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 <linux/uaccess.h>
  56 #include <linux/atomic.h>
  57
  58 #include "internal.h"
  59 #include "iostat.h"
  60 #include "pnfs.h"
  61
  62 #define NFSDBG_FACILITY         NFSDBG_VFS
  63
  64 static struct kmem_cache *nfs_direct_cachep;
  65
  66 /*
  67  * This represents a set of asynchronous requests that we're waiting on
  68  */
  69 struct nfs_direct_mirror {
  70         ssize_t count;
  71 };
  72
  73 struct nfs_direct_req {
  74         struct kref             kref;           /* release manager */
  75
  76         /* I/O parameters */
  77         struct nfs_open_context *ctx;           /* file open context info */
  78         struct nfs_lock_context *l_ctx;         /* Lock context info */
  79         struct kiocb *          iocb;           /* controlling i/o request */
  80         struct inode *          inode;          /* target file of i/o */
  81
  82         /* completion state */
  83         atomic_t                io_count;       /* i/os we're waiting for */
  84         spinlock_t              lock;           /* protect completion state */
  85
  86         struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
  87         int                     mirror_count;
  88
  89         ssize_t                 count,          /* bytes actually processed */
  90                                 max_count,      /* max expected count */
  91                                 bytes_left,     /* bytes left to be sent */
  92                                 io_start,       /* start of IO */
  93                                 error;          /* any reported error */
  94         struct completion       completion;     /* wait for i/o completion */
  95
  96         /* commit state */
  97         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  98         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  99         struct work_struct      work;
 100         int                     flags;
 101 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
 102 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
 103         struct nfs_writeverf    verf;           /* unstable write verifier */
 104 };
 105
 106 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 107 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 108 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
 109 static void nfs_direct_write_schedule_work(struct work_struct *work);
 110
 111 static inline void get_dreq(struct nfs_direct_req *dreq)
 112 {
 113         atomic_inc(&dreq->io_count);
 114 }
 115
 116 static inline int put_dreq(struct nfs_direct_req *dreq)
 117 {
 118         return atomic_dec_and_test(&dreq->io_count);
 119 }
 120
 121 static void
 122 nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
 123 {
 124         int i;
 125         ssize_t count;
 126
 127         WARN_ON_ONCE(dreq->count >= dreq->max_count);
 128
 129         if (dreq->mirror_count == 1) {
 130                 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
 131                 dreq->count += hdr->good_bytes;
 132         } else {
 133                 /* mirrored writes */
 134                 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
 135                 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
 136                         count = hdr->io_start + hdr->good_bytes - dreq->io_start;
 137                         dreq->mirrors[hdr->pgio_mirror_idx].count = count;
 138                 }
 139                 /* update the dreq->count by finding the minimum agreed count from all
 140                  * mirrors */
 141                 count = dreq->mirrors[0].count;
 142
 143                 for (i = 1; i < dreq->mirror_count; i++)
 144                         count = min(count, dreq->mirrors[i].count);
 145
 146                 dreq->count = count;
 147         }
 148 }
 149
 150 /*
 151  * nfs_direct_select_verf - select the right verifier
 152  * @dreq - direct request possibly spanning multiple servers
 153  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 154  * @commit_idx - commit bucket index for the DS
 155  *
 156  * returns the correct verifier to use given the role of the server
 157  */
 158 static struct nfs_writeverf *
 159 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 160                        struct nfs_client *ds_clp,
 161                        int commit_idx)
 162 {
 163         struct nfs_writeverf *verfp = &dreq->verf;
 164
 165 #ifdef CONFIG_NFS_V4_1
 166         /*
 167          * pNFS is in use, use the DS verf except commit_through_mds is set
 168          * for layout segment where nbuckets is zero.
