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         spin_lock(&dreq->lock);
 405         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
 406                 dreq->error = hdr->error;
 407
 408         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 409                 spin_unlock(&dreq->lock);
 410                 goto out_put;
 411         }
 412
 413         if (hdr->good_bytes != 0)
 414                 nfs_direct_good_bytes(dreq, hdr);
 415
 416         if (test_bit(NFS_IOHDR_EOF, &hdr->flags))
 417                 dreq->error = 0;
 418
 419         spin_unlock(&dreq->lock);
 420
 421         while (!list_empty(&hdr->pages)) {
 422                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 423                 struct page *page = req->wb_page;
 424
 425                 if (!PageCompound(page) && bytes < hdr->good_bytes &&
 426                     (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 427                         set_page_dirty(page);
 428                 bytes += req->wb_bytes;
 429                 nfs_list_remove_request(req);
 430                 nfs_release_request(req);
 431         }
 432 out_put:
 433         if (put_dreq(dreq))
 434                 nfs_direct_complete(dreq);
 435         hdr->release(hdr);
 436 }
 437
 438 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 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],
 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);
 541         return requested_bytes;
 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, requested;
 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         task_io_account_read(count);
 582
 583         result = -ENOMEM;
 584         dreq = nfs_direct_req_alloc();
 585         if (dreq == NULL)
 586                 goto out;
 587
 588         dreq->inode = inode;
 589         dreq->bytes_left = dreq->max_count = count;
 590         dreq->io_start = iocb->ki_pos;
 591         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 592         l_ctx = nfs_get_lock_context(dreq->ctx);
 593         if (IS_ERR(l_ctx)) {
 594                 result = PTR_ERR(l_ctx);
 595                 goto out_release;
 596         }
 597         dreq->l_ctx = l_ctx;
 598         if (!is_sync_kiocb(iocb))
 599                 dreq->iocb = iocb;
 600
 601         if (iter_is_iovec(iter))
 602                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 603
 604         nfs_start_io_direct(inode);
 605
 606         NFS_I(inode)->read_io += count;
 607         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 608
 609         nfs_end_io_direct(inode);
 610
 611         if (requested > 0) {
 612                 result = nfs_direct_wait(dreq);
 613                 if (result > 0) {
 614                         requested -= result;
 615                         iocb->ki_pos += result;
 616                 }
 617                 iov_iter_revert(iter, requested);
 618         } else {
 619                 result = requested;
 620         }
 621
 622 out_release:
 623         nfs_direct_req_release(dreq);
 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         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 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         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 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         dreq->verf.committed = NFS_INVALID_STABLE_HOW;
 656         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 657         for (i = 0; i < dreq->mirror_count; i++)
 658                 dreq->mirrors[i].count = 0;
 659         get_dreq(dreq);
 660
 661         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 662                               &nfs_direct_write_completion_ops);
 663         desc.pg_dreq = dreq;
 664
 665         req = nfs_list_entry(reqs.next);
 666         nfs_direct_setup_mirroring(dreq, &desc, req);
 667         if (desc.pg_error < 0) {
 668                 list_splice_init(&reqs, &failed);
 669                 goto out_failed;
 670         }
 671
 672         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 673                 /* Bump the transmission count */
 674                 req->wb_nio++;
 675                 if (!nfs_pageio_add_request(&desc, req)) {
 676                         nfs_list_move_request(req, &failed);
 677                         spin_lock(&cinfo.inode->i_lock);
 678                         dreq->flags = 0;
 679                         if (desc.pg_error < 0)
 680                                 dreq->error = desc.pg_error;
 681                         else
 682                                 dreq->error = -EIO;
 683                         spin_unlock(&cinfo.inode->i_lock);
 684                 }
 685                 nfs_release_request(req);
 686         }
 687         nfs_pageio_complete(&desc);
 688
 689 out_failed:
 690         while (!list_empty(&failed)) {
 691                 req = nfs_list_entry(failed.next);
 692                 nfs_list_remove_request(req);
 693                 nfs_unlock_and_release_request(req);
 694         }
 695
 696         if (put_dreq(dreq))
 697                 nfs_direct_write_complete(dreq);
 698 }
 699
 700 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 701 {
 702         struct nfs_direct_req *dreq = data->dreq;
 703         struct nfs_commit_info cinfo;
 704         struct nfs_page *req;
 705         int status = data->task.tk_status;
 706
 707         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 708         if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
 709                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 710
 711         while (!list_empty(&data->pages)) {
 712                 req = nfs_list_entry(data->pages.next);
 713                 nfs_list_remove_request(req);
 714                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 715                         /*
 716                          * Despite the reboot, the write was successful,
 717                          * so reset wb_nio.
