fs/nfs/direct.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/direct.c
   4  *
   5  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   6  *
   7  * High-performance uncached I/O for the Linux NFS client
   8  *
   9  * There are important applications whose performance or correctness
  10  * depends on uncached access to file data.  Database clusters
  11  * (multiple copies of the same instance running on separate hosts)
  12  * implement their own cache coherency protocol that subsumes file
  13  * system cache protocols.  Applications that process datasets
  14  * considerably larger than the client's memory do not always benefit
  15  * from a local cache.  A streaming video server, for instance, has no
  16  * need to cache the contents of a file.
  17  *
  18  * When an application requests uncached I/O, all read and write requests
  19  * are made directly to the server; data stored or fetched via these
  20  * requests is not cached in the Linux page cache.  The client does not
  21  * correct unaligned requests from applications.  All requested bytes are
  22  * held on permanent storage before a direct write system call returns to
  23  * an application.
  24  *
  25  * Solaris implements an uncached I/O facility called directio() that
  26  * is used for backups and sequential I/O to very large files.  Solaris
  27  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  28  * an undocumented mount option.
  29  *
  30  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  31  * help from Andrew Morton.
  32  *
  33  * 18 Dec 2001  Initial implementation for 2.4  --cel
  34  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  35  * 08 Jun 2003  Port to 2.5 APIs  --cel
  36  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  37  * 15 Sep 2004  Parallel async reads  --cel
  38  * 04 May 2005  support O_DIRECT with aio  --cel
  39  *
  40  */
  41
  42 #include <linux/errno.h>
  43 #include <linux/sched.h>
  44 #include <linux/kernel.h>
  45 #include <linux/file.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/kref.h>
  48 #include <linux/slab.h>
  49 #include <linux/task_io_accounting_ops.h>
  50 #include <linux/module.h>
  51
  52 #include <linux/nfs_fs.h>
  53 #include <linux/nfs_page.h>
  54 #include <linux/sunrpc/clnt.h>
  55
  56 #include <linux/uaccess.h>
  57 #include <linux/atomic.h>
  58
  59 #include "internal.h"
  60 #include "iostat.h"
  61 #include "pnfs.h"
  62
  63 #define NFSDBG_FACILITY         NFSDBG_VFS
  64
  65 static struct kmem_cache *nfs_direct_cachep;
  66
  67 /*
  68  * This represents a set of asynchronous requests that we're waiting on
  69  */
  70 struct nfs_direct_mirror {
  71         ssize_t count;
  72 };
  73
  74 struct nfs_direct_req {
  75         struct kref             kref;           /* release manager */
  76
  77         /* I/O parameters */
  78         struct nfs_open_context *ctx;           /* file open context info */
  79         struct nfs_lock_context *l_ctx;         /* Lock context info */
  80         struct kiocb *          iocb;           /* controlling i/o request */
  81         struct inode *          inode;          /* target file of i/o */
  82
  83         /* completion state */
  84         atomic_t                io_count;       /* i/os we're waiting for */
  85         spinlock_t              lock;           /* protect completion state */
  86
  87         struct nfs_direct_mirror mirrors[NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX];
  88         int                     mirror_count;
  89
  90         loff_t                  io_start;       /* Start offset for I/O */
  91         ssize_t                 count,          /* bytes actually processed */
  92                                 max_count,      /* max expected count */
  93                                 bytes_left,     /* bytes left to be sent */
  94                                 error;          /* any reported error */
  95         struct completion       completion;     /* wait for i/o completion */
  96
  97         /* commit state */
  98         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  99         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
 100         struct work_struct      work;
 101         int                     flags;
 102         /* for write */
 103 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
 104 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
 105         /* for read */
 106 #define NFS_ODIRECT_SHOULD_DIRTY        (3)     /* dirty user-space page after read */
 107         struct nfs_writeverf    verf;           /* unstable write verifier */
 108 };
 109
 110 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 111 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 112 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
 113 static void nfs_direct_write_schedule_work(struct work_struct *work);
 114
 115 static inline void get_dreq(struct nfs_direct_req *dreq)
 116 {
 117         atomic_inc(&dreq->io_count);
 118 }
 119
 120 static inline int put_dreq(struct nfs_direct_req *dreq)
 121 {
 122         return atomic_dec_and_test(&dreq->io_count);
 123 }
 124
 125 static void
 126 nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
 127 {
 128         int i;
