fs/nfs/direct.c

   1 /*
   2  * linux/fs/nfs/direct.c
   3  *
   4  * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
   5  *
   6  * High-performance uncached I/O for the Linux NFS client
   7  *
   8  * There are important applications whose performance or correctness
   9  * depends on uncached access to file data.  Database clusters
  10  * (multiple copies of the same instance running on separate hosts)
  11  * implement their own cache coherency protocol that subsumes file
  12  * system cache protocols.  Applications that process datasets
  13  * considerably larger than the client's memory do not always benefit
  14  * from a local cache.  A streaming video server, for instance, has no
  15  * need to cache the contents of a file.
  16  *
  17  * When an application requests uncached I/O, all read and write requests
  18  * are made directly to the server; data stored or fetched via these
  19  * requests is not cached in the Linux page cache.  The client does not
  20  * correct unaligned requests from applications.  All requested bytes are
  21  * held on permanent storage before a direct write system call returns to
  22  * an application.
  23  *
  24  * Solaris implements an uncached I/O facility called directio() that
  25  * is used for backups and sequential I/O to very large files.  Solaris
  26  * also supports uncaching whole NFS partitions with "-o forcedirectio,"
  27  * an undocumented mount option.
  28  *
  29  * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
  30  * help from Andrew Morton.
  31  *
  32  * 18 Dec 2001  Initial implementation for 2.4  --cel
  33  * 08 Jul 2002  Version for 2.4.19, with bug fixes --trondmy
  34  * 08 Jun 2003  Port to 2.5 APIs  --cel
  35  * 31 Mar 2004  Handle direct I/O without VFS support  --cel
  36  * 15 Sep 2004  Parallel async reads  --cel
  37  * 04 May 2005  support O_DIRECT with aio  --cel
  38  *
  39  */
  40
  41 #include <linux/errno.h>
  42 #include <linux/sched.h>
  43 #include <linux/kernel.h>
  44 #include <linux/file.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/kref.h>
  47 #include <linux/slab.h>
  48 #include <linux/task_io_accounting_ops.h>
  49 #include <linux/module.h>
  50
  51 #include <linux/nfs_fs.h>
  52 #include <linux/nfs_page.h>
  53 #include <linux/sunrpc/clnt.h>
  54
  55 #include <asm/uaccess.h>
  56 #include <linux/atomic.h>
  57
  58 #include "internal.h"
  59 #include "iostat.h"
  60 #include "pnfs.h"
  61
  62 #define NFSDBG_FACILITY         NFSDBG_VFS
  63
  64 static struct kmem_cache *nfs_direct_cachep;
  65
  66 /*
  67  * This represents a set of asynchronous requests that we're waiting on
  68  */
  69 struct nfs_direct_req {
  70         struct kref             kref;           /* release manager */
  71
  72         /* I/O parameters */
  73         struct nfs_open_context *ctx;           /* file open context info */
  74         struct nfs_lock_context *l_ctx;         /* Lock context info */
  75         struct kiocb *          iocb;           /* controlling i/o request */
  76         struct inode *          inode;          /* target file of i/o */
  77
  78         /* completion state */
  79         atomic_t                io_count;       /* i/os we're waiting for */
  80         spinlock_t              lock;           /* protect completion state */
  81         ssize_t                 count,          /* bytes actually processed */
  82                                 bytes_left,     /* bytes left to be sent */
  83                                 error;          /* any reported error */
  84         struct completion       completion;     /* wait for i/o completion */
  85
  86         /* commit state */
  87         struct nfs_mds_commit_info mds_cinfo;   /* Storage for cinfo */
  88         struct pnfs_ds_commit_info ds_cinfo;    /* Storage for cinfo */
  89         struct work_struct      work;
  90         int                     flags;
  91 #define NFS_ODIRECT_DO_COMMIT           (1)     /* an unstable reply was received */
  92 #define NFS_ODIRECT_RESCHED_WRITES      (2)     /* write verification failed */
  93         struct nfs_writeverf    verf;           /* unstable write verifier */
  94 };
  95
  96 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
  97 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
  98 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode);
  99 static void nfs_direct_write_schedule_work(struct work_struct *work);
 100
 101 static inline void get_dreq(struct nfs_direct_req *dreq)
 102 {
 103         atomic_inc(&dreq->io_count);
 104 }
 105
 106 static inline int put_dreq(struct nfs_direct_req *dreq)
 107 {
 108         return atomic_dec_and_test(&dreq->io_count);
 109 }
 110
 111 /**
 112  * nfs_direct_IO - NFS address space operation for direct I/O
 113  * @rw: direction (read or write)
 114  * @iocb: target I/O control block
 115  * @iov: array of vectors that define I/O buffer
 116  * @pos: offset in file to begin the operation
 117  * @nr_segs: size of iovec array
 118  *
 119  * The presence of this routine in the address space ops vector means
 120  * the NFS client supports direct I/O. However, for most direct IO, we
 121  * shunt off direct read and write requests before the VFS gets them,
 122  * so this method is only ever called for swap.
