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_handle_truncated(struct nfs_direct_req *dreq,
 127                             const struct nfs_pgio_header *hdr,
 128                             ssize_t dreq_len)
 129 {
 130         struct nfs_direct_mirror *mirror = &dreq->mirrors[hdr->pgio_mirror_idx];
 131
 132         if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
 133               test_bit(NFS_IOHDR_EOF, &hdr->flags)))
 134                 return;
 135         if (dreq->max_count >= dreq_len) {
 136                 dreq->max_count = dreq_len;
 137                 if (dreq->count > dreq_len)
 138                         dreq->count = dreq_len;
 139
 140                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
 141                         dreq->error = hdr->error;
 142                 else /* Clear outstanding error if this is EOF */
 143                         dreq->error = 0;
 144         }
 145         if (mirror->count > dreq_len)
 146                 mirror->count = dreq_len;
 147 }
 148
 149 static void
 150 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
 151                        const struct nfs_pgio_header *hdr)
 152 {
 153         struct nfs_direct_mirror *mirror = &dreq->mirrors[hdr->pgio_mirror_idx];
 154         loff_t hdr_end = hdr->io_start + hdr->good_bytes;
 155         ssize_t dreq_len = 0;
 156
 157         if (hdr_end > dreq->io_start)
 158                 dreq_len = hdr_end - dreq->io_start;
 159
 160         nfs_direct_handle_truncated(dreq, hdr, dreq_len);
 161
 162         if (dreq_len > dreq->max_count)
 163                 dreq_len = dreq->max_count;
 164
 165         if (mirror->count < dreq_len)
 166                 mirror->count = dreq_len;
 167         if (dreq->count < dreq_len)
 168                 dreq->count = dreq_len;
 169 }
 170
 171 /*
 172  * nfs_direct_select_verf - select the right verifier
 173  * @dreq - direct request possibly spanning multiple servers
 174  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 175  * @commit_idx - commit bucket index for the DS
 176  *
 177  * returns the correct verifier to use given the role of the server
 178  */
 179 static struct nfs_writeverf *
 180 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 181                        struct nfs_client *ds_clp,
 182                        int commit_idx)
 183 {
 184         struct nfs_writeverf *verfp = &dreq->verf;
 185
 186 #ifdef CONFIG_NFS_V4_1
 187         /*
 188          * pNFS is in use, use the DS verf except commit_through_mds is set
 189          * for layout segment where nbuckets is zero.
 190          */
 191         if (ds_clp && dreq->ds_cinfo.nbuckets > 0) {
 192                 if (commit_idx >= 0 && commit_idx < dreq->ds_cinfo.nbuckets)
 193                         verfp = &dreq->ds_cinfo.buckets[commit_idx].direct_verf;
 194                 else
 195                         WARN_ON_ONCE(1);
 196         }
 197 #endif
 198         return verfp;
 199 }
 200
 201
 202 /*
 203  * nfs_direct_set_hdr_verf - set the write/commit verifier
 204  * @dreq - direct request possibly spanning multiple servers
 205  * @hdr - pageio header to validate against previously seen verfs
 206  *
 207  * Set the server's (MDS or DS) "seen" verifier
 208  */
 209 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 210                                     struct nfs_pgio_header *hdr)
 211 {
 212         struct nfs_writeverf *verfp;
 213
 214         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 215         WARN_ON_ONCE(verfp->committed >= 0);
 216         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 217         WARN_ON_ONCE(verfp->committed < 0);
 218 }
 219
 220 static int nfs_direct_cmp_verf(const struct nfs_writeverf *v1,
 221                 const struct nfs_writeverf *v2)
 222 {
 223         return nfs_write_verifier_cmp(&v1->verifier, &v2->verifier);
 224 }
 225
 226 /*
 227  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 228  * @dreq - direct request possibly spanning multiple servers
 229  * @hdr - pageio header to validate against previously seen verf
 230  *
 231  * set the server's "seen" verf if not initialized.
