fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <[email protected]>
   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->data->ds_clp,
 152                                       hdr->data->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->data->ds_clp,
 173                                          hdr->data->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 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 182 /*
 183  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 184  * @dreq - direct request possibly spanning multiple servers
 185  * @data - commit data to validate against previously seen verf
 186  *
 187  * returns result of comparison between @data->verf and the verf of
 188  * the server used by @data (DS or MDS)
 189  */
 190 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 191                                            struct nfs_commit_data *data)
 192 {
 193         struct nfs_writeverf *verfp;
 194
 195         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 196                                          data->ds_commit_index);
 197         WARN_ON_ONCE(verfp->committed < 0);
 198         return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
 199 }
 200 #endif
 201
 202 /**
 203  * nfs_direct_IO - NFS address space operation for direct I/O
 204  * @rw: direction (read or write)
 205  * @iocb: target I/O control block
 206  * @iov: array of vectors that define I/O buffer
 207  * @pos: offset in file to begin the operation
 208  * @nr_segs: size of iovec array
 209  *
 210  * The presence of this routine in the address space ops vector means
 211  * the NFS client supports direct I/O. However, for most direct IO, we
 212  * shunt off direct read and write requests before the VFS gets them,
 213  * so this method is only ever called for swap.
 214  */
 215 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter, loff_t pos)
 216 {
 217 #ifndef CONFIG_NFS_SWAP
 218         dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
 219                         iocb->ki_filp, (long long) pos, iter->nr_segs);
 220
 221         return -EINVAL;
 222 #else
 223         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 224
 225         if (rw == READ || rw == KERNEL_READ)
 226                 return nfs_file_direct_read(iocb, iter, pos,
 227                                 rw == READ ? true : false);
 228         return nfs_file_direct_write(iocb, iter, pos,
 229                                 rw == WRITE ? true : false);
 230 #endif /* CONFIG_NFS_SWAP */
 231 }
 232
 233 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 234 {
 235         unsigned int i;
 236         for (i = 0; i < npages; i++)
 237                 page_cache_release(pages[i]);
 238 }
 239
 240 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 241                               struct nfs_direct_req *dreq)
 242 {
 243         cinfo->lock = &dreq->lock;
 244         cinfo->mds = &dreq->mds_cinfo;
 245         cinfo->ds = &dreq->ds_cinfo;
 246         cinfo->dreq = dreq;
 247         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 248 }
 249
 250 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 251 {
 252         struct nfs_direct_req *dreq;
 253
 254         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 255         if (!dreq)
 256                 return NULL;
 257
 258         kref_init(&dreq->kref);
 259         kref_get(&dreq->kref);
 260         init_completion(&dreq->completion);
 261         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 262         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 263         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 264         spin_lock_init(&dreq->lock);
 265
 266         return dreq;
 267 }
 268
 269 static void nfs_direct_req_free(struct kref *kref)
 270 {
 271         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 272
 273         if (dreq->l_ctx != NULL)
 274                 nfs_put_lock_context(dreq->l_ctx);
 275         if (dreq->ctx != NULL)
 276                 put_nfs_open_context(dreq->ctx);
 277         kmem_cache_free(nfs_direct_cachep, dreq);
 278 }
 279
 280 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 281 {
 282         kref_put(&dreq->kref, nfs_direct_req_free);
 283 }
 284
 285 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 286 {
 287         return dreq->bytes_left;
 288 }
 289 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 290
 291 /*
 292  * Collects and returns the final error value/byte-count.
 293  */
 294 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 295 {
 296         ssize_t result = -EIOCBQUEUED;
 297
 298         /* Async requests don't wait here */
 299         if (dreq->iocb)
 300                 goto out;
 301
 302         result = wait_for_completion_killable(&dreq->completion);
 303
 304         if (!result)
 305                 result = dreq->error;
 306         if (!result)
 307                 result = dreq->count;
 308
 309 out:
 310         return (ssize_t) result;
 311 }
 312
 313 /*
 314  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 315  * the iocb is still valid here if this is a synchronous request.
