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 <asm/uaccess.h>
  56 #include <linux/atomic.h>
  57
  58 #include "internal.h"
  59 #include "iostat.h"
  60 #include "pnfs.h"
  61
  62 #define NFSDBG_FACILITY         NFSDBG_VFS
  63
  64 static struct kmem_cache *nfs_direct_cachep;
  65
  66 /*
  67  * This represents a set of asynchronous requests that we're waiting on
  68  */
  69 struct nfs_direct_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                                 bytes_left,     /* bytes left to be sent */
  91                                 io_start,       /* start of IO */
  92                                 error;          /* any reported error */
  93         struct completion       completion;     /* wait for i/o completion */
  94
  95         /* commit state */
  96         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  97         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  98         struct work_struct      work;
  99         int                     flags;
 100 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
 101 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
 102         struct nfs_writeverf    verf;           /* unstable write verifier */
 103 };
 104
 105 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 106 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 107 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
 108 static void nfs_direct_write_schedule_work(struct work_struct *work);
 109
 110 static inline void get_dreq(struct nfs_direct_req *dreq)
 111 {
 112         atomic_inc(&dreq->io_count);
 113 }
 114
 115 static inline int put_dreq(struct nfs_direct_req *dreq)
 116 {
 117         return atomic_dec_and_test(&dreq->io_count);
 118 }
 119
 120 static void
 121 nfs_direct_good_bytes(struct nfs_direct_req *dreq, struct nfs_pgio_header *hdr)
 122 {
 123         int i;
 124         ssize_t count;
 125
 126         if (dreq->mirror_count == 1) {
 127                 dreq->mirrors[hdr->pgio_mirror_idx].count += hdr->good_bytes;
 128                 dreq->count += hdr->good_bytes;
 129         } else {
 130                 /* mirrored writes */
 131                 count = dreq->mirrors[hdr->pgio_mirror_idx].count;
 132                 if (count + dreq->io_start < hdr->io_start + hdr->good_bytes) {
 133                         count = hdr->io_start + hdr->good_bytes - dreq->io_start;
 134                         dreq->mirrors[hdr->pgio_mirror_idx].count = count;
 135                 }
 136                 /* update the dreq->count by finding the minimum agreed count from all
 137                  * mirrors */
 138                 count = dreq->mirrors[0].count;
 139
 140                 for (i = 1; i < dreq->mirror_count; i++)
 141                         count = min(count, dreq->mirrors[i].count);
 142
 143                 dreq->count = count;
 144         }
 145 }
 146
 147 /*
 148  * nfs_direct_select_verf - select the right verifier
 149  * @dreq - direct request possibly spanning multiple servers
 150  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 151  * @commit_idx - commit bucket index for the DS
 152  *
 153  * returns the correct verifier to use given the role of the server
 154  */
 155 static struct nfs_writeverf *
 156 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 157                        struct nfs_client *ds_clp,
 158                        int commit_idx)
 159 {
 160         struct nfs_writeverf *verfp = &dreq->verf;
 161
 162 #ifdef CONFIG_NFS_V4_1
 163         /*
 164          * pNFS is in use, use the DS verf except commit_through_mds is set
 165          * for layout segment where nbuckets is zero.
 166          */
 167         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 168                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 169                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 170                 else
 171                         WARN_ON_ONCE(1);
 172         }
 173 #endif
 174         return verfp;
 175 }
 176
 177
 178 /*
 179  * nfs_direct_set_hdr_verf - set the write/commit verifier
 180  * @dreq - direct request possibly spanning multiple servers
 181  * @hdr - pageio header to validate against previously seen verfs
 182  *
 183  * Set the server's (MDS or DS) "seen" verifier
 184  */
 185 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 186                                     struct nfs_pgio_header *hdr)
 187 {
 188         struct nfs_writeverf *verfp;
 189
 190         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 191         WARN_ON_ONCE(verfp->committed >= 0);
 192         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 193         WARN_ON_ONCE(verfp->committed < 0);
 194 }
 195
 196 /*
 197  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 198  * @dreq - direct request possibly spanning multiple servers
 199  * @hdr - pageio header to validate against previously seen verf
 200  *
 201  * set the server's "seen" verf if not initialized.
