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
  50 #include <linux/nfs_fs.h>
  51 #include <linux/nfs_page.h>
  52 #include <linux/sunrpc/clnt.h>
  53
  54 #include <asm/uaccess.h>
  55 #include <linux/atomic.h>
  56
  57 #include "internal.h"
  58 #include "iostat.h"
  59 #include "pnfs.h"
  60
  61 #define NFSDBG_FACILITY         NFSDBG_VFS
  62
  63 static struct kmem_cache *nfs_direct_cachep;
  64
  65 /*
  66  * This represents a set of asynchronous requests that we're waiting on
  67  */
  68 struct nfs_direct_req {
  69         struct kref             kref;           /* release manager */
  70
  71         /* I/O parameters */
  72         struct nfs_open_context *ctx;           /* file open context info */
  73         struct nfs_lock_context *l_ctx;         /* Lock context info */
  74         struct kiocb *          iocb;           /* controlling i/o request */
  75         struct inode *          inode;          /* target file of i/o */
  76
  77         /* completion state */
  78         atomic_t                io_count;       /* i/os we're waiting for */
  79         spinlock_t              lock;           /* protect completion state */
  80         ssize_t                 count,          /* bytes actually processed */
  81                                 error;          /* any reported error */
  82         struct completion       completion;     /* wait for i/o completion */
  83
  84         /* commit state */
  85         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  86         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  87         struct work_struct      work;
  88         int                     flags;
  89 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  90 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  91         struct nfs_writeverf    verf;           /* unstable write verifier */
  92 };
  93
  94 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  95 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  96 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  97 static void nfs_direct_write_schedule_work(struct work_struct *work);
  98
  99 static inline void get_dreq(struct nfs_direct_req *dreq)
 100 {
 101         atomic_inc(&dreq->io_count);
 102 }
 103
 104 static inline int put_dreq(struct nfs_direct_req *dreq)
 105 {
 106         return atomic_dec_and_test(&dreq->io_count);
 107 }
 108
 109 /**
 110  * nfs_direct_IO - NFS address space operation for direct I/O
 111  * @rw: direction (read or write)
 112  * @iocb: target I/O control block
 113  * @iov: array of vectors that define I/O buffer
 114  * @pos: offset in file to begin the operation
 115  * @nr_segs: size of iovec array
 116  *
 117  * The presence of this routine in the address space ops vector means
 118  * the NFS client supports direct I/O.  However, we shunt off direct
 119  * read and write requests before the VFS gets them, so this method
 120  * should never be called.
 121  */
 122 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 123 {
 124         dprintk("NFS: nfs_direct_IO (%s) off/no(%Ld/%lu) EINVAL\n",
 125                         iocb->ki_filp->f_path.dentry->d_name.name,
 126                         (long long) pos, nr_segs);
 127
 128         return -EINVAL;
 129 }
 130
 131 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 132 {
 133         unsigned int i;
 134         for (i = 0; i < npages; i++)
 135                 page_cache_release(pages[i]);
 136 }
 137
 138 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 139                               struct nfs_direct_req *dreq)
 140 {
 141         cinfo->lock = &dreq->lock;
 142         cinfo->mds = &dreq->mds_cinfo;
 143         cinfo->ds = &dreq->ds_cinfo;
 144         cinfo->dreq = dreq;
 145         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 146 }
 147
 148 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 149 {
 150         struct nfs_direct_req *dreq;
 151
 152         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 153         if (!dreq)
 154                 return NULL;
 155
 156         kref_init(&dreq->kref);
 157         kref_get(&dreq->kref);
 158         init_completion(&dreq->completion);
 159         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 160         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 161         spin_lock_init(&dreq->lock);
 162
 163         return dreq;
 164 }
 165
 166 static void nfs_direct_req_free(struct kref *kref)
 167 {
 168         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 169
 170         if (dreq->l_ctx != NULL)
 171                 nfs_put_lock_context(dreq->l_ctx);
 172         if (dreq->ctx != NULL)
 173                 put_nfs_open_context(dreq->ctx);
 174         kmem_cache_free(nfs_direct_cachep, dreq);
 175 }
 176
 177 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 178 {
 179         kref_put(&dreq->kref, nfs_direct_req_free);
 180 }
 181
 182 /*
 183  * Collects and returns the final error value/byte-count.