 129         ssize_t count;
 130
 131         WARN_ON_ONCE(dreq->count >= dreq->max_count);
 132
 133         if (dreq->mirror_count == 1) {
 134                 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
 135                 dreq->count += hdr->good_bytes;
 136         } else {
 137                 /* mirrored writes */
 138                 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
 139                 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
 140                         count = hdr->io_start + hdr->good_bytes - dreq->io_start;
 141                         dreq->mirrors[hdr->pgio_mirror_idx].count = count;
 142                 }
 143                 /* update the dreq->count by finding the minimum agreed count from all
 144                  * mirrors */
 145                 count = dreq->mirrors[0].count;
 146
 147                 for (i = 1; i < dreq->mirror_count; i++)
 148                         count = min(count, dreq->mirrors[i].count);
 149
 150                 dreq->count = count;
 151         }
 152 }
 153
 154 /*
 155  * nfs_direct_select_verf - select the right verifier
 156  * @dreq - direct request possibly spanning multiple servers
 157  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 158  * @commit_idx - commit bucket index for the DS
 159  *
 160  * returns the correct verifier to use given the role of the server
 161  */
 162 static struct nfs_writeverf *
 163 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 164                        struct nfs_client *ds_clp,
 165                        int commit_idx)
 166 {
 167         struct nfs_writeverf *verfp = &dreq->verf;
 168
 169 #ifdef CONFIG_NFS_V4_1
 170         /*
 171          * pNFS is in use, use the DS verf except commit_through_mds is set
 172          * for layout segment where nbuckets is zero.
 173          */
 174         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 175                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 176                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 177                 else
 178                         WARN_ON_ONCE(1);
 179         }
 180 #endif
 181         return verfp;
 182 }
 183
 184
 185 /*
 186  * nfs_direct_set_hdr_verf - set the write/commit verifier
 187  * @dreq - direct request possibly spanning multiple servers
 188  * @hdr - pageio header to validate against previously seen verfs
 189  *
 190  * Set the server's (MDS or DS) "seen" verifier
 191  */
 192 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 193                                     struct nfs_pgio_header *hdr)
 194 {
 195         struct nfs_writeverf *verfp;
 196
 197         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 198         WARN_ON_ONCE(verfp->committed >= 0);
 199         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 200         WARN_ON_ONCE(verfp->committed < 0);
 201 }
 202
 203 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
 204                 const struct nfs_writeverf *v2)
 205 {
 206         return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
 207 }
 208
 209 /*
 210  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 211  * @dreq - direct request possibly spanning multiple servers
 212  * @hdr - pageio header to validate against previously seen verf
 213  *
 214  * set the server's "seen" verf if not initialized.
 215  * returns result of comparison between @hdr->verf and the "seen"
 216  * verf of the server used by @hdr (DS or MDS)
 217  */
 218 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 219                                           struct nfs_pgio_header *hdr)
 220 {
 221         struct nfs_writeverf *verfp;
 222
 223         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 224         if (verfp->committed < 0) {
 225                 nfs_direct_set_hdr_verf(dreq, hdr);
 226                 return 0;
 227         }
 228         return nfs_direct_cmp_verf(verfp, &hdr->verf);
 229 }
 230
 231 /*
 232  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 233  * @dreq - direct request possibly spanning multiple servers
 234  * @data - commit data to validate against previously seen verf
 235  *
 236  * returns result of comparison between @data->verf and the verf of
 237  * the server used by @data (DS or MDS)
 238  */
 239 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 240                                            struct nfs_commit_data *data)
 241 {
 242         struct nfs_writeverf *verfp;
 243
 244         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 245                                          data->ds_commit_index);
 246
 247         /* verifier not set so always fail */
 248         if (verfp->committed < 0)
 249                 return 1;
 250
 251         return nfs_direct_cmp_verf(verfp, &data->verf);
 252 }
 253
 254 /**
 255  * nfs_direct_IO - NFS address space operation for direct I/O
 256  * @iocb: target I/O control block
 257  * @iter: I/O buffer
 258  *
 259  * The presence of this routine in the address space ops vector means
 260  * the NFS client supports direct I/O. However, for most direct IO, we
 261  * shunt off direct read and write requests before the VFS gets them,
 262  * so this method is only ever called for swap.