 202  * returns result of comparison between @hdr->verf and the "seen"
 203  * verf of the server used by @hdr (DS or MDS)
 204  */
 205 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 206                                           struct nfs_pgio_header *hdr)
 207 {
 208         struct nfs_writeverf *verfp;
 209
 210         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 211         if (verfp->committed < 0) {
 212                 nfs_direct_set_hdr_verf(dreq, hdr);
 213                 return 0;
 214         }
 215         return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 216 }
 217
 218 /*
 219  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 220  * @dreq - direct request possibly spanning multiple servers
 221  * @data - commit data to validate against previously seen verf
 222  *
 223  * returns result of comparison between @data->verf and the verf of
 224  * the server used by @data (DS or MDS)
 225  */
 226 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 227                                            struct nfs_commit_data *data)
 228 {
 229         struct nfs_writeverf *verfp;
 230
 231         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 232                                          data->ds_commit_index);
 233
 234         /* verifier not set so always fail */
 235         if (verfp->committed < 0)
 236                 return 1;
 237
 238         return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
 239 }
 240
 241 /**
 242  * nfs_direct_IO - NFS address space operation for direct I/O
 243  * @iocb: target I/O control block
 244  * @iov: array of vectors that define I/O buffer
 245  * @pos: offset in file to begin the operation
 246  * @nr_segs: size of iovec array
 247  *
 248  * The presence of this routine in the address space ops vector means
 249  * the NFS client supports direct I/O. However, for most direct IO, we
 250  * shunt off direct read and write requests before the VFS gets them,
 251  * so this method is only ever called for swap.
 252  */
 253 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 254 {
 255         struct inode *inode = iocb->ki_filp->f_mapping->host;
 256
 257         /* we only support swap file calling nfs_direct_IO */
 258         if (!IS_SWAPFILE(inode))
 259                 return 0;
 260
 261         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 262
 263         if (iov_iter_rw(iter) == READ)
 264                 return nfs_file_direct_read(iocb, iter);
 265         return nfs_file_direct_write(iocb, iter);
 266 }
 267
 268 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 269 {
 270         unsigned int i;
 271         for (i = 0; i < npages; i++)
 272                 put_page(pages[i]);
 273 }
 274
 275 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 276                               struct nfs_direct_req *dreq)
 277 {
 278         cinfo->lock = &dreq->inode->i_lock;
 279         cinfo->mds = &dreq->mds_cinfo;
 280         cinfo->ds = &dreq->ds_cinfo;
 281         cinfo->dreq = dreq;
 282         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 283 }
 284
 285 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 286                                              struct nfs_pageio_descriptor *pgio,
 287                                              struct nfs_page *req)
 288 {
 289         int mirror_count = 1;
 290
 291         if (pgio->pg_ops->pg_get_mirror_count)
 292                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 293
 294         dreq->mirror_count = mirror_count;
 295 }
 296
 297 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 298 {
 299         struct nfs_direct_req *dreq;
 300
 301         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 302         if (!dreq)
 303                 return NULL;
 304
 305         kref_init(&dreq->kref);
 306         kref_get(&dreq->kref);
 307         init_completion(&dreq->completion);
 308         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 309         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 310         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 311         dreq->mirror_count = 1;
 312         spin_lock_init(&dreq->lock);
 313
 314         return dreq;
 315 }
 316
 317 static void nfs_direct_req_free(struct kref *kref)
 318 {
 319         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 320
 321         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 322         if (dreq->l_ctx != NULL)
 323                 nfs_put_lock_context(dreq->l_ctx);
 324         if (dreq->ctx != NULL)
 325                 put_nfs_open_context(dreq->ctx);
 326         kmem_cache_free(nfs_direct_cachep, dreq);
 327 }
 328
 329 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 330 {
 331         kref_put(&dreq->kref, nfs_direct_req_free);
 332 }
 333
 334 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 335 {
 336         return dreq->bytes_left;
 337 }
 338 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 339
 340 /*
 341  * Collects and returns the final error value/byte-count.
