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_select_verf - select the right verifier
 113  * @dreq - direct request possibly spanning multiple servers
 114  * @ds_clp - nfs_client of data server or NULL if MDS / non-pnfs
 115  * @ds_idx - index of data server in data server list, only valid if ds_clp set
 116  *
 117  * returns the correct verifier to use given the role of the server
 118  */
 119 static struct nfs_writeverf *
 120 nfs_direct_select_verf(struct nfs_direct_req *dreq,
 121                        struct nfs_client *ds_clp,
 122                        int ds_idx)
 123 {
 124         struct nfs_writeverf *verfp = &dreq->verf;
 125
 126 #ifdef CONFIG_NFS_V4_1
 127         if (ds_clp) {
 128                 /* pNFS is in use, use the DS verf */
 129                 if (ds_idx >= 0 && ds_idx < dreq->ds_cinfo.nbuckets)
 130                         verfp = &dreq->ds_cinfo.buckets[ds_idx].direct_verf;
 131                 else
 132                         WARN_ON_ONCE(1);
 133         }
 134 #endif
 135         return verfp;
 136 }
 137
 138
 139 /*
 140  * nfs_direct_set_hdr_verf - set the write/commit verifier
 141  * @dreq - direct request possibly spanning multiple servers
 142  * @hdr - pageio header to validate against previously seen verfs
 143  *
 144  * Set the server's (MDS or DS) "seen" verifier
 145  */
 146 static void nfs_direct_set_hdr_verf(struct nfs_direct_req *dreq,
 147                                     struct nfs_pgio_header *hdr)
 148 {
 149         struct nfs_writeverf *verfp;
 150
 151         verfp = nfs_direct_select_verf(dreq, hdr->data->ds_clp,
 152                                       hdr->data->ds_idx);
 153         WARN_ON_ONCE(verfp->committed >= 0);
 154         memcpy(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 155         WARN_ON_ONCE(verfp->committed < 0);
 156 }
 157
 158 /*
 159  * nfs_direct_cmp_hdr_verf - compare verifier for pgio header
 160  * @dreq - direct request possibly spanning multiple servers
 161  * @hdr - pageio header to validate against previously seen verf
 162  *
 163  * set the server's "seen" verf if not initialized.
 164  * returns result of comparison between @hdr->verf and the "seen"
 165  * verf of the server used by @hdr (DS or MDS)
 166  */
 167 static int nfs_direct_set_or_cmp_hdr_verf(struct nfs_direct_req *dreq,
 168                                           struct nfs_pgio_header *hdr)
 169 {
 170         struct nfs_writeverf *verfp;
 171
 172         verfp = nfs_direct_select_verf(dreq, hdr->data->ds_clp,
 173                                          hdr->data->ds_idx);
 174         if (verfp->committed < 0) {
 175                 nfs_direct_set_hdr_verf(dreq, hdr);
 176                 return 0;
 177         }
 178         return memcmp(verfp, &hdr->verf, sizeof(struct nfs_writeverf));
 179 }
 180
 181 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 182 /*
 183  * nfs_direct_cmp_commit_data_verf - compare verifier for commit data
 184  * @dreq - direct request possibly spanning multiple servers
 185  * @data - commit data to validate against previously seen verf
 186  *
 187  * returns result of comparison between @data->verf and the verf of
 188  * the server used by @data (DS or MDS)
 189  */
 190 static int nfs_direct_cmp_commit_data_verf(struct nfs_direct_req *dreq,
 191                                            struct nfs_commit_data *data)
 192 {
 193         struct nfs_writeverf *verfp;
 194
 195         verfp = nfs_direct_select_verf(dreq, data->ds_clp,
 196                                          data->ds_commit_index);
 197         WARN_ON_ONCE(verfp->committed < 0);
 198         return memcmp(verfp, &data->verf, sizeof(struct nfs_writeverf));
 199 }
 200 #endif
 201
 202 /**
 203  * nfs_direct_IO - NFS address space operation for direct I/O
 204  * @rw: direction (read or write)
 205  * @iocb: target I/O control block
 206  * @iov: array of vectors that define I/O buffer
 207  * @pos: offset in file to begin the operation
 208  * @nr_segs: size of iovec array
 209  *
 210  * The presence of this routine in the address space ops vector means
 211  * the NFS client supports direct I/O. However, for most direct IO, we
 212  * shunt off direct read and write requests before the VFS gets them,
 213  * so this method is only ever called for swap.