 263  */
 264 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 265 {
 266         struct inode *inode = iocb->ki_filp->f_mapping->host;
 267
 268         /* we only support swap file calling nfs_direct_IO */
 269         if (!IS_SWAPFILE(inode))
 270                 return 0;
 271
 272         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 273
 274         if (iov_iter_rw(iter) == READ)
 275                 return nfs_file_direct_read(iocb, iter);
 276         return nfs_file_direct_write(iocb, iter);
 277 }
 278
 279 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 280 {
 281         unsigned int i;
 282         for (i = 0; i < npages; i++)
 283                 put_page(pages[i]);
 284 }
 285
 286 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 287                               struct nfs_direct_req *dreq)
 288 {
 289         cinfo->inode = dreq->inode;
 290         cinfo->mds = &dreq->mds_cinfo;
 291         cinfo->ds = &dreq->ds_cinfo;
 292         cinfo->dreq = dreq;
 293         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 294 }
 295
 296 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 297                                              struct nfs_pageio_descriptor *pgio,
 298                                              struct nfs_page *req)
 299 {
 300         int mirror_count = 1;
 301
 302         if (pgio->pg_ops->pg_get_mirror_count)
 303                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 304
 305         dreq->mirror_count = mirror_count;
 306 }
 307
 308 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 309 {
 310         struct nfs_direct_req *dreq;
 311
 312         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 313         if (!dreq)
 314                 return NULL;
 315
 316         kref_init(&dreq->kref);
 317         kref_get(&dreq->kref);
 318         init_completion(&dreq->completion);
 319         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 320         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 321         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 322         dreq->mirror_count = 1;
 323         spin_lock_init(&dreq->lock);
 324
 325         return dreq;
 326 }
 327
 328 static void nfs_direct_req_free(struct kref *kref)
 329 {
 330         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 331
 332         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 333         if (dreq->l_ctx != NULL)
 334                 nfs_put_lock_context(dreq->l_ctx);
 335         if (dreq->ctx != NULL)
 336                 put_nfs_open_context(dreq->ctx);
 337         kmem_cache_free(nfs_direct_cachep, dreq);
 338 }
 339
 340 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 341 {
 342         kref_put(&dreq->kref, nfs_direct_req_free);
 343 }
 344
 345 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 346 {
 347         return dreq->bytes_left;
 348 }
 349 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 350
 351 /*
 352  * Collects and returns the final error value/byte-count.
 353  */
 354 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 355 {
 356         ssize_t result = -EIOCBQUEUED;
 357
 358         /* Async requests don't wait here */
 359         if (dreq->iocb)
 360                 goto out;
 361
 362         result = wait_for_completion_killable(&dreq->completion);
 363
 364         if (!result) {
 365                 result = dreq->count;
 366                 WARN_ON_ONCE(dreq->count < 0);
 367         }
 368         if (!result)
 369                 result = dreq->error;
 370
 371 out:
 372         return (ssize_t) result;
 373 }
 374
 375 /*
 376  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 377  * the iocb is still valid here if this is a synchronous request.