 123  */
 124 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter, loff_t pos)
 125 {
 126 #ifndef CONFIG_NFS_SWAP
 127         dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
 128                         iocb->ki_filp, (long long) pos, iter->nr_segs);
 129
 130         return -EINVAL;
 131 #else
 132         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 133
 134         if (rw == READ || rw == KERNEL_READ)
 135                 return nfs_file_direct_read(iocb, iter, pos,
 136                                 rw == READ ? true : false);
 137         return nfs_file_direct_write(iocb, iter, pos,
 138                                 rw == WRITE ? true : false);
 139 #endif /* CONFIG_NFS_SWAP */
 140 }
 141
 142 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 143 {
 144         unsigned int i;
 145         for (i = 0; i < npages; i++)
 146                 page_cache_release(pages[i]);
 147 }
 148
 149 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 150                               struct nfs_direct_req *dreq)
 151 {
 152         cinfo->lock = &dreq->lock;
 153         cinfo->mds = &dreq->mds_cinfo;
 154         cinfo->ds = &dreq->ds_cinfo;
 155         cinfo->dreq = dreq;
 156         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 157 }
 158
 159 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 160 {
 161         struct nfs_direct_req *dreq;
 162
 163         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 164         if (!dreq)
 165                 return NULL;
 166
 167         kref_init(&dreq->kref);
 168         kref_get(&dreq->kref);
 169         init_completion(&dreq->completion);
 170         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 171         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 172         spin_lock_init(&dreq->lock);
 173
 174         return dreq;
 175 }
 176
 177 static void nfs_direct_req_free(struct kref *kref)
 178 {
 179         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 180
 181         if (dreq->l_ctx != NULL)
 182                 nfs_put_lock_context(dreq->l_ctx);
 183         if (dreq->ctx != NULL)
 184                 put_nfs_open_context(dreq->ctx);
 185         kmem_cache_free(nfs_direct_cachep, dreq);
 186 }
 187
 188 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 189 {
 190         kref_put(&dreq->kref, nfs_direct_req_free);
 191 }
 192
 193 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 194 {
 195         return dreq->bytes_left;
 196 }
 197 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 198
 199 /*
 200  * Collects and returns the final error value/byte-count.
 201  */
 202 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 203 {
 204         ssize_t result = -EIOCBQUEUED;
 205
 206         /* Async requests don't wait here */
 207         if (dreq->iocb)
 208                 goto out;
 209
 210         result = wait_for_completion_killable(&dreq->completion);
 211
 212         if (!result)
 213                 result = dreq->error;
 214         if (!result)
 215                 result = dreq->count;
 216
 217 out:
 218         return (ssize_t) result;
 219 }
 220
 221 /*
 222  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 223  * the iocb is still valid here if this is a synchronous request.