 214  */
 215 ssize_t nfs_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov, loff_t pos, unsigned long nr_segs)
 216 {
 217 #ifndef CONFIG_NFS_SWAP
 218         dprintk("NFS: nfs_direct_IO (%pD) off/no(%Ld/%lu) EINVAL\n",
 219                         iocb->ki_filp, (long long) pos, nr_segs);
 220
 221         return -EINVAL;
 222 #else
 223         VM_BUG_ON(iocb->ki_nbytes != PAGE_SIZE);
 224
 225         if (rw == READ || rw == KERNEL_READ)
 226                 return nfs_file_direct_read(iocb, iov, nr_segs, pos,
 227                                 rw == READ ? true : false);
 228         return nfs_file_direct_write(iocb, iov, nr_segs, pos,
 229                                 rw == WRITE ? true : false);
 230 #endif /* CONFIG_NFS_SWAP */
 231 }
 232
 233 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
 234 {
 235         unsigned int i;
 236         for (i = 0; i < npages; i++)
 237                 page_cache_release(pages[i]);
 238 }
 239
 240 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
 241                               struct nfs_direct_req *dreq)
 242 {
 243         cinfo->lock = &dreq->lock;
 244         cinfo->mds = &dreq->mds_cinfo;
 245         cinfo->ds = &dreq->ds_cinfo;
 246         cinfo->dreq = dreq;
 247         cinfo->completion_ops = &nfs_direct_commit_completion_ops;
 248 }
 249
 250 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
 251 {
 252         struct nfs_direct_req *dreq;
 253
 254         dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
 255         if (!dreq)
 256                 return NULL;
 257
 258         kref_init(&dreq->kref);
 259         kref_get(&dreq->kref);
 260         init_completion(&dreq->completion);
 261         INIT_LIST_HEAD(&dreq->mds_cinfo.list);
 262         dreq->verf.committed = NFS_INVALID_STABLE_HOW;  /* not set yet */
 263         INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
 264         spin_lock_init(&dreq->lock);
 265
 266         return dreq;
 267 }
 268
 269 static void nfs_direct_req_free(struct kref *kref)
 270 {
 271         struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
 272
 273         if (dreq->l_ctx != NULL)
 274                 nfs_put_lock_context(dreq->l_ctx);
 275         if (dreq->ctx != NULL)
 276                 put_nfs_open_context(dreq->ctx);
 277         kmem_cache_free(nfs_direct_cachep, dreq);
 278 }
 279
 280 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
 281 {
 282         kref_put(&dreq->kref, nfs_direct_req_free);
 283 }
 284
 285 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
 286 {
 287         return dreq->bytes_left;
 288 }
 289 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
 290
 291 /*
 292  * Collects and returns the final error value/byte-count.
 293  */
 294 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
 295 {
 296         ssize_t result = -EIOCBQUEUED;
 297
 298         /* Async requests don't wait here */
 299         if (dreq->iocb)
 300                 goto out;
 301
 302         result = wait_for_completion_killable(&dreq->completion);
 303
 304         if (!result)
 305                 result = dreq->error;
 306         if (!result)
 307                 result = dreq->count;
 308
 309 out:
 310         return (ssize_t) result;
 311 }
 312
 313 /*
 314  * Synchronous I/O uses a stack-allocated iocb.  Thus we can't trust
 315  * the iocb is still valid here if this is a synchronous request.
 316  */
 317 static void nfs_direct_complete(struct nfs_direct_req *dreq, bool write)
 318 {
 319         struct inode *inode = dreq->inode;
 320
 321         if (dreq->iocb && write) {
 322                 loff_t pos = dreq->iocb->ki_pos + dreq->count;
 323
 324                 spin_lock(&inode->i_lock);
 325                 if (i_size_read(inode) < pos)
 326                         i_size_write(inode, pos);
 327                 spin_unlock(&inode->i_lock);
 328         }
 329
 330         if (write)
 331                 nfs_zap_mapping(inode, inode->i_mapping);
 332
 333         inode_dio_done(inode);
 334
 335         if (dreq->iocb) {
 336                 long res = (long) dreq->error;
 337                 if (!res)
 338                         res = (long) dreq->count;
 339                 aio_complete(dreq->iocb, res, 0);
 340         }
 341
 342         complete_all(&dreq->completion);
 343
 344         nfs_direct_req_release(dreq);
 345 }
 346
 347 static void nfs_direct_readpage_release(struct nfs_page *req)
 348 {
 349         dprintk("NFS: direct read done (%s/%llu %d@%lld)\n",
 350                 req->wb_context->dentry->d_inode->i_sb->s_id,
 351                 (unsigned long long)NFS_FILEID(req->wb_context->dentry->d_inode),
 352                 req->wb_bytes,
 353                 (long long)req_offset(req));
 354         nfs_release_request(req);
 355 }
 356
 357 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
 358 {
 359         unsigned long bytes = 0;
 360         struct nfs_direct_req *dreq = hdr->dreq;
 361
 362         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 363                 goto out_put;
 364
 365         spin_lock(&dreq->lock);
 366         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && (hdr->good_bytes == 0))
 367                 dreq->error = hdr->error;
 368         else
 369                 dreq->count += hdr->good_bytes;
 370         spin_unlock(&dreq->lock);
 371
 372         while (!list_empty(&hdr->pages)) {
