empty(head)) {
 849                 req = nfs_list_entry(head->next);
 850                 nfs_list_remove_request(req);
 851                 nfs_unlock_and_release_request(req);
 852         }
 853 }
 854
 855 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 856 {
 857         struct nfs_direct_req *dreq = hdr->dreq;
 858
 859         spin_lock(&dreq->lock);
 860         if (dreq->error == 0) {
 861                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 862                 /* fake unstable write to let common nfs resend pages */
 863                 hdr->verf.committed = NFS_UNSTABLE;
 864                 hdr->good_bytes = hdr->args.count;
 865         }
 866         spin_unlock(&dreq->lock);
 867 }
 868
 869 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 870         .error_cleanup = nfs_write_sync_pgio_error,
 871         .init_hdr = nfs_direct_pgio_init,
 872         .completion = nfs_direct_write_completion,
 873         .reschedule_io = nfs_direct_write_reschedule_io,
 874 };
 875
 876
 877 /*
 878  * NB: Return the value of the first error return code.  Subsequent
 879  *     errors after the first one are ignored.
 880  */
 881 /*
 882  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 883  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 884  * bail and stop sending more writes.  Write length accounting is
 885  * handled automatically by nfs_direct_write_result().  Otherwise, if
 886  * no requests have been sent, just return an error.
 887  */
 888 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 889                                                struct iov_iter *iter,
 890                                                loff_t pos)
 891 {
 892         struct nfs_pageio_descriptor desc;
 893         struct inode *inode = dreq->inode;
 894         ssize_t result = 0;
 895         size_t requested_bytes = 0;
 896         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 897
 898         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 899                               &nfs_direct_write_completion_ops);
 900         desc.pg_dreq = dreq;
 901         get_dreq(dreq);
 902         inode_dio_begin(inode);
 903
 904         NFS_I(inode)->write_io += iov_iter_count(iter);
 905         while (iov_iter_count(iter)) {
 906                 struct page **pagevec;
 907                 size_t bytes;
 908                 size_t pgbase;
 909                 unsigned npages, i;
 910
 911                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 912                                                   wsize, &pgbase);
 913                 if (result < 0)
 914                         break;
 915
 916                 bytes = result;
 917                 iov_iter_advance(iter, bytes);
 918                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 919                 for (i = 0; i < npages; i++) {
 920                         struct nfs_page *req;
 921                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 922
 923                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 924                                                  pgbase, req_len);
 925                         if (IS_ERR(req)) {
 926                                 result = PTR_ERR(req);
 927                                 break;
 928                         }
 929
 930                         nfs_direct_setup_mirroring(dreq, &desc, req);
 931                         if (desc.pg_error < 0) {
 932                                 nfs_free_request(req);
 933                                 result = desc.pg_error;
 934                                 break;
 935                         }
 936
 937                         nfs_lock_request(req);
 938                         req->wb_index = pos >> PAGE_SHIFT;
 939                         req->wb_offset = pos & ~PAGE_MASK;
 940                         if (!nfs_pageio_add_request(&desc, req)) {
 941                                 result = desc.pg_error;
 942                                 nfs_unlock_and_release_request(req);
 943                                 break;
 944                         }
 945                         pgbase = 0;
 946                         bytes -= req_len;
 947                         requested_bytes += req_len;
 948                         pos += req_len;
 949                         dreq->bytes_left -= req_len;
 950                 }
 951                 nfs_direct_release_pages(pagevec, npages);
 952                 kvfree(pagevec);
 953                 if (result < 0)
 954                         break;
 955         }
 956         nfs_pageio_complete(&desc);
 957
 958         /*
 959          * If no bytes were started, return the error, and let the
 960          * generic layer handle the completion.
 961          */
 962         if (requested_bytes == 0) {
 963                 inode_dio_end(inode);
 964                 nfs_direct_req_release(dreq);
 965                 return result < 0 ? result : -EIO;
 966         }
 967
 968         if (put_dreq(dreq))
 969                 nfs_direct_write_complete(dreq, dreq->inode);
 970         return 0;
 971 }
 972
 973 /**
 974  * nfs_file_direct_write - file direct write operation for NFS files
 975  * @iocb: target I/O control block
 976  * @iter: vector of user buffers from which to write data
 977  *
 978  * We use this function for direct writes instead of calling
 979  * generic_file_aio_write() in order to avoid taking the inode
 980  * semaphore and updating the i_size.  The NFS server will set
 981  * the new i_size and this client must read the updated size
 982  * back into its cache.  We let the server do generic write
 983  * parameter checking and report problems.