 378  */
 379 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 380 {
 381         struct inode *inode = dreq->inode;
 382
 383         inode_dio_end(inode);
 384
 385         if (dreq->iocb) {
 386                 long res = (long) dreq->error;
 387                 if (dreq->count != 0) {
 388                         res = (long) dreq->count;
 389                         WARN_ON_ONCE(dreq->count < 0);
 390                 }
 391                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 392         }
 393
 394         complete(&dreq->completion);
 395
 396         nfs_direct_req_release(dreq);
 397 }
 398
 399 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 400 {
 401         unsigned long bytes = 0;
 402         struct nfs_direct_req *dreq = hdr->dreq;
 403
 404         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 405                 goto out_put;
 406
 407         spin_lock(&dreq->lock);
 408         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 409                 dreq->error = hdr->error;
 410         else
 411                 nfs_direct_good_bytes(dreq, hdr);
 412
 413         spin_unlock(&dreq->lock);
 414
 415         while (!list_empty(&hdr->pages)) {
 416                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 417                 struct page *page = req->wb_page;
 418
 419                 if (!PageCompound(page) && bytes < hdr->good_bytes &&
 420                     (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 421                         set_page_dirty(page);
 422                 bytes += req->wb_bytes;
 423                 nfs_list_remove_request(req);
 424                 nfs_release_request(req);
 425         }
 426 out_put:
 427         if (put_dreq(dreq))
 428                 nfs_direct_complete(dreq);
 429         hdr->release(hdr);
 430 }
 431
 432 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 433 {
 434         struct nfs_page *req;
 435
 436         while (!list_empty(head)) {
 437                 req = nfs_list_entry(head->next);
 438                 nfs_list_remove_request(req);
 439                 nfs_release_request(req);
 440         }
 441 }
 442
 443 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 444 {
 445         get_dreq(hdr->dreq);
 446 }
 447
 448 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 449         .error_cleanup = nfs_read_sync_pgio_error,
 450         .init_hdr = nfs_direct_pgio_init,
 451         .completion = nfs_direct_read_completion,
 452 };
 453
 454 /*
 455  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 456  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 457  * bail and stop sending more reads.  Read length accounting is
 458  * handled automatically by nfs_direct_read_result().  Otherwise, if
 459  * no requests have been sent, just return an error.
 460  */
 461
 462 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 463                                               struct iov_iter *iter,
 464                                               loff_t pos)
 465 {
 466         struct nfs_pageio_descriptor desc;
 467         struct inode *inode = dreq->inode;
 468         ssize_t result = -EINVAL;
 469         size_t requested_bytes = 0;
 470         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 471
 472         nfs_pageio_init_read(&desc, dreq->inode, false,
 473                              &nfs_direct_read_completion_ops);
 474         get_dreq(dreq);
 475         desc.pg_dreq = dreq;
 476         inode_dio_begin(inode);
 477
 478         while (iov_iter_count(iter)) {
 479                 struct page **pagevec;
 480                 size_t bytes;
 481                 size_t pgbase;
 482                 unsigned npages, i;
 483
 484                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 485                                                   rsize, &pgbase);
 486                 if (result < 0)
 487                         break;
 488
 489                 bytes = result;
 490                 iov_iter_advance(iter, bytes);
 491                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 492                 for (i = 0; i < npages; i++) {
 493                         struct nfs_page *req;
 494                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 495                         /* XXX do we need to do the eof zeroing found in async_filler? */
 496                         req = nfs_create_request(dreq->ctx, pagevec[i],
 497                                                  pgbase, req_len);
 498                         if (IS_ERR(req)) {
 499                                 result = PTR_ERR(req);
 500                                 break;
 501                         }
 502                         req->wb_index = pos >> PAGE_SHIFT;
 503                         req->wb_offset = pos & ~PAGE_MASK;
 504                         if (!nfs_pageio_add_request(&desc, req)) {
 505                                 result = desc.pg_error;
 506                                 nfs_release_request(req);
 507                                 break;
 508                         }
 509                         pgbase = 0;
 510                         bytes -= req_len;
 511                         requested_bytes += req_len;
 512                         pos += req_len;
 513                         dreq->bytes_left -= req_len;
 514                 }
 515                 nfs_direct_release_pages(pagevec, npages);
 516                 kvfree(pagevec);
 517                 if (result < 0)
 518                         break;
 519         }
 520
 521         nfs_pageio_complete(&desc);
 522
 523         /*
 524          * If no bytes were started, return the error, and let the
 525          * generic layer handle the completion.
 526          */
 527         if (requested_bytes == 0) {
 528                 inode_dio_end(inode);
 529                 nfs_direct_req_release(dreq);
 530                 return result < 0 ? result : -EIO;
 531         }
 532
 533         if (put_dreq(dreq))
 534                 nfs_direct_complete(dreq);
 535         return requested_bytes;
 536 }
 537
 538 /**
 539  * nfs_file_direct_read - file direct read operation for NFS files
 540  * @iocb: target I/O control block
 541  * @iter: vector of user buffers into which to read data
 542  *
 543  * We use this function for direct reads instead of calling
 544  * generic_file_aio_read() in order to avoid gfar's check to see if
 545  * the request starts before the end of the file.  For that check
 546  * to work, we must generate a GETATTR before each direct read, and
 547  * even then there is a window between the GETATTR and the subsequent
 548  * READ where the file size could change.  Our preference is simply
 549  * to do all reads the application wants, and the server will take
 550  * care of managing the end of file boundary.