 373                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 374                 struct page *page = req->wb_page;
 375
 376                 if (!PageCompound(page) && bytes < hdr->good_bytes)
 377                         set_page_dirty(page);
 378                 bytes += req->wb_bytes;
 379                 nfs_list_remove_request(req);
 380                 nfs_direct_readpage_release(req);
 381         }
 382 out_put:
 383         if (put_dreq(dreq))
 384                 nfs_direct_complete(dreq, false);
 385         hdr->release(hdr);
 386 }
 387
 388 static void nfs_read_sync_pgio_error(struct list_head *head)
 389 {
 390         struct nfs_page *req;
 391
 392         while (!list_empty(head)) {
 393                 req = nfs_list_entry(head->next);
 394                 nfs_list_remove_request(req);
 395                 nfs_release_request(req);
 396         }
 397 }
 398
 399 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
 400 {
 401         get_dreq(hdr->dreq);
 402 }
 403
 404 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
 405         .error_cleanup = nfs_read_sync_pgio_error,
 406         .init_hdr = nfs_direct_pgio_init,
 407         .completion = nfs_direct_read_completion,
 408 };
 409
 410 /*
 411  * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
 412  * operation.  If nfs_readdata_alloc() or get_user_pages() fails,
 413  * bail and stop sending more reads.  Read length accounting is
 414  * handled automatically by nfs_direct_read_result().  Otherwise, if
 415  * no requests have been sent, just return an error.
 416  */
 417 static ssize_t nfs_direct_read_schedule_segment(struct nfs_pageio_descriptor *desc,
 418                                                 const struct iovec *iov,
 419                                                 loff_t pos, bool uio)
 420 {
 421         struct nfs_direct_req *dreq = desc->pg_dreq;
 422         struct nfs_open_context *ctx = dreq->ctx;
 423         struct inode *inode = ctx->dentry->d_inode;
 424         unsigned long user_addr = (unsigned long)iov->iov_base;
 425         size_t count = iov->iov_len;
 426         size_t rsize = NFS_SERVER(inode)->rsize;
 427         unsigned int pgbase;
 428         int result;
 429         ssize_t started = 0;
 430         struct page **pagevec = NULL;
 431         unsigned int npages;
 432
 433         do {
 434                 size_t bytes;
 435                 int i;
 436
 437                 pgbase = user_addr & ~PAGE_MASK;
 438                 bytes = min(max_t(size_t, rsize, PAGE_SIZE), count);
 439
 440                 result = -ENOMEM;
 441                 npages = nfs_page_array_len(pgbase, bytes);
 442                 if (!pagevec)
 443                         pagevec = kmalloc(npages * sizeof(struct page *),
 444                                           GFP_KERNEL);
 445                 if (!pagevec)
 446                         break;
 447                 if (uio) {
 448                         down_read(&current->mm->mmap_sem);
 449                         result = get_user_pages(current, current->mm, user_addr,
 450                                         npages, 1, 0, pagevec, NULL);
 451                         up_read(&current->mm->mmap_sem);
 452                         if (result < 0)
 453                                 break;
 454                 } else {
 455                         WARN_ON(npages != 1);
 456                         result = get_kernel_page(user_addr, 1, pagevec);
 457                         if (WARN_ON(result != 1))
 458                                 break;
 459                 }
 460
 461                 if ((unsigned)result < npages) {
 462                         bytes = result * PAGE_SIZE;
 463                         if (bytes <= pgbase) {
 464                                 nfs_direct_release_pages(pagevec, result);
 465                                 break;
 466                         }
 467                         bytes -= pgbase;
 468                         npages = result;
 469                 }
 470
 471                 for (i = 0; i < npages; i++) {
 472                         struct nfs_page *req;
 473                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 474                         /* XXX do we need to do the eof zeroing found in async_filler? */
 475                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 476                                                  pgbase, req_len);
 477                         if (IS_ERR(req)) {
 478                                 result = PTR_ERR(req);
 479                                 break;
 480                         }
 481                         req->wb_index = pos >> PAGE_SHIFT;
 482                         req->wb_offset = pos & ~PAGE_MASK;
 483                         if (!nfs_pageio_add_request(desc, req)) {
 484                                 result = desc->pg_error;
 485                                 nfs_release_request(req);
 486                                 break;
 487                         }
 488                         pgbase = 0;
 489                         bytes -= req_len;
 490                         started += req_len;
 491                         user_addr += req_len;
 492                         pos += req_len;
 493                         count -= req_len;
 494                         dreq->bytes_left -= req_len;
 495                 }