 169          */
 170         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 171                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 172                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 173                 else
 174                         WARN_ON_ONCE(1);
 175         }
 176 #endif
 177         return verfp;
 178 }
 179
 180
 181 /*
 182  * nfs_direct_set_hdr_verf - set the write/commit verifier
 183  * @dreq - direct request possibly spanning multiple servers
 184  * @hdr - pageio header to validate against previously seen verfs
 185  *
 186  * Set the server's (MDS or DS) "seen" verifier
 187  */
 188 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 189                                     struct nfs_pgio_header *hdr)
 190 {
 191         struct nfs_writeverf *verfp;
 192
 193         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 194         WARN_ON_ONCE(verfp->committed >= 0);
 195         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 196         WARN_ON_ONCE(verfp->committed < 0);
 197 }
 198
 199 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
 200                 const struct nfs_writeverf *v2)
 201 {
 202         return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
 203 }
 204
 205 /*
 206  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 207  * @dreq - direct request possibly spanning multiple servers
 208  * @hdr - pageio header to validate against previously seen verf
 209  *
 210  * set the server's "seen" verf if not initialized.
 211  * returns result of comparison between @hdr->verf and the "seen"
 212  * verf of the server used by @hdr (DS or MDS)
 213  */
 214 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 215                                           struct nfs_pgio_header *hdr)
 216 {
 217         struct nfs_writeverf *verfp;
 218
 219         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 220         if (verfp->committed < 0) {
 221                 nfs_direct_set_hdr_verf(dreq, hdr);
 222                 return 0;
 223         }
 224         return nfs_direct_cmp_verf(verfp, &hdr->verf);
 225 }
 226
 227 /*
 228  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 229  * @dreq - direct request possibly spanning multiple servers
 230  * @data - commit data to validate against previously seen verf
 231  *
 232  * returns result of comparison between @data->verf and the verf of
 233  * the server used by @data (DS or MDS)
 234  */
 235 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 236                                            struct nfs_commit_data *data)
 237 {
 238         struct nfs_writeverf *verfp;
 239
 240         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 241                                          data->ds_commit_index);
 242
 243         /* verifier not set so always fail */
 244         if (verfp->committed < 0)
 245                 return 1;
 246
 247         return nfs_direct_cmp_verf(verfp, &data->verf);
 248 }
 249
 250 /**
 251  * nfs_direct_IO - NFS address space operation for direct I/O
 252  * @iocb: target I/O control block
 253  * @iter: I/O buffer
 254  *
 255  * The presence of this routine in the address space ops vector means
 256  * the NFS client supports direct I/O. However, for most direct IO, we
 257  * shunt off direct read and write requests before the VFS gets them,
 258  * so this method is only ever called for swap.
 259  */
 260 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 261 {
 262         struct inode *inode = iocb->ki_filp->f_mapping->host;
 263
 264         /* we only support swap file calling nfs_direct_IO */
 265         if (!IS_SWAPFILE(inode))
 266                 return 0;
 267
 268         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 269
 270         if (iov_iter_rw(iter) == READ)
 271                 return nfs_file_direct_read(iocb, iter);
 272         return nfs_file_direct_write(iocb, iter);
 273 }
 274
 275 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 276 {
 277         unsigned int i;
 278         for (i = 0; i < npages; i++)
 279                 put_page(pages[i]);
 280 }
 281
 282 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 283                               struct nfs_direct_req *dreq)
 284 {
 285         cinfo->inode = dreq->inode;
 286         cinfo->mds = &dreq->mds_cinfo;
 287         cinfo->ds = &dreq->ds_cinfo;
 288         cinfo->dreq = dreq;
 289         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 290 }
 291
 292 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 293                                              struct nfs_pageio_descriptor *pgio,
 294                                              struct nfs_page *req)
 295 {
 296         int mirror_count = 1;
 297
 298         if (pgio->pg_ops->pg_get_mirror_count)
 299                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 300
 301         dreq->mirror_count = mirror_count;
 302 }
 303
 304 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 305 {
 306         struct nfs_direct_req *dreq;
 307
 308         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 309         if (!dreq)
 310                 return NULL;
 311
 312         kref_init(&dreq->kref);
 313         kref_get(&dreq->kref);
 314         init_completion(&dreq->completion);
 315         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 316         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 317         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 318         dreq->mirror_count = 1;
 319         spin_lock_init(&dreq->lock);
 320
 321         return dreq;
 322 }
 323
 324 static void nfs_direct_req_free(struct kref *kref)
 325 {
 326         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 327
 328         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 329         if (dreq->l_ctx != NULL)
 330                 nfs_put_lock_context(dreq->l_ctx);
 331         if (dreq->ctx != NULL)
 332                 put_nfs_open_context(dreq->ctx);
 333         kmem_cache_free(nfs_direct_cachep, dreq);
 334 }
 335
 336 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 337 {
 338         kref_put(&dreq->kref, nfs_direct_req_free);
 339 }
 340
 341 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 342 {
 343         return dreq->bytes_left;
 344 }
 345 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 346
 347 /*
 348  * Collects and returns the final error value/byte-count.