 551  *
 552  * This function also eliminates unnecessarily updating the file's
 553  * atime locally, as the NFS server sets the file's atime, and this
 554  * client must read the updated atime from the server back into its
 555  * cache.
 556  */
 557 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
 558 {
 559         struct file *file = iocb->ki_filp;
 560         struct address_space *mapping = file->f_mapping;
 561         struct inode *inode = mapping->host;
 562         struct nfs_direct_req *dreq;
 563         struct nfs_lock_context *l_ctx;
 564         ssize_t result = -EINVAL, requested;
 565         size_t count = iov_iter_count(iter);
 566         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 567
 568         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 569                 file, count, (long long) iocb->ki_pos);
 570
 571         result = 0;
 572         if (!count)
 573                 goto out;
 574
 575         task_io_account_read(count);
 576
 577         result = -ENOMEM;
 578         dreq = nfs_direct_req_alloc();
 579         if (dreq == NULL)
 580                 goto out;
 581
 582         dreq->inode = inode;
 583         dreq->bytes_left = dreq->max_count = count;
 584         dreq->io_start = iocb->ki_pos;
 585         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 586         l_ctx = nfs_get_lock_context(dreq->ctx);
 587         if (IS_ERR(l_ctx)) {
 588                 result = PTR_ERR(l_ctx);
 589                 goto out_release;
 590         }
 591         dreq->l_ctx = l_ctx;
 592         if (!is_sync_kiocb(iocb))
 593                 dreq->iocb = iocb;
 594
 595         if (iter_is_iovec(iter))
 596                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 597
 598         nfs_start_io_direct(inode);
 599
 600         NFS_I(inode)->read_io += count;
 601         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 602
 603         nfs_end_io_direct(inode);
 604
 605         if (requested > 0) {
 606                 result = nfs_direct_wait(dreq);
 607                 if (result > 0) {
 608                         requested -= result;
 609                         iocb->ki_pos += result;
 610                 }
 611                 iov_iter_revert(iter, requested);
 612         } else {
 613                 result = requested;
 614         }
 615
 616 out_release:
 617         nfs_direct_req_release(dreq);
 618 out:
 619         return result;
 620 }
 621
 622 static void
 623 nfs_direct_write_scan_commit_list(struct inode *inode,
 624                                   struct list_head *list,
 625                                   struct nfs_commit_info *cinfo)
 626 {
 627         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 628 #ifdef CONFIG_NFS_V4_1
 629         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 630                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 631 #endif
 632         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 633         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 634 }
 635
 636 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 637 {
 638         struct nfs_pageio_descriptor desc;
 639         struct nfs_page *req, *tmp;
 640         LIST_HEAD(reqs);
 641         struct nfs_commit_info cinfo;
 642         LIST_HEAD(failed);
 643         int i;
 644
 645         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 646         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 647
 648         dreq->count = 0;
 649         dreq->verf.committed = NFS_INVALID_STABLE_HOW;
 650         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 651         for (i = 0; i < dreq->mirror_count; i++)
 652                 dreq->mirrors[i].count = 0;
 653         get_dreq(dreq);
 654
 655         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 656                               &nfs_direct_write_completion_ops);
 657         desc.pg_dreq = dreq;
 658
 659         req = nfs_list_entry(reqs.next);
 660         nfs_direct_setup_mirroring(dreq, &desc, req);
 661         if (desc.pg_error < 0) {
 662                 list_splice_init(&reqs, &failed);
 663                 goto out_failed;
 664         }
 665
 666         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 667                 /* Bump the transmission count */
 668                 req->wb_nio++;
 669                 if (!nfs_pageio_add_request(&desc, req)) {
 670                         nfs_list_move_request(req, &failed);
 671                         spin_lock(&cinfo.inode->i_lock);
 672                         dreq->flags = 0;
 673                         if (desc.pg_error < 0)
 674                                 dreq->error = desc.pg_error;
 675                         else
 676                                 dreq->error = -EIO;
 677                         spin_unlock(&cinfo.inode->i_lock);
 678                 }
 679                 nfs_release_request(req);
 680         }
 681         nfs_pageio_complete(&desc);
 682
 683 out_failed:
 684         while (!list_empty(&failed)) {
 685                 req = nfs_list_entry(failed.next);
 686                 nfs_list_remove_request(req);
 687                 nfs_unlock_and_release_request(req);
 688         }
 689
 690         if (put_dreq(dreq))