 496                 /* The nfs_page now hold references to these pages */
 497                 nfs_direct_release_pages(pagevec, npages);
 498         } while (count != 0 && result >= 0);
 499
 500         kfree(pagevec);
 501
 502         if (started)
 503                 return started;
 504         return result < 0 ? (ssize_t) result : -EFAULT;
 505 }
 506
 507 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
 508                                               const struct iovec *iov,
 509                                               unsigned long nr_segs,
 510                                               loff_t pos, bool uio)
 511 {
 512         struct nfs_pageio_descriptor desc;
 513         struct inode *inode = dreq->inode;
 514         ssize_t result = -EINVAL;
 515         size_t requested_bytes = 0;
 516         unsigned long seg;
 517
 518         nfs_pageio_init_read(&desc, dreq->inode, false,
 519                              &nfs_direct_read_completion_ops);
 520         get_dreq(dreq);
 521         desc.pg_dreq = dreq;
 522         atomic_inc(&inode->i_dio_count);
 523
 524         for (seg = 0; seg < nr_segs; seg++) {
 525                 const struct iovec *vec = &iov[seg];
 526                 result = nfs_direct_read_schedule_segment(&desc, vec, pos, uio);
 527                 if (result < 0)
 528                         break;
 529                 requested_bytes += result;
 530                 if ((size_t)result < vec->iov_len)
 531                         break;
 532                 pos += vec->iov_len;
 533         }
 534
 535         nfs_pageio_complete(&desc);
 536
 537         /*
 538          * If no bytes were started, return the error, and let the
 539          * generic layer handle the completion.
 540          */
 541         if (requested_bytes == 0) {
 542                 inode_dio_done(inode);
 543                 nfs_direct_req_release(dreq);
 544                 return result < 0 ? result : -EIO;
 545         }
 546
 547         if (put_dreq(dreq))
 548                 nfs_direct_complete(dreq, false);
 549         return 0;
 550 }
 551
 552 /**
 553  * nfs_file_direct_read - file direct read operation for NFS files
 554  * @iocb: target I/O control block
 555  * @iov: vector of user buffers into which to read data
 556  * @nr_segs: size of iov vector
 557  * @pos: byte offset in file where reading starts
 558  *
 559  * We use this function for direct reads instead of calling
 560  * generic_file_aio_read() in order to avoid gfar's check to see if
 561  * the request starts before the end of the file.  For that check
 562  * to work, we must generate a GETATTR before each direct read, and
 563  * even then there is a window between the GETATTR and the subsequent
 564  * READ where the file size could change.  Our preference is simply
 565  * to do all reads the application wants, and the server will take
 566  * care of managing the end of file boundary.
 567  *
 568  * This function also eliminates unnecessarily updating the file's
 569  * atime locally, as the NFS server sets the file's atime, and this
 570  * client must read the updated atime from the server back into its
 571  * cache.
 572  */
 573 ssize_t nfs_file_direct_read(struct kiocb *iocb, const struct iovec *iov,
 574                                 unsigned long nr_segs, loff_t pos, bool uio)
 575 {
 576         struct file *file = iocb->ki_filp;
 577         struct address_space *mapping = file->f_mapping;
 578         struct inode *inode = mapping->host;
 579         struct nfs_direct_req *dreq;
 580         struct nfs_lock_context *l_ctx;
 581         ssize_t result = -EINVAL;
 582         size_t count;
 583
 584         count = iov_length(iov, nr_segs);
 585         nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
 586
 587         dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
 588                 file, count, (long long) pos);
 589
 590         result = 0;
 591         if (!count)
 592                 goto out;
 593
 594         mutex_lock(&inode->i_mutex);
 595         result = nfs_sync_mapping(mapping);
 596         if (result)
 597                 goto out_unlock;
 598
 599         task_io_account_read(count);
 600
 601         result = -ENOMEM;
 602         dreq = nfs_direct_req_alloc();
 603         if (dreq == NULL)
 604                 goto out_unlock;
 605
 606         dreq->inode = inode;
 607         dreq->bytes_left = iov_length(iov, nr_segs);
 608         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
 609         l_ctx = nfs_get_lock_context(dreq->ctx);
 610         if (IS_ERR(l_ctx)) {
 611                 result = PTR_ERR(l_ctx);
 612                 goto out_release;
 613         }
 614         dreq->l_ctx = l_ctx;
 615         if (!is_sync_kiocb(iocb))
 616                 dreq->iocb = iocb;
 617
 618         NFS_I(inode)->read_io += iov_length(iov, nr_segs);
 619         result = nfs_direct_read_schedule_iovec(dreq, iov, nr_segs, pos, uio);
 620
 621         mutex_unlock(&inode->i_mutex);
 622
 623         if (!result) {
 624                 result = nfs_direct_wait(dreq);
 625                 if (result > 0)
 626                         iocb->ki_pos = pos + result;
 627         }
 628
 629         nfs_direct_req_release(dreq);
 630         return result;
 631
 632 out_release:
 633         nfs_direct_req_release(dreq);
 634 out_unlock:
 635         mutex_unlock(&inode->i_mutex);
 636 out:
 637         return result;
 638 }
 639
 640 #if IS_ENABLED(CONFIG_NFS_V3) || IS_ENABLED(CONFIG_NFS_V4)
 641 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
 642 {
 643         struct nfs_pageio_descriptor desc;
 644         struct nfs_page *req, *tmp;
 645         LIST_HEAD(reqs);
 646         struct nfs_commit_info cinfo;
 647         LIST_HEAD(failed);
 648
 649         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 650         pnfs_recover_commit_reqs(dreq->inode, &reqs, &cinfo);
 651         spin_lock(cinfo.lock);
 652         nfs_scan_commit_list(&cinfo.mds->list, &reqs, &cinfo, 0);
 653         spin_unlock(cinfo.lock);
 654
 655         dreq->count = 0;
 656         get_dreq(dreq);
 657
 658         nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
 659                               &nfs_direct_write_completion_ops);
 660         desc.pg_dreq = dreq;
 661
 662         list_for_each_entry_safe(req, tmp, &reqs, wb_list) {
 663                 if (!nfs_pageio_add_request(&desc, req)) {
 664                         nfs_list_remove_request(req);
 665                         nfs_list_add_request(req, &failed);
 666                         spin_lock(cinfo.lock);
 667                         dreq->flags = 0;
 668                         dreq->error = -EIO;
 669                         spin_unlock(cinfo.lock);
 670                 }
 671                 nfs_release_request(req);
 672         }
 673         nfs_pageio_complete(&desc);
 674
 675         while (!list_empty(&failed)) {
 676                 req = nfs_list_entry(failed.next);
 677                 nfs_list_remove_request(req);
 678                 nfs_unlock_and_release_request(req);
 679         }
 680
 681         if (put_dreq(dreq))
 682                 nfs_direct_write_complete(dreq, dreq->inode);
 683 }
 684
 685 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
 686 {
 687         struct nfs_direct_req *dreq = data->dreq;
 688         struct nfs_commit_info cinfo;
 689         struct nfs_page *req;
 690         int status = data->task.tk_status;
 691
 692         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 693         if (status < 0) {
 694                 dprintk("NFS: %5u commit failed with error %d.\n",
 695                         data->task.tk_pid, status);
 696                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 697         } else if (nfs_direct_cmp_commit_data_verf(dreq, data)) {
 698                 dprintk("NFS: %5u commit verify failed\n", data->task.tk_pid);
 699                 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 700         }
 701
 702         dprintk("NFS: %5u commit returned %d\n", data->task.tk_pid, status);
 703         while (!list_empty(&data->pages)) {
 704                 req = nfs_list_entry(data->pages.next);
 705                 nfs_list_remove_request(req);
 706                 if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES) {
 707                         /* Note the rewrite will go through mds */
 708                         nfs_mark_request_commit(req, NULL, &cinfo);
 709                 } else
 710                         nfs_release_request(req);
 711                 nfs_unlock_and_release_request(req);
 712         }
 713
 714         if (atomic_dec_and_test(&cinfo.mds->rpcs_out))
 715                 nfs_direct_write_complete(dreq, data->inode);
 716 }
 717
 718 static void nfs_direct_error_cleanup(struct nfs_inode *nfsi)
 719 {
 720         /* There is no lock to clear */
 721 }
 722
 723 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
 724         .completion = nfs_direct_commit_complete,
 725         .error_cleanup = nfs_direct_error_cleanup,
 726 };
 727
 728 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
 729 {
 730         int res;
 731         struct nfs_commit_info cinfo;
 732         LIST_HEAD(mds_list);
 733
 734         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 735         nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
 736         res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
 737         if (res < 0) /* res == -ENOMEM */
 738                 nfs_direct_write_reschedule(dreq);
 739 }
 740
 741 static void nfs_direct_write_schedule_work(struct work_struct *work)
 742 {
 743         struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
 744         int flags = dreq->flags;
 745
 746         dreq->flags = 0;
 747         switch (flags) {
 748                 case NFS_ODIRECT_DO_COMMIT:
 749                         nfs_direct_commit_schedule(dreq);
 750                         break;
 751                 case NFS_ODIRECT_RESCHED_WRITES:
 752                         nfs_direct_write_reschedule(dreq);
 753                         break;
 754                 default:
 755                         nfs_direct_complete(dreq, true);
 756         }
 757 }
 758
 759 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 760 {
 761         schedule_work(&dreq->work); /* Calls nfs_direct_write_schedule_work */
 762 }
 763
 764 #else
 765 static void nfs_direct_write_schedule_work(struct work_struct *work)
 766 {
 767 }
 768
 769 static void nfs_direct_write_complete(struct nfs_direct_req *dreq, struct inode *inode)
 770 {
 771         nfs_direct_complete(dreq, true);
 772 }
 773 #endif
 774
 775 /*
 776  * NB: Return the value of the first error return code.  Subsequent
 777  *     errors after the first one are ignored.
