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_req {
  70         struct kref             kref;           /* release manager */
  71
  72         /* I/O parameters */
  73         struct nfs_open_context *ctx;           /* file open context info */
  74         struct nfs_lock_context *l_ctx;         /* Lock context info */
  75         struct kiocb *          iocb;           /* controlling i/o request */
  76         struct inode *          inode;          /* target file of i/o */
  77
  78         /* completion state */
  79         atomic_t                io_count;       /* i/os we're waiting for */
  80         spinlock_t              lock;           /* protect completion state */
  81         ssize_t                 count,          /* bytes actually processed */
  82                                 bytes_left,     /* bytes left to be sent */
  83                                 error;          /* any reported error */
  84         struct completion       completion;     /* wait for i/o completion */
  85
  86         /* commit state */
  87         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  88         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  89         struct work_struct      work;
  90         int                     flags;
  91 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  92 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  93         struct nfs_writeverf    verf;           /* unstable write verifier */
  94 };
  95
  96 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  97 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  98 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  99 static void nfs_direct_write_schedule_work(struct work_struct *work);
 100
 101 static inline void get_dreq(struct nfs_direct_req *dreq)
 102 {
 103         atomic_inc(&dreq->io_count);
 104 }
 105
 106 static inline int put_dreq(struct nfs_direct_req *dreq)
 107 {
 108         return atomic_dec_and_test(&dreq->io_count);
 109 }
 110
 111 /*
 112  * nfs_direct_select_verf - select the right verifier
 113  * @dreq - direct request possibly spanning multiple servers
 114  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 115  * @ds_idx - index of data server in data server list, only valid if ds_clp set
 116  *
 117  * returns the correct verifier to use given the role of the server
 118  */
 119 static struct nfs_writeverf *
 120 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 121                        struct nfs_client *ds_clp,
 122                        int ds_idx)
 123 {
 124         struct nfs_writeverf *verfp = &dreq->verf;
 125
 126 #ifdef CONFIG_NFS_V4_1
 127         if (ds_clp) {
 128                 /* pNFS is in use, use the DS verf */
 129                 if (ds_idx >= 0 && ds_idx < dreq->ds_cinfo.nbuckets)
 130                         verfp = &dreq->ds_cinfo.buckets[ds_idx].direct_verf;
 131                 else
 132                         WARN_ON_ONCE(1);
 133         }
 134 #endif
 135         return verfp;
 136 }
 137
 138
 139 /*
 140  * nfs_direct_set_hdr_verf - set the write/commit verifier
 141  * @dreq - direct request possibly spanning multiple servers
 142  * @hdr - pageio header to validate against previously seen verfs
 143  *
 144  * Set the server's (MDS or DS) "seen" verifier
 145  */
 146 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 147                                     struct nfs_pgio_header *hdr)
 148 {
 149         struct nfs_writeverf *verfp;
 150
 151         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp,
 152                                       hdr->ds_idx);
 153         WARN_ON_ONCE(verfp->committed >= 0);
 154         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 155         WARN_ON_ONCE(verfp->committed < 0);
 156 }
 157
 158 /*
 159  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 160  * @dreq - direct request possibly spanning multiple servers
 161  * @hdr - pageio header to validate against previously seen verf
 162  *
 163  * set the server's "seen" verf if not initialized.
 164  * returns result of comparison between @hdr->verf and the "seen"
 165  * verf of the server used by @hdr (DS or MDS)
 166  */
 167 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 168                                           struct nfs_pgio_header *hdr)
 169 {
 170         struct nfs_writeverf *verfp;
 171
 172         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp,
 173                                          hdr->ds_idx);
 174         if (verfp->committed < 0) {
 175                 nfs_direct_set_hdr_verf(dreq, hdr);
 176                 return 0;
 177         }
 178         return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 179 }
 180
 181 /*
 182  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 183  * @dreq - direct request possibly spanning multiple servers
 184  * @data - commit data to validate against previously seen verf
 185  *
 186  * returns result of comparison between @data->verf and the verf of
 187  * the server used by @data (DS or MDS)
 188  */
 189 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 190                                            struct nfs_commit_data *data)
 191 {
 192         struct nfs_writeverf *verfp;
 193
 194         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 195                                          data->ds_commit_index);
 196         WARN_ON_ONCE(verfp->committed < 0);
 197         return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
 198 }
 199
 200 /**
 201  * nfs_direct_IO - NFS address space operation for direct I/O
 202  * @rw: direction (read or write)
 203  * @iocb: target I/O control block
 204  * @iov: array of vectors that define I/O buffer
 205  * @pos: offset in file to begin the operation
 206  * @nr_segs: size of iovec array
 207  *
 208  * The presence of this routine in the address space ops vector means
 209  * the NFS client supports direct I/O. However, for most direct IO, we
 210  * shunt off direct read and write requests before the VFS gets them,
 211  * so this method is only ever called for swap.