 691                 nfs_direct_write_complete(dreq);
 692 }
 693
 694 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 695 {
 696         struct nfs_direct_req *dreq = data->dreq;
 697         struct nfs_commit_info cinfo;
 698         struct nfs_page *req;
 699         int status = data->task.tk_status;
 700
 701         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 702         if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
 703                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 704
 705         while (!list_empty(&data->pages)) {
 706                 req = nfs_list_entry(data->pages.next);
 707                 nfs_list_remove_request(req);
 708                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 709                         /*
 710                          * Despite the reboot, the write was successful,
 711                          * so reset wb_nio.
 712                          */
 713                         req->wb_nio = 0;
 714                         /* Note the rewrite will go through mds */
 715                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 716                 } else
 717                         nfs_release_request(req);
 718                 nfs_unlock_and_release_request(req);
 719         }
 720
 721         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 722                 nfs_direct_write_complete(dreq);
 723 }
 724
 725 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 726                 struct nfs_page *req)
 727 {
 728         struct nfs_direct_req *dreq = cinfo->dreq;
 729
 730         spin_lock(&dreq->lock);
 731         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 732         spin_unlock(&dreq->lock);
 733         nfs_mark_request_commit(req, NULL, cinfo, 0);
 734 }
 735
 736 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 737         .completion = nfs_direct_commit_complete,
 738         .resched_write = nfs_direct_resched_write,
 739 };
 740
 741 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 742 {
 743         int res;
 744         struct nfs_commit_info cinfo;
 745         LIST_HEAD(mds_list);
 746
 747         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 748         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 749         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 750         if (res < 0) /* res == -ENOMEM */
 751                 nfs_direct_write_reschedule(dreq);
 752 }
 753
 754 static void nfs_direct_write_schedule_work(struct work_struct *work)
 755 {
 756         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 757         int flags = dreq->flags;
 758
 759         dreq->flags = 0;
 760         switch (flags) {
 761                 case NFS_ODIRECT_DO_COMMIT:
 762                         nfs_direct_commit_schedule(dreq);
 763                         break;
 764                 case NFS_ODIRECT_RESCHED_WRITES:
 765                         nfs_direct_write_reschedule(dreq);
 766                         break;
 767                 default:
 768                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 769                         nfs_direct_complete(dreq);
 770         }
 771 }
 772
 773 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 774 {
 775         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 776 }
 777
 778 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 779 {
 780         struct nfs_direct_req *dreq = hdr->dreq;
 781         struct nfs_commit_info cinfo;
 782         bool request_commit = false;
 783         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 784
 785         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 786                 goto out_put;
 787
 788         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 789
 790         spin_lock(&dreq->lock);
 791
 792         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
 793                 dreq->error = hdr->error;
 794         if (dreq->error == 0) {
 795                 nfs_direct_good_bytes(dreq, hdr);
 796                 if (nfs_write_need_commit(hdr)) {
 797                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 798                                 request_commit = true;
 799                         else if (dreq->flags == 0) {
 800                                 nfs_direct_set_hdr_verf(dreq, hdr);
 801                                 request_commit = true;
 802                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 803                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 804                                 request_commit = true;
 805                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 806                                         dreq->flags =
 807                                                 NFS_ODIRECT_RESCHED_WRITES;
 808                         }
 809                 }
 810         }
 811         spin_unlock(&dreq->lock);
 812
 813         while (!list_empty(&hdr->pages)) {
 814
 815                 req = nfs_list_entry(hdr->pages.next);
 816                 nfs_list_remove_request(req);
 817                 if (request_commit) {
 818                         kref_get(&req->wb_kref);