 224  */
 225 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 226 {
 227         struct inode *inode = dreq->inode;
 228
 229         if (dreq->iocb && write) {
 230                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 231
 232                 spin_lock(&inode->i_lock);
 233                 if (i_size_read(inode) < pos)
 234                         i_size_write(inode, pos);
 235                 spin_unlock(&inode->i_lock);
 236         }
 237
 238         if (write)
 239                 nfs_zap_mapping(inode, inode->i_mapping);
 240
 241         inode_dio_done(inode);
 242
 243         if (dreq->iocb) {
 244                 long res = (long) dreq->error;
 245                 if (!res)
 246                         res = (long) dreq->count;
 247                 aio_complete(dreq->iocb, res, 0);
 248         }
 249
 250         complete_all(&dreq->completion);
 251
 252         nfs_direct_req_release(dreq);
 253 }
 254
 255 static void nfs_direct_readpage_release(struct nfs_page *req)
 256 {
 257         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 258                 req->wb_context->dentry->d_inode->i_sb->s_id,
 259                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 260                 req->wb_bytes,
 261                 (long long)req_offset(req));
 262         nfs_release_request(req);
 263 }
 264
 265 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 266 {
 267         unsigned long bytes = 0;
 268         struct nfs_direct_req *dreq = hdr->dreq;
 269
 270         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 271                 goto out_put;
 272
 273         spin_lock(&dreq->lock);
 274         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 275                 dreq->error = hdr->error;
 276         else
 277                 dreq->count += hdr->good_bytes;
 278         spin_unlock(&dreq->lock);
 279
 280         while (!list_empty(&hdr->pages)) {
 281                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 282                 struct page *page = req->wb_page;
 283
 284                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 285                         set_page_dirty(page);
 286                 bytes += req->wb_bytes;
 287                 nfs_list_remove_request(req);
 288                 nfs_direct_readpage_release(req);
 289         }
 290 out_put:
 291         if (put_dreq(dreq))
 292                 nfs_direct_complete(dreq, false);
 293         hdr->release(hdr);
 294 }
 295
 296 static void nfs_read_sync_pgio_error(struct list_head *head)
 297 {
 298         struct nfs_page *req;
 299
 300         while (!list_empty(head)) {
 301                 req = nfs_list_entry(head->next);
 302                 nfs_list_remove_request(req);
 303                 nfs_release_request(req);
 304         }
 305 }
 306
 307 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 308 {
 309         get_dreq(hdr->dreq);
 310 }
 311
 312 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 313         .error_cleanup = nfs_read_sync_pgio_error,
 314         .init_hdr = nfs_direct_pgio_init,
 315         .completion = nfs_direct_read_completion,
 316 };
 317
 318 /*
 319  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 320  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 321  * bail and stop sending more reads.  Read length accounting is
 322  * handled automatically by nfs_direct_read_result().  Otherwise, if
 323  * no requests have been sent, just return an error.
 324  */
 325
 326 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 327                                               struct iov_iter *iter,
 328                                               loff_t pos)
 329 {
 330         struct nfs_pageio_descriptor desc;
 331         struct inode *inode = dreq->inode;
 332         ssize_t result = -EINVAL;
 333         size_t requested_bytes = 0;
 334         size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
 335
 336         NFS_PROTO(dreq->inode)->read_pageio_init(&desc, dreq->inode,
 337                              &nfs_direct_read_completion_ops);
 338         get_dreq(dreq);
 339         desc.pg_dreq = dreq;
 340         atomic_inc(&inode->i_dio_count);
 341
 342         while (iov_iter_count(iter)) {
 343                 struct page **pagevec;
 344                 size_t bytes;
 345                 size_t pgbase;
 346                 unsigned npages, i;
 347
 348                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 349                                                   rsize, &pgbase);
 350                 if (result < 0)
 351                         break;
 352
 353                 bytes = result;
 354                 iov_iter_advance(iter, bytes);
 355                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 356                 for (i = 0; i < npages; i++) {
 357                         struct nfs_page *req;
 358                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 359                         /* XXX do we need to do the eof zeroing found in async_filler? */
 360                         req = nfs_create_request(dreq->ctx, dreq->inode,
 361                                                  pagevec[i],
 362                                                  pgbase, req_len);
 363                         if (IS_ERR(req)) {
 364                                 result = PTR_ERR(req);
 365                                 break;
 366                         }
 367                         req->wb_index = pos >> PAGE_SHIFT;
 368                         req->wb_offset = pos & ~PAGE_MASK;
 369                         if (!nfs_pageio_add_request(&desc, req)) {
 370                                 result = desc.pg_error;
 371                                 nfs_release_request(req);
 372                                 break;
 373                         }
 374                         pgbase = 0;
 375                         bytes -= req_len;
 376                         requested_bytes += req_len;
 377                         pos += req_len;
 378                         dreq->bytes_left -= req_len;
 379                 }
 380                 nfs_direct_release_pages(pagevec, npages);
 381                 kvfree(pagevec);
 382                 if (result < 0)
 383                         break;
 384         }
 385
 386         nfs_pageio_complete(&desc);
 387
 388         /*
 389          * If no bytes were started, return the error, and let the
 390          * generic layer handle the completion.