 212  */
 213 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter, loff_t pos)
 214 {
 215 #ifndef CONFIG_NFS_SWAP
 216         dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
 217                         iocb->ki_filp, (long long) pos, iter->nr_segs);
 218
 219         return -EINVAL;
 220 #else
 221         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 222
 223         if (rw == READ)
 224                 return nfs_file_direct_read(iocb, iter, pos);
 225         return nfs_file_direct_write(iocb, iter, pos);
 226 #endif /* CONFIG_NFS_SWAP */
 227 }
 228
 229 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 230 {
 231         unsigned int i;
 232         for (i = 0; i < npages; i++)
 233                 page_cache_release(pages[i]);
 234 }
 235
 236 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 237                               struct nfs_direct_req *dreq)
 238 {
 239         cinfo->lock = &dreq->lock;
 240         cinfo->mds = &dreq->mds_cinfo;
 241         cinfo->ds = &dreq->ds_cinfo;
 242         cinfo->dreq = dreq;
 243         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 244 }
 245
 246 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 247 {
 248         struct nfs_direct_req *dreq;
 249
 250         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 251         if (!dreq)
 252                 return NULL;
 253
 254         kref_init(&dreq->kref);
 255         kref_get(&dreq->kref);
 256         init_completion(&dreq->completion);
 257         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 258         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 259         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 260         spin_lock_init(&dreq->lock);
 261
 262         return dreq;
 263 }
 264
 265 static void nfs_direct_req_free(struct kref *kref)
 266 {
 267         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 268
 269         if (dreq->l_ctx != NULL)
 270                 nfs_put_lock_context(dreq->l_ctx);
 271         if (dreq->ctx != NULL)
 272                 put_nfs_open_context(dreq->ctx);
 273         kmem_cache_free(nfs_direct_cachep, dreq);
 274 }
 275
 276 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 277 {
 278         kref_put(&dreq->kref, nfs_direct_req_free);
 279 }
 280
 281 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 282 {
 283         return dreq->bytes_left;
 284 }
 285 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 286
 287 /*
 288  * Collects and returns the final error value/byte-count.
 289  */
 290 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 291 {
 292         ssize_t result = -EIOCBQUEUED;
 293
 294         /* Async requests don't wait here */
 295         if (dreq->iocb)
 296                 goto out;
 297
 298         result = wait_for_completion_killable(&dreq->completion);
 299
 300         if (!result)
 301                 result = dreq->error;
 302         if (!result)
 303                 result = dreq->count;
 304
 305 out:
 306         return (ssize_t) result;
 307 }
 308
 309 /*
 310  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 311  * the iocb is still valid here if this is a synchronous request.
