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         struct inode *inode = iocb->ki_filp->f_mapping->host;
 216
 217         /* we only support swap file calling nfs_direct_IO */
 218         if (!IS_SWAPFILE(inode))
 219                 return 0;
 220
 221 #ifndef CONFIG_NFS_SWAP
 222         dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
 223                         iocb->ki_filp, (long long) pos, iter->nr_segs);
 224
 225         return -EINVAL;
 226 #else
 227         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 228
 229         if (rw == READ)
 230                 return nfs_file_direct_read(iocb, iter, pos);
 231         return nfs_file_direct_write(iocb, iter, pos);
 232 #endif /* CONFIG_NFS_SWAP */
 233 }
 234
 235 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 236 {
 237         unsigned int i;
 238         for (i = 0; i < npages; i++)
 239                 page_cache_release(pages[i]);
 240 }
 241
 242 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 243                               struct nfs_direct_req *dreq)
 244 {
 245         cinfo->lock = &dreq->lock;
 246         cinfo->mds = &dreq->mds_cinfo;
 247         cinfo->ds = &dreq->ds_cinfo;
 248         cinfo->dreq = dreq;
 249         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 250 }
 251
 252 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 253 {
 254         struct nfs_direct_req *dreq;
 255
 256         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 257         if (!dreq)
 258                 return NULL;
 259
 260         kref_init(&dreq->kref);
 261         kref_get(&dreq->kref);
 262         init_completion(&dreq->completion);
 263         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 264         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 265         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 266         spin_lock_init(&dreq->lock);
 267
 268         return dreq;
 269 }
 270
 271 static void nfs_direct_req_free(struct kref *kref)
 272 {
 273         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 274
 275         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 276         if (dreq->l_ctx != NULL)
 277                 nfs_put_lock_context(dreq->l_ctx);
 278         if (dreq->ctx != NULL)
 279                 put_nfs_open_context(dreq->ctx);
 280         kmem_cache_free(nfs_direct_cachep, dreq);
 281 }
 282
 283 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 284 {
 285         kref_put(&dreq->kref, nfs_direct_req_free);
 286 }
 287
 288 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 289 {
 290         return dreq->bytes_left;
 291 }
 292 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 293
 294 /*
 295  * Collects and returns the final error value/byte-count.
 296  */
 297 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 298 {
 299         ssize_t result = -EIOCBQUEUED;
 300
 301         /* Async requests don't wait here */
 302         if (dreq->iocb)
 303                 goto out;
 304
 305         result = wait_for_completion_killable(&dreq->completion);
 306
 307         if (!result)
 308                 result = dreq->error;
 309         if (!result)
 310                 result = dreq->count;
 311
 312 out:
 313         return (ssize_t) result;
 314 }
 315
 316 /*
 317  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 318  * the iocb is still valid here if this is a synchronous request.