 391          */
 392         if (requested_bytes == 0) {
 393                 inode_dio_done(inode);
 394                 nfs_direct_req_release(dreq);
 395                 return result < 0 ? result : -EIO;
 396         }
 397
 398         if (put_dreq(dreq))
 399                 nfs_direct_complete(dreq, false);
 400         return 0;
 401 }
 402
 403 /**
 404  * nfs_file_direct_read - file direct read operation for NFS files
 405  * @iocb: target I/O control block
 406  * @iter: vector of user buffers into which to read data
 407  * @pos: byte offset in file where reading starts
 408  *
 409  * We use this function for direct reads instead of calling
 410  * generic_file_aio_read() in order to avoid gfar's check to see if
 411  * the request starts before the end of the file.  For that check
 412  * to work, we must generate a GETATTR before each direct read, and
 413  * even then there is a window between the GETATTR and the subsequent
 414  * READ where the file size could change.  Our preference is simply
 415  * to do all reads the application wants, and the server will take
 416  * care of managing the end of file boundary.
 417  *
 418  * This function also eliminates unnecessarily updating the file's
 419  * atime locally, as the NFS server sets the file's atime, and this
 420  * client must read the updated atime from the server back into its
 421  * cache.
 422  */
 423 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
 424                                 loff_t pos, bool uio)
 425 {
 426         struct file *file = iocb->ki_filp;
 427         struct address_space *mapping = file->f_mapping;
 428         struct inode *inode = mapping->host;
 429         struct nfs_direct_req *dreq;
 430         struct nfs_lock_context *l_ctx;
 431         ssize_t result = -EINVAL;
 432         size_t count = iov_iter_count(iter);
 433         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 434
 435         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 436                 file, count, (long long) pos);
 437
 438         result = 0;
 439         if (!count)
 440                 goto out;
 441
 442         mutex_lock(&inode->i_mutex);
 443         result = nfs_sync_mapping(mapping);
 444         if (result)
 445                 goto out_unlock;
 446
 447         task_io_account_read(count);
 448
 449         result = -ENOMEM;
 450         dreq = nfs_direct_req_alloc();
 451         if (dreq == NULL)
 452                 goto out_unlock;
 453
 454         dreq->inode = inode;
 455         dreq->bytes_left = count;
 456         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 457         l_ctx = nfs_get_lock_context(dreq->ctx);
 458         if (IS_ERR(l_ctx)) {
 459                 result = PTR_ERR(l_ctx);
 460                 goto out_release;
 461         }
 462         dreq->l_ctx = l_ctx;
 463         if (!is_sync_kiocb(iocb))
 464                 dreq->iocb = iocb;
 465
 466         NFS_I(inode)->read_io += count;
 467         result = nfs_direct_read_schedule_iovec(dreq, iter, pos);
 468
 469         mutex_unlock(&inode->i_mutex);
 470
 471         if (!result) {
 472                 result = nfs_direct_wait(dreq);
 473                 if (result > 0)
 474                         iocb->ki_pos = pos + result;
 475         }
 476
 477         nfs_direct_req_release(dreq);
 478         return result;
 479
 480 out_release:
 481         nfs_direct_req_release(dreq);
 482 out_unlock:
 483         mutex_unlock(&inode->i_mutex);
 484 out:
 485         return result;
 486 }
 487
 488 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 489 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 490 {
 491         struct nfs_pageio_descriptor desc;
 492         struct nfs_page *req, *tmp;
 493         LIST_HEAD(reqs);
 494         struct nfs_commit_info cinfo;
 495         LIST_HEAD(failed);
 496
 497         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 498         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 499         spin_lock(cinfo.lock);
 500         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 501         spin_unlock(cinfo.lock);
 502
 503         dreq->count = 0;
 504         get_dreq(dreq);
 505
 506         NFS_PROTO(dreq->inode)->write_pageio_init(&desc, dreq->inode, FLUSH_STABLE,
 507                               &nfs_direct_write_completion_ops);
 508         desc.pg_dreq = dreq;
 509
 510         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 511                 if (!nfs_pageio_add_request(&desc, req)) {
 512                         nfs_list_remove_request(req);
 513                         nfs_list_add_request(req, &failed);
 514                         spin_lock(cinfo.lock);
 515                         dreq->flags = 0;
 516                         dreq->error = -EIO;
 517                         spin_unlock(cinfo.lock);
 518                 }
 519                 nfs_release_request(req);
 520         }
 521         nfs_pageio_complete(&desc);
 522
 523         while (!list_empty(&failed)) {
 524                 req = nfs_list_entry(failed.next);
 525                 nfs_list_remove_request(req);
 526                 nfs_unlock_and_release_request(req);
 527         }
 528
 529         if (put_dreq(dreq))
 530                 nfs_direct_write_complete(dreq, dreq->inode);
 531 }
 532
 533 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 534 {
 535         struct nfs_direct_req *dreq = data->dreq;
 536         struct nfs_commit_info cinfo;
 537         struct nfs_page *req;
 538         int status = data->task.tk_status;
 539
 540         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 541         if (status < 0) {
 542                 dprintk("NFS: %5u commit failed with error %d.\n",
 543                         data->task.tk_pid, status);
 544                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 545         } else if (memcmp(&dreq->verf, &data->verf, sizeof(data->verf))) {
 546                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 547                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 548         }