 184  */
 185 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 186 {
 187         ssize_t result = -EIOCBQUEUED;
 188
 189         /* Async requests don't wait here */
 190         if (dreq->iocb)
 191                 goto out;
 192
 193         result = wait_for_completion_killable(&dreq->completion);
 194
 195         if (!result)
 196                 result = dreq->error;
 197         if (!result)
 198                 result = dreq->count;
 199
 200 out:
 201         return (ssize_t) result;
 202 }
 203
 204 /*
 205  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 206  * the iocb is still valid here if this is a synchronous request.
 207  */
 208 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 209 {
 210         if (dreq->iocb) {
 211                 long res = (long) dreq->error;
 212                 if (!res)
 213                         res = (long) dreq->count;
 214                 aio_complete(dreq->iocb, res, 0);
 215         }
 216         complete_all(&dreq->completion);
 217
 218         nfs_direct_req_release(dreq);
 219 }
 220
 221 static void nfs_direct_readpage_release(struct nfs_page *req)
 222 {
 223         dprintk("NFS: direct read done (%s/%lld %d@%lld)\n",
 224                 req->wb_context->dentry->d_inode->i_sb->s_id,
 225                 (long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 226                 req->wb_bytes,
 227                 (long long)req_offset(req));
 228         nfs_release_request(req);
 229 }
 230
 231 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 232 {
 233         unsigned long bytes = 0;
 234         struct nfs_direct_req *dreq = hdr->dreq;
 235
 236         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 237                 goto out_put;
 238
 239         spin_lock(&dreq->lock);
 240         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 241                 dreq->error = hdr->error;
 242         else
 243                 dreq->count += hdr->good_bytes;
 244         spin_unlock(&dreq->lock);
 245
 246         while (!list_empty(&hdr->pages)) {
 247                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 248                 struct page *page = req->wb_page;
 249
 250                 if (test_bit(NFS_IOHDR_EOF, &hdr->flags)) {
 251                         if (bytes > hdr->good_bytes)
 252                                 zero_user(page, 0, PAGE_SIZE);
 253                         else if (hdr->good_bytes - bytes < PAGE_SIZE)
 254                                 zero_user_segment(page,
 255                                         hdr->good_bytes & ~PAGE_MASK,
 256                                         PAGE_SIZE);
 257                 }
 258                 if (!PageCompound(page)) {
 259                         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 260                                 if (bytes < hdr->good_bytes)
 261                                         set_page_dirty(page);
 262                         } else
 263                                 set_page_dirty(page);
 264                 }
 265                 bytes += req->wb_bytes;
 266                 nfs_list_remove_request(req);
 267                 nfs_direct_readpage_release(req);
 268         }
 269 out_put:
 270         if (put_dreq(dreq))
 271                 nfs_direct_complete(dreq);
 272         hdr->release(hdr);
 273 }
 274
 275 static void nfs_read_sync_pgio_error(struct list_head *head)
 276 {
 277         struct nfs_page *req;
 278
 279         while (!list_empty(head)) {
 280                 req = nfs_list_entry(head->next);
 281                 nfs_list_remove_request(req);
 282                 nfs_release_request(req);
 283         }
 284 }
 285
 286 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 287 {
 288         get_dreq(hdr->dreq);
 289 }
 290
 291 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 292         .error_cleanup = nfs_read_sync_pgio_error,
 293         .init_hdr = nfs_direct_pgio_init,
 294         .completion = nfs_direct_read_completion,
 295 };
 296
 297 /*
 298  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 299  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 300  * bail and stop sending more reads.  Read length accounting is
 301  * handled automatically by nfs_direct_read_result().  Otherwise, if
 302  * no requests have been sent, just return an error.