 342  */
 343 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 344 {
 345         ssize_t result = -EIOCBQUEUED;
 346
 347         /* Async requests don't wait here */
 348         if (dreq->iocb)
 349                 goto out;
 350
 351         result = wait_for_completion_killable(&dreq->completion);
 352
 353         if (!result)
 354                 result = dreq->error;
 355         if (!result)
 356                 result = dreq->count;
 357
 358 out:
 359         return (ssize_t) result;
 360 }
 361
 362 /*
 363  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 364  * the iocb is still valid here if this is a synchronous request.
 365  */
 366 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 367 {
 368         struct inode *inode = dreq->inode;
 369
 370         if (dreq->iocb && write) {
 371                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 372
 373                 spin_lock(&inode->i_lock);
 374                 if (i_size_read(inode) < pos)
 375                         i_size_write(inode, pos);
 376                 spin_unlock(&inode->i_lock);
 377         }
 378
 379         if (write)
 380                 nfs_zap_mapping(inode, inode->i_mapping);
 381
 382         inode_dio_end(inode);
 383
 384         if (dreq->iocb) {
 385                 long res = (long) dreq->error;
 386                 if (!res)
 387                         res = (long) dreq->count;
 388                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 389         }
 390
 391         complete_all(&dreq->completion);
 392
 393         nfs_direct_req_release(dreq);
 394 }
 395
 396 static void nfs_direct_readpage_release(struct nfs_page *req)
 397 {
 398         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 399                 req->wb_context->dentry->d_sb->s_id,
 400                 (unsigned long long)NFS_FILEID(d_inode(req->wb_context->dentry)),
 401                 req->wb_bytes,
 402                 (long long)req_offset(req));
 403         nfs_release_request(req);
 404 }
 405
 406 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 407 {
 408         unsigned long bytes = 0;
 409         struct nfs_direct_req *dreq = hdr->dreq;
 410
 411         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 412                 goto out_put;
 413
 414         spin_lock(&dreq->lock);
 415         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 416                 dreq->error = hdr->error;
 417         else
 418                 nfs_direct_good_bytes(dreq, hdr);
 419
 420         spin_unlock(&dreq->lock);
 421
 422         while (!list_empty(&hdr->pages)) {
 423                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 424                 struct page *page = req->wb_page;
 425
 426                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 427                         set_page_dirty(page);
 428                 bytes += req->wb_bytes;
 429                 nfs_list_remove_request(req);
 430                 nfs_direct_readpage_release(req);
 431         }
 432 out_put:
 433         if (put_dreq(dreq))
 434                 nfs_direct_complete(dreq, false);
 435         hdr->release(hdr);
 436 }
 437
 438 static void nfs_read_sync_pgio_error(struct list_head *head)
 439 {
 440         struct nfs_page *req;
 441
 442         while (!list_empty(head)) {
 443                 req = nfs_list_entry(head->next);
 444                 nfs_list_remove_request(req);
 445                 nfs_release_request(req);
 446         }
 447 }
 448
 449 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 450 {
 451         get_dreq(hdr->dreq);
 452 }
 453
 454 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 455         .error_cleanup = nfs_read_sync_pgio_error,
 456         .init_hdr = nfs_direct_pgio_init,
 457         .completion = nfs_direct_read_completion,
 458 };
 459
 460 /*
 461  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 462  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 463  * bail and stop sending more reads.  Read length accounting is
 464  * handled automatically by nfs_direct_read_result().  Otherwise, if
 465  * no requests have been sent, just return an error.