 549
 550         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 551         while (!list_empty(&data->pages)) {
 552                 req = nfs_list_entry(data->pages.next);
 553                 nfs_list_remove_request(req);
 554                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 555                         /* Note the rewrite will go through mds */
 556                         nfs_mark_request_commit(req, NULL, &cinfo);
 557                 } else
 558                         nfs_release_request(req);
 559                 nfs_unlock_and_release_request(req);
 560         }
 561
 562         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 563                 nfs_direct_write_complete(dreq, data->inode);
 564 }
 565
 566 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 567 {
 568         /* There is no lock to clear */
 569 }
 570
 571 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 572         .completion = nfs_direct_commit_complete,
 573         .error_cleanup = nfs_direct_error_cleanup,
 574 };
 575
 576 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 577 {
 578         int res;
 579         struct nfs_commit_info cinfo;
 580         LIST_HEAD(mds_list);
 581
 582         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 583         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 584         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 585         if (res < 0) /* res == -ENOMEM */
 586                 nfs_direct_write_reschedule(dreq);
 587 }
 588
 589 static void nfs_direct_write_schedule_work(struct work_struct *work)
 590 {
 591         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 592         int flags = dreq->flags;
 593
 594         dreq->flags = 0;
 595         switch (flags) {
 596                 case NFS_ODIRECT_DO_COMMIT:
 597                         nfs_direct_commit_schedule(dreq);
 598                         break;
 599                 case NFS_ODIRECT_RESCHED_WRITES:
 600                         nfs_direct_write_reschedule(dreq);
 601                         break;
 602                 default:
 603                         nfs_direct_complete(dreq, true);
 604         }
 605 }
 606
 607 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 608 {
 609         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 610 }
 611
 612 #else
 613 static void nfs_direct_write_schedule_work(struct work_struct *work)
 614 {
 615 }
 616
 617 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 618 {
 619         nfs_direct_complete(dreq, true);
 620 }
 621 #endif
 622
 623 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 624 {
 625         struct nfs_direct_req *dreq = hdr->dreq;
 626         struct nfs_commit_info cinfo;
 627         int bit = -1;
 628         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 629
 630         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 631                 goto out_put;
 632
 633         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 634
 635         spin_lock(&dreq->lock);
 636
 637         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 638                 dreq->flags = 0;
 639                 dreq->error = hdr->error;
 640         }
 641         if (dreq->error != 0)
 642                 bit = NFS_IOHDR_ERROR;
 643         else {
 644                 dreq->count += hdr->good_bytes;
 645                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 646                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 647                         bit = NFS_IOHDR_NEED_RESCHED;
 648                 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 649                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 650                                 bit = NFS_IOHDR_NEED_RESCHED;
 651                         else if (dreq->flags == 0) {
 652                                 memcpy(&dreq->verf, hdr->verf,
 653                                        sizeof(dreq->verf));
 654                                 bit = NFS_IOHDR_NEED_COMMIT;
 655                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 656                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 657                                 if (memcmp(&dreq->verf, hdr->verf, sizeof(dreq->verf))) {
 658                                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 659                                         bit = NFS_IOHDR_NEED_RESCHED;
 660                                 } else
 661                                         bit = NFS_IOHDR_NEED_COMMIT;
 662                         }
 663                 }
 664         }
 665         spin_unlock(&dreq->lock);
 666
 667         while (!list_empty(&hdr->pages)) {
 668                 req = nfs_list_entry(hdr->pages.next);
 669                 nfs_list_remove_request(req);
 670                 switch (bit) {
 671                 case NFS_IOHDR_NEED_RESCHED:
 672                 case NFS_IOHDR_NEED_COMMIT:
 673                         kref_get(&req->wb_kref);
 674                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 675                 }
 676                 nfs_unlock_and_release_request(req);
 677         }
 678
 679 out_put:
 680         if (put_dreq(dreq))
 681                 nfs_direct_write_complete(dreq, hdr->inode);
 682         hdr->release(hdr);
 683 }
 684
 685 static void nfs_write_sync_pgio_error(struct list_head *head)
 686 {
 687         struct nfs_page *req;
 688
 689         while (!list_empty(head)) {
 690                 req = nfs_list_entry(head->next);
 691                 nfs_list_remove_request(req);
 692                 nfs_unlock_and_release_request(req);
 693         }
 694 }
 695
 696 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 697         .error_cleanup = nfs_write_sync_pgio_error,
 698         .init_hdr = nfs_direct_pgio_init,
 699         .completion = nfs_direct_write_completion,
 700 };
 701
 702
 703 /*
 704  * NB: Return the value of the first error return code.  Subsequent
 705  *     errors after the first one are ignored.