 319  */
 320 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 321 {
 322         struct inode *inode = dreq->inode;
 323
 324         if (dreq->iocb && write) {
 325                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 326
 327                 spin_lock(&inode->i_lock);
 328                 if (i_size_read(inode) < pos)
 329                         i_size_write(inode, pos);
 330                 spin_unlock(&inode->i_lock);
 331         }
 332
 333         if (write)
 334                 nfs_zap_mapping(inode, inode->i_mapping);
 335
 336         inode_dio_done(inode);
 337
 338         if (dreq->iocb) {
 339                 long res = (long) dreq->error;
 340                 if (!res)
 341                         res = (long) dreq->count;
 342                 aio_complete(dreq->iocb, res, 0);
 343         }
 344
 345         complete_all(&dreq->completion);
 346
 347         nfs_direct_req_release(dreq);
 348 }
 349
 350 static void nfs_direct_readpage_release(struct nfs_page *req)
 351 {
 352         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 353                 req->wb_context->dentry->d_inode->i_sb->s_id,
 354                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 355                 req->wb_bytes,
 356                 (long long)req_offset(req));
 357         nfs_release_request(req);
 358 }
 359
 360 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 361 {
 362         unsigned long bytes = 0;
 363         struct nfs_direct_req *dreq = hdr->dreq;
 364
 365         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 366                 goto out_put;
 367
 368         spin_lock(&dreq->lock);
 369         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 370                 dreq->error = hdr->error;
 371         else
 372                 dreq->count += hdr->good_bytes;
 373         spin_unlock(&dreq->lock);
 374
 375         while (!list_empty(&hdr->pages)) {
 376                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 377                 struct page *page = req->wb_page;
 378
 379                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 380                         set_page_dirty(page);
 381                 bytes += req->wb_bytes;
 382                 nfs_list_remove_request(req);
 383                 nfs_direct_readpage_release(req);
 384         }
 385 out_put:
 386         if (put_dreq(dreq))
 387                 nfs_direct_complete(dreq, false);
 388         hdr->release(hdr);
 389 }
 390
 391 static void nfs_read_sync_pgio_error(struct list_head *head)
 392 {
 393         struct nfs_page *req;
 394
 395         while (!list_empty(head)) {
 396                 req = nfs_list_entry(head->next);
 397                 nfs_list_remove_request(req);
 398                 nfs_release_request(req);
 399         }
 400 }
 401
 402 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 403 {
 404         get_dreq(hdr->dreq);
 405 }
 406
 407 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 408         .error_cleanup = nfs_read_sync_pgio_error,
 409         .init_hdr = nfs_direct_pgio_init,
 410         .completion = nfs_direct_read_completion,
 411 };
 412
 413 /*
 414  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 415  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 416  * bail and stop sending more reads.  Read length accounting is
 417  * handled automatically by nfs_direct_read_result().  Otherwise, if
 418  * no requests have been sent, just return an error.
 419  */
 420
 421 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 422                                               struct iov_iter *iter,
 423                                               loff_t pos)
 424 {
 425         struct nfs_pageio_descriptor desc;
 426         struct inode *inode = dreq->inode;
 427         ssize_t result = -EINVAL;
 428         size_t requested_bytes = 0;
 429         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 430
 431         nfs_pageio_init_read(&desc, dreq->inode, false,
 432                              &nfs_direct_read_completion_ops);
 433         get_dreq(dreq);
 434         desc.pg_dreq = dreq;
 435         atomic_inc(&inode->i_dio_count);
 436
 437         while (iov_iter_count(iter)) {
 438                 struct page **pagevec;
 439                 size_t bytes;
 440                 size_t pgbase;
 441                 unsigned npages, i;
 442
 443                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 444                                                   rsize, &pgbase);
 445                 if (result < 0)
 446                         break;
 447
 448                 bytes = result;
 449                 iov_iter_advance(iter, bytes);
 450                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 451                 for (i = 0; i < npages; i++) {
 452                         struct nfs_page *req;
 453                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 454                         /* XXX do we need to do the eof zeroing found in async_filler? */
 455                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 456                                                  pgbase, req_len);
 457                         if (IS_ERR(req)) {
 458                                 result = PTR_ERR(req);
 459                                 break;
 460                         }
 461                         req->wb_index = pos >> PAGE_SHIFT;
 462                         req->wb_offset = pos & ~PAGE_MASK;
 463                         if (!nfs_pageio_add_request(&desc, req)) {
 464                                 result = desc.pg_error;
 465                                 nfs_release_request(req);
 466                                 break;
 467                         }
 468                         pgbase = 0;
 469                         bytes -= req_len;
 470                         requested_bytes += req_len;
 471                         pos += req_len;
 472                         dreq->bytes_left -= req_len;
 473                 }
 474                 nfs_direct_release_pages(pagevec, npages);
 475                 kvfree(pagevec);
 476                 if (result < 0)
 477                         break;
 478         }
 479
 480         nfs_pageio_complete(&desc);
 481
 482         /*
 483          * If no bytes were started, return the error, and let the
 484          * generic layer handle the completion.
