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 struct nfs_direct_req {
  68         struct kref             kref;           /* release manager */
  69
  70         /* I/O parameters */
  71         struct nfs_open_context *ctx;           /* file open context info */
  72         struct nfs_lock_context *l_ctx;         /* Lock context info */
  73         struct kiocb *          iocb;           /* controlling i/o request */
  74         struct inode *          inode;          /* target file of i/o */
  75
  76         /* completion state */
  77         atomic_t                io_count;       /* i/os we're waiting for */
  78         spinlock_t              lock;           /* protect completion state */
  79
  80         loff_t                  io_start;       /* Start offset for I/O */
  81         ssize_t                 count,          /* bytes actually processed */
  82                                 max_count,      /* max expected count */
  83                                 bytes_left,     /* bytes left to be sent */
  84                                 error;          /* any reported error */
  85         struct completion       completion;     /* wait for i/o completion */
  86
  87         /* commit state */
  88         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  89         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  90         struct work_struct      work;
  91         int                     flags;
  92         /* for write */
  93 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  94 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  95         /* for read */
  96 #define NFS_ODIRECT_SHOULD_DIRTY        (3)     /* dirty user-space page after read */
  97 #define NFS_ODIRECT_DONE                INT_MAX /* write verification failed */
  98 };
  99
 100 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
 101 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
 102 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
 103 static void nfs_direct_write_schedule_work(struct work_struct *work);
 104
 105 static inline void get_dreq(struct nfs_direct_req *dreq)
 106 {
 107         atomic_inc(&dreq->io_count);
 108 }
 109
 110 static inline int put_dreq(struct nfs_direct_req *dreq)
 111 {
 112         return atomic_dec_and_test(&dreq->io_count);
 113 }
 114
 115 static void
 116 nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
 117                             const struct nfs_pgio_header *hdr,
 118                             ssize_t dreq_len)
 119 {
 120         if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
 121               test_bit(NFS_IOHDR_EOF, &hdr->flags)))
 122                 return;
 123         if (dreq->max_count >= dreq_len) {
 124                 dreq->max_count = dreq_len;
 125                 if (dreq->count > dreq_len)
 126                         dreq->count = dreq_len;
 127
 128                 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
 129                         dreq->error = hdr->error;
 130                 else /* Clear outstanding error if this is EOF */
 131                         dreq->error = 0;
 132         }
 133 }
 134
 135 static void
 136 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
 137                        const struct nfs_pgio_header *hdr)
 138 {
 139         loff_t hdr_end = hdr->io_start + hdr->good_bytes;
 140         ssize_t dreq_len = 0;
 141
 142         if (hdr_end > dreq->io_start)
 143                 dreq_len = hdr_end - dreq->io_start;
 144
 145         nfs_direct_handle_truncated(dreq, hdr, dreq_len);
 146
 147         if (dreq_len > dreq->max_count)
 148                 dreq_len = dreq->max_count;
 149
 150         if (dreq->count < dreq_len)
 151                 dreq->count = dreq_len;
 152 }
 153
 154 /**
 155  * nfs_direct_IO - NFS address space operation for direct I/O
 156  * @iocb: target I/O control block
 157  * @iter: I/O buffer
 158  *
 159  * The presence of this routine in the address space ops vector means
 160  * the NFS client supports direct I/O. However, for most direct IO, we
 161  * shunt off direct read and write requests before the VFS gets them,
 162  * so this method is only ever called for swap.