 232  * returns result of comparison between @hdr->verf and the "seen"
 233  * verf of the server used by @hdr (DS or MDS)
 234  */
 235 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 236                                           struct nfs_pgio_header *hdr)
 237 {
 238         struct nfs_writeverf *verfp;
 239
 240         verfp = nfs_direct_select_verf(dreq, hdr->ds_clp, hdr->ds_commit_idx);
 241         if (verfp->committed < 0) {
 242                 nfs_direct_set_hdr_verf(dreq, hdr);
 243                 return 0;
 244         }
 245         return nfs_direct_cmp_verf(verfp, &hdr->verf);
 246 }
 247
 248 /*
 249  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 250  * @dreq - direct request possibly spanning multiple servers
 251  * @data - commit data to validate against previously seen verf
 252  *
 253  * returns result of comparison between @data->verf and the verf of
 254  * the server used by @data (DS or MDS)
 255  */
 256 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 257                                            struct nfs_commit_data *data)
 258 {
 259         struct nfs_writeverf *verfp;
 260
 261         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 262                                          data->ds_commit_index);
 263
 264         /* verifier not set so always fail */
 265         if (verfp->committed < 0)
 266                 return 1;
 267
 268         return nfs_direct_cmp_verf(verfp, &data->verf);
 269 }
 270
 271 /**
 272  * nfs_direct_IO - NFS address space operation for direct I/O
 273  * @iocb: target I/O control block
 274  * @iter: I/O buffer
 275  *
 276  * The presence of this routine in the address space ops vector means
 277  * the NFS client supports direct I/O. However, for most direct IO, we
 278  * shunt off direct read and write requests before the VFS gets them,
 279  * so this method is only ever called for swap.
 280  */
 281 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 282 {
 283         struct inode *inode = iocb->ki_filp->f_mapping->host;
 284
 285         /* we only support swap file calling nfs_direct_IO */
 286         if (!IS_SWAPFILE(inode))
 287                 return 0;
 288
 289         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 290
 291         if (iov_iter_rw(iter) == READ)
 292                 return nfs_file_direct_read(iocb, iter);
 293         return nfs_file_direct_write(iocb, iter);
 294 }
 295
 296 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 297 {
 298         unsigned int i;
 299         for (i = 0; i < npages; i++)
 300                 put_page(pages[i]);
 301 }
 302
 303 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 304                               struct nfs_direct_req *dreq)
 305 {
 306         cinfo->inode = dreq->inode;
 307         cinfo->mds = &dreq->mds_cinfo;
 308         cinfo->ds = &dreq->ds_cinfo;
 309         cinfo->dreq = dreq;
 310         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 311 }
 312
 313 static inline void nfs_direct_setup_mirroring(struct nfs_direct_req *dreq,
 314                                              struct nfs_pageio_descriptor *pgio,
 315                                              struct nfs_page *req)
 316 {
 317         int mirror_count = 1;
 318
 319         if (pgio->pg_ops->pg_get_mirror_count)
 320                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 321
 322         dreq->mirror_count = mirror_count;
 323 }
 324
 325 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 326 {
 327         struct nfs_direct_req *dreq;
 328
 329         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 330         if (!dreq)
 331                 return NULL;
 332
 333         kref_init(&dreq->kref);
 334         kref_get(&dreq->kref);
 335         init_completion(&dreq->completion);
 336         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 337         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 338         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 339         dreq->mirror_count = 1;
 340         spin_lock_init(&dreq->lock);
 341
 342         return dreq;
 343 }
 344
 345 static void nfs_direct_req_free(struct kref *kref)
 346 {
 347         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 348
 349         nfs_free_pnfs_ds_cinfo(&dreq->ds_cinfo);
 350         if (dreq->l_ctx != NULL)
 351                 nfs_put_lock_context(dreq->l_ctx);
 352         if (dreq->ctx != NULL)
 353                 put_nfs_open_context(dreq->ctx);
 354         kmem_cache_free(nfs_direct_cachep, dreq);
 355 }
 356
 357 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 358 {
 359         kref_put(&dreq->kref, nfs_direct_req_free);
 360 }
 361
 362 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 363 {
 364         return dreq->bytes_left;
 365 }
 366 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 367
 368 /*
 369  * Collects and returns the final error value/byte-count.