 485          */
 486         if (requested_bytes == 0) {
 487                 inode_dio_done(inode);
 488                 nfs_direct_req_release(dreq);
 489                 return result < 0 ? result : -EIO;
 490         }
 491
 492         if (put_dreq(dreq))
 493                 nfs_direct_complete(dreq, false);
 494         return 0;
 495 }
 496
 497 /**
 498  * nfs_file_direct_read - file direct read operation for NFS files
 499  * @iocb: target I/O control block
 500  * @iter: vector of user buffers into which to read data
 501  * @pos: byte offset in file where reading starts
 502  *
 503  * We use this function for direct reads instead of calling
 504  * generic_file_aio_read() in order to avoid gfar's check to see if
 505  * the request starts before the end of the file.  For that check
 506  * to work, we must generate a GETATTR before each direct read, and
 507  * even then there is a window between the GETATTR and the subsequent
 508  * READ where the file size could change.  Our preference is simply
 509  * to do all reads the application wants, and the server will take
 510  * care of managing the end of file boundary.
 511  *
 512  * This function also eliminates unnecessarily updating the file's
 513  * atime locally, as the NFS server sets the file's atime, and this
 514  * client must read the updated atime from the server back into its
 515  * cache.
 516  */
 517 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 518                                 loff_t pos)
 519 {
 520         struct file *file = iocb->ki_filp;
 521         struct address_space *mapping = file->f_mapping;
 522         struct inode *inode = mapping->host;
 523         struct nfs_direct_req *dreq;
 524         struct nfs_lock_context *l_ctx;
 525         ssize_t result = -EINVAL;
 526         size_t count = iov_iter_count(iter);
 527         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 528
 529         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 530                 file, count, (long long) pos);
 531
 532         result = 0;
 533         if (!count)
 534                 goto out;
 535
 536         mutex_lock(&inode->i_mutex);
 537         result = nfs_sync_mapping(mapping);
 538         if (result)
 539                 goto out_unlock;
 540
 541         task_io_account_read(count);
 542
 543         result = -ENOMEM;
 544         dreq = nfs_direct_req_alloc();
 545         if (dreq == NULL)
 546                 goto out_unlock;
 547
 548         dreq->inode = inode;
 549         dreq->bytes_left = count;
 550         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 551         l_ctx = nfs_get_lock_context(dreq->ctx);
 552         if (IS_ERR(l_ctx)) {
 553                 result = PTR_ERR(l_ctx);
 554                 goto out_release;
 555         }
 556         dreq->l_ctx = l_ctx;
 557         if (!is_sync_kiocb(iocb))
 558                 dreq->iocb = iocb;
 559
 560         NFS_I(inode)->read_io += count;
 561         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 562
 563         mutex_unlock(&inode->i_mutex);
 564
 565         if (!result) {
 566                 result = nfs_direct_wait(dreq);
 567                 if (result > 0)
 568                         iocb->ki_pos = pos + result;
 569         }
 570
 571         nfs_direct_req_release(dreq);
 572         return result;
 573
 574 out_release:
 575         nfs_direct_req_release(dreq);
 576 out_unlock:
 577         mutex_unlock(&inode->i_mutex);
 578 out:
 579         return result;
 580 }
 581
 582 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 583 {
 584         struct nfs_pageio_descriptor desc;
 585         struct nfs_page *req, *tmp;
 586         LIST_HEAD(reqs);
 587         struct nfs_commit_info cinfo;
 588         LIST_HEAD(failed);
 589
 590         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 591         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 592         spin_lock(cinfo.lock);
 593         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 594         spin_unlock(cinfo.lock);
 595
 596         dreq->count = 0;
 597         get_dreq(dreq);
 598
 599         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 600                               &nfs_direct_write_completion_ops);
 601         desc.pg_dreq = dreq;
 602
 603         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 604                 if (!nfs_pageio_add_request(&desc, req)) {
 605                         nfs_list_remove_request(req);
 606                         nfs_list_add_request(req, &failed);
 607                         spin_lock(cinfo.lock);
 608                         dreq->flags = 0;
 609                         dreq->error = -EIO;
 610                         spin_unlock(cinfo.lock);
 611                 }
 612                 nfs_release_request(req);
 613         }
 614         nfs_pageio_complete(&desc);
 615
 616         while (!list_empty(&failed)) {
 617                 req = nfs_list_entry(failed.next);
 618                 nfs_list_remove_request(req);
 619                 nfs_unlock_and_release_request(req);
 620         }
 621
 622         if (put_dreq(dreq))
 623                 nfs_direct_write_complete(dreq, dreq->inode);
 624 }
 625
 626 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 627 {
 628         struct nfs_direct_req *dreq = data->dreq;
 629         struct nfs_commit_info cinfo;
 630         struct nfs_page *req;
 631         int status = data->task.tk_status;
 632
 633         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 634         if (status < 0) {
 635                 dprintk("NFS: %5u commit failed with error %d.\n",