 466  */
 467
 468 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 469                                               struct iov_iter *iter,
 470                                               loff_t pos)
 471 {
 472         struct nfs_pageio_descriptor desc;
 473         struct inode *inode = dreq->inode;
 474         ssize_t result = -EINVAL;
 475         size_t requested_bytes = 0;
 476         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 477
 478         nfs_pageio_init_read(&desc, dreq->inode, false,
 479                              &nfs_direct_read_completion_ops);
 480         get_dreq(dreq);
 481         desc.pg_dreq = dreq;
 482         inode_dio_begin(inode);
 483
 484         while (iov_iter_count(iter)) {
 485                 struct page **pagevec;
 486                 size_t bytes;
 487                 size_t pgbase;
 488                 unsigned npages, i;
 489
 490                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 491                                                   rsize, &pgbase);
 492                 if (result < 0)
 493                         break;
 494
 495                 bytes = result;
 496                 iov_iter_advance(iter, bytes);
 497                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 498                 for (i = 0; i < npages; i++) {
 499                         struct nfs_page *req;
 500                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 501                         /* XXX do we need to do the eof zeroing found in async_filler? */
 502                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 503                                                  pgbase, req_len);
 504                         if (IS_ERR(req)) {
 505                                 result = PTR_ERR(req);
 506                                 break;
 507                         }
 508                         req->wb_index = pos >> PAGE_SHIFT;
 509                         req->wb_offset = pos & ~PAGE_MASK;
 510                         if (!nfs_pageio_add_request(&desc, req)) {
 511                                 result = desc.pg_error;
 512                                 nfs_release_request(req);
 513                                 break;
 514                         }
 515                         pgbase = 0;
 516                         bytes -= req_len;
 517                         requested_bytes += req_len;
 518                         pos += req_len;
 519                         dreq->bytes_left -= req_len;
 520                 }
 521                 nfs_direct_release_pages(pagevec, npages);
 522                 kvfree(pagevec);
 523                 if (result < 0)
 524                         break;
 525         }
 526
 527         nfs_pageio_complete(&desc);
 528
 529         /*
 530          * If no bytes were started, return the error, and let the
 531          * generic layer handle the completion.
 532          */
 533         if (requested_bytes == 0) {
 534                 inode_dio_end(inode);
 535                 nfs_direct_req_release(dreq);
 536                 return result < 0 ? result : -EIO;
 537         }
 538
 539         if (put_dreq(dreq))
 540                 nfs_direct_complete(dreq, false);
 541         return 0;
 542 }
 543
 544 /**
 545  * nfs_file_direct_read - file direct read operation for NFS files
 546  * @iocb: target I/O control block
 547  * @iter: vector of user buffers into which to read data
 548  *
 549  * We use this function for direct reads instead of calling
 550  * generic_file_aio_read() in order to avoid gfar's check to see if
 551  * the request starts before the end of the file.  For that check
 552  * to work, we must generate a GETATTR before each direct read, and
 553  * even then there is a window between the GETATTR and the subsequent
 554  * READ where the file size could change.  Our preference is simply
 555  * to do all reads the application wants, and the server will take
 556  * care of managing the end of file boundary.
 557  *
 558  * This function also eliminates unnecessarily updating the file's
 559  * atime locally, as the NFS server sets the file's atime, and this
 560  * client must read the updated atime from the server back into its
 561  * cache.