 312  */
 313 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 314 {
 315         struct inode *inode = dreq->inode;
 316
 317         if (dreq->iocb && write) {
 318                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 319
 320                 spin_lock(&inode->i_lock);
 321                 if (i_size_read(inode) < pos)
 322                         i_size_write(inode, pos);
 323                 spin_unlock(&inode->i_lock);
 324         }
 325
 326         if (write)
 327                 nfs_zap_mapping(inode, inode->i_mapping);
 328
 329         inode_dio_done(inode);
 330
 331         if (dreq->iocb) {
 332                 long res = (long) dreq->error;
 333                 if (!res)
 334                         res = (long) dreq->count;
 335                 aio_complete(dreq->iocb, res, 0);
 336         }
 337
 338         complete_all(&dreq->completion);
 339
 340         nfs_direct_req_release(dreq);
 341 }
 342
 343 static void nfs_direct_readpage_release(struct nfs_page *req)
 344 {
 345         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 346                 req->wb_context->dentry->d_inode->i_sb->s_id,
 347                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 348                 req->wb_bytes,
 349                 (long long)req_offset(req));
 350         nfs_release_request(req);
 351 }
 352
 353 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 354 {
 355         unsigned long bytes = 0;
 356         struct nfs_direct_req *dreq = hdr->dreq;
 357
 358         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 359                 goto out_put;
 360
 361         spin_lock(&dreq->lock);
 362         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 363                 dreq->error = hdr->error;
 364         else
 365                 dreq->count += hdr->good_bytes;
 366         spin_unlock(&dreq->lock);
 367
 368         while (!list_empty(&hdr->pages)) {
 369                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 370                 struct page *page = req->wb_page;
 371
 372                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 373                         set_page_dirty(page);
 374                 bytes += req->wb_bytes;
 375                 nfs_list_remove_request(req);
 376                 nfs_direct_readpage_release(req);
 377         }
 378 out_put:
 379         if (put_dreq(dreq))
 380                 nfs_direct_complete(dreq, false);
 381         hdr->release(hdr);
 382 }
 383
 384 static void nfs_read_sync_pgio_error(struct list_head *head)
 385 {
 386         struct nfs_page *req;
 387
 388         while (!list_empty(head)) {
 389                 req = nfs_list_entry(head->next);
 390                 nfs_list_remove_request(req);
 391                 nfs_release_request(req);
 392         }
 393 }
 394
 395 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 396 {
 397         get_dreq(hdr->dreq);
 398 }
 399
 400 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 401         .error_cleanup = nfs_read_sync_pgio_error,
 402         .init_hdr = nfs_direct_pgio_init,
 403         .completion = nfs_direct_read_completion,
 404 };
 405
 406 /*
 407  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 408  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 409  * bail and stop sending more reads.  Read length accounting is
 410  * handled automatically by nfs_direct_read_result().  Otherwise, if
 411  * no requests have been sent, just return an error.
 412  */
 413
 414 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 415                                               struct iov_iter *iter,
 416                                               loff_t pos)
 417 {
 418         struct nfs_pageio_descriptor desc;
 419         struct inode *inode = dreq->inode;
 420         ssize_t result = -EINVAL;
 421         size_t requested_bytes = 0;
 422         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 423
 424         nfs_pageio_init_read(&desc, dreq->inode, false,
 425                              &nfs_direct_read_completion_ops);
 426         get_dreq(dreq);
 427         desc.pg_dreq = dreq;
 428         atomic_inc(&inode->i_dio_count);
 429
 430         while (iov_iter_count(iter)) {
 431                 struct page **pagevec;
 432                 size_t bytes;
 433                 size_t pgbase;
 434                 unsigned npages, i;
 435
 436                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 437                                                   rsize, &pgbase);
 438                 if (result < 0)
 439                         break;
 440
 441                 bytes = result;
 442                 iov_iter_advance(iter, bytes);
 443                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 444                 for (i = 0; i < npages; i++) {
 445                         struct nfs_page *req;
 446                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 447                         /* XXX do we need to do the eof zeroing found in async_filler? */
 448                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 449                                                  pgbase, req_len);
 450                         if (IS_ERR(req)) {
 451                                 result = PTR_ERR(req);
 452                                 break;
 453                         }
 454                         req->wb_index = pos >> PAGE_SHIFT;
 455                         req->wb_offset = pos & ~PAGE_MASK;
 456                         if (!nfs_pageio_add_request(&desc, req)) {
 457                                 result = desc.pg_error;
 458                                 nfs_release_request(req);
 459                                 break;
 460                         }
 461                         pgbase = 0;
 462                         bytes -= req_len;
 463                         requested_bytes += req_len;
 464                         pos += req_len;
 465                         dreq->bytes_left -= req_len;
 466                 }
 467                 nfs_direct_release_pages(pagevec, npages);
 468                 kvfree(pagevec);
 469                 if (result < 0)
 470                         break;
 471         }
 472
 473         nfs_pageio_complete(&desc);
 474
 475         /*
 476          * If no bytes were started, return the error, and let the
 477          * generic layer handle the completion.