 819                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 820                                 hdr->ds_commit_idx);
 821                 }
 822                 nfs_unlock_and_release_request(req);
 823         }
 824
 825 out_put:
 826         if (put_dreq(dreq))
 827                 nfs_direct_write_complete(dreq);
 828         hdr->release(hdr);
 829 }
 830
 831 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 832 {
 833         struct nfs_page *req;
 834
 835         while (!list_empty(head)) {
 836                 req = nfs_list_entry(head->next);
 837                 nfs_list_remove_request(req);
 838                 nfs_unlock_and_release_request(req);
 839         }
 840 }
 841
 842 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 843 {
 844         struct nfs_direct_req *dreq = hdr->dreq;
 845
 846         spin_lock(&dreq->lock);
 847         if (dreq->error == 0) {
 848                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 849                 /* fake unstable write to let common nfs resend pages */
 850                 hdr->verf.committed = NFS_UNSTABLE;
 851                 hdr->good_bytes = hdr->args.count;
 852         }
 853         spin_unlock(&dreq->lock);
 854 }
 855
 856 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 857         .error_cleanup = nfs_write_sync_pgio_error,
 858         .init_hdr = nfs_direct_pgio_init,
 859         .completion = nfs_direct_write_completion,
 860         .reschedule_io = nfs_direct_write_reschedule_io,
 861 };
 862
 863
 864 /*
 865  * NB: Return the value of the first error return code.  Subsequent
 866  *     errors after the first one are ignored.
 867  */
 868 /*
 869  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 870  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 871  * bail and stop sending more writes.  Write length accounting is
 872  * handled automatically by nfs_direct_write_result().  Otherwise, if
 873  * no requests have been sent, just return an error.
 874  */
 875 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 876                                                struct iov_iter *iter,
 877                                                loff_t pos)
 878 {
 879         struct nfs_pageio_descriptor desc;
 880         struct inode *inode = dreq->inode;
 881         ssize_t result = 0;
 882         size_t requested_bytes = 0;
 883         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 884
 885         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 886                               &nfs_direct_write_completion_ops);
 887         desc.pg_dreq = dreq;
 888         get_dreq(dreq);
 889         inode_dio_begin(inode);
 890
 891         NFS_I(inode)->write_io += iov_iter_count(iter);
 892         while (iov_iter_count(iter)) {
 893                 struct page **pagevec;
 894                 size_t bytes;
 895                 size_t pgbase;
 896                 unsigned npages, i;
 897
 898                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 899                                                   wsize, &pgbase);
 900                 if (result < 0)
 901                         break;
 902
 903                 bytes = result;
 904                 iov_iter_advance(iter, bytes);
 905                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 906                 for (i = 0; i < npages; i++) {
 907                         struct nfs_page *req;
 908                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 909
 910                         req = nfs_create_request(dreq->ctx, pagevec[i],
 911                                                  pgbase, req_len);
 912                         if (IS_ERR(req)) {
 913                                 result = PTR_ERR(req);
 914                                 break;
 915                         }
 916
 917                         nfs_direct_setup_mirroring(dreq, &desc, req);
 918                         if (desc.pg_error < 0) {
 919                                 nfs_free_request(req);
 920                                 result = desc.pg_error;
 921                                 break;
 922                         }
 923
 924                         nfs_lock_request(req);
 925                         req->wb_index = pos >> PAGE_SHIFT;
 926                         req->wb_offset = pos & ~PAGE_MASK;
 927                         if (!nfs_pageio_add_request(&desc, req)) {
 928                                 result = desc.pg_error;
 929                                 nfs_unlock_and_release_request(req);
 930                                 break;
 931                         }
 932                         pgbase = 0;
 933                         bytes -= req_len;
 934                         requested_bytes += req_len;
 935                         pos += req_len;
 936                         dreq->bytes_left -= req_len;
 937                 }
 938                 nfs_direct_release_pages(pagevec, npages);
 939                 kvfree(pagevec);
 940                 if (result < 0)
 941                         break;
 942         }
 943         nfs_pageio_complete(&desc);
 944
 945         /*
 946          * If no bytes were started, return the error, and let the
 947          * generic layer handle the completion.