 706  */
 707 /*
 708  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 709  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 710  * bail and stop sending more writes.  Write length accounting is
 711  * handled automatically by nfs_direct_write_result().  Otherwise, if
 712  * no requests have been sent, just return an error.
 713  */
 714 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 715                                                struct iov_iter *iter,
 716                                                loff_t pos)
 717 {
 718         struct nfs_pageio_descriptor desc;
 719         struct inode *inode = dreq->inode;
 720         ssize_t result = 0;
 721         size_t requested_bytes = 0;
 722         size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
 723
 724         NFS_PROTO(inode)->write_pageio_init(&desc, inode, FLUSH_COND_STABLE,
 725                               &nfs_direct_write_completion_ops);
 726         desc.pg_dreq = dreq;
 727         get_dreq(dreq);
 728         atomic_inc(&inode->i_dio_count);
 729
 730         NFS_I(inode)->write_io += iov_iter_count(iter);
 731         while (iov_iter_count(iter)) {
 732                 struct page **pagevec;
 733                 size_t bytes;
 734                 size_t pgbase;
 735                 unsigned npages, i;
 736
 737                 result = iov_iter_get_pages_alloc(iter, &pagevec,
 738                                                   wsize, &pgbase);
 739                 if (result < 0)
 740                         break;
 741
 742                 bytes = result;
 743                 iov_iter_advance(iter, bytes);
 744                 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
 745                 for (i = 0; i < npages; i++) {
 746                         struct nfs_page *req;
 747                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 748
 749                         req = nfs_create_request(dreq->ctx, inode,
 750                                                  pagevec[i],
 751                                                  pgbase, req_len);
 752                         if (IS_ERR(req)) {
 753                                 result = PTR_ERR(req);
 754                                 break;
 755                         }
 756                         nfs_lock_request(req);
 757                         req->wb_index = pos >> PAGE_SHIFT;
 758                         req->wb_offset = pos & ~PAGE_MASK;
 759                         if (!nfs_pageio_add_request(&desc, req)) {
 760                                 result = desc.pg_error;
 761                                 nfs_unlock_and_release_request(req);
 762                                 break;
 763                         }
 764                         pgbase = 0;
 765                         bytes -= req_len;
 766                         requested_bytes += req_len;
 767                         pos += req_len;
 768                         dreq->bytes_left -= req_len;
 769                 }
 770                 nfs_direct_release_pages(pagevec, npages);
 771                 kvfree(pagevec);
 772                 if (result < 0)
 773                         break;
 774         }
 775         nfs_pageio_complete(&desc);
 776
 777         /*
 778          * If no bytes were started, return the error, and let the
 779          * generic layer handle the completion.
 780          */
 781         if (requested_bytes == 0) {
 782                 inode_dio_done(inode);
 783                 nfs_direct_req_release(dreq);
 784                 return result < 0 ? result : -EIO;
 785         }
 786
 787         if (put_dreq(dreq))
 788                 nfs_direct_write_complete(dreq, dreq->inode);
 789         return 0;
 790 }
 791
 792 /**
 793  * nfs_file_direct_write - file direct write operation for NFS files
 794  * @iocb: target I/O control block
 795  * @iter: vector of user buffers from which to write data
 796  * @pos: byte offset in file where writing starts
 797  *
 798  * We use this function for direct writes instead of calling
 799  * generic_file_aio_write() in order to avoid taking the inode
 800  * semaphore and updating the i_size.  The NFS server will set
 801  * the new i_size and this client must read the updated size
 802  * back into its cache.  We let the server do generic write
 803  * parameter checking and report problems.