 478          */
 479         if (requested_bytes == 0) {
 480                 inode_dio_done(inode);
 481                 nfs_direct_req_release(dreq);
 482                 return result < 0 ? result : -EIO;
 483         }
 484
 485         if (put_dreq(dreq))
 486                 nfs_direct_complete(dreq, false);
 487         return 0;
 488 }
 489
 490 /**
 491  * nfs_file_direct_read - file direct read operation for NFS files
 492  * @iocb: target I/O control block
 493  * @iter: vector of user buffers into which to read data
 494  * @pos: byte offset in file where reading starts
 495  *
 496  * We use this function for direct reads instead of calling
 497  * generic_file_aio_read() in order to avoid gfar's check to see if
 498  * the request starts before the end of the file.  For that check
 499  * to work, we must generate a GETATTR before each direct read, and
 500  * even then there is a window between the GETATTR and the subsequent
 501  * READ where the file size could change.  Our preference is simply
 502  * to do all reads the application wants, and the server will take
 503  * care of managing the end of file boundary.
 504  *
 505  * This function also eliminates unnecessarily updating the file's
 506  * atime locally, as the NFS server sets the file's atime, and this
 507  * client must read the updated atime from the server back into its
 508  * cache.
 509  */
 510 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 511                                 loff_t pos)
 512 {
 513         struct file *file = iocb->ki_filp;
 514         struct address_space *mapping = file->f_mapping;
 515         struct inode *inode = mapping->host;
 516         struct nfs_direct_req *dreq;
 517         struct nfs_lock_context *l_ctx;
 518         ssize_t result = -EINVAL;
 519         size_t count = iov_iter_count(iter);
 520         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 521
 522         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 523                 file, count, (long long) pos);
 524
 525         result = 0;
 526         if (!count)
 527                 goto out;
 528
 529         mutex_lock(&inode->i_mutex);
 530         result = nfs_sync_mapping(mapping);
 531         if (result)
 532                 goto out_unlock;
 533
 534         task_io_account_read(count);
 535
 536         result = -ENOMEM;
 537         dreq = nfs_direct_req_alloc();
 538         if (dreq == NULL)
 539                 goto out_unlock;
 540
 541         dreq->inode = inode;
 542         dreq->bytes_left = count;
 543         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 544         l_ctx = nfs_get_lock_context(dreq->ctx);
 545         if (IS_ERR(l_ctx)) {
 546                 result = PTR_ERR(l_ctx);
 547                 goto out_release;
 548         }
 549         dreq->l_ctx = l_ctx;
 550         if (!is_sync_kiocb(iocb))
 551                 dreq->iocb = iocb;
 552
 553         NFS_I(inode)->read_io += count;
 554         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 555
 556         mutex_unlock(&inode->i_mutex);
 557
 558         if (!result) {
 559                 result = nfs_direct_wait(dreq);
 560                 if (result > 0)
 561                         iocb->ki_pos = pos + result;
 562         }
 563
 564         nfs_direct_req_release(dreq);
 565         return result;
 566
 567 out_release:
 568         nfs_direct_req_release(dreq);
 569 out_unlock:
 570         mutex_unlock(&inode->i_mutex);
 571 out:
 572         return result;
 573 }
 574
 575 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 576 {
 577         struct nfs_pageio_descriptor desc;
 578         struct nfs_page *req, *tmp;
 579         LIST_HEAD(reqs);
 580         struct nfs_commit_info cinfo;
 581         LIST_HEAD(failed);
 582
 583         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 584         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 585         spin_lock(cinfo.lock);
 586         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 587         spin_unlock(cinfo.lock);
 588
 589         dreq->count = 0;
 590         get_dreq(dreq);
 591
 592         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 593                               &nfs_direct_write_completion_ops);
 594         desc.pg_dreq = dreq;
 595
 596         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 597                 if (!nfs_pageio_add_request(&desc, req)) {
 598                         nfs_list_remove_request(req);
 599                         nfs_list_add_request(req, &failed);
 600                         spin_lock(cinfo.lock);
 601                         dreq->flags = 0;
 602                         dreq->error = -EIO;
 603                         spin_unlock(cinfo.lock);
 604                 }
 605                 nfs_release_request(req);
 606         }
 607         nfs_pageio_complete(&desc);
 608
 609         while (!list_empty(&failed)) {
 610                 req = nfs_list_entry(failed.next);
 611                 nfs_list_remove_request(req);
 612                 nfs_unlock_and_release_request(req);
 613         }
 614
 615         if (put_dreq(dreq))
 616                 nfs_direct_write_complete(dreq, dreq->inode);
 617 }
 618
 619 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 620 {
 621         struct nfs_direct_req *dreq = data->dreq;
 622         struct nfs_commit_info cinfo;
 623         struct nfs_page *req;
 624         int status = data->task.tk_status;
 625
 626         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 627         if (status < 0) {
 628                 dprintk("NFS: %5u commit failed with error %d.\n",