 316  */
 317 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 318 {
 319         struct inode *inode = dreq->inode;
 320
 321         if (dreq->iocb && write) {
 322                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 323
 324                 spin_lock(&inode->i_lock);
 325                 if (i_size_read(inode) < pos)
 326                         i_size_write(inode, pos);
 327                 spin_unlock(&inode->i_lock);
 328         }
 329
 330         if (write)
 331                 nfs_zap_mapping(inode, inode->i_mapping);
 332
 333         inode_dio_done(inode);
 334
 335         if (dreq->iocb) {
 336                 long res = (long) dreq->error;
 337                 if (!res)
 338                         res = (long) dreq->count;
 339                 aio_complete(dreq->iocb, res, 0);
 340         }
 341
 342         complete_all(&dreq->completion);
 343
 344         nfs_direct_req_release(dreq);
 345 }
 346
 347 static void nfs_direct_readpage_release(struct nfs_page *req)
 348 {
 349         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 350                 req->wb_context->dentry->d_inode->i_sb->s_id,
 351                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 352                 req->wb_bytes,
 353                 (long long)req_offset(req));
 354         nfs_release_request(req);
 355 }
 356
 357 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 358 {
 359         unsigned long bytes = 0;
 360         struct nfs_direct_req *dreq = hdr->dreq;
 361
 362         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 363                 goto out_put;
 364
 365         spin_lock(&dreq->lock);
 366         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 367                 dreq->error = hdr->error;
 368         else
 369                 dreq->count += hdr->good_bytes;
 370         spin_unlock(&dreq->lock);
 371
 372         while (!list_empty(&hdr->pages)) {
 373                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 374                 struct page *page = req->wb_page;
 375
 376                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 377                         set_page_dirty(page);
 378                 bytes += req->wb_bytes;
 379                 nfs_list_remove_request(req);
 380                 nfs_direct_readpage_release(req);
 381         }
 382 out_put:
 383         if (put_dreq(dreq))
 384                 nfs_direct_complete(dreq, false);
 385         hdr->release(hdr);
 386 }
 387
 388 static void nfs_read_sync_pgio_error(struct list_head *head)
 389 {
 390         struct nfs_page *req;
 391
 392         while (!list_empty(head)) {
 393                 req = nfs_list_entry(head->next);
 394                 nfs_list_remove_request(req);
 395                 nfs_release_request(req);
 396         }
 397 }
 398
 399 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 400 {
 401         get_dreq(hdr->dreq);
 402 }
 403
 404 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 405         .error_cleanup = nfs_read_sync_pgio_error,
 406         .init_hdr = nfs_direct_pgio_init,
 407         .completion = nfs_direct_read_completion,
 408 };
 409
 410 /*
 411  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 412  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 413  * bail and stop sending more reads.  Read length accounting is
 414  * handled automatically by nfs_direct_read_result().  Otherwise, if
 415  * no requests have been sent, just return an error.
 416  */
 417
 418 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 419                                               struct iov_iter *iter,
 420                                               loff_t pos)
 421 {
 422         struct nfs_pageio_descriptor desc;
 423         struct inode *inode = dreq->inode;
 424         ssize_t result = -EINVAL;
 425         size_t requested_bytes = 0;
 426         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 427
 428         nfs_pageio_init_read(&desc, dreq->inode, false,
 429                              &nfs_direct_read_completion_ops);
 430         get_dreq(dreq);
 431         desc.pg_dreq = dreq;
 432         atomic_inc(&inode->i_dio_count);
 433
 434         while (iov_iter_count(iter)) {
 435                 struct page **pagevec;
 436                 size_t bytes;
 437                 size_t pgbase;
 438                 unsigned npages, i;
 439
 440                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 441                                                   rsize, &pgbase);
 442                 if (result < 0)
 443                         break;
 444
 445                 bytes = result;
 446                 iov_iter_advance(iter, bytes);
 447                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 448                 for (i = 0; i < npages; i++) {
 449                         struct nfs_page *req;
 450                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 451                         /* XXX do we need to do the eof zeroing found in async_filler? */
 452                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 453                                                  pgbase, req_len);
 454                         if (IS_ERR(req)) {
 455                                 result = PTR_ERR(req);
 456                                 break;
 457                         }
 458                         req->wb_index = pos >> PAGE_SHIFT;
 459                         req->wb_offset = pos & ~PAGE_MASK;
 460                         if (!nfs_pageio_add_request(&desc, req)) {
 461                                 result = desc.pg_error;
 462                                 nfs_release_request(req);
 463                                 break;
 464                         }
 465                         pgbase = 0;
 466                         bytes -= req_len;
 467                         requested_bytes += req_len;
 468                         pos += req_len;
 469                         dreq->bytes_left -= req_len;
 470                 }
 471                 nfs_direct_release_pages(pagevec, npages);
 472                 kvfree(pagevec);
 473                 if (result < 0)
 474                         break;
 475         }
 476
 477         nfs_pageio_complete(&desc);
 478
 479         /*
 480          * If no bytes were started, return the error, and let the
 481          * generic layer handle the completion.