 804  *
 805  * We eliminate local atime updates, see direct read above.
 806  *
 807  * We avoid unnecessary page cache invalidations for normal cached
 808  * readers of this file.
 809  *
 810  * Note that O_APPEND is not supported for NFS direct writes, as there
 811  * is no atomic O_APPEND write facility in the NFS protocol.
 812  */
 813 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
 814                                 loff_t pos, bool uio)
 815 {
 816         ssize_t result = -EINVAL;
 817         struct file *file = iocb->ki_filp;
 818         struct address_space *mapping = file->f_mapping;
 819         struct inode *inode = mapping->host;
 820         struct nfs_direct_req *dreq;
 821         struct nfs_lock_context *l_ctx;
 822         loff_t end;
 823         size_t count = iov_iter_count(iter);
 824         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 825
 826         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
 827
 828         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
 829                 file, count, (long long) pos);
 830
 831         result = generic_write_checks(file, &pos, &count, 0);
 832         if (result)
 833                 goto out;
 834
 835         result = -EINVAL;
 836         if ((ssize_t) count < 0)
 837                 goto out;
 838         result = 0;
 839         if (!count)
 840                 goto out;
 841
 842         mutex_lock(&inode->i_mutex);
 843
 844         result = nfs_sync_mapping(mapping);
 845         if (result)
 846                 goto out_unlock;
 847
 848         if (mapping->nrpages) {
 849                 result = invalidate_inode_pages2_range(mapping,
 850                                         pos >> PAGE_CACHE_SHIFT, end);
 851                 if (result)
 852                         goto out_unlock;
 853         }
 854
 855         task_io_account_write(count);
 856
 857         result = -ENOMEM;
 858         dreq = nfs_direct_req_alloc();
 859         if (!dreq)
 860                 goto out_unlock;
 861
 862         dreq->inode = inode;
 863         dreq->bytes_left = count;
 864         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 865         l_ctx = nfs_get_lock_context(dreq->ctx);
 866         if (IS_ERR(l_ctx)) {
 867                 result = PTR_ERR(l_ctx);
 868                 goto out_release;
 869         }
 870         dreq->l_ctx = l_ctx;
 871         if (!is_sync_kiocb(iocb))
 872                 dreq->iocb = iocb;
 873
 874         result = nfs_direct_write_schedule_iovec(dreq, iter, pos);
 875
 876         if (mapping->nrpages) {
 877                 invalidate_inode_pages2_range(mapping,
 878                                               pos >> PAGE_CACHE_SHIFT, end);
 879         }
 880
 881         mutex_unlock(&inode->i_mutex);
 882
 883         if (!result) {
 884                 result = nfs_direct_wait(dreq);
 885                 if (result > 0) {
 886                         struct inode *inode = mapping->host;
 887
 888                         iocb->ki_pos = pos + result;
 889                         spin_lock(&inode->i_lock);
 890                         if (i_size_read(inode) < iocb->ki_pos)
 891                                 i_size_write(inode, iocb->ki_pos);
 892                         spin_unlock(&inode->i_lock);
 893                 }
 894         }
 895         nfs_direct_req_release(dreq);
 896         return result;
 897
 898 out_release:
 899         nfs_direct_req_release(dreq);
 900 out_unlock:
 901         mutex_unlock(&inode->i_mutex);
 902 out:
 903         return result;
 904 }
 905
 906 /**
 907  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
 908  *
 909  */
 910 int __init nfs_init_directcache(void)
 911 {
 912         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
 913                                                 sizeof(struct nfs_direct_req),
 914                                                 0, (SLAB_RECLAIM_ACCOUNT|
 915                                                         SLAB_MEM_SPREAD),
 916                                                 NULL);
 917         if (nfs_direct_cachep == NULL)
 918                 return -ENOMEM;
 919
 920         return 0;
 921 }
 922
 923 /**
 924  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
 925  *
 926  */
 927 void nfs_destroy_directcache(void)
 928 {
 929         kmem_cache_destroy(nfs_direct_cachep);
 930 }