 370  */
 371 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 372 {
 373         ssize_t result = -EIOCBQUEUED;
 374
 375         /* Async requests don't wait here */
 376         if (dreq->iocb)
 377                 goto out;
 378
 379         result = wait_for_completion_killable(&dreq->completion);
 380
 381         if (!result) {
 382                 result = dreq->count;
 383                 WARN_ON_ONCE(dreq->count < 0);
 384         }
 385         if (!result)
 386                 result = dreq->error;
 387
 388 out:
 389         return (ssize_t) result;
 390 }
 391
 392 /*
 393  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 394  * the iocb is still valid here if this is a synchronous request.
 395  */
 396 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 397 {
 398         struct inode *inode = dreq->inode;
 399
 400         inode_dio_end(inode);
 401
 402         if (dreq->iocb) {
 403                 long res = (long) dreq->error;
 404                 if (dreq->count != 0) {
 405                         res = (long) dreq->count;
 406                         WARN_ON_ONCE(dreq->count < 0);
 407                 }
 408                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 409         }
 410
 411         complete(&dreq->completion);
 412
 413         nfs_direct_req_release(dreq);
 414 }
 415
 416 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 417 {
 418         unsigned long bytes = 0;
 419         struct nfs_direct_req *dreq = hdr->dreq;
 420
 421         spin_lock(&dreq->lock);
 422         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 423                 spin_unlock(&dreq->lock);
 424                 goto out_put;
 425         }
 426
 427         nfs_direct_count_bytes(dreq, hdr);
 428         spin_unlock(&dreq->lock);
 429
 430         while (!list_empty(&hdr->pages)) {
 431                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 432                 struct page *page = req->wb_page;
 433
 434                 if (!PageCompound(page) && bytes < hdr->good_bytes &&
 435                     (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 436                         set_page_dirty(page);
 437                 bytes += req->wb_bytes;
 438                 nfs_list_remove_request(req);
 439                 nfs_release_request(req);
 440         }
 441 out_put:
 442         if (put_dreq(dreq))
 443                 nfs_direct_complete(dreq);
 444         hdr->release(hdr);
 445 }
 446
 447 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 448 {
 449         struct nfs_page *req;
 450
 451         while (!list_empty(head)) {
 452                 req = nfs_list_entry(head->next);
 453                 nfs_list_remove_request(req);
 454                 nfs_release_request(req);
 455         }
 456 }
 457
 458 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 459 {
 460         get_dreq(hdr->dreq);
 461 }
 462
 463 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 464         .error_cleanup = nfs_read_sync_pgio_error,
 465         .init_hdr = nfs_direct_pgio_init,
 466         .completion = nfs_direct_read_completion,
 467 };
 468
 469 /*
 470  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 471  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 472  * bail and stop sending more reads.  Read length accounting is
 473  * handled automatically by nfs_direct_read_result().  Otherwise, if
 474  * no requests have been sent, just return an error.