 349  */
 350 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 351 {
 352         ssize_t result = -EIOCBQUEUED;
 353
 354         /* Async requests don't wait here */
 355         if (dreq->iocb)
 356                 goto out;
 357
 358         result = wait_for_completion_killable(&dreq->completion);
 359
 360         if (!result) {
 361                 result = dreq->count;
 362                 WARN_ON_ONCE(dreq->count < 0);
 363         }
 364         if (!result)
 365                 result = dreq->error;
 366
 367 out:
 368         return (ssize_t) result;
 369 }
 370
 371 /*
 372  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 373  * the iocb is still valid here if this is a synchronous request.
 374  */
 375 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 376 {
 377         struct inode *inode = dreq->inode;
 378
 379         inode_dio_end(inode);
 380
 381         if (dreq->iocb) {
 382                 long res = (long) dreq->error;
 383                 if (dreq->count != 0) {
 384                         res = (long) dreq->count;
 385                         WARN_ON_ONCE(dreq->count < 0);
 386                 }
 387                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 388         }
 389
 390         complete(&dreq->completion);
 391
 392         nfs_direct_req_release(dreq);
 393 }
 394
 395 static void nfs_direct_readpage_release(struct nfs_page *req)
 396 {
 397         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 398                 req->wb_context->dentry->d_sb->s_id,
 399                 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
 400                 req->wb_bytes,
 401                 (long long)req_offset(req));
 402         nfs_release_request(req);
 403 }
 404
 405 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 406 {
 407         unsigned long bytes = 0;
 408         struct nfs_direct_req *dreq = hdr->dreq;
 409
 410         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 411                 goto out_put;
 412
 413         spin_lock(&dreq->lock);
 414         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 415                 dreq->error = hdr->error;
 416         else
 417                 nfs_direct_good_bytes(dreq, hdr);
 418
 419         spin_unlock(&dreq->lock);
 420
 421         while (!list_empty(&hdr->pages)) {
 422                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 423                 struct page *page = req->wb_page;
 424
 425                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 426                         set_page_dirty(page);
 427                 bytes += req->wb_bytes;
 428                 nfs_list_remove_request(req);
 429                 nfs_direct_readpage_release(req);
 430         }
 431 out_put:
 432         if (put_dreq(dreq))
 433                 nfs_direct_complete(dreq);
 434         hdr->release(hdr);
 435 }
 436
 437 static void nfs_read_sync_pgio_error(struct list_head *head)
 438 {
 439         struct nfs_page *req;
 440
 441         while (!list_empty(head)) {
 442                 req = nfs_list_entry(head->next);
 443                 nfs_list_remove_request(req);
 444                 nfs_release_request(req);
 445         }
 446 }
 447
 448 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 449 {
 450         get_dreq(hdr->dreq);
 451 }
 452
 453 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 454         .error_cleanup = nfs_read_sync_pgio_error,
 455         .init_hdr = nfs_direct_pgio_init,
 456         .completion = nfs_direct_read_completion,
 457 };
 458
 459 /*
 460  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 461  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 462  * bail and stop sending more reads.  Read length accounting is
 463  * handled automatically by nfs_direct_read_result().  Otherwise, if
 464  * no requests have been sent, just return an error.