 163  */
 164 ssize_t nfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 165 {
 166         struct inode *inode = iocb->ki_filp->f_mapping->host;
 167
 168         /* we only support swap file calling nfs_direct_IO */
 169         if (!IS_SWAPFILE(inode))
 170                 return 0;
 171
 172         VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
 173
 174         if (iov_iter_rw(iter) == READ)
 175                 return nfs_file_direct_read(iocb, iter, true);
 176         return nfs_file_direct_write(iocb, iter, true);
 177 }
 178
 179 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 180 {
 181         unsigned int i;
 182         for (i = 0; i < npages; i++)
 183                 put_page(pages[i]);
 184 }
 185
 186 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 187                               struct nfs_direct_req *dreq)
 188 {
 189         cinfo->inode = dreq->inode;
 190         cinfo->mds = &dreq->mds_cinfo;
 191         cinfo->ds = &dreq->ds_cinfo;
 192         cinfo->dreq = dreq;
 193         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 194 }
 195
 196 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 197 {
 198         struct nfs_direct_req *dreq;
 199
 200         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 201         if (!dreq)
 202                 return NULL;
 203
 204         kref_init(&dreq->kref);
 205         kref_get(&dreq->kref);
 206         init_completion(&dreq->completion);
 207         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 208         pnfs_init_ds_commit_info(&dreq->ds_cinfo);
 209         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 210         spin_lock_init(&dreq->lock);
 211
 212         return dreq;
 213 }
 214
 215 static void nfs_direct_req_free(struct kref *kref)
 216 {
 217         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 218
 219         pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
 220         if (dreq->l_ctx != NULL)
 221                 nfs_put_lock_context(dreq->l_ctx);
 222         if (dreq->ctx != NULL)
 223                 put_nfs_open_context(dreq->ctx);
 224         kmem_cache_free(nfs_direct_cachep, dreq);
 225 }
 226
 227 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 228 {
 229         kref_put(&dreq->kref, nfs_direct_req_free);
 230 }
 231
 232 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 233 {
 234         return dreq->bytes_left;
 235 }
 236 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 237
 238 /*
 239  * Collects and returns the final error value/byte-count.
 240  */
 241 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 242 {
 243         ssize_t result = -EIOCBQUEUED;
 244
 245         /* Async requests don't wait here */
 246         if (dreq->iocb)
 247                 goto out;
 248
 249         result = wait_for_completion_killable(&dreq->completion);
 250
 251         if (!result) {
 252                 result = dreq->count;
 253                 WARN_ON_ONCE(dreq->count < 0);
 254         }
 255         if (!result)
 256                 result = dreq->error;
 257
 258 out:
 259         return (ssize_t) result;
 260 }
 261
 262 /*
 263  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 264  * the iocb is still valid here if this is a synchronous request.
 265  */
 266 static void nfs_direct_complete(struct nfs_direct_req *dreq)
 267 {
 268         struct inode *inode = dreq->inode;
 269
 270         inode_dio_end(inode);
 271
 272         if (dreq->iocb) {
 273                 long res = (long) dreq->error;
 274                 if (dreq->count != 0) {
 275                         res = (long) dreq->count;
 276                         WARN_ON_ONCE(dreq->count < 0);
 277                 }
 278                 dreq->iocb->ki_complete(dreq->iocb, res, 0);
 279         }
 280
 281         complete(&dreq->completion);
 282
 283         nfs_direct_req_release(dreq);
 284 }
 285
 286 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 287 {
 288         unsigned long bytes = 0;
 289         struct nfs_direct_req *dreq = hdr->dreq;
 290
 291         spin_lock(&dreq->lock);
 292         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 293                 spin_unlock(&dreq->lock);
 294                 goto out_put;
 295         }
 296
 297         nfs_direct_count_bytes(dreq, hdr);
 298         spin_unlock(&dreq->lock);
 299
 300         while (!list_empty(&hdr->pages)) {
 301                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 302                 struct page *page = req->wb_page;
 303
 304                 if (!PageCompound(page) && bytes < hdr->good_bytes &&
 305                     (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
 306                         set_page_dirty(page);
 307                 bytes += req->wb_bytes;
 308                 nfs_list_remove_request(req);
 309                 nfs_release_request(req);
 310         }
 311 out_put:
 312         if (put_dreq(dreq))
 313                 nfs_direct_complete(dreq);
 314         hdr->release(hdr);
 315 }
 316
 317 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
 318 {
 319         struct nfs_page *req;
 320
 321         while (!list_empty(head)) {
 322                 req = nfs_list_entry(head->next);
 323                 nfs_list_remove_request(req);
 324                 nfs_release_request(req);
 325         }
 326 }
 327
 328 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 329 {
 330         get_dreq(hdr->dreq);
 331 }
 332
 333 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 334         .error_cleanup = nfs_read_sync_pgio_error,
 335         .init_hdr = nfs_direct_pgio_init,
 336         .completion = nfs_direct_read_completion,
 337 };
 338
 339 /*
 340  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 341  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 342  * bail and stop sending more reads.  Read length accounting is
 343  * handled automatically by nfs_direct_read_result().  Otherwise, if
 344  * no requests have been sent, just return an error.