 984  *
 985  * We eliminate local atime updates, see direct read above.
 986  *
 987  * We avoid unnecessary page cache invalidations for normal cached
 988  * readers of this file.
 989  *
 990  * Note that O_APPEND is not supported for NFS direct writes, as there
 991  * is no atomic O_APPEND write facility in the NFS protocol.
 992  */
 993 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 994 {
 995         ssize_t result = -EINVAL;
 996         struct file *file = iocb->ki_filp;
 997         struct address_space *mapping = file->f_mapping;
 998         struct inode *inode = mapping->host;
 999         struct nfs_direct_req *dreq;
1000         struct nfs_lock_context *l_ctx;
1001         loff_t pos, end;
1002
1003         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1004                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
1005
1006         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES,
1007                       iov_iter_count(iter));
1008
1009         pos = iocb->ki_pos;
1010         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1011
1012         inode_lock(inode);
1013
1014         result = nfs_sync_mapping(mapping);
1015         if (result)
1016                 goto out_unlock;
1017
1018         if (mapping->nrpages) {
1019                 result = invalidate_inode_pages2_range(mapping,
1020                                         pos >> PAGE_SHIFT, end);
1021                 if (result)
1022                         goto out_unlock;
1023         }
1024
1025         task_io_account_write(iov_iter_count(iter));
1026
1027         result = -ENOMEM;
1028         dreq = nfs_direct_req_alloc();
1029         if (!dreq)
1030                 goto out_unlock;
1031
1032         dreq->inode = inode;
1033         dreq->bytes_left = dreq->max_count = iov_iter_count(iter);
1034         dreq->io_start = pos;
1035         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1036         l_ctx = nfs_get_lock_context(dreq->ctx);
1037         if (IS_ERR(l_ctx)) {
1038                 result = PTR_ERR(l_ctx);
1039                 goto out_release;
1040         }
1041         dreq->l_ctx = l_ctx;
1042         if (!is_sync_kiocb(iocb))
1043                 dreq->iocb = iocb;
1044
1045         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1046
1047         if (mapping->nrpages) {
1048                 invalidate_inode_pages2_range(mapping,
1049                                               pos >> PAGE_SHIFT, end);
1050         }
1051
1052         inode_unlock(inode);
1053
1054         if (!result) {
1055                 result = nfs_direct_wait(dreq);
1056                 if (result > 0) {
1057                         struct inode *inode = mapping->host;
1058
1059                         iocb->ki_pos = pos + result;
1060                         spin_lock(&inode->i_lock);
1061                         if (i_size_read(inode) < iocb->ki_pos)
1062                                 i_size_write(inode, iocb->ki_pos);
1063                         spin_unlock(&inode->i_lock);
1064
1065                         /* XXX: should check the generic_write_sync retval */
1066                         generic_write_sync(iocb, result);
1067                 }
1068         }
1069         nfs_direct_req_release(dreq);
1070         return result;
1071
1072 out_release:
1073         nfs_direct_req_release(dreq);
1074 out_unlock:
1075         inode_unlock(inode);
1076         return result;
1077 }
1078
1079 /**
1080  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1081  *
1082  */
1083 int __init nfs_init_directcache(void)
1084 {
1085         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1086                                                 sizeof(struct nfs_direct_req),
1087                                                 0, (SLAB_RECLAIM_ACCOUNT|
1088                                                         SLAB_MEM_SPREAD),
1089                                                 NULL);
1090         if (nfs_direct_cachep == NULL)
1091                 return -ENOMEM;
1092
1093         return 0;
1094 }
1095
1096 /**
1097  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1098  *
1099  */
1100 void nfs_destroy_directcache(void)
1101 {
1102         kmem_cache_destroy(nfs_direct_cachep);
1103 }