 303  */
 304 static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
 305                                                 const struct iovec *iov,
 306                                                 loff_t pos)
 307 {
 308         struct nfs_direct_req *dreq = desc->pg_dreq;
 309         struct nfs_open_context *ctx = dreq->ctx;
 310         struct inode *inode = ctx->dentry->d_inode;
 311         unsigned long user_addr = (unsigned long)iov->iov_base;
 312         size_t count = iov->iov_len;
 313         size_t rsize = NFS_SERVER(inode)->rsize;
 314         unsigned int pgbase;
 315         int result;
 316         ssize_t started = 0;
 317         struct page **pagevec = NULL;
 318         unsigned int npages;
 319
 320         do {
 321                 size_t bytes;
 322                 int i;
 323
 324                 pgbase = user_addr & ~PAGE_MASK;
 325                 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
 326
 327                 result = -ENOMEM;
 328                 npages = nfs_page_array_len(pgbase, bytes);
 329                 if (!pagevec)
 330                         pagevec = kmalloc(npages * sizeof(struct page *),
 331                                           GFP_KERNEL);
 332                 if (!pagevec)
 333                         break;
 334                 down_read(&current->mm->mmap_sem);
 335                 result = get_user_pages(current, current->mm, user_addr,
 336                                         npages, 1, 0, pagevec, NULL);
 337                 up_read(&current->mm->mmap_sem);
 338                 if (result < 0)
 339                         break;
 340                 if ((unsigned)result < npages) {
 341                         bytes = result * PAGE_SIZE;
 342                         if (bytes <= pgbase) {
 343                                 nfs_direct_release_pages(pagevec, result);
 344                                 break;
 345                         }
 346                         bytes -= pgbase;
 347                         npages = result;
 348                 }
 349
 350                 for (i = 0; i < npages; i++) {
 351                         struct nfs_page *req;
 352                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 353                         /* XXX do we need to do the eof zeroing found in async_filler? */
 354                         req = nfs_create_request(dreq->ctx, dreq->inode,
 355                                                  pagevec[i],
 356                                                  pgbase, req_len);
 357                         if (IS_ERR(req)) {
 358                                 result = PTR_ERR(req);
 359                                 break;
 360                         }
 361                         req->wb_index = pos >> PAGE_SHIFT;
 362                         req->wb_offset = pos & ~PAGE_MASK;
 363                         if (!nfs_pageio_add_request(desc, req)) {
 364                                 result = desc->pg_error;
 365                                 nfs_release_request(req);
 366                                 break;
 367                         }
 368                         pgbase = 0;
 369                         bytes -= req_len;
 370                         started += req_len;
 371                         user_addr += req_len;
 372                         pos += req_len;
 373                         count -= req_len;
 374                 }
 375                 /* The nfs_page now hold references to these pages */
 376                 nfs_direct_release_pages(pagevec, npages);
 377         } while (count != 0 && result >= 0);
 378
 379         kfree(pagevec);
 380
 381         if (started)
 382                 return started;
 383         return result < 0 ? (ssize_t) result : -EFAULT;
 384 }
 385
 386 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 387                                               const struct iovec *iov,
 388                                               unsigned long nr_segs,
 389                                               loff_t pos)
 390 {
 391         struct nfs_pageio_descriptor desc;
 392         ssize_t result = -EINVAL;
 393         size_t requested_bytes = 0;
 394         unsigned long seg;
 395
 396         nfs_pageio_init_read(&desc, dreq->inode,
 397                              &nfs_direct_read_completion_ops);
 398         get_dreq(dreq);
 399         desc.pg_dreq = dreq;
 400
 401         for (seg = 0; seg < nr_segs; seg++) {
 402                 const struct iovec *vec = &iov[seg];
 403                 result = nfs_direct_read_schedule_segment(&desc, vec, pos);
 404                 if (result < 0)
 405                         break;
 406                 requested_bytes += result;
 407                 if ((size_t)result < vec->iov_len)
 408                         break;
 409                 pos += vec->iov_len;
 410         }
 411
 412         nfs_pageio_complete(&desc);
 413
 414         /*
 415          * If no bytes were started, return the error, and let the
 416          * generic layer handle the completion.