 482          */
 483         if (requested_bytes == 0) {
 484                 inode_dio_done(inode);
 485                 nfs_direct_req_release(dreq);
 486                 return result < 0 ? result : -EIO;
 487         }
 488
 489         if (put_dreq(dreq))
 490                 nfs_direct_complete(dreq, false);
 491         return 0;
 492 }
 493
 494 /**
 495  * nfs_file_direct_read - file direct read operation for NFS files
 496  * @iocb: target I/O control block
 497  * @iter: vector of user buffers into which to read data
 498  * @pos: byte offset in file where reading starts
 499  *
 500  * We use this function for direct reads instead of calling
 501  * generic_file_aio_read() in order to avoid gfar's check to see if
 502  * the request starts before the end of the file.  For that check
 503  * to work, we must generate a GETATTR before each direct read, and
 504  * even then there is a window between the GETATTR and the subsequent
 505  * READ where the file size could change.  Our preference is simply
 506  * to do all reads the application wants, and the server will take
 507  * care of managing the end of file boundary.
 508  *
 509  * This function also eliminates unnecessarily updating the file's
 510  * atime locally, as the NFS server sets the file's atime, and this
 511  * client must read the updated atime from the server back into its
 512  * cache.
 513  */
 514 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 515                                 loff_t pos, bool uio)
 516 {
 517         struct file *file = iocb->ki_filp;
 518         struct address_space *mapping = file->f_mapping;
 519         struct inode *inode = mapping->host;
 520         struct nfs_direct_req *dreq;
 521         struct nfs_lock_context *l_ctx;
 522         ssize_t result = -EINVAL;
 523         size_t count = iov_iter_count(iter);
 524         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 525
 526         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 527                 file, count, (long long) pos);
 528
 529         result = 0;
 530         if (!count)
 531                 goto out;
 532
 533         mutex_lock(&inode->i_mutex);
 534         result = nfs_sync_mapping(mapping);
 535         if (result)
 536                 goto out_unlock;
 537
 538         task_io_account_read(count);
 539
 540         result = -ENOMEM;
 541         dreq = nfs_direct_req_alloc();
 542         if (dreq == NULL)
 543                 goto out_unlock;
 544
 545         dreq->inode = inode;
 546         dreq->bytes_left = count;
 547         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 548         l_ctx = nfs_get_lock_context(dreq->ctx);
 549         if (IS_ERR(l_ctx)) {
 550                 result = PTR_ERR(l_ctx);
 551                 goto out_release;
 552         }
 553         dreq->l_ctx = l_ctx;
 554         if (!is_sync_kiocb(iocb))
 555                 dreq->iocb = iocb;
 556
 557         NFS_I(inode)->read_io += count;
 558         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 559
 560         mutex_unlock(&inode->i_mutex);
 561
 562         if (!result) {
 563                 result = nfs_direct_wait(dreq);
 564                 if (result > 0)
 565                         iocb->ki_pos = pos + result;
 566         }
 567
 568         nfs_direct_req_release(dreq);
 569         return result;
 570
 571 out_release:
 572         nfs_direct_req_release(dreq);
 573 out_unlock:
 574         mutex_unlock(&inode->i_mutex);
 575 out:
 576         return result;
 577 }
 578
 579 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 580 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 581 {
 582         struct nfs_pageio_descriptor desc;
 583         struct nfs_page *req, *tmp;
 584         LIST_HEAD(reqs);
 585         struct nfs_commit_info cinfo;
 586         LIST_HEAD(failed);
 587
 588         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 589         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 590         spin_lock(cinfo.lock);
 591         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 592         spin_unlock(cinfo.lock);
 593
 594         dreq->count = 0;
 595         get_dreq(dreq);
 596
 597         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 598                               &nfs_direct_write_completion_ops);
 599         desc.pg_dreq = dreq;
 600
 601         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 602                 if (!nfs_pageio_add_request(&desc, req)) {
 603                         nfs_list_remove_request(req);
 604                         nfs_list_add_request(req, &failed);
 605                         spin_lock(cinfo.lock);
 606                         dreq->flags = 0;
 607                         dreq->error = -EIO;
 608                         spin_unlock(cinfo.lock);
 609                 }
 610                 nfs_release_request(req);
 611         }
 612         nfs_pageio_complete(&desc);
 613
 614         while (!list_empty(&failed)) {
 615                 req = nfs_list_entry(failed.next);
 616                 nfs_list_remove_request(req);
 617                 nfs_unlock_and_release_request(req);
 618         }
 619
 620         if (put_dreq(dreq))
 621                 nfs_direct_write_complete(dreq, dreq->inode);
 622 }
 623
 624 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 625 {
 626         struct nfs_direct_req *dreq = data->dreq;
 627         struct nfs_commit_info cinfo;
 628         struct nfs_page *req;