 562  */
 563 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
 564 {
 565         struct file *file = iocb->ki_filp;
 566         struct address_space *mapping = file->f_mapping;
 567         struct inode *inode = mapping->host;
 568         struct nfs_direct_req *dreq;
 569         struct nfs_lock_context *l_ctx;
 570         ssize_t result = -EINVAL;
 571         size_t count = iov_iter_count(iter);
 572         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 573
 574         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 575                 file, count, (long long) iocb->ki_pos);
 576
 577         result = 0;
 578         if (!count)
 579                 goto out;
 580
 581         inode_lock(inode);
 582         result = nfs_sync_mapping(mapping);
 583         if (result)
 584                 goto out_unlock;
 585
 586         task_io_account_read(count);
 587
 588         result = -ENOMEM;
 589         dreq = nfs_direct_req_alloc();
 590         if (dreq == NULL)
 591                 goto out_unlock;
 592
 593         dreq->inode = inode;
 594         dreq->bytes_left = count;
 595         dreq->io_start = iocb->ki_pos;
 596         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 597         l_ctx = nfs_get_lock_context(dreq->ctx);
 598         if (IS_ERR(l_ctx)) {
 599                 result = PTR_ERR(l_ctx);
 600                 goto out_release;
 601         }
 602         dreq->l_ctx = l_ctx;
 603         if (!is_sync_kiocb(iocb))
 604                 dreq->iocb = iocb;
 605
 606         NFS_I(inode)->read_io += count;
 607         result = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 608
 609         inode_unlock(inode);
 610
 611         if (!result) {
 612                 result = nfs_direct_wait(dreq);
 613                 if (result > 0)
 614                         iocb->ki_pos += result;
 615         }
 616
 617         nfs_direct_req_release(dreq);
 618         return result;
 619
 620 out_release:
 621         nfs_direct_req_release(dreq);
 622 out_unlock:
 623         inode_unlock(inode);
 624 out:
 625         return result;
 626 }
 627
 628 static void
 629 nfs_direct_write_scan_commit_list(struct inode *inode,
 630                                   struct list_head *list,
 631                                   struct nfs_commit_info *cinfo)
 632 {
 633         spin_lock(cinfo->lock);
 634 #ifdef CONFIG_NFS_V4_1
 635         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 636                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 637 #endif
 638         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 639         spin_unlock(cinfo->lock);
 640 }
 641
 642 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 643 {
 644         struct nfs_pageio_descriptor desc;
 645         struct nfs_page *req, *tmp;
 646         LIST_HEAD(reqs);
 647         struct nfs_commit_info cinfo;
 648         LIST_HEAD(failed);
 649         int i;
 650
 651         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 652         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 653
 654         dreq->count = 0;
 655         for (i = 0; i < dreq->mirror_count; i++)
 656                 dreq->mirrors[i].count = 0;
 657         get_dreq(dreq);
 658
 659         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 660                               &nfs_direct_write_completion_ops);
 661         desc.pg_dreq = dreq;
 662
 663         req = nfs_list_entry(reqs.next);
 664         nfs_direct_setup_mirroring(dreq, &desc, req);
 665         if (desc.pg_error < 0) {
 666                 list_splice_init(&reqs, &failed);
 667                 goto out_failed;
 668         }
 669
 670         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 671                 if (!nfs_pageio_add_request(&desc, req)) {
 672                         nfs_list_remove_request(req);
 673                         nfs_list_add_request(req, &failed);
 674                         spin_lock(cinfo.lock);
 675                         dreq->flags = 0;
 676                         if (desc.pg_error < 0)
 677                                 dreq->error = desc.pg_error;
 678                         else
 679                                 dreq->error = -EIO;
 680                         spin_unlock(cinfo.lock);
 681                 }
 682                 nfs_release_request(req);
 683         }
 684         nfs_pageio_complete(&desc);
 685
 686 out_failed:
 687         while (!list_empty(&failed)) {
 688                 req = nfs_list_entry(failed.next);
 689                 nfs_list_remove_request(req);
 690                 nfs_unlock_and_release_request(req);
 691         }
 692
 693         if (put_dreq(dreq))
 694                 nfs_direct_write_complete(dreq, dreq->inode);
 695 }
 696
 697 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 698 {
 699         struct nfs_direct_req *dreq = data->dreq;
 700         struct nfs_commit_info cinfo;
 701         struct nfs_page *req;
 702         int status = data->task.tk_status;
 703
 704         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 705         if (status < 0) {
 706                 dprintk("NFS: %5u commit failed with error %d.\n",
 707                         data->task.tk_pid, status);
 708                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 709         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 710                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 711                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 712         }
 713
 714         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 715         while (!list_empty(&data->pages)) {
 716                 req = nfs_list_entry(data->pages.next);
 717                 nfs_list_remove_request(req);
 718                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 719                         /* Note the rewrite will go through mds */