 778  */
 779 /*
 780  * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
 781  * operation.  If nfs_writedata_alloc() or get_user_pages() fails,
 782  * bail and stop sending more writes.  Write length accounting is
 783  * handled automatically by nfs_direct_write_result().  Otherwise, if
 784  * no requests have been sent, just return an error.
 785  */
 786 static ssize_t nfs_direct_write_schedule_segment(struct nfs_pageio_descriptor *desc,
 787                                                  const struct iovec *iov,
 788                                                  loff_t pos, bool uio)
 789 {
 790         struct nfs_direct_req *dreq = desc->pg_dreq;
 791         struct nfs_open_context *ctx = dreq->ctx;
 792         struct inode *inode = ctx->dentry->d_inode;
 793         unsigned long user_addr = (unsigned long)iov->iov_base;
 794         size_t count = iov->iov_len;
 795         size_t wsize = NFS_SERVER(inode)->wsize;
 796         unsigned int pgbase;
 797         int result;
 798         ssize_t started = 0;
 799         struct page **pagevec = NULL;
 800         unsigned int npages;
 801
 802         do {
 803                 size_t bytes;
 804                 int i;
 805
 806                 pgbase = user_addr & ~PAGE_MASK;
 807                 bytes = min(max_t(size_t, wsize, PAGE_SIZE), count);
 808
 809                 result = -ENOMEM;
 810                 npages = nfs_page_array_len(pgbase, bytes);
 811                 if (!pagevec)
 812                         pagevec = kmalloc(npages * sizeof(struct page *), GFP_KERNEL);
 813                 if (!pagevec)
 814                         break;
 815
 816                 if (uio) {
 817                         down_read(&current->mm->mmap_sem);
 818                         result = get_user_pages(current, current->mm, user_addr,
 819                                                 npages, 0, 0, pagevec, NULL);
 820                         up_read(&current->mm->mmap_sem);
 821                         if (result < 0)
 822                                 break;
 823                 } else {
 824                         WARN_ON(npages != 1);
 825                         result = get_kernel_page(user_addr, 0, pagevec);
 826                         if (WARN_ON(result != 1))
 827                                 break;
 828                 }
 829
 830                 if ((unsigned)result < npages) {
 831                         bytes = result * PAGE_SIZE;
 832                         if (bytes <= pgbase) {
 833                                 nfs_direct_release_pages(pagevec, result);
 834                                 break;
 835                         }
 836                         bytes -= pgbase;
 837                         npages = result;
 838                 }
 839
 840                 for (i = 0; i < npages; i++) {
 841                         struct nfs_page *req;
 842                         unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
 843
 844                         req = nfs_create_request(dreq->ctx, pagevec[i], NULL,
 845                                                  pgbase, req_len);
 846                         if (IS_ERR(req)) {
 847                                 result = PTR_ERR(req);
 848                                 break;
 849                         }
 850                         nfs_lock_request(req);
 851                         req->wb_index = pos >> PAGE_SHIFT;
 852                         req->wb_offset = pos & ~PAGE_MASK;
 853                         if (!nfs_pageio_add_request(desc, req)) {
 854                                 result = desc->pg_error;
 855                                 nfs_unlock_and_release_request(req);
 856                                 break;
 857                         }
 858                         pgbase = 0;
 859                         bytes -= req_len;
 860                         started += req_len;
 861                         user_addr += req_len;
 862                         pos += req_len;
 863                         count -= req_len;
 864                         dreq->bytes_left -= req_len;
 865                 }
 866                 /* The nfs_page now hold references to these pages */
 867                 nfs_direct_release_pages(pagevec, npages);
 868         } while (count != 0 && result >= 0);
 869
 870         kfree(pagevec);
 871
 872         if (started)
 873                 return started;
 874         return result < 0 ? (ssize_t) result : -EFAULT;
 875 }
 876
 877 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
 878 {
 879         struct nfs_direct_req *dreq = hdr->dreq;
 880         struct nfs_commit_info cinfo;
 881         int bit = -1;