 345  */
 346
 347 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 348                                               struct iov_iter *iter,
 349                                               loff_t pos)
 350 {
 351         struct nfs_pageio_descriptor desc;
 352         struct inode *inode = dreq->inode;
 353         ssize_t result = -EINVAL;
 354         size_t requested_bytes = 0;
 355         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 356
 357         nfs_pageio_init_read(&desc, dreq->inode, false,
 358                              &nfs_direct_read_completion_ops);
 359         get_dreq(dreq);
 360         desc.pg_dreq = dreq;
 361         inode_dio_begin(inode);
 362
 363         while (iov_iter_count(iter)) {
 364                 struct page **pagevec;
 365                 size_t bytes;
 366                 size_t pgbase;
 367                 unsigned npages, i;
 368
 369                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 370                                                   rsize, &pgbase);
 371                 if (result < 0)
 372                         break;
 373
 374                 bytes = result;
 375                 iov_iter_advance(iter, bytes);
 376                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 377                 for (i = 0; i < npages; i++) {
 378                         struct nfs_page *req;
 379                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 380                         /* XXX do we need to do the eof zeroing found in async_filler? */
 381                         req = nfs_create_request(dreq->ctx, pagevec[i],
 382                                                  pgbase, req_len);
 383                         if (IS_ERR(req)) {
 384                                 result = PTR_ERR(req);
 385                                 break;
 386                         }
 387                         req->wb_index = pos >> PAGE_SHIFT;
 388                         req->wb_offset = pos & ~PAGE_MASK;
 389                         if (!nfs_pageio_add_request(&desc, req)) {
 390                                 result = desc.pg_error;
 391                                 nfs_release_request(req);
 392                                 break;
 393                         }
 394                         pgbase = 0;
 395                         bytes -= req_len;
 396                         requested_bytes += req_len;
 397                         pos += req_len;
 398                         dreq->bytes_left -= req_len;
 399                 }
 400                 nfs_direct_release_pages(pagevec, npages);
 401                 kvfree(pagevec);
 402                 if (result < 0)
 403                         break;
 404         }
 405
 406         nfs_pageio_complete(&desc);
 407
 408         /*
 409          * If no bytes were started, return the error, and let the
 410          * generic layer handle the completion.
 411          */
 412         if (requested_bytes == 0) {
 413                 inode_dio_end(inode);
 414                 nfs_direct_req_release(dreq);
 415                 return result < 0 ? result : -EIO;
 416         }
 417
 418         if (put_dreq(dreq))
 419                 nfs_direct_complete(dreq);
 420         return requested_bytes;
 421 }
 422
 423 /**
 424  * nfs_file_direct_read - file direct read operation for NFS files
 425  * @iocb: target I/O control block
 426  * @iter: vector of user buffers into which to read data
 427  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 428  *
 429  * We use this function for direct reads instead of calling
 430  * generic_file_aio_read() in order to avoid gfar's check to see if
 431  * the request starts before the end of the file.  For that check
 432  * to work, we must generate a GETATTR before each direct read, and
 433  * even then there is a window between the GETATTR and the subsequent
 434  * READ where the file size could change.  Our preference is simply
 435  * to do all reads the application wants, and the server will take
 436  * care of managing the end of file boundary.
 437  *
 438  * This function also eliminates unnecessarily updating the file's
 439  * atime locally, as the NFS server sets the file's atime, and this
 440  * client must read the updated atime from the server back into its
 441  * cache.