 718                          */
 719                         req->wb_nio = 0;
 720                         /* Note the rewrite will go through mds */
 721                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 722                 } else
 723                         nfs_release_request(req);
 724                 nfs_unlock_and_release_request(req);
 725         }
 726
 727         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 728                 nfs_direct_write_complete(dreq);
 729 }
 730
 731 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 732                 struct nfs_page *req)
 733 {
 734         struct nfs_direct_req *dreq = cinfo->dreq;
 735
 736         spin_lock(&dreq->lock);
 737         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 738         spin_unlock(&dreq->lock);
 739         nfs_mark_request_commit(req, NULL, cinfo, 0);
 740 }
 741
 742 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 743         .completion = nfs_direct_commit_complete,
 744         .resched_write = nfs_direct_resched_write,
 745 };
 746
 747 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 748 {
 749         int res;
 750         struct nfs_commit_info cinfo;
 751         LIST_HEAD(mds_list);
 752
 753         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 754         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 755         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 756         if (res < 0) /* res == -ENOMEM */
 757                 nfs_direct_write_reschedule(dreq);
 758 }
 759
 760 static void nfs_direct_write_schedule_work(struct work_struct *work)
 761 {
 762         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 763         int flags = dreq->flags;
 764
 765         dreq->flags = 0;
 766         switch (flags) {
 767                 case NFS_ODIRECT_DO_COMMIT:
 768                         nfs_direct_commit_schedule(dreq);
 769                         break;
 770                 case NFS_ODIRECT_RESCHED_WRITES:
 771                         nfs_direct_write_reschedule(dreq);
 772                         break;
 773                 default:
 774                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 775                         nfs_direct_complete(dreq);
 776         }
 777 }
 778
 779 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 780 {
 781         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 782 }
 783
 784 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 785 {
 786         struct nfs_direct_req *dreq = hdr->dreq;
 787         struct nfs_commit_info cinfo;
 788         bool request_commit = false;
 789         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 790
 791         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 792
 793         spin_lock(&dreq->lock);
 794
 795         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
 796                 dreq->error = hdr->error;
 797
 798         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 799                 spin_unlock(&dreq->lock);
 800                 goto out_put;
 801         }
 802
 803         if (hdr->good_bytes != 0) {
 804                 nfs_direct_good_bytes(dreq, hdr);
 805                 if (nfs_write_need_commit(hdr)) {
 806                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 807                                 request_commit = true;
 808                         else if (dreq->flags == 0) {
 809                                 nfs_direct_set_hdr_verf(dreq, hdr);
 810                                 request_commit = true;
 811                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 812                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 813                                 request_commit = true;
 814                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 815                                         dreq->flags =
 816                                                 NFS_ODIRECT_RESCHED_WRITES;
 817                         }
 818                 }
 819         }
 820         spin_unlock(&dreq->lock);
 821
 822         while (!list_empty(&hdr->pages)) {
 823
 824                 req = nfs_list_entry(hdr->pages.next);
 825                 nfs_list_remove_request(req);
 826                 if (request_commit) {
 827                         kref_get(&req->wb_kref);
 828                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 829                                 hdr->ds_commit_idx);
 830                 }
 831                 nfs_unlock_and_release_request(req);
 832         }
 833
 834 out_put:
 835         if (put_dreq(dreq))
 836                 nfs_direct_write_complete(dreq);
 837         hdr->release(hdr);
 838 }
 839
 840 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 841 {
 842         struct nfs_page *req;