 629                         data->task.tk_pid, status);
 630                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 631         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 632                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 633                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 634         }
 635
 636         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 637         while (!list_empty(&data->pages)) {
 638                 req = nfs_list_entry(data->pages.next);
 639                 nfs_list_remove_request(req);
 640                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 641                         /* Note the rewrite will go through mds */
 642                         nfs_mark_request_commit(req, NULL, &cinfo);
 643                 } else
 644                         nfs_release_request(req);
 645                 nfs_unlock_and_release_request(req);
 646         }
 647
 648         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 649                 nfs_direct_write_complete(dreq, data->inode);
 650 }
 651
 652 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 653 {
 654         /* There is no lock to clear */
 655 }
 656
 657 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 658         .completion = nfs_direct_commit_complete,
 659         .error_cleanup = nfs_direct_error_cleanup,
 660 };
 661
 662 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 663 {
 664         int res;
 665         struct nfs_commit_info cinfo;
 666         LIST_HEAD(mds_list);
 667
 668         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 669         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 670         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 671         if (res < 0) /* res == -ENOMEM */
 672                 nfs_direct_write_reschedule(dreq);
 673 }
 674
 675 static void nfs_direct_write_schedule_work(struct work_struct *work)
 676 {
 677         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 678         int flags = dreq->flags;
 679
 680         dreq->flags = 0;
 681         switch (flags) {
 682                 case NFS_ODIRECT_DO_COMMIT:
 683                         nfs_direct_commit_schedule(dreq);
 684                         break;
 685                 case NFS_ODIRECT_RESCHED_WRITES:
 686                         nfs_direct_write_reschedule(dreq);
 687                         break;
 688                 default:
 689                         nfs_direct_complete(dreq, true);
 690         }
 691 }
 692
 693 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 694 {
 695         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 696 }
 697
 698 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 699 {
 700         struct nfs_direct_req *dreq = hdr->dreq;
 701         struct nfs_commit_info cinfo;
 702         bool request_commit = false;
 703         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 704
 705         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 706                 goto out_put;
 707
 708         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 709
 710         spin_lock(&dreq->lock);
 711
 712         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 713                 dreq->flags = 0;
 714                 dreq->error = hdr->error;
 715         }
 716         if (dreq->error == 0) {
 717                 dreq->count += hdr->good_bytes;
 718                 if (nfs_write_need_commit(hdr)) {
 719                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 720                                 request_commit = true;
 721                         else if (dreq->flags == 0) {
 722                                 nfs_direct_set_hdr_verf(dreq, hdr);
 723                                 request_commit = true;
 724                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 725                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 726                                 request_commit = true;
 727                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 728                                         dreq->flags =
 729                                                 NFS_ODIRECT_RESCHED_WRITES;
 730                         }
 731                 }
 732         }
 733         spin_unlock(&dreq->lock);
 734
 735         while (!list_empty(&hdr->pages)) {
 736
 737                 req = nfs_list_entry(hdr->pages.next);
 738                 nfs_list_remove_request(req);
 739                 if (request_commit) {
 740                         kref_get(&req->wb_kref);
 741                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 742                 }
 743                 nfs_unlock_and_release_request(req);
 744         }
 745
 746 out_put:
 747         if (put_dreq(dreq))
 748                 nfs_direct_write_complete(dreq, hdr->inode);
 749         hdr->release(hdr);
 750 }
 751
 752 static void nfs_write_sync_pgio_error(struct list_head *head)
 753 {
 754         struct nfs_page *req;
 755
 756         while (!list_empty(head)) {
 757                 req = nfs_list_entry(head->next);
 758                 nfs_list_remove_request(req);
 759                 nfs_unlock_and_release_request(req);
 760         }
 761 }
 762
 763 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 764         .error_cleanup = nfs_write_sync_pgio_error,
 765         .init_hdr = nfs_direct_pgio_init,
 766         .completion = nfs_direct_write_completion,
 767 };
 768
 769
 770 /*
 771  * NB: Return the value of the first error return code.  Subsequent
 772  *     errors after the first one are ignored.