 442  */
 443 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 444                              bool swap)
 445 {
 446         struct file *file = iocb->ki_filp;
 447         struct address_space *mapping = file->f_mapping;
 448         struct inode *inode = mapping->host;
 449         struct nfs_direct_req *dreq;
 450         struct nfs_lock_context *l_ctx;
 451         ssize_t result, requested;
 452         size_t count = iov_iter_count(iter);
 453         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 454
 455         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 456                 file, count, (long long) iocb->ki_pos);
 457
 458         result = 0;
 459         if (!count)
 460                 goto out;
 461
 462         task_io_account_read(count);
 463
 464         result = -ENOMEM;
 465         dreq = nfs_direct_req_alloc();
 466         if (dreq == NULL)
 467                 goto out;
 468
 469         dreq->inode = inode;
 470         dreq->bytes_left = dreq->max_count = count;
 471         dreq->io_start = iocb->ki_pos;
 472         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 473         l_ctx = nfs_get_lock_context(dreq->ctx);
 474         if (IS_ERR(l_ctx)) {
 475                 result = PTR_ERR(l_ctx);
 476                 nfs_direct_req_release(dreq);
 477                 goto out_release;
 478         }
 479         dreq->l_ctx = l_ctx;
 480         if (!is_sync_kiocb(iocb))
 481                 dreq->iocb = iocb;
 482
 483         if (iter_is_iovec(iter))
 484                 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
 485
 486         if (!swap)
 487                 nfs_start_io_direct(inode);
 488
 489         NFS_I(inode)->read_io += count;
 490         requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
 491
 492         if (!swap)
 493                 nfs_end_io_direct(inode);
 494
 495         if (requested > 0) {
 496                 result = nfs_direct_wait(dreq);
 497                 if (result > 0) {
 498                         requested -= result;
 499                         iocb->ki_pos += result;
 500                 }
 501                 iov_iter_revert(iter, requested);
 502         } else {
 503                 result = requested;
 504         }
 505
 506 out_release:
 507         nfs_direct_req_release(dreq);
 508 out:
 509         return result;
 510 }
 511
 512 static void
 513 nfs_direct_join_group(struct list_head *list, struct inode *inode)
 514 {
 515         struct nfs_page *req, *next;
 516
 517         list_for_each_entry(req, list, wb_list) {
 518                 if (req->wb_head != req || req->wb_this_page == req)
 519                         continue;
 520                 for (next = req->wb_this_page;
 521                                 next != req->wb_head;
 522                                 next = next->wb_this_page) {
 523                         nfs_list_remove_request(next);
 524                         nfs_release_request(next);
 525                 }
 526                 nfs_join_page_group(req, inode);
 527         }
 528 }
 529
 530 static void
 531 nfs_direct_write_scan_commit_list(struct inode *inode,
 532                                   struct list_head *list,
 533                                   struct nfs_commit_info *cinfo)
 534 {
 535         mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
 536         pnfs_recover_commit_reqs(list, cinfo);
 537         nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
 538         mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
 539 }
 540
 541 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 542 {
 543         struct nfs_pageio_descriptor desc;
 544         struct nfs_page *req, *tmp;
 545         LIST_HEAD(reqs);
 546         struct nfs_commit_info cinfo;
 547         LIST_HEAD(failed);
 548
 549         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 550         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 551
 552         nfs_direct_join_group(&reqs, dreq->inode);
 553
 554         dreq->count = 0;
 555         dreq->max_count = 0;
 556         list_for_each_entry(req, &reqs, wb_list)
 557                 dreq->max_count += req->wb_bytes;
 558         nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
 559         get_dreq(dreq);
 560
 561         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 562                               &nfs_direct_write_completion_ops);
 563         desc.pg_dreq = dreq;
 564
 565         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 566                 /* Bump the transmission count */
 567                 req->wb_nio++;
 568                 if (!nfs_pageio_add_request(&desc, req)) {
 569                         nfs_list_move_request(req, &failed);
 570                         spin_lock(&cinfo.inode->i_lock);
 571                         dreq->flags = 0;
 572                         if (desc.pg_error < 0)
 573                                 dreq->error = desc.pg_error;
 574                         else
 575                                 dreq->error = -EIO;
 576                         spin_unlock(&cinfo.inode->i_lock);
 577                 }
 578                 nfs_release_request(req);
 579         }
 580         nfs_pageio_complete(&desc);
 581
 582         while (!list_empty(&failed)) {
 583                 req = nfs_list_entry(failed.next);
 584                 nfs_list_remove_request(req);
 585                 nfs_unlock_and_release_request(req);
 586         }
 587
 588         if (put_dreq(dreq))
 589                 nfs_direct_write_complete(dreq);
 590 }
 591
 592 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 593 {
 594         const struct nfs_writeverf *verf = data->res.verf;
 595         struct nfs_direct_req *dreq = data->dreq;
 596         struct nfs_commit_info cinfo;
 597         struct nfs_page *req;
 598         int status = data->task.tk_status;
 599
 600         if (status < 0) {
 601                 /* Errors in commit are fatal */
 602                 dreq->error = status;
 603                 dreq->max_count = 0;
 604                 dreq->count = 0;
 605                 dreq->flags = NFS_ODIRECT_DONE;
 606         } else if (dreq->flags == NFS_ODIRECT_DONE)
 607                 status = dreq->error;
 608
 609         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 610
 611         while (!list_empty(&data->pages)) {
 612                 req = nfs_list_entry(data->pages.next);
 613                 nfs_list_remove_request(req);
 614                 if (status >= 0 && !nfs_write_match_verf(verf, req)) {
 615                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 616                         /*
 617                          * Despite the reboot, the write was successful,
 618                          * so reset wb_nio.