 629         int status = data->task.tk_status;
 630
 631         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 632         if (status < 0) {
 633                 dprintk("NFS: %5u commit failed with error %d.\n",
 634                         data->task.tk_pid, status);
 635                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 636         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 637                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 638                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 639         }
 640
 641         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 642         while (!list_empty(&data->pages)) {
 643                 req = nfs_list_entry(data->pages.next);
 644                 nfs_list_remove_request(req);
 645                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 646                         /* Note the rewrite will go through mds */
 647                         nfs_mark_request_commit(req, NULL, &cinfo);
 648                 } else
 649                         nfs_release_request(req);
 650                 nfs_unlock_and_release_request(req);
 651         }
 652
 653         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 654                 nfs_direct_write_complete(dreq, data->inode);
 655 }
 656
 657 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 658 {
 659         /* There is no lock to clear */
 660 }
 661
 662 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 663         .completion = nfs_direct_commit_complete,
 664         .error_cleanup = nfs_direct_error_cleanup,
 665 };
 666
 667 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 668 {
 669         int res;
 670         struct nfs_commit_info cinfo;
 671         LIST_HEAD(mds_list);
 672
 673         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 674         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 675         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 676         if (res < 0) /* res == -ENOMEM */
 677                 nfs_direct_write_reschedule(dreq);
 678 }
 679
 680 static void nfs_direct_write_schedule_work(struct work_struct *work)
 681 {
 682         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 683         int flags = dreq->flags;
 684
 685         dreq->flags = 0;
 686         switch (flags) {
 687                 case NFS_ODIRECT_DO_COMMIT:
 688                         nfs_direct_commit_schedule(dreq);
 689                         break;
 690                 case NFS_ODIRECT_RESCHED_WRITES:
 691                         nfs_direct_write_reschedule(dreq);
 692                         break;
 693                 default:
 694                         nfs_direct_complete(dreq, true);
 695         }
 696 }
 697
 698 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 699 {
 700         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 701 }
 702
 703 #else
 704 static void nfs_direct_write_schedule_work(struct work_struct *work)
 705 {
 706 }
 707
 708 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 709 {
 710         nfs_direct_complete(dreq, true);
 711 }
 712 #endif
 713
 714 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 715 {
 716         struct nfs_direct_req *dreq = hdr->dreq;
 717         struct nfs_commit_info cinfo;
 718         int bit = -1;
 719         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 720
 721         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 722                 goto out_put;
 723
 724         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 725
 726         spin_lock(&dreq->lock);
 727
 728         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 729                 dreq->flags = 0;
 730                 dreq->error = hdr->error;
 731         }
 732         if (dreq->error != 0)
 733                 bit = NFS_IOHDR_ERROR;
 734         else {
 735                 dreq->count += hdr->good_bytes;
 736                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 737                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 738                         bit = NFS_IOHDR_NEED_RESCHED;
 739                 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 740                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 741                                 bit = NFS_IOHDR_NEED_RESCHED;
 742                         else if (dreq->flags == 0) {
 743                                 nfs_direct_set_hdr_verf(dreq, hdr);
 744                                 bit = NFS_IOHDR_NEED_COMMIT;
 745                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 746                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 747                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr)) {
 748                                         dreq->flags =
 749                                                 NFS_ODIRECT_RESCHED_WRITES;
 750                                         bit = NFS_IOHDR_NEED_RESCHED;
 751                                 } else
 752                                         bit = NFS_IOHDR_NEED_COMMIT;
 753                         }
 754                 }
 755         }
 756         spin_unlock(&dreq->lock);
 757
 758         while (!list_empty(&hdr->pages)) {
 759                 bool do_destroy = true;
 760
 761                 req = nfs_list_entry(hdr->pages.next);
 762                 nfs_list_remove_request(req);
 763                 switch (bit) {
 764                 case NFS_IOHDR_NEED_RESCHED:
 765                 case NFS_IOHDR_NEED_COMMIT:
 766                         kref_get(&req->wb_kref);
 767                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 768                         do_destroy = false;
 769                 }
 770                 nfs_unlock_and_release_request(req);
 771         }
 772
 773 out_put:
 774         if (put_dreq(dreq))
 775                 nfs_direct_write_complete(dreq, hdr->inode);
 776         hdr->release(hdr);
 777 }
 778
 779 static void nfs_write_sync_pgio_error(struct list_head *head)
 780 {
 781         struct nfs_page *req;
 782
 783         while (!list_empty(head)) {
 784                 req = nfs_list_entry(head->next);
 785                 nfs_list_remove_request(req);
 786                 nfs_unlock_and_release_request(req);
 787         }
 788 }
 789
 790 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 791         .error_cleanup = nfs_write_sync_pgio_error,
 792         .init_hdr = nfs_direct_pgio_init,
 793         .completion = nfs_direct_write_completion,
 794 };
 795
 796
 797 /*
 798  * NB: Return the value of the first error return code.  Subsequent
 799  *     errors after the first one are ignored.
 800  */
 801 /*
 802  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 803  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 804  * bail and stop sending more writes.  Write length accounting is
 805  * handled automatically by nfs_direct_write_result().  Otherwise, if
 806  * no requests have been sent, just return an error.
 807  */
 808 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 809                                                struct iov_iter *iter,
 810                                                loff_t pos)
 811 {
 812         struct nfs_pageio_descriptor desc;
 813         struct inode *inode = dreq->inode;
 814         ssize_t result = 0;
 815         size_t requested_bytes = 0;
 816         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 817
 818         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 819                               &nfs_direct_write_completion_ops);
 820         desc.pg_dreq = dreq;
 821         get_dreq(dreq);
 822         atomic_inc(&inode->i_dio_count);
 823
 824         NFS_I(inode)->write_io += iov_iter_count(iter);
 825         while (iov_iter_count(iter)) {
 826                 struct page **pagevec;
 827                 size_t bytes;
 828                 size_t pgbase;
 829                 unsigned npages, i;
 830
 831                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 832                                                   wsize, &pgbase);
 833                 if (result < 0)
 834                         break;
 835
 836                 bytes = result;
 837                 iov_iter_advance(iter, bytes);
 838                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 839                 for (i = 0; i < npages; i++) {
 840                         struct nfs_page *req;
 841                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 842
 843                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 844                                                  pgbase, req_len);
 845                         if (IS_ERR(req)) {
 846                                 result = PTR_ERR(req);
 847                                 break;
 848                         }
 849                         nfs_lock_request(req);
 850                         req->wb_index = pos >> PAGE_SHIFT;
 851                         req->wb_offset = pos & ~PAGE_MASK;
 852                         if (!nfs_pageio_add_request(&desc, req)) {
 853                                 result = desc.pg_error;
 854                                 nfs_unlock_and_release_request(req);
 855                                 break;
 856                         }
 857                         pgbase = 0;
 858                         bytes -= req_len;
 859                         requested_bytes += req_len;
 860                         pos += req_len;
 861                         dreq->bytes_left -= req_len;
 862                 }
 863                 nfs_direct_release_pages(pagevec, npages);
 864                 kvfree(pagevec);
 865                 if (result < 0)
 866                         break;
 867         }
 868         nfs_pageio_complete(&desc);
 869
 870         /*
 871          * If no bytes were started, return the error, and let the
 872          * generic layer handle the completion.
 873          */
 874         if (requested_bytes == 0) {
 875                 inode_dio_done(inode);
 876                 nfs_direct_req_release(dreq);
 877                 return result < 0 ? result : -EIO;
 878         }
 879
 880         if (put_dreq(dreq))
 881                 nfs_direct_write_complete(dreq, dreq->inode);
 882         return 0;
 883 }
 884
 885 /**
 886  * nfs_file_direct_write - file direct write operation for NFS files
 887  * @iocb: target I/O control block
 888  * @iter: vector of user buffers from which to write data
 889  * @pos: byte offset in file where writing starts
 890  *
 891  * We use this function for direct writes instead of calling
 892  * generic_file_aio_write() in order to avoid taking the inode
 893  * semaphore and updating the i_size.  The NFS server will set
 894  * the new i_size and this client must read the updated size
 895  * back into its cache.  We let the server do generic write
 896  * parameter checking and report problems.