 773  */
 774 /*
 775  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 776  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 777  * bail and stop sending more writes.  Write length accounting is
 778  * handled automatically by nfs_direct_write_result().  Otherwise, if
 779  * no requests have been sent, just return an error.
 780  */
 781 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 782                                                struct iov_iter *iter,
 783                                                loff_t pos)
 784 {
 785         struct nfs_pageio_descriptor desc;
 786         struct inode *inode = dreq->inode;
 787         ssize_t result = 0;
 788         size_t requested_bytes = 0;
 789         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 790
 791         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 792                               &nfs_direct_write_completion_ops);
 793         desc.pg_dreq = dreq;
 794         get_dreq(dreq);
 795         atomic_inc(&inode->i_dio_count);
 796
 797         NFS_I(inode)->write_io += iov_iter_count(iter);
 798         while (iov_iter_count(iter)) {
 799                 struct page **pagevec;
 800                 size_t bytes;
 801                 size_t pgbase;
 802                 unsigned npages, i;
 803
 804                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 805                                                   wsize, &pgbase);
 806                 if (result < 0)
 807                         break;
 808
 809                 bytes = result;
 810                 iov_iter_advance(iter, bytes);
 811                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 812                 for (i = 0; i < npages; i++) {
 813                         struct nfs_page *req;
 814                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 815
 816                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 817                                                  pgbase, req_len);
 818                         if (IS_ERR(req)) {
 819                                 result = PTR_ERR(req);
 820                                 break;
 821                         }
 822                         nfs_lock_request(req);
 823                         req->wb_index = pos >> PAGE_SHIFT;
 824                         req->wb_offset = pos & ~PAGE_MASK;
 825                         if (!nfs_pageio_add_request(&desc, req)) {
 826                                 result = desc.pg_error;
 827                                 nfs_unlock_and_release_request(req);
 828                                 break;
 829                         }
 830                         pgbase = 0;
 831                         bytes -= req_len;
 832                         requested_bytes += req_len;
 833                         pos += req_len;
 834                         dreq->bytes_left -= req_len;
 835                 }
 836                 nfs_direct_release_pages(pagevec, npages);
 837                 kvfree(pagevec);
 838                 if (result < 0)
 839                         break;
 840         }
 841         nfs_pageio_complete(&desc);
 842
 843         /*
 844          * If no bytes were started, return the error, and let the
 845          * generic layer handle the completion.
 846          */
 847         if (requested_bytes == 0) {
 848                 inode_dio_done(inode);
 849                 nfs_direct_req_release(dreq);
 850                 return result < 0 ? result : -EIO;
 851         }
 852
 853         if (put_dreq(dreq))
 854                 nfs_direct_write_complete(dreq, dreq->inode);
 855         return 0;
 856 }
 857
 858 /**
 859  * nfs_file_direct_write - file direct write operation for NFS files
 860  * @iocb: target I/O control block
 861  * @iter: vector of user buffers from which to write data
 862  * @pos: byte offset in file where writing starts
 863  *
 864  * We use this function for direct writes instead of calling
 865  * generic_file_aio_write() in order to avoid taking the inode
 866  * semaphore and updating the i_size.  The NFS server will set
 867  * the new i_size and this client must read the updated size
 868  * back into its cache.  We let the server do generic write
 869  * parameter checking and report problems.