 636                         data->task.tk_pid, status);
 637                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 638         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 639                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 640                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 641         }
 642
 643         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 644         while (!list_empty(&data->pages)) {
 645                 req = nfs_list_entry(data->pages.next);
 646                 nfs_list_remove_request(req);
 647                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 648                         /* Note the rewrite will go through mds */
 649                         nfs_mark_request_commit(req, NULL, &cinfo);
 650                 } else
 651                         nfs_release_request(req);
 652                 nfs_unlock_and_release_request(req);
 653         }
 654
 655         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 656                 nfs_direct_write_complete(dreq, data->inode);
 657 }
 658
 659 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 660 {
 661         /* There is no lock to clear */
 662 }
 663
 664 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 665         .completion = nfs_direct_commit_complete,
 666         .error_cleanup = nfs_direct_error_cleanup,
 667 };
 668
 669 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 670 {
 671         int res;
 672         struct nfs_commit_info cinfo;
 673         LIST_HEAD(mds_list);
 674
 675         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 676         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 677         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 678         if (res < 0) /* res == -ENOMEM */
 679                 nfs_direct_write_reschedule(dreq);
 680 }
 681
 682 static void nfs_direct_write_schedule_work(struct work_struct *work)
 683 {
 684         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 685         int flags = dreq->flags;
 686
 687         dreq->flags = 0;
 688         switch (flags) {
 689                 case NFS_ODIRECT_DO_COMMIT:
 690                         nfs_direct_commit_schedule(dreq);
 691                         break;
 692                 case NFS_ODIRECT_RESCHED_WRITES:
 693                         nfs_direct_write_reschedule(dreq);
 694                         break;
 695                 default:
 696                         nfs_direct_complete(dreq, true);
 697         }
 698 }
 699
 700 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 701 {
 702         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 703 }
 704
 705 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 706 {
 707         struct nfs_direct_req *dreq = hdr->dreq;
 708         struct nfs_commit_info cinfo;
 709         bool request_commit = false;
 710         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 711
 712         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 713                 goto out_put;
 714
 715         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 716
 717         spin_lock(&dreq->lock);
 718
 719         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 720                 dreq->flags = 0;
 721                 dreq->error = hdr->error;
 722         }
 723         if (dreq->error == 0) {
 724                 dreq->count += hdr->good_bytes;
 725                 if (nfs_write_need_commit(hdr)) {
 726                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 727                                 request_commit = true;
 728                         else if (dreq->flags == 0) {
 729                                 nfs_direct_set_hdr_verf(dreq, hdr);
 730                                 request_commit = true;
 731                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 732                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 733                                 request_commit = true;
 734                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 735                                         dreq->flags =
 736                                                 NFS_ODIRECT_RESCHED_WRITES;
 737                         }
 738                 }
 739         }
 740         spin_unlock(&dreq->lock);
 741
 742         while (!list_empty(&hdr->pages)) {
 743
 744                 req = nfs_list_entry(hdr->pages.next);
 745                 nfs_list_remove_request(req);
 746                 if (request_commit) {
 747                         kref_get(&req->wb_kref);
 748                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 749                 }
 750                 nfs_unlock_and_release_request(req);
 751         }
 752
 753 out_put:
 754         if (put_dreq(dreq))
 755                 nfs_direct_write_complete(dreq, hdr->inode);
 756         hdr->release(hdr);
 757 }
 758
 759 static void nfs_write_sync_pgio_error(struct list_head *head)
 760 {
 761         struct nfs_page *req;
 762
 763         while (!list_empty(head)) {
 764                 req = nfs_list_entry(head->next);
 765                 nfs_list_remove_request(req);
 766                 nfs_unlock_and_release_request(req);
 767         }
 768 }
 769
 770 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 771         .error_cleanup = nfs_write_sync_pgio_error,
 772         .init_hdr = nfs_direct_pgio_init,
 773         .completion = nfs_direct_write_completion,
 774 };
 775
 776
 777 /*
 778  * NB: Return the value of the first error return code.  Subsequent
 779  *     errors after the first one are ignored.