 465  */
 466
 467 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 468                                               struct iov_iter *iter,
 469                                               loff_t pos)
 470 {
 471         struct nfs_pageio_descriptor desc;
 472         struct inode *inode = dreq->inode;
 473         ssize_t result = -EINVAL;
 474         size_t requested_bytes = 0;
 475         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 476
 477         nfs_pageio_init_read(&desc, dreq->inode, false,
 478                              &nfs_direct_read_completion_ops);
 479         get_dreq(dreq);
 480         desc.pg_dreq = dreq;
 481         inode_dio_begin(inode);
 482
 483         while (iov_iter_count(iter)) {
 484                 struct page **pagevec;
 485                 size_t bytes;
 486                 size_t pgbase;
 487                 unsigned npages, i;
 488
 489                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 490                                                   rsize, &pgbase);
 491                 if (result < 0)
 492                         break;
 493
 494                 bytes = result;
 495                 iov_iter_advance(iter, bytes);
 496                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 497                 for (i = 0; i < npages; i++) {
 498                         struct nfs_page *req;
 499                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 500                         /* XXX do we need to do the eof zeroing found in async_filler? */
 501                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 502                                                  pgbase, req_len);
 503                         if (IS_ERR(req)) {
 504                                 result = PTR_ERR(req);
 505                                 break;
 506                         }
 507                         req->wb_index = pos >> PAGE_SHIFT;
 508                         req->wb_offset = pos & ~PAGE_MASK;
 509                         if (!nfs_pageio_add_request(&desc, req)) {
 510                                 result = desc.pg_error;
 511                                 nfs_release_request(req);
 512                                 break;
 513                         }
 514                         pgbase = 0;
 515                         bytes -= req_len;
 516                         requested_bytes += req_len;
 517                         pos += req_len;
 518                         dreq->bytes_left -= req_len;
 519                 }
 520                 nfs_direct_release_pages(pagevec, npages);
 521                 kvfree(pagevec);
 522                 if (result < 0)
 523                         break;
 524         }
 525
 526         nfs_pageio_complete(&desc);
 527
 528         /*
 529          * If no bytes were started, return the error, and let the
 530          * generic layer handle the completion.
 531          */
 532         if (requested_bytes == 0) {
 533                 inode_dio_end(inode);
 534                 nfs_direct_req_release(dreq);
 535                 return result < 0 ? result : -EIO;
 536         }
 537
 538         if (put_dreq(dreq))
 539                 nfs_direct_complete(dreq);
 540         return requested_bytes;
 541 }
 542
 543 /**
 544  * nfs_file_direct_read - file direct read operation for NFS files
 545  * @iocb: target I/O control block
 546  * @iter: vector of user buffers into which to read data
 547  *
 548  * We use this function for direct reads instead of calling
 549  * generic_file_aio_read() in order to avoid gfar's check to see if
 550  * the request starts before the end of the file.  For that check
 551  * to work, we must generate a GETATTR before each direct read, and
 552  * even then there is a window between the GETATTR and the subsequent
 553  * READ where the file size could change.  Our preference is simply
 554  * to do all reads the application wants, and the server will take
 555  * care of managing the end of file boundary.
 556  *
 557  * This function also eliminates unnecessarily updating the file's
 558  * atime locally, as the NFS server sets the file's atime, and this
 559  * client must read the updated atime from the server back into its
 560  * cache.