 720                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 721                 } else
 722                         nfs_release_request(req);
 723                 nfs_unlock_and_release_request(req);
 724         }
 725
 726         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 727                 nfs_direct_write_complete(dreq, data->inode);
 728 }
 729
 730 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 731                 struct nfs_page *req)
 732 {
 733         struct nfs_direct_req *dreq = cinfo->dreq;
 734
 735         spin_lock(&dreq->lock);
 736         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 737         spin_unlock(&dreq->lock);
 738         nfs_mark_request_commit(req, NULL, cinfo, 0);
 739 }
 740
 741 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 742         .completion = nfs_direct_commit_complete,
 743         .resched_write = nfs_direct_resched_write,
 744 };
 745
 746 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 747 {
 748         int res;
 749         struct nfs_commit_info cinfo;
 750         LIST_HEAD(mds_list);
 751
 752         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 753         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 754         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 755         if (res < 0) /* res == -ENOMEM */
 756                 nfs_direct_write_reschedule(dreq);
 757 }
 758
 759 static void nfs_direct_write_schedule_work(struct work_struct *work)
 760 {
 761         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 762         int flags = dreq->flags;
 763
 764         dreq->flags = 0;
 765         switch (flags) {
 766                 case NFS_ODIRECT_DO_COMMIT:
 767                         nfs_direct_commit_schedule(dreq);
 768                         break;
 769                 case NFS_ODIRECT_RESCHED_WRITES:
 770                         nfs_direct_write_reschedule(dreq);
 771                         break;
 772                 default:
 773                         nfs_direct_complete(dreq, true);
 774         }
 775 }
 776
 777 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 778 {
 779         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 780 }
 781
 782 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 783 {
 784         struct nfs_direct_req *dreq = hdr->dreq;
 785         struct nfs_commit_info cinfo;
 786         bool request_commit = false;
 787         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 788
 789         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 790                 goto out_put;
 791
 792         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 793
 794         spin_lock(&dreq->lock);
 795
 796         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 797                 dreq->flags = 0;
 798                 dreq->error = hdr->error;
 799         }
 800         if (dreq->error == 0) {
 801                 nfs_direct_good_bytes(dreq, hdr);
 802                 if (nfs_write_need_commit(hdr)) {
 803                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 804                                 request_commit = true;
 805                         else if (dreq->flags == 0) {
 806                                 nfs_direct_set_hdr_verf(dreq, hdr);
 807                                 request_commit = true;
 808                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 809                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 810                                 request_commit = true;
 811                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 812                                         dreq->flags =
 813                                                 NFS_ODIRECT_RESCHED_WRITES;
 814                         }
 815                 }
 816         }
 817         spin_unlock(&dreq->lock);
 818
 819         while (!list_empty(&hdr->pages)) {
 820
 821                 req = nfs_list_entry(hdr->pages.next);
 822                 nfs_list_remove_request(req);
 823                 if (request_commit) {
 824                         kref_get(&req->wb_kref);
 825                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 826                                 hdr->ds_commit_idx);
 827                 }
 828                 nfs_unlock_and_release_request(req);
 829         }
 830
 831 out_put:
 832         if (put_dreq(dreq))
 833                 nfs_direct_write_complete(dreq, hdr->inode);
 834         hdr->release(hdr);
 835 }
 836
 837 static void nfs_write_sync_pgio_error(struct list_head *head)
 838 {
 839         struct nfs_page *req;
 840
 841         while (!list_empty(head)) {
 842                 req = nfs_list_entry(head->next);
 843                 nfs_list_remove_request(req);
 844                 nfs_unlock_and_release_request(req);
 845         }
 846 }
 847
 848 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 849 {
 850         struct nfs_direct_req *dreq = hdr->dreq;
 851
 852         spin_lock(&dreq->lock);
 853         if (dreq->error == 0) {
 854                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 855                 /* fake unstable write to let common nfs resend pages */
 856                 hdr->verf.committed = NFS_UNSTABLE;
 857                 hdr->good_bytes = hdr->args.count;
 858         }
 859         spin_unlock(&dreq->lock);
 860 }
 861
 862 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 863         .error_cleanup = nfs_write_sync_pgio_error,
 864         .init_hdr = nfs_direct_pgio_init,
 865         .completion = nfs_direct_write_completion,
 866         .reschedule_io = nfs_direct_write_reschedule_io,
 867 };
 868
 869
 870 /*
 871  * NB: Return the value of the first error return code.  Subsequent
 872  *     errors after the first one are ignored.