 417          */
 418         if (requested_bytes == 0) {
 419                 nfs_direct_req_release(dreq);
 420                 return result < 0 ? result : -EIO;
 421         }
 422
 423         if (put_dreq(dreq))
 424                 nfs_direct_complete(dreq);
 425         return 0;
 426 }
 427
 428 static ssize_t nfs_direct_read(struct kiocb *iocb, const struct iovec *iov,
 429                                unsigned long nr_segs, loff_t pos)
 430 {
 431         ssize_t result = -ENOMEM;
 432         struct inode *inode = iocb->ki_filp->f_mapping->host;
 433         struct nfs_direct_req *dreq;
 434
 435         dreq = nfs_direct_req_alloc();
 436         if (dreq == NULL)
 437                 goto out;
 438
 439         dreq->inode = inode;
 440         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 441         dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
 442         if (dreq->l_ctx == NULL)
 443                 goto out_release;
 444         if (!is_sync_kiocb(iocb))
 445                 dreq->iocb = iocb;
 446
 447         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos);
 448         if (!result)
 449                 result = nfs_direct_wait(dreq);
 450 out_release:
 451         nfs_direct_req_release(dreq);
 452 out:
 453         return result;
 454 }
 455
 456 #if defined(CONFIG_NFS_V3) || defined(CONFIG_NFS_V4)
 457 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 458 {
 459         struct nfs_pageio_descriptor desc;
 460         struct nfs_page *req, *tmp;
 461         LIST_HEAD(reqs);
 462         struct nfs_commit_info cinfo;
 463         LIST_HEAD(failed);
 464
 465         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 466         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 467         spin_lock(cinfo.lock);
 468         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 469         spin_unlock(cinfo.lock);
 470
 471         dreq->count = 0;
 472         get_dreq(dreq);
 473
 474         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE,
 475                               &nfs_direct_write_completion_ops);
 476         desc.pg_dreq = dreq;
 477
 478         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 479                 if (!nfs_pageio_add_request(&desc, req)) {
 480                         nfs_list_add_request(req, &failed);
 481                         spin_lock(cinfo.lock);
 482                         dreq->flags = 0;
 483                         dreq->error = -EIO;
 484                         spin_unlock(cinfo.lock);
 485                 }
 486         }
 487         nfs_pageio_complete(&desc);
 488
 489         while (!list_empty(&failed))
 490                 nfs_unlock_and_release_request(req);
 491
 492         if (put_dreq(dreq))
 493                 nfs_direct_write_complete(dreq, dreq->inode);
 494 }
 495
 496 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 497 {
 498         struct nfs_direct_req *dreq = data->dreq;
 499         struct nfs_commit_info cinfo;
 500         struct nfs_page *req;
 501         int status = data->task.tk_status;
 502
 503         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 504         if (status < 0) {
 505                 dprintk("NFS: %5u commit failed with error %d.\n",
 506                         data->task.tk_pid, status);
 507                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 508         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 509                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 510                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 511         }
 512
 513         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 514         while (!list_empty(&data->pages)) {
 515                 req = nfs_list_entry(data->pages.next);
 516                 nfs_list_remove_request(req);
 517                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 518                         /* Note the rewrite will go through mds */
 519                         kref_get(&req->wb_kref);
 520                         nfs_mark_request_commit(req, NULL, &cinfo);
 521                 }
 522                 nfs_unlock_and_release_request(req);
 523         }
 524
 525         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 526                 nfs_direct_write_complete(dreq, data->inode);
 527 }
 528
 529 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 530 {
 531         /* There is no lock to clear */
 532 }
 533
 534 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 535         .completion = nfs_direct_commit_complete,
 536         .error_cleanup = nfs_direct_error_cleanup,
 537 };
 538
 539 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 540 {
 541         int res;
 542         struct nfs_commit_info cinfo;
 543         LIST_HEAD(mds_list);
 544
 545         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 546         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 547         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 548         if (res < 0) /* res == -ENOMEM */
 549                 nfs_direct_write_reschedule(dreq);
 550 }
 551
 552 static void nfs_direct_write_schedule_work(struct work_struct *work)
 553 {
 554         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 555         int flags = dreq->flags;
 556
 557         dreq->flags = 0;
 558         switch (flags) {
 559                 case NFS_ODIRECT_DO_COMMIT:
 560                         nfs_direct_commit_schedule(dreq);
 561                         break;
 562                 case NFS_ODIRECT_RESCHED_WRITES:
 563                         nfs_direct_write_reschedule(dreq);
 564                         break;
 565                 default:
 566                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 567                         nfs_direct_complete(dreq);
 568         }
 569 }
 570
 571 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 572 {
 573         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 574 }
 575
 576 #else
 577 static void nfs_direct_write_schedule_work(struct work_struct *work)
 578 {
 579 }
 580
 581 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 582 {
 583         nfs_zap_mapping(inode, inode->i_mapping);
 584         nfs_direct_complete(dreq);
 585 }
 586 #endif
 587
 588 /*
 589  * NB: Return the value of the first error return code.  Subsequent
 590  *     errors after the first one are ignored.
 591  */
 592 /*
 593  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 594  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 595  * bail and stop sending more writes.  Write length accounting is
 596  * handled automatically by nfs_direct_write_result().  Otherwise, if
 597  * no requests have been sent, just return an error.