 561  */
 562 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
 563 {
 564         struct file *file = iocb->ki_filp;
 565         struct address_space *mapping = file->f_mapping;
 566         struct inode *inode = mapping->host;
 567         struct nfs_direct_req *dreq;
 568         struct nfs_lock_context *l_ctx;
 569         ssize_t result = -EINVAL, requested;
 570         size_t count = iov_iter_count(iter);
 571         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 572
 573         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 574                 file, count, (long long) iocb->ki_pos);
 575
 576         result = 0;
 577         if (!count)
 578                 goto out;
 579
 580         task_io_account_read(count);
 581
 582         result = -ENOMEM;
 583         dreq = nfs_direct_req_alloc();
 584         if (dreq == NULL)
 585                 goto out;
 586
 587         dreq->inode = inode;
 588         dreq->bytes_left = dreq->max_count = count;
 589         dreq->io_start = iocb->ki_pos;
 590         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 591         l_ctx = nfs_get_lock_context(dreq->ctx);
 592         if (IS_ERR(l_ctx)) {
 593                 result = PTR_ERR(l_ctx);
 594                 goto out_release;
 595         }
 596         dreq->l_ctx = l_ctx;
 597         if (!is_sync_kiocb(iocb))
 598                 dreq->iocb = iocb;
 599
 600         nfs_start_io_direct(inode);
 601
 602         NFS_I(inode)->read_io += count;
 603         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 604
 605         nfs_end_io_direct(inode);
 606
 607         if (requested > 0) {
 608                 result = nfs_direct_wait(dreq);
 609                 if (result > 0) {
 610                         requested -= result;
 611                         iocb->ki_pos += result;
 612                 }
 613                 iov_iter_revert(iter, requested);
 614         } else {
 615                 result = requested;
 616         }
 617
 618 out_release:
 619         nfs_direct_req_release(dreq);
 620 out:
 621         return result;
 622 }
 623
 624 static void
 625 nfs_direct_write_scan_commit_list(struct inode *inode,
 626                                   struct list_head *list,
 627                                   struct nfs_commit_info *cinfo)
 628 {
 629         spin_lock(&cinfo->inode->i_lock);
 630 #ifdef CONFIG_NFS_V4_1
 631         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 632                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 633 #endif
 634         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 635         spin_unlock(&cinfo->inode->i_lock);
 636 }
 637
 638 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 639 {
 640         struct nfs_pageio_descriptor desc;
 641         struct nfs_page *req, *tmp;
 642         LIST_HEAD(reqs);
 643         struct nfs_commit_info cinfo;
 644         LIST_HEAD(failed);
 645         int i;
 646
 647         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 648         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 649
 650         dreq->count = 0;
 651         dreq->verf.committed = NFS_INVALID_STABLE_HOW;
 652         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 653         for (i = 0; i < dreq->mirror_count; i++)
 654                 dreq->mirrors[i].count = 0;
 655         get_dreq(dreq);
 656
 657         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 658                               &nfs_direct_write_completion_ops);
 659         desc.pg_dreq = dreq;
 660
 661         req = nfs_list_entry(reqs.next);
 662         nfs_direct_setup_mirroring(dreq, &desc, req);
 663         if (desc.pg_error < 0) {
 664                 list_splice_init(&reqs, &failed);
 665                 goto out_failed;
 666         }
 667
 668         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 669                 if (!nfs_pageio_add_request(&desc, req)) {
 670                         nfs_list_remove_request(req);
 671                         nfs_list_add_request(req, &failed);
 672                         spin_lock(&cinfo.inode->i_lock);
 673                         dreq->flags = 0;
 674                         if (desc.pg_error < 0)
 675                                 dreq->error = desc.pg_error;
 676                         else
 677                                 dreq->error = -EIO;
 678                         spin_unlock(&cinfo.inode->i_lock);
 679                 }
 680                 nfs_release_request(req);
 681         }
 682         nfs_pageio_complete(&desc);
 683
 684 out_failed:
 685         while (!list_empty(&failed)) {
 686                 req = nfs_list_entry(failed.next);
 687                 nfs_list_remove_request(req);
 688                 nfs_unlock_and_release_request(req);
 689         }
 690
 691         if (put_dreq(dreq))
 692                 nfs_direct_write_complete(dreq);
 693 }
 694
 695 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 696 {
 697         struct nfs_direct_req *dreq = data->dreq;
 698         struct nfs_commit_info cinfo;
 699         struct nfs_page *req;
 700         int status = data->task.tk_status;
 701