 873  */
 874 /*
 875  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 876  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 877  * bail and stop sending more writes.  Write length accounting is
 878  * handled automatically by nfs_direct_write_result().  Otherwise, if
 879  * no requests have been sent, just return an error.
 880  */
 881 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 882                                                struct iov_iter *iter,
 883                                                loff_t pos)
 884 {
 885         struct nfs_pageio_descriptor desc;
 886         struct inode *inode = dreq->inode;
 887         ssize_t result = 0;
 888         size_t requested_bytes = 0;
 889         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 890
 891         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 892                               &nfs_direct_write_completion_ops);
 893         desc.pg_dreq = dreq;
 894         get_dreq(dreq);
 895         inode_dio_begin(inode);
 896
 897         NFS_I(inode)->write_io += iov_iter_count(iter);
 898         while (iov_iter_count(iter)) {
 899                 struct page **pagevec;
 900                 size_t bytes;
 901                 size_t pgbase;
 902                 unsigned npages, i;
 903
 904                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 905                                                   wsize, &pgbase);
 906                 if (result < 0)
 907                         break;
 908
 909                 bytes = result;
 910                 iov_iter_advance(iter, bytes);
 911                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 912                 for (i = 0; i < npages; i++) {
 913                         struct nfs_page *req;
 914                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 915
 916                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 917                                                  pgbase, req_len);
 918                         if (IS_ERR(req)) {
 919                                 result = PTR_ERR(req);
 920                                 break;
 921                         }
 922
 923                         nfs_direct_setup_mirroring(dreq, &desc, req);
 924                         if (desc.pg_error < 0) {
 925                                 nfs_free_request(req);
 926                                 result = desc.pg_error;
 927                                 break;
 928                         }
 929
 930                         nfs_lock_request(req);
 931                         req->wb_index = pos >> PAGE_SHIFT;
 932                         req->wb_offset = pos & ~PAGE_MASK;
 933                         if (!nfs_pageio_add_request(&desc, req)) {
 934                                 result = desc.pg_error;
 935                                 nfs_unlock_and_release_request(req);
 936                                 break;
 937                         }
 938                         pgbase = 0;
 939                         bytes -= req_len;
 940                         requested_bytes += req_len;
 941                         pos += req_len;
 942                         dreq->bytes_left -= req_len;
 943                 }
 944                 nfs_direct_release_pages(pagevec, npages);
 945                 kvfree(pagevec);
 946                 if (result < 0)
 947                         break;
 948         }
 949         nfs_pageio_complete(&desc);
 950
 951         /*
 952          * If no bytes were started, return the error, and let the
 953          * generic layer handle the completion.