 870  *
 871  * We eliminate local atime updates, see direct read above.
 872  *
 873  * We avoid unnecessary page cache invalidations for normal cached
 874  * readers of this file.
 875  *
 876  * Note that O_APPEND is not supported for NFS direct writes, as there
 877  * is no atomic O_APPEND write facility in the NFS protocol.
 878  */
 879 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 880                                 loff_t pos)
 881 {
 882         ssize_t result = -EINVAL;
 883         struct file *file = iocb->ki_filp;
 884         struct address_space *mapping = file->f_mapping;
 885         struct inode *inode = mapping->host;
 886         struct nfs_direct_req *dreq;
 887         struct nfs_lock_context *l_ctx;
 888         loff_t end;
 889         size_t count = iov_iter_count(iter);
 890         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 891
 892         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 893
 894         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 895                 file, count, (long long) pos);
 896
 897         result = generic_write_checks(file, &pos, &count, 0);
 898         if (result)
 899                 goto out;
 900
 901         result = -EINVAL;
 902         if ((ssize_t) count < 0)
 903                 goto out;
 904         result = 0;
 905         if (!count)
 906                 goto out;
 907
 908         mutex_lock(&inode->i_mutex);
 909
 910         result = nfs_sync_mapping(mapping);
 911         if (result)
 912                 goto out_unlock;
 913
 914         if (mapping->nrpages) {
 915                 result = invalidate_inode_pages2_range(mapping,
 916                                         pos >> PAGE_CACHE_SHIFT, end);
 917                 if (result)
 918                         goto out_unlock;
 919         }
 920
 921         task_io_account_write(count);
 922
 923         result = -ENOMEM;
 924         dreq = nfs_direct_req_alloc();
 925         if (!dreq)
 926                 goto out_unlock;
 927
 928         dreq->inode = inode;
 929         dreq->bytes_left = count;
 930         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 931         l_ctx = nfs_get_lock_context(dreq->ctx);
 932         if (IS_ERR(l_ctx)) {
 933                 result = PTR_ERR(l_ctx);
 934                 goto out_release;
 935         }
 936         dreq->l_ctx = l_ctx;
 937         if (!is_sync_kiocb(iocb))
 938                 dreq->iocb = iocb;
 939
 940         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
 941
 942         if (mapping->nrpages) {
 943                 invalidate_inode_pages2_range(mapping,
 944                                               pos >> PAGE_CACHE_SHIFT, end);
 945         }
 946
 947         mutex_unlock(&inode->i_mutex);
 948
 949         if (!result) {
 950                 result = nfs_direct_wait(dreq);
 951                 if (result > 0) {
 952                         struct inode *inode = mapping->host;
 953
 954                         iocb->ki_pos = pos + result;
 955                         spin_lock(&inode->i_lock);
 956                         if (i_size_read(inode) < iocb->ki_pos)
 957                                 i_size_write(inode, iocb->ki_pos);
 958                         spin_unlock(&inode->i_lock);
 959                 }
 960         }
 961         nfs_direct_req_release(dreq);
 962         return result;
 963
 964 out_release:
 965         nfs_direct_req_release(dreq);
 966 out_unlock:
 967         mutex_unlock(&inode->i_mutex);
 968 out:
 969         return result;
 970 }
 971
 972 /**
 973  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 974  *
 975  */
 976 int __init nfs_init_directcache(void)
 977 {
 978         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 979                                                 sizeof(struct nfs_direct_req),
 980                                                 0, (SLAB_RECLAIM_ACCOUNT|
 981                                                         SLAB_MEM_SPREAD),
 982                                                 NULL);
 983         if (nfs_direct_cachep == NULL)
 984                 return -ENOMEM;
 985
 986         return 0;
 987 }
 988
 989 /**
 990  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 991  *
 992  */
 993 void nfs_destroy_directcache(void)
 994 {
 995         kmem_cache_destroy(nfs_direct_cachep);
 996 }