 598  */
 599 static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
 600                                                  const struct iovec *iov,
 601                                                  loff_t pos)
 602 {
 603         struct nfs_direct_req *dreq = desc->pg_dreq;
 604         struct nfs_open_context *ctx = dreq->ctx;
 605         struct inode *inode = ctx->dentry->d_inode;
 606         unsigned long user_addr = (unsigned long)iov->iov_base;
 607         size_t count = iov->iov_len;
 608         size_t wsize = NFS_SERVER(inode)->wsize;
 609         unsigned int pgbase;
 610         int result;
 611         ssize_t started = 0;
 612         struct page **pagevec = NULL;
 613         unsigned int npages;
 614
 615         do {
 616                 size_t bytes;
 617                 int i;
 618
 619                 pgbase = user_addr & ~PAGE_MASK;
 620                 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
 621
 622                 result = -ENOMEM;
 623                 npages = nfs_page_array_len(pgbase, bytes);
 624                 if (!pagevec)
 625                         pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
 626                 if (!pagevec)
 627                         break;
 628
 629                 down_read(&current->mm->mmap_sem);
 630                 result = get_user_pages(current, current->mm, user_addr,
 631                                         npages, 0, 0, pagevec, NULL);
 632                 up_read(&current->mm->mmap_sem);
 633                 if (result < 0)
 634                         break;
 635
 636                 if ((unsigned)result < npages) {
 637                         bytes = result * PAGE_SIZE;
 638                         if (bytes <= pgbase) {
 639                                 nfs_direct_release_pages(pagevec, result);
 640                                 break;
 641                         }
 642                         bytes -= pgbase;
 643                         npages = result;
 644                 }
 645
 646                 for (i = 0; i < npages; i++) {
 647                         struct nfs_page *req;
 648                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 649
 650                         req = nfs_create_request(dreq->ctx, dreq->inode,
 651                                                  pagevec[i],
 652                                                  pgbase, req_len);
 653                         if (IS_ERR(req)) {
 654                                 result = PTR_ERR(req);
 655                                 break;
 656                         }
 657                         nfs_lock_request(req);
 658                         req->wb_index = pos >> PAGE_SHIFT;
 659                         req->wb_offset = pos & ~PAGE_MASK;
 660                         if (!nfs_pageio_add_request(desc, req)) {
 661                                 result = desc->pg_error;
 662                                 nfs_unlock_and_release_request(req);
 663                                 break;
 664                         }
 665                         pgbase = 0;
 666                         bytes -= req_len;
 667                         started += req_len;
 668                         user_addr += req_len;
 669                         pos += req_len;
 670                         count -= req_len;
 671                 }
 672                 /* The nfs_page now hold references to these pages */
 673                 nfs_direct_release_pages(pagevec, npages);
 674         } while (count != 0 && result >= 0);
 675
 676         kfree(pagevec);
 677
 678         if (started)
 679                 return started;
 680         return result < 0 ? (ssize_t) result : -EFAULT;
 681 }
 682
 683 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 684 {
 685         struct nfs_direct_req *dreq = hdr->dreq;
 686         struct nfs_commit_info cinfo;
 687         int bit = -1;
 688         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 689
 690         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 691                 goto out_put;
 692
 693         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 694
 695         spin_lock(&dreq->lock);
 696
 697         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 698                 dreq->flags = 0;
 699                 dreq->error = hdr->error;
 700         }
 701         if (dreq->error != 0)
 702                 bit = NFS_IOHDR_ERROR;
 703         else {
 704                 dreq->count += hdr->good_bytes;
 705                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 706                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 707                         bit = NFS_IOHDR_NEED_RESCHED;
 708                 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 709                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 710                                 bit = NFS_IOHDR_NEED_RESCHED;
 711                         else if (dreq->flags == 0) {
 712                                 memcpy(&dreq->verf, &req->wb_verf,
 713                                        sizeof(dreq->verf));
 714                                 bit = NFS_IOHDR_NEED_COMMIT;