 475  */
 476
 477 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 478                                               struct iov_iter *iter,
 479                                               loff_t pos)
 480 {
 481         struct nfs_pageio_descriptor desc;
 482         struct inode *inode = dreq->inode;
 483         ssize_t result = -EINVAL;
 484         size_t requested_bytes = 0;
 485         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 486
 487         nfs_pageio_init_read(&desc, dreq->inode, false,
 488                              &nfs_direct_read_completion_ops);
 489         get_dreq(dreq);
 490         desc.pg_dreq = dreq;
 491         inode_dio_begin(inode);
 492
 493         while (iov_iter_count(iter)) {
 494                 struct page **pagevec;
 495                 size_t bytes;
 496                 size_t pgbase;
 497                 unsigned npages, i;
 498
 499                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 500                                                   rsize, &pgbase);
 501                 if (result < 0)
 502                         break;
 503
 504                 bytes = result;
 505                 iov_iter_advance(iter, bytes);
 506                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 507                 for (i = 0; i < npages; i++) {
 508                         struct nfs_page *req;
 509                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 510                         /* XXX do we need to do the eof zeroing found in async_filler? */
 511                         req = nfs_create_request(dreq->ctx, pagevec[i],
 512                                                  pgbase, req_len);
 513                         if (IS_ERR(req)) {
 514                                 result = PTR_ERR(req);
 515                                 break;
 516                         }
 517                         req->wb_index = pos >> PAGE_SHIFT;
 518                         req->wb_offset = pos & ~PAGE_MASK;
 519                         if (!nfs_pageio_add_request(&desc, req)) {
 520                                 result = desc.pg_error;
 521                                 nfs_release_request(req);
 522                                 break;
 523                         }
 524                         pgbase = 0;
 525                         bytes -= req_len;
 526                         requested_bytes += req_len;
 527                         pos += req_len;
 528                         dreq->bytes_left -= req_len;
 529                 }
 530                 nfs_direct_release_pages(pagevec, npages);
 531                 kvfree(pagevec);
 532                 if (result < 0)
 533                         break;
 534         }
 535
 536         nfs_pageio_complete(&desc);
 537
 538         /*
 539          * If no bytes were started, return the error, and let the
 540          * generic layer handle the completion.
 541          */
 542         if (requested_bytes == 0) {
 543                 inode_dio_end(inode);
 544                 nfs_direct_req_release(dreq);
 545                 return result < 0 ? result : -EIO;
 546         }
 547
 548         if (put_dreq(dreq))
 549                 nfs_direct_complete(dreq);
 550         return requested_bytes;
 551 }
 552
 553 /**
 554  * nfs_file_direct_read - file direct read operation for NFS files
 555  * @iocb: target I/O control block
 556  * @iter: vector of user buffers into which to read data
 557  *
 558  * We use this function for direct reads instead of calling
 559  * generic_file_aio_read() in order to avoid gfar's check to see if
 560  * the request starts before the end of the file.  For that check
 561  * to work, we must generate a GETATTR before each direct read, and
 562  * even then there is a window between the GETATTR and the subsequent
 563  * READ where the file size could change.  Our preference is simply
 564  * to do all reads the application wants, and the server will take
 565  * care of managing the end of file boundary.
 566  *
 567  * This function also eliminates unnecessarily updating the file's
 568  * atime locally, as the NFS server sets the file's atime, and this
 569  * client must read the updated atime from the server back into its
 570  * cache.