 843
 844         while (!list_empty(head)) {
 845                 req = nfs_list_entry(head->next);
 846                 nfs_list_remove_request(req);
 847                 nfs_unlock_and_release_request(req);
 848         }
 849 }
 850
 851 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 852 {
 853         struct nfs_direct_req *dreq = hdr->dreq;
 854
 855         spin_lock(&dreq->lock);
 856         if (dreq->error == 0) {
 857                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 858                 /* fake unstable write to let common nfs resend pages */
 859                 hdr->verf.committed = NFS_UNSTABLE;
 860                 hdr->good_bytes = hdr->args.count;
 861         }
 862         spin_unlock(&dreq->lock);
 863 }
 864
 865 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 866         .error_cleanup = nfs_write_sync_pgio_error,
 867         .init_hdr = nfs_direct_pgio_init,
 868         .completion = nfs_direct_write_completion,
 869         .reschedule_io = nfs_direct_write_reschedule_io,
 870 };
 871
 872
 873 /*
 874  * NB: Return the value of the first error return code.  Subsequent
 875  *     errors after the first one are ignored.
 876  */
 877 /*
 878  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 879  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 880  * bail and stop sending more writes.  Write length accounting is
 881  * handled automatically by nfs_direct_write_result().  Otherwise, if
 882  * no requests have been sent, just return an error.
 883  */
 884 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 885                                                struct iov_iter *iter,
 886                                                loff_t pos)
 887 {
 888         struct nfs_pageio_descriptor desc;
 889         struct inode *inode = dreq->inode;
 890         ssize_t result = 0;
 891         size_t requested_bytes = 0;
 892         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 893
 894         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 895                               &nfs_direct_write_completion_ops);
 896         desc.pg_dreq = dreq;
 897         get_dreq(dreq);
 898         inode_dio_begin(inode);
 899
 900         NFS_I(inode)->write_io += iov_iter_count(iter);
 901         while (iov_iter_count(iter)) {
 902                 struct page **pagevec;
 903                 size_t bytes;
 904                 size_t pgbase;
 905                 unsigned npages, i;
 906
 907                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 908                                                   wsize, &pgbase);
 909                 if (result < 0)
 910                         break;
 911
 912                 bytes = result;
 913                 iov_iter_advance(iter, bytes);
 914                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 915                 for (i = 0; i < npages; i++) {
 916                         struct nfs_page *req;
 917                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 918
 919                         req = nfs_create_request(dreq->ctx, pagevec[i],
 920                                                  pgbase, req_len);
 921                         if (IS_ERR(req)) {
 922                                 result = PTR_ERR(req);
 923                                 break;
 924                         }
 925
 926                         nfs_direct_setup_mirroring(dreq, &desc, req);
 927                         if (desc.pg_error < 0) {
 928                                 nfs_free_request(req);
 929                                 result = desc.pg_error;
 930                                 break;
 931                         }
 932
 933                         nfs_lock_request(req);
 934                         req->wb_index = pos >> PAGE_SHIFT;
 935                         req->wb_offset = pos & ~PAGE_MASK;
 936                         if (!nfs_pageio_add_request(&desc, req)) {
 937                                 result = desc.pg_error;
 938                                 nfs_unlock_and_release_request(req);
 939                                 break;
 940                         }
 941                         pgbase = 0;
 942                         bytes -= req_len;
 943                         requested_bytes += req_len;
 944                         pos += req_len;
 945                         dreq->bytes_left -= req_len;
 946                 }
 947                 nfs_direct_release_pages(pagevec, npages);
 948                 kvfree(pagevec);
 949                 if (result < 0)
 950                         break;
 951         }
 952         nfs_pageio_complete(&desc);
 953
 954         /*
 955          * If no bytes were started, return the error, and let the
 956          * generic layer handle the completion.
