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