 571  */
 572 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter)
 573 {
 574         struct file *file = iocb->ki_filp;
 575         struct address_space *mapping = file->f_mapping;
 576         struct inode *inode = mapping->host;
 577         struct nfs_direct_req *dreq;
 578         struct nfs_lock_context *l_ctx;
 579         ssize_t result = -EINVAL, requested;
 580         size_t count = iov_iter_count(iter);
 581         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 582
 583         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 584                 file, count, (long long) iocb->ki_pos);
 585
 586         result = 0;
 587         if (!count)
 588                 goto out;
 589
 590         task_io_account_read(count);
 591
 592         result = -ENOMEM;
 593         dreq = nfs_direct_req_alloc();
 594         if (dreq == NULL)
 595                 goto out;
 596
 597         dreq->inode = inode;
 598         dreq->bytes_left = dreq->max_count = count;
 599         dreq->io_start = iocb->ki_pos;
 600         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 601         l_ctx = nfs_get_lock_context(dreq->ctx);
 602         if (IS_ERR(l_ctx)) {
 603                 result = PTR_ERR(l_ctx);
 604                 goto out_release;
 605         }
 606         dreq->l_ctx = l_ctx;
 607         if (!is_sync_kiocb(iocb))
 608                 dreq->iocb = iocb;
 609
 610         if (iter_is_iovec(iter))
 611                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 612
 613         nfs_start_io_direct(inode);
 614
 615         NFS_I(inode)->read_io += count;
 616         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 617
 618         nfs_end_io_direct(inode);
 619
 620         if (requested > 0) {
 621                 result = nfs_direct_wait(dreq);
 622                 if (result > 0) {
 623                         requested -= result;
 624                         iocb->ki_pos += result;
 625                 }
 626                 iov_iter_revert(iter, requested);
 627         } else {
 628                 result = requested;
 629         }
 630
 631 out_release:
 632         nfs_direct_req_release(dreq);
 633 out:
 634         return result;
 635 }
 636
 637 static void
 638 nfs_direct_write_scan_commit_list(struct inode *inode,
 639                                   struct list_head *list,
 640                                   struct nfs_commit_info *cinfo)
 641 {
 642         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 643 #ifdef CONFIG_NFS_V4_1
 644         if (cinfo->ds != NULL && cinfo->ds->nwritten != 0)
 645                 NFS_SERVER(inode)->pnfs_curr_ld->recover_commit_reqs(list, cinfo);
 646 #endif
 647         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 648         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 649 }
 650
 651 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 652 {
 653         struct nfs_pageio_descriptor desc;
 654         struct nfs_page *req, *tmp;
 655         LIST_HEAD(reqs);
 656         struct nfs_commit_info cinfo;
 657         LIST_HEAD(failed);
 658         int i;
 659
 660         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 661         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 662
 663         dreq->count = 0;
 664         dreq->max_count = 0;
 665         list_for_each_entry(req, &reqs, wb_list)
 666                 dreq->max_count += req->wb_bytes;
 667         dreq->verf.committed = NFS_INVALID_STABLE_HOW;
 668         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 669         for (i = 0; i < dreq->mirror_count; i++)
 670                 dreq->mirrors[i].count = 0;
 671         get_dreq(dreq);
 672
 673         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 674                               &nfs_direct_write_completion_ops);
 675         desc.pg_dreq = dreq;
 676
 677         req = nfs_list_entry(reqs.next);
 678         nfs_direct_setup_mirroring(dreq, &desc, req);
 679         if (desc.pg_error < 0) {
 680                 list_splice_init(&reqs, &failed);
 681                 goto out_failed;
 682         }
 683
 684         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 685                 /* Bump the transmission count */
 686                 req->wb_nio++;
 687                 if (!nfs_pageio_add_request(&desc, req)) {
 688                         nfs_list_move_request(req, &failed);
 689                         spin_lock(&cinfo.inode->i_lock);
 690                         dreq->flags = 0;
 691                         if (desc.pg_error < 0)
 692                                 dreq->error = desc.pg_error;
 693                         else
 694                                 dreq->error = -EIO;
 695                         spin_unlock(&cinfo.inode->i_lock);
 696                 }
 697                 nfs_release_request(req);
 698         }
 699         nfs_pageio_complete(&desc);
 700
 701 out_failed:
 702         while (!list_empty(&failed)) {
 703                 req = nfs_list_entry(failed.next);
 704                 nfs_list_remove_request(req);
 705                 nfs_unlock_and_release_request(req);
 706         }
 707
 708         if (put_dreq(dreq))
 709                 nfs_direct_write_complete(dreq);
 710 }
 711
 712 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 713 {
 714         struct nfs_direct_req *dreq = data->dreq;
 715         struct nfs_commit_info cinfo;
 716         struct nfs_page *req;
 717         int status = data->task.tk_status;
 718
 719         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 720         if (status < 0 || nfs_direct_cmp_commit_data_verf(dreq, data))
 721                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 722
 723         while (!list_empty(&data->pages)) {
 724                 req = nfs_list_entry(data->pages.next);
 725                 nfs_list_remove_request(req);
 726                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 727                         /*
 728                          * Despite the reboot, the write was successful,
 729                          * so reset wb_nio.