 882         struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 883
 884         if (test_bit(NFS_IOHDR_REDO, &hdr->flags))
 885                 goto out_put;
 886
 887         nfs_init_cinfo_from_dreq(&cinfo, dreq);
 888
 889         spin_lock(&dreq->lock);
 890
 891         if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
 892                 dreq->flags = 0;
 893                 dreq->error = hdr->error;
 894         }
 895         if (dreq->error != 0)
 896                 bit = NFS_IOHDR_ERROR;
 897         else {
 898                 dreq->count += hdr->good_bytes;
 899                 if (test_bit(NFS_IOHDR_NEED_RESCHED, &hdr->flags)) {
 900                         dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
 901                         bit = NFS_IOHDR_NEED_RESCHED;
 902                 } else if (test_bit(NFS_IOHDR_NEED_COMMIT, &hdr->flags)) {
 903                         if (dreq->flags == NFS_ODIRECT_RESCHED_WRITES)
 904                                 bit = NFS_IOHDR_NEED_RESCHED;
 905                         else if (dreq->flags == 0) {
 906                                 nfs_direct_set_hdr_verf(dreq, hdr);
 907                                 bit = NFS_IOHDR_NEED_COMMIT;
 908                                 dreq->flags = NFS_ODIRECT_DO_COMMIT;
 909                         } else if (dreq->flags == NFS_ODIRECT_DO_COMMIT) {
 910                                 if (nfs_direct_set_or_cmp_hdr_verf(dreq, hdr)) {
 911                                         dreq->flags =
 912                                                 NFS_ODIRECT_RESCHED_WRITES;
 913                                         bit = NFS_IOHDR_NEED_RESCHED;
 914                                 } else
 915                                         bit = NFS_IOHDR_NEED_COMMIT;
 916                         }
 917                 }
 918         }
 919         spin_unlock(&dreq->lock);
 920
 921         while (!list_empty(&hdr->pages)) {
 922                 bool do_destroy = true;
 923
 924                 req = nfs_list_entry(hdr->pages.next);
 925                 nfs_list_remove_request(req);
 926                 switch (bit) {
 927                 case NFS_IOHDR_NEED_RESCHED:
 928                 case NFS_IOHDR_NEED_COMMIT:
 929                         kref_get(&req->wb_kref);
 930                         nfs_mark_request_commit(req, hdr->lseg, &cinfo);
 931                         do_destroy = false;
 932                 }
 933                 nfs_unlock_and_release_request(req);
 934         }
 935
 936 out_put:
 937         if (put_dreq(dreq))
 938                 nfs_direct_write_complete(dreq, hdr->inode);
 939         hdr->release(hdr);
 940 }
 941
 942 static void nfs_write_sync_pgio_error(struct list_head *head)
 943 {
 944         struct nfs_page *req;
 945
 946         while (!list_empty(head)) {
 947                 req = nfs_list_entry(head->next);
 948                 nfs_list_remove_request(req);
 949                 nfs_unlock_and_release_request(req);
 950         }
 951 }
 952
 953 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
 954         .error_cleanup = nfs_write_sync_pgio_error,
 955         .init_hdr = nfs_direct_pgio_init,
 956         .completion = nfs_direct_write_completion,
 957 };
 958
 959 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
 960                                                const struct iovec *iov,
 961                                                unsigned long nr_segs,
 962                                                loff_t pos, bool uio)
 963 {
 964         struct nfs_pageio_descriptor desc;
 965         struct inode *inode = dreq->inode;
 966         ssize_t result = 0;
 967         size_t requested_bytes = 0;
 968         unsigned long seg;
 969
 970         nfs_pageio_init_write(&desc, inode, FLUSH_COND_STABLE, false,
 971                               &nfs_direct_write_completion_ops);
 972         desc.pg_dreq = dreq;
 973         get_dreq(dreq);
 974         atomic_inc(&inode->i_dio_count);
 975
 976         NFS_I(dreq->inode)->write_io += iov_length(iov, nr_segs);
 977         for (seg = 0; seg < nr_segs; seg++) {
 978                 const struct iovec *vec = &iov[seg];
 979                 result = nfs_direct_write_schedule_segment(&desc, vec, pos, uio);
 980                 if (result < 0)
 981                         break;
 982                 requested_bytes += result;
 983                 if ((size_t)result < vec->iov_len)
 984                         break;
 985                 pos += vec->iov_len;
 986         }
 987         nfs_pageio_complete(&desc);
 988
 989         /*
 990          * If no bytes were started, return the error, and let the
 991          * generic layer handle the completion.