 897  *
 898  * We eliminate local atime updates, see direct read above.
 899  *
 900  * We avoid unnecessary page cache invalidations for normal cached
 901  * readers of this file.
 902  *
 903  * Note that O_APPEND is not supported for NFS direct writes, as there
 904  * is no atomic O_APPEND write facility in the NFS protocol.
 905  */
 906 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 907                                 loff_t pos, bool uio)
 908 {
 909         ssize_t result = -EINVAL;
 910         struct file *file = iocb->ki_filp;
 911         struct address_space *mapping = file->f_mapping;
 912         struct inode *inode = mapping->host;
 913         struct nfs_direct_req *dreq;
 914         struct nfs_lock_context *l_ctx;
 915         loff_t end;
 916         size_t count = iov_iter_count(iter);
 917         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 918
 919         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 920
 921         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 922                 file, count, (long long) pos);
 923
 924         result = generic_write_checks(file, &pos, &count, 0);
 925         if (result)
 926                 goto out;
 927
 928         result = -EINVAL;
 929         if ((ssize_t) count < 0)
 930                 goto out;
 931         result = 0;
 932         if (!count)
 933                 goto out;
 934
 935         mutex_lock(&inode->i_mutex);
 936
 937         result = nfs_sync_mapping(mapping);
 938         if (result)
 939                 goto out_unlock;
 940
 941         if (mapping->nrpages) {
 942                 result = invalidate_inode_pages2_range(mapping,
 943                                         pos >> PAGE_CACHE_SHIFT, end);
 944                 if (result)
 945                         goto out_unlock;
 946         }
 947
 948         task_io_account_write(count);
 949
 950         result = -ENOMEM;
 951         dreq = nfs_direct_req_alloc();
 952         if (!dreq)
 953                 goto out_unlock;
 954
 955         dreq->inode = inode;
 956         dreq->bytes_left = count;
 957         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 958         l_ctx = nfs_get_lock_context(dreq->ctx);
 959         if (IS_ERR(l_ctx)) {
 960                 result = PTR_ERR(l_ctx);
 961                 goto out_release;
 962         }
 963         dreq->l_ctx = l_ctx;
 964         if (!is_sync_kiocb(iocb))
 965                 dreq->iocb = iocb;
 966
 967         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
 968
 969         if (mapping->nrpages) {
 970                 invalidate_inode_pages2_range(mapping,
 971                                               pos >> PAGE_CACHE_SHIFT, end);
 972         }
 973
 974         mutex_unlock(&inode->i_mutex);
 975
 976         if (!result) {
 977                 result = nfs_direct_wait(dreq);
 978                 if (result > 0) {
 979                         struct inode *inode = mapping->host;
 980
 981                         iocb->ki_pos = pos + result;
 982                         spin_lock(&inode->i_lock);
 983                         if (i_size_read(inode) < iocb->ki_pos)
 984                                 i_size_write(inode, iocb->ki_pos);
 985                         spin_unlock(&inode->i_lock);
 986                 }
 987         }
 988         nfs_direct_req_release(dreq);
 989         return result;
 990
 991 out_release:
 992         nfs_direct_req_release(dreq);
 993 out_unlock:
 994         mutex_unlock(&inode->i_mutex);
 995 out:
 996         return result;
 997 }
 998
 999 /**
1000  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1001  *
1002  */
1003 int __init nfs_init_directcache(void)
1004 {
1005         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1006                                                 sizeof(struct nfs_direct_req),
1007                                                 0, (SLAB_RECLAIM_ACCOUNT|
1008                                                         SLAB_MEM_SPREAD),
1009                                                 NULL);
1010         if (nfs_direct_cachep == NULL)
1011                 return -ENOMEM;
1012
1013         return 0;
1014 }
1015
1016 /**
1017  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1018  *
1019  */
1020 void nfs_destroy_directcache(void)
1021 {
1022         kmem_cache_destroy(nfs_direct_cachep);
1023 }