 780  */
 781 /*
 782  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 783  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 784  * bail and stop sending more writes.  Write length accounting is
 785  * handled automatically by nfs_direct_write_result().  Otherwise, if
 786  * no requests have been sent, just return an error.
 787  */
 788 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 789                                                struct iov_iter *iter,
 790                                                loff_t pos)
 791 {
 792         struct nfs_pageio_descriptor desc;
 793         struct inode *inode = dreq->inode;
 794         ssize_t result = 0;
 795         size_t requested_bytes = 0;
 796         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 797
 798         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 799                               &nfs_direct_write_completion_ops);
 800         desc.pg_dreq = dreq;
 801         get_dreq(dreq);
 802         atomic_inc(&inode->i_dio_count);
 803
 804         NFS_I(inode)->write_io += iov_iter_count(iter);
 805         while (iov_iter_count(iter)) {
 806                 struct page **pagevec;
 807                 size_t bytes;
 808                 size_t pgbase;
 809                 unsigned npages, i;
 810
 811                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 812                                                   wsize, &pgbase);
 813                 if (result < 0)
 814                         break;
 815
 816                 bytes = result;
 817                 iov_iter_advance(iter, bytes);
 818                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 819                 for (i = 0; i < npages; i++) {
 820                         struct nfs_page *req;
 821                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 822
 823                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 824                                                  pgbase, req_len);
 825                         if (IS_ERR(req)) {
 826                                 result = PTR_ERR(req);
 827                                 break;
 828                         }
 829                         nfs_lock_request(req);
 830                         req->wb_index = pos >> PAGE_SHIFT;
 831                         req->wb_offset = pos & ~PAGE_MASK;
 832                         if (!nfs_pageio_add_request(&desc, req)) {
 833                                 result = desc.pg_error;
 834                                 nfs_unlock_and_release_request(req);
 835                                 break;
 836                         }
 837                         pgbase = 0;
 838                         bytes -= req_len;
 839                         requested_bytes += req_len;
 840                         pos += req_len;
 841                         dreq->bytes_left -= req_len;
 842                 }
 843                 nfs_direct_release_pages(pagevec, npages);
 844                 kvfree(pagevec);
 845                 if (result < 0)
 846                         break;
 847         }
 848         nfs_pageio_complete(&desc);
 849
 850         /*
 851          * If no bytes were started, return the error, and let the
 852          * generic layer handle the completion.
 853          */
 854         if (requested_bytes == 0) {
 855                 inode_dio_done(inode);
 856                 nfs_direct_req_release(dreq);
 857                 return result < 0 ? result : -EIO;
 858         }
 859
 860         if (put_dreq(dreq))
 861                 nfs_direct_write_complete(dreq, dreq->inode);
 862         return 0;
 863 }
 864
 865 /**
 866  * nfs_file_direct_write - file direct write operation for NFS files
 867  * @iocb: target I/O control block
 868  * @iter: vector of user buffers from which to write data
 869  * @pos: byte offset in file where writing starts
 870  *
 871  * We use this function for direct writes instead of calling
 872  * generic_file_aio_write() in order to avoid taking the inode
 873  * semaphore and updating the i_size.  The NFS server will set
 874  * the new i_size and this client must read the updated size
 875  * back into its cache.  We let the server do generic write
 876  * parameter checking and report problems.