 730                          */
 731                         req->wb_nio = 0;
 732                         /* Note the rewrite will go through mds */
 733                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 734                 } else
 735                         nfs_release_request(req);
 736                 nfs_unlock_and_release_request(req);
 737         }
 738
 739         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 740                 nfs_direct_write_complete(dreq);
 741 }
 742
 743 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 744                 struct nfs_page *req)
 745 {
 746         struct nfs_direct_req *dreq = cinfo->dreq;
 747
 748         spin_lock(&dreq->lock);
 749         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 750         spin_unlock(&dreq->lock);
 751         nfs_mark_request_commit(req, NULL, cinfo, 0);
 752 }
 753
 754 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 755         .completion = nfs_direct_commit_complete,
 756         .resched_write = nfs_direct_resched_write,
 757 };
 758
 759 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 760 {
 761         int res;
 762         struct nfs_commit_info cinfo;
 763         LIST_HEAD(mds_list);
 764
 765         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 766         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 767         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 768         if (res < 0) /* res == -ENOMEM */
 769                 nfs_direct_write_reschedule(dreq);
 770 }
 771
 772 static void nfs_direct_write_schedule_work(struct work_struct *work)
 773 {
 774         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 775         int flags = dreq->flags;
 776
 777         dreq->flags = 0;
 778         switch (flags) {
 779                 case NFS_ODIRECT_DO_COMMIT:
 780                         nfs_direct_commit_schedule(dreq);
 781                         break;
 782                 case NFS_ODIRECT_RESCHED_WRITES:
 783                         nfs_direct_write_reschedule(dreq);
 784                         break;
 785                 default:
 786                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 787                         nfs_direct_complete(dreq);
 788         }
 789 }
 790
 791 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 792 {
 793         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 794 }
 795
 796 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 797 {
 798         struct nfs_direct_req *dreq = hdr->dreq;
 799         struct nfs_commit_info cinfo;
 800         bool request_commit = false;
 801         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 802
 803         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 804
 805         spin_lock(&dreq->lock);
 806         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 807                 spin_unlock(&dreq->lock);
 808                 goto out_put;
 809         }
 810
 811         nfs_direct_count_bytes(dreq, hdr);
 812         if (hdr->good_bytes != 0) {
 813                 if (nfs_write_need_commit(hdr)) {
 814                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 815                                 request_commit = true;
 816                         else if (dreq->flags == 0) {
 817                                 nfs_direct_set_hdr_verf(dreq, hdr);
 818                                 request_commit = true;
 819                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 820                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 821                                 request_commit = true;
 822                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr))
 823                                         dreq->flags =
 824                                                 NFS_ODIRECT_RESCHED_WRITES;
 825                         }
 826                 }
 827         }
 828         spin_unlock(&dreq->lock);
 829
 830         while (!list_empty(&hdr->pages)) {
 831
 832                 req = nfs_list_entry(hdr->pages.next);
 833                 nfs_list_remove_request(req);
 834                 if (request_commit) {
 835                         kref_get(&req->wb_kref);
 836                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 837                                 hdr->ds_commit_idx);
 838                 }
 839                 nfs_unlock_and_release_request(req);
 840         }
 841
 842 out_put:
 843         if (put_dreq(dreq))
 844                 nfs_direct_write_complete(dreq);
 845         hdr->release(hdr);
 846 }
 847
 848 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 849 {
 850         struct nfs_page *req;
 851
 852         while (!list_empty(head)) {
 853                 req = nfs_list_entry(head->next);
 854                 nfs_list_remove_request(req);
 855                 nfs_unlock_and_release_request(req);
 856         }
 857 }
 858
 859 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 860 {
 861         struct nfs_direct_req *dreq = hdr->dreq;
 862
 863         spin_lock(&dreq->lock);
 864         if (dreq->error == 0) {
 865                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 866                 /* fake unstable write to let common nfs resend pages */
 867                 hdr->verf.committed = NFS_UNSTABLE;
 868                 hdr->good_bytes = hdr->args.count;
 869         }
 870         spin_unlock(&dreq->lock);
 871 }
 872
 873 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 874         .error_cleanup = nfs_write_sync_pgio_error,
 875         .init_hdr = nfs_direct_pgio_init,
 876         .completion = nfs_direct_write_completion,
 877         .reschedule_io = nfs_direct_write_reschedule_io,
 878 };
 879
 880
 881 /*
 882  * NB: Return the value of the first error return code.  Subsequent
 883  *     errors after the first one are ignored.