 619                          */
 620                         req->wb_nio = 0;
 621                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 622                 } else /* Error or match */
 623                         nfs_release_request(req);
 624                 nfs_unlock_and_release_request(req);
 625         }
 626
 627         if (nfs_commit_end(cinfo.mds))
 628                 nfs_direct_write_complete(dreq);
 629 }
 630
 631 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
 632                 struct nfs_page *req)
 633 {
 634         struct nfs_direct_req *dreq = cinfo->dreq;
 635
 636         spin_lock(&dreq->lock);
 637         if (dreq->flags != NFS_ODIRECT_DONE)
 638                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 639         spin_unlock(&dreq->lock);
 640         nfs_mark_request_commit(req, NULL, cinfo, 0);
 641 }
 642
 643 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 644         .completion = nfs_direct_commit_complete,
 645         .resched_write = nfs_direct_resched_write,
 646 };
 647
 648 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 649 {
 650         int res;
 651         struct nfs_commit_info cinfo;
 652         LIST_HEAD(mds_list);
 653
 654         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 655         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 656         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 657         if (res < 0) /* res == -ENOMEM */
 658                 nfs_direct_write_reschedule(dreq);
 659 }
 660
 661 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
 662 {
 663         struct nfs_commit_info cinfo;
 664         struct nfs_page *req;
 665         LIST_HEAD(reqs);
 666
 667         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 668         nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
 669
 670         while (!list_empty(&reqs)) {
 671                 req = nfs_list_entry(reqs.next);
 672                 nfs_list_remove_request(req);
 673                 nfs_release_request(req);
 674                 nfs_unlock_and_release_request(req);
 675         }
 676 }
 677
 678 static void nfs_direct_write_schedule_work(struct work_struct *work)
 679 {
 680         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 681         int flags = dreq->flags;
 682
 683         dreq->flags = 0;
 684         switch (flags) {
 685                 case NFS_ODIRECT_DO_COMMIT:
 686                         nfs_direct_commit_schedule(dreq);
 687                         break;
 688                 case NFS_ODIRECT_RESCHED_WRITES:
 689                         nfs_direct_write_reschedule(dreq);
 690                         break;
 691                 default:
 692                         nfs_direct_write_clear_reqs(dreq);
 693                         nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
 694                         nfs_direct_complete(dreq);
 695         }
 696 }
 697
 698 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
 699 {
 700         queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
 701 }
 702
 703 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 704 {
 705         struct nfs_direct_req *dreq = hdr->dreq;
 706         struct nfs_commit_info cinfo;
 707         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 708         int flags = NFS_ODIRECT_DONE;
 709
 710         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 711
 712         spin_lock(&dreq->lock);
 713         if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 714                 spin_unlock(&dreq->lock);
 715                 goto out_put;
 716         }
 717
 718         nfs_direct_count_bytes(dreq, hdr);
 719         if (hdr->good_bytes != 0 && nfs_write_need_commit(hdr)) {
 720                 if (!dreq->flags)
 721                         dreq->flags = NFS_ODIRECT_DO_COMMIT;
 722                 flags = dreq->flags;
 723         }
 724         spin_unlock(&dreq->lock);
 725
 726         while (!list_empty(&hdr->pages)) {
 727
 728                 req = nfs_list_entry(hdr->pages.next);
 729                 nfs_list_remove_request(req);
 730                 if (flags == NFS_ODIRECT_DO_COMMIT) {
 731                         kref_get(&req->wb_kref);
 732                         memcpy(&req->wb_verf, &hdr->verf.verifier,
 733                                sizeof(req->wb_verf));
 734                         nfs_mark_request_commit(req, hdr->lseg, &cinfo,
 735                                 hdr->ds_commit_idx);
 736                 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
 737                         kref_get(&req->wb_kref);
 738                         nfs_mark_request_commit(req, NULL, &cinfo, 0);
 739                 }
 740                 nfs_unlock_and_release_request(req);
 741         }
 742
 743 out_put:
 744         if (put_dreq(dreq))
 745                 nfs_direct_write_complete(dreq);
 746         hdr->release(hdr);