 702         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 703         if (status < 0) {
 704                 dprintk("NFS: %5u commit failed with error %d.\n",
 705                         data->task.tk_pid, status);
 706                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 707         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 708                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 709                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 710         }
 711
 712         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 713         while (!list_empty(&data->pages)) {
 714                 req = nfs_list_entry(data->pages.next);
 715                 nfs_list_remove_request(req);
 716                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 717                         /* Note the rewrite will go through mds */
 718                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 719                 } else
 720                         nfs_release_request(req);
 721                 nfs_unlock_and_release_request(req);
 722         }
 723
 724         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 725                 nfs_direct_write_complete(dreq);
 726 }
 727
 728 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 729                 struct nfs_page *req)
 730 {
 731         struct nfs_direct_req *dreq = cinfo->dreq;
 732
 733         spin_lock(&dreq->lock);
 734         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 735         spin_unlock(&dreq->lock);
 736         nfs_mark_request_commit(req, NULL, cinfo, 0);
 737 }
 738
 739 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 740         .completion = nfs_direct_commit_complete,
 741         .resched_write = nfs_direct_resched_write,
 742 };
 743
 744 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 745 {
 746         int res;
 747         struct nfs_commit_info cinfo;
 748         LIST_HEAD(mds_list);
 749
 750         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 751         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 752         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 753         if (res < 0) /* res == -ENOMEM */
 754                 nfs_direct_write_reschedule(dreq);
 755 }
 756
 757 static void nfs_direct_write_schedule_work(struct work_struct *work)
 758 {
 759         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 760         int flags = dreq->flags;
 761
 762         dreq->flags = 0;
 763         switch (flags) {
 764                 case NFS_ODIRECT_DO_COMMIT:
 765                         nfs_direct_commit_schedule(dreq);
 766                         break;
 767                 case NFS_ODIRECT_RESCHED_WRITES:
 768                         nfs_direct_write_reschedule(dreq);
 769                         break;
 770                 default:
 771                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 772                         nfs_direct_complete(dreq);
 773         }
 774 }
 775
 776 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 777 {
 778         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 779 }
 780
 781 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 782 {
 783         struct nfs_direct_req *dreq = hdr->dreq;
 784         struct nfs_commit_info cinfo;
 785         bool request_commit = false;
 786         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 787
 788         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 789                 goto out_put;
 790
 791         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 792
 793         spin_lock(&dreq->lock);
 794
 795         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 796                 dreq->flags = 0;
 797                 dreq->error = hdr->error;
 798         }
 799         if (dreq->error == 0) {
 800                 nfs_direct_good_bytes(dreq, hdr);
 801                 if (nfs_write_need_commit(hdr)) {
 802                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 803                                 request_commit = true;
 804                         else if (dreq->flags == 0) {
 805                                 nfs_direct_set_hdr_verf(dreq, hdr);
 806                                 request_commit = true;
 807                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 808                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 809                                 request_commit = true;
 810                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 811                                         dreq->flags =
 812                                                 NFS_ODIRECT_RESCHED_WRITES;
 813                         }
 814                 }
 815         }
 816         spin_unlock(&dreq->lock);
 817
 818         while (!list_empty(&hdr->pages)) {
 819
 820                 req = nfs_list_entry(hdr->pages.next);
 821                 nfs_list_remove_request(req);
 822                 if (request_commit) {
 823                         kref_get(&req->wb_kref);
 824                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 825                                 hdr->ds_commit_idx);