 884  */
 885 /*
 886  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 887  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 888  * bail and stop sending more writes.  Write length accounting is
 889  * handled automatically by nfs_direct_write_result().  Otherwise, if
 890  * no requests have been sent, just return an error.
 891  */
 892 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 893                                                struct iov_iter *iter,
 894                                                loff_t pos)
 895 {
 896         struct nfs_pageio_descriptor desc;
 897         struct inode *inode = dreq->inode;
 898         ssize_t result = 0;
 899         size_t requested_bytes = 0;
 900         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 901
 902         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 903                               &nfs_direct_write_completion_ops);
 904         desc.pg_dreq = dreq;
 905         get_dreq(dreq);
 906         inode_dio_begin(inode);
 907
 908         NFS_I(inode)->write_io += iov_iter_count(iter);
 909         while (iov_iter_count(iter)) {
 910                 struct page **pagevec;
 911                 size_t bytes;
 912                 size_t pgbase;
 913                 unsigned npages, i;
 914
 915                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 916                                                   wsize, &pgbase);
 917                 if (result < 0)
 918                         break;
 919
 920                 bytes = result;
 921                 iov_iter_advance(iter, bytes);
 922                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 923                 for (i = 0; i < npages; i++) {
 924                         struct nfs_page *req;
 925                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 926
 927                         req = nfs_create_request(dreq->ctx, pagevec[i],
 928                                                  pgbase, req_len);
 929                         if (IS_ERR(req)) {
 930                                 result = PTR_ERR(req);
 931                                 break;
 932                         }
 933
 934                         nfs_direct_setup_mirroring(dreq, &desc, req);
 935                         if (desc.pg_error < 0) {
 936                                 nfs_free_request(req);
 937                                 result = desc.pg_error;
 938                                 break;
 939                         }
 940
 941                         nfs_lock_request(req);
 942                         req->wb_index = pos >> PAGE_SHIFT;
 943                         req->wb_offset = pos & ~PAGE_MASK;
 944                         if (!nfs_pageio_add_request(&desc, req)) {
 945                                 result = desc.pg_error;
 946                                 nfs_unlock_and_release_request(req);
 947                                 break;
 948                         }
 949                         pgbase = 0;
 950                         bytes -= req_len;
 951                         requested_bytes += req_len;
 952                         pos += req_len;
 953                         dreq->bytes_left -= req_len;
 954                 }
 955                 nfs_direct_release_pages(pagevec, npages);
 956                 kvfree(pagevec);
 957                 if (result < 0)
 958                         break;
 959         }
 960         nfs_pageio_complete(&desc);
 961
 962         /*
 963          * If no bytes were started, return the error, and let the
 964          * generic layer handle the completion.