 715                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 716                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 717                                 if (memcmp(&dreq->verf, &req->wb_verf, sizeof(dreq->verf))) {
 718                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 719                                         bit = NFS_IOHDR_NEED_RESCHED;
 720                                 } else
 721                                         bit = NFS_IOHDR_NEED_COMMIT;
 722                         }
 723                 }
 724         }
 725         spin_unlock(&dreq->lock);
 726
 727         while (!list_empty(&hdr->pages)) {
 728                 req = nfs_list_entry(hdr->pages.next);
 729                 nfs_list_remove_request(req);
 730                 switch (bit) {
 731                 case NFS_IOHDR_NEED_RESCHED:
 732                 case NFS_IOHDR_NEED_COMMIT:
 733                         kref_get(&req->wb_kref);
 734                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 735                 }
 736                 nfs_unlock_and_release_request(req);
 737         }
 738
 739 out_put:
 740         if (put_dreq(dreq))
 741                 nfs_direct_write_complete(dreq, hdr->inode);
 742         hdr->release(hdr);
 743 }
 744
 745 static void nfs_write_sync_pgio_error(struct list_head *head)
 746 {
 747         struct nfs_page *req;
 748
 749         while (!list_empty(head)) {
 750                 req = nfs_list_entry(head->next);
 751                 nfs_list_remove_request(req);
 752                 nfs_unlock_and_release_request(req);
 753         }
 754 }
 755
 756 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 757         .error_cleanup = nfs_write_sync_pgio_error,
 758         .init_hdr = nfs_direct_pgio_init,
 759         .completion = nfs_direct_write_completion,
 760 };
 761
 762 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 763                                                const struct iovec *iov,
 764                                                unsigned long nr_segs,
 765                                                loff_t pos)
 766 {
 767         struct nfs_pageio_descriptor desc;
 768         ssize_t result = 0;
 769         size_t requested_bytes = 0;
 770         unsigned long seg;
 771
 772         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_COND_STABLE,
 773                               &nfs_direct_write_completion_ops);
 774         desc.pg_dreq = dreq;
 775         get_dreq(dreq);
 776
 777         for (seg = 0; seg < nr_segs; seg++) {
 778                 const struct iovec *vec = &iov[seg];
 779                 result = nfs_direct_write_schedule_segment(&desc, vec, pos);
 780                 if (result < 0)
 781                         break;
 782                 requested_bytes += result;
 783                 if ((size_t)result < vec->iov_len)
 784                         break;
 785                 pos += vec->iov_len;
 786         }
 787         nfs_pageio_complete(&desc);
 788
 789         /*
 790          * If no bytes were started, return the error, and let the
 791          * generic layer handle the completion.
 792          */
 793         if (requested_bytes == 0) {
 794                 nfs_direct_req_release(dreq);
 795                 return result < 0 ? result : -EIO;
 796         }
 797
 798         if (put_dreq(dreq))
 799                 nfs_direct_write_complete(dreq, dreq->inode);
 800         return 0;
 801 }
 802
 803 static ssize_t nfs_direct_write(struct kiocb *iocb, const struct iovec *iov,
 804                                 unsigned long nr_segs, loff_t pos,
 805                                 size_t count)
 806 {
 807         ssize_t result = -ENOMEM;
 808         struct inode *inode = iocb->ki_filp->f_mapping->host;
 809         struct nfs_direct_req *dreq;
 810
 811         dreq = nfs_direct_req_alloc();
 812         if (!dreq)
 813                 goto out;
 814
 815         dreq->inode = inode;
 816         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 817         dreq->l_ctx = nfs_get_lock_context(dreq->ctx);
 818         if (dreq->l_ctx == NULL)
 819                 goto out_release;
 820         if (!is_sync_kiocb(iocb))
 821                 dreq->iocb = iocb;
 822
 823         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos);
 824         if (!result)
 825                 result = nfs_direct_wait(dreq);
 826 out_release:
 827         nfs_direct_req_release(dreq);
 828 out:
 829         return result;
 830 }
 831
 832 /**
 833  * nfs_file_direct_read - file direct read operation for NFS files
 834  * @iocb: target I/O control block
 835  * @iov: vector of user buffers into which to read data
 836  * @nr_segs: size of iov vector
 837  * @pos: byte offset in file where reading starts
 838  *
 839  * We use this function for direct reads instead of calling
 840  * generic_file_aio_read() in order to avoid gfar's check to see if
 841  * the request starts before the end of the file.  For that check
 842  * to work, we must generate a GETATTR before each direct read, and
 843  * even then there is a window between the GETATTR and the subsequent
 844  * READ where the file size could change.  Our preference is simply
 845  * to do all reads the application wants, and the server will take
 846  * care of managing the end of file boundary.