 992          */
 993         if (requested_bytes == 0) {
 994                 inode_dio_done(inode);
 995                 nfs_direct_req_release(dreq);
 996                 return result < 0 ? result : -EIO;
 997         }
 998
 999         if (put_dreq(dreq))
1000                 nfs_direct_write_complete(dreq, dreq->inode);
1001         return 0;
1002 }
1003
1004 /**
1005  * nfs_file_direct_write - file direct write operation for NFS files
1006  * @iocb: target I/O control block
1007  * @iov: vector of user buffers from which to write data
1008  * @nr_segs: size of iov vector
1009  * @pos: byte offset in file where writing starts
1010  *
1011  * We use this function for direct writes instead of calling
1012  * generic_file_aio_write() in order to avoid taking the inode
1013  * semaphore and updating the i_size.  The NFS server will set
1014  * the new i_size and this client must read the updated size
1015  * back into its cache.  We let the server do generic write
1016  * parameter checking and report problems.
1017  *
1018  * We eliminate local atime updates, see direct read above.
1019  *
1020  * We avoid unnecessary page cache invalidations for normal cached
1021  * readers of this file.
1022  *
1023  * Note that O_APPEND is not supported for NFS direct writes, as there
1024  * is no atomic O_APPEND write facility in the NFS protocol.
1025  */
1026 ssize_t nfs_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1027                                 unsigned long nr_segs, loff_t pos, bool uio)
1028 {
1029         ssize_t result = -EINVAL;
1030         struct file *file = iocb->ki_filp;
1031         struct address_space *mapping = file->f_mapping;
1032         struct inode *inode = mapping->host;
1033         struct nfs_direct_req *dreq;
1034         struct nfs_lock_context *l_ctx;
1035         loff_t end;
1036         size_t count;
1037
1038         count = iov_length(iov, nr_segs);
1039         end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1040
1041         nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
1042
1043         dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
1044                 file, count, (long long) pos);
1045
1046         result = generic_write_checks(file, &pos, &count, 0);
1047         if (result)
1048                 goto out;
1049
1050         result = -EINVAL;
1051         if ((ssize_t) count < 0)
1052                 goto out;
1053         result = 0;
1054         if (!count)
1055                 goto out;
1056
1057         mutex_lock(&inode->i_mutex);
1058
1059         result = nfs_sync_mapping(mapping);
1060         if (result)
1061                 goto out_unlock;
1062
1063         if (mapping->nrpages) {
1064                 result = invalidate_inode_pages2_range(mapping,
1065                                         pos >> PAGE_CACHE_SHIFT, end);
1066                 if (result)
1067                         goto out_unlock;
1068         }
1069
1070         task_io_account_write(count);
1071
1072         result = -ENOMEM;
1073         dreq = nfs_direct_req_alloc();
1074         if (!dreq)
1075                 goto out_unlock;
1076
1077         dreq->inode = inode;
1078         dreq->bytes_left = count;
1079         dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
1080         l_ctx = nfs_get_lock_context(dreq->ctx);
1081         if (IS_ERR(l_ctx)) {
1082                 result = PTR_ERR(l_ctx);
1083                 goto out_release;
1084         }
1085         dreq->l_ctx = l_ctx;
1086         if (!is_sync_kiocb(iocb))
1087                 dreq->iocb = iocb;
1088
1089         result = nfs_direct_write_schedule_iovec(dreq, iov, nr_segs, pos, uio);
1090
1091         if (mapping->nrpages) {
1092                 invalidate_inode_pages2_range(mapping,
1093                                               pos >> PAGE_CACHE_SHIFT, end);
1094         }
1095
1096         mutex_unlock(&inode->i_mutex);
1097
1098         if (!result) {
1099                 result = nfs_direct_wait(dreq);
1100                 if (result > 0) {
1101                         struct inode *inode = mapping->host;
1102
1103                         iocb->ki_pos = pos + result;
1104                         spin_lock(&inode->i_lock);
1105                         if (i_size_read(inode) < iocb->ki_pos)
1106                                 i_size_write(inode, iocb->ki_pos);
1107                         spin_unlock(&inode->i_lock);
1108                 }
1109         }
1110         nfs_direct_req_release(dreq);
1111         return result;
1112
1113 out_release:
1114         nfs_direct_req_release(dreq);
1115 out_unlock:
1116         mutex_unlock(&inode->i_mutex);
1117 out:
1118         return result;
1119 }
1120
1121 /**
1122  * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1123  *
1124  */
1125 int __init nfs_init_directcache(void)
1126 {
1127         nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1128                                                 sizeof(struct nfs_direct_req),
1129                                                 0, (SLAB_RECLAIM_ACCOUNT|
1130                                                         SLAB_MEM_SPREAD),
1131                                                 NULL);
1132         if (nfs_direct_cachep == NULL)
1133                 return -ENOMEM;
1134
1135         return 0;
1136 }
1137
1138 /**
1139  * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1140  *
1141  */
1142 void nfs_destroy_directcache(void)
1143 {
1144         kmem_cache_destroy(nfs_direct_cachep);
1145 }