 965          */
 966         if (requested_bytes == 0) {
 967                 inode_dio_end(inode);
 968                 nfs_direct_req_release(dreq);
 969                 return result < 0 ? result : -EIO;
 970         }
 971
 972         if (put_dreq(dreq))
 973                 nfs_direct_write_complete(dreq);
 974         return requested_bytes;
 975 }
 976
 977 /**
 978  * nfs_file_direct_write - file direct write operation for NFS files
 979  * @iocb: target I/O control block
 980  * @iter: vector of user buffers from which to write data
 981  *
 982  * We use this function for direct writes instead of calling
 983  * generic_file_aio_write() in order to avoid taking the inode
 984  * semaphore and updating the i_size.  The NFS server will set
 985  * the new i_size and this client must read the updated size
 986  * back into its cache.  We let the server do generic write
 987  * parameter checking and report problems.
 988  *
 989  * We eliminate local atime updates, see direct read above.
 990  *
 991  * We avoid unnecessary page cache invalidations for normal cached
 992  * readers of this file.
 993  *
 994  * Note that O_APPEND is not supported for NFS direct writes, as there
 995  * is no atomic O_APPEND write facility in the NFS protocol.
 996  */
 997 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter)
 998 {
 999         ssize_t result = -EINVAL, requested;
1000         size_t count;
1001         struct file *file = iocb->ki_filp;
1002         struct address_space *mapping = file->f_mapping;
1003         struct inode *inode = mapping->host;
1004         struct nfs_direct_req *dreq;
1005         struct nfs_lock_context *l_ctx;
1006         loff_t pos, end;
1007
1008         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1009                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
1010
1011         result = generic_write_checks(iocb, iter);
1012         if (result <= 0)
1013                 return result;
1014         count = result;
1015         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1016
1017         pos = iocb->ki_pos;
1018         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
1019
1020         task_io_account_write(count);
1021
1022         result = -ENOMEM;
1023         dreq = nfs_direct_req_alloc();
1024         if (!dreq)
1025                 goto out;
1026
1027         dreq->inode = inode;
1028         dreq->bytes_left = dreq->max_count = count;
1029         dreq->io_start = pos;
1030         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1031         l_ctx = nfs_get_lock_context(dreq->ctx);
1032         if (IS_ERR(l_ctx)) {
1033                 result = PTR_ERR(l_ctx);
1034                 goto out_release;
1035         }
1036         dreq->l_ctx = l_ctx;
1037         if (!is_sync_kiocb(iocb))
1038                 dreq->iocb = iocb;
1039
1040         nfs_start_io_direct(inode);
1041
1042         requested = nfs_direct_write_schedule_iovec(dreq, iter, pos);
1043
1044         if (mapping->nrpages) {
1045                 invalidate_inode_pages2_range(mapping,
1046                                               pos >> PAGE_SHIFT, end);
1047         }
1048
1049         nfs_end_io_direct(inode);
1050
1051         if (requested > 0) {
1052                 result = nfs_direct_wait(dreq);
1053                 if (result > 0) {
1054                         requested -= result;
1055                         iocb->ki_pos = pos + result;
1056                         /* XXX: should check the generic_write_sync retval */
1057                         generic_write_sync(iocb, result);
1058                 }
1059                 iov_iter_revert(iter, requested);
1060         } else {
1061                 result = requested;
1062         }
1063 out_release:
1064         nfs_direct_req_release(dreq);
1065 out:
1066         return result;
1067 }
1068
1069 /**
1070  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1071  *
1072  */
1073 int __init nfs_init_directcache(void)
1074 {
1075         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1076                                                 sizeof(struct nfs_direct_req),
1077                                                 0, (SLAB_RECLAIM_ACCOUNT|
1078                                                         SLAB_MEM_SPREAD),
1079                                                 NULL);
1080         if (nfs_direct_cachep == NULL)
1081                 return -ENOMEM;
1082
1083         return 0;
1084 }
1085
1086 /**
1087  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1088  *
1089  */
1090 void nfs_destroy_directcache(void)
1091 {
1092         kmem_cache_destroy(nfs_direct_cachep);
1093 }