 847  *
 848  * This function also eliminates unnecessarily updating the file's
 849  * atime locally, as the NFS server sets the file's atime, and this
 850  * client must read the updated atime from the server back into its
 851  * cache.
 852  */
 853 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 854                                 unsigned long nr_segs, loff_t pos)
 855 {
 856         ssize_t retval = -EINVAL;
 857         struct file *file = iocb->ki_filp;
 858         struct address_space *mapping = file->f_mapping;
 859         size_t count;
 860
 861         count = iov_length(iov, nr_segs);
 862         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 863
 864         dfprintk(FILE, "NFS: direct read(%s/%s, %zd@%Ld)\n",
 865                 file->f_path.dentry->d_parent->d_name.name,
 866                 file->f_path.dentry->d_name.name,
 867                 count, (long long) pos);
 868
 869         retval = 0;
 870         if (!count)
 871                 goto out;
 872
 873         retval = nfs_sync_mapping(mapping);
 874         if (retval)
 875                 goto out;
 876
 877         task_io_account_read(count);
 878
 879         retval = nfs_direct_read(iocb, iov, nr_segs, pos);
 880         if (retval > 0)
 881                 iocb->ki_pos = pos + retval;
 882
 883 out:
 884         return retval;
 885 }
 886
 887 /**
 888  * nfs_file_direct_write - file direct write operation for NFS files
 889  * @iocb: target I/O control block
 890  * @iov: vector of user buffers from which to write data
 891  * @nr_segs: size of iov vector
 892  * @pos: byte offset in file where writing starts
 893  *
 894  * We use this function for direct writes instead of calling
 895  * generic_file_aio_write() in order to avoid taking the inode
 896  * semaphore and updating the i_size.  The NFS server will set
 897  * the new i_size and this client must read the updated size
 898  * back into its cache.  We let the server do generic write
 899  * parameter checking and report problems.
 900  *
 901  * We eliminate local atime updates, see direct read above.
 902  *
 903  * We avoid unnecessary page cache invalidations for normal cached
 904  * readers of this file.
 905  *
 906  * Note that O_APPEND is not supported for NFS direct writes, as there
 907  * is no atomic O_APPEND write facility in the NFS protocol.
 908  */
 909 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
 910                                 unsigned long nr_segs, loff_t pos)
 911 {
 912         ssize_t retval = -EINVAL;
 913         struct file *file = iocb->ki_filp;
 914         struct address_space *mapping = file->f_mapping;
 915         size_t count;
 916
 917         count = iov_length(iov, nr_segs);
 918         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 919
 920         dfprintk(FILE, "NFS: direct write(%s/%s, %zd@%Ld)\n",
 921                 file->f_path.dentry->d_parent->d_name.name,
 922                 file->f_path.dentry->d_name.name,
 923                 count, (long long) pos);
 924
 925         retval = generic_write_checks(file, &pos, &count, 0);
 926         if (retval)
 927                 goto out;
 928
 929         retval = -EINVAL;
 930         if ((ssize_t) count < 0)
 931                 goto out;
 932         retval = 0;
 933         if (!count)
 934                 goto out;
 935
 936         retval = nfs_sync_mapping(mapping);
 937         if (retval)
 938                 goto out;
 939
 940         task_io_account_write(count);
 941
 942         retval = nfs_direct_write(iocb, iov, nr_segs, pos, count);
 943         if (retval > 0) {
 944                 struct inode *inode = mapping->host;
 945
 946                 iocb->ki_pos = pos + retval;
 947                 spin_lock(&inode->i_lock);
 948                 if (i_size_read(inode) < iocb->ki_pos)
 949                         i_size_write(inode, iocb->ki_pos);
 950                 spin_unlock(&inode->i_lock);
 951         }
 952 out:
 953         return retval;
 954 }
 955
 956 /**
 957  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 958  *
 959  */
 960 int __init nfs_init_directcache(void)
 961 {
 962         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 963                                                 sizeof(struct nfs_direct_req),
 964                                                 0, (SLAB_RECLAIM_ACCOUNT|
 965                                                         SLAB_MEM_SPREAD),
 966                                                 NULL);
 967         if (nfs_direct_cachep == NULL)
 968                 return -ENOMEM;
 969
 970         return 0;
 971 }
 972
 973 /**
 974  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 975  *
 976  */
 977 void nfs_destroy_directcache(void)
 978 {
 979         kmem_cache_destroy(nfs_direct_cachep);
 980 }