 826                 }
 827                 nfs_unlock_and_release_request(req);
 828         }
 829
 830 out_put:
 831         if (put_dreq(dreq))
 832                 nfs_direct_write_complete(dreq);
 833         hdr->release(hdr);
 834 }
 835
 836 static void nfs_write_sync_pgio_error(struct list_head *head)
 837 {
 838         struct nfs_page *req;
 839
 840         while (!list_empty(head)) {
 841                 req = nfs_list_entry(head->next);
 842                 nfs_list_remove_request(req);
 843                 nfs_unlock_and_release_request(req);
 844         }
 845 }
 846
 847 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 848 {
 849         struct nfs_direct_req *dreq = hdr->dreq;
 850
 851         spin_lock(&dreq->lock);
 852         if (dreq->error == 0) {
 853                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 854                 /* fake unstable write to let common nfs resend pages */
 855                 hdr->verf.committed = NFS_UNSTABLE;
 856                 hdr->good_bytes = hdr->args.count;
 857         }
 858         spin_unlock(&dreq->lock);
 859 }
 860
 861 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 862         .error_cleanup = nfs_write_sync_pgio_error,
 863         .init_hdr = nfs_direct_pgio_init,
 864         .completion = nfs_direct_write_completion,
 865         .reschedule_io = nfs_direct_write_reschedule_io,
 866 };
 867
 868
 869 /*
 870  * NB: Return the value of the first error return code.  Subsequent
 871  *     errors after the first one are ignored.
 872  */
 873 /*
 874  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 875  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 876  * bail and stop sending more writes.  Write length accounting is
 877  * handled automatically by nfs_direct_write_result().  Otherwise, if
 878  * no requests have been sent, just return an error.
 879  */
 880 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 881                                                struct iov_iter *iter,
 882                                                loff_t pos)
 883 {
 884         struct nfs_pageio_descriptor desc;
 885         struct inode *inode = dreq->inode;
 886         ssize_t result = 0;
 887         size_t requested_bytes = 0;
 888         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 889
 890         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 891                               &nfs_direct_write_completion_ops);
 892         desc.pg_dreq = dreq;
 893         get_dreq(dreq);
 894         inode_dio_begin(inode);
 895
 896         NFS_I(inode)->write_io += iov_iter_count(iter);
 897         while (iov_iter_count(iter)) {
 898                 struct page **pagevec;
 899                 size_t bytes;
 900                 size_t pgbase;
 901                 unsigned npages, i;
 902
 903                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 904                                                   wsize, &pgbase);
 905                 if (result < 0)
 906                         break;
 907
 908                 bytes = result;
 909                 iov_iter_advance(iter, bytes);
 910                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 911                 for (i = 0; i < npages; i++) {
 912                         struct nfs_page *req;
 913                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 914
 915                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 916                                                  pgbase, req_len);
 917                         if (IS_ERR(req)) {
 918                                 result = PTR_ERR(req);
 919                                 break;
 920                         }
 921
 922                         nfs_direct_setup_mirroring(dreq, &desc, req);
 923                         if (desc.pg_error < 0) {
 924                                 nfs_free_request(req);
 925                                 result = desc.pg_error;
 926                                 break;
 927                         }
 928
 929                         nfs_lock_request(req);
 930                         req->wb_index = pos >> PAGE_SHIFT;
 931                         req->wb_offset = pos & ~PAGE_MASK;
 932                         if (!nfs_pageio_add_request(&desc, req)) {
 933                                 result = desc.pg_error;
 934                                 nfs_unlock_and_release_request(req);
 935                                 break;
 936                         }
 937                         pgbase = 0;
 938                         bytes -= req_len;
 939                         requested_bytes += req_len;
 940                         pos += req_len;
 941                         dreq->bytes_left -= req_len;
 942                 }
 943                 nfs_direct_release_pages(pagevec, npages);
 944                 kvfree(pagevec);
 945                 if (result < 0)
 946                         break;
 947         }
 948         nfs_pageio_complete(&desc);
 949
 950         /*
 951          * If no bytes were started, return the error, and let the
 952          * generic layer handle the completion.