 957          */
 958         if (requested_bytes == 0) {
 959                 inode_dio_end(inode);
 960                 nfs_direct_req_release(dreq);
 961                 return result < 0 ? result : -EIO;
 962         }
 963
 964         if (put_dreq(dreq))
 965                 nfs_direct_write_complete(dreq);
 966         return requested_bytes;
 967 }
 968
 969 /**
 970  * nfs_file_direct_write - file direct write operation for NFS files
 971  * @iocb: target I/O control block
 972  * @iter: vector of user buffers from which to write data
 973  *
 974  * We use this function for direct writes instead of calling
 975  * generic_file_aio_write() in order to avoid taking the inode
 976  * semaphore and updating the i_size.  The NFS server will set
 977  * the new i_size and this client must read the updated size
 978  * back into its cache.  We let the server do generic write
 979  * parameter checking and report problems.
 980  *
 981  * We eliminate local atime updates, see direct read above.
 982  *
 983  * We avoid unnecessary page cache invalidations for normal cached
 984  * readers of this file.
 985  *
 986  * Note that O_APPEND is not supported for NFS direct writes, as there
 987  * is no atomic O_APPEND write facility in the NFS protocol.
 988  */
 989 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 990 {
 991         ssize_t result = -EINVAL, requested;
 992         size_t count;
 993         struct file *file = iocb->ki_filp;
 994         struct address_space *mapping = file->f_mapping;
 995         struct inode *inode = mapping->host;
 996         struct nfs_direct_req *dreq;
 997         struct nfs_lock_context *l_ctx;
 998         loff_t pos, end;
 999
1000         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1001                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
1002
1003         result = generic_write_checks(iocb, iter);
1004         if (result <= 0)
1005                 return result;
1006         count = result;
1007         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1008
1009         pos = iocb->ki_pos;
1010         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1011
1012         task_io_account_write(count);
1013
1014         result = -ENOMEM;
1015         dreq = nfs_direct_req_alloc();
1016         if (!dreq)
1017                 goto out;
1018
1019         dreq->inode = inode;
1020         dreq->bytes_left = dreq->max_count = count;
1021         dreq->io_start = pos;
1022         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1023         l_ctx = nfs_get_lock_context(dreq->ctx);
1024         if (IS_ERR(l_ctx)) {
1025                 result = PTR_ERR(l_ctx);
1026                 goto out_release;
1027         }
1028         dreq->l_ctx = l_ctx;
1029         if (!is_sync_kiocb(iocb))
1030                 dreq->iocb = iocb;
1031
1032         nfs_start_io_direct(inode);
1033
1034         requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1035
1036         if (mapping->nrpages) {
1037                 invalidate_inode_pages2_range(mapping,
1038                                               pos >> PAGE_SHIFT, end);
1039         }
1040
1041         nfs_end_io_direct(inode);
1042
1043         if (requested > 0) {
1044                 result = nfs_direct_wait(dreq);
1045                 if (result > 0) {
1046                         requested -= result;
1047                         iocb->ki_pos = pos + result;
1048                         /* XXX: should check the generic_write_sync retval */
1049                         generic_write_sync(iocb, result);
1050                 }
1051                 iov_iter_revert(iter, requested);
1052         } else {
1053                 result = requested;
1054         }
1055 out_release:
1056         nfs_direct_req_release(dreq);
1057 out:
1058         return result;
1059 }
1060
1061 /**
1062  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1063  *
1064  */
1065 int __init nfs_init_directcache(void)
1066 {
1067         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1068                                                 sizeof(struct nfs_direct_req),
1069                                                 0, (SLAB_RECLAIM_ACCOUNT|
1070                                                         SLAB_MEM_SPREAD),
1071                                                 NULL);
1072         if (nfs_direct_cachep == NULL)
1073                 return -ENOMEM;
1074
1075         return 0;
1076 }
1077
1078 /**
1079  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1080  *
1081  */
1082 void nfs_destroy_directcache(void)
1083 {
1084         kmem_cache_destroy(nfs_direct_cachep);
1085 }