 877  *
 878  * We eliminate local atime updates, see direct read above.
 879  *
 880  * We avoid unnecessary page cache invalidations for normal cached
 881  * readers of this file.
 882  *
 883  * Note that O_APPEND is not supported for NFS direct writes, as there
 884  * is no atomic O_APPEND write facility in the NFS protocol.
 885  */
 886 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 887                                 loff_t pos)
 888 {
 889         ssize_t result = -EINVAL;
 890         struct file *file = iocb->ki_filp;
 891         struct address_space *mapping = file->f_mapping;
 892         struct inode *inode = mapping->host;
 893         struct nfs_direct_req *dreq;
 894         struct nfs_lock_context *l_ctx;
 895         loff_t end;
 896         size_t count = iov_iter_count(iter);
 897         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 898
 899         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 900
 901         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 902                 file, count, (long long) pos);
 903
 904         result = generic_write_checks(file, &pos, &count, 0);
 905         if (result)
 906                 goto out;
 907
 908         result = -EINVAL;
 909         if ((ssize_t) count < 0)
 910                 goto out;
 911         result = 0;
 912         if (!count)
 913                 goto out;
 914
 915         mutex_lock(&inode->i_mutex);
 916
 917         result = nfs_sync_mapping(mapping);
 918         if (result)
 919                 goto out_unlock;
 920
 921         if (mapping->nrpages) {
 922                 result = invalidate_inode_pages2_range(mapping,
 923                                         pos >> PAGE_CACHE_SHIFT, end);
 924                 if (result)
 925                         goto out_unlock;
 926         }
 927
 928         task_io_account_write(count);
 929
 930         result = -ENOMEM;
 931         dreq = nfs_direct_req_alloc();
 932         if (!dreq)
 933                 goto out_unlock;
 934
 935         dreq->inode = inode;
 936         dreq->bytes_left = count;
 937         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 938         l_ctx = nfs_get_lock_context(dreq->ctx);
 939         if (IS_ERR(l_ctx)) {
 940                 result = PTR_ERR(l_ctx);
 941                 goto out_release;
 942         }
 943         dreq->l_ctx = l_ctx;
 944         if (!is_sync_kiocb(iocb))
 945                 dreq->iocb = iocb;
 946
 947         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
 948
 949         if (mapping->nrpages) {
 950                 invalidate_inode_pages2_range(mapping,
 951                                               pos >> PAGE_CACHE_SHIFT, end);
 952         }
 953
 954         mutex_unlock(&inode->i_mutex);
 955
 956         if (!result) {
 957                 result = nfs_direct_wait(dreq);
 958                 if (result > 0) {
 959                         struct inode *inode = mapping->host;
 960
 961                         iocb->ki_pos = pos + result;
 962                         spin_lock(&inode->i_lock);
 963                         if (i_size_read(inode) < iocb->ki_pos)
 964                                 i_size_write(inode, iocb->ki_pos);
 965                         spin_unlock(&inode->i_lock);
 966                 }
 967         }
 968         nfs_direct_req_release(dreq);
 969         return result;
 970
 971 out_release:
 972         nfs_direct_req_release(dreq);
 973 out_unlock:
 974         mutex_unlock(&inode->i_mutex);
 975 out:
 976         return result;
 977 }
 978
 979 /**
 980  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 981  *
 982  */
 983 int __init nfs_init_directcache(void)
 984 {
 985         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 986                                                 sizeof(struct nfs_direct_req),
 987                                                 0, (SLAB_RECLAIM_ACCOUNT|
 988                                                         SLAB_MEM_SPREAD),
 989                                                 NULL);
 990         if (nfs_direct_cachep == NULL)
 991                 return -ENOMEM;
 992
 993         return 0;
 994 }
 995
 996 /**
 997  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 998  *
 999  */
1000 void nfs_destroy_directcache(void)
1001 {
1002         kmem_cache_destroy(nfs_direct_cachep);
1003 }