 747 }
 748
 749 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
 750 {
 751         struct nfs_page *req;
 752
 753         while (!list_empty(head)) {
 754                 req = nfs_list_entry(head->next);
 755                 nfs_list_remove_request(req);
 756                 nfs_unlock_and_release_request(req);
 757         }
 758 }
 759
 760 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
 761 {
 762         struct nfs_direct_req *dreq = hdr->dreq;
 763
 764         spin_lock(&dreq->lock);
 765         if (dreq->error == 0) {
 766                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 767                 /* fake unstable write to let common nfs resend pages */
 768                 hdr->verf.committed = NFS_UNSTABLE;
 769                 hdr->good_bytes = hdr->args.offset + hdr->args.count -
 770                         hdr->io_start;
 771         }
 772         spin_unlock(&dreq->lock);
 773 }
 774
 775 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 776         .error_cleanup = nfs_write_sync_pgio_error,
 777         .init_hdr = nfs_direct_pgio_init,
 778         .completion = nfs_direct_write_completion,
 779         .reschedule_io = nfs_direct_write_reschedule_io,
 780 };
 781
 782
 783 /*
 784  * NB: Return the value of the first error return code.  Subsequent
 785  *     errors after the first one are ignored.
 786  */
 787 /*
 788  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 789  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 790  * bail and stop sending more writes.  Write length accounting is
 791  * handled automatically by nfs_direct_write_result().  Otherwise, if
 792  * no requests have been sent, just return an error.
 793  */
 794 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 795                                                struct iov_iter *iter,
 796                                                loff_t pos, int ioflags)
 797 {
 798         struct nfs_pageio_descriptor desc;
 799         struct inode *inode = dreq->inode;
 800         ssize_t result = 0;
 801         size_t requested_bytes = 0;
 802         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 803
 804         nfs_pageio_init_write(&desc, inode, ioflags, false,
 805                               &nfs_direct_write_completion_ops);
 806         desc.pg_dreq = dreq;
 807         get_dreq(dreq);
 808         inode_dio_begin(inode);
 809
 810         NFS_I(inode)->write_io += iov_iter_count(iter);
 811         while (iov_iter_count(iter)) {
 812                 struct page **pagevec;
 813                 size_t bytes;
 814                 size_t pgbase;
 815                 unsigned npages, i;
 816
 817                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 818                                                   wsize, &pgbase);
 819                 if (result < 0)
 820                         break;
 821
 822                 bytes = result;
 823                 iov_iter_advance(iter, bytes);
 824                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 825                 for (i = 0; i < npages; i++) {
 826                         struct nfs_page *req;
 827                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 828
 829                         req = nfs_create_request(dreq->ctx, pagevec[i],
 830                                                  pgbase, req_len);
 831                         if (IS_ERR(req)) {
 832                                 result = PTR_ERR(req);
 833                                 break;
 834                         }
 835
 836                         if (desc.pg_error < 0) {
 837                                 nfs_free_request(req);
 838                                 result = desc.pg_error;
 839                                 break;
 840                         }
 841
 842                         nfs_lock_request(req);
 843                         req->wb_index = pos >> PAGE_SHIFT;
 844                         req->wb_offset = pos & ~PAGE_MASK;
 845                         if (!nfs_pageio_add_request(&desc, req)) {
 846                                 result = desc.pg_error;
 847                                 nfs_unlock_and_release_request(req);
 848                                 break;
 849                         }
 850                         pgbase = 0;
 851                         bytes -= req_len;
 852                         requested_bytes += req_len;
 853                         pos += req_len;
 854                         dreq->bytes_left -= req_len;
 855                 }
 856                 nfs_direct_release_pages(pagevec, npages);
 857                 kvfree(pagevec);
 858                 if (result < 0)
 859                         break;
 860         }
 861         nfs_pageio_complete(&desc);
 862
 863         /*
 864          * If no bytes were started, return the error, and let the
 865          * generic layer handle the completion.