 953          */
 954         if (requested_bytes == 0) {
 955                 inode_dio_end(inode);
 956                 nfs_direct_req_release(dreq);
 957                 return result < 0 ? result : -EIO;
 958         }
 959
 960         if (put_dreq(dreq))
 961                 nfs_direct_write_complete(dreq);
 962         return requested_bytes;
 963 }
 964
 965 /**
 966  * nfs_file_direct_write - file direct write operation for NFS files
 967  * @iocb: target I/O control block
 968  * @iter: vector of user buffers from which to write data
 969  *
 970  * We use this function for direct writes instead of calling
 971  * generic_file_aio_write() in order to avoid taking the inode
 972  * semaphore and updating the i_size.  The NFS server will set
 973  * the new i_size and this client must read the updated size
 974  * back into its cache.  We let the server do generic write
 975  * parameter checking and report problems.
 976  *
 977  * We eliminate local atime updates, see direct read above.
 978  *
 979  * We avoid unnecessary page cache invalidations for normal cached
 980  * readers of this file.
 981  *
 982  * Note that O_APPEND is not supported for NFS direct writes, as there
 983  * is no atomic O_APPEND write facility in the NFS protocol.
 984  */
 985 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 986 {
 987         ssize_t result = -EINVAL, requested;
 988         size_t count;
 989         struct file *file = iocb->ki_filp;
 990         struct address_space *mapping = file->f_mapping;
 991         struct inode *inode = mapping->host;
 992         struct nfs_direct_req *dreq;
 993         struct nfs_lock_context *l_ctx;
 994         loff_t pos, end;
 995
 996         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 997                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
 998
 999         result = generic_write_checks(iocb, iter);
1000         if (result <= 0)
1001                 return result;
1002         count = result;
1003         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1004
1005         pos = iocb->ki_pos;
1006         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1007
1008         task_io_account_write(count);
1009
1010         result = -ENOMEM;
1011         dreq = nfs_direct_req_alloc();
1012         if (!dreq)
1013                 goto out;
1014
1015         dreq->inode = inode;
1016         dreq->bytes_left = dreq->max_count = count;
1017         dreq->io_start = pos;
1018         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1019         l_ctx = nfs_get_lock_context(dreq->ctx);
1020         if (IS_ERR(l_ctx)) {
1021                 result = PTR_ERR(l_ctx);
1022                 goto out_release;
1023         }
1024         dreq->l_ctx = l_ctx;
1025         if (!is_sync_kiocb(iocb))
1026                 dreq->iocb = iocb;
1027
1028         nfs_start_io_direct(inode);
1029
1030         requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1031
1032         if (mapping->nrpages) {
1033                 invalidate_inode_pages2_range(mapping,
1034                                               pos >> PAGE_SHIFT, end);
1035         }
1036
1037         nfs_end_io_direct(inode);
1038
1039         if (requested > 0) {
1040                 result = nfs_direct_wait(dreq);
1041                 if (result > 0) {
1042                         requested -= result;
1043                         iocb->ki_pos = pos + result;
1044                         /* XXX: should check the generic_write_sync retval */
1045                         generic_write_sync(iocb, result);
1046                 }
1047                 iov_iter_revert(iter, requested);
1048         } else {
1049                 result = requested;
1050         }
1051 out_release:
1052         nfs_direct_req_release(dreq);
1053 out:
1054         return result;
1055 }
1056
1057 /**
1058  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1059  *
1060  */
1061 int __init nfs_init_directcache(void)
1062 {
1063         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1064                                                 sizeof(struct nfs_direct_req),
1065                                                 0, (SLAB_RECLAIM_ACCOUNT|
1066                                                         SLAB_MEM_SPREAD),
1067                                                 NULL);
1068         if (nfs_direct_cachep == NULL)
1069                 return -ENOMEM;
1070
1071         return 0;
1072 }
1073
1074 /**
1075  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1076  *
1077  */
1078 void nfs_destroy_directcache(void)
1079 {
1080         kmem_cache_destroy(nfs_direct_cachep);
1081 }