 948          */
 949         if (requested_bytes == 0) {
 950                 inode_dio_end(inode);
 951                 nfs_direct_req_release(dreq);
 952                 return result < 0 ? result : -EIO;
 953         }
 954
 955         if (put_dreq(dreq))
 956                 nfs_direct_write_complete(dreq);
 957         return requested_bytes;
 958 }
 959
 960 /**
 961  * nfs_file_direct_write - file direct write operation for NFS files
 962  * @iocb: target I/O control block
 963  * @iter: vector of user buffers from which to write data
 964  *
 965  * We use this function for direct writes instead of calling
 966  * generic_file_aio_write() in order to avoid taking the inode
 967  * semaphore and updating the i_size.  The NFS server will set
 968  * the new i_size and this client must read the updated size
 969  * back into its cache.  We let the server do generic write
 970  * parameter checking and report problems.
 971  *
 972  * We eliminate local atime updates, see direct read above.
 973  *
 974  * We avoid unnecessary page cache invalidations for normal cached
 975  * readers of this file.
 976  *
 977  * Note that O_APPEND is not supported for NFS direct writes, as there
 978  * is no atomic O_APPEND write facility in the NFS protocol.
 979  */
 980 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 981 {
 982         ssize_t result = -EINVAL, requested;
 983         size_t count;
 984         struct file *file = iocb->ki_filp;
 985         struct address_space *mapping = file->f_mapping;
 986         struct inode *inode = mapping->host;
 987         struct nfs_direct_req *dreq;
 988         struct nfs_lock_context *l_ctx;
 989         loff_t pos, end;
 990
 991         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 992                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
 993
 994         result = generic_write_checks(iocb, iter);
 995         if (result <= 0)
 996                 return result;
 997         count = result;
 998         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 999
1000         pos = iocb->ki_pos;
1001         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1002
1003         task_io_account_write(count);
1004
1005         result = -ENOMEM;
1006         dreq = nfs_direct_req_alloc();
1007         if (!dreq)
1008                 goto out;
1009
1010         dreq->inode = inode;
1011         dreq->bytes_left = dreq->max_count = count;
1012         dreq->io_start = pos;
1013         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1014         l_ctx = nfs_get_lock_context(dreq->ctx);
1015         if (IS_ERR(l_ctx)) {
1016                 result = PTR_ERR(l_ctx);
1017                 goto out_release;
1018         }
1019         dreq->l_ctx = l_ctx;
1020         if (!is_sync_kiocb(iocb))
1021                 dreq->iocb = iocb;
1022
1023         nfs_start_io_direct(inode);
1024
1025         requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1026
1027         if (mapping->nrpages) {
1028                 invalidate_inode_pages2_range(mapping,
1029                                               pos >> PAGE_SHIFT, end);
1030         }
1031
1032         nfs_end_io_direct(inode);
1033
1034         if (requested > 0) {
1035                 result = nfs_direct_wait(dreq);
1036                 if (result > 0) {
1037                         requested -= result;
1038                         iocb->ki_pos = pos + result;
1039                         /* XXX: should check the generic_write_sync retval */
1040                         generic_write_sync(iocb, result);
1041                 }
1042                 iov_iter_revert(iter, requested);
1043         } else {
1044                 result = requested;
1045         }
1046 out_release:
1047         nfs_direct_req_release(dreq);
1048 out:
1049         return result;
1050 }
1051
1052 /**
1053  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1054  *
1055  */
1056 int __init nfs_init_directcache(void)
1057 {
1058         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1059                                                 sizeof(struct nfs_direct_req),
1060                                                 0, (SLAB_RECLAIM_ACCOUNT|
1061                                                         SLAB_MEM_SPREAD),
1062                                                 NULL);
1063         if (nfs_direct_cachep == NULL)
1064                 return -ENOMEM;
1065
1066         return 0;
1067 }
1068
1069 /**
1070  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1071  *
1072  */
1073 void nfs_destroy_directcache(void)
1074 {
1075         kmem_cache_destroy(nfs_direct_cachep);
1076 }