 954          */
 955         if (requested_bytes == 0) {
 956                 inode_dio_end(inode);
 957                 nfs_direct_req_release(dreq);
 958                 return result < 0 ? result : -EIO;
 959         }
 960
 961         if (put_dreq(dreq))
 962                 nfs_direct_write_complete(dreq, dreq->inode);
 963         return 0;
 964 }
 965
 966 /**
 967  * nfs_file_direct_write - file direct write operation for NFS files
 968  * @iocb: target I/O control block
 969  * @iter: vector of user buffers from which to write data
 970  *
 971  * We use this function for direct writes instead of calling
 972  * generic_file_aio_write() in order to avoid taking the inode
 973  * semaphore and updating the i_size.  The NFS server will set
 974  * the new i_size and this client must read the updated size
 975  * back into its cache.  We let the server do generic write
 976  * parameter checking and report problems.
 977  *
 978  * We eliminate local atime updates, see direct read above.
 979  *
 980  * We avoid unnecessary page cache invalidations for normal cached
 981  * readers of this file.
 982  *
 983  * Note that O_APPEND is not supported for NFS direct writes, as there
 984  * is no atomic O_APPEND write facility in the NFS protocol.
 985  */
 986 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 987 {
 988         ssize_t result = -EINVAL;
 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         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES,
1000                       iov_iter_count(iter));
1001
1002         pos = iocb->ki_pos;
1003         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1004
1005         inode_lock(inode);
1006
1007         result = nfs_sync_mapping(mapping);
1008         if (result)
1009                 goto out_unlock;
1010
1011         if (mapping->nrpages) {
1012                 result = invalidate_inode_pages2_range(mapping,
1013                                         pos >> PAGE_SHIFT, end);
1014                 if (result)
1015                         goto out_unlock;
1016         }
1017
1018         task_io_account_write(iov_iter_count(iter));
1019
1020         result = -ENOMEM;
1021         dreq = nfs_direct_req_alloc();
1022         if (!dreq)
1023                 goto out_unlock;
1024
1025         dreq->inode = inode;
1026         dreq->bytes_left = iov_iter_count(iter);
1027         dreq->io_start = pos;
1028         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1029         l_ctx = nfs_get_lock_context(dreq->ctx);
1030         if (IS_ERR(l_ctx)) {
1031                 result = PTR_ERR(l_ctx);
1032                 goto out_release;
1033         }
1034         dreq->l_ctx = l_ctx;
1035         if (!is_sync_kiocb(iocb))
1036                 dreq->iocb = iocb;
1037
1038         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1039
1040         if (mapping->nrpages) {
1041                 invalidate_inode_pages2_range(mapping,
1042                                               pos >> PAGE_SHIFT, end);
1043         }
1044
1045         inode_unlock(inode);
1046
1047         if (!result) {
1048                 result = nfs_direct_wait(dreq);
1049                 if (result > 0) {
1050                         struct inode *inode = mapping->host;
1051
1052                         iocb->ki_pos = pos + result;
1053                         spin_lock(&inode->i_lock);
1054                         if (i_size_read(inode) < iocb->ki_pos)
1055                                 i_size_write(inode, iocb->ki_pos);
1056                         spin_unlock(&inode->i_lock);
1057
1058                         /* XXX: should check the generic_write_sync retval */
1059                         generic_write_sync(iocb, result);
1060                 }
1061         }
1062         nfs_direct_req_release(dreq);
1063         return result;
1064
1065 out_release:
1066         nfs_direct_req_release(dreq);
1067 out_unlock:
1068         inode_unlock(inode);
1069         return result;
1070 }
1071
1072 /**
1073  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1074  *
1075  */
1076 int __init nfs_init_directcache(void)
1077 {
1078         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1079                                                 sizeof(struct nfs_direct_req),
1080                                                 0, (SLAB_RECLAIM_ACCOUNT|
1081                                                         SLAB_MEM_SPREAD),
1082                                                 NULL);
1083         if (nfs_direct_cachep == NULL)
1084                 return -ENOMEM;
1085
1086         return 0;
1087 }
1088
1089 /**
1090  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1091  *
1092  */
1093 void nfs_destroy_directcache(void)
1094 {
1095         kmem_cache_destroy(nfs_direct_cachep);
1096 }