 866          */
 867         if (requested_bytes == 0) {
 868                 inode_dio_end(inode);
 869                 nfs_direct_req_release(dreq);
 870                 return result < 0 ? result : -EIO;
 871         }
 872
 873         if (put_dreq(dreq))
 874                 nfs_direct_write_complete(dreq);
 875         return requested_bytes;
 876 }
 877
 878 /**
 879  * nfs_file_direct_write - file direct write operation for NFS files
 880  * @iocb: target I/O control block
 881  * @iter: vector of user buffers from which to write data
 882  * @swap: flag indicating this is swap IO, not O_DIRECT IO
 883  *
 884  * We use this function for direct writes instead of calling
 885  * generic_file_aio_write() in order to avoid taking the inode
 886  * semaphore and updating the i_size.  The NFS server will set
 887  * the new i_size and this client must read the updated size
 888  * back into its cache.  We let the server do generic write
 889  * parameter checking and report problems.
 890  *
 891  * We eliminate local atime updates, see direct read above.
 892  *
 893  * We avoid unnecessary page cache invalidations for normal cached
 894  * readers of this file.
 895  *
 896  * Note that O_APPEND is not supported for NFS direct writes, as there
 897  * is no atomic O_APPEND write facility in the NFS protocol.
 898  */
 899 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 900                               bool swap)
 901 {
 902         ssize_t result, requested;
 903         size_t count;
 904         struct file *file = iocb->ki_filp;
 905         struct address_space *mapping = file->f_mapping;
 906         struct inode *inode = mapping->host;
 907         struct nfs_direct_req *dreq;
 908         struct nfs_lock_context *l_ctx;
 909         loff_t pos, end;
 910
 911         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 912                 file, iov_iter_count(iter), (long long) iocb->ki_pos);
 913
 914         if (swap)
 915                 /* bypass generic checks */
 916                 result =  iov_iter_count(iter);
 917         else
 918                 result = generic_write_checks(iocb, iter);
 919         if (result <= 0)
 920                 return result;
 921         count = result;
 922         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 923
 924         pos = iocb->ki_pos;
 925         end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
 926
 927         task_io_account_write(count);
 928
 929         result = -ENOMEM;
 930         dreq = nfs_direct_req_alloc();
 931         if (!dreq)
 932                 goto out;
 933
 934         dreq->inode = inode;
 935         dreq->bytes_left = dreq->max_count = count;
 936         dreq->io_start = pos;
 937         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 938         l_ctx = nfs_get_lock_context(dreq->ctx);
 939         if (IS_ERR(l_ctx)) {
 940                 result = PTR_ERR(l_ctx);
 941                 nfs_direct_req_release(dreq);
 942                 goto out_release;
 943         }
 944         dreq->l_ctx = l_ctx;
 945         if (!is_sync_kiocb(iocb))
 946                 dreq->iocb = iocb;
 947         pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
 948
 949         if (swap) {
 950                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
 951                                                             FLUSH_STABLE);
 952         } else {
 953                 nfs_start_io_direct(inode);
 954
 955                 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
 956                                                             FLUSH_COND_STABLE);
 957
 958                 if (mapping->nrpages) {
 959                         invalidate_inode_pages2_range(mapping,
 960                                                       pos >> PAGE_SHIFT, end);
 961                 }
 962
 963                 nfs_end_io_direct(inode);
 964         }
 965
 966         if (requested > 0) {
 967                 result = nfs_direct_wait(dreq);
 968                 if (result > 0) {
 969                         requested -= result;
 970                         iocb->ki_pos = pos + result;
 971                         /* XXX: should check the generic_write_sync retval */
 972                         generic_write_sync(iocb, result);
 973                 }
 974                 iov_iter_revert(iter, requested);
 975         } else {
 976                 result = requested;
 977         }
 978 out_release:
 979         nfs_direct_req_release(dreq);
 980 out:
 981         return result;
 982 }
 983
 984 /**
 985  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 986  *
 987  */
 988 int __init nfs_init_directcache(void)
 989 {
 990         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 991                                                 sizeof(struct nfs_direct_req),
 992                                                 0, (SLAB_RECLAIM_ACCOUNT|
 993                                                         SLAB_MEM_SPREAD),
 994                                                 NULL);
 995         if (nfs_direct_cachep == NULL)
 996                 return -ENOMEM;
 997
 998         return 0;
 999 }
1000
1001 /**
1002  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1003  *
1004  */
1005 void nfs_destroy_directcache(void)
1006 {
1007         kmem_cache_destroy(nfs_direct_cachep);
1008 }